Hermes parser: clarify string arguments comment (JSON vs plain)

Previously _truncate_context() mutated the shared messages list, which could drop older turns and break reward computation/debugging.

Now we keep messages as the full trajectory and apply truncation to a copy (prompt_messages) for each model call.

.clinerules

@@ -0,0 +1,115 @@
+# Cline's Memory Bank
+
+I am Cline, an expert software engineer with a unique characteristic: my memory resets completely between sessions. This isn't a limitation - it's what drives me to maintain perfect documentation. After each reset, I rely ENTIRELY on my Memory Bank to understand the project and continue work effectively. I MUST read ALL memory bank files at the start of EVERY task - this is not optional.
+
+## Memory Bank Structure
+
+The Memory Bank consists of core files and optional context files, all in Markdown format. Files build upon each other in a clear hierarchy:
+
+flowchart TD
+    PB[projectbrief.md] --> PC[productContext.md]
+    PB --> SP[systemPatterns.md]
+    PB --> TC[techContext.md]
+
+    PC --> AC[activeContext.md]
+    SP --> AC
+    TC --> AC
+
+    AC --> P[progress.md]
+
+### Core Files (Required)
+1. `projectbrief.md`
+   - Foundation document that shapes all other files
+   - Created at project start if it doesn't exist
+   - Defines core requirements and goals
+   - Source of truth for project scope
+
+2. `productContext.md`
+   - Why this project exists
+   - Problems it solves
+   - How it should work
+   - User experience goals
+
+3. `activeContext.md`
+   - Current work focus
+   - Recent changes
+   - Next steps
+   - Active decisions and considerations
+   - Important patterns and preferences
+   - Learnings and project insights
+
+4. `systemPatterns.md`
+   - System architecture
+   - Key technical decisions
+   - Design patterns in use
+   - Component relationships
+   - Critical implementation paths
+
+5. `techContext.md`
+   - Technologies used
+   - Development setup
+   - Technical constraints
+   - Dependencies
+   - Tool usage patterns
+
+6. `progress.md`
+   - What works
+   - What's left to build
+   - Current status
+   - Known issues
+   - Evolution of project decisions
+
+### Additional Context
+Create additional files/folders within memory-bank/ when they help organize:
+- Complex feature documentation
+- Integration specifications
+- API documentation
+- Testing strategies
+- Deployment procedures
+
+## Core Workflows
+
+### Plan Mode
+flowchart TD
+    Start[Start] --> ReadFiles[Read Memory Bank]
+    ReadFiles --> CheckFiles{Files Complete?}
+
+    CheckFiles -->|No| Plan[Create Plan]
+    Plan --> Document[Document in Chat]
+
+    CheckFiles -->|Yes| Verify[Verify Context]
+    Verify --> Strategy[Develop Strategy]
+    Strategy --> Present[Present Approach]
+
+### Act Mode
+flowchart TD
+    Start[Start] --> Context[Check Memory Bank]
+    Context --> Update[Update Documentation]
+    Update --> Execute[Execute Task]
+    Execute --> Document[Document Changes]
+
+## Documentation Updates
+
+Memory Bank updates occur when:
+1. Discovering new project patterns
+2. After implementing significant changes
+3. When user requests with **update memory bank** (MUST review ALL files)
+4. When context needs clarification
+
+flowchart TD
+    Start[Update Process]
+
+    subgraph Process
+        P1[Review ALL Files]
+        P2[Document Current State]
+        P3[Clarify Next Steps]
+        P4[Document Insights & Patterns]
+
+        P1 --> P2 --> P3 --> P4
+    end
+
+    Start --> Process
+
+Note: When triggered by **update memory bank**, I MUST review every memory bank file, even if some don't require updates. Focus particularly on activeContext.md and progress.md as they track current state.
+
+REMEMBER: After every memory reset, I begin completely fresh. The Memory Bank is my only link to previous work. It must be maintained with precision and clarity, as my effectiveness depends entirely on its accuracy.

.dockerignore

@@ -1,15 +0,0 @@
-# Git
-.git
-.gitignore
-.gitmodules
-
-# Dependencies
-node_modules
-
-# CI/CD
-.github
-
-# Environment files
-.env
-
-*.md

.env.example

@@ -1,125 +1,110 @@
 # Hermes Agent Environment Configuration
 # Copy this file to .env and fill in your API keys
 
+# =============================================================================
+# CORE SETTINGS
+# =============================================================================
+# Agent backend:
+# - openai  : default Hermes-Agent loop (OpenAI function-calling via OpenAI SDK)
+# - atropos : Atroposlib ServerManager/ManagedServer-backed loop (training/env integration)
+HERMES_BACKEND=openai
+
+
+# =============================================================================
+# LOCAL / SELF-HOSTED OPENAI-COMPATIBLE ENDPOINTS (vLLM, SGLang, llama.cpp, etc.)
+# =============================================================================
+# For local development (matches the Atropos test env defaults):
+# ATROPOS_SERVER_BASE_URL=http://127.0.0.1:8080
+# ATROPOS_SERVER_MODEL=hermes-4-36b
+# For hosted inference (Nous Research inference API):
+ATROPOS_SERVER_BASE_URL=
+ATROPOS_SERVER_MODEL=
+ATROPOS_TOKENIZER_NAME=
+# Set this to your Nous API key (Bearer token).
+ATROPOS_SERVER_API_KEY=
+
+# Debugging (prints to stdout; use with care)
+# HERMES_DEBUG_ATROPOS_REQUEST=1
+# HERMES_DEBUG_ATROPOS_RESPONSE=1
+# HERMES_DEBUG_OPENAI_REQUEST=1
+# HERMES_DEBUG_OPENAI_RESPONSE=1
+
+# =============================================================================
+# LOCAL / SELF-HOSTED OPENAI-COMPATIBLE ENDPOINTS (vLLM, SGLang, llama.cpp, etc.)
+# =============================================================================
+# If you set ATROPOS_SERVER_BASE_URL or OPENAI_BASE_URL, Hermes will use it instead
+# of OpenRouter.
+#
+# Local server convenience (base URL without /v1):
+# llama.cpp example (see `Hermes-Agent/scripts/launch_llama_cpp_hermes_4_36b.sh`):
+# ATROPOS_SERVER_BASE_URL=http://127.0.0.1:8080
+# ATROPOS_SERVER_MODEL=hermes-4-36b
+# ATROPOS_TOKENIZER_NAME=NousResearch/Hermes-4.3-36B
+# ATROPOS_SERVER_API_KEY=local
+#
+# Hosted Nous inference API:
+# ATROPOS_SERVER_BASE_URL=https://inference-api.nousresearch.com
+# ATROPOS_SERVER_MODEL=Hermes-4.3-36B
+# ATROPOS_TOKENIZER_NAME=NousResearch/Hermes-4.3-36B
+# ATROPOS_SERVER_API_KEY=sk-... (Bearer token)
+#
+# If you plan to run GRPO-style group sampling (e.g. `--env.group_size 4`) against
+# llama.cpp, start the server with at least that many slots, e.g.:
+#   LLAMA_CPP_PARALLEL=4 Hermes-Agent/scripts/launch_llama_cpp_hermes_4_36b.sh
+#
+# Generic OpenAI-compatible (base URL should include /v1):
+# OPENAI_BASE_URL=http://127.0.0.1:8080/v1
+# OPENAI_API_KEY=local
+
 # =============================================================================
 # LLM PROVIDER (OpenRouter)
 # =============================================================================
 # OpenRouter provides access to many models through one API
 # All LLM calls go through OpenRouter - no direct provider keys needed
 # Get your key at: https://openrouter.ai/keys
-# OPENROUTER_API_KEY=
+OPENROUTER_BASE_URL=https://openrouter.ai/api/v1
+OPENROUTER_API_KEY=
 
-# Default model is configured in ~/.hermes/config.yaml (model.default).
-# Use 'hermes model' or 'hermes setup' to change it.
-# LLM_MODEL is no longer read from .env — this line is kept for reference only.
-# LLM_MODEL=anthropic/claude-opus-4.6
-
-# =============================================================================
-# LLM PROVIDER (z.ai / GLM)
-# =============================================================================
-# z.ai provides access to ZhipuAI GLM models (GLM-4-Plus, etc.)
-# Get your key at: https://z.ai or https://open.bigmodel.cn
-# GLM_API_KEY=
-# GLM_BASE_URL=https://api.z.ai/api/paas/v4  # Override default base URL
-
-# =============================================================================
-# LLM PROVIDER (Kimi / Moonshot)
-# =============================================================================
-# Kimi Code provides access to Moonshot AI coding models (kimi-k2.5, etc.)
-# Get your key at: https://platform.kimi.ai (Kimi Code console)
-# Keys prefixed sk-kimi- use the Kimi Code API (api.kimi.com) by default.
-# Legacy keys from platform.moonshot.ai need KIMI_BASE_URL override below.
-# KIMI_API_KEY=
-# KIMI_BASE_URL=https://api.kimi.com/coding/v1  # Default for sk-kimi- keys
-# KIMI_BASE_URL=https://api.moonshot.ai/v1      # For legacy Moonshot keys
-# KIMI_BASE_URL=https://api.moonshot.cn/v1       # For Moonshot China keys
-
-# =============================================================================
-# LLM PROVIDER (MiniMax)
-# =============================================================================
-# MiniMax provides access to MiniMax models (global endpoint)
-# Get your key at: https://www.minimax.io
-# MINIMAX_API_KEY=
-# MINIMAX_BASE_URL=https://api.minimax.io/v1  # Override default base URL
-
-# MiniMax China endpoint (for users in mainland China)
-# MINIMAX_CN_API_KEY=
-# MINIMAX_CN_BASE_URL=https://api.minimaxi.com/v1  # Override default base URL
-
-# =============================================================================
-# LLM PROVIDER (OpenCode Zen)
-# =============================================================================
-# OpenCode Zen provides curated, tested models (GPT, Claude, Gemini, MiniMax, GLM, Kimi)
-# Pay-as-you-go pricing. Get your key at: https://opencode.ai/auth
-# OPENCODE_ZEN_API_KEY=
-# OPENCODE_ZEN_BASE_URL=https://opencode.ai/zen/v1  # Override default base URL
-
-# =============================================================================
-# LLM PROVIDER (OpenCode Go)
-# =============================================================================
-# OpenCode Go provides access to open models (GLM-5, Kimi K2.5, MiniMax M2.5)
-# $10/month subscription. Get your key at: https://opencode.ai/auth
-# OPENCODE_GO_API_KEY=
-
-# =============================================================================
-# LLM PROVIDER (Hugging Face Inference Providers)
-# =============================================================================
-# Hugging Face routes to 20+ open models via unified OpenAI-compatible endpoint.
-# Free tier included ($0.10/month), no markup on provider rates.
-# Get your token at: https://huggingface.co/settings/tokens
-# Required permission: "Make calls to Inference Providers"
-# HF_TOKEN=
-# OPENCODE_GO_BASE_URL=https://opencode.ai/zen/go/v1  # Override default base URL
+# Default model to use (OpenRouter format: provider/model)
+# Examples: anthropic/claude-opus-4.6, openai/gpt-4o, google/gemini-2.0-flash, zhipuai/glm-4-plus
+LLM_MODEL=anthropic/claude-opus-4.6
 
 # =============================================================================
 # TOOL API KEYS
 # =============================================================================
 
-# Exa API Key - AI-native web search and contents
-# Get at: https://exa.ai
-# EXA_API_KEY=
-
-# Parallel API Key - AI-native web search and extract
-# Get at: https://parallel.ai
-# PARALLEL_API_KEY=
-
 # Firecrawl API Key - Web search, extract, and crawl
 # Get at: https://firecrawl.dev/
-# FIRECRAWL_API_KEY=
+FIRECRAWL_API_KEY=
 
+# Nous Research API Key - Vision analysis and multi-model reasoning
+# Get at: https://inference-api.nousresearch.com/
+NOUS_API_KEY=
 
 # FAL.ai API Key - Image generation
 # Get at: https://fal.ai/
-# FAL_KEY=
-
-# Honcho - Cross-session AI-native user modeling (optional)
-# Builds a persistent understanding of the user across sessions and tools.
-# Get at: https://app.honcho.dev
-# Also requires ~/.honcho/config.json with enabled=true (see README).
-# HONCHO_API_KEY=
+FAL_KEY=
 
 # =============================================================================
-# TERMINAL TOOL CONFIGURATION
+# TERMINAL TOOL CONFIGURATION (mini-swe-agent backend)
 # =============================================================================
 # Backend type: "local", "singularity", "docker", "modal", or "ssh"
-# Terminal backend is configured in ~/.hermes/config.yaml (terminal.backend).
-# Use 'hermes setup' or 'hermes config set terminal.backend docker' to change.
-# Supported: local, docker, singularity, modal, ssh
-#
-# Only override here if you need to force a backend without touching config.yaml:
-# TERMINAL_ENV=local
+# - local: Runs directly on your machine (fastest, no isolation)
+# - ssh: Runs on remote server via SSH (great for sandboxing - agent can't touch its own code)
+# - singularity: Runs in Apptainer/Singularity containers (HPC clusters, no root needed)
+# - docker: Runs in Docker containers (isolated, requires Docker + docker group)
+# - modal: Runs in Modal cloud sandboxes (scalable, requires Modal account)
+TERMINAL_ENV=local
+
 
 # Container images (for singularity/docker/modal backends)
-# TERMINAL_DOCKER_IMAGE=nikolaik/python-nodejs:python3.11-nodejs20
-# TERMINAL_SINGULARITY_IMAGE=docker://nikolaik/python-nodejs:python3.11-nodejs20
-TERMINAL_MODAL_IMAGE=nikolaik/python-nodejs:python3.11-nodejs20
-
+TERMINAL_DOCKER_IMAGE=python:3.11
+TERMINAL_SINGULARITY_IMAGE=docker://python:3.11
+TERMINAL_MODAL_IMAGE=python:3.11
 
 # Working directory for terminal commands
-# For local backend: "." means current directory (resolved automatically)
-# For remote backends (ssh/docker/modal/singularity): use an absolute path
-#   INSIDE the target environment, or leave unset for the backend's default
-#   (/root for modal, / for docker, ~ for ssh). Do NOT use a host-local path.
+# For CLI: "." means current directory (resolved automatically from config.yaml)
+# For containers (docker/singularity/modal): absolute path inside the container
 # Usually managed by config.yaml (terminal.cwd) — uncomment to override
 # TERMINAL_CWD=.
 
@@ -163,12 +148,87 @@ TERMINAL_LIFETIME_SECONDS=300
 # SUDO_PASSWORD=your_password_here
 
 # =============================================================================
-# MODAL CLOUD BACKEND (Optional - for TERMINAL_ENV=modal)
+# MODAL CLOUD BACKEND (for TERMINAL_ENV=modal)
 # =============================================================================
-# Modal uses CLI authentication, not environment variables.
-# Run: pip install modal && modal setup
-# This will authenticate via browser and store credentials locally.
-# No API key needed in .env - Modal handles auth automatically.
+# Modal provides cloud sandboxes with per-second billing and auto-scaling.
+# This implementation uses a warm pool of sandboxes for cost efficiency.
+#
+# SETUP:
+#   pip install modal && modal setup
+#   (Authenticates via browser, stores credentials locally)
+#
+# FEATURES:
+# - Auto-scaling warm sandbox pool (no cold start after first use)
+# - Named sandbox recovery (reconnects after restart)
+# - Profile-based heterogeneous environments (CPU, GPU, different images)
+# - Server-side idle_timeout protection against orphaned sandboxes
+
+# Modal app name (groups all sandboxes, used for recovery)
+TERMINAL_MODAL_APP_NAME=hermes-sandbox
+
+# Default profile when none specified
+TERMINAL_MODAL_DEFAULT_PROFILE=default
+
+# Profile config file (optional - YAML format, see modal_profiles.yaml)
+# TERMINAL_MODAL_PROFILES_FILE=modal_profiles.yaml
+
+# --- Default Profile Settings (used if no YAML file) ---
+# These apply when no profile is specified or for the "default" profile
+TERMINAL_MODAL_IMAGE=python:3.11
+TERMINAL_MODAL_MIN_POOL=1
+TERMINAL_MODAL_MAX_POOL=5
+TERMINAL_MODAL_IDLE_TIMEOUT=120
+TERMINAL_MODAL_MAX_LIFETIME=3600
+TERMINAL_MODAL_SCALE_DOWN_IDLE=180
+
+# --- Custom Profile Example: pytorch-gpu ---
+# Uncomment to enable a GPU profile for ML tasks
+# Usage: terminal_tool("python train.py", profile="pytorch-gpu")
+#
+# TERMINAL_MODAL_PROFILE_pytorch_gpu_IMAGE=pytorch/pytorch:2.1.0-cuda12.1-cudnn8-runtime
+# TERMINAL_MODAL_PROFILE_pytorch_gpu_GPU=T4
+# TERMINAL_MODAL_PROFILE_pytorch_gpu_MEMORY=16384
+# TERMINAL_MODAL_PROFILE_pytorch_gpu_MIN_POOL=0
+# TERMINAL_MODAL_PROFILE_pytorch_gpu_MAX_POOL=2
+# TERMINAL_MODAL_PROFILE_pytorch_gpu_IDLE_TIMEOUT=60
+
+# --- Custom Profile Example: node ---
+# Uncomment to enable a Node.js profile
+# Usage: terminal_tool("npm test", profile="node")
+#
+# TERMINAL_MODAL_PROFILE_node_IMAGE=node:18
+# TERMINAL_MODAL_PROFILE_node_MIN_POOL=0
+# TERMINAL_MODAL_PROFILE_node_MAX_POOL=3
+
+# =============================================================================
+# MODAL SECRETS (Secure credential injection)
+# =============================================================================
+# Modal Secrets allow you to securely pass API keys, passwords, and other
+# sensitive data to your sandboxes without exposing them in code or logs.
+#
+# SETUP SECRETS:
+#   1. Via Dashboard: https://modal.com/secrets
+#   2. Via CLI: modal secret create my-secret KEY1=value1 KEY2=value2
+#   3. Via CLI with env: modal secret create my-secret API_KEY="$API_KEY"
+#
+# LIST SECRETS:
+#   modal secret list
+#
+# DELETE SECRETS:
+#   modal secret delete my-secret
+
+# Global secrets applied to ALL profiles (comma-separated secret names)
+# These secrets must be created on Modal dashboard or via CLI first
+# TERMINAL_MODAL_SECRETS=my-api-keys,database-creds
+
+# Per-profile secrets (comma-separated secret names)
+# TERMINAL_MODAL_PROFILE_pytorch_gpu_SECRETS=huggingface-token,wandb-key
+
+# Per-profile environment variables (semicolon-separated KEY=VALUE pairs)
+# TERMINAL_MODAL_PROFILE_default_ENV_VARS=DEBUG=1;LOG_LEVEL=info
+
+# Load local .env file into sandbox (useful for development)
+# TERMINAL_MODAL_PROFILE_default_USE_DOTENV=true
 
 # =============================================================================
 # BROWSER TOOL CONFIGURATION (agent-browser + Browserbase)
@@ -182,10 +242,10 @@ TERMINAL_LIFETIME_SECONDS=300
 
 # Browserbase API Key - Cloud browser execution
 # Get at: https://browserbase.com/
-# BROWSERBASE_API_KEY=
+BROWSERBASE_API_KEY=
 
 # Browserbase Project ID - From your Browserbase dashboard
-# BROWSERBASE_PROJECT_ID=
+BROWSERBASE_PROJECT_ID=
 
 # Enable residential proxies for better CAPTCHA solving (default: true)
 # Routes traffic through residential IPs, significantly improves success rate
@@ -211,70 +271,16 @@ BROWSER_INACTIVITY_TIMEOUT=120
 # Contains full conversation history in trajectory format for debugging/replay
 
 # =============================================================================
-# VOICE TRANSCRIPTION & OPENAI TTS
+# LEGACY/OPTIONAL API KEYS
 # =============================================================================
-# Required for voice message transcription (Whisper) and OpenAI TTS voices.
-# Uses OpenAI's API directly (not via OpenRouter).
-# Named VOICE_TOOLS_OPENAI_KEY to avoid interference with OpenRouter.
-# Get at: https://platform.openai.com/api-keys
-# VOICE_TOOLS_OPENAI_KEY=
 
-# =============================================================================
-# SLACK INTEGRATION
-# =============================================================================
-# Slack Bot Token - From Slack App settings (OAuth & Permissions)
-# Get at: https://api.slack.com/apps
-# SLACK_BOT_TOKEN=xoxb-...
+# Morph API Key - For legacy Hecate terminal backend (terminal-hecate tool)
+# Get at: https://morph.so/
+MORPH_API_KEY=
 
-# Slack App Token - For Socket Mode (App-Level Tokens in Slack App settings)
-# SLACK_APP_TOKEN=xapp-...
-
-# Slack allowed users (comma-separated Slack user IDs)
-# SLACK_ALLOWED_USERS=
-
-# =============================================================================
-# TELEGRAM INTEGRATION
-# =============================================================================
-# Telegram Bot Token - From @BotFather (https://t.me/BotFather)
-# TELEGRAM_BOT_TOKEN=
-# TELEGRAM_ALLOWED_USERS=                  # Comma-separated user IDs
-# TELEGRAM_HOME_CHANNEL=                   # Default chat for cron delivery
-# TELEGRAM_HOME_CHANNEL_NAME=              # Display name for home channel
-
-# Webhook mode (optional — for cloud deployments like Fly.io/Railway)
-# Default is long polling. Setting TELEGRAM_WEBHOOK_URL switches to webhook mode.
-# TELEGRAM_WEBHOOK_URL=https://my-app.fly.dev/telegram
-# TELEGRAM_WEBHOOK_PORT=8443
-# TELEGRAM_WEBHOOK_SECRET=                 # Recommended for production
-
-# WhatsApp (built-in Baileys bridge — run `hermes whatsapp` to pair)
-# WHATSAPP_ENABLED=false
-# WHATSAPP_ALLOWED_USERS=15551234567
-
-# Email (IMAP/SMTP — send and receive emails as Hermes)
-# For Gmail: enable 2FA → create App Password at https://myaccount.google.com/apppasswords
-# EMAIL_ADDRESS=hermes@gmail.com
-# EMAIL_PASSWORD=xxxx xxxx xxxx xxxx
-# EMAIL_IMAP_HOST=imap.gmail.com
-# EMAIL_IMAP_PORT=993
-# EMAIL_SMTP_HOST=smtp.gmail.com
-# EMAIL_SMTP_PORT=587
-# EMAIL_POLL_INTERVAL=15
-# EMAIL_ALLOWED_USERS=your@email.com
-# EMAIL_HOME_ADDRESS=your@email.com
-
-# Gateway-wide: allow ALL users without an allowlist (default: false = deny)
-# Only set to true if you intentionally want open access.
-# GATEWAY_ALLOW_ALL_USERS=false
-
-# =============================================================================
-# RESPONSE PACING
-# =============================================================================
-# Human-like delays between message chunks on messaging platforms.
-# Makes the bot feel less robotic.
-# HERMES_HUMAN_DELAY_MODE=off     # off | natural | custom
-# HERMES_HUMAN_DELAY_MIN_MS=800   # Min delay in ms (custom mode)
-# HERMES_HUMAN_DELAY_MAX_MS=2500  # Max delay in ms (custom mode)
+# Hecate VM Settings (only if using terminal-hecate tool)
+HECATE_VM_LIFETIME_SECONDS=300
+HECATE_DEFAULT_SNAPSHOT_ID=snapshot_p5294qxt
 
 # =============================================================================
 # DEBUG OPTIONS
@@ -290,10 +296,9 @@ IMAGE_TOOLS_DEBUG=false
 # When conversation approaches model's context limit, middle turns are
 # automatically summarized to free up space.
 #
-# Context compression is configured in ~/.hermes/config.yaml under compression:
 # CONTEXT_COMPRESSION_ENABLED=true        # Enable auto-compression (default: true)
 # CONTEXT_COMPRESSION_THRESHOLD=0.85      # Compress at 85% of context limit
-# Model is set via compression.summary_model in config.yaml (default: google/gemini-3-flash-preview)
+# CONTEXT_COMPRESSION_MODEL=google/gemini-2.0-flash-001  # Fast model for summaries
 
 # =============================================================================
 # RL TRAINING (Tinker + Atropos)
@@ -303,49 +308,12 @@ IMAGE_TOOLS_DEBUG=false
 
 # Tinker API Key - RL training service
 # Get at: https://tinker-console.thinkingmachines.ai/keys
-# TINKER_API_KEY=
+TINKER_API_KEY=
 
 # Weights & Biases API Key - Experiment tracking and metrics
 # Get at: https://wandb.ai/authorize
-# WANDB_API_KEY=
+WANDB_API_KEY=
 
 # RL API Server URL (default: http://localhost:8080)
 # Change if running the rl-server on a different host/port
 # RL_API_URL=http://localhost:8080
-
-# =============================================================================
-# SKILLS HUB (GitHub integration for skill search/install/publish)
-# =============================================================================
-
-# GitHub Personal Access Token — for higher API rate limits on skill search/install
-# Get at: https://github.com/settings/tokens (Fine-grained recommended)
-# GITHUB_TOKEN=ghp_xxxxxxxxxxxxxxxxxxxx
-
-# GitHub App credentials (optional — for bot identity on PRs)
-# GITHUB_APP_ID=
-# GITHUB_APP_PRIVATE_KEY_PATH=
-# GITHUB_APP_INSTALLATION_ID=
-
-# Groq API key (free tier — used for Whisper STT in voice mode)
-# GROQ_API_KEY=
-
-# =============================================================================
-# STT PROVIDER SELECTION
-# =============================================================================
-# Default STT provider is "local" (faster-whisper) — runs on your machine, no API key needed.
-# Install with: pip install faster-whisper
-# Model downloads automatically on first use (~150 MB for "base").
-# To use cloud providers instead, set GROQ_API_KEY or VOICE_TOOLS_OPENAI_KEY above.
-# Provider priority: local > groq > openai
-# Configure in config.yaml: stt.provider: local | groq | openai
-
-# =============================================================================
-# STT ADVANCED OVERRIDES (optional)
-# =============================================================================
-# Override default STT models per provider (normally set via stt.model in config.yaml)
-# STT_GROQ_MODEL=whisper-large-v3-turbo
-# STT_OPENAI_MODEL=whisper-1
-
-# Override STT provider endpoints (for proxies or self-hosted instances)
-# GROQ_BASE_URL=https://api.groq.com/openai/v1
-# STT_OPENAI_BASE_URL=https://api.openai.com/v1

.envrc

@@ -1 +0,0 @@
-use flake

.github/ISSUE_TEMPLATE/bug_report.yml

@@ -1,144 +0,0 @@
-name: "🐛 Bug Report"
-description: Report a bug — something that's broken, crashes, or behaves incorrectly.
-title: "[Bug]: "
-labels: ["bug"]
-body:
-  - type: markdown
-    attributes:
-      value: |
-        Thanks for reporting a bug! Please fill out the sections below so we can reproduce and fix it quickly.
-
-        **Before submitting**, please:
-        - [ ] Search [existing issues](https://github.com/NousResearch/hermes-agent/issues) to avoid duplicates
-        - [ ] Update to the latest version (`hermes update`) and confirm the bug still exists
-
-  - type: textarea
-    id: description
-    attributes:
-      label: Bug Description
-      description: A clear description of what's broken. Include error messages, tracebacks, or screenshots if relevant.
-      placeholder: |
-        What happened? What did you expect to happen instead?
-    validations:
-      required: true
-
-  - type: textarea
-    id: reproduction
-    attributes:
-      label: Steps to Reproduce
-      description: Minimal steps to trigger the bug. The more specific, the faster we can fix it.
-      placeholder: |
-        1. Run `hermes chat`
-        2. Send the message "..."
-        3. Agent calls tool X
-        4. Error appears: ...
-    validations:
-      required: true
-
-  - type: textarea
-    id: expected
-    attributes:
-      label: Expected Behavior
-      description: What should have happened instead?
-    validations:
-      required: true
-
-  - type: textarea
-    id: actual
-    attributes:
-      label: Actual Behavior
-      description: What actually happened? Include full error output if available.
-    validations:
-      required: true
-
-  - type: dropdown
-    id: component
-    attributes:
-      label: Affected Component
-      description: Which part of Hermes is affected?
-      multiple: true
-      options:
-        - CLI (interactive chat)
-        - Gateway (Telegram/Discord/Slack/WhatsApp)
-        - Setup / Installation
-        - Tools (terminal, file ops, web, code execution, etc.)
-        - Skills (skill loading, skill hub, skill guard)
-        - Agent Core (conversation loop, context compression, memory)
-        - Configuration (config.yaml, .env, hermes setup)
-        - Other
-    validations:
-      required: true
-
-  - type: dropdown
-    id: platform
-    attributes:
-      label: Messaging Platform (if gateway-related)
-      description: Which platform adapter is affected?
-      multiple: true
-      options:
-        - N/A (CLI only)
-        - Telegram
-        - Discord
-        - Slack
-        - WhatsApp
-
-  - type: input
-    id: os
-    attributes:
-      label: Operating System
-      description: e.g. Ubuntu 24.04, macOS 15.2, Windows 11
-      placeholder: Ubuntu 24.04
-    validations:
-      required: true
-
-  - type: input
-    id: python-version
-    attributes:
-      label: Python Version
-      description: Output of `python --version`
-      placeholder: "3.11.9"
-    validations:
-      required: true
-
-  - type: input
-    id: hermes-version
-    attributes:
-      label: Hermes Version
-      description: Output of `hermes version`
-      placeholder: "2.1.0"
-    validations:
-      required: true
-
-  - type: textarea
-    id: logs
-    attributes:
-      label: Relevant Logs / Traceback
-      description: Paste any error output, traceback, or log messages. This will be auto-formatted as code.
-      render: shell
-
-  - type: textarea
-    id: root-cause
-    attributes:
-      label: Root Cause Analysis (optional)
-      description: |
-        If you've dug into the code and identified the root cause, share it here.
-        Include file paths, line numbers, and code snippets if possible. This massively speeds up fixes.
-      placeholder: |
-        The bug is in `gateway/run.py` line 949. `len(history)` counts session_meta entries
-        but `agent_messages` was built from filtered history...
-
-  - type: textarea
-    id: proposed-fix
-    attributes:
-      label: Proposed Fix (optional)
-      description: If you have a fix in mind (or a PR ready), describe it here.
-      placeholder: |
-        Replace `.get()` with `.pop()` on line 289 of `gateway/platforms/base.py`
-        to actually clear the pending message after retrieval.
-
-  - type: checkboxes
-    id: pr-ready
-    attributes:
-      label: Are you willing to submit a PR for this?
-      options:
-        - label: I'd like to fix this myself and submit a PR

.github/ISSUE_TEMPLATE/config.yml

@@ -1,11 +0,0 @@
-blank_issues_enabled: true
-contact_links:
-  - name: 💬 Nous Research Discord
-    url: https://discord.gg/NousResearch
-    about: For quick questions, showcasing projects, sharing skills, and community chat.
-  - name: 📖 Documentation
-    url: https://github.com/NousResearch/hermes-agent/blob/main/README.md
-    about: Check the README and docs before opening an issue.
-  - name: 🤝 Contributing Guide
-    url: https://github.com/NousResearch/hermes-agent/blob/main/CONTRIBUTING.md
-    about: Read this before submitting a PR.

.github/ISSUE_TEMPLATE/feature_request.yml

@@ -1,73 +0,0 @@
-name: "✨ Feature Request"
-description: Suggest a new feature or improvement.
-title: "[Feature]: "
-labels: ["enhancement"]
-body:
-  - type: markdown
-    attributes:
-      value: |
-        Thanks for the suggestion! Before submitting, please consider:
-
-        - **Is this a new skill?** Most capabilities should be [skills, not tools](https://github.com/NousResearch/hermes-agent/blob/main/CONTRIBUTING.md#should-it-be-a-skill-or-a-tool). If it's a specialized integration (crypto, NFT, niche SaaS), it belongs on the Skills Hub, not bundled.
-        - **Search [existing issues](https://github.com/NousResearch/hermes-agent/issues)** — someone may have already proposed this.
-
-  - type: textarea
-    id: problem
-    attributes:
-      label: Problem or Use Case
-      description: What problem does this solve? What are you trying to do that you can't today?
-      placeholder: |
-        I'm trying to use Hermes with [provider/platform/workflow] but currently
-        there's no way to...
-    validations:
-      required: true
-
-  - type: textarea
-    id: solution
-    attributes:
-      label: Proposed Solution
-      description: How do you think this should work? Be as specific as you can — CLI flags, config options, UI behavior.
-      placeholder: |
-        Add a `--foo` flag to `hermes chat` that enables...
-        Or: Add a config key `bar.baz` that controls...
-    validations:
-      required: true
-
-  - type: textarea
-    id: alternatives
-    attributes:
-      label: Alternatives Considered
-      description: What other approaches did you consider? Why is the proposed solution better?
-
-  - type: dropdown
-    id: type
-    attributes:
-      label: Feature Type
-      options:
-        - New tool
-        - New bundled skill
-        - CLI improvement
-        - Gateway / messaging improvement
-        - Configuration option
-        - Performance / reliability
-        - Developer experience (tests, docs, CI)
-        - Other
-    validations:
-      required: true
-
-  - type: dropdown
-    id: scope
-    attributes:
-      label: Scope
-      description: How big is this change?
-      options:
-        - Small (single file, < 50 lines)
-        - Medium (few files, < 300 lines)
-        - Large (new module or significant refactor)
-
-  - type: checkboxes
-    id: pr-ready
-    attributes:
-      label: Contribution
-      options:
-        - label: I'd like to implement this myself and submit a PR

.github/ISSUE_TEMPLATE/setup_help.yml

@@ -1,100 +0,0 @@
-name: "🔧 Setup / Installation Help"
-description: Having trouble installing or configuring Hermes? Ask here.
-title: "[Setup]: "
-labels: ["setup"]
-body:
-  - type: markdown
-    attributes:
-      value: |
-        Sorry you're having trouble! Please fill out the details below so we can help.
-
-        **Quick checks first:**
-        - Run `hermes doctor` and include the output below
-        - Try `hermes update` to get the latest version
-        - Check the [README troubleshooting section](https://github.com/NousResearch/hermes-agent#troubleshooting)
-        - For general questions, consider the [Nous Research Discord](https://discord.gg/NousResearch) for faster help
-
-  - type: textarea
-    id: description
-    attributes:
-      label: What's Going Wrong?
-      description: Describe what you're trying to do and where it fails.
-      placeholder: |
-        I ran `hermes setup` and selected Nous Portal, but when I try to
-        start the gateway I get...
-    validations:
-      required: true
-
-  - type: textarea
-    id: steps
-    attributes:
-      label: Steps Taken
-      description: What did you do? Include the exact commands you ran.
-      placeholder: |
-        1. Ran the install script: `curl -fsSL ... | bash`
-        2. Ran `hermes setup` and chose "Quick setup"
-        3. Selected OpenRouter, entered API key
-        4. Ran `hermes chat` and got error...
-    validations:
-      required: true
-
-  - type: dropdown
-    id: install-method
-    attributes:
-      label: Installation Method
-      options:
-        - Install script (curl | bash)
-        - Manual clone + pip/uv install
-        - PowerShell installer (Windows)
-        - Docker
-        - Other
-    validations:
-      required: true
-
-  - type: input
-    id: os
-    attributes:
-      label: Operating System
-      placeholder: Ubuntu 24.04 / macOS 15.2 / Windows 11
-    validations:
-      required: true
-
-  - type: input
-    id: python-version
-    attributes:
-      label: Python Version
-      description: Output of `python --version` (or `python3 --version`)
-      placeholder: "3.11.9"
-
-  - type: input
-    id: hermes-version
-    attributes:
-      label: Hermes Version
-      description: Output of `hermes version` (if install got that far)
-      placeholder: "2.1.0"
-
-  - type: textarea
-    id: doctor-output
-    attributes:
-      label: Output of `hermes doctor`
-      description: Run `hermes doctor` and paste the full output. This will be auto-formatted.
-      render: shell
-
-  - type: textarea
-    id: error-output
-    attributes:
-      label: Full Error Output
-      description: Paste the complete error message or traceback. This will be auto-formatted.
-      render: shell
-    validations:
-      required: true
-
-  - type: textarea
-    id: tried
-    attributes:
-      label: What I've Already Tried
-      description: List any fixes or workarounds you've already attempted.
-      placeholder: |
-        - Ran `hermes update`
-        - Tried reinstalling with `pip install -e ".[all]"`
-        - Checked that OPENROUTER_API_KEY is set in ~/.hermes/.env

.github/PULL_REQUEST_TEMPLATE.md

@@ -1,75 +0,0 @@
-## What does this PR do?
-
-<!-- Describe the change clearly. What problem does it solve? Why is this approach the right one? -->
-
-
-
-## Related Issue
-
-<!-- Link the issue this PR addresses. If no issue exists, consider creating one first. -->
-
-Fixes #
-
-## Type of Change
-
-<!-- Check the one that applies. -->
-
-- [ ] 🐛 Bug fix (non-breaking change that fixes an issue)
-- [ ] ✨ New feature (non-breaking change that adds functionality)
-- [ ] 🔒 Security fix
-- [ ] 📝 Documentation update
-- [ ] ✅ Tests (adding or improving test coverage)
-- [ ] ♻️ Refactor (no behavior change)
-- [ ] 🎯 New skill (bundled or hub)
-
-## Changes Made
-
-<!-- List the specific changes. Include file paths for code changes. -->
-
-- 
-
-## How to Test
-
-<!-- Steps to verify this change works. For bugs: reproduction steps + proof that the fix works. -->
-
-1. 
-2. 
-3. 
-
-## Checklist
-
-<!-- Complete these before requesting review. -->
-
-### Code
-
-- [ ] I've read the [Contributing Guide](https://github.com/NousResearch/hermes-agent/blob/main/CONTRIBUTING.md)
-- [ ] My commit messages follow [Conventional Commits](https://www.conventionalcommits.org/) (`fix(scope):`, `feat(scope):`, etc.)
-- [ ] I searched for [existing PRs](https://github.com/NousResearch/hermes-agent/pulls) to make sure this isn't a duplicate
-- [ ] My PR contains **only** changes related to this fix/feature (no unrelated commits)
-- [ ] I've run `pytest tests/ -q` and all tests pass
-- [ ] I've added tests for my changes (required for bug fixes, strongly encouraged for features)
-- [ ] I've tested on my platform: <!-- e.g. Ubuntu 24.04, macOS 15.2, Windows 11 -->
-
-### Documentation & Housekeeping
-
-<!-- Check all that apply. It's OK to check "N/A" if a category doesn't apply to your change. -->
-
-- [ ] I've updated relevant documentation (README, `docs/`, docstrings) — or N/A
-- [ ] I've updated `cli-config.yaml.example` if I added/changed config keys — or N/A
-- [ ] I've updated `CONTRIBUTING.md` or `AGENTS.md` if I changed architecture or workflows — or N/A
-- [ ] I've considered cross-platform impact (Windows, macOS) per the [compatibility guide](https://github.com/NousResearch/hermes-agent/blob/main/CONTRIBUTING.md#cross-platform-compatibility) — or N/A
-- [ ] I've updated tool descriptions/schemas if I changed tool behavior — or N/A
-
-## For New Skills
-
-<!-- Only fill this out if you're adding a skill. Delete this section otherwise. -->
-
-- [ ] This skill is **broadly useful** to most users (if bundled) — see [Contributing Guide](https://github.com/NousResearch/hermes-agent/blob/main/CONTRIBUTING.md#should-the-skill-be-bundled)
-- [ ] SKILL.md follows the [standard format](https://github.com/NousResearch/hermes-agent/blob/main/CONTRIBUTING.md#skillmd-format) (frontmatter, trigger conditions, steps, pitfalls)
-- [ ] No external dependencies that aren't already available (prefer stdlib, curl, existing Hermes tools)
-- [ ] I've tested the skill end-to-end: `hermes --toolsets skills -q "Use the X skill to do Y"`
-
-## Screenshots / Logs
-
-<!-- If applicable, add screenshots or log output showing the fix/feature in action. -->
-

.github/workflows/deploy-site.yml

@@ -1,74 +0,0 @@
-name: Deploy Site
-
-on:
-  push:
-    branches: [main]
-    paths:
-      - 'website/**'
-      - 'landingpage/**'
-      - 'skills/**'
-      - 'optional-skills/**'
-      - '.github/workflows/deploy-site.yml'
-  workflow_dispatch:
-
-permissions:
-  pages: write
-  id-token: write
-
-concurrency:
-  group: pages
-  cancel-in-progress: false
-
-jobs:
-  build-and-deploy:
-    # Only run on the upstream repository, not on forks
-    if: github.repository == 'NousResearch/hermes-agent'
-    runs-on: ubuntu-latest
-    environment:
-      name: github-pages
-      url: ${{ steps.deploy.outputs.page_url }}
-    steps:
-      - uses: actions/checkout@v4
-
-      - uses: actions/setup-node@v4
-        with:
-          node-version: 20
-          cache: npm
-          cache-dependency-path: website/package-lock.json
-
-      - uses: actions/setup-python@v5
-        with:
-          python-version: '3.11'
-
-      - name: Install PyYAML for skill extraction
-        run: pip install pyyaml
-
-      - name: Extract skill metadata for dashboard
-        run: python3 website/scripts/extract-skills.py
-
-      - name: Install dependencies
-        run: npm ci
-        working-directory: website
-
-      - name: Build Docusaurus
-        run: npm run build
-        working-directory: website
-
-      - name: Stage deployment
-        run: |
-          mkdir -p _site/docs
-          # Landing page at root
-          cp -r landingpage/* _site/
-          # Docusaurus at /docs/
-          cp -r website/build/* _site/docs/
-          # CNAME so GitHub Pages keeps the custom domain between deploys
-          echo "hermes-agent.nousresearch.com" > _site/CNAME
-
-      - name: Upload artifact
-        uses: actions/upload-pages-artifact@v3
-        with:
-          path: _site
-
-      - name: Deploy to GitHub Pages
-        id: deploy
-        uses: actions/deploy-pages@v4

.github/workflows/docker-publish.yml

@@ -1,79 +0,0 @@
-name: Docker Build and Publish
-
-on:
-  push:
-    branches: [main]
-  pull_request:
-    branches: [main]
-  release:
-    types: [published]
-
-concurrency:
-  group: docker-${{ github.ref }}
-  cancel-in-progress: true
-
-jobs:
-  build-and-push:
-    # Only run on the upstream repository, not on forks
-    if: github.repository == 'NousResearch/hermes-agent'
-    runs-on: ubuntu-latest
-    timeout-minutes: 30
-    steps:
-      - name: Checkout code
-        uses: actions/checkout@v4
-        with:
-          submodules: recursive
-
-      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v3
-
-      - name: Build image
-        uses: docker/build-push-action@v6
-        with:
-          context: .
-          file: Dockerfile
-          load: true
-          tags: nousresearch/hermes-agent:test
-          cache-from: type=gha
-          cache-to: type=gha,mode=max
-
-      - name: Test image starts
-        run: |
-          docker run --rm \
-            -v /tmp/hermes-test:/opt/data \
-            --entrypoint /opt/hermes/docker/entrypoint.sh \
-            nousresearch/hermes-agent:test --help
-
-      - name: Log in to Docker Hub
-        if: github.event_name == 'push' && github.ref == 'refs/heads/main' || github.event_name == 'release'
-        uses: docker/login-action@v3
-        with:
-          username: ${{ secrets.DOCKERHUB_USERNAME }}
-          password: ${{ secrets.DOCKERHUB_TOKEN }}
-
-      - name: Push image (main branch)
-        if: github.event_name == 'push' && github.ref == 'refs/heads/main'
-        uses: docker/build-push-action@v6
-        with:
-          context: .
-          file: Dockerfile
-          push: true
-          tags: |
-            nousresearch/hermes-agent:latest
-            nousresearch/hermes-agent:${{ github.sha }}
-          cache-from: type=gha
-          cache-to: type=gha,mode=max
-
-      - name: Push image (release)
-        if: github.event_name == 'release'
-        uses: docker/build-push-action@v6
-        with:
-          context: .
-          file: Dockerfile
-          push: true
-          tags: |
-            nousresearch/hermes-agent:latest
-            nousresearch/hermes-agent:${{ github.event.release.tag_name }}
-            nousresearch/hermes-agent:${{ github.sha }}
-          cache-from: type=gha
-          cache-to: type=gha,mode=max

.github/workflows/docs-site-checks.yml

@@ -1,42 +0,0 @@
-name: Docs Site Checks
-
-on:
-  pull_request:
-    paths:
-      - 'website/**'
-      - '.github/workflows/docs-site-checks.yml'
-  workflow_dispatch:
-
-jobs:
-  docs-site-checks:
-    runs-on: ubuntu-latest
-    steps:
-      - uses: actions/checkout@v4
-
-      - uses: actions/setup-node@v4
-        with:
-          node-version: 20
-          cache: npm
-          cache-dependency-path: website/package-lock.json
-
-      - name: Install website dependencies
-        run: npm ci
-        working-directory: website
-
-      - uses: actions/setup-python@v5
-        with:
-          python-version: '3.11'
-
-      - name: Install Python dependencies
-        run: python -m pip install ascii-guard pyyaml
-
-      - name: Extract skill metadata for dashboard
-        run: python3 website/scripts/extract-skills.py
-
-      - name: Lint docs diagrams
-        run: npm run lint:diagrams
-        working-directory: website
-
-      - name: Build Docusaurus
-        run: npm run build
-        working-directory: website

.github/workflows/nix.yml

@@ -1,40 +0,0 @@
-name: Nix
-
-on:
-  push:
-    branches: [main]
-  pull_request:
-    paths:
-      - 'flake.nix'
-      - 'flake.lock'
-      - 'nix/**'
-      - 'pyproject.toml'
-      - 'uv.lock'
-      - 'hermes_cli/**'
-      - 'run_agent.py'
-      - 'acp_adapter/**'
-
-concurrency:
-  group: nix-${{ github.ref }}
-  cancel-in-progress: true
-
-jobs:
-  nix:
-    strategy:
-      matrix:
-        os: [ubuntu-latest, macos-latest]
-    runs-on: ${{ matrix.os }}
-    timeout-minutes: 30
-    steps:
-      - uses: actions/checkout@v4
-      - uses: DeterminateSystems/nix-installer-action@main
-      - uses: DeterminateSystems/magic-nix-cache-action@main
-      - name: Check flake
-        if: runner.os == 'Linux'
-        run: nix flake check --print-build-logs
-      - name: Build package
-        if: runner.os == 'Linux'
-        run: nix build --print-build-logs
-      - name: Evaluate flake (macOS)
-        if: runner.os == 'macOS'
-        run: nix flake show --json > /dev/null

.github/workflows/supply-chain-audit.yml

@@ -1,192 +0,0 @@
-name: Supply Chain Audit
-
-on:
-  pull_request:
-    types: [opened, synchronize, reopened]
-
-permissions:
-  pull-requests: write
-  contents: read
-
-jobs:
-  scan:
-    name: Scan PR for supply chain risks
-    runs-on: ubuntu-latest
-    steps:
-      - name: Checkout
-        uses: actions/checkout@v4
-        with:
-          fetch-depth: 0
-
-      - name: Scan diff for suspicious patterns
-        id: scan
-        env:
-          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-        run: |
-          set -euo pipefail
-
-          BASE="${{ github.event.pull_request.base.sha }}"
-          HEAD="${{ github.event.pull_request.head.sha }}"
-
-          # Get the full diff (added lines only)
-          DIFF=$(git diff "$BASE".."$HEAD" -- . ':!uv.lock' ':!*.lock' ':!package-lock.json' ':!yarn.lock' || true)
-
-          FINDINGS=""
-          CRITICAL=false
-
-          # --- .pth files (auto-execute on Python startup) ---
-          PTH_FILES=$(git diff --name-only "$BASE".."$HEAD" | grep '\.pth$' || true)
-          if [ -n "$PTH_FILES" ]; then
-            CRITICAL=true
-            FINDINGS="${FINDINGS}
-          ### 🚨 CRITICAL: .pth file added or modified
-          Python \`.pth\` files in \`site-packages/\` execute automatically when the interpreter starts — no import required. This is the exact mechanism used in the [litellm supply chain attack](https://github.com/BerriAI/litellm/issues/24512).
-
-          **Files:**
-          \`\`\`
-          ${PTH_FILES}
-          \`\`\`
-          "
-          fi
-
-          # --- base64 + exec/eval combo (the litellm attack pattern) ---
-          B64_EXEC_HITS=$(echo "$DIFF" | grep -n '^\+' | grep -iE 'base64\.(b64decode|decodebytes|urlsafe_b64decode)' | grep -iE 'exec\(|eval\(' | head -10 || true)
-          if [ -n "$B64_EXEC_HITS" ]; then
-            CRITICAL=true
-            FINDINGS="${FINDINGS}
-          ### 🚨 CRITICAL: base64 decode + exec/eval combo
-          This is the exact pattern used in the [litellm supply chain attack](https://github.com/BerriAI/litellm/issues/24512) — base64-decoded strings passed to exec/eval to hide credential-stealing payloads.
-
-          **Matches:**
-          \`\`\`
-          ${B64_EXEC_HITS}
-          \`\`\`
-          "
-          fi
-
-          # --- base64 decode/encode (alone — legitimate uses exist) ---
-          B64_HITS=$(echo "$DIFF" | grep -n '^\+' | grep -iE 'base64\.(b64decode|b64encode|decodebytes|encodebytes|urlsafe_b64decode)|atob\(|btoa\(|Buffer\.from\(.*base64' | head -20 || true)
-          if [ -n "$B64_HITS" ]; then
-            FINDINGS="${FINDINGS}
-          ### ⚠️ WARNING: base64 encoding/decoding detected
-          Base64 has legitimate uses (images, JWT, etc.) but is also commonly used to obfuscate malicious payloads. Verify the usage is appropriate.
-
-          **Matches (first 20):**
-          \`\`\`
-          ${B64_HITS}
-          \`\`\`
-          "
-          fi
-
-          # --- exec/eval with string arguments ---
-          EXEC_HITS=$(echo "$DIFF" | grep -n '^\+' | grep -E '(exec|eval)\s*\(' | grep -v '^\+\s*#' | grep -v 'test_\|mock\|assert\|# ' | head -20 || true)
-          if [ -n "$EXEC_HITS" ]; then
-            FINDINGS="${FINDINGS}
-          ### ⚠️ WARNING: exec() or eval() usage
-          Dynamic code execution can hide malicious behavior, especially when combined with base64 or network fetches.
-
-          **Matches (first 20):**
-          \`\`\`
-          ${EXEC_HITS}
-          \`\`\`
-          "
-          fi
-
-          # --- subprocess with encoded/obfuscated commands ---
-          PROC_HITS=$(echo "$DIFF" | grep -n '^\+' | grep -E 'subprocess\.(Popen|call|run)\s*\(' | grep -iE 'base64|decode|encode|\\x|chr\(' | head -10 || true)
-          if [ -n "$PROC_HITS" ]; then
-            CRITICAL=true
-            FINDINGS="${FINDINGS}
-          ### 🚨 CRITICAL: subprocess with encoded/obfuscated command
-          Subprocess calls with encoded arguments are a strong indicator of payload execution.
-
-          **Matches:**
-          \`\`\`
-          ${PROC_HITS}
-          \`\`\`
-          "
-          fi
-
-          # --- Network calls to non-standard domains ---
-          EXFIL_HITS=$(echo "$DIFF" | grep -n '^\+' | grep -iE 'requests\.(post|put)\(|httpx\.(post|put)\(|urllib\.request\.urlopen' | grep -v '^\+\s*#' | grep -v 'test_\|mock\|assert' | head -10 || true)
-          if [ -n "$EXFIL_HITS" ]; then
-            FINDINGS="${FINDINGS}
-          ### ⚠️ WARNING: Outbound network calls (POST/PUT)
-          Outbound POST/PUT requests in new code could be data exfiltration. Verify the destination URLs are legitimate.
-
-          **Matches (first 10):**
-          \`\`\`
-          ${EXFIL_HITS}
-          \`\`\`
-          "
-          fi
-
-          # --- setup.py / setup.cfg install hooks ---
-          SETUP_HITS=$(git diff --name-only "$BASE".."$HEAD" | grep -E '(setup\.py|setup\.cfg|__init__\.pth|sitecustomize\.py|usercustomize\.py)$' || true)
-          if [ -n "$SETUP_HITS" ]; then
-            FINDINGS="${FINDINGS}
-          ### ⚠️ WARNING: Install hook files modified
-          These files can execute code during package installation or interpreter startup.
-
-          **Files:**
-          \`\`\`
-          ${SETUP_HITS}
-          \`\`\`
-          "
-          fi
-
-          # --- Compile/marshal/pickle (code object injection) ---
-          MARSHAL_HITS=$(echo "$DIFF" | grep -n '^\+' | grep -iE 'marshal\.loads|pickle\.loads|compile\(' | grep -v '^\+\s*#' | grep -v 'test_\|re\.compile\|ast\.compile' | head -10 || true)
-          if [ -n "$MARSHAL_HITS" ]; then
-            FINDINGS="${FINDINGS}
-          ### ⚠️ WARNING: marshal/pickle/compile usage
-          These can deserialize or construct executable code objects.
-
-          **Matches:**
-          \`\`\`
-          ${MARSHAL_HITS}
-          \`\`\`
-          "
-          fi
-
-          # --- Output results ---
-          if [ -n "$FINDINGS" ]; then
-            echo "found=true" >> "$GITHUB_OUTPUT"
-            if [ "$CRITICAL" = true ]; then
-              echo "critical=true" >> "$GITHUB_OUTPUT"
-            else
-              echo "critical=false" >> "$GITHUB_OUTPUT"
-            fi
-            # Write findings to a file (multiline env vars are fragile)
-            echo "$FINDINGS" > /tmp/findings.md
-          else
-            echo "found=false" >> "$GITHUB_OUTPUT"
-            echo "critical=false" >> "$GITHUB_OUTPUT"
-          fi
-
-      - name: Post warning comment
-        if: steps.scan.outputs.found == 'true'
-        env:
-          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-        run: |
-          SEVERITY="⚠️ Supply Chain Risk Detected"
-          if [ "${{ steps.scan.outputs.critical }}" = "true" ]; then
-            SEVERITY="🚨 CRITICAL Supply Chain Risk Detected"
-          fi
-
-          BODY="## ${SEVERITY}
-
-          This PR contains patterns commonly associated with supply chain attacks. This does **not** mean the PR is malicious — but these patterns require careful human review before merging.
-
-          $(cat /tmp/findings.md)
-
-          ---
-          *Automated scan triggered by [supply-chain-audit](/.github/workflows/supply-chain-audit.yml). If this is a false positive, a maintainer can approve after manual review.*"
-
-          gh pr comment "${{ github.event.pull_request.number }}" --body "$BODY"
-
-      - name: Fail on critical findings
-        if: steps.scan.outputs.critical == 'true'
-        run: |
-          echo "::error::CRITICAL supply chain risk patterns detected in this PR. See the PR comment for details."
-          exit 1

.github/workflows/tests.yml

@@ -1,70 +0,0 @@
-name: Tests
-
-on:
-  push:
-    branches: [main]
-  pull_request:
-    branches: [main]
-
-# Cancel in-progress runs for the same PR/branch
-concurrency:
-  group: tests-${{ github.ref }}
-  cancel-in-progress: true
-
-jobs:
-  test:
-    runs-on: ubuntu-latest
-    timeout-minutes: 10
-    steps:
-      - name: Checkout code
-        uses: actions/checkout@v4
-
-      - name: Install uv
-        uses: astral-sh/setup-uv@v5
-
-      - name: Set up Python 3.11
-        run: uv python install 3.11
-
-      - name: Install dependencies
-        run: |
-          uv venv .venv --python 3.11
-          source .venv/bin/activate
-          uv pip install -e ".[all,dev]"
-
-      - name: Run tests
-        run: |
-          source .venv/bin/activate
-          python -m pytest tests/ -q --ignore=tests/integration --ignore=tests/e2e --tb=short -n auto
-        env:
-          # Ensure tests don't accidentally call real APIs
-          OPENROUTER_API_KEY: ""
-          OPENAI_API_KEY: ""
-          NOUS_API_KEY: ""
-
-  e2e:
-    runs-on: ubuntu-latest
-    timeout-minutes: 10
-    steps:
-      - name: Checkout code
-        uses: actions/checkout@v4
-
-      - name: Install uv
-        uses: astral-sh/setup-uv@v5
-
-      - name: Set up Python 3.11
-        run: uv python install 3.11
-
-      - name: Install dependencies
-        run: |
-          uv venv .venv --python 3.11
-          source .venv/bin/activate
-          uv pip install -e ".[all,dev]"
-
-      - name: Run e2e tests
-        run: |
-          source .venv/bin/activate
-          python -m pytest tests/e2e/ -v --tb=short
-        env:
-          OPENROUTER_API_KEY: ""
-          OPENAI_API_KEY: ""
-          NOUS_API_KEY: ""

.gitignore

@@ -1,60 +1,72 @@
-/venv/
-/_pycache/
-*.pyc*
-__pycache__/
-.venv/
-.vscode/
-.env
-.env.local
-.env.development.local
-.env.test.local
-.env.production.local
-.env.development
-.env.test
-export*
-__pycache__/model_tools.cpython-310.pyc
-__pycache__/web_tools.cpython-310.pyc
-logs/
-data/
-.pytest_cache/
-tmp/
-temp_vision_images/
-hermes-*/*
-examples/
-tests/quick_test_dataset.jsonl
-tests/sample_dataset.jsonl
-run_datagen_kimik2-thinking.sh
-run_datagen_megascience_glm4-6.sh
-run_datagen_sonnet.sh
-source-data/*
-run_datagen_megascience_glm4-6.sh
-data/*
-node_modules/
-browser-use/
-agent-browser/
-# Private keys
-*.ppk
-*.pem
-privvy*
-images/
-__pycache__/
-hermes_agent.egg-info/
-wandb/
-testlogs
-
-# CLI config (may contain sensitive SSH paths)
-cli-config.yaml
-
-# Skills Hub state (lives in ~/.hermes/skills/.hub/ at runtime, but just in case)
-skills/.hub/
-ignored/
-.worktrees/
-environments/benchmarks/evals/
-
-# Release script temp files
-.release_notes.md
-mini-swe-agent/
-
-# Nix
-.direnv/
-result
+/venv/
+/_pycache/
+hecate/
+hecate-lib/
+*.pyc*
+__pycache__/
+.venv/
+.vscode/
+.env
+.env.local
+.env.development.local
+.env.test.local
+.env.production.local
+.env.development
+.env.test
+export*
+__pycache__/model_tools.cpython-310.pyc
+__pycache__/web_tools.cpython-310.pyc
+logs/
+data/
+.pytest_cache/
+tmp/
+temp_vision_images/
+hermes-*/*
+examples/
+tests/quick_test_dataset.jsonl
+tests/sample_dataset.jsonl
+run_datagen_kimik2-thinking.sh
+run_datagen_megascience_glm4-6.sh
+run_datagen_sonnet.sh
+source-data/*
+run_datagen_megascience_glm4-6.sh
+data/*
+node_modules/
+browser-use/
+agent-browser/
+# Private keys
+*.ppk
+*.pem
+privvy*
+images/
+__pycache__/
+hermes_agent.egg-info/
+wandb/
+testlogs
+
+# CLI config (may contain sensitive SSH paths)
+cli-config.yaml
+
+.DS_Store
+
+# artifacts
+*.jsonl
+*.html
+*.json
+*.log
+*.csv
+
+# Singularity/Apptainer images (large binary files)
+*.sif
+
+# Test files
+test_singularity_*.py
+test_*.py
+!tests/test_*.py
+
+# Nomad data
+/tmp/NomadClient*/
+
+*.egg-info*
+wandb
+logs

.gitmodules

@@ -1,3 +1,6 @@
+[submodule "mini-swe-agent"]
+	path = mini-swe-agent
+	url = https://github.com/SWE-agent/mini-swe-agent
 [submodule "tinker-atropos"]
 	path = tinker-atropos
 	url = https://github.com/nousresearch/tinker-atropos

.plans/openai-api-server.md

@@ -1,291 +0,0 @@
-# OpenAI-Compatible API Server for Hermes Agent
-
-## Motivation
-
-Every major chat frontend (Open WebUI 126k★, LobeChat 73k★, LibreChat 34k★,
-AnythingLLM 56k★, NextChat 87k★, ChatBox 39k★, Jan 26k★, HF Chat-UI 8k★,
-big-AGI 7k★) connects to backends via the OpenAI-compatible REST API with
-SSE streaming. By exposing this endpoint, hermes-agent becomes instantly
-usable as a backend for all of them — no custom adapters needed.
-
-## What It Enables
-
-```
-┌──────────────────┐
-│  Open WebUI      │──┐
-│  LobeChat        │  │    POST /v1/chat/completions
-│  LibreChat       │  ├──► Authorization: Bearer <key>     ┌─────────────────┐
-│  AnythingLLM     │  │    {"messages": [...]}             │  hermes-agent   │
-│  NextChat        │  │                                    │  gateway        │
-│  Any OAI client  │──┘    ◄── SSE streaming response      │  (API server)   │
-└──────────────────┘                                        └─────────────────┘
-```
-
-A user would:
-1. Set `API_SERVER_ENABLED=true` in `~/.hermes/.env`
-2. Run `hermes gateway` (API server starts alongside Telegram/Discord/etc.)
-3. Point Open WebUI (or any frontend) at `http://localhost:8642/v1`
-4. Chat with hermes-agent through any OpenAI-compatible UI
-
-## Endpoints
-
-| Method | Path | Purpose |
-|--------|------|---------|
-| POST | `/v1/chat/completions` | Chat with the agent (streaming + non-streaming) |
-| GET | `/v1/models` | List available "models" (returns hermes-agent as a model) |
-| GET | `/health` | Health check |
-
-## Architecture
-
-### Option A: Gateway Platform Adapter (recommended)
-
-Create `gateway/platforms/api_server.py` as a new platform adapter that
-extends `BasePlatformAdapter`. This is the cleanest approach because:
-
-- Reuses all gateway infrastructure (session management, auth, context building)
-- Runs in the same async loop as other adapters
-- Gets message handling, interrupt support, and session persistence for free
-- Follows the established pattern (like Telegram, Discord, etc.)
-- Uses `aiohttp.web` (already a dependency) for the HTTP server
-
-The adapter would start an `aiohttp.web.Application` server in `connect()`
-and route incoming HTTP requests through the standard `handle_message()` pipeline.
-
-### Option B: Standalone Component
-
-A separate HTTP server class in `gateway/api_server.py` that creates its own
-AIAgent instances directly. Simpler but duplicates session/auth logic.
-
-**Recommendation: Option A** — fits the existing architecture, less code to
-maintain, gets all gateway features for free.
-
-## Request/Response Format
-
-### Chat Completions (non-streaming)
-
-```
-POST /v1/chat/completions
-Authorization: Bearer hermes-api-key-here
-Content-Type: application/json
-
-{
-  "model": "hermes-agent",
-  "messages": [
-    {"role": "system", "content": "You are a helpful assistant."},
-    {"role": "user", "content": "What files are in the current directory?"}
-  ],
-  "stream": false,
-  "temperature": 0.7
-}
-```
-
-Response:
-```json
-{
-  "id": "chatcmpl-abc123",
-  "object": "chat.completion",
-  "created": 1710000000,
-  "model": "hermes-agent",
-  "choices": [{
-    "index": 0,
-    "message": {
-      "role": "assistant",
-      "content": "Here are the files in the current directory:\n..."
-    },
-    "finish_reason": "stop"
-  }],
-  "usage": {
-    "prompt_tokens": 50,
-    "completion_tokens": 200,
-    "total_tokens": 250
-  }
-}
-```
-
-### Chat Completions (streaming)
-
-Same request with `"stream": true`. Response is SSE:
-
-```
-data: {"id":"chatcmpl-abc123","object":"chat.completion.chunk","choices":[{"index":0,"delta":{"role":"assistant"},"finish_reason":null}]}
-
-data: {"id":"chatcmpl-abc123","object":"chat.completion.chunk","choices":[{"index":0,"delta":{"content":"Here "},"finish_reason":null}]}
-
-data: {"id":"chatcmpl-abc123","object":"chat.completion.chunk","choices":[{"index":0,"delta":{"content":"are "},"finish_reason":null}]}
-
-data: {"id":"chatcmpl-abc123","object":"chat.completion.chunk","choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}
-
-data: [DONE]
-```
-
-### Models List
-
-```
-GET /v1/models
-Authorization: Bearer hermes-api-key-here
-```
-
-Response:
-```json
-{
-  "object": "list",
-  "data": [{
-    "id": "hermes-agent",
-    "object": "model",
-    "created": 1710000000,
-    "owned_by": "hermes-agent"
-  }]
-}
-```
-
-## Key Design Decisions
-
-### 1. Session Management
-
-The OpenAI API is stateless — each request includes the full conversation.
-But hermes-agent sessions have persistent state (memory, skills, tool context).
-
-**Approach: Hybrid**
-- Default: Stateless. Each request is independent. The `messages` array IS
-  the conversation. No session persistence between requests.
-- Opt-in persistent sessions via `X-Session-ID` header. When provided, the
-  server maintains session state across requests (conversation history,
-  memory context, tool state). This enables richer agent behavior.
-- The session ID also enables interrupt support — a subsequent request with
-  the same session ID while one is running triggers an interrupt.
-
-### 2. Streaming
-
-The agent's `run_conversation()` is synchronous and returns the full response.
-For real SSE streaming, we need to emit chunks as they're generated.
-
-**Phase 1 (MVP):** Run agent in a thread, return the complete response as
-a single SSE chunk + `[DONE]`. This works with all frontends — they just see
-a fast single-chunk response. Not true streaming but functional.
-
-**Phase 2:** Add a response callback to AIAgent that emits text chunks as the
-LLM generates them. The API server captures these via a queue and streams them
-as SSE events. This gives real token-by-token streaming.
-
-**Phase 3:** Stream tool execution progress too — emit tool call/result events
-as the agent works, giving frontends visibility into what the agent is doing.
-
-### 3. Tool Transparency
-
-Two modes:
-- **Opaque (default):** Frontends see only the final response. Tool calls
-  happen server-side and are invisible. Best for general-purpose UIs.
-- **Transparent (opt-in via header):** Tool calls are emitted as OpenAI-format
-  tool_call/tool_result messages in the stream. Useful for agent-aware frontends.
-
-### 4. Authentication
-
-- Bearer token via `Authorization: Bearer <key>` header
-- Token configured via `API_SERVER_KEY` env var
-- Optional: allow unauthenticated local-only access (127.0.0.1 bind)
-- Follows the same pattern as other platform adapters
-
-### 5. Model Mapping
-
-Frontends send `"model": "hermes-agent"` (or whatever). The actual LLM model
-used is configured server-side in config.yaml. The API server maps any
-requested model name to the configured hermes-agent model.
-
-Optionally, allow model passthrough: if the frontend sends
-`"model": "anthropic/claude-sonnet-4"`, the agent uses that model. Controlled
-by a config flag.
-
-## Configuration
-
-```yaml
-# In config.yaml
-api_server:
-  enabled: true
-  port: 8642
-  host: "127.0.0.1"        # localhost only by default
-  key: "your-secret-key"   # or via API_SERVER_KEY env var
-  allow_model_override: false  # let clients choose the model
-  max_concurrent: 5         # max simultaneous requests
-```
-
-Environment variables:
-```bash
-API_SERVER_ENABLED=true
-API_SERVER_PORT=8642
-API_SERVER_HOST=127.0.0.1
-API_SERVER_KEY=your-secret-key
-```
-
-## Implementation Plan
-
-### Phase 1: MVP (non-streaming) — PR
-
-1. `gateway/platforms/api_server.py` — new adapter
-   - aiohttp.web server with endpoints:
-     - `POST /v1/chat/completions` — Chat Completions API (universal compat)
-     - `POST /v1/responses` — Responses API (server-side state, tool preservation)
-     - `GET /v1/models` — list available models
-     - `GET /health` — health check
-   - Bearer token auth middleware
-   - Non-streaming responses (run agent, return full result)
-   - Chat Completions: stateless, messages array is the conversation
-   - Responses API: server-side conversation storage via previous_response_id
-     - Store full internal conversation (including tool calls) keyed by response ID
-     - On subsequent requests, reconstruct full context from stored chain
-   - Frontend system prompt layered on top of hermes-agent's core prompt
-
-2. `gateway/config.py` — add `Platform.API_SERVER` enum + config
-
-3. `gateway/run.py` — register adapter in `_create_adapter()`
-
-4. Tests in `tests/gateway/test_api_server.py`
-
-### Phase 2: SSE Streaming
-
-1. Add response streaming to both endpoints
-   - Chat Completions: `choices[0].delta.content` SSE format
-   - Responses API: semantic events (response.output_text.delta, etc.)
-   - Run agent in thread, collect output via callback queue
-   - Handle client disconnect (cancel agent)
-
-2. Add `stream_callback` parameter to `AIAgent.run_conversation()`
-
-### Phase 3: Enhanced Features
-
-1. Tool call transparency mode (opt-in)
-2. Model passthrough/override
-3. Concurrent request limiting
-4. Usage tracking / rate limiting
-5. CORS headers for browser-based frontends
-6. GET /v1/responses/{id} — retrieve stored response
-7. DELETE /v1/responses/{id} — delete stored response
-
-## Files Changed
-
-| File | Change |
-|------|--------|
-| `gateway/platforms/api_server.py` | NEW — main adapter (~300 lines) |
-| `gateway/config.py` | Add Platform.API_SERVER + config (~20 lines) |
-| `gateway/run.py` | Register adapter in _create_adapter() (~10 lines) |
-| `tests/gateway/test_api_server.py` | NEW — tests (~200 lines) |
-| `cli-config.yaml.example` | Add api_server section |
-| `README.md` | Mention API server in platform list |
-
-## Compatibility Matrix
-
-Once implemented, hermes-agent works as a drop-in backend for:
-
-| Frontend | Stars | How to Connect |
-|----------|-------|---------------|
-| Open WebUI | 126k | Settings → Connections → Add OpenAI API, URL: `http://localhost:8642/v1` |
-| NextChat | 87k | BASE_URL env var |
-| LobeChat | 73k | Custom provider endpoint |
-| AnythingLLM | 56k | LLM Provider → Generic OpenAI |
-| Oobabooga | 42k | Already a backend, not a frontend |
-| ChatBox | 39k | API Host setting |
-| LibreChat | 34k | librechat.yaml custom endpoint |
-| Chatbot UI | 29k | Custom API endpoint |
-| Jan | 26k | Remote model config |
-| AionUI | 18k | Custom API endpoint |
-| HF Chat-UI | 8k | OPENAI_BASE_URL env var |
-| big-AGI | 7k | Custom endpoint |

.plans/streaming-support.md

@@ -1,705 +0,0 @@
-# Streaming LLM Response Support for Hermes Agent
-
-## Overview
-
-Add token-by-token streaming of LLM responses across all platforms. When enabled,
-users see the response typing out live instead of waiting for the full generation.
-Streaming is opt-in via config, defaults to off, and all existing non-streaming
-code paths remain intact as the default.
-
-## Design Principles
-
-1. **Feature-flagged**: `streaming.enabled: true` in config.yaml. Off by default.
-   When off, all existing code paths are unchanged — zero risk to current behavior.
-2. **Callback-based**: A simple `stream_callback(text_delta: str)` function injected
-   into AIAgent. The agent doesn't know or care what the consumer does with tokens.
-3. **Graceful degradation**: If the provider doesn't support streaming, or streaming
-   fails for any reason, silently fall back to the non-streaming path.
-4. **Platform-agnostic core**: The streaming mechanism in AIAgent works the same
-   regardless of whether the consumer is CLI, Telegram, Discord, or the API server.
-
----
-
-## Architecture
-
-```
-                              stream_callback(delta)
-                                    │
-  ┌─────────────┐    ┌─────────────▼──────────────┐
-  │  LLM API    │    │      queue.Queue()          │
-  │  (stream)   │───►│  thread-safe bridge between │
-  │             │    │  agent thread & consumer    │
-  └─────────────┘    └─────────────┬──────────────┘
-                                   │
-                    ┌──────────────┼──────────────┐
-                    │              │              │
-              ┌─────▼─────┐ ┌─────▼─────┐ ┌─────▼─────┐
-              │    CLI     │ │  Gateway  │ │ API Server│
-              │ print to   │ │ edit msg  │ │ SSE event │
-              │ terminal   │ │ on Tg/Dc  │ │ to client │
-              └───────────┘ └───────────┘ └───────────┘
-```
-
-The agent runs in a thread. The callback puts tokens into a thread-safe queue.
-Each consumer reads the queue in its own context (async task, main thread, etc.).
-
----
-
-## Configuration
-
-### config.yaml
-
-```yaml
-streaming:
-  enabled: false          # Master switch. Default off.
-  # Per-platform overrides (optional):
-  # cli: true             # Override for CLI only
-  # telegram: true        # Override for Telegram only
-  # discord: false        # Keep Discord non-streaming
-  # api_server: true      # Override for API server
-```
-
-### Environment variables
-
-```
-HERMES_STREAMING_ENABLED=true    # Master switch via env
-```
-
-### How the flag is read
-
-- **CLI**: `load_cli_config()` reads `streaming.enabled`, sets env var. AIAgent
-  checks at init time.
-- **Gateway**: `_run_agent()` reads config, decides whether to pass
-  `stream_callback` to the AIAgent constructor.
-- **API server**: For Chat Completions `stream=true` requests, always uses streaming
-  regardless of config (the client is explicitly requesting it). For non-stream
-  requests, uses config.
-
-### Precedence
-
-1. API server: client's `stream` field overrides everything
-2. Per-platform config override (e.g., `streaming.telegram: true`)
-3. Master `streaming.enabled` flag
-4. Default: off
-
----
-
-## Implementation Plan
-
-### Phase 1: Core streaming infrastructure in AIAgent
-
-**File: run_agent.py**
-
-#### 1a. Add stream_callback parameter to __init__ (~5 lines)
-
-```python
-def __init__(self, ..., stream_callback: callable = None, ...):
-    self.stream_callback = stream_callback
-```
-
-No other init changes. The callback is optional — when None, everything
-works exactly as before.
-
-#### 1b. Add _run_streaming_chat_completion() method (~65 lines)
-
-New method for Chat Completions API streaming:
-
-```python
-def _run_streaming_chat_completion(self, api_kwargs: dict):
-    """Stream a chat completion, emitting text tokens via stream_callback.
-    
-    Returns a fake response object compatible with the non-streaming code path.
-    Falls back to non-streaming on any error.
-    """
-    stream_kwargs = dict(api_kwargs)
-    stream_kwargs["stream"] = True
-    stream_kwargs["stream_options"] = {"include_usage": True}
-    
-    accumulated_content = []
-    accumulated_tool_calls = {}  # index -> {id, name, arguments}
-    final_usage = None
-    
-    try:
-        stream = self.client.chat.completions.create(**stream_kwargs)
-        
-        for chunk in stream:
-            if not chunk.choices:
-                # Usage-only chunk (final)
-                if chunk.usage:
-                    final_usage = chunk.usage
-                continue
-            
-            delta = chunk.choices[0].delta
-            
-            # Text content — emit via callback
-            if delta.content:
-                accumulated_content.append(delta.content)
-                if self.stream_callback:
-                    try:
-                        self.stream_callback(delta.content)
-                    except Exception:
-                        pass
-            
-            # Tool call deltas — accumulate silently
-            if delta.tool_calls:
-                for tc_delta in delta.tool_calls:
-                    idx = tc_delta.index
-                    if idx not in accumulated_tool_calls:
-                        accumulated_tool_calls[idx] = {
-                            "id": tc_delta.id or "",
-                            "name": "", "arguments": ""
-                        }
-                    if tc_delta.function:
-                        if tc_delta.function.name:
-                            accumulated_tool_calls[idx]["name"] = tc_delta.function.name
-                        if tc_delta.function.arguments:
-                            accumulated_tool_calls[idx]["arguments"] += tc_delta.function.arguments
-        
-        # Build fake response compatible with existing code
-        tool_calls = []
-        for idx in sorted(accumulated_tool_calls):
-            tc = accumulated_tool_calls[idx]
-            if tc["name"]:
-                tool_calls.append(SimpleNamespace(
-                    id=tc["id"], type="function",
-                    function=SimpleNamespace(name=tc["name"], arguments=tc["arguments"]),
-                ))
-        
-        return SimpleNamespace(
-            choices=[SimpleNamespace(
-                message=SimpleNamespace(
-                    content="".join(accumulated_content) or "",
-                    tool_calls=tool_calls or None,
-                    role="assistant",
-                ),
-                finish_reason="tool_calls" if tool_calls else "stop",
-            )],
-            usage=final_usage,
-            model=self.model,
-        )
-    
-    except Exception as e:
-        logger.debug("Streaming failed, falling back to non-streaming: %s", e)
-        return self.client.chat.completions.create(**api_kwargs)
-```
-
-#### 1c. Modify _run_codex_stream() for Responses API (~10 lines)
-
-The method already iterates the stream. Add callback emission:
-
-```python
-def _run_codex_stream(self, api_kwargs: dict):
-    with self.client.responses.stream(**api_kwargs) as stream:
-        for event in stream:
-            # Emit text deltas if streaming callback is set
-            if self.stream_callback and hasattr(event, 'type'):
-                if event.type == 'response.output_text.delta':
-                    try:
-                        self.stream_callback(event.delta)
-                    except Exception:
-                        pass
-        return stream.get_final_response()
-```
-
-#### 1d. Modify _interruptible_api_call() (~5 lines)
-
-Add the streaming branch:
-
-```python
-def _call():
-    try:
-        if self.api_mode == "codex_responses":
-            result["response"] = self._run_codex_stream(api_kwargs)
-        elif self.stream_callback is not None:
-            result["response"] = self._run_streaming_chat_completion(api_kwargs)
-        else:
-            result["response"] = self.client.chat.completions.create(**api_kwargs)
-    except Exception as e:
-        result["error"] = e
-```
-
-#### 1e. Signal end-of-stream to consumers (~5 lines)
-
-After the API call returns, signal the callback that streaming is done
-so consumers can finalize (remove cursor, close SSE, etc.):
-
-```python
-# In run_conversation(), after _interruptible_api_call returns:
-if self.stream_callback:
-    try:
-        self.stream_callback(None)  # None = end of stream signal
-    except Exception:
-        pass
-```
-
-Consumers check: `if delta is None: finalize()`
-
-**Tests for Phase 1:** (~150 lines)
-- Test _run_streaming_chat_completion with mocked stream
-- Test fallback to non-streaming on error
-- Test tool_call accumulation during streaming
-- Test stream_callback receives correct deltas
-- Test None signal at end of stream
-- Test streaming disabled when callback is None
-
----
-
-### Phase 2: Gateway consumers (Telegram, Discord, etc.)
-
-**File: gateway/run.py**
-
-#### 2a. Read streaming config (~15 lines)
-
-In `_run_agent()`, before creating the AIAgent:
-
-```python
-# Read streaming config
-_streaming_enabled = False
-try:
-    # Check per-platform override first
-    platform_key = source.platform.value if source.platform else ""
-    _stream_cfg = {}  # loaded from config.yaml streaming section
-    if _stream_cfg.get(platform_key) is not None:
-        _streaming_enabled = bool(_stream_cfg[platform_key])
-    else:
-        _streaming_enabled = bool(_stream_cfg.get("enabled", False))
-except Exception:
-    pass
-# Env var override
-if os.getenv("HERMES_STREAMING_ENABLED", "").lower() in ("true", "1", "yes"):
-    _streaming_enabled = True
-```
-
-#### 2b. Set up queue + callback (~15 lines)
-
-```python
-_stream_q = None
-_stream_done = None
-_stream_msg_id = [None]  # mutable ref for the async task
-
-if _streaming_enabled:
-    import queue as _q
-    _stream_q = _q.Queue()
-    _stream_done = threading.Event()
-    
-    def _on_token(delta):
-        if delta is None:
-            _stream_done.set()
-        else:
-            _stream_q.put(delta)
-```
-
-Pass `stream_callback=_on_token` to the AIAgent constructor.
-
-#### 2c. Telegram/Discord stream preview task (~50 lines)
-
-```python
-async def stream_preview():
-    """Progressively edit a message with streaming tokens."""
-    if not _stream_q:
-        return
-    adapter = self.adapters.get(source.platform)
-    if not adapter:
-        return
-    
-    accumulated = []
-    token_count = 0
-    last_edit = 0.0
-    MIN_TOKENS = 20          # Don't show until enough context
-    EDIT_INTERVAL = 1.5      # Respect Telegram rate limits
-    
-    try:
-        while not _stream_done.is_set():
-            try:
-                chunk = _stream_q.get(timeout=0.1)
-                accumulated.append(chunk)
-                token_count += 1
-            except queue.Empty:
-                continue
-            
-            now = time.monotonic()
-            if token_count >= MIN_TOKENS and (now - last_edit) >= EDIT_INTERVAL:
-                preview = "".join(accumulated) + " ▌"
-                if _stream_msg_id[0] is None:
-                    r = await adapter.send(
-                        chat_id=source.chat_id,
-                        content=preview,
-                        metadata=_thread_metadata,
-                    )
-                    if r.success and r.message_id:
-                        _stream_msg_id[0] = r.message_id
-                else:
-                    await adapter.edit_message(
-                        chat_id=source.chat_id,
-                        message_id=_stream_msg_id[0],
-                        content=preview,
-                    )
-                last_edit = now
-        
-        # Drain remaining tokens
-        while not _stream_q.empty():
-            accumulated.append(_stream_q.get_nowait())
-        
-        # Final edit — remove cursor, show complete text
-        if _stream_msg_id[0] and accumulated:
-            await adapter.edit_message(
-                chat_id=source.chat_id,
-                message_id=_stream_msg_id[0],
-                content="".join(accumulated),
-            )
-    
-    except asyncio.CancelledError:
-        # Clean up on cancel
-        if _stream_msg_id[0] and accumulated:
-            try:
-                await adapter.edit_message(
-                    chat_id=source.chat_id,
-                    message_id=_stream_msg_id[0],
-                    content="".join(accumulated),
-                )
-            except Exception:
-                pass
-    except Exception as e:
-        logger.debug("stream_preview error: %s", e)
-```
-
-#### 2d. Skip final send if already streamed (~10 lines)
-
-In `_process_message_background()` (base.py), after getting the response,
-if streaming was active and `_stream_msg_id[0]` is set, the final response
-was already delivered via progressive edits. Skip the normal `self.send()`
-call to avoid duplicating the message.
-
-This is the most delicate integration point — we need to communicate from
-the gateway's `_run_agent` back to the base adapter's response sender that
-the response was already delivered. Options:
-
-- **Option A**: Return a special marker in the result dict:
-  `result["_streamed_msg_id"] = _stream_msg_id[0]`
-  The base adapter checks this and skips `send()`.
-  
-- **Option B**: Edit the already-sent message with the final response
-  (which may differ slightly from accumulated tokens due to think-block
-  stripping, etc.) and don't send a new one.
-
-- **Option C**: The stream preview task handles the FULL final response
-  (including any post-processing), and the handler returns None to skip
-  the normal send path.
-
-Recommended: **Option A** — cleanest separation. The result dict already
-carries metadata; adding one more field is low-risk.
-
-**Platform-specific considerations:**
-
-| Platform | Edit support | Rate limits | Streaming approach |
-|----------|-------------|-------------|-------------------|
-| Telegram | ✅ edit_message_text | ~20 edits/min | Edit every 1.5s |
-| Discord | ✅ message.edit | 5 edits/5s per message | Edit every 1.2s |
-| Slack | ✅ chat.update | Tier 3 (~50/min) | Edit every 1.5s |
-| WhatsApp | ❌ no edit support | N/A | Skip streaming, use normal path |
-| HomeAssistant | ❌ no edit | N/A | Skip streaming |
-| API Server | ✅ SSE native | No limit | Real SSE events |
-
-WhatsApp and HomeAssistant fall back to non-streaming automatically because
-they don't support message editing.
-
-**Tests for Phase 2:** (~100 lines)
-- Test stream_preview sends/edits correctly
-- Test skip-final-send when streaming delivered
-- Test WhatsApp/HA graceful fallback
-- Test streaming disabled per-platform config
-- Test thread_id metadata forwarded in stream messages
-
----
-
-### Phase 3: CLI streaming
-
-**File: cli.py**
-
-#### 3a. Set up callback in the CLI chat loop (~20 lines)
-
-In `_chat_once()` or wherever the agent is invoked:
-
-```python
-if streaming_enabled:
-    _stream_q = queue.Queue()
-    _stream_done = threading.Event()
-    
-    def _cli_stream_callback(delta):
-        if delta is None:
-            _stream_done.set()
-        else:
-            _stream_q.put(delta)
-    
-    agent.stream_callback = _cli_stream_callback
-```
-
-#### 3b. Token display thread/task (~30 lines)
-
-Start a thread that reads the queue and prints tokens:
-
-```python
-def _stream_display():
-    """Print tokens to terminal as they arrive."""
-    first_token = True
-    while not _stream_done.is_set():
-        try:
-            delta = _stream_q.get(timeout=0.1)
-        except queue.Empty:
-            continue
-        if first_token:
-            # Print response box top border
-            _cprint(f"\n{top}")
-            first_token = False
-        sys.stdout.write(delta)
-        sys.stdout.flush()
-    # Drain remaining
-    while not _stream_q.empty():
-        sys.stdout.write(_stream_q.get_nowait())
-    sys.stdout.flush()
-    # Print bottom border
-    _cprint(f"\n\n{bot}")
-```
-
-**Integration challenge: prompt_toolkit**
-
-The CLI uses prompt_toolkit which controls the terminal. Writing directly
-to stdout while prompt_toolkit is active can cause display corruption.
-The existing KawaiiSpinner already solves this by using prompt_toolkit's
-`patch_stdout` context. The streaming display would need to do the same.
-
-Alternative: use `_cprint()` for each token chunk (routes through
-prompt_toolkit's renderer). But this might be slow for individual tokens.
-
-Recommended approach: accumulate tokens in small batches (e.g., every 50ms)
-and `_cprint()` the batch. This balances display responsiveness with
-prompt_toolkit compatibility.
-
-**Tests for Phase 3:** (~50 lines)
-- Test CLI streaming callback setup
-- Test response box borders with streaming
-- Test fallback when streaming disabled
-
----
-
-### Phase 4: API Server real streaming
-
-**File: gateway/platforms/api_server.py**
-
-Replace the pseudo-streaming `_write_sse_chat_completion()` with real
-token-by-token SSE when the agent supports it.
-
-#### 4a. Wire streaming callback for stream=true requests (~20 lines)
-
-```python
-if stream:
-    _stream_q = queue.Queue()
-    
-    def _api_stream_callback(delta):
-        _stream_q.put(delta)  # None = done
-    
-    # Pass callback to _run_agent
-    result, usage = await self._run_agent(
-        ..., stream_callback=_api_stream_callback,
-    )
-```
-
-#### 4b. Real SSE writer (~40 lines)
-
-```python
-async def _write_real_sse(self, request, completion_id, model, stream_q):
-    response = web.StreamResponse(
-        headers={"Content-Type": "text/event-stream", "Cache-Control": "no-cache"},
-    )
-    await response.prepare(request)
-    
-    # Role chunk
-    await response.write(...)
-    
-    # Stream content chunks as they arrive
-    while True:
-        try:
-            delta = await asyncio.get_event_loop().run_in_executor(
-                None, lambda: stream_q.get(timeout=0.1)
-            )
-        except queue.Empty:
-            continue
-        
-        if delta is None:  # End of stream
-            break
-        
-        chunk = {"id": completion_id, "object": "chat.completion.chunk", ...
-                 "choices": [{"delta": {"content": delta}, ...}]}
-        await response.write(f"data: {json.dumps(chunk)}\n\n".encode())
-    
-    # Finish + [DONE]
-    await response.write(...)
-    await response.write(b"data: [DONE]\n\n")
-    return response
-```
-
-**Challenge: concurrent execution**
-
-The agent runs in a thread executor. SSE writing happens in the async event
-loop. The queue bridges them. But `_run_agent()` currently awaits the full
-result before returning. For real streaming, we need to start the agent in
-the background and stream tokens while it runs:
-
-```python
-# Start agent in background
-agent_task = asyncio.create_task(self._run_agent_async(...))
-
-# Stream tokens while agent runs
-await self._write_real_sse(request, ..., stream_q)
-
-# Agent is done by now (stream_q received None)
-result, usage = await agent_task
-```
-
-This requires splitting `_run_agent` into an async version that doesn't
-block waiting for the result, or running it in a separate task.
-
-**Responses API SSE format:**
-
-For `/v1/responses` with `stream=true`, the SSE events are different:
-
-```
-event: response.output_text.delta
-data: {"type":"response.output_text.delta","delta":"Hello"}
-
-event: response.completed  
-data: {"type":"response.completed","response":{...}}
-```
-
-This needs a separate SSE writer that emits Responses API format events.
-
-**Tests for Phase 4:** (~80 lines)
-- Test real SSE streaming with mocked agent
-- Test SSE event format (Chat Completions vs Responses)
-- Test client disconnect during streaming
-- Test fallback to pseudo-streaming when callback not available
-
----
-
-## Integration Issues & Edge Cases
-
-### 1. Tool calls during streaming
-
-When the model returns tool calls instead of text, no text tokens are emitted.
-The stream_callback is simply never called with text. After tools execute, the
-next API call may produce the final text response — streaming picks up again.
-
-The stream preview task needs to handle this: if no tokens arrive during a
-tool-call round, don't send/edit any message. The tool progress messages
-continue working as before.
-
-### 2. Duplicate messages
-
-The biggest risk: the agent sends the final response normally (via the
-existing send path) AND the stream preview already showed it. The user
-sees the response twice.
-
-Prevention: when streaming is active and tokens were delivered, the final
-response send must be suppressed. The `result["_streamed_msg_id"]` marker
-tells the base adapter to skip its normal send.
-
-### 3. Response post-processing
-
-The final response may differ from the accumulated streamed tokens:
-- Think block stripping (`<think>...</think>` removed)
-- Trailing whitespace cleanup
-- Tool result media tag appending
-
-The stream preview shows raw tokens. The final edit should use the
-post-processed version. This means the final edit (removing the cursor)
-should use the post-processed `final_response`, not just the accumulated
-stream text.
-
-### 4. Context compression during streaming
-
-If the agent triggers context compression mid-conversation, the streaming
-tokens from BEFORE compression are from a different context than those
-after. This isn't a problem in practice — compression happens between
-API calls, not during streaming.
-
-### 5. Interrupt during streaming
-
-User sends a new message while streaming → interrupt. The stream is killed
-(HTTP connection closed), accumulated tokens are shown as-is (no cursor),
-and the interrupt message is processed normally. This is already handled by
-`_interruptible_api_call` closing the client.
-
-### 6. Multi-model / fallback
-
-If the primary model fails and the agent falls back to a different model,
-streaming state resets. The fallback call may or may not support streaming.
-The graceful fallback in `_run_streaming_chat_completion` handles this.
-
-### 7. Rate limiting on edits
-
-Telegram: ~20 edits/minute (~1 every 3 seconds to be safe)
-Discord: 5 edits per 5 seconds per message
-Slack: ~50 API calls/minute
-
-The 1.5s edit interval is conservative enough for all platforms. If we get
-429 rate limit errors on edits, just skip that edit cycle and try next time.
-
----
-
-## Files Changed Summary
-
-| File | Phase | Changes |
-|------|-------|---------|
-| `run_agent.py` | 1 | +stream_callback param, +_run_streaming_chat_completion(), modify _run_codex_stream(), modify _interruptible_api_call() |
-| `gateway/run.py` | 2 | +streaming config reader, +queue/callback setup, +stream_preview task, +skip-final-send logic |
-| `gateway/platforms/base.py` | 2 | +check for _streamed_msg_id in response handler |
-| `cli.py` | 3 | +streaming setup, +token display, +response box integration |
-| `gateway/platforms/api_server.py` | 4 | +real SSE writer, +streaming callback wiring |
-| `hermes_cli/config.py` | 1 | +streaming config defaults |
-| `cli-config.yaml.example` | 1 | +streaming section |
-| `tests/test_streaming.py` | 1-4 | NEW — ~380 lines of tests |
-
-**Total new code**: ~500 lines across all phases
-**Total test code**: ~380 lines
-
----
-
-## Rollout Plan
-
-1. **Phase 1** (core): Merge to main. Streaming disabled by default.
-   Zero impact on existing behavior. Can be tested with env var.
-
-2. **Phase 2** (gateway): Merge to main. Test on Telegram manually.
-   Enable per-platform: `streaming.telegram: true` in config.
-
-3. **Phase 3** (CLI): Merge to main. Test in terminal.
-   Enable: `streaming.cli: true` or `streaming.enabled: true`.
-
-4. **Phase 4** (API server): Merge to main. Test with Open WebUI.
-   Auto-enabled when client sends `stream: true`.
-
-Each phase is independently mergeable and testable. Streaming stays
-off by default throughout. Once all phases are stable, consider
-changing the default to enabled.
-
----
-
-## Config Reference (final state)
-
-```yaml
-# config.yaml
-streaming:
-  enabled: false          # Master switch (default: off)
-  cli: true               # Per-platform override
-  telegram: true
-  discord: true
-  slack: true
-  api_server: true        # API server always streams when client requests it
-  edit_interval: 1.5      # Seconds between message edits (default: 1.5)
-  min_tokens: 20          # Tokens before first display (default: 20)
-```
-
-```bash
-# Environment variable override
-HERMES_STREAMING_ENABLED=true
-```

AGENTS.md

@@ -1,469 +1,533 @@
 # Hermes Agent - Development Guide
 
-Instructions for AI coding assistants and developers working on the hermes-agent codebase.
+Instructions for AI coding assistants (GitHub Copilot, Cursor, etc.) and human developers.
+
+Hermes-Agent is an AI agent harness with tool-calling capabilities, interactive CLI, messaging integrations, and scheduled tasks.
 
 ## Development Environment
 
+**IMPORTANT**: Always use the virtual environment if it exists:
 ```bash
-source venv/bin/activate  # ALWAYS activate before running Python
+source venv/bin/activate  # Before running any Python commands
 ```
 
 ## Project Structure
 
 ```
 hermes-agent/
-├── run_agent.py          # AIAgent class — core conversation loop
-├── model_tools.py        # Tool orchestration, _discover_tools(), handle_function_call()
-├── toolsets.py           # Toolset definitions, _HERMES_CORE_TOOLS list
-├── cli.py                # HermesCLI class — interactive CLI orchestrator
-├── hermes_state.py       # SessionDB — SQLite session store (FTS5 search)
-├── agent/                # Agent internals
-│   ├── prompt_builder.py     # System prompt assembly
-│   ├── context_compressor.py # Auto context compression
-│   ├── prompt_caching.py     # Anthropic prompt caching
-│   ├── auxiliary_client.py   # Auxiliary LLM client (vision, summarization)
-│   ├── model_metadata.py     # Model context lengths, token estimation
-│   ├── models_dev.py         # models.dev registry integration (provider-aware context)
-│   ├── display.py            # KawaiiSpinner, tool preview formatting
-│   ├── skill_commands.py     # Skill slash commands (shared CLI/gateway)
-│   └── trajectory.py         # Trajectory saving helpers
-├── hermes_cli/           # CLI subcommands and setup
-│   ├── main.py           # Entry point — all `hermes` subcommands
-│   ├── config.py         # DEFAULT_CONFIG, OPTIONAL_ENV_VARS, migration
-│   ├── commands.py       # Slash command definitions + SlashCommandCompleter
-│   ├── callbacks.py      # Terminal callbacks (clarify, sudo, approval)
+├── hermes_cli/           # Unified CLI commands
+│   ├── main.py           # Entry point, command dispatcher
 │   ├── setup.py          # Interactive setup wizard
-│   ├── skin_engine.py    # Skin/theme engine — CLI visual customization
-│   ├── skills_config.py  # `hermes skills` — enable/disable skills per platform
-│   ├── tools_config.py   # `hermes tools` — enable/disable tools per platform
-│   ├── skills_hub.py     # `/skills` slash command (search, browse, install)
-│   ├── models.py         # Model catalog, provider model lists
-│   ├── model_switch.py   # Shared /model switch pipeline (CLI + gateway)
-│   └── auth.py           # Provider credential resolution
-├── tools/                # Tool implementations (one file per tool)
-│   ├── registry.py       # Central tool registry (schemas, handlers, dispatch)
-│   ├── approval.py       # Dangerous command detection
-│   ├── terminal_tool.py  # Terminal orchestration
-│   ├── process_registry.py # Background process management
-│   ├── file_tools.py     # File read/write/search/patch
-│   ├── web_tools.py      # Web search/extract (Parallel + Firecrawl)
-│   ├── browser_tool.py   # Browserbase browser automation
-│   ├── code_execution_tool.py # execute_code sandbox
-│   ├── delegate_tool.py  # Subagent delegation
-│   ├── mcp_tool.py       # MCP client (~1050 lines)
-│   └── environments/     # Terminal backends (local, docker, ssh, modal, daytona, singularity)
-├── gateway/              # Messaging platform gateway
-│   ├── run.py            # Main loop, slash commands, message dispatch
-│   ├── session.py        # SessionStore — conversation persistence
-│   └── platforms/        # Adapters: telegram, discord, slack, whatsapp, homeassistant, signal
-├── acp_adapter/          # ACP server (VS Code / Zed / JetBrains integration)
-├── cron/                 # Scheduler (jobs.py, scheduler.py)
-├── environments/         # RL training environments (Atropos)
-├── tests/                # Pytest suite (~3000 tests)
+│   ├── config.py         # Config management & migration
+│   ├── status.py         # Status display
+│   ├── doctor.py         # Diagnostics
+│   ├── gateway.py        # Gateway management
+│   ├── uninstall.py      # Uninstaller
+│   └── cron.py           # Cron job management
+├── tools/                # Tool implementations
+├── gateway/              # Messaging platform adapters
+├── cron/                 # Scheduler implementation
+├── skills/               # Knowledge documents
+├── cli.py                # Interactive CLI (Rich UI)
+├── run_agent.py          # Agent runner with AIAgent class
+├── model_tools.py        # Tool schemas and handlers
+├── toolsets.py           # Tool groupings
+├── toolset_distributions.py  # Probability-based tool selection
 └── batch_runner.py       # Parallel batch processing
 ```
 
-**User config:** `~/.hermes/config.yaml` (settings), `~/.hermes/.env` (API keys)
+**User Configuration** (stored in `~/.hermes/`):
+- `~/.hermes/config.yaml` - Settings (model, terminal, toolsets, etc.)
+- `~/.hermes/.env` - API keys and secrets
 
 ## File Dependency Chain
 
 ```
-tools/registry.py  (no deps — imported by all tool files)
-       ↑
-tools/*.py  (each calls registry.register() at import time)
-       ↑
-model_tools.py  (imports tools/registry + triggers tool discovery)
-       ↑
-run_agent.py, cli.py, batch_runner.py, environments/
+tools/*.py → tools/__init__.py → model_tools.py → toolsets.py → toolset_distributions.py
+                                       ↑
+run_agent.py ──────────────────────────┘
+cli.py → run_agent.py (uses AIAgent with quiet_mode=True)
+batch_runner.py → run_agent.py + toolset_distributions.py
 ```
 
+Always ensure consistency between tools, model_tools.py, and toolsets.py when changing any of them.
+
 ---
 
-## AIAgent Class (run_agent.py)
+## AIAgent Class
+
+The main agent is implemented in `run_agent.py`:
 
 ```python
 class AIAgent:
-    def __init__(self,
-        model: str = "anthropic/claude-opus-4.6",
-        max_iterations: int = 90,
+    def __init__(
+        self,
+        model: str = "anthropic/claude-sonnet-4",
+        api_key: str = None,
+        base_url: str = "https://openrouter.ai/api/v1",
+        max_iterations: int = 60,        # Max tool-calling loops
         enabled_toolsets: list = None,
         disabled_toolsets: list = None,
-        quiet_mode: bool = False,
-        save_trajectories: bool = False,
-        platform: str = None,           # "cli", "telegram", etc.
-        session_id: str = None,
-        skip_context_files: bool = False,
-        skip_memory: bool = False,
-        # ... plus provider, api_mode, callbacks, routing params
-    ): ...
-
-    def chat(self, message: str) -> str:
-        """Simple interface — returns final response string."""
-
-    def run_conversation(self, user_message: str, system_message: str = None,
-                         conversation_history: list = None, task_id: str = None) -> dict:
-        """Full interface — returns dict with final_response + messages."""
+        verbose_logging: bool = False,
+        quiet_mode: bool = False,         # Suppress progress output
+        tool_progress_callback: callable = None,  # Called on each tool use
+    ):
+        # Initialize OpenAI client, load tools based on toolsets
+        ...
+    
+    def chat(self, user_message: str, task_id: str = None) -> str:
+        # Main entry point - runs the agent loop
+        ...
 ```
 
 ### Agent Loop
 
-The core loop is inside `run_conversation()` — entirely synchronous:
+The core loop in `_run_agent_loop()`:
+
+```
+1. Add user message to conversation
+2. Call LLM with tools
+3. If LLM returns tool calls:
+   - Execute each tool
+   - Add tool results to conversation
+   - Go to step 2
+4. If LLM returns text response:
+   - Return response to user
+```
 
 ```python
-while api_call_count < self.max_iterations and self.iteration_budget.remaining > 0:
-    response = client.chat.completions.create(model=model, messages=messages, tools=tool_schemas)
+while turns < max_turns:
+    response = client.chat.completions.create(
+        model=model,
+        messages=messages,
+        tools=tool_schemas,
+    )
+    
     if response.tool_calls:
         for tool_call in response.tool_calls:
-            result = handle_function_call(tool_call.name, tool_call.args, task_id)
+            result = await execute_tool(tool_call)
             messages.append(tool_result_message(result))
-        api_call_count += 1
+        turns += 1
     else:
         return response.content
 ```
 
-Messages follow OpenAI format: `{"role": "system/user/assistant/tool", ...}`. Reasoning content is stored in `assistant_msg["reasoning"]`.
+### Conversation Management
+
+Messages are stored as a list of dicts following OpenAI format:
+
+```python
+messages = [
+    {"role": "system", "content": "You are a helpful assistant..."},
+    {"role": "user", "content": "Search for Python tutorials"},
+    {"role": "assistant", "content": None, "tool_calls": [...]},
+    {"role": "tool", "tool_call_id": "...", "content": "..."},
+    {"role": "assistant", "content": "Here's what I found..."},
+]
+```
+
+### Reasoning Model Support
+
+For models that support chain-of-thought reasoning:
+- Extract `reasoning_content` from API responses
+- Store in `assistant_msg["reasoning"]` for trajectory export
+- Pass back via `reasoning_content` field on subsequent turns
 
 ---
 
 ## CLI Architecture (cli.py)
 
-- **Rich** for banner/panels, **prompt_toolkit** for input with autocomplete
-- **KawaiiSpinner** (`agent/display.py`) — animated faces during API calls, `┊` activity feed for tool results
-- `load_cli_config()` in cli.py merges hardcoded defaults + user config YAML
-- **Skin engine** (`hermes_cli/skin_engine.py`) — data-driven CLI theming; initialized from `display.skin` config key at startup; skins customize banner colors, spinner faces/verbs/wings, tool prefix, response box, branding text
-- `process_command()` is a method on `HermesCLI` — dispatches on canonical command name resolved via `resolve_command()` from the central registry
-- Skill slash commands: `agent/skill_commands.py` scans `~/.hermes/skills/`, injects as **user message** (not system prompt) to preserve prompt caching
+The interactive CLI uses:
+- **Rich** - For the welcome banner and styled panels
+- **prompt_toolkit** - For fixed input area with history and `patch_stdout`
+- **KawaiiSpinner** (in run_agent.py) - Animated feedback during API calls and tool execution
 
-### Slash Command Registry (`hermes_cli/commands.py`)
+Key components:
+- `HermesCLI` class - Main CLI controller with commands and conversation loop
+- `load_cli_config()` - Loads config, sets environment variables for terminal
+- `build_welcome_banner()` - Displays ASCII art logo, tools, and skills summary
+- `/commands` - Process user commands like `/help`, `/clear`, `/personality`, etc.
 
-All slash commands are defined in a central `COMMAND_REGISTRY` list of `CommandDef` objects. Every downstream consumer derives from this registry automatically:
+CLI uses `quiet_mode=True` when creating AIAgent to suppress verbose logging.
 
-- **CLI** — `process_command()` resolves aliases via `resolve_command()`, dispatches on canonical name
-- **Gateway** — `GATEWAY_KNOWN_COMMANDS` frozenset for hook emission, `resolve_command()` for dispatch
-- **Gateway help** — `gateway_help_lines()` generates `/help` output
-- **Telegram** — `telegram_bot_commands()` generates the BotCommand menu
-- **Slack** — `slack_subcommand_map()` generates `/hermes` subcommand routing
-- **Autocomplete** — `COMMANDS` flat dict feeds `SlashCommandCompleter`
-- **CLI help** — `COMMANDS_BY_CATEGORY` dict feeds `show_help()`
+### Adding CLI Commands
 
-### Adding a Slash Command
+1. Add to `COMMANDS` dict with description
+2. Add handler in `process_command()` method
+3. For persistent settings, use `save_config_value()` to update config
 
-1. Add a `CommandDef` entry to `COMMAND_REGISTRY` in `hermes_cli/commands.py`:
-```python
-CommandDef("mycommand", "Description of what it does", "Session",
-           aliases=("mc",), args_hint="[arg]"),
+---
+
+## Hermes CLI Commands
+
+The unified `hermes` command provides all functionality:
+
+| Command | Description |
+|---------|-------------|
+| `hermes` | Interactive chat (default) |
+| `hermes chat -q "..."` | Single query mode |
+| `hermes setup` | Configure API keys and settings |
+| `hermes config` | View current configuration |
+| `hermes config edit` | Open config in editor |
+| `hermes config set KEY VAL` | Set a specific value |
+| `hermes config check` | Check for missing config |
+| `hermes config migrate` | Prompt for missing config interactively |
+| `hermes status` | Show configuration status |
+| `hermes doctor` | Diagnose issues |
+| `hermes update` | Update to latest (checks for new config) |
+| `hermes uninstall` | Uninstall (can keep configs for reinstall) |
+| `hermes gateway` | Start messaging gateway |
+| `hermes cron list` | View scheduled jobs |
+| `hermes version` | Show version info |
+
+---
+
+## Messaging Gateway
+
+The gateway connects Hermes to Telegram, Discord, and WhatsApp.
+
+### Configuration (in `~/.hermes/.env`):
+
+```bash
+# Telegram
+TELEGRAM_BOT_TOKEN=123456:ABC-DEF...      # From @BotFather
+TELEGRAM_ALLOWED_USERS=123456789,987654   # Comma-separated user IDs (from @userinfobot)
+
+# Discord  
+DISCORD_BOT_TOKEN=MTIz...                 # From Developer Portal
+DISCORD_ALLOWED_USERS=123456789012345678  # Comma-separated user IDs
+
+# Agent Behavior
+HERMES_MAX_ITERATIONS=60                  # Max tool-calling iterations
+MESSAGING_CWD=/home/myuser                # Terminal working directory for messaging
+
+# Tool Progress (optional)
+HERMES_TOOL_PROGRESS=true                 # Send progress messages
+HERMES_TOOL_PROGRESS_MODE=new             # "new" or "all"
 ```
-2. Add handler in `HermesCLI.process_command()` in `cli.py`:
-```python
-elif canonical == "mycommand":
-    self._handle_mycommand(cmd_original)
-```
-3. If the command is available in the gateway, add a handler in `gateway/run.py`:
-```python
-if canonical == "mycommand":
-    return await self._handle_mycommand(event)
-```
-4. For persistent settings, use `save_config_value()` in `cli.py`
 
-**CommandDef fields:**
-- `name` — canonical name without slash (e.g. `"background"`)
-- `description` — human-readable description
-- `category` — one of `"Session"`, `"Configuration"`, `"Tools & Skills"`, `"Info"`, `"Exit"`
-- `aliases` — tuple of alternative names (e.g. `("bg",)`)
-- `args_hint` — argument placeholder shown in help (e.g. `"<prompt>"`, `"[name]"`)
-- `cli_only` — only available in the interactive CLI
-- `gateway_only` — only available in messaging platforms
-- `gateway_config_gate` — config dotpath (e.g. `"display.tool_progress_command"`); when set on a `cli_only` command, the command becomes available in the gateway if the config value is truthy. `GATEWAY_KNOWN_COMMANDS` always includes config-gated commands so the gateway can dispatch them; help/menus only show them when the gate is open.
+### Working Directory Behavior
 
-**Adding an alias** requires only adding it to the `aliases` tuple on the existing `CommandDef`. No other file changes needed — dispatch, help text, Telegram menu, Slack mapping, and autocomplete all update automatically.
+- **CLI (`hermes` command)**: Uses current directory (`.` → `os.getcwd()`)
+- **Messaging (Telegram/Discord)**: Uses `MESSAGING_CWD` (default: home directory)
+
+This is intentional: CLI users are in a terminal and expect the agent to work in their current directory, while messaging users need a consistent starting location.
+
+### Security (User Allowlists):
+
+**IMPORTANT**: Without an allowlist, anyone who finds your bot can use it!
+
+The gateway checks `{PLATFORM}_ALLOWED_USERS` environment variables:
+- If set: Only listed user IDs can interact with the bot
+- If unset: All users are allowed (dangerous with terminal access!)
+
+Users can find their IDs:
+- **Telegram**: Message [@userinfobot](https://t.me/userinfobot)
+- **Discord**: Enable Developer Mode, right-click name → Copy ID
+
+### Tool Progress Notifications
+
+When `HERMES_TOOL_PROGRESS=true`, the bot sends status messages as it works:
+- `💻 \`ls -la\`...` (terminal commands show the actual command)
+- `🔍 web_search...`
+- `📄 web_extract...`
+
+Modes:
+- `new`: Only when switching to a different tool (less spam)
+- `all`: Every single tool call
+
+### Typing Indicator
+
+The gateway keeps the "typing..." indicator active throughout processing, refreshing every 4 seconds. This lets users know the bot is working even during long tool-calling sequences.
+
+### Platform Toolsets:
+
+Each platform has a dedicated toolset in `toolsets.py`:
+- `hermes-telegram`: Full tools including terminal (with safety checks)
+- `hermes-discord`: Full tools including terminal
+- `hermes-whatsapp`: Full tools including terminal
+
+---
+
+## Configuration System
+
+Configuration files are stored in `~/.hermes/` for easy user access:
+- `~/.hermes/config.yaml` - All settings (model, terminal, compression, etc.)
+- `~/.hermes/.env` - API keys and secrets
+
+### Adding New Configuration Options
+
+When adding new configuration variables, you MUST follow this process:
+
+#### For config.yaml options:
+
+1. Add to `DEFAULT_CONFIG` in `hermes_cli/config.py`
+2. **CRITICAL**: Bump `_config_version` in `DEFAULT_CONFIG` when adding required fields
+3. This triggers migration prompts for existing users on next `hermes update` or `hermes setup`
+
+Example:
+```python
+DEFAULT_CONFIG = {
+    # ... existing config ...
+    
+    "new_feature": {
+        "enabled": True,
+        "option": "default_value",
+    },
+    
+    # BUMP THIS when adding required fields
+    "_config_version": 2,  # Was 1, now 2
+}
+```
+
+#### For .env variables (API keys/secrets):
+
+1. Add to `REQUIRED_ENV_VARS` or `OPTIONAL_ENV_VARS` in `hermes_cli/config.py`
+2. Include metadata for the migration system:
+
+```python
+OPTIONAL_ENV_VARS = {
+    # ... existing vars ...
+    "NEW_API_KEY": {
+        "description": "What this key is for",
+        "prompt": "Display name in prompts",
+        "url": "https://where-to-get-it.com/",
+        "tools": ["tools_it_enables"],  # What tools need this
+        "password": True,  # Mask input
+    },
+}
+```
+
+#### Update related files:
+
+- `hermes_cli/setup.py` - Add prompts in the setup wizard
+- `cli-config.yaml.example` - Add example with comments
+- Update README.md if user-facing
+
+### Config Version Migration
+
+The system uses `_config_version` to detect outdated configs:
+
+1. `check_for_missing_config()` compares user config to `DEFAULT_CONFIG`
+2. `migrate_config()` interactively prompts for missing values
+3. Called automatically by `hermes update` and optionally by `hermes setup`
+
+---
+
+## Environment Variables
+
+API keys are loaded from `~/.hermes/.env`:
+- `OPENROUTER_API_KEY` - Main LLM API access (primary provider)
+- `FIRECRAWL_API_KEY` - Web search/extract tools
+- `BROWSERBASE_API_KEY` / `BROWSERBASE_PROJECT_ID` - Browser automation
+- `FAL_KEY` - Image generation (FLUX model)
+- `NOUS_API_KEY` - Vision and Mixture-of-Agents tools
+
+Terminal tool configuration (in `~/.hermes/config.yaml`):
+- `terminal.backend` - Backend: local, docker, singularity, modal, or ssh
+- `terminal.cwd` - Working directory for CLI ("." = current directory)
+- `terminal.docker_image` - Image for Docker backend
+- `terminal.singularity_image` - Image for Singularity backend
+- `terminal.modal_image` - Image for Modal backend
+- SSH: `TERMINAL_SSH_HOST`, `TERMINAL_SSH_USER`, `TERMINAL_SSH_KEY` in .env
+
+Agent behavior (in `~/.hermes/.env`):
+- `HERMES_MAX_ITERATIONS` - Max tool-calling iterations (default: 60)
+- `MESSAGING_CWD` - Working directory for messaging platforms (default: ~)
+- `HERMES_TOOL_PROGRESS` - Enable tool progress messages (`true`/`false`)
+- `HERMES_TOOL_PROGRESS_MODE` - Progress mode: `new` (tool changes) or `all`
+
+### Dangerous Command Approval
+
+The terminal tool includes safety checks for potentially destructive commands (e.g., `rm -rf`, `DROP TABLE`, `chmod 777`, etc.):
+
+**Behavior by Backend:**
+- **Docker/Singularity/Modal**: Commands run unrestricted (isolated containers)
+- **Local/SSH**: Dangerous commands trigger approval flow
+
+**Approval Flow (CLI):**
+```
+⚠️  Potentially dangerous command detected: recursive delete
+    rm -rf /tmp/test
+
+    [o]nce  |  [s]ession  |  [a]lways  |  [d]eny
+    Choice [o/s/a/D]: 
+```
+
+**Approval Flow (Messaging):**
+- Command is blocked with explanation
+- Agent explains the command was blocked for safety
+- User must add the pattern to their allowlist via `hermes config edit` or run the command directly on their machine
+
+**Configuration:**
+- `command_allowlist` in `~/.hermes/config.yaml` stores permanently allowed patterns
+- Add patterns via "always" approval or edit directly
+
+**Sudo Handling (Messaging):**
+- If sudo fails over messaging, output includes tip to add `SUDO_PASSWORD` to `~/.hermes/.env`
 
 ---
 
 ## Adding New Tools
 
-Requires changes in **3 files**:
+Follow this strict order to maintain consistency:
+
+1. Create `tools/your_tool.py` with:
+   - Handler function (sync or async) returning a JSON string via `json.dumps()`
+   - `check_*_requirements()` function to verify dependencies (e.g., API keys)
+   - Schema definition following OpenAI function-calling format
+
+2. Export in `tools/__init__.py`:
+   - Import the handler and check function
+   - Add to `__all__` list
+
+3. Register in `model_tools.py`:
+   - Add to `TOOLSET_REQUIREMENTS` if it needs API keys
+   - Create `get_*_tool_definitions()` function or add to existing
+   - Add routing in `handle_function_call()` dispatcher
+   - Update `get_all_tool_names()` with the tool name
+   - Update `get_toolset_for_tool()` mapping
+   - Update `get_available_toolsets()` and `check_toolset_requirements()`
+
+4. Add to toolset in `toolsets.py`:
+   - Add to existing toolset or create new one in TOOLSETS dict
+
+5. If the tool requires an API key:
+   - Add to `OPTIONAL_ENV_VARS` in `hermes_cli/config.py`
+   - The tool will be auto-disabled if the key is missing
+
+6. Optionally add to `toolset_distributions.py` for batch processing
+
+### Tool Implementation Pattern
 
-**1. Create `tools/your_tool.py`:**
 ```python
-import json, os
-from tools.registry import registry
+# tools/example_tool.py
+import json
+import os
 
-def check_requirements() -> bool:
+def check_example_requirements() -> bool:
+    """Check if required API keys/dependencies are available."""
     return bool(os.getenv("EXAMPLE_API_KEY"))
 
 def example_tool(param: str, task_id: str = None) -> str:
-    return json.dumps({"success": True, "data": "..."})
-
-registry.register(
-    name="example_tool",
-    toolset="example",
-    schema={"name": "example_tool", "description": "...", "parameters": {...}},
-    handler=lambda args, **kw: example_tool(param=args.get("param", ""), task_id=kw.get("task_id")),
-    check_fn=check_requirements,
-    requires_env=["EXAMPLE_API_KEY"],
-)
+    """Execute the tool and return JSON string result."""
+    try:
+        result = {"success": True, "data": "..."}
+        return json.dumps(result, ensure_ascii=False)
+    except Exception as e:
+        return json.dumps({"error": str(e)}, ensure_ascii=False)
 ```
 
-**2. Add import** in `model_tools.py` `_discover_tools()` list.
+All tool handlers MUST return a JSON string. Never return raw dicts.
 
-**3. Add to `toolsets.py`** — either `_HERMES_CORE_TOOLS` (all platforms) or a new toolset.
+### Dynamic Tool Availability
 
-The registry handles schema collection, dispatch, availability checking, and error wrapping. All handlers MUST return a JSON string.
-
-**Path references in tool schemas**: If the schema description mentions file paths (e.g. default output directories), use `display_hermes_home()` to make them profile-aware. The schema is generated at import time, which is after `_apply_profile_override()` sets `HERMES_HOME`.
-
-**State files**: If a tool stores persistent state (caches, logs, checkpoints), use `get_hermes_home()` for the base directory — never `Path.home() / ".hermes"`. This ensures each profile gets its own state.
-
-**Agent-level tools** (todo, memory): intercepted by `run_agent.py` before `handle_function_call()`. See `todo_tool.py` for the pattern.
-
----
-
-## Adding Configuration
-
-### config.yaml options:
-1. Add to `DEFAULT_CONFIG` in `hermes_cli/config.py`
-2. Bump `_config_version` (currently 5) to trigger migration for existing users
-
-### .env variables:
-1. Add to `OPTIONAL_ENV_VARS` in `hermes_cli/config.py` with metadata:
-```python
-"NEW_API_KEY": {
-    "description": "What it's for",
-    "prompt": "Display name",
-    "url": "https://...",
-    "password": True,
-    "category": "tool",  # provider, tool, messaging, setting
-},
-```
-
-### Config loaders (two separate systems):
-
-| Loader | Used by | Location |
-|--------|---------|----------|
-| `load_cli_config()` | CLI mode | `cli.py` |
-| `load_config()` | `hermes tools`, `hermes setup` | `hermes_cli/config.py` |
-| Direct YAML load | Gateway | `gateway/run.py` |
-
----
-
-## Skin/Theme System
-
-The skin engine (`hermes_cli/skin_engine.py`) provides data-driven CLI visual customization. Skins are **pure data** — no code changes needed to add a new skin.
-
-### Architecture
-
-```
-hermes_cli/skin_engine.py    # SkinConfig dataclass, built-in skins, YAML loader
-~/.hermes/skins/*.yaml       # User-installed custom skins (drop-in)
-```
-
-- `init_skin_from_config()` — called at CLI startup, reads `display.skin` from config
-- `get_active_skin()` — returns cached `SkinConfig` for the current skin
-- `set_active_skin(name)` — switches skin at runtime (used by `/skin` command)
-- `load_skin(name)` — loads from user skins first, then built-ins, then falls back to default
-- Missing skin values inherit from the `default` skin automatically
-
-### What skins customize
-
-| Element | Skin Key | Used By |
-|---------|----------|---------|
-| Banner panel border | `colors.banner_border` | `banner.py` |
-| Banner panel title | `colors.banner_title` | `banner.py` |
-| Banner section headers | `colors.banner_accent` | `banner.py` |
-| Banner dim text | `colors.banner_dim` | `banner.py` |
-| Banner body text | `colors.banner_text` | `banner.py` |
-| Response box border | `colors.response_border` | `cli.py` |
-| Spinner faces (waiting) | `spinner.waiting_faces` | `display.py` |
-| Spinner faces (thinking) | `spinner.thinking_faces` | `display.py` |
-| Spinner verbs | `spinner.thinking_verbs` | `display.py` |
-| Spinner wings (optional) | `spinner.wings` | `display.py` |
-| Tool output prefix | `tool_prefix` | `display.py` |
-| Per-tool emojis | `tool_emojis` | `display.py` → `get_tool_emoji()` |
-| Agent name | `branding.agent_name` | `banner.py`, `cli.py` |
-| Welcome message | `branding.welcome` | `cli.py` |
-| Response box label | `branding.response_label` | `cli.py` |
-| Prompt symbol | `branding.prompt_symbol` | `cli.py` |
-
-### Built-in skins
-
-- `default` — Classic Hermes gold/kawaii (the current look)
-- `ares` — Crimson/bronze war-god theme with custom spinner wings
-- `mono` — Clean grayscale monochrome
-- `slate` — Cool blue developer-focused theme
-
-### Adding a built-in skin
-
-Add to `_BUILTIN_SKINS` dict in `hermes_cli/skin_engine.py`:
+Tools are automatically disabled when their API keys are missing:
 
 ```python
-"mytheme": {
-    "name": "mytheme",
-    "description": "Short description",
-    "colors": { ... },
-    "spinner": { ... },
-    "branding": { ... },
-    "tool_prefix": "┊",
-},
+# In model_tools.py
+TOOLSET_REQUIREMENTS = {
+    "web": {"env_vars": ["FIRECRAWL_API_KEY"]},
+    "browser": {"env_vars": ["BROWSERBASE_API_KEY", "BROWSERBASE_PROJECT_ID"]},
+    "creative": {"env_vars": ["FAL_KEY"]},
+}
 ```
 
-### User skins (YAML)
+The `check_tool_availability()` function determines which tools to include.
 
-Users create `~/.hermes/skins/<name>.yaml`:
+### Stateful Tools
 
-```yaml
-name: cyberpunk
-description: Neon-soaked terminal theme
-
-colors:
-  banner_border: "#FF00FF"
-  banner_title: "#00FFFF"
-  banner_accent: "#FF1493"
-
-spinner:
-  thinking_verbs: ["jacking in", "decrypting", "uploading"]
-  wings:
-    - ["⟨⚡", "⚡⟩"]
-
-branding:
-  agent_name: "Cyber Agent"
-  response_label: " ⚡ Cyber "
-
-tool_prefix: "▏"
-```
-
-Activate with `/skin cyberpunk` or `display.skin: cyberpunk` in config.yaml.
+Tools that maintain state (terminal, browser) require:
+- `task_id` parameter for session isolation between concurrent tasks
+- `cleanup_*()` function to release resources
+- Cleanup is called automatically in run_agent.py after conversation completes
 
 ---
 
-## Important Policies
-### Prompt Caching Must Not Break
+## Trajectory Format
 
-Hermes-Agent ensures caching remains valid throughout a conversation. **Do NOT implement changes that would:**
-- Alter past context mid-conversation
-- Change toolsets mid-conversation
-- Reload memories or rebuild system prompts mid-conversation
+Conversations are saved in ShareGPT format for training:
+```json
+{"from": "system", "value": "System prompt with <tools>...</tools>"}
+{"from": "human", "value": "User message"}
+{"from": "gpt", "value": "<think>reasoning</think>\n<tool_call>{...}</tool_call>"}
+{"from": "tool", "value": "<tool_response>{...}</tool_response>"}
+{"from": "gpt", "value": "Final response"}
+```
 
-Cache-breaking forces dramatically higher costs. The ONLY time we alter context is during context compression.
+Tool calls use `<tool_call>` XML tags, responses use `<tool_response>` tags, reasoning uses `<think>` tags.
 
-### Working Directory Behavior
-- **CLI**: Uses current directory (`.` → `os.getcwd()`)
-- **Messaging**: Uses `MESSAGING_CWD` env var (default: home directory)
+### Trajectory Export
 
-### Background Process Notifications (Gateway)
-
-When `terminal(background=true, check_interval=...)` is used, the gateway runs a watcher that
-pushes status updates to the user's chat. Control verbosity with `display.background_process_notifications`
-in config.yaml (or `HERMES_BACKGROUND_NOTIFICATIONS` env var):
-
-- `all` — running-output updates + final message (default)
-- `result` — only the final completion message
-- `error` — only the final message when exit code != 0
-- `off` — no watcher messages at all
-
----
-
-## Profiles: Multi-Instance Support
-
-Hermes supports **profiles** — multiple fully isolated instances, each with its own
-`HERMES_HOME` directory (config, API keys, memory, sessions, skills, gateway, etc.).
-
-The core mechanism: `_apply_profile_override()` in `hermes_cli/main.py` sets
-`HERMES_HOME` before any module imports. All 119+ references to `get_hermes_home()`
-automatically scope to the active profile.
-
-### Rules for profile-safe code
-
-1. **Use `get_hermes_home()` for all HERMES_HOME paths.** Import from `hermes_constants`.
-   NEVER hardcode `~/.hermes` or `Path.home() / ".hermes"` in code that reads/writes state.
-   ```python
-   # GOOD
-   from hermes_constants import get_hermes_home
-   config_path = get_hermes_home() / "config.yaml"
-
-   # BAD — breaks profiles
-   config_path = Path.home() / ".hermes" / "config.yaml"
-   ```
-
-2. **Use `display_hermes_home()` for user-facing messages.** Import from `hermes_constants`.
-   This returns `~/.hermes` for default or `~/.hermes/profiles/<name>` for profiles.
-   ```python
-   # GOOD
-   from hermes_constants import display_hermes_home
-   print(f"Config saved to {display_hermes_home()}/config.yaml")
-
-   # BAD — shows wrong path for profiles
-   print("Config saved to ~/.hermes/config.yaml")
-   ```
-
-3. **Module-level constants are fine** — they cache `get_hermes_home()` at import time,
-   which is AFTER `_apply_profile_override()` sets the env var. Just use `get_hermes_home()`,
-   not `Path.home() / ".hermes"`.
-
-4. **Tests that mock `Path.home()` must also set `HERMES_HOME`** — since code now uses
-   `get_hermes_home()` (reads env var), not `Path.home() / ".hermes"`:
-   ```python
-   with patch.object(Path, "home", return_value=tmp_path), \
-        patch.dict(os.environ, {"HERMES_HOME": str(tmp_path / ".hermes")}):
-       ...
-   ```
-
-5. **Gateway platform adapters should use token locks** — if the adapter connects with
-   a unique credential (bot token, API key), call `acquire_scoped_lock()` from
-   `gateway.status` in the `connect()`/`start()` method and `release_scoped_lock()` in
-   `disconnect()`/`stop()`. This prevents two profiles from using the same credential.
-   See `gateway/platforms/telegram.py` for the canonical pattern.
-
-6. **Profile operations are HOME-anchored, not HERMES_HOME-anchored** — `_get_profiles_root()`
-   returns `Path.home() / ".hermes" / "profiles"`, NOT `get_hermes_home() / "profiles"`.
-   This is intentional — it lets `hermes -p coder profile list` see all profiles regardless
-   of which one is active.
-
-## Known Pitfalls
-
-### DO NOT hardcode `~/.hermes` paths
-Use `get_hermes_home()` from `hermes_constants` for code paths. Use `display_hermes_home()`
-for user-facing print/log messages. Hardcoding `~/.hermes` breaks profiles — each profile
-has its own `HERMES_HOME` directory. This was the source of 5 bugs fixed in PR #3575.
-
-### DO NOT use `simple_term_menu` for interactive menus
-Rendering bugs in tmux/iTerm2 — ghosting on scroll. Use `curses` (stdlib) instead. See `hermes_cli/tools_config.py` for the pattern.
-
-### DO NOT use `\033[K` (ANSI erase-to-EOL) in spinner/display code
-Leaks as literal `?[K` text under `prompt_toolkit`'s `patch_stdout`. Use space-padding: `f"\r{line}{' ' * pad}"`.
-
-### `_last_resolved_tool_names` is a process-global in `model_tools.py`
-`_run_single_child()` in `delegate_tool.py` saves and restores this global around subagent execution. If you add new code that reads this global, be aware it may be temporarily stale during child agent runs.
-
-### DO NOT hardcode cross-tool references in schema descriptions
-Tool schema descriptions must not mention tools from other toolsets by name (e.g., `browser_navigate` saying "prefer web_search"). Those tools may be unavailable (missing API keys, disabled toolset), causing the model to hallucinate calls to non-existent tools. If a cross-reference is needed, add it dynamically in `get_tool_definitions()` in `model_tools.py` — see the `browser_navigate` / `execute_code` post-processing blocks for the pattern.
-
-### Tests must not write to `~/.hermes/`
-The `_isolate_hermes_home` autouse fixture in `tests/conftest.py` redirects `HERMES_HOME` to a temp dir. Never hardcode `~/.hermes/` paths in tests.
-
-**Profile tests**: When testing profile features, also mock `Path.home()` so that
-`_get_profiles_root()` and `_get_default_hermes_home()` resolve within the temp dir.
-Use the pattern from `tests/hermes_cli/test_profiles.py`:
 ```python
-@pytest.fixture
-def profile_env(tmp_path, monkeypatch):
-    home = tmp_path / ".hermes"
-    home.mkdir()
-    monkeypatch.setattr(Path, "home", lambda: tmp_path)
-    monkeypatch.setenv("HERMES_HOME", str(home))
-    return home
+agent = AIAgent(save_trajectories=True)
+agent.chat("Do something")
+# Saves to trajectories/*.jsonl in ShareGPT format
 ```
 
 ---
 
-## Testing
+## Batch Processing (batch_runner.py)
+
+For processing multiple prompts:
+- Parallel execution with multiprocessing
+- Content-based resume for fault tolerance (matches on prompt text, not indices)
+- Toolset distributions control probabilistic tool availability per prompt
+- Output: `data/<run_name>/trajectories.jsonl` (combined) + individual batch files
 
 ```bash
-source venv/bin/activate
-python -m pytest tests/ -q          # Full suite (~3000 tests, ~3 min)
-python -m pytest tests/test_model_tools.py -q   # Toolset resolution
-python -m pytest tests/test_cli_init.py -q       # CLI config loading
-python -m pytest tests/gateway/ -q               # Gateway tests
-python -m pytest tests/tools/ -q                 # Tool-level tests
+python batch_runner.py \
+    --dataset_file=prompts.jsonl \
+    --batch_size=20 \
+    --num_workers=4 \
+    --run_name=my_run
 ```
 
-Always run the full suite before pushing changes.
+---
+
+## Skills System
+
+Skills are on-demand knowledge documents the agent can load. Located in `skills/` directory:
+
+```
+skills/
+├── mlops/                    # Category folder
+│   ├── axolotl/             # Skill folder
+│   │   ├── SKILL.md         # Main instructions (required)
+│   │   ├── references/      # Additional docs, API specs
+│   │   └── templates/       # Output formats, configs
+│   └── vllm/
+│       └── SKILL.md
+└── example-skill/
+    └── SKILL.md
+```
+
+**Progressive disclosure** (token-efficient):
+1. `skills_categories()` - List category names (~50 tokens)
+2. `skills_list(category)` - Name + description per skill (~3k tokens)
+3. `skill_view(name)` - Full content + tags + linked files
+
+SKILL.md files use YAML frontmatter:
+```yaml
+---
+name: skill-name
+description: Brief description for listing
+tags: [tag1, tag2]
+related_skills: [other-skill]
+version: 1.0.0
+---
+# Skill Content...
+```
+
+Tool files: `tools/skills_tool.py` → `model_tools.py` → `toolsets.py`
+
+---
+
+## Testing Changes
+
+After making changes:
+
+1. Run `hermes doctor` to check setup
+2. Run `hermes config check` to verify config
+3. Test with `hermes chat -q "test message"`
+4. For new config options, test fresh install: `rm -rf ~/.hermes && hermes setup`

CONTRIBUTING.md

@@ -1,660 +0,0 @@
-# Contributing to Hermes Agent
-
-Thank you for contributing to Hermes Agent! This guide covers everything you need: setting up your dev environment, understanding the architecture, deciding what to build, and getting your PR merged.
-
----
-
-## Contribution Priorities
-
-We value contributions in this order:
-
-1. **Bug fixes** — crashes, incorrect behavior, data loss. Always top priority.
-2. **Cross-platform compatibility** — Windows, macOS, different Linux distros, different terminal emulators. We want Hermes to work everywhere.
-3. **Security hardening** — shell injection, prompt injection, path traversal, privilege escalation. See [Security](#security-considerations).
-4. **Performance and robustness** — retry logic, error handling, graceful degradation.
-5. **New skills** — but only broadly useful ones. See [Should it be a Skill or a Tool?](#should-it-be-a-skill-or-a-tool)
-6. **New tools** — rarely needed. Most capabilities should be skills. See below.
-7. **Documentation** — fixes, clarifications, new examples.
-
----
-
-## Should it be a Skill or a Tool?
-
-This is the most common question for new contributors. The answer is almost always **skill**.
-
-### Make it a Skill when:
-
-- The capability can be expressed as instructions + shell commands + existing tools
-- It wraps an external CLI or API that the agent can call via `terminal` or `web_extract`
-- It doesn't need custom Python integration or API key management baked into the agent
-- Examples: arXiv search, git workflows, Docker management, PDF processing, email via CLI tools
-
-### Make it a Tool when:
-
-- It requires end-to-end integration with API keys, auth flows, or multi-component configuration managed by the agent harness
-- It needs custom processing logic that must execute precisely every time (not "best effort" from LLM interpretation)
-- It handles binary data, streaming, or real-time events that can't go through the terminal
-- Examples: browser automation (Browserbase session management), TTS (audio encoding + platform delivery), vision analysis (base64 image handling)
-
-### Should the Skill be bundled?
-
-Bundled skills (in `skills/`) ship with every Hermes install. They should be **broadly useful to most users**:
-
-- Document handling, web research, common dev workflows, system administration
-- Used regularly by a wide range of people
-
-If your skill is official and useful but not universally needed (e.g., a paid service integration, a heavyweight dependency), put it in **`optional-skills/`** — it ships with the repo but isn't activated by default. Users can discover it via `hermes skills browse` (labeled "official") and install it with `hermes skills install` (no third-party warning, builtin trust).
-
-If your skill is specialized, community-contributed, or niche, it's better suited for a **Skills Hub** — upload it to a skills registry and share it in the [Nous Research Discord](https://discord.gg/NousResearch). Users can install it with `hermes skills install`.
-
----
-
-## Development Setup
-
-### Prerequisites
-
-| Requirement | Notes |
-|-------------|-------|
-| **Git** | With `--recurse-submodules` support |
-| **Python 3.11+** | uv will install it if missing |
-| **uv** | Fast Python package manager ([install](https://docs.astral.sh/uv/)) |
-| **Node.js 18+** | Optional — needed for browser tools and WhatsApp bridge |
-
-### Clone and install
-
-```bash
-git clone --recurse-submodules https://github.com/NousResearch/hermes-agent.git
-cd hermes-agent
-
-# Create venv with Python 3.11
-uv venv venv --python 3.11
-export VIRTUAL_ENV="$(pwd)/venv"
-
-# Install with all extras (messaging, cron, CLI menus, dev tools)
-uv pip install -e ".[all,dev]"
-
-# Optional: RL training submodule
-# git submodule update --init tinker-atropos && uv pip install -e "./tinker-atropos"
-
-# Optional: browser tools
-npm install
-```
-
-### Configure for development
-
-```bash
-mkdir -p ~/.hermes/{cron,sessions,logs,memories,skills}
-cp cli-config.yaml.example ~/.hermes/config.yaml
-touch ~/.hermes/.env
-
-# Add at minimum an LLM provider key:
-echo 'OPENROUTER_API_KEY=sk-or-v1-your-key' >> ~/.hermes/.env
-```
-
-### Run
-
-```bash
-# Symlink for global access
-mkdir -p ~/.local/bin
-ln -sf "$(pwd)/venv/bin/hermes" ~/.local/bin/hermes
-
-# Verify
-hermes doctor
-hermes chat -q "Hello"
-```
-
-### Run tests
-
-```bash
-pytest tests/ -v
-```
-
----
-
-## Project Structure
-
-```
-hermes-agent/
-├── run_agent.py              # AIAgent class — core conversation loop, tool dispatch, session persistence
-├── cli.py                    # HermesCLI class — interactive TUI, prompt_toolkit integration
-├── model_tools.py            # Tool orchestration (thin layer over tools/registry.py)
-├── toolsets.py               # Tool groupings and presets (hermes-cli, hermes-telegram, etc.)
-├── hermes_state.py           # SQLite session database with FTS5 full-text search, session titles
-├── batch_runner.py           # Parallel batch processing for trajectory generation
-│
-├── agent/                    # Agent internals (extracted modules)
-│   ├── prompt_builder.py         # System prompt assembly (identity, skills, context files, memory)
-│   ├── context_compressor.py     # Auto-summarization when approaching context limits
-│   ├── auxiliary_client.py       # Resolves auxiliary OpenAI clients (summarization, vision)
-│   ├── display.py                # KawaiiSpinner, tool progress formatting
-│   ├── model_metadata.py         # Model context lengths, token estimation
-│   └── trajectory.py             # Trajectory saving helpers
-│
-├── hermes_cli/               # CLI command implementations
-│   ├── main.py                   # Entry point, argument parsing, command dispatch
-│   ├── config.py                 # Config management, migration, env var definitions
-│   ├── setup.py                  # Interactive setup wizard
-│   ├── auth.py                   # Provider resolution, OAuth, Nous Portal
-│   ├── models.py                 # OpenRouter model selection lists
-│   ├── banner.py                 # Welcome banner, ASCII art
-│   ├── commands.py               # Central slash command registry (CommandDef), autocomplete, gateway helpers
-│   ├── callbacks.py              # Interactive callbacks (clarify, sudo, approval)
-│   ├── doctor.py                 # Diagnostics
-│   ├── skills_hub.py             # Skills Hub CLI + /skills slash command
-│   └── skin_engine.py            # Skin/theme engine — data-driven CLI visual customization
-│
-├── tools/                    # Tool implementations (self-registering)
-│   ├── registry.py               # Central tool registry (schemas, handlers, dispatch)
-│   ├── approval.py               # Dangerous command detection + per-session approval
-│   ├── terminal_tool.py          # Terminal orchestration (sudo, env lifecycle, backends)
-│   ├── file_operations.py        # read_file, write_file, search, patch, etc.
-│   ├── web_tools.py              # web_search, web_extract (Parallel/Firecrawl + Gemini summarization)
-│   ├── vision_tools.py           # Image analysis via multimodal models
-│   ├── delegate_tool.py          # Subagent spawning and parallel task execution
-│   ├── code_execution_tool.py    # Sandboxed Python with RPC tool access
-│   ├── session_search_tool.py    # Search past conversations with FTS5 + summarization
-│   ├── cronjob_tools.py          # Scheduled task management
-│   ├── skill_tools.py            # Skill search, load, manage
-│   └── environments/             # Terminal execution backends
-│       ├── base.py                   # BaseEnvironment ABC
-│       ├── local.py, docker.py, ssh.py, singularity.py, modal.py, daytona.py
-│
-├── gateway/                  # Messaging gateway
-│   ├── run.py                    # GatewayRunner — platform lifecycle, message routing, cron
-│   ├── config.py                 # Platform configuration resolution
-│   ├── session.py                # Session store, context prompts, reset policies
-│   └── platforms/                # Platform adapters
-│       ├── telegram.py, discord_adapter.py, slack.py, whatsapp.py
-│
-├── scripts/                  # Installer and bridge scripts
-│   ├── install.sh                # Linux/macOS installer
-│   ├── install.ps1               # Windows PowerShell installer
-│   └── whatsapp-bridge/          # Node.js WhatsApp bridge (Baileys)
-│
-├── skills/                   # Bundled skills (copied to ~/.hermes/skills/ on install)
-├── optional-skills/          # Official optional skills (discoverable via hub, not activated by default)
-├── environments/             # RL training environments (Atropos integration)
-├── tests/                    # Test suite
-├── website/                  # Documentation site (hermes-agent.nousresearch.com)
-│
-├── cli-config.yaml.example   # Example configuration (copied to ~/.hermes/config.yaml)
-└── AGENTS.md                 # Development guide for AI coding assistants
-```
-
-### User configuration (stored in `~/.hermes/`)
-
-| Path | Purpose |
-|------|---------|
-| `~/.hermes/config.yaml` | Settings (model, terminal, toolsets, compression, etc.) |
-| `~/.hermes/.env` | API keys and secrets |
-| `~/.hermes/auth.json` | OAuth credentials (Nous Portal) |
-| `~/.hermes/skills/` | All active skills (bundled + hub-installed + agent-created) |
-| `~/.hermes/memories/` | Persistent memory (MEMORY.md, USER.md) |
-| `~/.hermes/state.db` | SQLite session database |
-| `~/.hermes/sessions/` | JSON session logs |
-| `~/.hermes/cron/` | Scheduled job data |
-| `~/.hermes/whatsapp/session/` | WhatsApp bridge credentials |
-
----
-
-## Architecture Overview
-
-### Core Loop
-
-```
-User message → AIAgent._run_agent_loop()
-  ├── Build system prompt (prompt_builder.py)
-  ├── Build API kwargs (model, messages, tools, reasoning config)
-  ├── Call LLM (OpenAI-compatible API)
-  ├── If tool_calls in response:
-  │     ├── Execute each tool via registry dispatch
-  │     ├── Add tool results to conversation
-  │     └── Loop back to LLM call
-  ├── If text response:
-  │     ├── Persist session to DB
-  │     └── Return final_response
-  └── Context compression if approaching token limit
-```
-
-### Key Design Patterns
-
-- **Self-registering tools**: Each tool file calls `registry.register()` at import time. `model_tools.py` triggers discovery by importing all tool modules.
-- **Toolset grouping**: Tools are grouped into toolsets (`web`, `terminal`, `file`, `browser`, etc.) that can be enabled/disabled per platform.
-- **Session persistence**: All conversations are stored in SQLite (`hermes_state.py`) with full-text search and unique session titles. JSON logs go to `~/.hermes/sessions/`.
-- **Ephemeral injection**: System prompts and prefill messages are injected at API call time, never persisted to the database or logs.
-- **Provider abstraction**: The agent works with any OpenAI-compatible API. Provider resolution happens at init time (Nous Portal OAuth, OpenRouter API key, or custom endpoint).
-- **Provider routing**: When using OpenRouter, `provider_routing` in config.yaml controls provider selection (sort by throughput/latency/price, allow/ignore specific providers, data retention policies). These are injected as `extra_body.provider` in API requests.
-
----
-
-## Code Style
-
-- **PEP 8** with practical exceptions (we don't enforce strict line length)
-- **Comments**: Only when explaining non-obvious intent, trade-offs, or API quirks. Don't narrate what the code does — `# increment counter` adds nothing
-- **Error handling**: Catch specific exceptions. Log with `logger.warning()`/`logger.error()` — use `exc_info=True` for unexpected errors so stack traces appear in logs
-- **Cross-platform**: Never assume Unix. See [Cross-Platform Compatibility](#cross-platform-compatibility)
-
----
-
-## Adding a New Tool
-
-Before writing a tool, ask: [should this be a skill instead?](#should-it-be-a-skill-or-a-tool)
-
-Tools self-register with the central registry. Each tool file co-locates its schema, handler, and registration:
-
-```python
-"""my_tool — Brief description of what this tool does."""
-
-import json
-from tools.registry import registry
-
-
-def my_tool(param1: str, param2: int = 10, **kwargs) -> str:
-    """Handler. Returns a string result (often JSON)."""
-    result = do_work(param1, param2)
-    return json.dumps(result)
-
-
-MY_TOOL_SCHEMA = {
-    "type": "function",
-    "function": {
-        "name": "my_tool",
-        "description": "What this tool does and when the agent should use it.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "param1": {"type": "string", "description": "What param1 is"},
-                "param2": {"type": "integer", "description": "What param2 is", "default": 10},
-            },
-            "required": ["param1"],
-        },
-    },
-}
-
-
-def _check_requirements() -> bool:
-    """Return True if this tool's dependencies are available."""
-    return True
-
-
-registry.register(
-    name="my_tool",
-    toolset="my_toolset",
-    schema=MY_TOOL_SCHEMA,
-    handler=lambda args, **kw: my_tool(**args, **kw),
-    check_fn=_check_requirements,
-)
-```
-
-Then add the import to `model_tools.py` in the `_modules` list:
-
-```python
-_modules = [
-    # ... existing modules ...
-    "tools.my_tool",
-]
-```
-
-If it's a new toolset, add it to `toolsets.py` and to the relevant platform presets.
-
----
-
-## Adding a Skill
-
-Bundled skills live in `skills/` organized by category. Official optional skills use the same structure in `optional-skills/`:
-
-```
-skills/
-├── research/
-│   └── arxiv/
-│       ├── SKILL.md              # Required: main instructions
-│       └── scripts/              # Optional: helper scripts
-│           └── search_arxiv.py
-├── productivity/
-│   └── ocr-and-documents/
-│       ├── SKILL.md
-│       ├── scripts/
-│       └── references/
-└── ...
-```
-
-### SKILL.md format
-
-```markdown
----
-name: my-skill
-description: Brief description (shown in skill search results)
-version: 1.0.0
-author: Your Name
-license: MIT
-platforms: [macos, linux]          # Optional — restrict to specific OS platforms
-                                   #   Valid: macos, linux, windows
-                                   #   Omit to load on all platforms (default)
-required_environment_variables:    # Optional — secure setup-on-load metadata
-  - name: MY_API_KEY
-    prompt: API key
-    help: Where to get it
-    required_for: full functionality
-prerequisites:                     # Optional legacy runtime requirements
-  env_vars: [MY_API_KEY]           #   Backward-compatible alias for required env vars
-  commands: [curl, jq]             #   Advisory only; does not hide the skill
-metadata:
-  hermes:
-    tags: [Category, Subcategory, Keywords]
-    related_skills: [other-skill-name]
-    fallback_for_toolsets: [web]       # Optional — show only when toolset is unavailable
-    requires_toolsets: [terminal]      # Optional — show only when toolset is available
----
-
-# Skill Title
-
-Brief intro.
-
-## When to Use
-Trigger conditions — when should the agent load this skill?
-
-## Quick Reference
-Table of common commands or API calls.
-
-## Procedure
-Step-by-step instructions the agent follows.
-
-## Pitfalls
-Known failure modes and how to handle them.
-
-## Verification
-How the agent confirms it worked.
-```
-
-### Platform-specific skills
-
-Skills can declare which OS platforms they support via the `platforms` frontmatter field. Skills with this field are automatically hidden from the system prompt, `skills_list()`, and slash commands on incompatible platforms.
-
-```yaml
-platforms: [macos]            # macOS only (e.g., iMessage, Apple Reminders)
-platforms: [macos, linux]     # macOS and Linux
-platforms: [windows]          # Windows only
-```
-
-If the field is omitted or empty, the skill loads on all platforms (backward compatible). See `skills/apple/` for examples of macOS-only skills.
-
-### Conditional skill activation
-
-Skills can declare conditions that control when they appear in the system prompt, based on which tools and toolsets are available in the current session. This is primarily used for **fallback skills** — alternatives that should only be shown when a primary tool is unavailable.
-
-Four fields are supported under `metadata.hermes`:
-
-```yaml
-metadata:
-  hermes:
-    fallback_for_toolsets: [web]      # Show ONLY when these toolsets are unavailable
-    requires_toolsets: [terminal]     # Show ONLY when these toolsets are available
-    fallback_for_tools: [web_search]  # Show ONLY when these specific tools are unavailable
-    requires_tools: [terminal]        # Show ONLY when these specific tools are available
-```
-
-**Semantics:**
-- `fallback_for_*`: The skill is a backup. It is **hidden** when the listed tools/toolsets are available, and **shown** when they are unavailable. Use this for free alternatives to premium tools.
-- `requires_*`: The skill needs certain tools to function. It is **hidden** when the listed tools/toolsets are unavailable. Use this for skills that depend on specific capabilities (e.g., a skill that only makes sense with terminal access).
-- If both are specified, both conditions must be satisfied for the skill to appear.
-- If neither is specified, the skill is always shown (backward compatible).
-
-**Examples:**
-
-```yaml
-# DuckDuckGo search — shown when Firecrawl (web toolset) is unavailable
-metadata:
-  hermes:
-    fallback_for_toolsets: [web]
-
-# Smart home skill — only useful when terminal is available
-metadata:
-  hermes:
-    requires_toolsets: [terminal]
-
-# Local browser fallback — shown when Browserbase is unavailable
-metadata:
-  hermes:
-    fallback_for_toolsets: [browser]
-```
-
-The filtering happens at prompt build time in `agent/prompt_builder.py`. The `build_skills_system_prompt()` function receives the set of available tools and toolsets from the agent and uses `_skill_should_show()` to evaluate each skill's conditions.
-
-### Skill setup metadata
-
-Skills can declare secure setup-on-load metadata via the `required_environment_variables` frontmatter field. Missing values do not hide the skill from discovery; they trigger a CLI-only secure prompt when the skill is actually loaded.
-
-```yaml
-required_environment_variables:
-  - name: TENOR_API_KEY
-    prompt: Tenor API key
-    help: Get a key from https://developers.google.com/tenor
-    required_for: full functionality
-```
-
-The user may skip setup and keep loading the skill. Hermes only exposes metadata (`stored_as`, `skipped`, `validated`) to the model — never the secret value.
-
-Legacy `prerequisites.env_vars` remains supported and is normalized into the new representation.
-
-```yaml
-prerequisites:
-  env_vars: [TENOR_API_KEY]       # Legacy alias for required_environment_variables
-  commands: [curl, jq]            # Advisory CLI checks
-```
-
-Gateway and messaging sessions never collect secrets in-band; they instruct the user to run `hermes setup` or update `~/.hermes/.env` locally.
-
-**When to declare required environment variables:**
-- The skill uses an API key or token that should be collected securely at load time
-- The skill can still be useful if the user skips setup, but may degrade gracefully
-
-**When to declare command prerequisites:**
-- The skill relies on a CLI tool that may not be installed (e.g., `himalaya`, `openhue`, `ddgs`)
-- Treat command checks as guidance, not discovery-time hiding
-
-See `skills/gifs/gif-search/` and `skills/email/himalaya/` for examples.
-
-### Skill guidelines
-
-- **No external dependencies unless absolutely necessary.** Prefer stdlib Python, curl, and existing Hermes tools (`web_extract`, `terminal`, `read_file`).
-- **Progressive disclosure.** Put the most common workflow first. Edge cases and advanced usage go at the bottom.
-- **Include helper scripts** for XML/JSON parsing or complex logic — don't expect the LLM to write parsers inline every time.
-- **Test it.** Run `hermes --toolsets skills -q "Use the X skill to do Y"` and verify the agent follows the instructions correctly.
-
----
-
-## Adding a Skin / Theme
-
-Hermes uses a data-driven skin system — no code changes needed to add a new skin.
-
-**Option A: User skin (YAML file)**
-
-Create `~/.hermes/skins/<name>.yaml`:
-
-```yaml
-name: mytheme
-description: Short description of the theme
-
-colors:
-  banner_border: "#HEX"     # Panel border color
-  banner_title: "#HEX"      # Panel title color
-  banner_accent: "#HEX"     # Section header color
-  banner_dim: "#HEX"        # Muted/dim text color
-  banner_text: "#HEX"       # Body text color
-  response_border: "#HEX"   # Response box border
-
-spinner:
-  waiting_faces: ["(⚔)", "(⛨)"]
-  thinking_faces: ["(⚔)", "(⌁)"]
-  thinking_verbs: ["forging", "plotting"]
-  wings:                     # Optional left/right decorations
-    - ["⟪⚔", "⚔⟫"]
-
-branding:
-  agent_name: "My Agent"
-  welcome: "Welcome message"
-  response_label: " ⚔ Agent "
-  prompt_symbol: "⚔ ❯ "
-
-tool_prefix: "╎"             # Tool output line prefix
-```
-
-All fields are optional — missing values inherit from the default skin.
-
-**Option B: Built-in skin**
-
-Add to `_BUILTIN_SKINS` dict in `hermes_cli/skin_engine.py`. Use the same schema as above but as a Python dict. Built-in skins ship with the package and are always available.
-
-**Activating:**
-- CLI: `/skin mytheme` or set `display.skin: mytheme` in config.yaml
-- Config: `display: { skin: mytheme }`
-
-See `hermes_cli/skin_engine.py` for the full schema and existing skins as examples.
-
----
-
-## Cross-Platform Compatibility
-
-Hermes runs on Linux, macOS, and Windows. When writing code that touches the OS:
-
-### Critical rules
-
-1. **`termios` and `fcntl` are Unix-only.** Always catch both `ImportError` and `NotImplementedError`:
-   ```python
-   try:
-       from simple_term_menu import TerminalMenu
-       menu = TerminalMenu(options)
-       idx = menu.show()
-   except (ImportError, NotImplementedError):
-       # Fallback: numbered menu for Windows
-       for i, opt in enumerate(options):
-           print(f"  {i+1}. {opt}")
-       idx = int(input("Choice: ")) - 1
-   ```
-
-2. **File encoding.** Windows may save `.env` files in `cp1252`. Always handle encoding errors:
-   ```python
-   try:
-       load_dotenv(env_path)
-   except UnicodeDecodeError:
-       load_dotenv(env_path, encoding="latin-1")
-   ```
-
-3. **Process management.** `os.setsid()`, `os.killpg()`, and signal handling differ on Windows. Use platform checks:
-   ```python
-   import platform
-   if platform.system() != "Windows":
-       kwargs["preexec_fn"] = os.setsid
-   ```
-
-4. **Path separators.** Use `pathlib.Path` instead of string concatenation with `/`.
-
-5. **Shell commands in installers.** If you change `scripts/install.sh`, check if the equivalent change is needed in `scripts/install.ps1`.
-
----
-
-## Security Considerations
-
-Hermes has terminal access. Security matters.
-
-### Existing protections
-
-| Layer | Implementation |
-|-------|---------------|
-| **Sudo password piping** | Uses `shlex.quote()` to prevent shell injection |
-| **Dangerous command detection** | Regex patterns in `tools/approval.py` with user approval flow |
-| **Cron prompt injection** | Scanner in `tools/cronjob_tools.py` blocks instruction-override patterns |
-| **Write deny list** | Protected paths (`~/.ssh/authorized_keys`, `/etc/shadow`) resolved via `os.path.realpath()` to prevent symlink bypass |
-| **Skills guard** | Security scanner for hub-installed skills (`tools/skills_guard.py`) |
-| **Code execution sandbox** | `execute_code` child process runs with API keys stripped from environment |
-| **Container hardening** | Docker: all capabilities dropped, no privilege escalation, PID limits, size-limited tmpfs |
-
-### When contributing security-sensitive code
-
-- **Always use `shlex.quote()`** when interpolating user input into shell commands
-- **Resolve symlinks** with `os.path.realpath()` before path-based access control checks
-- **Don't log secrets.** API keys, tokens, and passwords should never appear in log output
-- **Catch broad exceptions** around tool execution so a single failure doesn't crash the agent loop
-- **Test on all platforms** if your change touches file paths, process management, or shell commands
-
-If your PR affects security, note it explicitly in the description.
-
----
-
-## Pull Request Process
-
-### Branch naming
-
-```
-fix/description        # Bug fixes
-feat/description       # New features
-docs/description       # Documentation
-test/description       # Tests
-refactor/description   # Code restructuring
-```
-
-### Before submitting
-
-1. **Run tests**: `pytest tests/ -v`
-2. **Test manually**: Run `hermes` and exercise the code path you changed
-3. **Check cross-platform impact**: If you touch file I/O, process management, or terminal handling, consider Windows and macOS
-4. **Keep PRs focused**: One logical change per PR. Don't mix a bug fix with a refactor with a new feature.
-
-### PR description
-
-Include:
-- **What** changed and **why**
-- **How to test** it (reproduction steps for bugs, usage examples for features)
-- **What platforms** you tested on
-- Reference any related issues
-
-### Commit messages
-
-We use [Conventional Commits](https://www.conventionalcommits.org/):
-
-```
-<type>(<scope>): <description>
-```
-
-| Type | Use for |
-|------|---------|
-| `fix` | Bug fixes |
-| `feat` | New features |
-| `docs` | Documentation |
-| `test` | Tests |
-| `refactor` | Code restructuring (no behavior change) |
-| `chore` | Build, CI, dependency updates |
-
-Scopes: `cli`, `gateway`, `tools`, `skills`, `agent`, `install`, `whatsapp`, `security`, etc.
-
-Examples:
-```
-fix(cli): prevent crash in save_config_value when model is a string
-feat(gateway): add WhatsApp multi-user session isolation
-fix(security): prevent shell injection in sudo password piping
-test(tools): add unit tests for file_operations
-```
-
----
-
-## Reporting Issues
-
-- Use [GitHub Issues](https://github.com/NousResearch/hermes-agent/issues)
-- Include: OS, Python version, Hermes version (`hermes version`), full error traceback
-- Include steps to reproduce
-- Check existing issues before creating duplicates
-- For security vulnerabilities, please report privately
-
----
-
-## Community
-
-- **Discord**: [discord.gg/NousResearch](https://discord.gg/NousResearch) — for questions, showcasing projects, and sharing skills
-- **GitHub Discussions**: For design proposals and architecture discussions
-- **Skills Hub**: Upload specialized skills to a registry and share them with the community
-
----
-
-## License
-
-By contributing, you agree that your contributions will be licensed under the [MIT License](LICENSE).

Dockerfile

@@ -1,25 +0,0 @@
-FROM debian:13.4
-
-# Install system dependencies in one layer, clear APT cache
-RUN apt-get update && \
-    apt-get install -y --no-install-recommends \
-        build-essential nodejs npm python3 python3-pip ripgrep ffmpeg gcc python3-dev libffi-dev && \
-    rm -rf /var/lib/apt/lists/*
-
-COPY . /opt/hermes
-WORKDIR /opt/hermes
-
-# Install Python and Node dependencies in one layer, no cache
-RUN pip install --no-cache-dir -e ".[all]" --break-system-packages && \
-    npm install --prefer-offline --no-audit && \
-    npx playwright install --with-deps chromium --only-shell && \
-    cd /opt/hermes/scripts/whatsapp-bridge && \
-    npm install --prefer-offline --no-audit && \
-    npm cache clean --force
-
-WORKDIR /opt/hermes
-RUN chmod +x /opt/hermes/docker/entrypoint.sh
-
-ENV HERMES_HOME=/opt/data
-VOLUME [ "/opt/data" ]
-ENTRYPOINT [ "/opt/hermes/docker/entrypoint.sh" ]

LICENSE

@@ -1,21 +0,0 @@
-MIT License
-
-Copyright (c) 2025 Nous Research
-
-Permission is hereby granted, free of charge, to any person obtaining a copy
-of this software and associated documentation files (the "Software"), to deal
-in the Software without restriction, including without limitation the rights
-to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
-copies of the Software, and to permit persons to whom the Software is
-furnished to do so, subject to the following conditions:
-
-The above copyright notice and this permission notice shall be included in all
-copies or substantial portions of the Software.
-
-THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
-AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
-LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
-SOFTWARE.

MANIFEST.in

@@ -1,4 +0,0 @@
-graft skills
-graft optional-skills
-global-exclude __pycache__
-global-exclude *.py[cod]

RELEASE_v0.2.0.md

@@ -1,383 +0,0 @@
-# Hermes Agent v0.2.0 (v2026.3.12)
-
-**Release Date:** March 12, 2026
-
-> First tagged release since v0.1.0 (the initial pre-public foundation). In just over two weeks, Hermes Agent went from a small internal project to a full-featured AI agent platform — thanks to an explosion of community contributions. This release covers **216 merged pull requests** from **63 contributors**, resolving **119 issues**.
-
----
-
-## ✨ Highlights
-
-- **Multi-Platform Messaging Gateway** — Telegram, Discord, Slack, WhatsApp, Signal, Email (IMAP/SMTP), and Home Assistant platforms with unified session management, media attachments, and per-platform tool configuration.
-
-- **MCP (Model Context Protocol) Client** — Native MCP support with stdio and HTTP transports, reconnection, resource/prompt discovery, and sampling (server-initiated LLM requests). ([#291](https://github.com/NousResearch/hermes-agent/pull/291) — @0xbyt4, [#301](https://github.com/NousResearch/hermes-agent/pull/301), [#753](https://github.com/NousResearch/hermes-agent/pull/753))
-
-- **Skills Ecosystem** — 70+ bundled and optional skills across 15+ categories with a Skills Hub for community discovery, per-platform enable/disable, conditional activation based on tool availability, and prerequisite validation. ([#743](https://github.com/NousResearch/hermes-agent/pull/743) — @teyrebaz33, [#785](https://github.com/NousResearch/hermes-agent/pull/785) — @teyrebaz33)
-
-- **Centralized Provider Router** — Unified `call_llm()`/`async_call_llm()` API replaces scattered provider logic across vision, summarization, compression, and trajectory saving. All auxiliary consumers route through a single code path with automatic credential resolution. ([#1003](https://github.com/NousResearch/hermes-agent/pull/1003))
-
-- **ACP Server** — VS Code, Zed, and JetBrains editor integration via the Agent Communication Protocol standard. ([#949](https://github.com/NousResearch/hermes-agent/pull/949))
-
-- **CLI Skin/Theme Engine** — Data-driven visual customization: banners, spinners, colors, branding. 7 built-in skins + custom YAML skins.
-
-- **Git Worktree Isolation** — `hermes -w` launches isolated agent sessions in git worktrees for safe parallel work on the same repo. ([#654](https://github.com/NousResearch/hermes-agent/pull/654))
-
-- **Filesystem Checkpoints & Rollback** — Automatic snapshots before destructive operations with `/rollback` to restore. ([#824](https://github.com/NousResearch/hermes-agent/pull/824))
-
-- **3,289 Tests** — From near-zero test coverage to a comprehensive test suite covering agent, gateway, tools, cron, and CLI.
-
----
-
-## 🏗️ Core Agent & Architecture
-
-### Provider & Model Support
-- Centralized provider router with `resolve_provider_client()` + `call_llm()` API ([#1003](https://github.com/NousResearch/hermes-agent/pull/1003))
-- Nous Portal as first-class provider in setup ([#644](https://github.com/NousResearch/hermes-agent/issues/644))
-- OpenAI Codex (Responses API) with ChatGPT subscription support ([#43](https://github.com/NousResearch/hermes-agent/pull/43)) — @grp06
-- Codex OAuth vision support + multimodal content adapter
-- Validate `/model` against live API instead of hardcoded lists
-- Self-hosted Firecrawl support ([#460](https://github.com/NousResearch/hermes-agent/pull/460)) — @caentzminger
-- Kimi Code API support ([#635](https://github.com/NousResearch/hermes-agent/pull/635)) — @christomitov
-- MiniMax model ID update ([#473](https://github.com/NousResearch/hermes-agent/pull/473)) — @tars90percent
-- OpenRouter provider routing configuration (provider_preferences)
-- Nous credential refresh on 401 errors ([#571](https://github.com/NousResearch/hermes-agent/pull/571), [#269](https://github.com/NousResearch/hermes-agent/pull/269)) — @rewbs
-- z.ai/GLM, Kimi/Moonshot, MiniMax, Azure OpenAI as first-class providers
-- Unified `/model` and `/provider` into single view
-
-### Agent Loop & Conversation
-- Simple fallback model for provider resilience ([#740](https://github.com/NousResearch/hermes-agent/pull/740))
-- Shared iteration budget across parent + subagent delegation
-- Iteration budget pressure via tool result injection
-- Configurable subagent provider/model with full credential resolution
-- Handle 413 payload-too-large via compression instead of aborting ([#153](https://github.com/NousResearch/hermes-agent/pull/153)) — @tekelala
-- Retry with rebuilt payload after compression ([#616](https://github.com/NousResearch/hermes-agent/pull/616)) — @tripledoublev
-- Auto-compress pathologically large gateway sessions ([#628](https://github.com/NousResearch/hermes-agent/issues/628))
-- Tool call repair middleware — auto-lowercase and invalid tool handler
-- Reasoning effort configuration and `/reasoning` command ([#921](https://github.com/NousResearch/hermes-agent/pull/921))
-- Detect and block file re-read/search loops after context compression ([#705](https://github.com/NousResearch/hermes-agent/pull/705)) — @0xbyt4
-
-### Session & Memory
-- Session naming with unique titles, auto-lineage, rich listing, and resume by name ([#720](https://github.com/NousResearch/hermes-agent/pull/720))
-- Interactive session browser with search filtering ([#733](https://github.com/NousResearch/hermes-agent/pull/733))
-- Display previous messages when resuming a session ([#734](https://github.com/NousResearch/hermes-agent/pull/734))
-- Honcho AI-native cross-session user modeling ([#38](https://github.com/NousResearch/hermes-agent/pull/38)) — @erosika
-- Proactive async memory flush on session expiry
-- Smart context length probing with persistent caching + banner display
-- `/resume` command for switching to named sessions in gateway
-- Session reset policy for messaging platforms
-
----
-
-## 📱 Messaging Platforms (Gateway)
-
-### Telegram
-- Native file attachments: send_document + send_video
-- Document file processing for PDF, text, and Office files — @tekelala
-- Forum topic session isolation ([#766](https://github.com/NousResearch/hermes-agent/pull/766)) — @spanishflu-est1918
-- Browser screenshot sharing via MEDIA: protocol ([#657](https://github.com/NousResearch/hermes-agent/pull/657))
-- Location support for find-nearby skill
-- TTS voice message accumulation fix ([#176](https://github.com/NousResearch/hermes-agent/pull/176)) — @Bartok9
-- Improved error handling and logging ([#763](https://github.com/NousResearch/hermes-agent/pull/763)) — @aydnOktay
-- Italic regex newline fix + 43 format tests ([#204](https://github.com/NousResearch/hermes-agent/pull/204)) — @0xbyt4
-
-### Discord
-- Channel topic included in session context ([#248](https://github.com/NousResearch/hermes-agent/pull/248)) — @Bartok9
-- DISCORD_ALLOW_BOTS config for bot message filtering ([#758](https://github.com/NousResearch/hermes-agent/pull/758))
-- Document and video support ([#784](https://github.com/NousResearch/hermes-agent/pull/784))
-- Improved error handling and logging ([#761](https://github.com/NousResearch/hermes-agent/pull/761)) — @aydnOktay
-
-### Slack
-- App_mention 404 fix + document/video support ([#784](https://github.com/NousResearch/hermes-agent/pull/784))
-- Structured logging replacing print statements — @aydnOktay
-
-### WhatsApp
-- Native media sending — images, videos, documents ([#292](https://github.com/NousResearch/hermes-agent/pull/292)) — @satelerd
-- Multi-user session isolation ([#75](https://github.com/NousResearch/hermes-agent/pull/75)) — @satelerd
-- Cross-platform port cleanup replacing Linux-only fuser ([#433](https://github.com/NousResearch/hermes-agent/pull/433)) — @Farukest
-- DM interrupt key mismatch fix ([#350](https://github.com/NousResearch/hermes-agent/pull/350)) — @Farukest
-
-### Signal
-- Full Signal messenger gateway via signal-cli-rest-api ([#405](https://github.com/NousResearch/hermes-agent/issues/405))
-- Media URL support in message events ([#871](https://github.com/NousResearch/hermes-agent/pull/871))
-
-### Email (IMAP/SMTP)
-- New email gateway platform — @0xbyt4
-
-### Home Assistant
-- REST tools + WebSocket gateway integration ([#184](https://github.com/NousResearch/hermes-agent/pull/184)) — @0xbyt4
-- Service discovery and enhanced setup
-- Toolset mapping fix ([#538](https://github.com/NousResearch/hermes-agent/pull/538)) — @Himess
-
-### Gateway Core
-- Expose subagent tool calls and thinking to users ([#186](https://github.com/NousResearch/hermes-agent/pull/186)) — @cutepawss
-- Configurable background process watcher notifications ([#840](https://github.com/NousResearch/hermes-agent/pull/840))
-- `edit_message()` for Telegram/Discord/Slack with fallback
-- `/compress`, `/usage`, `/update` slash commands
-- Eliminated 3x SQLite message duplication in gateway sessions ([#873](https://github.com/NousResearch/hermes-agent/pull/873))
-- Stabilize system prompt across gateway turns for cache hits ([#754](https://github.com/NousResearch/hermes-agent/pull/754))
-- MCP server shutdown on gateway exit ([#796](https://github.com/NousResearch/hermes-agent/pull/796)) — @0xbyt4
-- Pass session_db to AIAgent, fixing session_search error ([#108](https://github.com/NousResearch/hermes-agent/pull/108)) — @Bartok9
-- Persist transcript changes in /retry, /undo; fix /reset attribute ([#217](https://github.com/NousResearch/hermes-agent/pull/217)) — @Farukest
-- UTF-8 encoding fix preventing Windows crashes ([#369](https://github.com/NousResearch/hermes-agent/pull/369)) — @ch3ronsa
-
----
-
-## 🖥️ CLI & User Experience
-
-### Interactive CLI
-- Data-driven skin/theme engine — 7 built-in skins (default, ares, mono, slate, poseidon, sisyphus, charizard) + custom YAML skins
-- `/personality` command with custom personality + disable support ([#773](https://github.com/NousResearch/hermes-agent/pull/773)) — @teyrebaz33
-- User-defined quick commands that bypass the agent loop ([#746](https://github.com/NousResearch/hermes-agent/pull/746)) — @teyrebaz33
-- `/reasoning` command for effort level and display toggle ([#921](https://github.com/NousResearch/hermes-agent/pull/921))
-- `/verbose` slash command to toggle debug at runtime ([#94](https://github.com/NousResearch/hermes-agent/pull/94)) — @cesareth
-- `/insights` command — usage analytics, cost estimation & activity patterns ([#552](https://github.com/NousResearch/hermes-agent/pull/552))
-- `/background` command for managing background processes
-- `/help` formatting with command categories
-- Bell-on-complete — terminal bell when agent finishes ([#738](https://github.com/NousResearch/hermes-agent/pull/738))
-- Up/down arrow history navigation
-- Clipboard image paste (Alt+V / Ctrl+V)
-- Loading indicators for slow slash commands ([#882](https://github.com/NousResearch/hermes-agent/pull/882))
-- Spinner flickering fix under patch_stdout ([#91](https://github.com/NousResearch/hermes-agent/pull/91)) — @0xbyt4
-- `--quiet/-Q` flag for programmatic single-query mode
-- `--fuck-it-ship-it` flag to bypass all approval prompts ([#724](https://github.com/NousResearch/hermes-agent/pull/724)) — @dmahan93
-- Tools summary flag ([#767](https://github.com/NousResearch/hermes-agent/pull/767)) — @luisv-1
-- Terminal blinking fix on SSH ([#284](https://github.com/NousResearch/hermes-agent/pull/284)) — @ygd58
-- Multi-line paste detection fix ([#84](https://github.com/NousResearch/hermes-agent/pull/84)) — @0xbyt4
-
-### Setup & Configuration
-- Modular setup wizard with section subcommands and tool-first UX
-- Container resource configuration prompts
-- Backend validation for required binaries
-- Config migration system (currently v7)
-- API keys properly routed to .env instead of config.yaml ([#469](https://github.com/NousResearch/hermes-agent/pull/469)) — @ygd58
-- Atomic write for .env to prevent API key loss on crash ([#954](https://github.com/NousResearch/hermes-agent/pull/954))
-- `hermes tools` — per-platform tool enable/disable with curses UI
-- `hermes doctor` for health checks across all configured providers
-- `hermes update` with auto-restart for gateway service
-- Show update-available notice in CLI banner
-- Multiple named custom providers
-- Shell config detection improvement for PATH setup ([#317](https://github.com/NousResearch/hermes-agent/pull/317)) — @mehmetkr-31
-- Consistent HERMES_HOME and .env path resolution ([#51](https://github.com/NousResearch/hermes-agent/pull/51), [#48](https://github.com/NousResearch/hermes-agent/pull/48)) — @deankerr
-- Docker backend fix on macOS + subagent auth for Nous Portal ([#46](https://github.com/NousResearch/hermes-agent/pull/46)) — @rsavitt
-
----
-
-## 🔧 Tool System
-
-### MCP (Model Context Protocol)
-- Native MCP client with stdio + HTTP transports ([#291](https://github.com/NousResearch/hermes-agent/pull/291) — @0xbyt4, [#301](https://github.com/NousResearch/hermes-agent/pull/301))
-- Sampling support — server-initiated LLM requests ([#753](https://github.com/NousResearch/hermes-agent/pull/753))
-- Resource and prompt discovery
-- Automatic reconnection and security hardening
-- Banner integration, `/reload-mcp` command
-- `hermes tools` UI integration
-
-### Browser
-- Local browser backend — zero-cost headless Chromium (no Browserbase needed)
-- Console/errors tool, annotated screenshots, auto-recording, dogfood QA skill ([#745](https://github.com/NousResearch/hermes-agent/pull/745))
-- Screenshot sharing via MEDIA: on all messaging platforms ([#657](https://github.com/NousResearch/hermes-agent/pull/657))
-
-### Terminal & Execution
-- `execute_code` sandbox with json_parse, shell_quote, retry helpers
-- Docker: custom volume mounts ([#158](https://github.com/NousResearch/hermes-agent/pull/158)) — @Indelwin
-- Daytona cloud sandbox backend ([#451](https://github.com/NousResearch/hermes-agent/pull/451)) — @rovle
-- SSH backend fix ([#59](https://github.com/NousResearch/hermes-agent/pull/59)) — @deankerr
-- Shell noise filtering and login shell execution for environment consistency
-- Head+tail truncation for execute_code stdout overflow
-- Configurable background process notification modes
-
-### File Operations
-- Filesystem checkpoints and `/rollback` command ([#824](https://github.com/NousResearch/hermes-agent/pull/824))
-- Structured tool result hints (next-action guidance) for patch and search_files ([#722](https://github.com/NousResearch/hermes-agent/issues/722))
-- Docker volumes passed to sandbox container config ([#687](https://github.com/NousResearch/hermes-agent/pull/687)) — @manuelschipper
-
----
-
-## 🧩 Skills Ecosystem
-
-### Skills System
-- Per-platform skill enable/disable ([#743](https://github.com/NousResearch/hermes-agent/pull/743)) — @teyrebaz33
-- Conditional skill activation based on tool availability ([#785](https://github.com/NousResearch/hermes-agent/pull/785)) — @teyrebaz33
-- Skill prerequisites — hide skills with unmet dependencies ([#659](https://github.com/NousResearch/hermes-agent/pull/659)) — @kshitijk4poor
-- Optional skills — shipped but not activated by default
-- `hermes skills browse` — paginated hub browsing
-- Skills sub-category organization
-- Platform-conditional skill loading
-- Atomic skill file writes ([#551](https://github.com/NousResearch/hermes-agent/pull/551)) — @aydnOktay
-- Skills sync data loss prevention ([#563](https://github.com/NousResearch/hermes-agent/pull/563)) — @0xbyt4
-- Dynamic skill slash commands for CLI and gateway
-
-### New Skills (selected)
-- **ASCII Art** — pyfiglet (571 fonts), cowsay, image-to-ascii ([#209](https://github.com/NousResearch/hermes-agent/pull/209)) — @0xbyt4
-- **ASCII Video** — Full production pipeline ([#854](https://github.com/NousResearch/hermes-agent/pull/854)) — @SHL0MS
-- **DuckDuckGo Search** — Firecrawl fallback ([#267](https://github.com/NousResearch/hermes-agent/pull/267)) — @gamedevCloudy; DDGS API expansion ([#598](https://github.com/NousResearch/hermes-agent/pull/598)) — @areu01or00
-- **Solana Blockchain** — Wallet balances, USD pricing, token names ([#212](https://github.com/NousResearch/hermes-agent/pull/212)) — @gizdusum
-- **AgentMail** — Agent-owned email inboxes ([#330](https://github.com/NousResearch/hermes-agent/pull/330)) — @teyrebaz33
-- **Polymarket** — Prediction market data (read-only) ([#629](https://github.com/NousResearch/hermes-agent/pull/629))
-- **OpenClaw Migration** — Official migration tool ([#570](https://github.com/NousResearch/hermes-agent/pull/570)) — @unmodeled-tyler
-- **Domain Intelligence** — Passive recon: subdomains, SSL, WHOIS, DNS ([#136](https://github.com/NousResearch/hermes-agent/pull/136)) — @FurkanL0
-- **Superpowers** — Software development skills ([#137](https://github.com/NousResearch/hermes-agent/pull/137)) — @kaos35
-- **Hermes-Atropos** — RL environment development skill ([#815](https://github.com/NousResearch/hermes-agent/pull/815))
-- Plus: arXiv search, OCR/documents, Excalidraw diagrams, YouTube transcripts, GIF search, Pokémon player, Minecraft modpack server, OpenHue (Philips Hue), Google Workspace, Notion, PowerPoint, Obsidian, find-nearby, and 40+ MLOps skills
-
----
-
-## 🔒 Security & Reliability
-
-### Security Hardening
-- Path traversal fix in skill_view — prevented reading arbitrary files ([#220](https://github.com/NousResearch/hermes-agent/issues/220)) — @Farukest
-- Shell injection prevention in sudo password piping ([#65](https://github.com/NousResearch/hermes-agent/pull/65)) — @leonsgithub
-- Dangerous command detection: multiline bypass fix ([#233](https://github.com/NousResearch/hermes-agent/pull/233)) — @Farukest; tee/process substitution patterns ([#280](https://github.com/NousResearch/hermes-agent/pull/280)) — @dogiladeveloper
-- Symlink boundary check fix in skills_guard ([#386](https://github.com/NousResearch/hermes-agent/pull/386)) — @Farukest
-- Symlink bypass fix in write deny list on macOS ([#61](https://github.com/NousResearch/hermes-agent/pull/61)) — @0xbyt4
-- Multi-word prompt injection bypass prevention ([#192](https://github.com/NousResearch/hermes-agent/pull/192)) — @0xbyt4
-- Cron prompt injection scanner bypass fix ([#63](https://github.com/NousResearch/hermes-agent/pull/63)) — @0xbyt4
-- Enforce 0600/0700 file permissions on sensitive files ([#757](https://github.com/NousResearch/hermes-agent/pull/757))
-- .env file permissions restricted to owner-only ([#529](https://github.com/NousResearch/hermes-agent/pull/529)) — @Himess
-- `--force` flag properly blocked from overriding dangerous verdicts ([#388](https://github.com/NousResearch/hermes-agent/pull/388)) — @Farukest
-- FTS5 query sanitization + DB connection leak fix ([#565](https://github.com/NousResearch/hermes-agent/pull/565)) — @0xbyt4
-- Expand secret redaction patterns + config toggle to disable
-- In-memory permanent allowlist to prevent data leak ([#600](https://github.com/NousResearch/hermes-agent/pull/600)) — @alireza78a
-
-### Atomic Writes (data loss prevention)
-- sessions.json ([#611](https://github.com/NousResearch/hermes-agent/pull/611)) — @alireza78a
-- Cron jobs ([#146](https://github.com/NousResearch/hermes-agent/pull/146)) — @alireza78a
-- .env config ([#954](https://github.com/NousResearch/hermes-agent/pull/954))
-- Process checkpoints ([#298](https://github.com/NousResearch/hermes-agent/pull/298)) — @aydnOktay
-- Batch runner ([#297](https://github.com/NousResearch/hermes-agent/pull/297)) — @aydnOktay
-- Skill files ([#551](https://github.com/NousResearch/hermes-agent/pull/551)) — @aydnOktay
-
-### Reliability
-- Guard all print() against OSError for systemd/headless environments ([#963](https://github.com/NousResearch/hermes-agent/pull/963))
-- Reset all retry counters at start of run_conversation ([#607](https://github.com/NousResearch/hermes-agent/pull/607)) — @0xbyt4
-- Return deny on approval callback timeout instead of None ([#603](https://github.com/NousResearch/hermes-agent/pull/603)) — @0xbyt4
-- Fix None message content crashes across codebase ([#277](https://github.com/NousResearch/hermes-agent/pull/277))
-- Fix context overrun crash with local LLM backends ([#403](https://github.com/NousResearch/hermes-agent/pull/403)) — @ch3ronsa
-- Prevent `_flush_sentinel` from leaking to external APIs ([#227](https://github.com/NousResearch/hermes-agent/pull/227)) — @Farukest
-- Prevent conversation_history mutation in callers ([#229](https://github.com/NousResearch/hermes-agent/pull/229)) — @Farukest
-- Fix systemd restart loop ([#614](https://github.com/NousResearch/hermes-agent/pull/614)) — @voidborne-d
-- Close file handles and sockets to prevent fd leaks ([#568](https://github.com/NousResearch/hermes-agent/pull/568) — @alireza78a, [#296](https://github.com/NousResearch/hermes-agent/pull/296) — @alireza78a, [#709](https://github.com/NousResearch/hermes-agent/pull/709) — @memosr)
-- Prevent data loss in clipboard PNG conversion ([#602](https://github.com/NousResearch/hermes-agent/pull/602)) — @0xbyt4
-- Eliminate shell noise from terminal output ([#293](https://github.com/NousResearch/hermes-agent/pull/293)) — @0xbyt4
-- Timezone-aware now() for prompt, cron, and execute_code ([#309](https://github.com/NousResearch/hermes-agent/pull/309)) — @areu01or00
-
-### Windows Compatibility
-- Guard POSIX-only process functions ([#219](https://github.com/NousResearch/hermes-agent/pull/219)) — @Farukest
-- Windows native support via Git Bash + ZIP-based update fallback
-- pywinpty for PTY support ([#457](https://github.com/NousResearch/hermes-agent/pull/457)) — @shitcoinsherpa
-- Explicit UTF-8 encoding on all config/data file I/O ([#458](https://github.com/NousResearch/hermes-agent/pull/458)) — @shitcoinsherpa
-- Windows-compatible path handling ([#354](https://github.com/NousResearch/hermes-agent/pull/354), [#390](https://github.com/NousResearch/hermes-agent/pull/390)) — @Farukest
-- Regex-based search output parsing for drive-letter paths ([#533](https://github.com/NousResearch/hermes-agent/pull/533)) — @Himess
-- Auth store file lock for Windows ([#455](https://github.com/NousResearch/hermes-agent/pull/455)) — @shitcoinsherpa
-
----
-
-## 🐛 Notable Bug Fixes
-
-- Fix DeepSeek V3 tool call parser silently dropping multi-line JSON arguments ([#444](https://github.com/NousResearch/hermes-agent/pull/444)) — @PercyDikec
-- Fix gateway transcript losing 1 message per turn due to offset mismatch ([#395](https://github.com/NousResearch/hermes-agent/pull/395)) — @PercyDikec
-- Fix /retry command silently discarding the agent's final response ([#441](https://github.com/NousResearch/hermes-agent/pull/441)) — @PercyDikec
-- Fix max-iterations retry returning empty string after think-block stripping ([#438](https://github.com/NousResearch/hermes-agent/pull/438)) — @PercyDikec
-- Fix max-iterations retry using hardcoded max_tokens ([#436](https://github.com/NousResearch/hermes-agent/pull/436)) — @Farukest
-- Fix Codex status dict key mismatch ([#448](https://github.com/NousResearch/hermes-agent/pull/448)) and visibility filter ([#446](https://github.com/NousResearch/hermes-agent/pull/446)) — @PercyDikec
-- Strip \<think\> blocks from final user-facing responses ([#174](https://github.com/NousResearch/hermes-agent/pull/174)) — @Bartok9
-- Fix \<think\> block regex stripping visible content when model discusses tags literally ([#786](https://github.com/NousResearch/hermes-agent/issues/786))
-- Fix Mistral 422 errors from leftover finish_reason in assistant messages ([#253](https://github.com/NousResearch/hermes-agent/pull/253)) — @Sertug17
-- Fix OPENROUTER_API_KEY resolution order across all code paths ([#295](https://github.com/NousResearch/hermes-agent/pull/295)) — @0xbyt4
-- Fix OPENAI_BASE_URL API key priority ([#420](https://github.com/NousResearch/hermes-agent/pull/420)) — @manuelschipper
-- Fix Anthropic "prompt is too long" 400 error not detected as context length error ([#813](https://github.com/NousResearch/hermes-agent/issues/813))
-- Fix SQLite session transcript accumulating duplicate messages — 3-4x token inflation ([#860](https://github.com/NousResearch/hermes-agent/issues/860))
-- Fix setup wizard skipping API key prompts on first install ([#748](https://github.com/NousResearch/hermes-agent/pull/748))
-- Fix setup wizard showing OpenRouter model list for Nous Portal ([#575](https://github.com/NousResearch/hermes-agent/pull/575)) — @PercyDikec
-- Fix provider selection not persisting when switching via hermes model ([#881](https://github.com/NousResearch/hermes-agent/pull/881))
-- Fix Docker backend failing when docker not in PATH on macOS ([#889](https://github.com/NousResearch/hermes-agent/pull/889))
-- Fix ClawHub Skills Hub adapter for API endpoint changes ([#286](https://github.com/NousResearch/hermes-agent/pull/286)) — @BP602
-- Fix Honcho auto-enable when API key is present ([#243](https://github.com/NousResearch/hermes-agent/pull/243)) — @Bartok9
-- Fix duplicate 'skills' subparser crash on Python 3.11+ ([#898](https://github.com/NousResearch/hermes-agent/issues/898))
-- Fix memory tool entry parsing when content contains section sign ([#162](https://github.com/NousResearch/hermes-agent/pull/162)) — @aydnOktay
-- Fix piped install silently aborting when interactive prompts fail ([#72](https://github.com/NousResearch/hermes-agent/pull/72)) — @cutepawss
-- Fix false positives in recursive delete detection ([#68](https://github.com/NousResearch/hermes-agent/pull/68)) — @cutepawss
-- Fix Ruff lint warnings across codebase ([#608](https://github.com/NousResearch/hermes-agent/pull/608)) — @JackTheGit
-- Fix Anthropic native base URL fail-fast ([#173](https://github.com/NousResearch/hermes-agent/pull/173)) — @adavyas
-- Fix install.sh creating ~/.hermes before moving Node.js directory ([#53](https://github.com/NousResearch/hermes-agent/pull/53)) — @JoshuaMart
-- Fix SystemExit traceback during atexit cleanup on Ctrl+C ([#55](https://github.com/NousResearch/hermes-agent/pull/55)) — @bierlingm
-- Restore missing MIT license file ([#620](https://github.com/NousResearch/hermes-agent/pull/620)) — @stablegenius49
-
----
-
-## 🧪 Testing
-
-- **3,289 tests** across agent, gateway, tools, cron, and CLI
-- Parallelized test suite with pytest-xdist ([#802](https://github.com/NousResearch/hermes-agent/pull/802)) — @OutThisLife
-- Unit tests batch 1: 8 core modules ([#60](https://github.com/NousResearch/hermes-agent/pull/60)) — @0xbyt4
-- Unit tests batch 2: 8 more modules ([#62](https://github.com/NousResearch/hermes-agent/pull/62)) — @0xbyt4
-- Unit tests batch 3: 8 untested modules ([#191](https://github.com/NousResearch/hermes-agent/pull/191)) — @0xbyt4
-- Unit tests batch 4: 5 security/logic-critical modules ([#193](https://github.com/NousResearch/hermes-agent/pull/193)) — @0xbyt4
-- AIAgent (run_agent.py) unit tests ([#67](https://github.com/NousResearch/hermes-agent/pull/67)) — @0xbyt4
-- Trajectory compressor tests ([#203](https://github.com/NousResearch/hermes-agent/pull/203)) — @0xbyt4
-- Clarify tool tests ([#121](https://github.com/NousResearch/hermes-agent/pull/121)) — @Bartok9
-- Telegram format tests — 43 tests for italic/bold/code rendering ([#204](https://github.com/NousResearch/hermes-agent/pull/204)) — @0xbyt4
-- Vision tools type hints + 42 tests ([#792](https://github.com/NousResearch/hermes-agent/pull/792))
-- Compressor tool-call boundary regression tests ([#648](https://github.com/NousResearch/hermes-agent/pull/648)) — @intertwine
-- Test structure reorganization ([#34](https://github.com/NousResearch/hermes-agent/pull/34)) — @0xbyt4
-- Shell noise elimination + fix 36 test failures ([#293](https://github.com/NousResearch/hermes-agent/pull/293)) — @0xbyt4
-
----
-
-## 🔬 RL & Evaluation Environments
-
-- WebResearchEnv — Multi-step web research RL environment ([#434](https://github.com/NousResearch/hermes-agent/pull/434)) — @jackx707
-- Modal sandbox concurrency limits to avoid deadlocks ([#621](https://github.com/NousResearch/hermes-agent/pull/621)) — @voteblake
-- Hermes-atropos-environments bundled skill ([#815](https://github.com/NousResearch/hermes-agent/pull/815))
-- Local vLLM instance support for evaluation — @dmahan93
-- YC-Bench long-horizon agent benchmark environment
-- OpenThoughts-TBLite evaluation environment and scripts
-
----
-
-## 📚 Documentation
-
-- Full documentation website (Docusaurus) with 37+ pages
-- Comprehensive platform setup guides for Telegram, Discord, Slack, WhatsApp, Signal, Email
-- AGENTS.md — development guide for AI coding assistants
-- CONTRIBUTING.md ([#117](https://github.com/NousResearch/hermes-agent/pull/117)) — @Bartok9
-- Slash commands reference ([#142](https://github.com/NousResearch/hermes-agent/pull/142)) — @Bartok9
-- Comprehensive AGENTS.md accuracy audit ([#732](https://github.com/NousResearch/hermes-agent/pull/732))
-- Skin/theme system documentation
-- MCP documentation and examples
-- Docs accuracy audit — 35+ corrections
-- Documentation typo fixes ([#825](https://github.com/NousResearch/hermes-agent/pull/825), [#439](https://github.com/NousResearch/hermes-agent/pull/439)) — @JackTheGit
-- CLI config precedence and terminology standardization ([#166](https://github.com/NousResearch/hermes-agent/pull/166), [#167](https://github.com/NousResearch/hermes-agent/pull/167), [#168](https://github.com/NousResearch/hermes-agent/pull/168)) — @Jr-kenny
-- Telegram token regex documentation ([#713](https://github.com/NousResearch/hermes-agent/pull/713)) — @VolodymyrBg
-
----
-
-## 👥 Contributors
-
-Thank you to the 63 contributors who made this release possible! In just over two weeks, the Hermes Agent community came together to ship an extraordinary amount of work.
-
-### Core
-- **@teknium1** — 43 PRs: Project lead, core architecture, provider router, sessions, skills, CLI, documentation
-
-### Top Community Contributors
-- **@0xbyt4** — 40 PRs: MCP client, Home Assistant, security fixes (symlink, prompt injection, cron), extensive test coverage (6 batches), ascii-art skill, shell noise elimination, skills sync, Telegram formatting, and dozens more
-- **@Farukest** — 16 PRs: Security hardening (path traversal, dangerous command detection, symlink boundary), Windows compatibility (POSIX guards, path handling), WhatsApp fixes, max-iterations retry, gateway fixes
-- **@aydnOktay** — 11 PRs: Atomic writes (process checkpoints, batch runner, skill files), error handling improvements across Telegram, Discord, code execution, transcription, TTS, and skills
-- **@Bartok9** — 9 PRs: CONTRIBUTING.md, slash commands reference, Discord channel topics, think-block stripping, TTS fix, Honcho fix, session count fix, clarify tests
-- **@PercyDikec** — 7 PRs: DeepSeek V3 parser fix, /retry response discard, gateway transcript offset, Codex status/visibility, max-iterations retry, setup wizard fix
-- **@teyrebaz33** — 5 PRs: Skills enable/disable system, quick commands, personality customization, conditional skill activation
-- **@alireza78a** — 5 PRs: Atomic writes (cron, sessions), fd leak prevention, security allowlist, code execution socket cleanup
-- **@shitcoinsherpa** — 3 PRs: Windows support (pywinpty, UTF-8 encoding, auth store lock)
-- **@Himess** — 3 PRs: Cron/HomeAssistant/Daytona fix, Windows drive-letter parsing, .env permissions
-- **@satelerd** — 2 PRs: WhatsApp native media, multi-user session isolation
-- **@rovle** — 1 PR: Daytona cloud sandbox backend (4 commits)
-- **@erosika** — 1 PR: Honcho AI-native memory integration
-- **@dmahan93** — 1 PR: --fuck-it-ship-it flag + RL environment work
-- **@SHL0MS** — 1 PR: ASCII video skill
-
-### All Contributors
-@0xbyt4, @BP602, @Bartok9, @Farukest, @FurkanL0, @Himess, @Indelwin, @JackTheGit, @JoshuaMart, @Jr-kenny, @OutThisLife, @PercyDikec, @SHL0MS, @Sertug17, @VencentSoliman, @VolodymyrBg, @adavyas, @alireza78a, @areu01or00, @aydnOktay, @batuhankocyigit, @bierlingm, @caentzminger, @cesareth, @ch3ronsa, @christomitov, @cutepawss, @deankerr, @dmahan93, @dogiladeveloper, @dragonkhoi, @erosika, @gamedevCloudy, @gizdusum, @grp06, @intertwine, @jackx707, @jdblackstar, @johnh4098, @kaos35, @kshitijk4poor, @leonsgithub, @luisv-1, @manuelschipper, @mehmetkr-31, @memosr, @PeterFile, @rewbs, @rovle, @rsavitt, @satelerd, @spanishflu-est1918, @stablegenius49, @tars90percent, @tekelala, @teknium1, @teyrebaz33, @tripledoublev, @unmodeled-tyler, @voidborne-d, @voteblake, @ygd58
-
----
-
-**Full Changelog**: [v0.1.0...v2026.3.12](https://github.com/NousResearch/hermes-agent/compare/v0.1.0...v2026.3.12)

RELEASE_v0.3.0.md

@@ -1,377 +0,0 @@
-# Hermes Agent v0.3.0 (v2026.3.17)
-
-**Release Date:** March 17, 2026
-
-> The streaming, plugins, and provider release — unified real-time token delivery, first-class plugin architecture, rebuilt provider system with Vercel AI Gateway, native Anthropic provider, smart approvals, live Chrome CDP browser connect, ACP IDE integration, Honcho memory, voice mode, persistent shell, and 50+ bug fixes across every platform.
-
----
-
-## ✨ Highlights
-
-- **Unified Streaming Infrastructure** — Real-time token-by-token delivery in CLI and all gateway platforms. Responses stream as they're generated instead of arriving as a block. ([#1538](https://github.com/NousResearch/hermes-agent/pull/1538))
-
-- **First-Class Plugin Architecture** — Drop Python files into `~/.hermes/plugins/` to extend Hermes with custom tools, commands, and hooks. No forking required. ([#1544](https://github.com/NousResearch/hermes-agent/pull/1544), [#1555](https://github.com/NousResearch/hermes-agent/pull/1555))
-
-- **Native Anthropic Provider** — Direct Anthropic API calls with Claude Code credential auto-discovery, OAuth PKCE flows, and native prompt caching. No OpenRouter middleman needed. ([#1097](https://github.com/NousResearch/hermes-agent/pull/1097))
-
-- **Smart Approvals + /stop Command** — Codex-inspired approval system that learns which commands are safe and remembers your preferences. `/stop` kills the current agent run immediately. ([#1543](https://github.com/NousResearch/hermes-agent/pull/1543))
-
-- **Honcho Memory Integration** — Async memory writes, configurable recall modes, session title integration, and multi-user isolation in gateway mode. By @erosika. ([#736](https://github.com/NousResearch/hermes-agent/pull/736))
-
-- **Voice Mode** — Push-to-talk in CLI, voice notes in Telegram/Discord, Discord voice channel support, and local Whisper transcription via faster-whisper. ([#1299](https://github.com/NousResearch/hermes-agent/pull/1299), [#1185](https://github.com/NousResearch/hermes-agent/pull/1185), [#1429](https://github.com/NousResearch/hermes-agent/pull/1429))
-
-- **Concurrent Tool Execution** — Multiple independent tool calls now run in parallel via ThreadPoolExecutor, significantly reducing latency for multi-tool turns. ([#1152](https://github.com/NousResearch/hermes-agent/pull/1152))
-
-- **PII Redaction** — When `privacy.redact_pii` is enabled, personally identifiable information is automatically scrubbed before sending context to LLM providers. ([#1542](https://github.com/NousResearch/hermes-agent/pull/1542))
-
-- **`/browser connect` via CDP** — Attach browser tools to a live Chrome instance through Chrome DevTools Protocol. Debug, inspect, and interact with pages you already have open. ([#1549](https://github.com/NousResearch/hermes-agent/pull/1549))
-
-- **Vercel AI Gateway Provider** — Route Hermes through Vercel's AI Gateway for access to their model catalog and infrastructure. ([#1628](https://github.com/NousResearch/hermes-agent/pull/1628))
-
-- **Centralized Provider Router** — Rebuilt provider system with `call_llm` API, unified `/model` command, auto-detect provider on model switch, and direct endpoint overrides for auxiliary/delegation clients. ([#1003](https://github.com/NousResearch/hermes-agent/pull/1003), [#1506](https://github.com/NousResearch/hermes-agent/pull/1506), [#1375](https://github.com/NousResearch/hermes-agent/pull/1375))
-
-- **ACP Server (IDE Integration)** — VS Code, Zed, and JetBrains can now connect to Hermes as an agent backend, with full slash command support. ([#1254](https://github.com/NousResearch/hermes-agent/pull/1254), [#1532](https://github.com/NousResearch/hermes-agent/pull/1532))
-
-- **Persistent Shell Mode** — Local and SSH terminal backends can maintain shell state across tool calls — cd, env vars, and aliases persist. By @alt-glitch. ([#1067](https://github.com/NousResearch/hermes-agent/pull/1067), [#1483](https://github.com/NousResearch/hermes-agent/pull/1483))
-
-- **Agentic On-Policy Distillation (OPD)** — New RL training environment for distilling agent policies, expanding the Atropos training ecosystem. ([#1149](https://github.com/NousResearch/hermes-agent/pull/1149))
-
----
-
-## 🏗️ Core Agent & Architecture
-
-### Provider & Model Support
-- **Centralized provider router** with `call_llm` API and unified `/model` command — switch models and providers seamlessly ([#1003](https://github.com/NousResearch/hermes-agent/pull/1003))
-- **Vercel AI Gateway** provider support ([#1628](https://github.com/NousResearch/hermes-agent/pull/1628))
-- **Auto-detect provider** when switching models via `/model` ([#1506](https://github.com/NousResearch/hermes-agent/pull/1506))
-- **Direct endpoint overrides** for auxiliary and delegation clients — point vision/subagent calls at specific endpoints ([#1375](https://github.com/NousResearch/hermes-agent/pull/1375))
-- **Native Anthropic auxiliary vision** — use Claude's native vision API instead of routing through OpenAI-compatible endpoints ([#1377](https://github.com/NousResearch/hermes-agent/pull/1377))
-- Anthropic OAuth flow improvements — auto-run `claude setup-token`, reauthentication, PKCE state persistence, identity fingerprinting ([#1132](https://github.com/NousResearch/hermes-agent/pull/1132), [#1360](https://github.com/NousResearch/hermes-agent/pull/1360), [#1396](https://github.com/NousResearch/hermes-agent/pull/1396), [#1597](https://github.com/NousResearch/hermes-agent/pull/1597))
-- Fix adaptive thinking without `budget_tokens` for Claude 4.6 models — by @ASRagab ([#1128](https://github.com/NousResearch/hermes-agent/pull/1128))
-- Fix Anthropic cache markers through adapter — by @brandtcormorant ([#1216](https://github.com/NousResearch/hermes-agent/pull/1216))
-- Retry Anthropic 429/529 errors and surface details to users — by @0xbyt4 ([#1585](https://github.com/NousResearch/hermes-agent/pull/1585))
-- Fix Anthropic adapter max_tokens, fallback crash, proxy base_url — by @0xbyt4 ([#1121](https://github.com/NousResearch/hermes-agent/pull/1121))
-- Fix DeepSeek V3 parser dropping multiple parallel tool calls — by @mr-emmett-one ([#1365](https://github.com/NousResearch/hermes-agent/pull/1365), [#1300](https://github.com/NousResearch/hermes-agent/pull/1300))
-- Accept unlisted models with warning instead of rejecting ([#1047](https://github.com/NousResearch/hermes-agent/pull/1047), [#1102](https://github.com/NousResearch/hermes-agent/pull/1102))
-- Skip reasoning params for unsupported OpenRouter models ([#1485](https://github.com/NousResearch/hermes-agent/pull/1485))
-- MiniMax Anthropic API compatibility fix ([#1623](https://github.com/NousResearch/hermes-agent/pull/1623))
-- Custom endpoint `/models` verification and `/v1` base URL suggestion ([#1480](https://github.com/NousResearch/hermes-agent/pull/1480))
-- Resolve delegation providers from `custom_providers` config ([#1328](https://github.com/NousResearch/hermes-agent/pull/1328))
-- Kimi model additions and User-Agent fix ([#1039](https://github.com/NousResearch/hermes-agent/pull/1039))
-- Strip `call_id`/`response_item_id` for Mistral compatibility ([#1058](https://github.com/NousResearch/hermes-agent/pull/1058))
-
-### Agent Loop & Conversation
-- **Anthropic Context Editing API** support ([#1147](https://github.com/NousResearch/hermes-agent/pull/1147))
-- Improved context compaction handoff summaries — compressor now preserves more actionable state ([#1273](https://github.com/NousResearch/hermes-agent/pull/1273))
-- Sync session_id after mid-run context compression ([#1160](https://github.com/NousResearch/hermes-agent/pull/1160))
-- Session hygiene threshold tuned to 50% for more proactive compression ([#1096](https://github.com/NousResearch/hermes-agent/pull/1096), [#1161](https://github.com/NousResearch/hermes-agent/pull/1161))
-- Include session ID in system prompt via `--pass-session-id` flag ([#1040](https://github.com/NousResearch/hermes-agent/pull/1040))
-- Prevent closed OpenAI client reuse across retries ([#1391](https://github.com/NousResearch/hermes-agent/pull/1391))
-- Sanitize chat payloads and provider precedence ([#1253](https://github.com/NousResearch/hermes-agent/pull/1253))
-- Handle dict tool call arguments from Codex and local backends ([#1393](https://github.com/NousResearch/hermes-agent/pull/1393), [#1440](https://github.com/NousResearch/hermes-agent/pull/1440))
-
-### Memory & Sessions
-- **Improve memory prioritization** — user preferences and corrections weighted above procedural knowledge ([#1548](https://github.com/NousResearch/hermes-agent/pull/1548))
-- Tighter memory and session recall guidance in system prompts ([#1329](https://github.com/NousResearch/hermes-agent/pull/1329))
-- Persist CLI token counts to session DB for `/insights` ([#1498](https://github.com/NousResearch/hermes-agent/pull/1498))
-- Keep Honcho recall out of the cached system prefix ([#1201](https://github.com/NousResearch/hermes-agent/pull/1201))
-- Correct `seed_ai_identity` to use `session.add_messages()` ([#1475](https://github.com/NousResearch/hermes-agent/pull/1475))
-- Isolate Honcho session routing for multi-user gateway ([#1500](https://github.com/NousResearch/hermes-agent/pull/1500))
-
----
-
-## 📱 Messaging Platforms (Gateway)
-
-### Gateway Core
-- **System gateway service mode** — run as a system-level systemd service, not just user-level ([#1371](https://github.com/NousResearch/hermes-agent/pull/1371))
-- **Gateway install scope prompts** — choose user vs system scope during setup ([#1374](https://github.com/NousResearch/hermes-agent/pull/1374))
-- **Reasoning hot reload** — change reasoning settings without restarting the gateway ([#1275](https://github.com/NousResearch/hermes-agent/pull/1275))
-- Default group sessions to per-user isolation — no more shared state across users in group chats ([#1495](https://github.com/NousResearch/hermes-agent/pull/1495), [#1417](https://github.com/NousResearch/hermes-agent/pull/1417))
-- Harden gateway restart recovery ([#1310](https://github.com/NousResearch/hermes-agent/pull/1310))
-- Cancel active runs during shutdown ([#1427](https://github.com/NousResearch/hermes-agent/pull/1427))
-- SSL certificate auto-detection for NixOS and non-standard systems ([#1494](https://github.com/NousResearch/hermes-agent/pull/1494))
-- Auto-detect D-Bus session bus for `systemctl --user` on headless servers ([#1601](https://github.com/NousResearch/hermes-agent/pull/1601))
-- Auto-enable systemd linger during gateway install on headless servers ([#1334](https://github.com/NousResearch/hermes-agent/pull/1334))
-- Fall back to module entrypoint when `hermes` is not on PATH ([#1355](https://github.com/NousResearch/hermes-agent/pull/1355))
-- Fix dual gateways on macOS launchd after `hermes update` ([#1567](https://github.com/NousResearch/hermes-agent/pull/1567))
-- Remove recursive ExecStop from systemd units ([#1530](https://github.com/NousResearch/hermes-agent/pull/1530))
-- Prevent logging handler accumulation in gateway mode ([#1251](https://github.com/NousResearch/hermes-agent/pull/1251))
-- Restart on retryable startup failures — by @jplew ([#1517](https://github.com/NousResearch/hermes-agent/pull/1517))
-- Backfill model on gateway sessions after agent runs ([#1306](https://github.com/NousResearch/hermes-agent/pull/1306))
-- PID-based gateway kill and deferred config write ([#1499](https://github.com/NousResearch/hermes-agent/pull/1499))
-
-### Telegram
-- Buffer media groups to prevent self-interruption from photo bursts ([#1341](https://github.com/NousResearch/hermes-agent/pull/1341), [#1422](https://github.com/NousResearch/hermes-agent/pull/1422))
-- Retry on transient TLS failures during connect and send ([#1535](https://github.com/NousResearch/hermes-agent/pull/1535))
-- Harden polling conflict handling ([#1339](https://github.com/NousResearch/hermes-agent/pull/1339))
-- Escape chunk indicators and inline code in MarkdownV2 ([#1478](https://github.com/NousResearch/hermes-agent/pull/1478), [#1626](https://github.com/NousResearch/hermes-agent/pull/1626))
-- Check updater/app state before disconnect ([#1389](https://github.com/NousResearch/hermes-agent/pull/1389))
-
-### Discord
-- `/thread` command with `auto_thread` config and media metadata fixes ([#1178](https://github.com/NousResearch/hermes-agent/pull/1178))
-- Auto-thread on @mention, skip mention text in bot threads ([#1438](https://github.com/NousResearch/hermes-agent/pull/1438))
-- Retry without reply reference for system messages ([#1385](https://github.com/NousResearch/hermes-agent/pull/1385))
-- Preserve native document and video attachment support ([#1392](https://github.com/NousResearch/hermes-agent/pull/1392))
-- Defer discord adapter annotations to avoid optional import crashes ([#1314](https://github.com/NousResearch/hermes-agent/pull/1314))
-
-### Slack
-- Thread handling overhaul — progress messages, responses, and session isolation all respect threads ([#1103](https://github.com/NousResearch/hermes-agent/pull/1103))
-- Formatting, reactions, user resolution, and command improvements ([#1106](https://github.com/NousResearch/hermes-agent/pull/1106))
-- Fix MAX_MESSAGE_LENGTH 3900 → 39000 ([#1117](https://github.com/NousResearch/hermes-agent/pull/1117))
-- File upload fallback preserves thread context — by @0xbyt4 ([#1122](https://github.com/NousResearch/hermes-agent/pull/1122))
-- Improve setup guidance ([#1387](https://github.com/NousResearch/hermes-agent/pull/1387))
-
-### Email
-- Fix IMAP UID tracking and SMTP TLS verification ([#1305](https://github.com/NousResearch/hermes-agent/pull/1305))
-- Add `skip_attachments` option via config.yaml ([#1536](https://github.com/NousResearch/hermes-agent/pull/1536))
-
-### Home Assistant
-- Event filtering closed by default ([#1169](https://github.com/NousResearch/hermes-agent/pull/1169))
-
----
-
-## 🖥️ CLI & User Experience
-
-### Interactive CLI
-- **Persistent CLI status bar** — always-visible model, provider, and token counts ([#1522](https://github.com/NousResearch/hermes-agent/pull/1522))
-- **File path autocomplete** in the input prompt ([#1545](https://github.com/NousResearch/hermes-agent/pull/1545))
-- **`/plan` command** — generate implementation plans from specs ([#1372](https://github.com/NousResearch/hermes-agent/pull/1372), [#1381](https://github.com/NousResearch/hermes-agent/pull/1381))
-- **Major `/rollback` improvements** — richer checkpoint history, clearer UX ([#1505](https://github.com/NousResearch/hermes-agent/pull/1505))
-- **Preload CLI skills on launch** — skills are ready before the first prompt ([#1359](https://github.com/NousResearch/hermes-agent/pull/1359))
-- **Centralized slash command registry** — all commands defined once, consumed everywhere ([#1603](https://github.com/NousResearch/hermes-agent/pull/1603))
-- `/bg` alias for `/background` ([#1590](https://github.com/NousResearch/hermes-agent/pull/1590))
-- Prefix matching for slash commands — `/mod` resolves to `/model` ([#1320](https://github.com/NousResearch/hermes-agent/pull/1320))
-- `/new`, `/reset`, `/clear` now start genuinely fresh sessions ([#1237](https://github.com/NousResearch/hermes-agent/pull/1237))
-- Accept session ID prefixes for session actions ([#1425](https://github.com/NousResearch/hermes-agent/pull/1425))
-- TUI prompt and accent output now respect active skin ([#1282](https://github.com/NousResearch/hermes-agent/pull/1282))
-- Centralize tool emoji metadata in registry + skin integration ([#1484](https://github.com/NousResearch/hermes-agent/pull/1484))
-- "View full command" option added to dangerous command approval — by @teknium1 based on design by community ([#887](https://github.com/NousResearch/hermes-agent/pull/887))
-- Non-blocking startup update check and banner deduplication ([#1386](https://github.com/NousResearch/hermes-agent/pull/1386))
-- `/reasoning` command output ordering and inline think extraction fixes ([#1031](https://github.com/NousResearch/hermes-agent/pull/1031))
-- Verbose mode shows full untruncated output ([#1472](https://github.com/NousResearch/hermes-agent/pull/1472))
-- Fix `/status` to report live state and tokens ([#1476](https://github.com/NousResearch/hermes-agent/pull/1476))
-- Seed a default global SOUL.md ([#1311](https://github.com/NousResearch/hermes-agent/pull/1311))
-
-### Setup & Configuration
-- **OpenClaw migration** during first-time setup — by @kshitijk4poor ([#981](https://github.com/NousResearch/hermes-agent/pull/981))
-- `hermes claw migrate` command + migration docs ([#1059](https://github.com/NousResearch/hermes-agent/pull/1059))
-- Smart vision setup that respects the user's chosen provider ([#1323](https://github.com/NousResearch/hermes-agent/pull/1323))
-- Handle headless setup flows end-to-end ([#1274](https://github.com/NousResearch/hermes-agent/pull/1274))
-- Prefer curses over `simple_term_menu` in setup.py ([#1487](https://github.com/NousResearch/hermes-agent/pull/1487))
-- Show effective model and provider in `/status` ([#1284](https://github.com/NousResearch/hermes-agent/pull/1284))
-- Config set examples use placeholder syntax ([#1322](https://github.com/NousResearch/hermes-agent/pull/1322))
-- Reload .env over stale shell overrides ([#1434](https://github.com/NousResearch/hermes-agent/pull/1434))
-- Fix is_coding_plan NameError crash — by @0xbyt4 ([#1123](https://github.com/NousResearch/hermes-agent/pull/1123))
-- Add missing packages to setuptools config — by @alt-glitch ([#912](https://github.com/NousResearch/hermes-agent/pull/912))
-- Installer: clarify why sudo is needed at every prompt ([#1602](https://github.com/NousResearch/hermes-agent/pull/1602))
-
----
-
-## 🔧 Tool System
-
-### Terminal & Execution
-- **Persistent shell mode** for local and SSH backends — maintain shell state across tool calls — by @alt-glitch ([#1067](https://github.com/NousResearch/hermes-agent/pull/1067), [#1483](https://github.com/NousResearch/hermes-agent/pull/1483))
-- **Tirith pre-exec command scanning** — security layer that analyzes commands before execution ([#1256](https://github.com/NousResearch/hermes-agent/pull/1256))
-- Strip Hermes provider env vars from all subprocess environments ([#1157](https://github.com/NousResearch/hermes-agent/pull/1157), [#1172](https://github.com/NousResearch/hermes-agent/pull/1172), [#1399](https://github.com/NousResearch/hermes-agent/pull/1399), [#1419](https://github.com/NousResearch/hermes-agent/pull/1419)) — initial fix by @eren-karakus0
-- SSH preflight check ([#1486](https://github.com/NousResearch/hermes-agent/pull/1486))
-- Docker backend: make cwd workspace mount explicit opt-in ([#1534](https://github.com/NousResearch/hermes-agent/pull/1534))
-- Add project root to PYTHONPATH in execute_code sandbox ([#1383](https://github.com/NousResearch/hermes-agent/pull/1383))
-- Eliminate execute_code progress spam on gateway platforms ([#1098](https://github.com/NousResearch/hermes-agent/pull/1098))
-- Clearer docker backend preflight errors ([#1276](https://github.com/NousResearch/hermes-agent/pull/1276))
-
-### Browser
-- **`/browser connect`** — attach browser tools to a live Chrome instance via CDP ([#1549](https://github.com/NousResearch/hermes-agent/pull/1549))
-- Improve browser cleanup, local browser PATH setup, and screenshot recovery ([#1333](https://github.com/NousResearch/hermes-agent/pull/1333))
-
-### MCP
-- **Selective tool loading** with utility policies — filter which MCP tools are available ([#1302](https://github.com/NousResearch/hermes-agent/pull/1302))
-- Auto-reload MCP tools when `mcp_servers` config changes without restart ([#1474](https://github.com/NousResearch/hermes-agent/pull/1474))
-- Resolve npx stdio connection failures ([#1291](https://github.com/NousResearch/hermes-agent/pull/1291))
-- Preserve MCP toolsets when saving platform tool config ([#1421](https://github.com/NousResearch/hermes-agent/pull/1421))
-
-### Vision
-- Unify vision backend gating ([#1367](https://github.com/NousResearch/hermes-agent/pull/1367))
-- Surface actual error reason instead of generic message ([#1338](https://github.com/NousResearch/hermes-agent/pull/1338))
-- Make Claude image handling work end-to-end ([#1408](https://github.com/NousResearch/hermes-agent/pull/1408))
-
-### Cron
-- **Compress cron management into one tool** — single `cronjob` tool replaces multiple commands ([#1343](https://github.com/NousResearch/hermes-agent/pull/1343))
-- Suppress duplicate cron sends to auto-delivery targets ([#1357](https://github.com/NousResearch/hermes-agent/pull/1357))
-- Persist cron sessions to SQLite ([#1255](https://github.com/NousResearch/hermes-agent/pull/1255))
-- Per-job runtime overrides (provider, model, base_url) ([#1398](https://github.com/NousResearch/hermes-agent/pull/1398))
-- Atomic write in `save_job_output` to prevent data loss on crash ([#1173](https://github.com/NousResearch/hermes-agent/pull/1173))
-- Preserve thread context for `deliver=origin` ([#1437](https://github.com/NousResearch/hermes-agent/pull/1437))
-
-### Patch Tool
-- Avoid corrupting pipe chars in V4A patch apply ([#1286](https://github.com/NousResearch/hermes-agent/pull/1286))
-- Permissive `block_anchor` thresholds and unicode normalization ([#1539](https://github.com/NousResearch/hermes-agent/pull/1539))
-
-### Delegation
-- Add observability metadata to subagent results (model, tokens, duration, tool trace) ([#1175](https://github.com/NousResearch/hermes-agent/pull/1175))
-
----
-
-## 🧩 Skills Ecosystem
-
-### Skills System
-- **Integrate skills.sh** as a hub source alongside ClawHub ([#1303](https://github.com/NousResearch/hermes-agent/pull/1303))
-- Secure skill env setup on load ([#1153](https://github.com/NousResearch/hermes-agent/pull/1153))
-- Honor policy table for dangerous verdicts ([#1330](https://github.com/NousResearch/hermes-agent/pull/1330))
-- Harden ClawHub skill search exact matches ([#1400](https://github.com/NousResearch/hermes-agent/pull/1400))
-- Fix ClawHub skill install — use `/download` ZIP endpoint ([#1060](https://github.com/NousResearch/hermes-agent/pull/1060))
-- Avoid mislabeling local skills as builtin — by @arceus77-7 ([#862](https://github.com/NousResearch/hermes-agent/pull/862))
-
-### New Skills
-- **Linear** project management ([#1230](https://github.com/NousResearch/hermes-agent/pull/1230))
-- **X/Twitter** via x-cli ([#1285](https://github.com/NousResearch/hermes-agent/pull/1285))
-- **Telephony** — Twilio, SMS, and AI calls ([#1289](https://github.com/NousResearch/hermes-agent/pull/1289))
-- **1Password** — by @arceus77-7 ([#883](https://github.com/NousResearch/hermes-agent/pull/883), [#1179](https://github.com/NousResearch/hermes-agent/pull/1179))
-- **NeuroSkill BCI** integration ([#1135](https://github.com/NousResearch/hermes-agent/pull/1135))
-- **Blender MCP** for 3D modeling ([#1531](https://github.com/NousResearch/hermes-agent/pull/1531))
-- **OSS Security Forensics** ([#1482](https://github.com/NousResearch/hermes-agent/pull/1482))
-- **Parallel CLI** research skill ([#1301](https://github.com/NousResearch/hermes-agent/pull/1301))
-- **OpenCode** CLI skill ([#1174](https://github.com/NousResearch/hermes-agent/pull/1174))
-- **ASCII Video** skill refactored — by @SHL0MS ([#1213](https://github.com/NousResearch/hermes-agent/pull/1213), [#1598](https://github.com/NousResearch/hermes-agent/pull/1598))
-
----
-
-## 🎙️ Voice Mode
-
-- Voice mode foundation — push-to-talk CLI, Telegram/Discord voice notes ([#1299](https://github.com/NousResearch/hermes-agent/pull/1299))
-- Free local Whisper transcription via faster-whisper ([#1185](https://github.com/NousResearch/hermes-agent/pull/1185))
-- Discord voice channel reliability fixes ([#1429](https://github.com/NousResearch/hermes-agent/pull/1429))
-- Restore local STT fallback for gateway voice notes ([#1490](https://github.com/NousResearch/hermes-agent/pull/1490))
-- Honor `stt.enabled: false` across gateway transcription ([#1394](https://github.com/NousResearch/hermes-agent/pull/1394))
-- Fix bogus incapability message on Telegram voice notes (Issue [#1033](https://github.com/NousResearch/hermes-agent/issues/1033))
-
----
-
-## 🔌 ACP (IDE Integration)
-
-- Restore ACP server implementation ([#1254](https://github.com/NousResearch/hermes-agent/pull/1254))
-- Support slash commands in ACP adapter ([#1532](https://github.com/NousResearch/hermes-agent/pull/1532))
-
----
-
-## 🧪 RL Training
-
-- **Agentic On-Policy Distillation (OPD)** environment — new RL training environment for agent policy distillation ([#1149](https://github.com/NousResearch/hermes-agent/pull/1149))
-- Make tinker-atropos RL training fully optional ([#1062](https://github.com/NousResearch/hermes-agent/pull/1062))
-
----
-
-## 🔒 Security & Reliability
-
-### Security Hardening
-- **Tirith pre-exec command scanning** — static analysis of terminal commands before execution ([#1256](https://github.com/NousResearch/hermes-agent/pull/1256))
-- **PII redaction** when `privacy.redact_pii` is enabled ([#1542](https://github.com/NousResearch/hermes-agent/pull/1542))
-- Strip Hermes provider/gateway/tool env vars from all subprocess environments ([#1157](https://github.com/NousResearch/hermes-agent/pull/1157), [#1172](https://github.com/NousResearch/hermes-agent/pull/1172), [#1399](https://github.com/NousResearch/hermes-agent/pull/1399), [#1419](https://github.com/NousResearch/hermes-agent/pull/1419))
-- Docker cwd workspace mount now explicit opt-in — never auto-mount host directories ([#1534](https://github.com/NousResearch/hermes-agent/pull/1534))
-- Escape parens and braces in fork bomb regex pattern ([#1397](https://github.com/NousResearch/hermes-agent/pull/1397))
-- Harden `.worktreeinclude` path containment ([#1388](https://github.com/NousResearch/hermes-agent/pull/1388))
-- Use description as `pattern_key` to prevent approval collisions ([#1395](https://github.com/NousResearch/hermes-agent/pull/1395))
-
-### Reliability
-- Guard init-time stdio writes ([#1271](https://github.com/NousResearch/hermes-agent/pull/1271))
-- Session log writes reuse shared atomic JSON helper ([#1280](https://github.com/NousResearch/hermes-agent/pull/1280))
-- Atomic temp cleanup protected on interrupts ([#1401](https://github.com/NousResearch/hermes-agent/pull/1401))
-
----
-
-## 🐛 Notable Bug Fixes
-
-- **`/status` always showing 0 tokens** — now reports live state (Issue [#1465](https://github.com/NousResearch/hermes-agent/issues/1465), [#1476](https://github.com/NousResearch/hermes-agent/pull/1476))
-- **Custom model endpoints not working** — restored config-saved endpoint resolution (Issue [#1460](https://github.com/NousResearch/hermes-agent/issues/1460), [#1373](https://github.com/NousResearch/hermes-agent/pull/1373))
-- **MCP tools not visible until restart** — auto-reload on config change (Issue [#1036](https://github.com/NousResearch/hermes-agent/issues/1036), [#1474](https://github.com/NousResearch/hermes-agent/pull/1474))
-- **`hermes tools` removing MCP tools** — preserve MCP toolsets when saving (Issue [#1247](https://github.com/NousResearch/hermes-agent/issues/1247), [#1421](https://github.com/NousResearch/hermes-agent/pull/1421))
-- **Terminal subprocesses inheriting `OPENAI_BASE_URL`** breaking external tools (Issue [#1002](https://github.com/NousResearch/hermes-agent/issues/1002), [#1399](https://github.com/NousResearch/hermes-agent/pull/1399))
-- **Background process lost on gateway restart** — improved recovery (Issue [#1144](https://github.com/NousResearch/hermes-agent/issues/1144))
-- **Cron jobs not persisting state** — now stored in SQLite (Issue [#1416](https://github.com/NousResearch/hermes-agent/issues/1416), [#1255](https://github.com/NousResearch/hermes-agent/pull/1255))
-- **Cronjob `deliver: origin` not preserving thread context** (Issue [#1219](https://github.com/NousResearch/hermes-agent/issues/1219), [#1437](https://github.com/NousResearch/hermes-agent/pull/1437))
-- **Gateway systemd service failing to auto-restart** when browser processes orphaned (Issue [#1617](https://github.com/NousResearch/hermes-agent/issues/1617))
-- **`/background` completion report cut off in Telegram** (Issue [#1443](https://github.com/NousResearch/hermes-agent/issues/1443))
-- **Model switching not taking effect** (Issue [#1244](https://github.com/NousResearch/hermes-agent/issues/1244), [#1183](https://github.com/NousResearch/hermes-agent/pull/1183))
-- **`hermes doctor` reporting cronjob as unavailable** (Issue [#878](https://github.com/NousResearch/hermes-agent/issues/878), [#1180](https://github.com/NousResearch/hermes-agent/pull/1180))
-- **WhatsApp bridge messages not received** from mobile (Issue [#1142](https://github.com/NousResearch/hermes-agent/issues/1142))
-- **Setup wizard hanging on headless SSH** (Issue [#905](https://github.com/NousResearch/hermes-agent/issues/905), [#1274](https://github.com/NousResearch/hermes-agent/pull/1274))
-- **Log handler accumulation** degrading gateway performance (Issue [#990](https://github.com/NousResearch/hermes-agent/issues/990), [#1251](https://github.com/NousResearch/hermes-agent/pull/1251))
-- **Gateway NULL model in DB** (Issue [#987](https://github.com/NousResearch/hermes-agent/issues/987), [#1306](https://github.com/NousResearch/hermes-agent/pull/1306))
-- **Strict endpoints rejecting replayed tool_calls** (Issue [#893](https://github.com/NousResearch/hermes-agent/issues/893))
-- **Remaining hardcoded `~/.hermes` paths** — all now respect `HERMES_HOME` (Issue [#892](https://github.com/NousResearch/hermes-agent/issues/892), [#1233](https://github.com/NousResearch/hermes-agent/pull/1233))
-- **Delegate tool not working with custom inference providers** (Issue [#1011](https://github.com/NousResearch/hermes-agent/issues/1011), [#1328](https://github.com/NousResearch/hermes-agent/pull/1328))
-- **Skills Guard blocking official skills** (Issue [#1006](https://github.com/NousResearch/hermes-agent/issues/1006), [#1330](https://github.com/NousResearch/hermes-agent/pull/1330))
-- **Setup writing provider before model selection** (Issue [#1182](https://github.com/NousResearch/hermes-agent/issues/1182))
-- **`GatewayConfig.get()` AttributeError** crashing all message handling (Issue [#1158](https://github.com/NousResearch/hermes-agent/issues/1158), [#1287](https://github.com/NousResearch/hermes-agent/pull/1287))
-- **`/update` hard-failing with "command not found"** (Issue [#1049](https://github.com/NousResearch/hermes-agent/issues/1049))
-- **Image analysis failing silently** (Issue [#1034](https://github.com/NousResearch/hermes-agent/issues/1034), [#1338](https://github.com/NousResearch/hermes-agent/pull/1338))
-- **API `BadRequestError` from `'dict'` object has no attribute `'strip'`** (Issue [#1071](https://github.com/NousResearch/hermes-agent/issues/1071))
-- **Slash commands requiring exact full name** — now uses prefix matching (Issue [#928](https://github.com/NousResearch/hermes-agent/issues/928), [#1320](https://github.com/NousResearch/hermes-agent/pull/1320))
-- **Gateway stops responding when terminal is closed on headless** (Issue [#1005](https://github.com/NousResearch/hermes-agent/issues/1005))
-
----
-
-## 🧪 Testing
-
-- Cover empty cached Anthropic tool-call turns ([#1222](https://github.com/NousResearch/hermes-agent/pull/1222))
-- Fix stale CI assumptions in parser and quick-command coverage ([#1236](https://github.com/NousResearch/hermes-agent/pull/1236))
-- Fix gateway async tests without implicit event loop ([#1278](https://github.com/NousResearch/hermes-agent/pull/1278))
-- Make gateway async tests xdist-safe ([#1281](https://github.com/NousResearch/hermes-agent/pull/1281))
-- Cross-timezone naive timestamp regression for cron ([#1319](https://github.com/NousResearch/hermes-agent/pull/1319))
-- Isolate codex provider tests from local env ([#1335](https://github.com/NousResearch/hermes-agent/pull/1335))
-- Lock retry replacement semantics ([#1379](https://github.com/NousResearch/hermes-agent/pull/1379))
-- Improve error logging in session search tool — by @aydnOktay ([#1533](https://github.com/NousResearch/hermes-agent/pull/1533))
-
----
-
-## 📚 Documentation
-
-- Comprehensive SOUL.md guide ([#1315](https://github.com/NousResearch/hermes-agent/pull/1315))
-- Voice mode documentation ([#1316](https://github.com/NousResearch/hermes-agent/pull/1316), [#1362](https://github.com/NousResearch/hermes-agent/pull/1362))
-- Provider contribution guide ([#1361](https://github.com/NousResearch/hermes-agent/pull/1361))
-- ACP and internal systems implementation guides ([#1259](https://github.com/NousResearch/hermes-agent/pull/1259))
-- Expand Docusaurus coverage across CLI, tools, skills, and skins ([#1232](https://github.com/NousResearch/hermes-agent/pull/1232))
-- Terminal backend and Windows troubleshooting ([#1297](https://github.com/NousResearch/hermes-agent/pull/1297))
-- Skills hub reference section ([#1317](https://github.com/NousResearch/hermes-agent/pull/1317))
-- Checkpoint, /rollback, and git worktrees guide ([#1493](https://github.com/NousResearch/hermes-agent/pull/1493), [#1524](https://github.com/NousResearch/hermes-agent/pull/1524))
-- CLI status bar and /usage reference ([#1523](https://github.com/NousResearch/hermes-agent/pull/1523))
-- Fallback providers + /background command docs ([#1430](https://github.com/NousResearch/hermes-agent/pull/1430))
-- Gateway service scopes docs ([#1378](https://github.com/NousResearch/hermes-agent/pull/1378))
-- Slack thread reply behavior docs ([#1407](https://github.com/NousResearch/hermes-agent/pull/1407))
-- Redesigned landing page with Nous blue palette — by @austinpickett ([#974](https://github.com/NousResearch/hermes-agent/pull/974))
-- Fix several documentation typos — by @JackTheGit ([#953](https://github.com/NousResearch/hermes-agent/pull/953))
-- Stabilize website diagrams ([#1405](https://github.com/NousResearch/hermes-agent/pull/1405))
-- CLI vs messaging quick reference in README ([#1491](https://github.com/NousResearch/hermes-agent/pull/1491))
-- Add search to Docusaurus ([#1053](https://github.com/NousResearch/hermes-agent/pull/1053))
-- Home Assistant integration docs ([#1170](https://github.com/NousResearch/hermes-agent/pull/1170))
-
----
-
-## 👥 Contributors
-
-### Core
-- **@teknium1** — 220+ PRs spanning every area of the codebase
-
-### Top Community Contributors
-
-- **@0xbyt4** (4 PRs) — Anthropic adapter fixes (max_tokens, fallback crash, 429/529 retry), Slack file upload thread context, setup NameError fix
-- **@erosika** (1 PR) — Honcho memory integration: async writes, memory modes, session title integration
-- **@SHL0MS** (2 PRs) — ASCII video skill design patterns and refactoring
-- **@alt-glitch** (2 PRs) — Persistent shell mode for local/SSH backends, setuptools packaging fix
-- **@arceus77-7** (2 PRs) — 1Password skill, fix skills list mislabeling
-- **@kshitijk4poor** (1 PR) — OpenClaw migration during setup wizard
-- **@ASRagab** (1 PR) — Fix adaptive thinking for Claude 4.6 models
-- **@eren-karakus0** (1 PR) — Strip Hermes provider env vars from subprocess environment
-- **@mr-emmett-one** (1 PR) — Fix DeepSeek V3 parser multi-tool call support
-- **@jplew** (1 PR) — Gateway restart on retryable startup failures
-- **@brandtcormorant** (1 PR) — Fix Anthropic cache control for empty text blocks
-- **@aydnOktay** (1 PR) — Improve error logging in session search tool
-- **@austinpickett** (1 PR) — Landing page redesign with Nous blue palette
-- **@JackTheGit** (1 PR) — Documentation typo fixes
-
-### All Contributors
-
-@0xbyt4, @alt-glitch, @arceus77-7, @ASRagab, @austinpickett, @aydnOktay, @brandtcormorant, @eren-karakus0, @erosika, @JackTheGit, @jplew, @kshitijk4poor, @mr-emmett-one, @SHL0MS, @teknium1
-
----
-
-**Full Changelog**: [v2026.3.12...v2026.3.17](https://github.com/NousResearch/hermes-agent/compare/v2026.3.12...v2026.3.17)

RELEASE_v0.4.0.md

@@ -1,400 +0,0 @@
-# Hermes Agent v0.4.0 (v2026.3.23)
-
-**Release Date:** March 23, 2026
-
-> The platform expansion release — OpenAI-compatible API server, 6 new messaging adapters, 4 new inference providers, MCP server management with OAuth 2.1, @ context references, gateway prompt caching, streaming enabled by default, and a sweeping reliability pass with 200+ bug fixes.
-
----
-
-## ✨ Highlights
-
-- **OpenAI-compatible API server** — Expose Hermes as an `/v1/chat/completions` endpoint with a new `/api/jobs` REST API for cron job management, hardened with input limits, field whitelists, SQLite-backed response persistence, and CORS origin protection ([#1756](https://github.com/NousResearch/hermes-agent/pull/1756), [#2450](https://github.com/NousResearch/hermes-agent/pull/2450), [#2456](https://github.com/NousResearch/hermes-agent/pull/2456), [#2451](https://github.com/NousResearch/hermes-agent/pull/2451), [#2472](https://github.com/NousResearch/hermes-agent/pull/2472))
-
-- **6 new messaging platform adapters** — Signal, DingTalk, SMS (Twilio), Mattermost, Matrix, and Webhook adapters join Telegram, Discord, and WhatsApp. Gateway auto-reconnects failed platforms with exponential backoff ([#2206](https://github.com/NousResearch/hermes-agent/pull/2206), [#1685](https://github.com/NousResearch/hermes-agent/pull/1685), [#1688](https://github.com/NousResearch/hermes-agent/pull/1688), [#1683](https://github.com/NousResearch/hermes-agent/pull/1683), [#2166](https://github.com/NousResearch/hermes-agent/pull/2166), [#2584](https://github.com/NousResearch/hermes-agent/pull/2584))
-
-- **@ context references** — Claude Code-style `@file` and `@url` context injection with tab completions in the CLI ([#2343](https://github.com/NousResearch/hermes-agent/pull/2343), [#2482](https://github.com/NousResearch/hermes-agent/pull/2482))
-
-- **4 new inference providers** — GitHub Copilot (OAuth + token validation), Alibaba Cloud / DashScope, Kilo Code, and OpenCode Zen/Go ([#1924](https://github.com/NousResearch/hermes-agent/pull/1924), [#1879](https://github.com/NousResearch/hermes-agent/pull/1879) by @mchzimm, [#1673](https://github.com/NousResearch/hermes-agent/pull/1673), [#1666](https://github.com/NousResearch/hermes-agent/pull/1666), [#1650](https://github.com/NousResearch/hermes-agent/pull/1650))
-
-- **MCP server management CLI** — `hermes mcp` commands for installing, configuring, and authenticating MCP servers with full OAuth 2.1 PKCE flow ([#2465](https://github.com/NousResearch/hermes-agent/pull/2465))
-
-- **Gateway prompt caching** — Cache AIAgent instances per session, preserving Anthropic prompt cache across turns for dramatic cost reduction on long conversations ([#2282](https://github.com/NousResearch/hermes-agent/pull/2282), [#2284](https://github.com/NousResearch/hermes-agent/pull/2284), [#2361](https://github.com/NousResearch/hermes-agent/pull/2361))
-
-- **Context compression overhaul** — Structured summaries with iterative updates, token-budget tail protection, configurable summary endpoint, and fallback model support ([#2323](https://github.com/NousResearch/hermes-agent/pull/2323), [#1727](https://github.com/NousResearch/hermes-agent/pull/1727), [#2224](https://github.com/NousResearch/hermes-agent/pull/2224))
-
-- **Streaming enabled by default** — CLI streaming on by default with proper spinner/tool progress display during streaming mode, plus extensive linebreak and concatenation fixes ([#2340](https://github.com/NousResearch/hermes-agent/pull/2340), [#2161](https://github.com/NousResearch/hermes-agent/pull/2161), [#2258](https://github.com/NousResearch/hermes-agent/pull/2258))
-
----
-
-## 🖥️ CLI & User Experience
-
-### New Commands & Interactions
-- **@ context completions** — Tab-completable `@file`/`@url` references that inject file content or web pages into the conversation ([#2482](https://github.com/NousResearch/hermes-agent/pull/2482), [#2343](https://github.com/NousResearch/hermes-agent/pull/2343))
-- **`/statusbar`** — Toggle a persistent config bar showing model + provider info in the prompt ([#2240](https://github.com/NousResearch/hermes-agent/pull/2240), [#1917](https://github.com/NousResearch/hermes-agent/pull/1917))
-- **`/queue`** — Queue prompts for the agent without interrupting the current run ([#2191](https://github.com/NousResearch/hermes-agent/pull/2191), [#2469](https://github.com/NousResearch/hermes-agent/pull/2469))
-- **`/permission`** — Switch approval mode dynamically during a session ([#2207](https://github.com/NousResearch/hermes-agent/pull/2207))
-- **`/browser`** — Interactive browser sessions from the CLI ([#2273](https://github.com/NousResearch/hermes-agent/pull/2273), [#1814](https://github.com/NousResearch/hermes-agent/pull/1814))
-- **`/cost`** — Live pricing and usage tracking in gateway mode ([#2180](https://github.com/NousResearch/hermes-agent/pull/2180))
-- **`/approve` and `/deny`** — Replaced bare text approval in gateway with explicit commands ([#2002](https://github.com/NousResearch/hermes-agent/pull/2002))
-
-### Streaming & Display
-- Streaming enabled by default in CLI ([#2340](https://github.com/NousResearch/hermes-agent/pull/2340))
-- Show spinners and tool progress during streaming mode ([#2161](https://github.com/NousResearch/hermes-agent/pull/2161))
-- Show reasoning/thinking blocks when `show_reasoning` enabled ([#2118](https://github.com/NousResearch/hermes-agent/pull/2118))
-- Context pressure warnings for CLI and gateway ([#2159](https://github.com/NousResearch/hermes-agent/pull/2159))
-- Fix: streaming chunks concatenated without whitespace ([#2258](https://github.com/NousResearch/hermes-agent/pull/2258))
-- Fix: iteration boundary linebreak prevents stream concatenation ([#2413](https://github.com/NousResearch/hermes-agent/pull/2413))
-- Fix: defer streaming linebreak to prevent blank line stacking ([#2473](https://github.com/NousResearch/hermes-agent/pull/2473))
-- Fix: suppress spinner animation in non-TTY environments ([#2216](https://github.com/NousResearch/hermes-agent/pull/2216))
-- Fix: display provider and endpoint in API error messages ([#2266](https://github.com/NousResearch/hermes-agent/pull/2266))
-- Fix: resolve garbled ANSI escape codes in status printouts ([#2448](https://github.com/NousResearch/hermes-agent/pull/2448))
-- Fix: update gold ANSI color to true-color format ([#2246](https://github.com/NousResearch/hermes-agent/pull/2246))
-- Fix: normalize toolset labels and use skin colors in banner ([#1912](https://github.com/NousResearch/hermes-agent/pull/1912))
-
-### CLI Polish
-- Fix: prevent 'Press ENTER to continue...' on exit ([#2555](https://github.com/NousResearch/hermes-agent/pull/2555))
-- Fix: flush stdout during agent loop to prevent macOS display freeze ([#1654](https://github.com/NousResearch/hermes-agent/pull/1654))
-- Fix: show human-readable error when `hermes setup` hits permissions error ([#2196](https://github.com/NousResearch/hermes-agent/pull/2196))
-- Fix: `/stop` command crash + UnboundLocalError in streaming media delivery ([#2463](https://github.com/NousResearch/hermes-agent/pull/2463))
-- Fix: allow custom/local endpoints without API key ([#2556](https://github.com/NousResearch/hermes-agent/pull/2556))
-- Fix: Kitty keyboard protocol Shift+Enter for Ghostty/WezTerm (attempted + reverted due to prompt_toolkit crash) ([#2345](https://github.com/NousResearch/hermes-agent/pull/2345), [#2349](https://github.com/NousResearch/hermes-agent/pull/2349))
-
-### Configuration
-- **`${ENV_VAR}` substitution** in config.yaml ([#2684](https://github.com/NousResearch/hermes-agent/pull/2684))
-- **Real-time config reload** — config.yaml changes apply without restart ([#2210](https://github.com/NousResearch/hermes-agent/pull/2210))
-- **`custom_models.yaml`** for user-managed model additions ([#2214](https://github.com/NousResearch/hermes-agent/pull/2214))
-- **Priority-based context file selection** + CLAUDE.md support ([#2301](https://github.com/NousResearch/hermes-agent/pull/2301))
-- **Merge nested YAML sections** instead of replacing on config update ([#2213](https://github.com/NousResearch/hermes-agent/pull/2213))
-- Fix: config.yaml provider key overrides env var silently ([#2272](https://github.com/NousResearch/hermes-agent/pull/2272))
-- Fix: log warning instead of silently swallowing config.yaml errors ([#2683](https://github.com/NousResearch/hermes-agent/pull/2683))
-- Fix: disabled toolsets re-enable themselves after `hermes tools` ([#2268](https://github.com/NousResearch/hermes-agent/pull/2268))
-- Fix: platform default toolsets silently override tool deselection ([#2624](https://github.com/NousResearch/hermes-agent/pull/2624))
-- Fix: honor bare YAML `approvals.mode: off` ([#2620](https://github.com/NousResearch/hermes-agent/pull/2620))
-- Fix: `hermes update` use `.[all]` extras with fallback ([#1728](https://github.com/NousResearch/hermes-agent/pull/1728))
-- Fix: `hermes update` prompt before resetting working tree on stash conflicts ([#2390](https://github.com/NousResearch/hermes-agent/pull/2390))
-- Fix: use git pull --rebase in update/install to avoid divergent branch error ([#2274](https://github.com/NousResearch/hermes-agent/pull/2274))
-- Fix: add zprofile fallback and create zshrc on fresh macOS installs ([#2320](https://github.com/NousResearch/hermes-agent/pull/2320))
-- Fix: remove `ANTHROPIC_BASE_URL` env var to avoid collisions ([#1675](https://github.com/NousResearch/hermes-agent/pull/1675))
-- Fix: don't ask IMAP password if already in keyring or env ([#2212](https://github.com/NousResearch/hermes-agent/pull/2212))
-- Fix: OpenCode Zen/Go show OpenRouter models instead of their own ([#2277](https://github.com/NousResearch/hermes-agent/pull/2277))
-
----
-
-## 🏗️ Core Agent & Architecture
-
-### New Providers
-- **GitHub Copilot** — Full OAuth auth, API routing, token validation, and 400k context. ([#1924](https://github.com/NousResearch/hermes-agent/pull/1924), [#1896](https://github.com/NousResearch/hermes-agent/pull/1896), [#1879](https://github.com/NousResearch/hermes-agent/pull/1879) by @mchzimm, [#2507](https://github.com/NousResearch/hermes-agent/pull/2507))
-- **Alibaba Cloud / DashScope** — Full integration with DashScope v1 runtime, model dot preservation, and 401 auth fixes ([#1673](https://github.com/NousResearch/hermes-agent/pull/1673), [#2332](https://github.com/NousResearch/hermes-agent/pull/2332), [#2459](https://github.com/NousResearch/hermes-agent/pull/2459))
-- **Kilo Code** — First-class inference provider ([#1666](https://github.com/NousResearch/hermes-agent/pull/1666))
-- **OpenCode Zen and OpenCode Go** — New provider backends ([#1650](https://github.com/NousResearch/hermes-agent/pull/1650), [#2393](https://github.com/NousResearch/hermes-agent/pull/2393) by @0xbyt4)
-- **NeuTTS** — Local TTS provider backend with built-in setup flow, replacing the old optional skill ([#1657](https://github.com/NousResearch/hermes-agent/pull/1657), [#1664](https://github.com/NousResearch/hermes-agent/pull/1664))
-
-### Provider Improvements
-- **Eager fallback** to backup model on rate-limit errors ([#1730](https://github.com/NousResearch/hermes-agent/pull/1730))
-- **Endpoint metadata** for custom model context and pricing; query local servers for actual context window size ([#1906](https://github.com/NousResearch/hermes-agent/pull/1906), [#2091](https://github.com/NousResearch/hermes-agent/pull/2091) by @dusterbloom)
-- **Context length detection overhaul** — models.dev integration, provider-aware resolution, fuzzy matching for custom endpoints, `/v1/props` for llama.cpp ([#2158](https://github.com/NousResearch/hermes-agent/pull/2158), [#2051](https://github.com/NousResearch/hermes-agent/pull/2051), [#2403](https://github.com/NousResearch/hermes-agent/pull/2403))
-- **Model catalog updates** — gpt-5.4-mini, gpt-5.4-nano, healer-alpha, haiku-4.5, minimax-m2.7, claude 4.6 at 1M context ([#1913](https://github.com/NousResearch/hermes-agent/pull/1913), [#1915](https://github.com/NousResearch/hermes-agent/pull/1915), [#1900](https://github.com/NousResearch/hermes-agent/pull/1900), [#2155](https://github.com/NousResearch/hermes-agent/pull/2155), [#2474](https://github.com/NousResearch/hermes-agent/pull/2474))
-- **Custom endpoint improvements** — `model.base_url` in config.yaml, `api_mode` override for responses API, allow endpoints without API key, fail fast on missing keys ([#2330](https://github.com/NousResearch/hermes-agent/pull/2330), [#1651](https://github.com/NousResearch/hermes-agent/pull/1651), [#2556](https://github.com/NousResearch/hermes-agent/pull/2556), [#2445](https://github.com/NousResearch/hermes-agent/pull/2445), [#1994](https://github.com/NousResearch/hermes-agent/pull/1994), [#1998](https://github.com/NousResearch/hermes-agent/pull/1998))
-- Inject model and provider into system prompt ([#1929](https://github.com/NousResearch/hermes-agent/pull/1929))
-- Tie `api_mode` to provider config instead of env var ([#1656](https://github.com/NousResearch/hermes-agent/pull/1656))
-- Fix: prevent Anthropic token leaking to third-party `anthropic_messages` providers ([#2389](https://github.com/NousResearch/hermes-agent/pull/2389))
-- Fix: prevent Anthropic fallback from inheriting non-Anthropic `base_url` ([#2388](https://github.com/NousResearch/hermes-agent/pull/2388))
-- Fix: `auxiliary_is_nous` flag never resets — leaked Nous tags to other providers ([#1713](https://github.com/NousResearch/hermes-agent/pull/1713))
-- Fix: Anthropic `tool_choice 'none'` still allowed tool calls ([#1714](https://github.com/NousResearch/hermes-agent/pull/1714))
-- Fix: Mistral parser nested JSON fallback extraction ([#2335](https://github.com/NousResearch/hermes-agent/pull/2335))
-- Fix: MiniMax 401 auth resolved by defaulting to `anthropic_messages` ([#2103](https://github.com/NousResearch/hermes-agent/pull/2103))
-- Fix: case-insensitive model family matching ([#2350](https://github.com/NousResearch/hermes-agent/pull/2350))
-- Fix: ignore placeholder provider keys in activation checks ([#2358](https://github.com/NousResearch/hermes-agent/pull/2358))
-- Fix: Preserve Ollama model:tag colons in context length detection ([#2149](https://github.com/NousResearch/hermes-agent/pull/2149))
-- Fix: recognize Claude Code OAuth credentials in startup gate ([#1663](https://github.com/NousResearch/hermes-agent/pull/1663))
-- Fix: detect Claude Code version dynamically for OAuth user-agent ([#1670](https://github.com/NousResearch/hermes-agent/pull/1670))
-- Fix: OAuth flag stale after refresh/fallback ([#1890](https://github.com/NousResearch/hermes-agent/pull/1890))
-- Fix: auxiliary client skips expired Codex JWT ([#2397](https://github.com/NousResearch/hermes-agent/pull/2397))
-
-### Agent Loop
-- **Gateway prompt caching** — Cache AIAgent per session, keep assistant turns, fix session restore ([#2282](https://github.com/NousResearch/hermes-agent/pull/2282), [#2284](https://github.com/NousResearch/hermes-agent/pull/2284), [#2361](https://github.com/NousResearch/hermes-agent/pull/2361))
-- **Context compression overhaul** — Structured summaries, iterative updates, token-budget tail protection, configurable `summary_base_url` ([#2323](https://github.com/NousResearch/hermes-agent/pull/2323), [#1727](https://github.com/NousResearch/hermes-agent/pull/1727), [#2224](https://github.com/NousResearch/hermes-agent/pull/2224))
-- **Pre-call sanitization and post-call tool guardrails** ([#1732](https://github.com/NousResearch/hermes-agent/pull/1732))
-- **Auto-recover** from provider-rejected `tool_choice` by retrying without ([#2174](https://github.com/NousResearch/hermes-agent/pull/2174))
-- **Background memory/skill review** replaces inline nudges ([#2235](https://github.com/NousResearch/hermes-agent/pull/2235))
-- **SOUL.md as primary agent identity** instead of hardcoded default ([#1922](https://github.com/NousResearch/hermes-agent/pull/1922))
-- Fix: prevent silent tool result loss during context compression ([#1993](https://github.com/NousResearch/hermes-agent/pull/1993))
-- Fix: handle empty/null function arguments in tool call recovery ([#2163](https://github.com/NousResearch/hermes-agent/pull/2163))
-- Fix: handle API refusal responses gracefully instead of crashing ([#2156](https://github.com/NousResearch/hermes-agent/pull/2156))
-- Fix: prevent stuck agent loop on malformed tool calls ([#2114](https://github.com/NousResearch/hermes-agent/pull/2114))
-- Fix: return JSON parse error to model instead of dispatching with empty args ([#2342](https://github.com/NousResearch/hermes-agent/pull/2342))
-- Fix: consecutive assistant message merge drops content on mixed types ([#1703](https://github.com/NousResearch/hermes-agent/pull/1703))
-- Fix: message role alternation violations in JSON recovery and error handler ([#1722](https://github.com/NousResearch/hermes-agent/pull/1722))
-- Fix: `compression_attempts` resets each iteration — allowed unlimited compressions ([#1723](https://github.com/NousResearch/hermes-agent/pull/1723))
-- Fix: `length_continue_retries` never resets — later truncations got fewer retries ([#1717](https://github.com/NousResearch/hermes-agent/pull/1717))
-- Fix: compressor summary role violated consecutive-role constraint ([#1720](https://github.com/NousResearch/hermes-agent/pull/1720), [#1743](https://github.com/NousResearch/hermes-agent/pull/1743))
-- Fix: remove hardcoded `gemini-3-flash-preview` as default summary model ([#2464](https://github.com/NousResearch/hermes-agent/pull/2464))
-- Fix: correctly handle empty tool results ([#2201](https://github.com/NousResearch/hermes-agent/pull/2201))
-- Fix: crash on None entry in `tool_calls` list ([#2209](https://github.com/NousResearch/hermes-agent/pull/2209) by @0xbyt4, [#2316](https://github.com/NousResearch/hermes-agent/pull/2316))
-- Fix: per-thread persistent event loops in worker threads ([#2214](https://github.com/NousResearch/hermes-agent/pull/2214) by @jquesnelle)
-- Fix: prevent 'event loop already running' when async tools run in parallel ([#2207](https://github.com/NousResearch/hermes-agent/pull/2207))
-- Fix: strip ANSI at the source — clean terminal output before it reaches the model ([#2115](https://github.com/NousResearch/hermes-agent/pull/2115))
-- Fix: skip top-level `cache_control` on role:tool for OpenRouter ([#2391](https://github.com/NousResearch/hermes-agent/pull/2391))
-- Fix: delegate tool — save parent tool names before child construction mutates global ([#2083](https://github.com/NousResearch/hermes-agent/pull/2083) by @ygd58, [#1894](https://github.com/NousResearch/hermes-agent/pull/1894))
-- Fix: only strip last assistant message if empty string ([#2326](https://github.com/NousResearch/hermes-agent/pull/2326))
-
-### Session & Memory
-- **Session search** and management slash commands ([#2198](https://github.com/NousResearch/hermes-agent/pull/2198))
-- **Auto session titles** and `.hermes.md` project config ([#1712](https://github.com/NousResearch/hermes-agent/pull/1712))
-- Fix: concurrent memory writes silently drop entries — added file locking ([#1726](https://github.com/NousResearch/hermes-agent/pull/1726))
-- Fix: search all sources by default in `session_search` ([#1892](https://github.com/NousResearch/hermes-agent/pull/1892))
-- Fix: handle hyphenated FTS5 queries and preserve quoted literals ([#1776](https://github.com/NousResearch/hermes-agent/pull/1776))
-- Fix: skip corrupt lines in `load_transcript` instead of crashing ([#1744](https://github.com/NousResearch/hermes-agent/pull/1744))
-- Fix: normalize session keys to prevent case-sensitive duplicates ([#2157](https://github.com/NousResearch/hermes-agent/pull/2157))
-- Fix: prevent `session_search` crash when no sessions exist ([#2194](https://github.com/NousResearch/hermes-agent/pull/2194))
-- Fix: reset token counters on new session for accurate usage display ([#2101](https://github.com/NousResearch/hermes-agent/pull/2101) by @InB4DevOps)
-- Fix: prevent stale memory overwrites by flush agent ([#2687](https://github.com/NousResearch/hermes-agent/pull/2687))
-- Fix: remove synthetic error message injection, fix session resume after repeated failures ([#2303](https://github.com/NousResearch/hermes-agent/pull/2303))
-- Fix: quiet mode with `--resume` now passes conversation_history ([#2357](https://github.com/NousResearch/hermes-agent/pull/2357))
-- Fix: unify resume logic in batch mode ([#2331](https://github.com/NousResearch/hermes-agent/pull/2331))
-
-### Honcho Memory
-- Honcho config fixes and @ context reference integration ([#2343](https://github.com/NousResearch/hermes-agent/pull/2343))
-- Self-hosted / Docker configuration documentation ([#2475](https://github.com/NousResearch/hermes-agent/pull/2475))
-
----
-
-## 📱 Messaging Platforms (Gateway)
-
-### New Platform Adapters
-- **Signal Messenger** — Full adapter with attachment handling, group message filtering, and Note to Self echo-back protection ([#2206](https://github.com/NousResearch/hermes-agent/pull/2206), [#2400](https://github.com/NousResearch/hermes-agent/pull/2400), [#2297](https://github.com/NousResearch/hermes-agent/pull/2297), [#2156](https://github.com/NousResearch/hermes-agent/pull/2156))
-- **DingTalk** — Adapter with gateway wiring and setup docs ([#1685](https://github.com/NousResearch/hermes-agent/pull/1685), [#1690](https://github.com/NousResearch/hermes-agent/pull/1690), [#1692](https://github.com/NousResearch/hermes-agent/pull/1692))
-- **SMS (Twilio)** ([#1688](https://github.com/NousResearch/hermes-agent/pull/1688))
-- **Mattermost** — With @-mention-only channel filter ([#1683](https://github.com/NousResearch/hermes-agent/pull/1683), [#2443](https://github.com/NousResearch/hermes-agent/pull/2443))
-- **Matrix** — With vision support and image caching ([#1683](https://github.com/NousResearch/hermes-agent/pull/1683), [#2520](https://github.com/NousResearch/hermes-agent/pull/2520))
-- **Webhook** — Platform adapter for external event triggers ([#2166](https://github.com/NousResearch/hermes-agent/pull/2166))
-- **OpenAI-compatible API server** — `/v1/chat/completions` endpoint with `/api/jobs` cron management ([#1756](https://github.com/NousResearch/hermes-agent/pull/1756), [#2450](https://github.com/NousResearch/hermes-agent/pull/2450), [#2456](https://github.com/NousResearch/hermes-agent/pull/2456))
-
-### Telegram Improvements
-- MarkdownV2 support — strikethrough, spoiler, blockquotes, escape parentheses/braces/backslashes/backticks ([#2199](https://github.com/NousResearch/hermes-agent/pull/2199), [#2200](https://github.com/NousResearch/hermes-agent/pull/2200) by @llbn, [#2386](https://github.com/NousResearch/hermes-agent/pull/2386))
-- Auto-detect HTML tags and use `parse_mode=HTML` ([#1709](https://github.com/NousResearch/hermes-agent/pull/1709))
-- Telegram group vision support + thread-based sessions ([#2153](https://github.com/NousResearch/hermes-agent/pull/2153))
-- Auto-reconnect polling after network interruption ([#2517](https://github.com/NousResearch/hermes-agent/pull/2517))
-- Aggregate split text messages before dispatching ([#1674](https://github.com/NousResearch/hermes-agent/pull/1674))
-- Fix: streaming config bridge, not-modified, flood control ([#1782](https://github.com/NousResearch/hermes-agent/pull/1782), [#1783](https://github.com/NousResearch/hermes-agent/pull/1783))
-- Fix: edited_message event crashes ([#2074](https://github.com/NousResearch/hermes-agent/pull/2074))
-- Fix: retry 409 polling conflicts before giving up ([#2312](https://github.com/NousResearch/hermes-agent/pull/2312))
-- Fix: topic delivery via `platform:chat_id:thread_id` format ([#2455](https://github.com/NousResearch/hermes-agent/pull/2455))
-
-### Discord Improvements
-- Document caching and text-file injection ([#2503](https://github.com/NousResearch/hermes-agent/pull/2503))
-- Persistent typing indicator for DMs ([#2468](https://github.com/NousResearch/hermes-agent/pull/2468))
-- Discord DM vision — inline images + attachment analysis ([#2186](https://github.com/NousResearch/hermes-agent/pull/2186))
-- Persist thread participation across gateway restarts ([#1661](https://github.com/NousResearch/hermes-agent/pull/1661))
-- Fix: gateway crash on non-ASCII guild names ([#2302](https://github.com/NousResearch/hermes-agent/pull/2302))
-- Fix: thread permission errors ([#2073](https://github.com/NousResearch/hermes-agent/pull/2073))
-- Fix: slash event routing in threads ([#2460](https://github.com/NousResearch/hermes-agent/pull/2460))
-- Fix: remove bugged followup messages + `/ask` command ([#1836](https://github.com/NousResearch/hermes-agent/pull/1836))
-- Fix: graceful WebSocket reconnection ([#2127](https://github.com/NousResearch/hermes-agent/pull/2127))
-- Fix: voice channel TTS when streaming enabled ([#2322](https://github.com/NousResearch/hermes-agent/pull/2322))
-
-### WhatsApp & Other Adapters
-- WhatsApp: outbound `send_message` routing ([#1769](https://github.com/NousResearch/hermes-agent/pull/1769) by @sai-samarth), LID format self-chat ([#1667](https://github.com/NousResearch/hermes-agent/pull/1667)), `reply_prefix` config fix ([#1923](https://github.com/NousResearch/hermes-agent/pull/1923)), restart on bridge child exit ([#2334](https://github.com/NousResearch/hermes-agent/pull/2334)), image/bridge improvements ([#2181](https://github.com/NousResearch/hermes-agent/pull/2181))
-- Matrix: correct `reply_to_message_id` parameter ([#1895](https://github.com/NousResearch/hermes-agent/pull/1895)), bare media types fix ([#1736](https://github.com/NousResearch/hermes-agent/pull/1736))
-- Mattermost: MIME types for media attachments ([#2329](https://github.com/NousResearch/hermes-agent/pull/2329))
-
-### Gateway Core
-- **Auto-reconnect** failed platforms with exponential backoff ([#2584](https://github.com/NousResearch/hermes-agent/pull/2584))
-- **Notify users when session auto-resets** ([#2519](https://github.com/NousResearch/hermes-agent/pull/2519))
-- **Reply-to message context** for out-of-session replies ([#1662](https://github.com/NousResearch/hermes-agent/pull/1662))
-- **Ignore unauthorized DMs** config option ([#1919](https://github.com/NousResearch/hermes-agent/pull/1919))
-- Fix: `/reset` in thread-mode resets global session instead of thread ([#2254](https://github.com/NousResearch/hermes-agent/pull/2254))
-- Fix: deliver MEDIA: files after streaming responses ([#2382](https://github.com/NousResearch/hermes-agent/pull/2382))
-- Fix: cap interrupt recursion depth to prevent resource exhaustion ([#1659](https://github.com/NousResearch/hermes-agent/pull/1659))
-- Fix: detect stopped processes and release stale locks on `--replace` ([#2406](https://github.com/NousResearch/hermes-agent/pull/2406), [#1908](https://github.com/NousResearch/hermes-agent/pull/1908))
-- Fix: PID-based wait with force-kill for gateway restart ([#1902](https://github.com/NousResearch/hermes-agent/pull/1902))
-- Fix: prevent `--replace` mode from killing the caller process ([#2185](https://github.com/NousResearch/hermes-agent/pull/2185))
-- Fix: `/model` shows active fallback model instead of config default ([#1660](https://github.com/NousResearch/hermes-agent/pull/1660))
-- Fix: `/title` command fails when session doesn't exist in SQLite yet ([#2379](https://github.com/NousResearch/hermes-agent/pull/2379) by @ten-jampa)
-- Fix: process `/queue`'d messages after agent completion ([#2469](https://github.com/NousResearch/hermes-agent/pull/2469))
-- Fix: strip orphaned `tool_results` + let `/reset` bypass running agent ([#2180](https://github.com/NousResearch/hermes-agent/pull/2180))
-- Fix: prevent agents from starting gateway outside systemd management ([#2617](https://github.com/NousResearch/hermes-agent/pull/2617))
-- Fix: prevent systemd restart storm on gateway connection failure ([#2327](https://github.com/NousResearch/hermes-agent/pull/2327))
-- Fix: include resolved node path in systemd unit ([#1767](https://github.com/NousResearch/hermes-agent/pull/1767) by @sai-samarth)
-- Fix: send error details to user in gateway outer exception handler ([#1966](https://github.com/NousResearch/hermes-agent/pull/1966))
-- Fix: improve error handling for 429 usage limits and 500 context overflow ([#1839](https://github.com/NousResearch/hermes-agent/pull/1839))
-- Fix: add all missing platform allowlist env vars to startup warning check ([#2628](https://github.com/NousResearch/hermes-agent/pull/2628))
-- Fix: media delivery fails for file paths containing spaces ([#2621](https://github.com/NousResearch/hermes-agent/pull/2621))
-- Fix: duplicate session-key collision in multi-platform gateway ([#2171](https://github.com/NousResearch/hermes-agent/pull/2171))
-- Fix: Matrix and Mattermost never report as connected ([#1711](https://github.com/NousResearch/hermes-agent/pull/1711))
-- Fix: PII redaction config never read — missing yaml import ([#1701](https://github.com/NousResearch/hermes-agent/pull/1701))
-- Fix: NameError on skill slash commands ([#1697](https://github.com/NousResearch/hermes-agent/pull/1697))
-- Fix: persist watcher metadata in checkpoint for crash recovery ([#1706](https://github.com/NousResearch/hermes-agent/pull/1706))
-- Fix: pass `message_thread_id` in send_image_file, send_document, send_video ([#2339](https://github.com/NousResearch/hermes-agent/pull/2339))
-- Fix: media-group aggregation on rapid successive photo messages ([#2160](https://github.com/NousResearch/hermes-agent/pull/2160))
-
----
-
-## 🔧 Tool System
-
-### MCP Enhancements
-- **MCP server management CLI** + OAuth 2.1 PKCE auth ([#2465](https://github.com/NousResearch/hermes-agent/pull/2465))
-- **Expose MCP servers as standalone toolsets** ([#1907](https://github.com/NousResearch/hermes-agent/pull/1907))
-- **Interactive MCP tool configuration** in `hermes tools` ([#1694](https://github.com/NousResearch/hermes-agent/pull/1694))
-- Fix: MCP-OAuth port mismatch, path traversal, and shared handler state ([#2552](https://github.com/NousResearch/hermes-agent/pull/2552))
-- Fix: preserve MCP tool registrations across session resets ([#2124](https://github.com/NousResearch/hermes-agent/pull/2124))
-- Fix: concurrent file access crash + duplicate MCP registration ([#2154](https://github.com/NousResearch/hermes-agent/pull/2154))
-- Fix: normalise MCP schemas + expand session list columns ([#2102](https://github.com/NousResearch/hermes-agent/pull/2102))
-- Fix: `tool_choice` `mcp_` prefix handling ([#1775](https://github.com/NousResearch/hermes-agent/pull/1775))
-
-### Web Tool Backends
-- **Tavily** as web search/extract/crawl backend ([#1731](https://github.com/NousResearch/hermes-agent/pull/1731))
-- **Parallel** as alternative web search/extract backend ([#1696](https://github.com/NousResearch/hermes-agent/pull/1696))
-- **Configurable web backend** — Firecrawl/BeautifulSoup/Playwright selection ([#2256](https://github.com/NousResearch/hermes-agent/pull/2256))
-- Fix: whitespace-only env vars bypass web backend detection ([#2341](https://github.com/NousResearch/hermes-agent/pull/2341))
-
-### New Tools
-- **IMAP email** reading and sending ([#2173](https://github.com/NousResearch/hermes-agent/pull/2173))
-- **STT (speech-to-text)** tool using Whisper API ([#2072](https://github.com/NousResearch/hermes-agent/pull/2072))
-- **Route-aware pricing estimates** ([#1695](https://github.com/NousResearch/hermes-agent/pull/1695))
-
-### Tool Improvements
-- TTS: `base_url` support for OpenAI TTS provider ([#2064](https://github.com/NousResearch/hermes-agent/pull/2064) by @hanai)
-- Vision: configurable timeout, tilde expansion in file paths, DM vision with multi-image and base64 fallback ([#2480](https://github.com/NousResearch/hermes-agent/pull/2480), [#2585](https://github.com/NousResearch/hermes-agent/pull/2585), [#2211](https://github.com/NousResearch/hermes-agent/pull/2211))
-- Browser: race condition fix in session creation ([#1721](https://github.com/NousResearch/hermes-agent/pull/1721)), TypeError on unexpected LLM params ([#1735](https://github.com/NousResearch/hermes-agent/pull/1735))
-- File tools: strip ANSI escape codes from write_file and patch content ([#2532](https://github.com/NousResearch/hermes-agent/pull/2532)), include pagination args in repeated search key ([#1824](https://github.com/NousResearch/hermes-agent/pull/1824) by @cutepawss), improve fuzzy matching accuracy + position calculation refactor ([#2096](https://github.com/NousResearch/hermes-agent/pull/2096), [#1681](https://github.com/NousResearch/hermes-agent/pull/1681))
-- Code execution: resource leak and double socket close fix ([#2381](https://github.com/NousResearch/hermes-agent/pull/2381))
-- Delegate: thread safety for concurrent subagent delegation ([#1672](https://github.com/NousResearch/hermes-agent/pull/1672)), preserve parent agent's tool list after delegation ([#1778](https://github.com/NousResearch/hermes-agent/pull/1778))
-- Fix: make concurrent tool batching path-aware for file mutations ([#1914](https://github.com/NousResearch/hermes-agent/pull/1914))
-- Fix: chunk long messages in `send_message_tool` before platform dispatch ([#1646](https://github.com/NousResearch/hermes-agent/pull/1646))
-- Fix: add missing 'messaging' toolset ([#1718](https://github.com/NousResearch/hermes-agent/pull/1718))
-- Fix: prevent unavailable tool names from leaking into model schemas ([#2072](https://github.com/NousResearch/hermes-agent/pull/2072))
-- Fix: pass visited set by reference to prevent diamond dependency duplication ([#2311](https://github.com/NousResearch/hermes-agent/pull/2311))
-- Fix: Daytona sandbox lookup migrated from `find_one` to `get/list` ([#2063](https://github.com/NousResearch/hermes-agent/pull/2063) by @rovle)
-
----
-
-## 🧩 Skills Ecosystem
-
-### Skills System Improvements
-- **Agent-created skills** — Caution-level findings allowed, dangerous skills ask instead of block ([#1840](https://github.com/NousResearch/hermes-agent/pull/1840), [#2446](https://github.com/NousResearch/hermes-agent/pull/2446))
-- **`--yes` flag** to bypass confirmation in `/skills install` and uninstall ([#1647](https://github.com/NousResearch/hermes-agent/pull/1647))
-- **Disabled skills respected** across banner, system prompt, and slash commands ([#1897](https://github.com/NousResearch/hermes-agent/pull/1897))
-- Fix: skills custom_tools import crash + sandbox file_tools integration ([#2239](https://github.com/NousResearch/hermes-agent/pull/2239))
-- Fix: agent-created skills with pip requirements crash on install ([#2145](https://github.com/NousResearch/hermes-agent/pull/2145))
-- Fix: race condition in `Skills.__init__` when `hub.yaml` missing ([#2242](https://github.com/NousResearch/hermes-agent/pull/2242))
-- Fix: validate skill metadata before install and block duplicates ([#2241](https://github.com/NousResearch/hermes-agent/pull/2241))
-- Fix: skills hub inspect/resolve — 4 bugs in inspect, redirects, discovery, tap list ([#2447](https://github.com/NousResearch/hermes-agent/pull/2447))
-- Fix: agent-created skills keep working after session reset ([#2121](https://github.com/NousResearch/hermes-agent/pull/2121))
-
-### New Skills
-- **OCR-and-documents** — PDF/DOCX/XLS/PPTX/image OCR with optional GPU ([#2236](https://github.com/NousResearch/hermes-agent/pull/2236), [#2461](https://github.com/NousResearch/hermes-agent/pull/2461))
-- **Huggingface-hub** bundled skill ([#1921](https://github.com/NousResearch/hermes-agent/pull/1921))
-- **Sherlock OSINT** username search ([#1671](https://github.com/NousResearch/hermes-agent/pull/1671))
-- **Meme-generation** — Image generator with Pillow ([#2344](https://github.com/NousResearch/hermes-agent/pull/2344))
-- **Bioinformatics** gateway skill — index to 400+ bio skills ([#2387](https://github.com/NousResearch/hermes-agent/pull/2387))
-- **Inference.sh** skill (terminal-based) ([#1686](https://github.com/NousResearch/hermes-agent/pull/1686))
-- **Base blockchain** optional skill ([#1643](https://github.com/NousResearch/hermes-agent/pull/1643))
-- **3D-model-viewer** optional skill ([#2226](https://github.com/NousResearch/hermes-agent/pull/2226))
-- **FastMCP** optional skill ([#2113](https://github.com/NousResearch/hermes-agent/pull/2113))
-- **Hermes-agent-setup** skill ([#1905](https://github.com/NousResearch/hermes-agent/pull/1905))
-
----
-
-## 🔌 Plugin System Enhancements
-
-- **TUI extension hooks** — Build custom CLIs on top of Hermes ([#2333](https://github.com/NousResearch/hermes-agent/pull/2333))
-- **`hermes plugins install/remove/list`** commands ([#2337](https://github.com/NousResearch/hermes-agent/pull/2337))
-- **Slash command registration** for plugins ([#2359](https://github.com/NousResearch/hermes-agent/pull/2359))
-- **`session:end` lifecycle event** hook ([#1725](https://github.com/NousResearch/hermes-agent/pull/1725))
-- Fix: require opt-in for project plugin discovery ([#2215](https://github.com/NousResearch/hermes-agent/pull/2215))
-
----
-
-## 🔒 Security & Reliability
-
-### Security
-- **SSRF protection** for vision_tools and web_tools ([#2679](https://github.com/NousResearch/hermes-agent/pull/2679))
-- **Shell injection prevention** in `_expand_path` via `~user` path suffix ([#2685](https://github.com/NousResearch/hermes-agent/pull/2685))
-- **Block untrusted browser-origin** API server access ([#2451](https://github.com/NousResearch/hermes-agent/pull/2451))
-- **Block sandbox backend creds** from subprocess env ([#1658](https://github.com/NousResearch/hermes-agent/pull/1658))
-- **Block @ references** from reading secrets outside workspace ([#2601](https://github.com/NousResearch/hermes-agent/pull/2601) by @Gutslabs)
-- **Malicious code pattern pre-exec scanner** for terminal_tool ([#2245](https://github.com/NousResearch/hermes-agent/pull/2245))
-- **Harden terminal safety** and sandbox file writes ([#1653](https://github.com/NousResearch/hermes-agent/pull/1653))
-- **PKCE verifier leak** fix + OAuth refresh Content-Type ([#1775](https://github.com/NousResearch/hermes-agent/pull/1775))
-- **Eliminate SQL string formatting** in `execute()` calls ([#2061](https://github.com/NousResearch/hermes-agent/pull/2061) by @dusterbloom)
-- **Harden jobs API** — input limits, field whitelist, startup check ([#2456](https://github.com/NousResearch/hermes-agent/pull/2456))
-
-### Reliability
-- Thread locks on 4 SessionDB methods ([#1704](https://github.com/NousResearch/hermes-agent/pull/1704))
-- File locking for concurrent memory writes ([#1726](https://github.com/NousResearch/hermes-agent/pull/1726))
-- Handle OpenRouter errors gracefully ([#2112](https://github.com/NousResearch/hermes-agent/pull/2112))
-- Guard print() calls against OSError ([#1668](https://github.com/NousResearch/hermes-agent/pull/1668))
-- Safely handle non-string inputs in redacting formatter ([#2392](https://github.com/NousResearch/hermes-agent/pull/2392), [#1700](https://github.com/NousResearch/hermes-agent/pull/1700))
-- ACP: preserve session provider on model switch, persist sessions to disk ([#2380](https://github.com/NousResearch/hermes-agent/pull/2380), [#2071](https://github.com/NousResearch/hermes-agent/pull/2071))
-- API server: persist ResponseStore to SQLite across restarts ([#2472](https://github.com/NousResearch/hermes-agent/pull/2472))
-- Fix: `fetch_nous_models` always TypeError from positional args ([#1699](https://github.com/NousResearch/hermes-agent/pull/1699))
-- Fix: resolve merge conflict markers in cli.py breaking startup ([#2347](https://github.com/NousResearch/hermes-agent/pull/2347))
-- Fix: `minisweagent_path.py` missing from wheel ([#2098](https://github.com/NousResearch/hermes-agent/pull/2098) by @JiwaniZakir)
-
-### Cron System
-- **`[SILENT]` response** — cron agents can suppress delivery ([#1833](https://github.com/NousResearch/hermes-agent/pull/1833))
-- **Scale missed-job grace window** with schedule frequency ([#2449](https://github.com/NousResearch/hermes-agent/pull/2449))
-- **Recover recent one-shot jobs** ([#1918](https://github.com/NousResearch/hermes-agent/pull/1918))
-- Fix: normalize `repeat<=0` to None — jobs deleted after first run when LLM passes -1 ([#2612](https://github.com/NousResearch/hermes-agent/pull/2612) by @Mibayy)
-- Fix: Matrix added to scheduler delivery platform_map ([#2167](https://github.com/NousResearch/hermes-agent/pull/2167) by @buntingszn)
-- Fix: naive ISO timestamps without timezone — jobs fire at wrong time ([#1729](https://github.com/NousResearch/hermes-agent/pull/1729))
-- Fix: `get_due_jobs` reads `jobs.json` twice — race condition ([#1716](https://github.com/NousResearch/hermes-agent/pull/1716))
-- Fix: silent jobs return empty response for delivery skip ([#2442](https://github.com/NousResearch/hermes-agent/pull/2442))
-- Fix: stop injecting cron outputs into gateway session history ([#2313](https://github.com/NousResearch/hermes-agent/pull/2313))
-- Fix: close abandoned coroutine when `asyncio.run()` raises RuntimeError ([#2317](https://github.com/NousResearch/hermes-agent/pull/2317))
-
----
-
-## 🧪 Testing
-
-- Resolve all consistently failing tests ([#2488](https://github.com/NousResearch/hermes-agent/pull/2488))
-- Replace `FakePath` with `monkeypatch` for Python 3.12 compat ([#2444](https://github.com/NousResearch/hermes-agent/pull/2444))
-- Align Hermes setup and full-suite expectations ([#1710](https://github.com/NousResearch/hermes-agent/pull/1710))
-
----
-
-## 📚 Documentation
-
-- Comprehensive docs update for recent features ([#1693](https://github.com/NousResearch/hermes-agent/pull/1693), [#2183](https://github.com/NousResearch/hermes-agent/pull/2183))
-- Alibaba Cloud and DingTalk setup guides ([#1687](https://github.com/NousResearch/hermes-agent/pull/1687), [#1692](https://github.com/NousResearch/hermes-agent/pull/1692))
-- Detailed skills documentation ([#2244](https://github.com/NousResearch/hermes-agent/pull/2244))
-- Honcho self-hosted / Docker configuration ([#2475](https://github.com/NousResearch/hermes-agent/pull/2475))
-- Context length detection FAQ and quickstart references ([#2179](https://github.com/NousResearch/hermes-agent/pull/2179))
-- Fix docs inconsistencies across reference and user guides ([#1995](https://github.com/NousResearch/hermes-agent/pull/1995))
-- Fix MCP install commands — use uv, not bare pip ([#1909](https://github.com/NousResearch/hermes-agent/pull/1909))
-- Replace ASCII diagrams with Mermaid/lists ([#2402](https://github.com/NousResearch/hermes-agent/pull/2402))
-- Gemini OAuth provider implementation plan ([#2467](https://github.com/NousResearch/hermes-agent/pull/2467))
-- Discord Server Members Intent marked as required ([#2330](https://github.com/NousResearch/hermes-agent/pull/2330))
-- Fix MDX build error in api-server.md ([#1787](https://github.com/NousResearch/hermes-agent/pull/1787))
-- Align venv path to match installer ([#2114](https://github.com/NousResearch/hermes-agent/pull/2114))
-- New skills added to hub index ([#2281](https://github.com/NousResearch/hermes-agent/pull/2281))
-
----
-
-## 👥 Contributors
-
-### Core
-- **@teknium1** (Teknium) — 280 PRs
-
-### Community Contributors
-- **@mchzimm** (to_the_max) — GitHub Copilot provider integration ([#1879](https://github.com/NousResearch/hermes-agent/pull/1879))
-- **@jquesnelle** (Jeffrey Quesnelle) — Per-thread persistent event loops fix ([#2214](https://github.com/NousResearch/hermes-agent/pull/2214))
-- **@llbn** (lbn) — Telegram MarkdownV2 strikethrough, spoiler, blockquotes, and escape fixes ([#2199](https://github.com/NousResearch/hermes-agent/pull/2199), [#2200](https://github.com/NousResearch/hermes-agent/pull/2200))
-- **@dusterbloom** — SQL injection prevention + local server context window querying ([#2061](https://github.com/NousResearch/hermes-agent/pull/2061), [#2091](https://github.com/NousResearch/hermes-agent/pull/2091))
-- **@0xbyt4** — Anthropic tool_calls None guard + OpenCode-Go provider config fix ([#2209](https://github.com/NousResearch/hermes-agent/pull/2209), [#2393](https://github.com/NousResearch/hermes-agent/pull/2393))
-- **@sai-samarth** (Saisamarth) — WhatsApp send_message routing + systemd node path ([#1769](https://github.com/NousResearch/hermes-agent/pull/1769), [#1767](https://github.com/NousResearch/hermes-agent/pull/1767))
-- **@Gutslabs** (Guts) — Block @ references from reading secrets ([#2601](https://github.com/NousResearch/hermes-agent/pull/2601))
-- **@Mibayy** (Mibay) — Cron job repeat normalization ([#2612](https://github.com/NousResearch/hermes-agent/pull/2612))
-- **@ten-jampa** (Tenzin Jampa) — Gateway /title command fix ([#2379](https://github.com/NousResearch/hermes-agent/pull/2379))
-- **@cutepawss** (lila) — File tools search pagination fix ([#1824](https://github.com/NousResearch/hermes-agent/pull/1824))
-- **@hanai** (Hanai) — OpenAI TTS base_url support ([#2064](https://github.com/NousResearch/hermes-agent/pull/2064))
-- **@rovle** (Lovre Pešut) — Daytona sandbox API migration ([#2063](https://github.com/NousResearch/hermes-agent/pull/2063))
-- **@buntingszn** (bunting szn) — Matrix cron delivery support ([#2167](https://github.com/NousResearch/hermes-agent/pull/2167))
-- **@InB4DevOps** — Token counter reset on new session ([#2101](https://github.com/NousResearch/hermes-agent/pull/2101))
-- **@JiwaniZakir** (Zakir Jiwani) — Missing file in wheel fix ([#2098](https://github.com/NousResearch/hermes-agent/pull/2098))
-- **@ygd58** (buray) — Delegate tool parent tool names fix ([#2083](https://github.com/NousResearch/hermes-agent/pull/2083))
-
----
-
-**Full Changelog**: [v2026.3.17...v2026.3.23](https://github.com/NousResearch/hermes-agent/compare/v2026.3.17...v2026.3.23)

RELEASE_v0.5.0.md

@@ -1,348 +0,0 @@
-# Hermes Agent v0.5.0 (v2026.3.28)
-
-**Release Date:** March 28, 2026
-
-> The hardening release — Hugging Face provider, /model command overhaul, Telegram Private Chat Topics, native Modal SDK, plugin lifecycle hooks, tool-use enforcement for GPT models, Nix flake, 50+ security and reliability fixes, and a comprehensive supply chain audit.
-
----
-
-## ✨ Highlights
-
-- **Nous Portal now supports 400+ models** — The Nous Research inference portal has expanded dramatically, giving Hermes Agent users access to over 400 models through a single provider endpoint
-
-- **Hugging Face as a first-class inference provider** — Full integration with HF Inference API including curated agentic model picker that maps to OpenRouter analogues, live `/models` endpoint probe, and setup wizard flow ([#3419](https://github.com/NousResearch/hermes-agent/pull/3419), [#3440](https://github.com/NousResearch/hermes-agent/pull/3440))
-
-- **Telegram Private Chat Topics** — Project-based conversations with functional skill binding per topic, enabling isolated workflows within a single Telegram chat ([#3163](https://github.com/NousResearch/hermes-agent/pull/3163))
-
-- **Native Modal SDK backend** — Replaced swe-rex dependency with native Modal SDK (`Sandbox.create.aio` + `exec.aio`), eliminating tunnels and simplifying the Modal terminal backend ([#3538](https://github.com/NousResearch/hermes-agent/pull/3538))
-
-- **Plugin lifecycle hooks activated** — `pre_llm_call`, `post_llm_call`, `on_session_start`, and `on_session_end` hooks now fire in the agent loop and CLI/gateway, completing the plugin hook system ([#3542](https://github.com/NousResearch/hermes-agent/pull/3542))
-
-- **Improved OpenAI Model Reliability** — Added `GPT_TOOL_USE_GUIDANCE` to prevent GPT models from describing intended actions instead of making tool calls, plus automatic stripping of stale budget warnings from conversation history that caused models to avoid tools across turns ([#3528](https://github.com/NousResearch/hermes-agent/pull/3528))
-
-- **Nix flake** — Full uv2nix build, NixOS module with persistent container mode, auto-generated config keys from Python source, and suffix PATHs for agent-friendliness ([#20](https://github.com/NousResearch/hermes-agent/pull/20), [#3274](https://github.com/NousResearch/hermes-agent/pull/3274), [#3061](https://github.com/NousResearch/hermes-agent/pull/3061)) by @alt-glitch
-
-- **Supply chain hardening** — Removed compromised `litellm` dependency, pinned all dependency version ranges, regenerated `uv.lock` with hashes, added CI workflow scanning PRs for supply chain attack patterns, and bumped deps to fix CVEs ([#2796](https://github.com/NousResearch/hermes-agent/pull/2796), [#2810](https://github.com/NousResearch/hermes-agent/pull/2810), [#2812](https://github.com/NousResearch/hermes-agent/pull/2812), [#2816](https://github.com/NousResearch/hermes-agent/pull/2816), [#3073](https://github.com/NousResearch/hermes-agent/pull/3073))
-
-- **Anthropic output limits fix** — Replaced hardcoded 16K `max_tokens` with per-model native output limits (128K for Opus 4.6, 64K for Sonnet 4.6), fixing "Response truncated" and thinking-budget exhaustion on direct Anthropic API ([#3426](https://github.com/NousResearch/hermes-agent/pull/3426), [#3444](https://github.com/NousResearch/hermes-agent/pull/3444))
-
----
-
-## 🏗️ Core Agent & Architecture
-
-### New Provider: Hugging Face
-- First-class Hugging Face Inference API integration with auth, setup wizard, and model picker ([#3419](https://github.com/NousResearch/hermes-agent/pull/3419))
-- Curated model list mapping OpenRouter agentic defaults to HF equivalents — providers with 8+ curated models skip live `/models` probe for speed ([#3440](https://github.com/NousResearch/hermes-agent/pull/3440))
-- Added glm-5-turbo to Z.AI provider model list ([#3095](https://github.com/NousResearch/hermes-agent/pull/3095))
-
-### Provider & Model Improvements
-- `/model` command overhaul — extracted shared `switch_model()` pipeline for CLI and gateway, custom endpoint support, provider-aware routing ([#2795](https://github.com/NousResearch/hermes-agent/pull/2795), [#2799](https://github.com/NousResearch/hermes-agent/pull/2799))
-- Removed `/model` slash command from CLI and gateway in favor of `hermes model` subcommand ([#3080](https://github.com/NousResearch/hermes-agent/pull/3080))
-- Preserve `custom` provider instead of silently remapping to `openrouter` ([#2792](https://github.com/NousResearch/hermes-agent/pull/2792))
-- Read root-level `provider` and `base_url` from config.yaml into model config ([#3112](https://github.com/NousResearch/hermes-agent/pull/3112))
-- Align Nous Portal model slugs with OpenRouter naming ([#3253](https://github.com/NousResearch/hermes-agent/pull/3253))
-- Fix Alibaba provider default endpoint and model list ([#3484](https://github.com/NousResearch/hermes-agent/pull/3484))
-- Allow MiniMax users to override `/v1` → `/anthropic` auto-correction ([#3553](https://github.com/NousResearch/hermes-agent/pull/3553))
-- Migrate OAuth token refresh to `platform.claude.com` with fallback ([#3246](https://github.com/NousResearch/hermes-agent/pull/3246))
-
-### Agent Loop & Conversation
-- **Improved OpenAI model reliability** — `GPT_TOOL_USE_GUIDANCE` prevents GPT models from describing actions instead of calling tools + automatic budget warning stripping from history ([#3528](https://github.com/NousResearch/hermes-agent/pull/3528))
-- **Surface lifecycle events** — All retry, fallback, and compression events now surface to the user as formatted messages ([#3153](https://github.com/NousResearch/hermes-agent/pull/3153))
-- **Anthropic output limits** — Per-model native output limits instead of hardcoded 16K `max_tokens` ([#3426](https://github.com/NousResearch/hermes-agent/pull/3426))
-- **Thinking-budget exhaustion detection** — Skip useless continuation retries when model uses all output tokens on reasoning ([#3444](https://github.com/NousResearch/hermes-agent/pull/3444))
-- Always prefer streaming for API calls to prevent hung subagents ([#3120](https://github.com/NousResearch/hermes-agent/pull/3120))
-- Restore safe non-streaming fallback after stream failures ([#3020](https://github.com/NousResearch/hermes-agent/pull/3020))
-- Give subagents independent iteration budgets ([#3004](https://github.com/NousResearch/hermes-agent/pull/3004))
-- Update `api_key` in `_try_activate_fallback` for subagent auth ([#3103](https://github.com/NousResearch/hermes-agent/pull/3103))
-- Graceful return on max retries instead of crashing thread ([untagged commit](https://github.com/NousResearch/hermes-agent))
-- Count compression restarts toward retry limit ([#3070](https://github.com/NousResearch/hermes-agent/pull/3070))
-- Include tool tokens in preflight estimate, guard context probe persistence ([#3164](https://github.com/NousResearch/hermes-agent/pull/3164))
-- Update context compressor limits after fallback activation ([#3305](https://github.com/NousResearch/hermes-agent/pull/3305))
-- Validate empty user messages to prevent Anthropic API 400 errors ([#3322](https://github.com/NousResearch/hermes-agent/pull/3322))
-- GLM reasoning-only and max-length handling ([#3010](https://github.com/NousResearch/hermes-agent/pull/3010))
-- Increase API timeout default from 900s to 1800s for slow-thinking models ([#3431](https://github.com/NousResearch/hermes-agent/pull/3431))
-- Send `max_tokens` for Claude/OpenRouter + retry SSE connection errors ([#3497](https://github.com/NousResearch/hermes-agent/pull/3497))
-- Prevent AsyncOpenAI/httpx cross-loop deadlock in gateway mode ([#2701](https://github.com/NousResearch/hermes-agent/pull/2701)) by @ctlst
-
-### Streaming & Reasoning
-- **Persist reasoning across gateway session turns** with new schema v6 columns (`reasoning`, `reasoning_details`, `codex_reasoning_items`) ([#2974](https://github.com/NousResearch/hermes-agent/pull/2974))
-- Detect and kill stale SSE connections ([untagged commit](https://github.com/NousResearch/hermes-agent))
-- Fix stale stream detector race causing spurious `RemoteProtocolError` ([untagged commit](https://github.com/NousResearch/hermes-agent))
-- Skip duplicate callback for `<think>`-extracted reasoning during streaming ([#3116](https://github.com/NousResearch/hermes-agent/pull/3116))
-- Preserve reasoning fields in `rewrite_transcript` ([#3311](https://github.com/NousResearch/hermes-agent/pull/3311))
-- Preserve Gemini thought signatures in streamed tool calls ([#2997](https://github.com/NousResearch/hermes-agent/pull/2997))
-- Ensure first delta is fired during reasoning updates ([untagged commit](https://github.com/NousResearch/hermes-agent))
-
-### Session & Memory
-- **Session search recent sessions mode** — Omit query to browse recent sessions with titles, previews, and timestamps ([#2533](https://github.com/NousResearch/hermes-agent/pull/2533))
-- **Session config surfacing** on `/new`, `/reset`, and auto-reset ([#3321](https://github.com/NousResearch/hermes-agent/pull/3321))
-- **Third-party session isolation** — `--source` flag for isolating sessions by origin ([#3255](https://github.com/NousResearch/hermes-agent/pull/3255))
-- Add `/resume` CLI handler, session log truncation guard, `reopen_session` API ([#3315](https://github.com/NousResearch/hermes-agent/pull/3315))
-- Clear compressor summary and turn counter on `/clear` and `/new` ([#3102](https://github.com/NousResearch/hermes-agent/pull/3102))
-- Surface silent SessionDB failures that cause session data loss ([#2999](https://github.com/NousResearch/hermes-agent/pull/2999))
-- Session search fallback preview on summarization failure ([#3478](https://github.com/NousResearch/hermes-agent/pull/3478))
-- Prevent stale memory overwrites by flush agent ([#2687](https://github.com/NousResearch/hermes-agent/pull/2687))
-
-### Context Compression
-- Replace dead `summary_target_tokens` with ratio-based scaling ([#2554](https://github.com/NousResearch/hermes-agent/pull/2554))
-- Expose `compression.target_ratio`, `protect_last_n`, and `threshold` in `DEFAULT_CONFIG` ([untagged commit](https://github.com/NousResearch/hermes-agent))
-- Restore sane defaults and cap summary at 12K tokens ([untagged commit](https://github.com/NousResearch/hermes-agent))
-- Preserve transcript on `/compress` and hygiene compression ([#3556](https://github.com/NousResearch/hermes-agent/pull/3556))
-- Update context pressure warnings and token estimates after compaction ([untagged commit](https://github.com/NousResearch/hermes-agent))
-
-### Architecture & Dependencies
-- **Remove mini-swe-agent dependency** — Inline Docker and Modal backends directly ([#2804](https://github.com/NousResearch/hermes-agent/pull/2804))
-- **Replace swe-rex with native Modal SDK** for Modal backend ([#3538](https://github.com/NousResearch/hermes-agent/pull/3538))
-- **Plugin lifecycle hooks** — `pre_llm_call`, `post_llm_call`, `on_session_start`, `on_session_end` now fire in the agent loop ([#3542](https://github.com/NousResearch/hermes-agent/pull/3542))
-- Fix plugin toolsets invisible in `hermes tools` and standalone processes ([#3457](https://github.com/NousResearch/hermes-agent/pull/3457))
-- Consolidate `get_hermes_home()` and `parse_reasoning_effort()` ([#3062](https://github.com/NousResearch/hermes-agent/pull/3062))
-- Remove unused Hermes-native PKCE OAuth flow ([#3107](https://github.com/NousResearch/hermes-agent/pull/3107))
-- Remove ~100 unused imports across 55 files ([#3016](https://github.com/NousResearch/hermes-agent/pull/3016))
-- Fix 154 f-strings, simplify getattr/URL patterns, remove dead code ([#3119](https://github.com/NousResearch/hermes-agent/pull/3119))
-
----
-
-## 📱 Messaging Platforms (Gateway)
-
-### Telegram
-- **Private Chat Topics** — Project-based conversations with functional skill binding per topic, enabling isolated workflows within a single Telegram chat ([#3163](https://github.com/NousResearch/hermes-agent/pull/3163))
-- **Auto-discover fallback IPs via DNS-over-HTTPS** when `api.telegram.org` is unreachable ([#3376](https://github.com/NousResearch/hermes-agent/pull/3376))
-- **Configurable reply threading mode** ([#2907](https://github.com/NousResearch/hermes-agent/pull/2907))
-- Fall back to no `thread_id` on "Message thread not found" BadRequest ([#3390](https://github.com/NousResearch/hermes-agent/pull/3390))
-- Self-reschedule reconnect when `start_polling` fails after 502 ([#3268](https://github.com/NousResearch/hermes-agent/pull/3268))
-
-### Discord
-- Stop phantom typing indicator after agent turn completes ([#3003](https://github.com/NousResearch/hermes-agent/pull/3003))
-
-### Slack
-- Send tool call progress messages to correct Slack thread ([#3063](https://github.com/NousResearch/hermes-agent/pull/3063))
-- Scope progress thread fallback to Slack only ([#3488](https://github.com/NousResearch/hermes-agent/pull/3488))
-
-### WhatsApp
-- Download documents, audio, and video media from messages ([#2978](https://github.com/NousResearch/hermes-agent/pull/2978))
-
-### Matrix
-- Add missing Matrix entry in `PLATFORMS` dict ([#3473](https://github.com/NousResearch/hermes-agent/pull/3473))
-- Harden e2ee access-token handling ([#3562](https://github.com/NousResearch/hermes-agent/pull/3562))
-- Add backoff for `SyncError` in sync loop ([#3280](https://github.com/NousResearch/hermes-agent/pull/3280))
-
-### Signal
-- Track SSE keepalive comments as connection activity ([#3316](https://github.com/NousResearch/hermes-agent/pull/3316))
-
-### Email
-- Prevent unbounded growth of `_seen_uids` in EmailAdapter ([#3490](https://github.com/NousResearch/hermes-agent/pull/3490))
-
-### Gateway Core
-- **Config-gated `/verbose` command** for messaging platforms — toggle tool output verbosity from chat ([#3262](https://github.com/NousResearch/hermes-agent/pull/3262))
-- **Background review notifications** delivered to user chat ([#3293](https://github.com/NousResearch/hermes-agent/pull/3293))
-- **Retry transient send failures** and notify user on exhaustion ([#3288](https://github.com/NousResearch/hermes-agent/pull/3288))
-- Recover from hung agents — `/stop` hard-kills session lock ([#3104](https://github.com/NousResearch/hermes-agent/pull/3104))
-- Thread-safe `SessionStore` — protect `_entries` with `threading.Lock` ([#3052](https://github.com/NousResearch/hermes-agent/pull/3052))
-- Fix gateway token double-counting with cached agents — use absolute set instead of increment ([#3306](https://github.com/NousResearch/hermes-agent/pull/3306), [#3317](https://github.com/NousResearch/hermes-agent/pull/3317))
-- Fingerprint full auth token in agent cache signature ([#3247](https://github.com/NousResearch/hermes-agent/pull/3247))
-- Silence background agent terminal output ([#3297](https://github.com/NousResearch/hermes-agent/pull/3297))
-- Include per-platform `ALLOW_ALL` and `SIGNAL_GROUP` in startup allowlist check ([#3313](https://github.com/NousResearch/hermes-agent/pull/3313))
-- Include user-local bin paths in systemd unit PATH ([#3527](https://github.com/NousResearch/hermes-agent/pull/3527))
-- Track background task references in `GatewayRunner` ([#3254](https://github.com/NousResearch/hermes-agent/pull/3254))
-- Add request timeouts to HA, Email, Mattermost, SMS adapters ([#3258](https://github.com/NousResearch/hermes-agent/pull/3258))
-- Add media download retry to Mattermost, Slack, and base cache ([#3323](https://github.com/NousResearch/hermes-agent/pull/3323))
-- Detect virtualenv path instead of hardcoding `venv/` ([#2797](https://github.com/NousResearch/hermes-agent/pull/2797))
-- Use `TERMINAL_CWD` for context file discovery, not process cwd ([untagged commit](https://github.com/NousResearch/hermes-agent))
-- Stop loading hermes repo AGENTS.md into gateway sessions (~10k wasted tokens) ([#2891](https://github.com/NousResearch/hermes-agent/pull/2891))
-
----
-
-## 🖥️ CLI & User Experience
-
-### Interactive CLI
-- **Configurable busy input mode** + fix `/queue` always working ([#3298](https://github.com/NousResearch/hermes-agent/pull/3298))
-- **Preserve user input on multiline paste** ([#3065](https://github.com/NousResearch/hermes-agent/pull/3065))
-- **Tool generation callback** — streaming "preparing terminal…" updates during tool argument generation ([untagged commit](https://github.com/NousResearch/hermes-agent))
-- Show tool progress for substantive tools, not just "preparing" ([untagged commit](https://github.com/NousResearch/hermes-agent))
-- Buffer reasoning preview chunks and fix duplicate display ([#3013](https://github.com/NousResearch/hermes-agent/pull/3013))
-- Prevent reasoning box from rendering 3x during tool-calling loops ([#3405](https://github.com/NousResearch/hermes-agent/pull/3405))
-- Eliminate "Event loop is closed" / "Press ENTER to continue" during idle — three-layer fix with `neuter_async_httpx_del()`, custom exception handler, and stale client cleanup ([#3398](https://github.com/NousResearch/hermes-agent/pull/3398))
-- Fix status bar shows 26K instead of 260K for token counts with trailing zeros ([#3024](https://github.com/NousResearch/hermes-agent/pull/3024))
-- Fix status bar duplicates and degrades during long sessions ([#3291](https://github.com/NousResearch/hermes-agent/pull/3291))
-- Refresh TUI before background task output to prevent status bar overlap ([#3048](https://github.com/NousResearch/hermes-agent/pull/3048))
-- Suppress KawaiiSpinner animation under `patch_stdout` ([#2994](https://github.com/NousResearch/hermes-agent/pull/2994))
-- Skip KawaiiSpinner when TUI handles tool progress ([#2973](https://github.com/NousResearch/hermes-agent/pull/2973))
-- Guard `isatty()` against closed streams via `_is_tty` property ([#3056](https://github.com/NousResearch/hermes-agent/pull/3056))
-- Ensure single closure of streaming boxes during tool generation ([untagged commit](https://github.com/NousResearch/hermes-agent))
-- Cap context pressure percentage at 100% in display ([#3480](https://github.com/NousResearch/hermes-agent/pull/3480))
-- Clean up HTML error messages in CLI display ([#3069](https://github.com/NousResearch/hermes-agent/pull/3069))
-- Show HTTP status code and 400 body in API error output ([#3096](https://github.com/NousResearch/hermes-agent/pull/3096))
-- Extract useful info from HTML error pages, dump debug on max retries ([untagged commit](https://github.com/NousResearch/hermes-agent))
-- Prevent TypeError on startup when `base_url` is None ([#3068](https://github.com/NousResearch/hermes-agent/pull/3068))
-- Prevent update crash in non-TTY environments ([#3094](https://github.com/NousResearch/hermes-agent/pull/3094))
-- Handle EOFError in sessions delete/prune confirmation prompts ([#3101](https://github.com/NousResearch/hermes-agent/pull/3101))
-- Catch KeyboardInterrupt during `flush_memories` on exit and in exit cleanup handlers ([#3025](https://github.com/NousResearch/hermes-agent/pull/3025), [#3257](https://github.com/NousResearch/hermes-agent/pull/3257))
-- Guard `.strip()` against None values from YAML config ([#3552](https://github.com/NousResearch/hermes-agent/pull/3552))
-- Guard `config.get()` against YAML null values to prevent AttributeError ([#3377](https://github.com/NousResearch/hermes-agent/pull/3377))
-- Store asyncio task references to prevent GC mid-execution ([#3267](https://github.com/NousResearch/hermes-agent/pull/3267))
-
-### Setup & Configuration
-- Use explicit key mapping for returning-user menu dispatch instead of positional index ([#3083](https://github.com/NousResearch/hermes-agent/pull/3083))
-- Use `sys.executable` for pip in update commands to fix PEP 668 ([#3099](https://github.com/NousResearch/hermes-agent/pull/3099))
-- Harden `hermes update` against diverged history, non-main branches, and gateway edge cases ([#3492](https://github.com/NousResearch/hermes-agent/pull/3492))
-- OpenClaw migration overwrites defaults and setup wizard skips imported sections — fixed ([#3282](https://github.com/NousResearch/hermes-agent/pull/3282))
-- Stop recursive AGENTS.md walk, load top-level only ([#3110](https://github.com/NousResearch/hermes-agent/pull/3110))
-- Add macOS Homebrew paths to browser and terminal PATH resolution ([#2713](https://github.com/NousResearch/hermes-agent/pull/2713))
-- YAML boolean handling for `tool_progress` config ([#3300](https://github.com/NousResearch/hermes-agent/pull/3300))
-- Reset default SOUL.md to baseline identity text ([#3159](https://github.com/NousResearch/hermes-agent/pull/3159))
-- Reject relative cwd paths for container terminal backends ([untagged commit](https://github.com/NousResearch/hermes-agent))
-- Add explicit `hermes-api-server` toolset for API server platform ([#3304](https://github.com/NousResearch/hermes-agent/pull/3304))
-- Reorder setup wizard providers — OpenRouter first ([untagged commit](https://github.com/NousResearch/hermes-agent))
-
----
-
-## 🔧 Tool System
-
-### API Server
-- **Idempotency-Key support**, body size limit, and OpenAI error envelope ([#2903](https://github.com/NousResearch/hermes-agent/pull/2903))
-- Allow Idempotency-Key in CORS headers ([#3530](https://github.com/NousResearch/hermes-agent/pull/3530))
-- Cancel orphaned agent + true interrupt on SSE disconnect ([#3427](https://github.com/NousResearch/hermes-agent/pull/3427))
-- Fix streaming breaks when agent makes tool calls ([#2985](https://github.com/NousResearch/hermes-agent/pull/2985))
-
-### Terminal & File Operations
-- Handle addition-only hunks in V4A patch parser ([#3325](https://github.com/NousResearch/hermes-agent/pull/3325))
-- Exponential backoff for persistent shell polling ([#2996](https://github.com/NousResearch/hermes-agent/pull/2996))
-- Add timeout to subprocess calls in `context_references` ([#3469](https://github.com/NousResearch/hermes-agent/pull/3469))
-
-### Browser & Vision
-- Handle 402 insufficient credits error in vision tool ([#2802](https://github.com/NousResearch/hermes-agent/pull/2802))
-- Fix `browser_vision` ignores `auxiliary.vision.timeout` config ([#2901](https://github.com/NousResearch/hermes-agent/pull/2901))
-- Make browser command timeout configurable via config.yaml ([#2801](https://github.com/NousResearch/hermes-agent/pull/2801))
-
-### MCP
-- MCP toolset resolution for runtime and config ([#3252](https://github.com/NousResearch/hermes-agent/pull/3252))
-- Add MCP tool name collision protection ([#3077](https://github.com/NousResearch/hermes-agent/pull/3077))
-
-### Auxiliary LLM
-- Guard aux LLM calls against None content + reasoning fallback + retry ([#3449](https://github.com/NousResearch/hermes-agent/pull/3449))
-- Catch ImportError from `build_anthropic_client` in vision auto-detection ([#3312](https://github.com/NousResearch/hermes-agent/pull/3312))
-
-### Other Tools
-- Add request timeouts to `send_message_tool` HTTP calls ([#3162](https://github.com/NousResearch/hermes-agent/pull/3162)) by @memosr
-- Auto-repair `jobs.json` with invalid control characters ([#3537](https://github.com/NousResearch/hermes-agent/pull/3537))
-- Enable fine-grained tool streaming for Claude/OpenRouter ([#3497](https://github.com/NousResearch/hermes-agent/pull/3497))
-
----
-
-## 🧩 Skills Ecosystem
-
-### Skills System
-- **Env var passthrough** for skills and user config — skills can declare environment variables to pass through ([#2807](https://github.com/NousResearch/hermes-agent/pull/2807))
-- Cache skills prompt with shared `skill_utils` module for faster TTFT ([#3421](https://github.com/NousResearch/hermes-agent/pull/3421))
-- Avoid redundant file re-read for skill conditions ([#2992](https://github.com/NousResearch/hermes-agent/pull/2992))
-- Use Git Trees API to prevent silent subdirectory loss during install ([#2995](https://github.com/NousResearch/hermes-agent/pull/2995))
-- Fix skills-sh install for deeply nested repo structures ([#2980](https://github.com/NousResearch/hermes-agent/pull/2980))
-- Handle null metadata in skill frontmatter ([untagged commit](https://github.com/NousResearch/hermes-agent))
-- Preserve trust for skills-sh identifiers + reduce resolution churn ([#3251](https://github.com/NousResearch/hermes-agent/pull/3251))
-- Agent-created skills were incorrectly treated as untrusted community content — fixed ([untagged commit](https://github.com/NousResearch/hermes-agent))
-
-### New Skills
-- **G0DM0D3 godmode jailbreaking skill** + docs ([#3157](https://github.com/NousResearch/hermes-agent/pull/3157))
-- **Docker management skill** added to optional-skills ([#3060](https://github.com/NousResearch/hermes-agent/pull/3060))
-- **OpenClaw migration v2** — 17 new modules, terminal recap for migrating from OpenClaw to Hermes ([#2906](https://github.com/NousResearch/hermes-agent/pull/2906))
-
----
-
-## 🔒 Security & Reliability
-
-### Security Hardening
-- **SSRF protection** added to `browser_navigate` ([#3058](https://github.com/NousResearch/hermes-agent/pull/3058))
-- **SSRF protection** added to `vision_tools` and `web_tools` (hardened) ([#2679](https://github.com/NousResearch/hermes-agent/pull/2679))
-- **Restrict subagent toolsets** to parent's enabled set ([#3269](https://github.com/NousResearch/hermes-agent/pull/3269))
-- **Prevent zip-slip path traversal** in self-update ([#3250](https://github.com/NousResearch/hermes-agent/pull/3250))
-- **Prevent shell injection** in `_expand_path` via `~user` path suffix ([#2685](https://github.com/NousResearch/hermes-agent/pull/2685))
-- **Normalize input** before dangerous command detection ([#3260](https://github.com/NousResearch/hermes-agent/pull/3260))
-- Make tirith block verdicts approvable instead of hard-blocking ([#3428](https://github.com/NousResearch/hermes-agent/pull/3428))
-- Remove compromised `litellm`/`typer`/`platformdirs` from deps ([#2796](https://github.com/NousResearch/hermes-agent/pull/2796))
-- Pin all dependency version ranges ([#2810](https://github.com/NousResearch/hermes-agent/pull/2810))
-- Regenerate `uv.lock` with hashes, use lockfile in setup ([#2812](https://github.com/NousResearch/hermes-agent/pull/2812))
-- Bump dependencies to fix CVEs + regenerate `uv.lock` ([#3073](https://github.com/NousResearch/hermes-agent/pull/3073))
-- Supply chain audit CI workflow for PR scanning ([#2816](https://github.com/NousResearch/hermes-agent/pull/2816))
-
-### Reliability
-- **SQLite WAL write-lock contention** causing 15-20s TUI freeze — fixed ([#3385](https://github.com/NousResearch/hermes-agent/pull/3385))
-- **SQLite concurrency hardening** + session transcript integrity ([#3249](https://github.com/NousResearch/hermes-agent/pull/3249))
-- Prevent recurring cron job re-fire on gateway crash/restart loop ([#3396](https://github.com/NousResearch/hermes-agent/pull/3396))
-- Mark cron session as ended after job completes ([#2998](https://github.com/NousResearch/hermes-agent/pull/2998))
-
----
-
-## ⚡ Performance
-
-- **TTFT startup optimizations** — salvaged easy-win startup improvements ([#3395](https://github.com/NousResearch/hermes-agent/pull/3395))
-- Cache skills prompt with shared `skill_utils` module ([#3421](https://github.com/NousResearch/hermes-agent/pull/3421))
-- Avoid redundant file re-read for skill conditions in prompt builder ([#2992](https://github.com/NousResearch/hermes-agent/pull/2992))
-
----
-
-## 🐛 Notable Bug Fixes
-
-- Fix gateway token double-counting with cached agents ([#3306](https://github.com/NousResearch/hermes-agent/pull/3306), [#3317](https://github.com/NousResearch/hermes-agent/pull/3317))
-- Fix "Event loop is closed" / "Press ENTER to continue" during idle sessions ([#3398](https://github.com/NousResearch/hermes-agent/pull/3398))
-- Fix reasoning box rendering 3x during tool-calling loops ([#3405](https://github.com/NousResearch/hermes-agent/pull/3405))
-- Fix status bar shows 26K instead of 260K for token counts ([#3024](https://github.com/NousResearch/hermes-agent/pull/3024))
-- Fix `/queue` always working regardless of config ([#3298](https://github.com/NousResearch/hermes-agent/pull/3298))
-- Fix phantom Discord typing indicator after agent turn ([#3003](https://github.com/NousResearch/hermes-agent/pull/3003))
-- Fix Slack progress messages appearing in wrong thread ([#3063](https://github.com/NousResearch/hermes-agent/pull/3063))
-- Fix WhatsApp media downloads (documents, audio, video) ([#2978](https://github.com/NousResearch/hermes-agent/pull/2978))
-- Fix Telegram "Message thread not found" killing progress messages ([#3390](https://github.com/NousResearch/hermes-agent/pull/3390))
-- Fix OpenClaw migration overwriting defaults ([#3282](https://github.com/NousResearch/hermes-agent/pull/3282))
-- Fix returning-user setup menu dispatching wrong section ([#3083](https://github.com/NousResearch/hermes-agent/pull/3083))
-- Fix `hermes update` PEP 668 "externally-managed-environment" error ([#3099](https://github.com/NousResearch/hermes-agent/pull/3099))
-- Fix subagents hitting `max_iterations` prematurely via shared budget ([#3004](https://github.com/NousResearch/hermes-agent/pull/3004))
-- Fix YAML boolean handling for `tool_progress` config ([#3300](https://github.com/NousResearch/hermes-agent/pull/3300))
-- Fix `config.get()` crashes on YAML null values ([#3377](https://github.com/NousResearch/hermes-agent/pull/3377))
-- Fix `.strip()` crash on None values from YAML config ([#3552](https://github.com/NousResearch/hermes-agent/pull/3552))
-- Fix hung agents on gateway — `/stop` now hard-kills session lock ([#3104](https://github.com/NousResearch/hermes-agent/pull/3104))
-- Fix `_custom` provider silently remapped to `openrouter` ([#2792](https://github.com/NousResearch/hermes-agent/pull/2792))
-- Fix Matrix missing from `PLATFORMS` dict ([#3473](https://github.com/NousResearch/hermes-agent/pull/3473))
-- Fix Email adapter unbounded `_seen_uids` growth ([#3490](https://github.com/NousResearch/hermes-agent/pull/3490))
-
----
-
-## 🧪 Testing
-
-- Pin `agent-client-protocol` < 0.9 to handle breaking upstream release ([#3320](https://github.com/NousResearch/hermes-agent/pull/3320))
-- Catch anthropic ImportError in vision auto-detection tests ([#3312](https://github.com/NousResearch/hermes-agent/pull/3312))
-- Update retry-exhaust test for new graceful return behavior ([#3320](https://github.com/NousResearch/hermes-agent/pull/3320))
-- Add regression tests for null metadata frontmatter ([untagged commit](https://github.com/NousResearch/hermes-agent))
-
----
-
-## 📚 Documentation
-
-- Update all docs for `/model` command overhaul and custom provider support ([#2800](https://github.com/NousResearch/hermes-agent/pull/2800))
-- Fix stale and incorrect documentation across 18 files ([#2805](https://github.com/NousResearch/hermes-agent/pull/2805))
-- Document 9 previously undocumented features ([#2814](https://github.com/NousResearch/hermes-agent/pull/2814))
-- Add missing skills, CLI commands, and messaging env vars to docs ([#2809](https://github.com/NousResearch/hermes-agent/pull/2809))
-- Fix api-server response storage documentation — SQLite, not in-memory ([#2819](https://github.com/NousResearch/hermes-agent/pull/2819))
-- Quote pip install extras to fix zsh glob errors ([#2815](https://github.com/NousResearch/hermes-agent/pull/2815))
-- Unify hooks documentation — add plugin hooks to hooks page, add `session:end` event ([untagged commit](https://github.com/NousResearch/hermes-agent))
-- Clarify two-mode behavior in `session_search` schema description ([untagged commit](https://github.com/NousResearch/hermes-agent))
-- Fix Discord Public Bot setting for Discord-provided invite link ([#3519](https://github.com/NousResearch/hermes-agent/pull/3519)) by @mehmoodosman
-- Revise v0.4.0 changelog — fix feature attribution, reorder sections ([untagged commit](https://github.com/NousResearch/hermes-agent))
-
----
-
-## 👥 Contributors
-
-### Core
-- **@teknium1** — 157 PRs covering the full scope of this release
-
-### Community Contributors
-- **@alt-glitch** (Siddharth Balyan) — 2 PRs: Nix flake with uv2nix build, NixOS module, and persistent container mode ([#20](https://github.com/NousResearch/hermes-agent/pull/20)); auto-generated config keys and suffix PATHs for Nix builds ([#3061](https://github.com/NousResearch/hermes-agent/pull/3061), [#3274](https://github.com/NousResearch/hermes-agent/pull/3274))
-- **@ctlst** — 1 PR: Prevent AsyncOpenAI/httpx cross-loop deadlock in gateway mode ([#2701](https://github.com/NousResearch/hermes-agent/pull/2701))
-- **@memosr** (memosr.eth) — 1 PR: Add request timeouts to `send_message_tool` HTTP calls ([#3162](https://github.com/NousResearch/hermes-agent/pull/3162))
-- **@mehmoodosman** (Osman Mehmood) — 1 PR: Fix Discord docs for Public Bot setting ([#3519](https://github.com/NousResearch/hermes-agent/pull/3519))
-
-### All Contributors
-@alt-glitch, @ctlst, @mehmoodosman, @memosr, @teknium1
-
----
-
-**Full Changelog**: [v2026.3.23...v2026.3.28](https://github.com/NousResearch/hermes-agent/compare/v2026.3.23...v2026.3.28)

RELEASE_v0.6.0.md

@@ -1,249 +0,0 @@
-# Hermes Agent v0.6.0 (v2026.3.30)
-
-**Release Date:** March 30, 2026
-
-> The multi-instance release — Profiles for running isolated agent instances, MCP server mode, Docker container, fallback provider chains, two new messaging platforms (Feishu/Lark and WeCom), Telegram webhook mode, Slack multi-workspace OAuth, 95 PRs and 16 resolved issues in 2 days.
-
----
-
-## ✨ Highlights
-
-- **Profiles — Multi-Instance Hermes** — Run multiple isolated Hermes instances from the same installation. Each profile gets its own config, memory, sessions, skills, and gateway service. Create with `hermes profile create`, switch with `hermes -p <name>`, export/import for sharing. Full token-lock isolation prevents two profiles from using the same bot credential. ([#3681](https://github.com/NousResearch/hermes-agent/pull/3681))
-
-- **MCP Server Mode** — Expose Hermes conversations and sessions to any MCP-compatible client (Claude Desktop, Cursor, VS Code, etc.) via `hermes mcp serve`. Browse conversations, read messages, search across sessions, and manage attachments — all through the Model Context Protocol. Supports both stdio and Streamable HTTP transports. ([#3795](https://github.com/NousResearch/hermes-agent/pull/3795))
-
-- **Docker Container** — Official Dockerfile for running Hermes Agent in a container. Supports both CLI and gateway modes with volume-mounted config. ([#3668](https://github.com/NousResearch/hermes-agent/pull/3668), closes [#850](https://github.com/NousResearch/hermes-agent/issues/850))
-
-- **Ordered Fallback Provider Chain** — Configure multiple inference providers with automatic failover. When your primary provider returns errors or is unreachable, Hermes automatically tries the next provider in the chain. Configure via `fallback_providers` in config.yaml. ([#3813](https://github.com/NousResearch/hermes-agent/pull/3813), closes [#1734](https://github.com/NousResearch/hermes-agent/issues/1734))
-
-- **Feishu/Lark Platform Support** — Full gateway adapter for Feishu (飞书) and Lark with event subscriptions, message cards, group chat, image/file attachments, and interactive card callbacks. ([#3799](https://github.com/NousResearch/hermes-agent/pull/3799), [#3817](https://github.com/NousResearch/hermes-agent/pull/3817), closes [#1788](https://github.com/NousResearch/hermes-agent/issues/1788))
-
-- **WeCom (Enterprise WeChat) Platform Support** — New gateway adapter for WeCom (企业微信) with text/image/voice messages, group chats, and callback verification. ([#3847](https://github.com/NousResearch/hermes-agent/pull/3847))
-
-- **Slack Multi-Workspace OAuth** — Connect a single Hermes gateway to multiple Slack workspaces via OAuth token file. Each workspace gets its own bot token, resolved dynamically per incoming event. ([#3903](https://github.com/NousResearch/hermes-agent/pull/3903))
-
-- **Telegram Webhook Mode & Group Controls** — Run the Telegram adapter in webhook mode as an alternative to polling — faster response times and better for production deployments behind a reverse proxy. New group mention gating controls when the bot responds: always, only when @mentioned, or via regex triggers. ([#3880](https://github.com/NousResearch/hermes-agent/pull/3880), [#3870](https://github.com/NousResearch/hermes-agent/pull/3870))
-
-- **Exa Search Backend** — Add Exa as an alternative web search and content extraction backend alongside Firecrawl and DuckDuckGo. Set `EXA_API_KEY` and configure as preferred backend. ([#3648](https://github.com/NousResearch/hermes-agent/pull/3648))
-
-- **Skills & Credentials on Remote Backends** — Mount skill directories and credential files into Modal and Docker containers, so remote terminal sessions have access to the same skills and secrets as local execution. ([#3890](https://github.com/NousResearch/hermes-agent/pull/3890), [#3671](https://github.com/NousResearch/hermes-agent/pull/3671), closes [#3665](https://github.com/NousResearch/hermes-agent/issues/3665), [#3433](https://github.com/NousResearch/hermes-agent/issues/3433))
-
----
-
-## 🏗️ Core Agent & Architecture
-
-### Provider & Model Support
-- **Ordered fallback provider chain** — automatic failover across multiple configured providers ([#3813](https://github.com/NousResearch/hermes-agent/pull/3813))
-- **Fix api_mode on provider switch** — switching providers via `hermes model` now correctly clears stale `api_mode` instead of hardcoding `chat_completions`, fixing 404s for providers with Anthropic-compatible endpoints ([#3726](https://github.com/NousResearch/hermes-agent/pull/3726), [#3857](https://github.com/NousResearch/hermes-agent/pull/3857), closes [#3685](https://github.com/NousResearch/hermes-agent/issues/3685))
-- **Stop silent OpenRouter fallback** — when no provider is configured, Hermes now raises a clear error instead of silently routing to OpenRouter ([#3807](https://github.com/NousResearch/hermes-agent/pull/3807), [#3862](https://github.com/NousResearch/hermes-agent/pull/3862))
-- **Gemini 3.1 preview models** — added to OpenRouter and Nous Portal catalogs ([#3803](https://github.com/NousResearch/hermes-agent/pull/3803), closes [#3753](https://github.com/NousResearch/hermes-agent/issues/3753))
-- **Gemini direct API context length** — full context length resolution for direct Google AI endpoints ([#3876](https://github.com/NousResearch/hermes-agent/pull/3876))
-- **gpt-5.4-mini** added to Codex fallback catalog ([#3855](https://github.com/NousResearch/hermes-agent/pull/3855))
-- **Curated model lists preferred** over live API probe when the probe returns fewer models ([#3856](https://github.com/NousResearch/hermes-agent/pull/3856), [#3867](https://github.com/NousResearch/hermes-agent/pull/3867))
-- **User-friendly 429 rate limit messages** with Retry-After countdown ([#3809](https://github.com/NousResearch/hermes-agent/pull/3809))
-- **Auxiliary client placeholder key** for local servers without auth requirements ([#3842](https://github.com/NousResearch/hermes-agent/pull/3842))
-- **INFO-level logging** for auxiliary provider resolution ([#3866](https://github.com/NousResearch/hermes-agent/pull/3866))
-
-### Agent Loop & Conversation
-- **Subagent status reporting** — reports `completed` status when summary exists instead of generic failure ([#3829](https://github.com/NousResearch/hermes-agent/pull/3829))
-- **Session log file updated during compression** — prevents stale file references after context compression ([#3835](https://github.com/NousResearch/hermes-agent/pull/3835))
-- **Omit empty tools param** — sends no `tools` parameter when empty instead of `None`, fixing compatibility with strict providers ([#3820](https://github.com/NousResearch/hermes-agent/pull/3820))
-
-### Profiles & Multi-Instance
-- **Profiles system** — `hermes profile create/list/switch/delete/export/import/rename`. Each profile gets isolated HERMES_HOME, gateway service, CLI wrapper. Token locks prevent credential collisions. Tab completion for profile names. ([#3681](https://github.com/NousResearch/hermes-agent/pull/3681))
-- **Profile-aware display paths** — all user-facing `~/.hermes` paths replaced with `display_hermes_home()` to show the correct profile directory ([#3623](https://github.com/NousResearch/hermes-agent/pull/3623))
-- **Lazy display_hermes_home imports** — prevents `ImportError` during `hermes update` when modules cache stale bytecode ([#3776](https://github.com/NousResearch/hermes-agent/pull/3776))
-- **HERMES_HOME for protected paths** — `.env` write-deny path now respects HERMES_HOME instead of hardcoded `~/.hermes` ([#3840](https://github.com/NousResearch/hermes-agent/pull/3840))
-
----
-
-## 📱 Messaging Platforms (Gateway)
-
-### New Platforms
-- **Feishu/Lark** — Full adapter with event subscriptions, message cards, group chat, image/file attachments, interactive card callbacks ([#3799](https://github.com/NousResearch/hermes-agent/pull/3799), [#3817](https://github.com/NousResearch/hermes-agent/pull/3817))
-- **WeCom (Enterprise WeChat)** — Text/image/voice messages, group chats, callback verification ([#3847](https://github.com/NousResearch/hermes-agent/pull/3847))
-
-### Telegram
-- **Webhook mode** — run as webhook endpoint instead of polling for production deployments ([#3880](https://github.com/NousResearch/hermes-agent/pull/3880))
-- **Group mention gating & regex triggers** — configurable bot response behavior in groups: always, @mention-only, or regex-matched ([#3870](https://github.com/NousResearch/hermes-agent/pull/3870))
-- **Gracefully handle deleted reply targets** — no more crashes when the message being replied to was deleted ([#3858](https://github.com/NousResearch/hermes-agent/pull/3858), closes [#3229](https://github.com/NousResearch/hermes-agent/issues/3229))
-
-### Discord
-- **Message processing reactions** — adds a reaction emoji while processing and removes it when done, giving visual feedback in channels ([#3871](https://github.com/NousResearch/hermes-agent/pull/3871))
-- **DISCORD_IGNORE_NO_MENTION** — skip messages that @mention other users/bots but not Hermes ([#3640](https://github.com/NousResearch/hermes-agent/pull/3640))
-- **Clean up deferred "thinking..."** — properly removes the "thinking..." indicator after slash commands complete ([#3674](https://github.com/NousResearch/hermes-agent/pull/3674), closes [#3595](https://github.com/NousResearch/hermes-agent/issues/3595))
-
-### Slack
-- **Multi-workspace OAuth** — connect to multiple Slack workspaces from a single gateway via OAuth token file ([#3903](https://github.com/NousResearch/hermes-agent/pull/3903))
-
-### WhatsApp
-- **Persistent aiohttp session** — reuse HTTP sessions across requests instead of creating new ones per message ([#3818](https://github.com/NousResearch/hermes-agent/pull/3818))
-- **LID↔phone alias resolution** — correctly match Linked ID and phone number formats in allowlists ([#3830](https://github.com/NousResearch/hermes-agent/pull/3830))
-- **Skip reply prefix in bot mode** — cleaner message formatting when running as a WhatsApp bot ([#3931](https://github.com/NousResearch/hermes-agent/pull/3931))
-
-### Matrix
-- **Native voice messages via MSC3245** — send voice messages as proper Matrix voice events instead of file attachments ([#3877](https://github.com/NousResearch/hermes-agent/pull/3877))
-
-### Mattermost
-- **Configurable mention behavior** — respond to messages without requiring @mention ([#3664](https://github.com/NousResearch/hermes-agent/pull/3664))
-
-### Signal
-- **URL-encode phone numbers** and correct attachment RPC parameter — fixes delivery failures with certain phone number formats ([#3670](https://github.com/NousResearch/hermes-agent/pull/3670)) — @kshitijk4poor
-
-### Email
-- **Close SMTP/IMAP connections on failure** — prevents connection leaks during error scenarios ([#3804](https://github.com/NousResearch/hermes-agent/pull/3804))
-
-### Gateway Core
-- **Atomic config writes** — use atomic file writes for config.yaml to prevent data loss during crashes ([#3800](https://github.com/NousResearch/hermes-agent/pull/3800))
-- **Home channel env overrides** — apply environment variable overrides for home channels consistently ([#3796](https://github.com/NousResearch/hermes-agent/pull/3796), [#3808](https://github.com/NousResearch/hermes-agent/pull/3808))
-- **Replace print() with logger** — BasePlatformAdapter now uses proper logging instead of print statements ([#3669](https://github.com/NousResearch/hermes-agent/pull/3669))
-- **Cron delivery labels** — resolve human-friendly delivery labels via channel directory ([#3860](https://github.com/NousResearch/hermes-agent/pull/3860), closes [#1945](https://github.com/NousResearch/hermes-agent/issues/1945))
-- **Cron [SILENT] tightening** — prevent agents from prefixing reports with [SILENT] to suppress delivery ([#3901](https://github.com/NousResearch/hermes-agent/pull/3901))
-- **Background task media delivery** and vision download timeout fixes ([#3919](https://github.com/NousResearch/hermes-agent/pull/3919))
-- **Boot-md hook** — example built-in hook to run a BOOT.md file on gateway startup ([#3733](https://github.com/NousResearch/hermes-agent/pull/3733))
-
----
-
-## 🖥️ CLI & User Experience
-
-### Interactive CLI
-- **Configurable tool preview length** — show full file paths by default instead of truncating at 40 chars ([#3841](https://github.com/NousResearch/hermes-agent/pull/3841))
-- **Tool token context display** — `hermes tools` checklist now shows estimated token cost per toolset ([#3805](https://github.com/NousResearch/hermes-agent/pull/3805))
-- **/bg spinner TUI fix** — route background task spinner through the TUI widget to prevent status bar collision ([#3643](https://github.com/NousResearch/hermes-agent/pull/3643))
-- **Prevent status bar wrapping** into duplicate rows ([#3883](https://github.com/NousResearch/hermes-agent/pull/3883)) — @kshitijk4poor
-- **Handle closed stdout ValueError** in safe print paths — fixes crashes when stdout is closed during gateway thread shutdown ([#3843](https://github.com/NousResearch/hermes-agent/pull/3843), closes [#3534](https://github.com/NousResearch/hermes-agent/issues/3534))
-- **Remove input() from /tools disable** — eliminates freeze in terminal when disabling tools ([#3918](https://github.com/NousResearch/hermes-agent/pull/3918))
-- **TTY guard for interactive CLI commands** — prevent CPU spin when launched without a terminal ([#3933](https://github.com/NousResearch/hermes-agent/pull/3933))
-- **Argparse entrypoint** — use argparse in the top-level launcher for cleaner error handling ([#3874](https://github.com/NousResearch/hermes-agent/pull/3874))
-- **Lazy-initialized tools show yellow** in banner instead of red, reducing false alarm about "missing" tools ([#3822](https://github.com/NousResearch/hermes-agent/pull/3822))
-- **Honcho tools shown in banner** when configured ([#3810](https://github.com/NousResearch/hermes-agent/pull/3810))
-
-### Setup & Configuration
-- **Auto-install matrix-nio** during `hermes setup` when Matrix is selected ([#3802](https://github.com/NousResearch/hermes-agent/pull/3802), [#3873](https://github.com/NousResearch/hermes-agent/pull/3873))
-- **Session export stdout support** — export sessions to stdout with `-` for piping ([#3641](https://github.com/NousResearch/hermes-agent/pull/3641), closes [#3609](https://github.com/NousResearch/hermes-agent/issues/3609))
-- **Configurable approval timeouts** — set how long dangerous command approval prompts wait before auto-denying ([#3886](https://github.com/NousResearch/hermes-agent/pull/3886), closes [#3765](https://github.com/NousResearch/hermes-agent/issues/3765))
-- **Clear __pycache__ during update** — prevents stale bytecode ImportError after `hermes update` ([#3819](https://github.com/NousResearch/hermes-agent/pull/3819))
-
----
-
-## 🔧 Tool System
-
-### MCP
-- **MCP Server Mode** — `hermes mcp serve` exposes conversations, sessions, and attachments to MCP clients via stdio or Streamable HTTP ([#3795](https://github.com/NousResearch/hermes-agent/pull/3795))
-- **Dynamic tool discovery** — respond to `notifications/tools/list_changed` events to pick up new tools from MCP servers without reconnecting ([#3812](https://github.com/NousResearch/hermes-agent/pull/3812))
-- **Non-deprecated HTTP transport** — switched from `sse_client` to `streamable_http_client` ([#3646](https://github.com/NousResearch/hermes-agent/pull/3646))
-
-### Web Tools
-- **Exa search backend** — alternative to Firecrawl and DuckDuckGo for web search and extraction ([#3648](https://github.com/NousResearch/hermes-agent/pull/3648))
-
-### Browser
-- **Guard against None LLM responses** in browser snapshot and vision tools ([#3642](https://github.com/NousResearch/hermes-agent/pull/3642))
-
-### Terminal & Remote Backends
-- **Mount skill directories** into Modal and Docker containers ([#3890](https://github.com/NousResearch/hermes-agent/pull/3890))
-- **Mount credential files** into remote backends with mtime+size caching ([#3671](https://github.com/NousResearch/hermes-agent/pull/3671))
-- **Preserve partial output** when commands time out instead of losing everything ([#3868](https://github.com/NousResearch/hermes-agent/pull/3868))
-- **Stop marking persisted env vars as missing** on remote backends ([#3650](https://github.com/NousResearch/hermes-agent/pull/3650))
-
-### Audio
-- **.aac format support** in transcription tool ([#3865](https://github.com/NousResearch/hermes-agent/pull/3865), closes [#1963](https://github.com/NousResearch/hermes-agent/issues/1963))
-- **Audio download retry** — retry logic for `cache_audio_from_url` matching the existing image download pattern ([#3401](https://github.com/NousResearch/hermes-agent/pull/3401)) — @binhnt92
-
-### Vision
-- **Reject non-image files** and enforce website-only policy for vision analysis ([#3845](https://github.com/NousResearch/hermes-agent/pull/3845))
-
-### Tool Schema
-- **Ensure name field** always present in tool definitions, fixing `KeyError: 'name'` crashes ([#3811](https://github.com/NousResearch/hermes-agent/pull/3811), closes [#3729](https://github.com/NousResearch/hermes-agent/issues/3729))
-
-### ACP (Editor Integration)
-- **Complete session management surface** for VS Code/Zed/JetBrains clients — proper task lifecycle, cancel support, session persistence ([#3675](https://github.com/NousResearch/hermes-agent/pull/3675))
-
----
-
-## 🧩 Skills & Plugins
-
-### Skills System
-- **External skill directories** — configure additional skill directories via `skills.external_dirs` in config.yaml ([#3678](https://github.com/NousResearch/hermes-agent/pull/3678))
-- **Category path traversal blocked** — prevents `../` attacks in skill category names ([#3844](https://github.com/NousResearch/hermes-agent/pull/3844))
-- **parallel-cli moved to optional-skills** — reduces default skill footprint ([#3673](https://github.com/NousResearch/hermes-agent/pull/3673)) — @kshitijk4poor
-
-### New Skills
-- **memento-flashcards** — spaced repetition flashcard system ([#3827](https://github.com/NousResearch/hermes-agent/pull/3827))
-- **songwriting-and-ai-music** — songwriting craft and AI music generation prompts ([#3834](https://github.com/NousResearch/hermes-agent/pull/3834))
-- **SiYuan Note** — integration with SiYuan note-taking app ([#3742](https://github.com/NousResearch/hermes-agent/pull/3742))
-- **Scrapling** — web scraping skill using Scrapling library ([#3742](https://github.com/NousResearch/hermes-agent/pull/3742))
-- **one-three-one-rule** — communication framework skill ([#3797](https://github.com/NousResearch/hermes-agent/pull/3797))
-
-### Plugin System
-- **Plugin enable/disable commands** — `hermes plugins enable/disable <name>` for managing plugin state without removing them ([#3747](https://github.com/NousResearch/hermes-agent/pull/3747))
-- **Plugin message injection** — plugins can now inject messages into the conversation stream on behalf of the user via `ctx.inject_message()` ([#3778](https://github.com/NousResearch/hermes-agent/pull/3778)) — @winglian
-- **Honcho self-hosted support** — allow local Honcho instances without requiring an API key ([#3644](https://github.com/NousResearch/hermes-agent/pull/3644))
-
----
-
-## 🔒 Security & Reliability
-
-### Security Hardening
-- **Hardened dangerous command detection** — expanded pattern matching for risky shell commands and added file tool path guards for sensitive locations (`/etc/`, `/boot/`, docker.sock) ([#3872](https://github.com/NousResearch/hermes-agent/pull/3872))
-- **Sensitive path write checks** in approval system — catch writes to system config files through file tools, not just terminal ([#3859](https://github.com/NousResearch/hermes-agent/pull/3859))
-- **Secret redaction expansion** — now covers ElevenLabs, Tavily, and Exa API keys ([#3920](https://github.com/NousResearch/hermes-agent/pull/3920))
-- **Vision file rejection** — reject non-image files passed to vision analysis to prevent information disclosure ([#3845](https://github.com/NousResearch/hermes-agent/pull/3845))
-- **Category path traversal blocking** — prevent directory traversal in skill category names ([#3844](https://github.com/NousResearch/hermes-agent/pull/3844))
-
-### Reliability
-- **Atomic config.yaml writes** — prevent data loss during gateway crashes ([#3800](https://github.com/NousResearch/hermes-agent/pull/3800))
-- **Clear __pycache__ on update** — prevent stale bytecode from causing ImportError after updates ([#3819](https://github.com/NousResearch/hermes-agent/pull/3819))
-- **Lazy imports for update safety** — prevent ImportError chains during `hermes update` when modules reference new functions ([#3776](https://github.com/NousResearch/hermes-agent/pull/3776))
-- **Restore terminalbench2 from patch corruption** — recovered file damaged by patch tool's secret redaction ([#3801](https://github.com/NousResearch/hermes-agent/pull/3801))
-- **Terminal timeout preserves partial output** — no more lost command output on timeout ([#3868](https://github.com/NousResearch/hermes-agent/pull/3868))
-
----
-
-## 🐛 Notable Bug Fixes
-
-- **OpenClaw migration model config overwrite** — migration no longer overwrites model config dict with a string ([#3924](https://github.com/NousResearch/hermes-agent/pull/3924)) — @0xbyt4
-- **OpenClaw migration expanded** — covers full data footprint including sessions, cron, memory ([#3869](https://github.com/NousResearch/hermes-agent/pull/3869))
-- **Telegram deleted reply targets** — gracefully handle replies to deleted messages instead of crashing ([#3858](https://github.com/NousResearch/hermes-agent/pull/3858))
-- **Discord "thinking..." persistence** — properly cleans up deferred response indicators ([#3674](https://github.com/NousResearch/hermes-agent/pull/3674))
-- **WhatsApp LID↔phone aliases** — fixes allowlist matching failures with Linked ID format ([#3830](https://github.com/NousResearch/hermes-agent/pull/3830))
-- **Signal URL-encoded phone numbers** — fixes delivery failures with certain formats ([#3670](https://github.com/NousResearch/hermes-agent/pull/3670))
-- **Email connection leaks** — properly close SMTP/IMAP connections on error ([#3804](https://github.com/NousResearch/hermes-agent/pull/3804))
-- **_safe_print ValueError** — no more gateway thread crashes on closed stdout ([#3843](https://github.com/NousResearch/hermes-agent/pull/3843))
-- **Tool schema KeyError 'name'** — ensure name field always present in tool definitions ([#3811](https://github.com/NousResearch/hermes-agent/pull/3811))
-- **api_mode stale on provider switch** — correctly clear when switching providers via `hermes model` ([#3857](https://github.com/NousResearch/hermes-agent/pull/3857))
-
----
-
-## 🧪 Testing
-
-- Resolved 10+ CI failures across hooks, tiktoken, plugins, and skill tests ([#3848](https://github.com/NousResearch/hermes-agent/pull/3848), [#3721](https://github.com/NousResearch/hermes-agent/pull/3721), [#3936](https://github.com/NousResearch/hermes-agent/pull/3936))
-
----
-
-## 📚 Documentation
-
-- **Comprehensive OpenClaw migration guide** — step-by-step guide for migrating from OpenClaw/Claw3D to Hermes Agent ([#3864](https://github.com/NousResearch/hermes-agent/pull/3864), [#3900](https://github.com/NousResearch/hermes-agent/pull/3900))
-- **Credential file passthrough docs** — document how to forward credential files and env vars to remote backends ([#3677](https://github.com/NousResearch/hermes-agent/pull/3677))
-- **DuckDuckGo requirements clarified** — note runtime dependency on duckduckgo-search package ([#3680](https://github.com/NousResearch/hermes-agent/pull/3680))
-- **Skills catalog updated** — added red-teaming category and optional skills listing ([#3745](https://github.com/NousResearch/hermes-agent/pull/3745))
-- **Feishu docs MDX fix** — escape angle-bracket URLs that break Docusaurus build ([#3902](https://github.com/NousResearch/hermes-agent/pull/3902))
-
----
-
-## 👥 Contributors
-
-### Core
-- **@teknium1** — 90 PRs across all subsystems
-
-### Community Contributors
-- **@kshitijk4poor** — 3 PRs: Signal phone number fix ([#3670](https://github.com/NousResearch/hermes-agent/pull/3670)), parallel-cli to optional-skills ([#3673](https://github.com/NousResearch/hermes-agent/pull/3673)), status bar wrapping fix ([#3883](https://github.com/NousResearch/hermes-agent/pull/3883))
-- **@winglian** — 1 PR: Plugin message injection interface ([#3778](https://github.com/NousResearch/hermes-agent/pull/3778))
-- **@binhnt92** — 1 PR: Audio download retry logic ([#3401](https://github.com/NousResearch/hermes-agent/pull/3401))
-- **@0xbyt4** — 1 PR: OpenClaw migration model config fix ([#3924](https://github.com/NousResearch/hermes-agent/pull/3924))
-
-### Issues Resolved from Community
-@Material-Scientist ([#850](https://github.com/NousResearch/hermes-agent/issues/850)), @hanxu98121 ([#1734](https://github.com/NousResearch/hermes-agent/issues/1734)), @penwyp ([#1788](https://github.com/NousResearch/hermes-agent/issues/1788)), @dan-and ([#1945](https://github.com/NousResearch/hermes-agent/issues/1945)), @AdrianScott ([#1963](https://github.com/NousResearch/hermes-agent/issues/1963)), @clawdbot47 ([#3229](https://github.com/NousResearch/hermes-agent/issues/3229)), @alanfwilliams ([#3404](https://github.com/NousResearch/hermes-agent/issues/3404)), @kentimsit ([#3433](https://github.com/NousResearch/hermes-agent/issues/3433)), @hayka-pacha ([#3534](https://github.com/NousResearch/hermes-agent/issues/3534)), @primmer ([#3595](https://github.com/NousResearch/hermes-agent/issues/3595)), @dagelf ([#3609](https://github.com/NousResearch/hermes-agent/issues/3609)), @HenkDz ([#3685](https://github.com/NousResearch/hermes-agent/issues/3685)), @tmdgusya ([#3729](https://github.com/NousResearch/hermes-agent/issues/3729)), @TypQxQ ([#3753](https://github.com/NousResearch/hermes-agent/issues/3753)), @acsezen ([#3765](https://github.com/NousResearch/hermes-agent/issues/3765))
-
----
-
-**Full Changelog**: [v2026.3.28...v2026.3.30](https://github.com/NousResearch/hermes-agent/compare/v2026.3.28...v2026.3.30)

TODO.md

@@ -0,0 +1,589 @@
+# Hermes Agent - Future Improvements
+
+> Ideas for enhancing the agent's capabilities, generated from self-analysis of the codebase.
+
+---
+
+## 1. Subagent Architecture (Context Isolation) 🎯
+
+**Problem:** Long-running tools (terminal commands, browser automation, complex file operations) consume massive context. A single `ls -la` can add hundreds of lines. Browser snapshots, debugging sessions, and iterative terminal work quickly bloat the main conversation, leaving less room for actual reasoning.
+
+**Solution:** The main agent becomes an **orchestrator** that delegates context-heavy tasks to **subagents**.
+
+**Architecture:**
+```
+┌─────────────────────────────────────────────────────────────────┐
+│  ORCHESTRATOR (main agent)                                      │
+│  - Receives user request                                        │
+│  - Plans approach                                               │
+│  - Delegates heavy tasks to subagents                           │
+│  - Receives summarized results                                  │
+│  - Maintains clean, focused context                             │
+└─────────────────────────────────────────────────────────────────┘
+         │                    │                    │
+         ▼                    ▼                    ▼
+┌─────────────────┐  ┌─────────────────┐  ┌─────────────────┐
+│ TERMINAL AGENT  │  │ BROWSER AGENT   │  │ CODE AGENT      │
+│ - terminal tool │  │ - browser tools │  │ - file tools    │
+│ - file tools    │  │ - web_search    │  │ - terminal      │
+│                 │  │ - web_extract   │  │                 │
+│ Isolated context│  │ Isolated context│  │ Isolated context│
+│ Returns summary │  │ Returns summary │  │ Returns summary │
+└─────────────────┘  └─────────────────┘  └─────────────────┘
+```
+
+**How it works:**
+1. User asks: "Set up a new Python project with FastAPI and tests"
+2. Orchestrator plans: "I need to create files, install deps, write code"
+3. Orchestrator calls: `terminal_task(goal="Create venv, install fastapi pytest", context="New project in ~/myapp")`
+4. **Subagent spawns** with fresh context, only terminal/file tools
+5. Subagent iterates (may take 10+ tool calls, lots of output)
+6. Subagent completes → returns summary: "Created venv, installed fastapi==0.109.0, pytest==8.0.0"
+7. Orchestrator receives **only the summary**, context stays clean
+8. Orchestrator continues with next subtask
+
+**Key tools to implement:**
+- [ ] `terminal_task(goal, context, cwd?)` - Delegate terminal/shell work
+- [ ] `browser_task(goal, context, start_url?)` - Delegate web research/automation  
+- [ ] `code_task(goal, context, files?)` - Delegate code writing/modification
+- [ ] Generic `delegate_task(goal, context, toolsets=[])` - Flexible delegation
+
+**Implementation details:**
+- [ ] Subagent uses same `run_agent.py` but with:
+  - Fresh/empty conversation history
+  - Limited toolset (only what's needed)
+  - Smaller max_iterations (focused task)
+  - Task-specific system prompt
+- [ ] Subagent returns structured result:
+  ```python
+  {
+    "success": True,
+    "summary": "Installed 3 packages, created 2 files",
+    "details": "Optional longer explanation if needed",
+    "artifacts": ["~/myapp/requirements.txt", "~/myapp/main.py"],  # Files created
+    "errors": []  # Any issues encountered
+  }
+  ```
+- [ ] Orchestrator sees only the summary in its context
+- [ ] Full subagent transcript saved separately for debugging
+
+**Benefits:**
+- 🧹 **Clean context** - Orchestrator stays focused, doesn't drown in tool output
+- 📊 **Better token efficiency** - 50 terminal outputs → 1 summary paragraph
+- 🎯 **Focused subagents** - Each agent has just the tools it needs
+- 🔄 **Parallel potential** - Independent subtasks could run concurrently
+- 🐛 **Easier debugging** - Each subtask has its own isolated transcript
+
+**When to use subagents vs direct tools:**
+- **Subagent**: Multi-step tasks, iteration likely, lots of output expected
+- **Direct**: Quick one-off commands, simple file reads, user needs to see output
+
+**Files to modify:** `run_agent.py` (add orchestration mode), new `tools/delegate_tools.py`, new `subagent_runner.py`
+
+---
+
+## 2. Planning & Task Management 📋
+
+**Problem:** Agent handles tasks reactively without explicit planning. Complex multi-step tasks lack structure, progress tracking, and the ability to decompose work into manageable chunks.
+
+**Ideas:**
+- [ ] **Task decomposition tool** - Break complex requests into subtasks:
+  ```
+  User: "Set up a new Python project with FastAPI, tests, and Docker"
+  
+  Agent creates plan:
+  ├── 1. Create project structure and requirements.txt
+  ├── 2. Implement FastAPI app skeleton
+  ├── 3. Add pytest configuration and initial tests
+  ├── 4. Create Dockerfile and docker-compose.yml
+  └── 5. Verify everything works together
+  ```
+  - Each subtask becomes a trackable unit
+  - Agent can report progress: "Completed 3/5 tasks"
+  
+- [ ] **Progress checkpoints** - Periodic self-assessment:
+  - After N tool calls or time elapsed, pause to evaluate
+  - "What have I accomplished? What remains? Am I on track?"
+  - Detect if stuck in loops or making no progress
+  - Could trigger replanning if approach isn't working
+  
+- [ ] **Explicit plan storage** - Persist plan in conversation:
+  - Store as structured data (not just in context)
+  - Update status as tasks complete
+  - User can ask "What's the plan?" or "What's left?"
+  - Survives context compression (plans are protected)
+
+- [ ] **Failure recovery with replanning** - When things go wrong:
+  - Record what failed and why
+  - Revise plan to work around the issue
+  - "Step 3 failed because X, adjusting approach to Y"
+  - Prevents repeating failed strategies
+
+**Files to modify:** `run_agent.py` (add planning hooks), new `tools/planning_tool.py`
+
+---
+
+## 3. Dynamic Skills Expansion 📚
+
+**Problem:** Skills system is elegant but static. Skills must be manually created and added.
+
+**Ideas:**
+- [ ] **Skill acquisition from successful tasks** - After completing a complex task:
+  - "This approach worked well. Save as a skill?"
+  - Extract: goal, steps taken, tools used, key decisions
+  - Generate SKILL.md automatically
+  - Store in user's skills directory
+  
+- [ ] **Skill templates** - Common patterns that can be parameterized:
+  ```markdown
+  # Debug {language} Error
+  1. Reproduce the error
+  2. Search for error message: `web_search("{error_message} {language}")`
+  3. Check common causes: {common_causes}
+  4. Apply fix and verify
+  ```
+  
+- [ ] **Skill chaining** - Combine skills for complex workflows:
+  - Skills can reference other skills as dependencies
+  - "To do X, first apply skill Y, then skill Z"
+  - Directed graph of skill dependencies
+
+**Files to modify:** `tools/skills_tool.py`, `skills/` directory structure, new `skill_generator.py`
+
+---
+
+## 4. Interactive Clarifying Questions Tool ❓
+
+**Problem:** Agent sometimes makes assumptions or guesses when it should ask the user. Currently can only ask via text, which gets lost in long outputs.
+
+**Ideas:**
+- [ ] **Multiple-choice prompt tool** - Let agent present structured choices to user:
+  ```
+  ask_user_choice(
+    question="Should the language switcher enable only German or all languages?",
+    choices=[
+      "Only enable German - works immediately",
+      "Enable all, mark untranslated - show fallback notice",
+      "Let me specify something else"
+    ]
+  )
+  ```
+  - Renders as interactive terminal UI with arrow key / Tab navigation
+  - User selects option, result returned to agent
+  - Up to 4 choices + optional free-text option
+  
+- [ ] **Implementation:**
+  - Use `inquirer` or `questionary` Python library for rich terminal prompts
+  - Tool returns selected option text (or user's custom input)
+  - **CLI-only** - only works when running via `cli.py` (not API/programmatic use)
+  - Graceful fallback: if not in interactive mode, return error asking agent to rephrase as text
+  
+- [ ] **Use cases:**
+  - Clarify ambiguous requirements before starting work
+  - Confirm destructive operations with clear options
+  - Let user choose between implementation approaches
+  - Checkpoint complex multi-step workflows
+
+**Files to modify:** New `tools/ask_user_tool.py`, `cli.py` (detect interactive mode), `model_tools.py`
+
+---
+
+## 5. Collaborative Problem Solving 🤝
+
+**Problem:** Interaction is command/response. Complex problems benefit from dialogue.
+
+**Ideas:**
+- [ ] **Assumption surfacing** - Make implicit assumptions explicit:
+  - "I'm assuming you want Python 3.11+. Correct?"
+  - "This solution assumes you have sudo access..."
+  - Let user correct before going down wrong path
+
+- [ ] **Checkpoint & confirm** - For high-stakes operations:
+  - "About to delete 47 files. Here's the list - proceed?"
+  - "This will modify your database. Want a backup first?"
+  - Configurable threshold for when to ask
+
+**Files to modify:** `run_agent.py`, system prompt configuration
+
+---
+
+## 6. Project-Local Context 💾
+
+**Problem:** Valuable context lost between sessions.
+
+**Ideas:**
+- [ ] **Project awareness** - Remember project-specific context:
+  - Store `.hermes/context.md` in project directory
+  - "This is a Django project using PostgreSQL"
+  - Coding style preferences, deployment setup, etc.
+  - Load automatically when working in that directory
+
+- [ ] **Handoff notes** - Leave notes for future sessions:
+  - Write to `.hermes/notes.md` in project
+  - "TODO for next session: finish implementing X"
+  - "Known issues: Y doesn't work on Windows"
+
+**Files to modify:** New `project_context.py`, auto-load in `run_agent.py`
+
+## 6. Tools & Skills Wishlist 🧰
+
+*Things that would need new tool implementations (can't do well with current tools):*
+
+### High-Impact
+
+- [ ] **Audio/Video Transcription** 🎬 *(See also: Section 16 for detailed spec)*
+  - Transcribe audio files, podcasts, YouTube videos
+  - Extract key moments from video
+  - Voice memo transcription for messaging integrations
+  - *Provider options: Whisper API, Deepgram, local Whisper*
+  
+- [ ] **Diagram Rendering** 📊
+  - Render Mermaid/PlantUML to actual images
+  - Can generate the code, but rendering requires external service or tool
+  - "Show me how these components connect" → actual visual diagram
+
+### Medium-Impact
+
+- [ ] **Canvas / Visual Workspace** 🖼️
+  - Agent-controlled visual panel for rendering interactive UI
+  - Inspired by OpenClaw's Canvas feature
+  - **Capabilities:**
+    - `present` / `hide` - Show/hide the canvas panel
+    - `navigate` - Load HTML files or URLs into the canvas
+    - `eval` - Execute JavaScript in the canvas context
+    - `snapshot` - Capture the rendered UI as an image
+  - **Use cases:**
+    - Display generated HTML/CSS/JS previews
+    - Show interactive data visualizations (charts, graphs)
+    - Render diagrams (Mermaid → rendered output)
+    - Present structured information in rich format
+    - A2UI-style component system for structured agent UI
+  - **Implementation options:**
+    - Electron-based panel for CLI
+    - WebSocket-connected web app
+    - VS Code webview extension
+  - *Would let agent "show" things rather than just describe them*
+
+- [ ] **Document Generation** 📄
+  - Create styled PDFs, Word docs, presentations
+  - *Can do basic PDF via terminal tools, but limited*
+
+- [ ] **Diff/Patch Tool** 📝
+  - Surgical code modifications with preview
+  - "Change line 45-50 to X" without rewriting whole file
+  - Show diffs before applying
+  - *Can use `diff`/`patch` but a native tool would be safer*
+
+### Skills to Create
+
+- [ ] **Domain-specific skill packs:**
+  - DevOps/Infrastructure (Terraform, K8s, AWS)
+  - Data Science workflows (EDA, model training)
+  - Security/pentesting procedures
+  
+- [ ] **Framework-specific skills:**
+  - React/Vue/Angular patterns
+  - Django/Rails/Express conventions
+  - Database optimization playbooks
+
+- [ ] **Troubleshooting flowcharts:**
+  - "Docker container won't start" → decision tree
+  - "Production is slow" → systematic diagnosis
+
+---
+
+## 7. Messaging Platform Integrations 💬 ✅ COMPLETE
+
+**Problem:** Agent currently only works via `cli.py` which requires direct terminal access. Users may want to interact via messaging apps from their phone or other devices.
+
+**Architecture:**
+- `run_agent.py` already accepts `conversation_history` parameter and returns updated messages ✅
+- Need: persistent session storage, platform monitors, session key resolution
+
+**Implementation approach:**
+```
+┌─────────────────────────────────────────────────────────────┐
+│  Platform Monitor (e.g., telegram_monitor.py)               │
+│  ├─ Long-running daemon connecting to messaging platform    │
+│  ├─ On message: resolve session key → load history from disk│
+│  ├─ Call run_agent.py with loaded history                   │
+│  ├─ Save updated history back to disk (JSONL)               │
+│  └─ Send response back to platform                          │
+└─────────────────────────────────────────────────────────────┘
+```
+
+**Platform support (each user sets up their own credentials):**
+- [x] **Telegram** - via `python-telegram-bot`
+  - Bot token from @BotFather
+  - Easiest to set up, good for personal use
+- [x] **Discord** - via `discord.py`
+  - Bot token from Discord Developer Portal
+  - Can work in servers (group sessions) or DMs
+- [x] **WhatsApp** - via Node.js bridge (whatsapp-web.js/baileys)
+  - Requires Node.js bridge setup
+  - More complex, but reaches most people
+
+**Session management:**
+- [x] **Session store** - JSONL persistence per session key
+  - `~/.hermes/sessions/{session_id}.jsonl`
+  - Session keys: `agent:main:telegram:dm`, `agent:main:discord:group:123`, etc.
+- [x] **Session expiry** - Configurable reset policies
+  - Daily reset (default 4am) OR idle timeout (default 2 hours)
+  - Manual reset via `/reset` or `/new` command in chat
+  - Per-platform and per-type overrides
+- [x] **Session continuity** - Conversations persist across messages until reset
+
+**Files created:** `gateway/`, `gateway/platforms/`, `gateway/config.py`, `gateway/session.py`, `gateway/delivery.py`, `gateway/run.py`
+
+**Configuration:**
+- Environment variables: `TELEGRAM_BOT_TOKEN`, `DISCORD_BOT_TOKEN`, etc.
+- Config file: `~/.hermes/gateway.json`
+- CLI commands: `/platforms` to check status, `--gateway` to start
+
+**Dynamic context injection:**
+- Agent knows its source platform and chat
+- Agent knows connected platforms and home channels
+- Agent can deliver cron outputs to specific platforms
+
+---
+
+## 8. Text-to-Speech (TTS) 🔊
+
+**Problem:** Agent can only respond with text. Some users prefer audio responses (accessibility, hands-free use, podcasts).
+
+**Ideas:**
+- [ ] **TTS tool** - Generate audio files from text
+  ```python
+  tts_generate(text="Here's your summary...", voice="nova", output="summary.mp3")
+  ```
+  - Returns path to generated audio file
+  - For messaging integrations: can send as voice message
+  
+- [ ] **Provider options:**
+  - Edge TTS (free, good quality, many voices)
+  - OpenAI TTS (paid, excellent quality)
+  - ElevenLabs (paid, best quality, voice cloning)
+  - Local options (Coqui TTS, Bark)
+  
+- [ ] **Modes:**
+  - On-demand: User explicitly asks "read this to me"
+  - Auto-TTS: Configurable to always generate audio for responses
+  - Long-text handling: Summarize or chunk very long responses
+  
+- [ ] **Integration with messaging:**
+  - When enabled, can send voice notes instead of/alongside text
+  - User preference per channel
+
+**Files to create:** `tools/tts_tool.py`, config in `cli-config.yaml`
+
+---
+
+## 13. Speech-to-Text / Audio Transcription 🎤
+
+**Problem:** Users may want to send voice memos instead of typing. Agent is blind to audio content.
+
+**Ideas:**
+- [ ] **Voice memo transcription** - For messaging integrations
+  - User sends voice message → transcribe → process as text
+  - Seamless: user speaks, agent responds
+  
+- [ ] **Audio/video file transcription** - Existing idea, expanded:
+  - Transcribe local audio files (mp3, wav, m4a)
+  - Transcribe YouTube videos (download audio → transcribe)
+  - Extract key moments with timestamps
+  
+- [ ] **Provider options:**
+  - OpenAI Whisper API (good quality, cheap)
+  - Deepgram (fast, good for real-time)
+  - Local Whisper (free, runs on GPU)
+  - Groq Whisper (fast, free tier available)
+  
+- [ ] **Tool interface:**
+  ```python
+  transcribe(source="audio.mp3")  # Local file
+  transcribe(source="https://youtube.com/...")  # YouTube
+  transcribe(source="voice_message", data=bytes)  # Voice memo
+  ```
+
+**Files to create:** `tools/transcribe_tool.py`, integrate with messaging monitors
+
+### Plugin/Extension System 🔌
+
+**Concept:** Allow users to add custom tools/skills without modifying core code.
+
+**Why interesting:**
+- Community contributions
+- Organization-specific tools
+- Clean separation of core vs. extensions
+
+**Open questions:**
+- Security implications of loading arbitrary code
+- Versioning and compatibility
+- Discovery and installation UX
+
+---
+
+## Recently Completed ✅
+
+### Dangerous Command Approval System
+**Implemented:** Dangerous command detection and approval for terminal tool.
+
+**Features:**
+- Pattern-based detection of dangerous commands (rm -rf, DROP TABLE, chmod 777, etc.)
+- CLI prompt with options: `[o]nce | [s]ession | [a]lways | [d]eny`
+- Session caching (approved patterns don't re-prompt)
+- Permanent allowlist in `~/.hermes/config.yaml`
+- Force flag for agent to bypass after user confirmation
+- Skip check for isolated backends (Docker, Singularity, Modal)
+- Helpful sudo failure messages for messaging platforms
+
+**Files:** `tools/terminal_tool.py`, `model_tools.py`, `hermes_cli/config.py`
+
+---
+
+## 14. Learning Machine / Dynamic Memory System 🧠
+
+*Inspired by [Dash](~/agent-codebases/dash) - a self-learning data agent.*
+
+**Problem:** Agent starts fresh every session. Valuable learnings from debugging, error patterns, successful approaches, and user preferences are lost.
+
+**Dash's Key Insight:** Separate **Knowledge** (static, curated) from **Learnings** (dynamic, discovered):
+
+| System | What It Stores | How It Evolves |
+|--------|---------------|----------------|
+| **Knowledge** (Skills) | Validated approaches, templates, best practices | Curated by user |
+| **Learnings** | Error patterns, gotchas, discovered fixes | Managed automatically |
+
+**Tools to implement:**
+- [ ] `save_learning(topic, learning, context?)` - Record a discovered pattern
+  ```python
+  save_learning(
+    topic="python-ssl",
+    learning="On Ubuntu 22.04, SSL certificate errors often fixed by: apt install ca-certificates",
+    context="Debugging requests SSL failure"
+  )
+  ```
+- [ ] `search_learnings(query)` - Find relevant past learnings
+  ```python
+  search_learnings("SSL certificate error Python")
+  # Returns: "On Ubuntu 22.04, SSL certificate errors often fixed by..."
+  ```
+
+**User Profile & Memory:**
+- [ ] `user_profile` - Structured facts about user preferences
+  ```yaml
+  # ~/.hermes/user_profile.yaml
+  coding_style:
+    python_formatter: black
+    type_hints: always
+    test_framework: pytest
+  preferences:
+    verbosity: detailed
+    confirm_destructive: true
+  environment:
+    os: linux
+    shell: bash
+    default_python: 3.11
+  ```
+- [ ] `user_memory` - Unstructured observations the agent learns
+  ```yaml
+  # ~/.hermes/user_memory.yaml
+  - "User prefers tabs over spaces despite black's defaults"
+  - "User's main project is ~/work/myapp - a Django app"
+  - "User often works late - don't ask about timezone"
+  ```
+
+**When to learn:**
+- After fixing an error that took multiple attempts
+- When user corrects the agent's approach
+- When a workaround is discovered for a tool limitation
+- When user expresses a preference
+
+**Storage:** Vector database (ChromaDB) or simple YAML with embedding search.
+
+**Files to create:** `tools/learning_tools.py`, `learning/store.py`, `~/.hermes/learnings/`
+
+---
+
+## 15. Layered Context Architecture 📊
+
+*Inspired by Dash's "Six Layers of Context" - grounding responses in multiple sources.*
+
+**Problem:** Context sources are ad-hoc. No clear hierarchy or strategy for what context to include when.
+
+**Proposed Layers for Hermes:**
+
+| Layer | Source | When Loaded | Example |
+|-------|--------|-------------|---------|
+| 1. **Project Context** | `.hermes/context.md` | Auto on cwd | "This is a FastAPI project using PostgreSQL" |
+| 2. **Skills** | `skills/*.md` | On request | "How to set up React project" |
+| 3. **User Profile** | `~/.hermes/user_profile.yaml` | Always | "User prefers pytest, uses black" |
+| 4. **Learnings** | `~/.hermes/learnings/` | Semantic search | "SSL fix for Ubuntu" |
+| 5. **External Knowledge** | Web search, docs | On demand | Current API docs, Stack Overflow |
+| 6. **Runtime Introspection** | Tool calls | Real-time | File contents, terminal output |
+
+**Benefits:**
+- Clear mental model for what context is available
+- Prioritization: local > learned > external
+- Debugging: "Why did agent do X?" → check which layers contributed
+
+**Files to modify:** `run_agent.py` (context loading), new `context/layers.py`
+
+---
+
+## 16. Evaluation System with LLM Grading 📏
+
+*Inspired by Dash's evaluation framework.*
+
+**Problem:** `batch_runner.py` runs test cases but lacks quality assessment.
+
+**Dash's Approach:**
+- **String matching** (default) - Check if expected strings appear
+- **LLM grader** (-g flag) - GPT evaluates response quality
+- **Result comparison** (-r flag) - Compare against golden output
+
+**Implementation for Hermes:**
+
+- [ ] **Test case format:**
+  ```python
+  TestCase(
+    name="create_python_project",
+    prompt="Create a new Python project with FastAPI and tests",
+    expected_strings=["requirements.txt", "main.py", "test_"],  # Basic check
+    golden_actions=["write:main.py", "write:requirements.txt", "terminal:pip install"],
+    grader_criteria="Should create complete project structure with working code"
+  )
+  ```
+
+- [ ] **LLM grader mode:**
+  ```python
+  def grade_response(response: str, criteria: str) -> Grade:
+      """Use GPT to evaluate response quality."""
+      prompt = f"""
+      Evaluate this agent response against the criteria.
+      Criteria: {criteria}
+      Response: {response}
+      
+      Score (1-5) and explain why.
+      """
+      # Returns: Grade(score=4, explanation="Created all files but tests are minimal")
+  ```
+
+- [ ] **Action comparison mode:**
+  - Record tool calls made during test
+  - Compare against expected actions
+  - "Expected terminal call to pip install, got npm install"
+
+- [ ] **CLI flags:**
+  ```bash
+  python batch_runner.py eval test_cases.yaml       # String matching
+  python batch_runner.py eval test_cases.yaml -g    # + LLM grading
+  python batch_runner.py eval test_cases.yaml -r    # + Result comparison
+  python batch_runner.py eval test_cases.yaml -v    # Verbose (show responses)
+  ```
+
+**Files to modify:** `batch_runner.py`, new `evals/test_cases.py`, new `evals/grader.py`
+
+---
+
+*Last updated: $(date +%Y-%m-%d)* 🤖

acp_adapter/__init__.py

@@ -1 +0,0 @@
-"""ACP (Agent Communication Protocol) adapter for hermes-agent."""

acp_adapter/__main__.py

@@ -1,5 +0,0 @@
-"""Allow running the ACP adapter as ``python -m acp_adapter``."""
-
-from .entry import main
-
-main()

acp_adapter/auth.py

@@ -1,24 +0,0 @@
-"""ACP auth helpers — detect the currently configured Hermes provider."""
-
-from __future__ import annotations
-
-from typing import Optional
-
-
-def detect_provider() -> Optional[str]:
-    """Resolve the active Hermes runtime provider, or None if unavailable."""
-    try:
-        from hermes_cli.runtime_provider import resolve_runtime_provider
-        runtime = resolve_runtime_provider()
-        api_key = runtime.get("api_key")
-        provider = runtime.get("provider")
-        if isinstance(api_key, str) and api_key.strip() and isinstance(provider, str) and provider.strip():
-            return provider.strip().lower()
-    except Exception:
-        return None
-    return None
-
-
-def has_provider() -> bool:
-    """Return True if Hermes can resolve any runtime provider credentials."""
-    return detect_provider() is not None

acp_adapter/entry.py

@@ -1,86 +0,0 @@
-"""CLI entry point for the hermes-agent ACP adapter.
-
-Loads environment variables from ``~/.hermes/.env``, configures logging
-to write to stderr (so stdout is reserved for ACP JSON-RPC transport),
-and starts the ACP agent server.
-
-Usage::
-
-    python -m acp_adapter.entry
-    # or
-    hermes acp
-    # or
-    hermes-acp
-"""
-
-import asyncio
-import logging
-import os
-import sys
-from pathlib import Path
-from hermes_constants import get_hermes_home
-
-
-def _setup_logging() -> None:
-    """Route all logging to stderr so stdout stays clean for ACP stdio."""
-    handler = logging.StreamHandler(sys.stderr)
-    handler.setFormatter(
-        logging.Formatter(
-            "%(asctime)s [%(levelname)s] %(name)s: %(message)s",
-            datefmt="%Y-%m-%d %H:%M:%S",
-        )
-    )
-    root = logging.getLogger()
-    root.handlers.clear()
-    root.addHandler(handler)
-    root.setLevel(logging.INFO)
-
-    # Quiet down noisy libraries
-    logging.getLogger("httpx").setLevel(logging.WARNING)
-    logging.getLogger("httpcore").setLevel(logging.WARNING)
-    logging.getLogger("openai").setLevel(logging.WARNING)
-
-
-def _load_env() -> None:
-    """Load .env from HERMES_HOME (default ``~/.hermes``)."""
-    from hermes_cli.env_loader import load_hermes_dotenv
-
-    hermes_home = get_hermes_home()
-    loaded = load_hermes_dotenv(hermes_home=hermes_home)
-    if loaded:
-        for env_file in loaded:
-            logging.getLogger(__name__).info("Loaded env from %s", env_file)
-    else:
-        logging.getLogger(__name__).info(
-            "No .env found at %s, using system env", hermes_home / ".env"
-        )
-
-
-def main() -> None:
-    """Entry point: load env, configure logging, run the ACP agent."""
-    _setup_logging()
-    _load_env()
-
-    logger = logging.getLogger(__name__)
-    logger.info("Starting hermes-agent ACP adapter")
-
-    # Ensure the project root is on sys.path so ``from run_agent import AIAgent`` works
-    project_root = str(Path(__file__).resolve().parent.parent)
-    if project_root not in sys.path:
-        sys.path.insert(0, project_root)
-
-    import acp
-    from .server import HermesACPAgent
-
-    agent = HermesACPAgent()
-    try:
-        asyncio.run(acp.run_agent(agent, use_unstable_protocol=True))
-    except KeyboardInterrupt:
-        logger.info("Shutting down (KeyboardInterrupt)")
-    except Exception:
-        logger.exception("ACP agent crashed")
-        sys.exit(1)
-
-
-if __name__ == "__main__":
-    main()

acp_adapter/events.py

@@ -1,171 +0,0 @@
-"""Callback factories for bridging AIAgent events to ACP notifications.
-
-Each factory returns a callable with the signature that AIAgent expects
-for its callbacks. Internally, the callbacks push ACP session updates
-to the client via ``conn.session_update()`` using
-``asyncio.run_coroutine_threadsafe()`` (since AIAgent runs in a worker
-thread while the event loop lives on the main thread).
-"""
-
-import asyncio
-import json
-import logging
-from collections import deque
-from typing import Any, Callable, Deque, Dict
-
-import acp
-
-from .tools import (
-    build_tool_complete,
-    build_tool_start,
-    make_tool_call_id,
-)
-
-logger = logging.getLogger(__name__)
-
-
-def _send_update(
-    conn: acp.Client,
-    session_id: str,
-    loop: asyncio.AbstractEventLoop,
-    update: Any,
-) -> None:
-    """Fire-and-forget an ACP session update from a worker thread."""
-    try:
-        future = asyncio.run_coroutine_threadsafe(
-            conn.session_update(session_id, update), loop
-        )
-        future.result(timeout=5)
-    except Exception:
-        logger.debug("Failed to send ACP update", exc_info=True)
-
-
-# ------------------------------------------------------------------
-# Tool progress callback
-# ------------------------------------------------------------------
-
-def make_tool_progress_cb(
-    conn: acp.Client,
-    session_id: str,
-    loop: asyncio.AbstractEventLoop,
-    tool_call_ids: Dict[str, Deque[str]],
-) -> Callable:
-    """Create a ``tool_progress_callback`` for AIAgent.
-
-    Signature expected by AIAgent::
-
-        tool_progress_callback(name: str, preview: str, args: dict)
-
-    Emits ``ToolCallStart`` for each tool invocation and tracks IDs in a FIFO
-    queue per tool name so duplicate/parallel same-name calls still complete
-    against the correct ACP tool call.
-    """
-
-    def _tool_progress(name: str, preview: str, args: Any = None) -> None:
-        if isinstance(args, str):
-            try:
-                args = json.loads(args)
-            except (json.JSONDecodeError, TypeError):
-                args = {"raw": args}
-        if not isinstance(args, dict):
-            args = {}
-
-        tc_id = make_tool_call_id()
-        queue = tool_call_ids.get(name)
-        if queue is None:
-            queue = deque()
-            tool_call_ids[name] = queue
-        elif isinstance(queue, str):
-            queue = deque([queue])
-            tool_call_ids[name] = queue
-        queue.append(tc_id)
-
-        update = build_tool_start(tc_id, name, args)
-        _send_update(conn, session_id, loop, update)
-
-    return _tool_progress
-
-
-# ------------------------------------------------------------------
-# Thinking callback
-# ------------------------------------------------------------------
-
-def make_thinking_cb(
-    conn: acp.Client,
-    session_id: str,
-    loop: asyncio.AbstractEventLoop,
-) -> Callable:
-    """Create a ``thinking_callback`` for AIAgent."""
-
-    def _thinking(text: str) -> None:
-        if not text:
-            return
-        update = acp.update_agent_thought_text(text)
-        _send_update(conn, session_id, loop, update)
-
-    return _thinking
-
-
-# ------------------------------------------------------------------
-# Step callback
-# ------------------------------------------------------------------
-
-def make_step_cb(
-    conn: acp.Client,
-    session_id: str,
-    loop: asyncio.AbstractEventLoop,
-    tool_call_ids: Dict[str, Deque[str]],
-) -> Callable:
-    """Create a ``step_callback`` for AIAgent.
-
-    Signature expected by AIAgent::
-
-        step_callback(api_call_count: int, prev_tools: list)
-    """
-
-    def _step(api_call_count: int, prev_tools: Any = None) -> None:
-        if prev_tools and isinstance(prev_tools, list):
-            for tool_info in prev_tools:
-                tool_name = None
-                result = None
-
-                if isinstance(tool_info, dict):
-                    tool_name = tool_info.get("name") or tool_info.get("function_name")
-                    result = tool_info.get("result") or tool_info.get("output")
-                elif isinstance(tool_info, str):
-                    tool_name = tool_info
-
-                queue = tool_call_ids.get(tool_name or "")
-                if isinstance(queue, str):
-                    queue = deque([queue])
-                    tool_call_ids[tool_name] = queue
-                if tool_name and queue:
-                    tc_id = queue.popleft()
-                    update = build_tool_complete(
-                        tc_id, tool_name, result=str(result) if result is not None else None
-                    )
-                    _send_update(conn, session_id, loop, update)
-                    if not queue:
-                        tool_call_ids.pop(tool_name, None)
-
-    return _step
-
-
-# ------------------------------------------------------------------
-# Agent message callback
-# ------------------------------------------------------------------
-
-def make_message_cb(
-    conn: acp.Client,
-    session_id: str,
-    loop: asyncio.AbstractEventLoop,
-) -> Callable:
-    """Create a callback that streams agent response text to the editor."""
-
-    def _message(text: str) -> None:
-        if not text:
-            return
-        update = acp.update_agent_message_text(text)
-        _send_update(conn, session_id, loop, update)
-
-    return _message

acp_adapter/permissions.py

@@ -1,77 +0,0 @@
-"""ACP permission bridging — maps ACP approval requests to hermes approval callbacks."""
-
-from __future__ import annotations
-
-import asyncio
-import logging
-from concurrent.futures import TimeoutError as FutureTimeout
-from typing import Callable
-
-from acp.schema import (
-    AllowedOutcome,
-    PermissionOption,
-)
-
-logger = logging.getLogger(__name__)
-
-# Maps ACP PermissionOptionKind -> hermes approval result strings
-_KIND_TO_HERMES = {
-    "allow_once": "once",
-    "allow_always": "always",
-    "reject_once": "deny",
-    "reject_always": "deny",
-}
-
-
-def make_approval_callback(
-    request_permission_fn: Callable,
-    loop: asyncio.AbstractEventLoop,
-    session_id: str,
-    timeout: float = 60.0,
-) -> Callable[[str, str], str]:
-    """
-    Return a hermes-compatible ``approval_callback(command, description) -> str``
-    that bridges to the ACP client's ``request_permission`` call.
-
-    Args:
-        request_permission_fn: The ACP connection's ``request_permission`` coroutine.
-        loop: The event loop on which the ACP connection lives.
-        session_id: Current ACP session id.
-        timeout: Seconds to wait for a response before auto-denying.
-    """
-
-    def _callback(command: str, description: str) -> str:
-        options = [
-            PermissionOption(option_id="allow_once", kind="allow_once", name="Allow once"),
-            PermissionOption(option_id="allow_always", kind="allow_always", name="Allow always"),
-            PermissionOption(option_id="deny", kind="reject_once", name="Deny"),
-        ]
-        import acp as _acp
-
-        tool_call = _acp.start_tool_call("perm-check", command, kind="execute")
-
-        coro = request_permission_fn(
-            session_id=session_id,
-            tool_call=tool_call,
-            options=options,
-        )
-
-        try:
-            future = asyncio.run_coroutine_threadsafe(coro, loop)
-            response = future.result(timeout=timeout)
-        except (FutureTimeout, Exception) as exc:
-            logger.warning("Permission request timed out or failed: %s", exc)
-            return "deny"
-
-        outcome = response.outcome
-        if isinstance(outcome, AllowedOutcome):
-            option_id = outcome.option_id
-            # Look up the kind from our options list
-            for opt in options:
-                if opt.option_id == option_id:
-                    return _KIND_TO_HERMES.get(opt.kind, "deny")
-            return "once"  # fallback for unknown option_id
-        else:
-            return "deny"
-
-    return _callback

acp_adapter/server.py

@@ -1,608 +0,0 @@
-"""ACP agent server — exposes Hermes Agent via the Agent Client Protocol."""
-
-from __future__ import annotations
-
-import asyncio
-import logging
-from collections import defaultdict, deque
-from concurrent.futures import ThreadPoolExecutor
-from typing import Any, Deque, Optional
-
-import acp
-from acp.schema import (
-    AgentCapabilities,
-    AuthenticateResponse,
-    AuthMethod,
-    ClientCapabilities,
-    EmbeddedResourceContentBlock,
-    ForkSessionResponse,
-    ImageContentBlock,
-    AudioContentBlock,
-    Implementation,
-    InitializeResponse,
-    ListSessionsResponse,
-    LoadSessionResponse,
-    McpServerHttp,
-    McpServerSse,
-    McpServerStdio,
-    NewSessionResponse,
-    PromptResponse,
-    ResumeSessionResponse,
-    SetSessionConfigOptionResponse,
-    SetSessionModelResponse,
-    SetSessionModeResponse,
-    ResourceContentBlock,
-    SessionCapabilities,
-    SessionForkCapabilities,
-    SessionListCapabilities,
-    SessionInfo,
-    TextContentBlock,
-    Usage,
-)
-
-from acp_adapter.auth import detect_provider, has_provider
-from acp_adapter.events import (
-    make_message_cb,
-    make_step_cb,
-    make_thinking_cb,
-    make_tool_progress_cb,
-)
-from acp_adapter.permissions import make_approval_callback
-from acp_adapter.session import SessionManager, SessionState
-
-logger = logging.getLogger(__name__)
-
-try:
-    from hermes_cli import __version__ as HERMES_VERSION
-except Exception:
-    HERMES_VERSION = "0.0.0"
-
-# Thread pool for running AIAgent (synchronous) in parallel.
-_executor = ThreadPoolExecutor(max_workers=4, thread_name_prefix="acp-agent")
-
-
-def _extract_text(
-    prompt: list[
-        TextContentBlock
-        | ImageContentBlock
-        | AudioContentBlock
-        | ResourceContentBlock
-        | EmbeddedResourceContentBlock
-    ],
-) -> str:
-    """Extract plain text from ACP content blocks."""
-    parts: list[str] = []
-    for block in prompt:
-        if isinstance(block, TextContentBlock):
-            parts.append(block.text)
-        elif hasattr(block, "text"):
-            parts.append(str(block.text))
-        # Non-text blocks are ignored for now.
-    return "\n".join(parts)
-
-
-class HermesACPAgent(acp.Agent):
-    """ACP Agent implementation wrapping Hermes AIAgent."""
-
-    def __init__(self, session_manager: SessionManager | None = None):
-        super().__init__()
-        self.session_manager = session_manager or SessionManager()
-        self._conn: Optional[acp.Client] = None
-
-    # ---- Connection lifecycle -----------------------------------------------
-
-    def on_connect(self, conn: acp.Client) -> None:
-        """Store the client connection for sending session updates."""
-        self._conn = conn
-        logger.info("ACP client connected")
-
-    async def _register_session_mcp_servers(
-        self,
-        state: SessionState,
-        mcp_servers: list[McpServerStdio | McpServerHttp | McpServerSse] | None,
-    ) -> None:
-        """Register ACP-provided MCP servers and refresh the agent tool surface."""
-        if not mcp_servers:
-            return
-
-        try:
-            from tools.mcp_tool import register_mcp_servers
-
-            config_map: dict[str, dict] = {}
-            for server in mcp_servers:
-                name = server.name
-                if isinstance(server, McpServerStdio):
-                    config = {
-                        "command": server.command,
-                        "args": list(server.args),
-                        "env": {item.name: item.value for item in server.env},
-                    }
-                else:
-                    config = {
-                        "url": server.url,
-                        "headers": {item.name: item.value for item in server.headers},
-                    }
-                config_map[name] = config
-
-            await asyncio.to_thread(register_mcp_servers, config_map)
-        except Exception:
-            logger.warning(
-                "Session %s: failed to register ACP MCP servers",
-                state.session_id,
-                exc_info=True,
-            )
-            return
-
-        try:
-            from model_tools import get_tool_definitions
-
-            enabled_toolsets = getattr(state.agent, "enabled_toolsets", None) or ["hermes-acp"]
-            disabled_toolsets = getattr(state.agent, "disabled_toolsets", None)
-            state.agent.tools = get_tool_definitions(
-                enabled_toolsets=enabled_toolsets,
-                disabled_toolsets=disabled_toolsets,
-                quiet_mode=True,
-            )
-            state.agent.valid_tool_names = {
-                tool["function"]["name"] for tool in state.agent.tools or []
-            }
-            invalidate = getattr(state.agent, "_invalidate_system_prompt", None)
-            if callable(invalidate):
-                invalidate()
-            logger.info(
-                "Session %s: refreshed tool surface after ACP MCP registration (%d tools)",
-                state.session_id,
-                len(state.agent.tools or []),
-            )
-        except Exception:
-            logger.warning(
-                "Session %s: failed to refresh tool surface after ACP MCP registration",
-                state.session_id,
-                exc_info=True,
-            )
-
-    # ---- ACP lifecycle ------------------------------------------------------
-
-    async def initialize(
-        self,
-        protocol_version: int | None = None,
-        client_capabilities: ClientCapabilities | None = None,
-        client_info: Implementation | None = None,
-        **kwargs: Any,
-    ) -> InitializeResponse:
-        resolved_protocol_version = (
-            protocol_version if isinstance(protocol_version, int) else acp.PROTOCOL_VERSION
-        )
-        provider = detect_provider()
-        auth_methods = None
-        if provider:
-            auth_methods = [
-                AuthMethod(
-                    id=provider,
-                    name=f"{provider} runtime credentials",
-                    description=f"Authenticate Hermes using the currently configured {provider} runtime credentials.",
-                )
-            ]
-
-        client_name = client_info.name if client_info else "unknown"
-        logger.info(
-            "Initialize from %s (protocol v%s)",
-            client_name,
-            resolved_protocol_version,
-        )
-
-        return InitializeResponse(
-            protocol_version=acp.PROTOCOL_VERSION,
-            agent_info=Implementation(name="hermes-agent", version=HERMES_VERSION),
-            agent_capabilities=AgentCapabilities(
-                session_capabilities=SessionCapabilities(
-                    fork=SessionForkCapabilities(),
-                    list=SessionListCapabilities(),
-                ),
-            ),
-            auth_methods=auth_methods,
-        )
-
-    async def authenticate(self, method_id: str, **kwargs: Any) -> AuthenticateResponse | None:
-        if has_provider():
-            return AuthenticateResponse()
-        return None
-
-    # ---- Session management -------------------------------------------------
-
-    async def new_session(
-        self,
-        cwd: str,
-        mcp_servers: list | None = None,
-        **kwargs: Any,
-    ) -> NewSessionResponse:
-        state = self.session_manager.create_session(cwd=cwd)
-        await self._register_session_mcp_servers(state, mcp_servers)
-        logger.info("New session %s (cwd=%s)", state.session_id, cwd)
-        return NewSessionResponse(session_id=state.session_id)
-
-    async def load_session(
-        self,
-        cwd: str,
-        session_id: str,
-        mcp_servers: list | None = None,
-        **kwargs: Any,
-    ) -> LoadSessionResponse | None:
-        state = self.session_manager.update_cwd(session_id, cwd)
-        if state is None:
-            logger.warning("load_session: session %s not found", session_id)
-            return None
-        await self._register_session_mcp_servers(state, mcp_servers)
-        logger.info("Loaded session %s", session_id)
-        return LoadSessionResponse()
-
-    async def resume_session(
-        self,
-        cwd: str,
-        session_id: str,
-        mcp_servers: list | None = None,
-        **kwargs: Any,
-    ) -> ResumeSessionResponse:
-        state = self.session_manager.update_cwd(session_id, cwd)
-        if state is None:
-            logger.warning("resume_session: session %s not found, creating new", session_id)
-            state = self.session_manager.create_session(cwd=cwd)
-        await self._register_session_mcp_servers(state, mcp_servers)
-        logger.info("Resumed session %s", state.session_id)
-        return ResumeSessionResponse()
-
-    async def cancel(self, session_id: str, **kwargs: Any) -> None:
-        state = self.session_manager.get_session(session_id)
-        if state and state.cancel_event:
-            state.cancel_event.set()
-            try:
-                if getattr(state, "agent", None) and hasattr(state.agent, "interrupt"):
-                    state.agent.interrupt()
-            except Exception:
-                logger.debug("Failed to interrupt ACP session %s", session_id, exc_info=True)
-            logger.info("Cancelled session %s", session_id)
-
-    async def fork_session(
-        self,
-        cwd: str,
-        session_id: str,
-        mcp_servers: list | None = None,
-        **kwargs: Any,
-    ) -> ForkSessionResponse:
-        state = self.session_manager.fork_session(session_id, cwd=cwd)
-        new_id = state.session_id if state else ""
-        if state is not None:
-            await self._register_session_mcp_servers(state, mcp_servers)
-        logger.info("Forked session %s -> %s", session_id, new_id)
-        return ForkSessionResponse(session_id=new_id)
-
-    async def list_sessions(
-        self,
-        cursor: str | None = None,
-        cwd: str | None = None,
-        **kwargs: Any,
-    ) -> ListSessionsResponse:
-        infos = self.session_manager.list_sessions()
-        sessions = [
-            SessionInfo(session_id=s["session_id"], cwd=s["cwd"])
-            for s in infos
-        ]
-        return ListSessionsResponse(sessions=sessions)
-
-    # ---- Prompt (core) ------------------------------------------------------
-
-    async def prompt(
-        self,
-        prompt: list[
-            TextContentBlock
-            | ImageContentBlock
-            | AudioContentBlock
-            | ResourceContentBlock
-            | EmbeddedResourceContentBlock
-        ],
-        session_id: str,
-        **kwargs: Any,
-    ) -> PromptResponse:
-        """Run Hermes on the user's prompt and stream events back to the editor."""
-        state = self.session_manager.get_session(session_id)
-        if state is None:
-            logger.error("prompt: session %s not found", session_id)
-            return PromptResponse(stop_reason="refusal")
-
-        user_text = _extract_text(prompt).strip()
-        if not user_text:
-            return PromptResponse(stop_reason="end_turn")
-
-        # Intercept slash commands — handle locally without calling the LLM
-        if user_text.startswith("/"):
-            response_text = self._handle_slash_command(user_text, state)
-            if response_text is not None:
-                if self._conn:
-                    update = acp.update_agent_message_text(response_text)
-                    await self._conn.session_update(session_id, update)
-                return PromptResponse(stop_reason="end_turn")
-
-        logger.info("Prompt on session %s: %s", session_id, user_text[:100])
-
-        conn = self._conn
-        loop = asyncio.get_running_loop()
-
-        if state.cancel_event:
-            state.cancel_event.clear()
-
-        tool_call_ids: dict[str, Deque[str]] = defaultdict(deque)
-        previous_approval_cb = None
-
-        if conn:
-            tool_progress_cb = make_tool_progress_cb(conn, session_id, loop, tool_call_ids)
-            thinking_cb = make_thinking_cb(conn, session_id, loop)
-            step_cb = make_step_cb(conn, session_id, loop, tool_call_ids)
-            message_cb = make_message_cb(conn, session_id, loop)
-            approval_cb = make_approval_callback(conn.request_permission, loop, session_id)
-        else:
-            tool_progress_cb = None
-            thinking_cb = None
-            step_cb = None
-            message_cb = None
-            approval_cb = None
-
-        agent = state.agent
-        agent.tool_progress_callback = tool_progress_cb
-        agent.thinking_callback = thinking_cb
-        agent.step_callback = step_cb
-        agent.message_callback = message_cb
-
-        if approval_cb:
-            try:
-                from tools import terminal_tool as _terminal_tool
-                previous_approval_cb = getattr(_terminal_tool, "_approval_callback", None)
-                _terminal_tool.set_approval_callback(approval_cb)
-            except Exception:
-                logger.debug("Could not set ACP approval callback", exc_info=True)
-
-        def _run_agent() -> dict:
-            try:
-                result = agent.run_conversation(
-                    user_message=user_text,
-                    conversation_history=state.history,
-                    task_id=session_id,
-                )
-                return result
-            except Exception as e:
-                logger.exception("Agent error in session %s", session_id)
-                return {"final_response": f"Error: {e}", "messages": state.history}
-            finally:
-                if approval_cb:
-                    try:
-                        from tools import terminal_tool as _terminal_tool
-                        _terminal_tool.set_approval_callback(previous_approval_cb)
-                    except Exception:
-                        logger.debug("Could not restore approval callback", exc_info=True)
-
-        try:
-            result = await loop.run_in_executor(_executor, _run_agent)
-        except Exception:
-            logger.exception("Executor error for session %s", session_id)
-            return PromptResponse(stop_reason="end_turn")
-
-        if result.get("messages"):
-            state.history = result["messages"]
-            # Persist updated history so sessions survive process restarts.
-            self.session_manager.save_session(session_id)
-
-        final_response = result.get("final_response", "")
-        if final_response and conn:
-            update = acp.update_agent_message_text(final_response)
-            await conn.session_update(session_id, update)
-
-        usage = None
-        usage_data = result.get("usage")
-        if usage_data and isinstance(usage_data, dict):
-            usage = Usage(
-                input_tokens=usage_data.get("prompt_tokens", 0),
-                output_tokens=usage_data.get("completion_tokens", 0),
-                total_tokens=usage_data.get("total_tokens", 0),
-                thought_tokens=usage_data.get("reasoning_tokens"),
-                cached_read_tokens=usage_data.get("cached_tokens"),
-            )
-
-        stop_reason = "cancelled" if state.cancel_event and state.cancel_event.is_set() else "end_turn"
-        return PromptResponse(stop_reason=stop_reason, usage=usage)
-
-    # ---- Slash commands (headless) -------------------------------------------
-
-    _SLASH_COMMANDS = {
-        "help": "Show available commands",
-        "model": "Show or change current model",
-        "tools": "List available tools",
-        "context": "Show conversation context info",
-        "reset": "Clear conversation history",
-        "compact": "Compress conversation context",
-        "version": "Show Hermes version",
-    }
-
-    def _handle_slash_command(self, text: str, state: SessionState) -> str | None:
-        """Dispatch a slash command and return the response text.
-
-        Returns ``None`` for unrecognized commands so they fall through
-        to the LLM (the user may have typed ``/something`` as prose).
-        """
-        parts = text.split(maxsplit=1)
-        cmd = parts[0].lstrip("/").lower()
-        args = parts[1].strip() if len(parts) > 1 else ""
-
-        handler = {
-            "help": self._cmd_help,
-            "model": self._cmd_model,
-            "tools": self._cmd_tools,
-            "context": self._cmd_context,
-            "reset": self._cmd_reset,
-            "compact": self._cmd_compact,
-            "version": self._cmd_version,
-        }.get(cmd)
-
-        if handler is None:
-            return None  # not a known command — let the LLM handle it
-
-        try:
-            return handler(args, state)
-        except Exception as e:
-            logger.error("Slash command /%s error: %s", cmd, e, exc_info=True)
-            return f"Error executing /{cmd}: {e}"
-
-    def _cmd_help(self, args: str, state: SessionState) -> str:
-        lines = ["Available commands:", ""]
-        for cmd, desc in self._SLASH_COMMANDS.items():
-            lines.append(f"  /{cmd:10s}  {desc}")
-        lines.append("")
-        lines.append("Unrecognized /commands are sent to the model as normal messages.")
-        return "\n".join(lines)
-
-    def _cmd_model(self, args: str, state: SessionState) -> str:
-        if not args:
-            model = state.model or getattr(state.agent, "model", "unknown")
-            provider = getattr(state.agent, "provider", None) or "auto"
-            return f"Current model: {model}\nProvider: {provider}"
-
-        new_model = args.strip()
-        target_provider = None
-        current_provider = getattr(state.agent, "provider", None) or "openrouter"
-
-        # Auto-detect provider for the requested model
-        try:
-            from hermes_cli.models import parse_model_input, detect_provider_for_model
-            target_provider, new_model = parse_model_input(new_model, current_provider)
-            if target_provider == current_provider:
-                detected = detect_provider_for_model(new_model, current_provider)
-                if detected:
-                    target_provider, new_model = detected
-        except Exception:
-            logger.debug("Provider detection failed, using model as-is", exc_info=True)
-
-        state.model = new_model
-        state.agent = self.session_manager._make_agent(
-            session_id=state.session_id,
-            cwd=state.cwd,
-            model=new_model,
-            requested_provider=target_provider or current_provider,
-        )
-        self.session_manager.save_session(state.session_id)
-        provider_label = getattr(state.agent, "provider", None) or target_provider or current_provider
-        logger.info("Session %s: model switched to %s", state.session_id, new_model)
-        return f"Model switched to: {new_model}\nProvider: {provider_label}"
-
-    def _cmd_tools(self, args: str, state: SessionState) -> str:
-        try:
-            from model_tools import get_tool_definitions
-            toolsets = getattr(state.agent, "enabled_toolsets", None) or ["hermes-acp"]
-            tools = get_tool_definitions(enabled_toolsets=toolsets, quiet_mode=True)
-            if not tools:
-                return "No tools available."
-            lines = [f"Available tools ({len(tools)}):"]
-            for t in tools:
-                name = t.get("function", {}).get("name", "?")
-                desc = t.get("function", {}).get("description", "")
-                # Truncate long descriptions
-                if len(desc) > 80:
-                    desc = desc[:77] + "..."
-                lines.append(f"  {name}: {desc}")
-            return "\n".join(lines)
-        except Exception as e:
-            return f"Could not list tools: {e}"
-
-    def _cmd_context(self, args: str, state: SessionState) -> str:
-        n_messages = len(state.history)
-        if n_messages == 0:
-            return "Conversation is empty (no messages yet)."
-        # Count by role
-        roles: dict[str, int] = {}
-        for msg in state.history:
-            role = msg.get("role", "unknown")
-            roles[role] = roles.get(role, 0) + 1
-        lines = [
-            f"Conversation: {n_messages} messages",
-            f"  user: {roles.get('user', 0)}, assistant: {roles.get('assistant', 0)}, "
-            f"tool: {roles.get('tool', 0)}, system: {roles.get('system', 0)}",
-        ]
-        model = state.model or getattr(state.agent, "model", "")
-        if model:
-            lines.append(f"Model: {model}")
-        return "\n".join(lines)
-
-    def _cmd_reset(self, args: str, state: SessionState) -> str:
-        state.history.clear()
-        self.session_manager.save_session(state.session_id)
-        return "Conversation history cleared."
-
-    def _cmd_compact(self, args: str, state: SessionState) -> str:
-        if not state.history:
-            return "Nothing to compress — conversation is empty."
-        try:
-            agent = state.agent
-            if hasattr(agent, "compress_context"):
-                agent.compress_context(state.history)
-                self.session_manager.save_session(state.session_id)
-                return f"Context compressed. Messages: {len(state.history)}"
-            return "Context compression not available for this agent."
-        except Exception as e:
-            return f"Compression failed: {e}"
-
-    def _cmd_version(self, args: str, state: SessionState) -> str:
-        return f"Hermes Agent v{HERMES_VERSION}"
-
-    # ---- Model switching (ACP protocol method) -------------------------------
-
-    async def set_session_model(
-        self, model_id: str, session_id: str, **kwargs: Any
-    ) -> SetSessionModelResponse | None:
-        """Switch the model for a session (called by ACP protocol)."""
-        state = self.session_manager.get_session(session_id)
-        if state:
-            state.model = model_id
-            current_provider = getattr(state.agent, "provider", None)
-            current_base_url = getattr(state.agent, "base_url", None)
-            current_api_mode = getattr(state.agent, "api_mode", None)
-            state.agent = self.session_manager._make_agent(
-                session_id=session_id,
-                cwd=state.cwd,
-                model=model_id,
-                requested_provider=current_provider,
-                base_url=current_base_url,
-                api_mode=current_api_mode,
-            )
-            self.session_manager.save_session(session_id)
-            logger.info("Session %s: model switched to %s", session_id, model_id)
-            return SetSessionModelResponse()
-        logger.warning("Session %s: model switch requested for missing session", session_id)
-        return None
-
-    async def set_session_mode(
-        self, mode_id: str, session_id: str, **kwargs: Any
-    ) -> SetSessionModeResponse | None:
-        """Persist the editor-requested mode so ACP clients do not fail on mode switches."""
-        state = self.session_manager.get_session(session_id)
-        if state is None:
-            logger.warning("Session %s: mode switch requested for missing session", session_id)
-            return None
-        setattr(state, "mode", mode_id)
-        self.session_manager.save_session(session_id)
-        logger.info("Session %s: mode switched to %s", session_id, mode_id)
-        return SetSessionModeResponse()
-
-    async def set_config_option(
-        self, config_id: str, session_id: str, value: str, **kwargs: Any
-    ) -> SetSessionConfigOptionResponse | None:
-        """Accept ACP config option updates even when Hermes has no typed ACP config surface yet."""
-        state = self.session_manager.get_session(session_id)
-        if state is None:
-            logger.warning("Session %s: config update requested for missing session", session_id)
-            return None
-
-        options = getattr(state, "config_options", None)
-        if not isinstance(options, dict):
-            options = {}
-        options[str(config_id)] = value
-        setattr(state, "config_options", options)
-        self.session_manager.save_session(session_id)
-        logger.info("Session %s: config option %s updated", session_id, config_id)
-        return SetSessionConfigOptionResponse(config_options=[])

acp_adapter/session.py

@@ -1,461 +0,0 @@
-"""ACP session manager — maps ACP sessions to Hermes AIAgent instances.
-
-Sessions are persisted to the shared SessionDB (``~/.hermes/state.db``) so they
-survive process restarts and appear in ``session_search``.  When the editor
-reconnects after idle/restart, the ``load_session`` / ``resume_session`` calls
-find the persisted session in the database and restore the full conversation
-history.
-"""
-from __future__ import annotations
-
-from hermes_constants import get_hermes_home
-
-import copy
-import json
-import logging
-import uuid
-from dataclasses import dataclass, field
-from threading import Lock
-from typing import Any, Dict, List, Optional
-
-logger = logging.getLogger(__name__)
-
-
-def _register_task_cwd(task_id: str, cwd: str) -> None:
-    """Bind a task/session id to the editor's working directory for tools."""
-    if not task_id:
-        return
-    try:
-        from tools.terminal_tool import register_task_env_overrides
-        register_task_env_overrides(task_id, {"cwd": cwd})
-    except Exception:
-        logger.debug("Failed to register ACP task cwd override", exc_info=True)
-
-
-def _clear_task_cwd(task_id: str) -> None:
-    """Remove task-specific cwd overrides for an ACP session."""
-    if not task_id:
-        return
-    try:
-        from tools.terminal_tool import clear_task_env_overrides
-        clear_task_env_overrides(task_id)
-    except Exception:
-        logger.debug("Failed to clear ACP task cwd override", exc_info=True)
-
-
-@dataclass
-class SessionState:
-    """Tracks per-session state for an ACP-managed Hermes agent."""
-
-    session_id: str
-    agent: Any  # AIAgent instance
-    cwd: str = "."
-    model: str = ""
-    history: List[Dict[str, Any]] = field(default_factory=list)
-    cancel_event: Any = None  # threading.Event
-
-
-class SessionManager:
-    """Thread-safe manager for ACP sessions backed by Hermes AIAgent instances.
-
-    Sessions are held in-memory for fast access **and** persisted to the
-    shared SessionDB so they survive process restarts and are searchable
-    via ``session_search``.
-    """
-
-    def __init__(self, agent_factory=None, db=None):
-        """
-        Args:
-            agent_factory: Optional callable that creates an AIAgent-like object.
-                           Used by tests. When omitted, a real AIAgent is created
-                           using the current Hermes runtime provider configuration.
-            db:            Optional SessionDB instance. When omitted, the default
-                           SessionDB (``~/.hermes/state.db``) is lazily created.
-        """
-        self._sessions: Dict[str, SessionState] = {}
-        self._lock = Lock()
-        self._agent_factory = agent_factory
-        self._db_instance = db  # None → lazy-init on first use
-
-    # ---- public API ---------------------------------------------------------
-
-    def create_session(self, cwd: str = ".") -> SessionState:
-        """Create a new session with a unique ID and a fresh AIAgent."""
-        import threading
-
-        session_id = str(uuid.uuid4())
-        agent = self._make_agent(session_id=session_id, cwd=cwd)
-        state = SessionState(
-            session_id=session_id,
-            agent=agent,
-            cwd=cwd,
-            model=getattr(agent, "model", "") or "",
-            cancel_event=threading.Event(),
-        )
-        with self._lock:
-            self._sessions[session_id] = state
-        _register_task_cwd(session_id, cwd)
-        self._persist(state)
-        logger.info("Created ACP session %s (cwd=%s)", session_id, cwd)
-        return state
-
-    def get_session(self, session_id: str) -> Optional[SessionState]:
-        """Return the session for *session_id*, or ``None``.
-
-        If the session is not in memory but exists in the database (e.g. after
-        a process restart), it is transparently restored.
-        """
-        with self._lock:
-            state = self._sessions.get(session_id)
-        if state is not None:
-            return state
-        # Attempt to restore from database.
-        return self._restore(session_id)
-
-    def remove_session(self, session_id: str) -> bool:
-        """Remove a session from memory and database. Returns True if it existed."""
-        with self._lock:
-            existed = self._sessions.pop(session_id, None) is not None
-        db_existed = self._delete_persisted(session_id)
-        if existed or db_existed:
-            _clear_task_cwd(session_id)
-        return existed or db_existed
-
-    def fork_session(self, session_id: str, cwd: str = ".") -> Optional[SessionState]:
-        """Deep-copy a session's history into a new session."""
-        import threading
-
-        original = self.get_session(session_id)  # checks DB too
-        if original is None:
-            return None
-
-        new_id = str(uuid.uuid4())
-        agent = self._make_agent(
-            session_id=new_id,
-            cwd=cwd,
-            model=original.model or None,
-        )
-        state = SessionState(
-            session_id=new_id,
-            agent=agent,
-            cwd=cwd,
-            model=getattr(agent, "model", original.model) or original.model,
-            history=copy.deepcopy(original.history),
-            cancel_event=threading.Event(),
-        )
-        with self._lock:
-            self._sessions[new_id] = state
-        _register_task_cwd(new_id, cwd)
-        self._persist(state)
-        logger.info("Forked ACP session %s -> %s", session_id, new_id)
-        return state
-
-    def list_sessions(self) -> List[Dict[str, Any]]:
-        """Return lightweight info dicts for all sessions (memory + database)."""
-        # Collect in-memory sessions first.
-        with self._lock:
-            seen_ids = set(self._sessions.keys())
-            results = [
-                {
-                    "session_id": s.session_id,
-                    "cwd": s.cwd,
-                    "model": s.model,
-                    "history_len": len(s.history),
-                }
-                for s in self._sessions.values()
-            ]
-
-        # Merge any persisted sessions not currently in memory.
-        db = self._get_db()
-        if db is not None:
-            try:
-                rows = db.search_sessions(source="acp", limit=1000)
-                for row in rows:
-                    sid = row["id"]
-                    if sid in seen_ids:
-                        continue
-                    # Extract cwd from model_config JSON.
-                    cwd = "."
-                    mc = row.get("model_config")
-                    if mc:
-                        try:
-                            cwd = json.loads(mc).get("cwd", ".")
-                        except (json.JSONDecodeError, TypeError):
-                            pass
-                    results.append({
-                        "session_id": sid,
-                        "cwd": cwd,
-                        "model": row.get("model") or "",
-                        "history_len": row.get("message_count") or 0,
-                    })
-            except Exception:
-                logger.debug("Failed to list ACP sessions from DB", exc_info=True)
-
-        return results
-
-    def update_cwd(self, session_id: str, cwd: str) -> Optional[SessionState]:
-        """Update the working directory for a session and its tool overrides."""
-        state = self.get_session(session_id)  # checks DB too
-        if state is None:
-            return None
-        state.cwd = cwd
-        _register_task_cwd(session_id, cwd)
-        self._persist(state)
-        return state
-
-    def cleanup(self) -> None:
-        """Remove all sessions (memory and database) and clear task-specific cwd overrides."""
-        with self._lock:
-            session_ids = list(self._sessions.keys())
-            self._sessions.clear()
-        for session_id in session_ids:
-            _clear_task_cwd(session_id)
-            self._delete_persisted(session_id)
-        # Also remove any DB-only ACP sessions not currently in memory.
-        db = self._get_db()
-        if db is not None:
-            try:
-                rows = db.search_sessions(source="acp", limit=10000)
-                for row in rows:
-                    sid = row["id"]
-                    _clear_task_cwd(sid)
-                    db.delete_session(sid)
-            except Exception:
-                logger.debug("Failed to cleanup ACP sessions from DB", exc_info=True)
-
-    def save_session(self, session_id: str) -> None:
-        """Persist the current state of a session to the database.
-
-        Called by the server after prompt completion, slash commands that
-        mutate history, and model switches.
-        """
-        with self._lock:
-            state = self._sessions.get(session_id)
-        if state is not None:
-            self._persist(state)
-
-    # ---- persistence via SessionDB ------------------------------------------
-
-    def _get_db(self):
-        """Lazily initialise and return the SessionDB instance.
-
-        Returns ``None`` if the DB is unavailable (e.g. import error in a
-        minimal test environment).
-
-        Note: we resolve ``HERMES_HOME`` dynamically rather than relying on
-        the module-level ``DEFAULT_DB_PATH`` constant, because that constant
-        is evaluated at import time and won't reflect env-var changes made
-        later (e.g. by the test fixture ``_isolate_hermes_home``).
-        """
-        if self._db_instance is not None:
-            return self._db_instance
-        try:
-            import os
-            from pathlib import Path
-            from hermes_state import SessionDB
-            hermes_home = get_hermes_home()
-            self._db_instance = SessionDB(db_path=hermes_home / "state.db")
-            return self._db_instance
-        except Exception:
-            logger.debug("SessionDB unavailable for ACP persistence", exc_info=True)
-            return None
-
-    def _persist(self, state: SessionState) -> None:
-        """Write session state to the database.
-
-        Creates the session record if it doesn't exist, then replaces all
-        stored messages with the current in-memory history.
-        """
-        db = self._get_db()
-        if db is None:
-            return
-
-        # Ensure model is a plain string (not a MagicMock or other proxy).
-        model_str = str(state.model) if state.model else None
-        session_meta = {"cwd": state.cwd}
-        provider = getattr(state.agent, "provider", None)
-        base_url = getattr(state.agent, "base_url", None)
-        api_mode = getattr(state.agent, "api_mode", None)
-        if isinstance(provider, str) and provider.strip():
-            session_meta["provider"] = provider.strip()
-        if isinstance(base_url, str) and base_url.strip():
-            session_meta["base_url"] = base_url.strip()
-        if isinstance(api_mode, str) and api_mode.strip():
-            session_meta["api_mode"] = api_mode.strip()
-        cwd_json = json.dumps(session_meta)
-
-        try:
-            # Ensure the session record exists.
-            existing = db.get_session(state.session_id)
-            if existing is None:
-                db.create_session(
-                    session_id=state.session_id,
-                    source="acp",
-                    model=model_str,
-                    model_config={"cwd": state.cwd},
-                )
-            else:
-                # Update model_config (contains cwd) if changed.
-                try:
-                    with db._lock:
-                        db._conn.execute(
-                            "UPDATE sessions SET model_config = ?, model = COALESCE(?, model) WHERE id = ?",
-                            (cwd_json, model_str, state.session_id),
-                        )
-                        db._conn.commit()
-                except Exception:
-                    logger.debug("Failed to update ACP session metadata", exc_info=True)
-
-            # Replace stored messages with current history.
-            db.clear_messages(state.session_id)
-            for msg in state.history:
-                db.append_message(
-                    session_id=state.session_id,
-                    role=msg.get("role", "user"),
-                    content=msg.get("content"),
-                    tool_name=msg.get("tool_name") or msg.get("name"),
-                    tool_calls=msg.get("tool_calls"),
-                    tool_call_id=msg.get("tool_call_id"),
-                )
-        except Exception:
-            logger.warning("Failed to persist ACP session %s", state.session_id, exc_info=True)
-
-    def _restore(self, session_id: str) -> Optional[SessionState]:
-        """Load a session from the database into memory, recreating the AIAgent."""
-        import threading
-
-        db = self._get_db()
-        if db is None:
-            return None
-
-        try:
-            row = db.get_session(session_id)
-        except Exception:
-            logger.debug("Failed to query DB for ACP session %s", session_id, exc_info=True)
-            return None
-
-        if row is None:
-            return None
-
-        # Only restore ACP sessions.
-        if row.get("source") != "acp":
-            return None
-
-        # Extract cwd from model_config.
-        cwd = "."
-        requested_provider = row.get("billing_provider")
-        restored_base_url = row.get("billing_base_url")
-        restored_api_mode = None
-        mc = row.get("model_config")
-        if mc:
-            try:
-                meta = json.loads(mc)
-                if isinstance(meta, dict):
-                    cwd = meta.get("cwd", ".")
-                    requested_provider = meta.get("provider") or requested_provider
-                    restored_base_url = meta.get("base_url") or restored_base_url
-                    restored_api_mode = meta.get("api_mode") or restored_api_mode
-            except (json.JSONDecodeError, TypeError):
-                pass
-
-        model = row.get("model") or None
-
-        # Load conversation history.
-        try:
-            history = db.get_messages_as_conversation(session_id)
-        except Exception:
-            logger.warning("Failed to load messages for ACP session %s", session_id, exc_info=True)
-            history = []
-
-        try:
-            agent = self._make_agent(
-                session_id=session_id,
-                cwd=cwd,
-                model=model,
-                requested_provider=requested_provider,
-                base_url=restored_base_url,
-                api_mode=restored_api_mode,
-            )
-        except Exception:
-            logger.warning("Failed to recreate agent for ACP session %s", session_id, exc_info=True)
-            return None
-
-        state = SessionState(
-            session_id=session_id,
-            agent=agent,
-            cwd=cwd,
-            model=model or getattr(agent, "model", "") or "",
-            history=history,
-            cancel_event=threading.Event(),
-        )
-        with self._lock:
-            self._sessions[session_id] = state
-        _register_task_cwd(session_id, cwd)
-        logger.info("Restored ACP session %s from DB (%d messages)", session_id, len(history))
-        return state
-
-    def _delete_persisted(self, session_id: str) -> bool:
-        """Delete a session from the database. Returns True if it existed."""
-        db = self._get_db()
-        if db is None:
-            return False
-        try:
-            return db.delete_session(session_id)
-        except Exception:
-            logger.debug("Failed to delete ACP session %s from DB", session_id, exc_info=True)
-            return False
-
-    # ---- internal -----------------------------------------------------------
-
-    def _make_agent(
-        self,
-        *,
-        session_id: str,
-        cwd: str,
-        model: str | None = None,
-        requested_provider: str | None = None,
-        base_url: str | None = None,
-        api_mode: str | None = None,
-    ):
-        if self._agent_factory is not None:
-            return self._agent_factory()
-
-        from run_agent import AIAgent
-        from hermes_cli.config import load_config
-        from hermes_cli.runtime_provider import resolve_runtime_provider
-
-        config = load_config()
-        model_cfg = config.get("model")
-        default_model = ""
-        config_provider = None
-        if isinstance(model_cfg, dict):
-            default_model = str(model_cfg.get("default") or default_model)
-            config_provider = model_cfg.get("provider")
-        elif isinstance(model_cfg, str) and model_cfg.strip():
-            default_model = model_cfg.strip()
-
-        kwargs = {
-            "platform": "acp",
-            "enabled_toolsets": ["hermes-acp"],
-            "quiet_mode": True,
-            "session_id": session_id,
-            "model": model or default_model,
-        }
-
-        try:
-            runtime = resolve_runtime_provider(requested=requested_provider or config_provider)
-            kwargs.update(
-                {
-                    "provider": runtime.get("provider"),
-                    "api_mode": api_mode or runtime.get("api_mode"),
-                    "base_url": base_url or runtime.get("base_url"),
-                    "api_key": runtime.get("api_key"),
-                    "command": runtime.get("command"),
-                    "args": list(runtime.get("args") or []),
-                }
-            )
-        except Exception:
-            logger.debug("ACP session falling back to default provider resolution", exc_info=True)
-
-        _register_task_cwd(session_id, cwd)
-        return AIAgent(**kwargs)

acp_adapter/tools.py

@@ -1,215 +0,0 @@
-"""ACP tool-call helpers for mapping hermes tools to ACP ToolKind and building content."""
-
-from __future__ import annotations
-
-import uuid
-from typing import Any, Dict, List, Optional
-
-import acp
-from acp.schema import (
-    ToolCallLocation,
-    ToolCallStart,
-    ToolCallProgress,
-    ToolKind,
-)
-
-# ---------------------------------------------------------------------------
-# Map hermes tool names -> ACP ToolKind
-# ---------------------------------------------------------------------------
-
-TOOL_KIND_MAP: Dict[str, ToolKind] = {
-    # File operations
-    "read_file": "read",
-    "write_file": "edit",
-    "patch": "edit",
-    "search_files": "search",
-    # Terminal / execution
-    "terminal": "execute",
-    "process": "execute",
-    "execute_code": "execute",
-    # Web / fetch
-    "web_search": "fetch",
-    "web_extract": "fetch",
-    # Browser
-    "browser_navigate": "fetch",
-    "browser_click": "execute",
-    "browser_type": "execute",
-    "browser_snapshot": "read",
-    "browser_vision": "read",
-    "browser_scroll": "execute",
-    "browser_press": "execute",
-    "browser_back": "execute",
-    "browser_close": "execute",
-    "browser_get_images": "read",
-    # Agent internals
-    "delegate_task": "execute",
-    "vision_analyze": "read",
-    "image_generate": "execute",
-    "text_to_speech": "execute",
-    # Thinking / meta
-    "_thinking": "think",
-}
-
-
-def get_tool_kind(tool_name: str) -> ToolKind:
-    """Return the ACP ToolKind for a hermes tool, defaulting to 'other'."""
-    return TOOL_KIND_MAP.get(tool_name, "other")
-
-
-def make_tool_call_id() -> str:
-    """Generate a unique tool call ID."""
-    return f"tc-{uuid.uuid4().hex[:12]}"
-
-
-def build_tool_title(tool_name: str, args: Dict[str, Any]) -> str:
-    """Build a human-readable title for a tool call."""
-    if tool_name == "terminal":
-        cmd = args.get("command", "")
-        if len(cmd) > 80:
-            cmd = cmd[:77] + "..."
-        return f"terminal: {cmd}"
-    if tool_name == "read_file":
-        return f"read: {args.get('path', '?')}"
-    if tool_name == "write_file":
-        return f"write: {args.get('path', '?')}"
-    if tool_name == "patch":
-        mode = args.get("mode", "replace")
-        path = args.get("path", "?")
-        return f"patch ({mode}): {path}"
-    if tool_name == "search_files":
-        return f"search: {args.get('pattern', '?')}"
-    if tool_name == "web_search":
-        return f"web search: {args.get('query', '?')}"
-    if tool_name == "web_extract":
-        urls = args.get("urls", [])
-        if urls:
-            return f"extract: {urls[0]}" + (f" (+{len(urls)-1})" if len(urls) > 1 else "")
-        return "web extract"
-    if tool_name == "delegate_task":
-        goal = args.get("goal", "")
-        if goal and len(goal) > 60:
-            goal = goal[:57] + "..."
-        return f"delegate: {goal}" if goal else "delegate task"
-    if tool_name == "execute_code":
-        return "execute code"
-    if tool_name == "vision_analyze":
-        return f"analyze image: {args.get('question', '?')[:50]}"
-    return tool_name
-
-
-# ---------------------------------------------------------------------------
-# Build ACP content objects for tool-call events
-# ---------------------------------------------------------------------------
-
-
-def build_tool_start(
-    tool_call_id: str,
-    tool_name: str,
-    arguments: Dict[str, Any],
-) -> ToolCallStart:
-    """Create a ToolCallStart event for the given hermes tool invocation."""
-    kind = get_tool_kind(tool_name)
-    title = build_tool_title(tool_name, arguments)
-    locations = extract_locations(arguments)
-
-    if tool_name == "patch":
-        mode = arguments.get("mode", "replace")
-        if mode == "replace":
-            path = arguments.get("path", "")
-            old = arguments.get("old_string", "")
-            new = arguments.get("new_string", "")
-            content = [acp.tool_diff_content(path=path, new_text=new, old_text=old)]
-        else:
-            # Patch mode — show the patch content as text
-            patch_text = arguments.get("patch", "")
-            content = [acp.tool_content(acp.text_block(patch_text))]
-        return acp.start_tool_call(
-            tool_call_id, title, kind=kind, content=content, locations=locations,
-            raw_input=arguments,
-        )
-
-    if tool_name == "write_file":
-        path = arguments.get("path", "")
-        file_content = arguments.get("content", "")
-        content = [acp.tool_diff_content(path=path, new_text=file_content)]
-        return acp.start_tool_call(
-            tool_call_id, title, kind=kind, content=content, locations=locations,
-            raw_input=arguments,
-        )
-
-    if tool_name == "terminal":
-        command = arguments.get("command", "")
-        content = [acp.tool_content(acp.text_block(f"$ {command}"))]
-        return acp.start_tool_call(
-            tool_call_id, title, kind=kind, content=content, locations=locations,
-            raw_input=arguments,
-        )
-
-    if tool_name == "read_file":
-        path = arguments.get("path", "")
-        content = [acp.tool_content(acp.text_block(f"Reading {path}"))]
-        return acp.start_tool_call(
-            tool_call_id, title, kind=kind, content=content, locations=locations,
-            raw_input=arguments,
-        )
-
-    if tool_name == "search_files":
-        pattern = arguments.get("pattern", "")
-        target = arguments.get("target", "content")
-        content = [acp.tool_content(acp.text_block(f"Searching for '{pattern}' ({target})"))]
-        return acp.start_tool_call(
-            tool_call_id, title, kind=kind, content=content, locations=locations,
-            raw_input=arguments,
-        )
-
-    # Generic fallback
-    import json
-    try:
-        args_text = json.dumps(arguments, indent=2, default=str)
-    except (TypeError, ValueError):
-        args_text = str(arguments)
-    content = [acp.tool_content(acp.text_block(args_text))]
-    return acp.start_tool_call(
-        tool_call_id, title, kind=kind, content=content, locations=locations,
-        raw_input=arguments,
-    )
-
-
-def build_tool_complete(
-    tool_call_id: str,
-    tool_name: str,
-    result: Optional[str] = None,
-) -> ToolCallProgress:
-    """Create a ToolCallUpdate (progress) event for a completed tool call."""
-    kind = get_tool_kind(tool_name)
-
-    # Truncate very large results for the UI
-    display_result = result or ""
-    if len(display_result) > 5000:
-        display_result = display_result[:4900] + f"\n... ({len(result)} chars total, truncated)"
-
-    content = [acp.tool_content(acp.text_block(display_result))]
-    return acp.update_tool_call(
-        tool_call_id,
-        kind=kind,
-        status="completed",
-        content=content,
-        raw_output=result,
-    )
-
-
-# ---------------------------------------------------------------------------
-# Location extraction
-# ---------------------------------------------------------------------------
-
-
-def extract_locations(
-    arguments: Dict[str, Any],
-) -> List[ToolCallLocation]:
-    """Extract file-system locations from tool arguments."""
-    locations: List[ToolCallLocation] = []
-    path = arguments.get("path")
-    if path:
-        line = arguments.get("offset") or arguments.get("line")
-        locations.append(ToolCallLocation(path=path, line=line))
-    return locations

acp_registry/agent.json

@@ -1,12 +0,0 @@
-{
-  "schema_version": 1,
-  "name": "hermes-agent",
-  "display_name": "Hermes Agent",
-  "description": "AI agent by Nous Research with 90+ tools, persistent memory, and multi-platform support",
-  "icon": "icon.svg",
-  "distribution": {
-    "type": "command",
-    "command": "hermes",
-    "args": ["acp"]
-  }
-}

acp_registry/icon.svg

@@ -1,25 +0,0 @@
-<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 64 64" width="64" height="64">
-  <defs>
-    <linearGradient id="gold" x1="0%" y1="0%" x2="0%" y2="100%">
-      <stop offset="0%" style="stop-color:#F5C542;stop-opacity:1" />
-      <stop offset="100%" style="stop-color:#D4961C;stop-opacity:1" />
-    </linearGradient>
-  </defs>
-  <!-- Staff -->
-  <rect x="30" y="10" width="4" height="46" rx="2" fill="url(#gold)" />
-  <!-- Wings (left) -->
-  <path d="M30 18 C24 14, 14 14, 10 18 C14 16, 22 16, 28 20" fill="#F5C542" opacity="0.9" />
-  <path d="M30 22 C26 19, 18 19, 14 22 C18 20, 24 20, 28 24" fill="#D4961C" opacity="0.8" />
-  <!-- Wings (right) -->
-  <path d="M34 18 C40 14, 50 14, 54 18 C50 16, 42 16, 36 20" fill="#F5C542" opacity="0.9" />
-  <path d="M34 22 C38 19, 46 19, 50 22 C46 20, 40 20, 36 24" fill="#D4961C" opacity="0.8" />
-  <!-- Left serpent -->
-  <path d="M32 48 C22 44, 20 38, 26 34 C20 36, 18 42, 24 46 C18 40, 22 30, 30 28 C24 32, 22 38, 28 42"
-        fill="none" stroke="#F5C542" stroke-width="2.5" stroke-linecap="round" />
-  <!-- Right serpent -->
-  <path d="M32 48 C42 44, 44 38, 38 34 C44 36, 46 42, 40 46 C46 40, 42 30, 34 28 C40 32, 42 38, 36 42"
-        fill="none" stroke="#D4961C" stroke-width="2.5" stroke-linecap="round" />
-  <!-- Orb at top -->
-  <circle cx="32" cy="10" r="4" fill="#F5C542" />
-  <circle cx="32" cy="10" r="2" fill="#FFF8E1" opacity="0.7" />
-</svg>

agent/__init__.py

@@ -1,6 +0,0 @@
-"""Agent internals -- extracted modules from run_agent.py.
-
-These modules contain pure utility functions and self-contained classes
-that were previously embedded in the 3,600-line run_agent.py. Extracting
-them makes run_agent.py focused on the AIAgent orchestrator class.
-"""

agent/builtin_memory_provider.py

@@ -1,113 +0,0 @@
-"""BuiltinMemoryProvider — wraps MEMORY.md / USER.md as a MemoryProvider.
-
-Always registered as the first provider. Cannot be disabled or removed.
-This is the existing Hermes memory system exposed through the provider
-interface for compatibility with the MemoryManager.
-
-The actual storage logic lives in tools/memory_tool.py (MemoryStore).
-This provider is a thin adapter that delegates to MemoryStore and
-exposes the memory tool schema.
-"""
-
-from __future__ import annotations
-
-import json
-import logging
-from typing import Any, Dict, List, Optional
-
-from agent.memory_provider import MemoryProvider
-
-logger = logging.getLogger(__name__)
-
-
-class BuiltinMemoryProvider(MemoryProvider):
-    """Built-in file-backed memory (MEMORY.md + USER.md).
-
-    Always active, never disabled by other providers. The `memory` tool
-    is handled by run_agent.py's agent-level tool interception (not through
-    the normal registry), so get_tool_schemas() returns an empty list —
-    the memory tool is already wired separately.
-    """
-
-    def __init__(
-        self,
-        memory_store=None,
-        memory_enabled: bool = False,
-        user_profile_enabled: bool = False,
-    ):
-        self._store = memory_store
-        self._memory_enabled = memory_enabled
-        self._user_profile_enabled = user_profile_enabled
-
-    @property
-    def name(self) -> str:
-        return "builtin"
-
-    def is_available(self) -> bool:
-        """Built-in memory is always available."""
-        return True
-
-    def initialize(self, session_id: str, **kwargs) -> None:
-        """Load memory from disk if not already loaded."""
-        if self._store is not None:
-            self._store.load_from_disk()
-
-    def system_prompt_block(self) -> str:
-        """Return MEMORY.md and USER.md content for the system prompt.
-
-        Uses the frozen snapshot captured at load time. This ensures the
-        system prompt stays stable throughout a session (preserving the
-        prompt cache), even though the live entries may change via tool calls.
-        """
-        if not self._store:
-            return ""
-
-        parts = []
-        if self._memory_enabled:
-            mem_block = self._store.format_for_system_prompt("memory")
-            if mem_block:
-                parts.append(mem_block)
-        if self._user_profile_enabled:
-            user_block = self._store.format_for_system_prompt("user")
-            if user_block:
-                parts.append(user_block)
-
-        return "\n\n".join(parts)
-
-    def prefetch(self, query: str, *, session_id: str = "") -> str:
-        """Built-in memory doesn't do query-based recall — it's injected via system_prompt_block."""
-        return ""
-
-    def sync_turn(self, user_content: str, assistant_content: str, *, session_id: str = "") -> None:
-        """Built-in memory doesn't auto-sync turns — writes happen via the memory tool."""
-
-    def get_tool_schemas(self) -> List[Dict[str, Any]]:
-        """Return empty list.
-
-        The `memory` tool is an agent-level intercepted tool, handled
-        specially in run_agent.py before normal tool dispatch. It's not
-        part of the standard tool registry. We don't duplicate it here.
-        """
-        return []
-
-    def handle_tool_call(self, tool_name: str, args: Dict[str, Any], **kwargs) -> str:
-        """Not used — the memory tool is intercepted in run_agent.py."""
-        return json.dumps({"error": "Built-in memory tool is handled by the agent loop"})
-
-    def shutdown(self) -> None:
-        """No cleanup needed — files are saved on every write."""
-
-    # -- Property access for backward compatibility --------------------------
-
-    @property
-    def store(self):
-        """Access the underlying MemoryStore for legacy code paths."""
-        return self._store
-
-    @property
-    def memory_enabled(self) -> bool:
-        return self._memory_enabled
-
-    @property
-    def user_profile_enabled(self) -> bool:
-        return self._user_profile_enabled

agent/context_compressor.py

@@ -1,676 +0,0 @@
-"""Automatic context window compression for long conversations.
-
-Self-contained class with its own OpenAI client for summarization.
-Uses auxiliary model (cheap/fast) to summarize middle turns while
-protecting head and tail context.
-
-Improvements over v1:
-  - Structured summary template (Goal, Progress, Decisions, Files, Next Steps)
-  - Iterative summary updates (preserves info across multiple compactions)
-  - Token-budget tail protection instead of fixed message count
-  - Tool output pruning before LLM summarization (cheap pre-pass)
-  - Scaled summary budget (proportional to compressed content)
-  - Richer tool call/result detail in summarizer input
-"""
-
-import logging
-from typing import Any, Dict, List, Optional
-
-from agent.auxiliary_client import call_llm
-from agent.model_metadata import (
-    get_model_context_length,
-    estimate_messages_tokens_rough,
-)
-
-logger = logging.getLogger(__name__)
-
-SUMMARY_PREFIX = (
-    "[CONTEXT COMPACTION] Earlier turns in this conversation were compacted "
-    "to save context space. The summary below describes work that was "
-    "already completed, and the current session state may still reflect "
-    "that work (for example, files may already be changed). Use the summary "
-    "and the current state to continue from where things left off, and "
-    "avoid repeating work:"
-)
-LEGACY_SUMMARY_PREFIX = "[CONTEXT SUMMARY]:"
-
-# Minimum tokens for the summary output
-_MIN_SUMMARY_TOKENS = 2000
-# Proportion of compressed content to allocate for summary
-_SUMMARY_RATIO = 0.20
-# Absolute ceiling for summary tokens (even on very large context windows)
-_SUMMARY_TOKENS_CEILING = 12_000
-
-# Placeholder used when pruning old tool results
-_PRUNED_TOOL_PLACEHOLDER = "[Old tool output cleared to save context space]"
-
-# Chars per token rough estimate
-_CHARS_PER_TOKEN = 4
-
-
-class ContextCompressor:
-    """Compresses conversation context when approaching the model's context limit.
-
-    Algorithm:
-      1. Prune old tool results (cheap, no LLM call)
-      2. Protect head messages (system prompt + first exchange)
-      3. Protect tail messages by token budget (most recent ~20K tokens)
-      4. Summarize middle turns with structured LLM prompt
-      5. On subsequent compactions, iteratively update the previous summary
-    """
-
-    def __init__(
-        self,
-        model: str,
-        threshold_percent: float = 0.50,
-        protect_first_n: int = 3,
-        protect_last_n: int = 20,
-        summary_target_ratio: float = 0.20,
-        quiet_mode: bool = False,
-        summary_model_override: str = None,
-        base_url: str = "",
-        api_key: str = "",
-        config_context_length: int | None = None,
-        provider: str = "",
-    ):
-        self.model = model
-        self.base_url = base_url
-        self.api_key = api_key
-        self.provider = provider
-        self.threshold_percent = threshold_percent
-        self.protect_first_n = protect_first_n
-        self.protect_last_n = protect_last_n
-        self.summary_target_ratio = max(0.10, min(summary_target_ratio, 0.80))
-        self.quiet_mode = quiet_mode
-
-        self.context_length = get_model_context_length(
-            model, base_url=base_url, api_key=api_key,
-            config_context_length=config_context_length,
-            provider=provider,
-        )
-        self.threshold_tokens = int(self.context_length * threshold_percent)
-        self.compression_count = 0
-
-        # Derive token budgets: ratio is relative to the threshold, not total context
-        target_tokens = int(self.threshold_tokens * self.summary_target_ratio)
-        self.tail_token_budget = target_tokens
-        self.max_summary_tokens = min(
-            int(self.context_length * 0.05), _SUMMARY_TOKENS_CEILING,
-        )
-
-        if not quiet_mode:
-            logger.info(
-                "Context compressor initialized: model=%s context_length=%d "
-                "threshold=%d (%.0f%%) target_ratio=%.0f%% tail_budget=%d "
-                "provider=%s base_url=%s",
-                model, self.context_length, self.threshold_tokens,
-                threshold_percent * 100, self.summary_target_ratio * 100,
-                self.tail_token_budget,
-                provider or "none", base_url or "none",
-            )
-        self._context_probed = False  # True after a step-down from context error
-
-        self.last_prompt_tokens = 0
-        self.last_completion_tokens = 0
-        self.last_total_tokens = 0
-
-        self.summary_model = summary_model_override or ""
-
-        # Stores the previous compaction summary for iterative updates
-        self._previous_summary: Optional[str] = None
-
-    def update_from_response(self, usage: Dict[str, Any]):
-        """Update tracked token usage from API response."""
-        self.last_prompt_tokens = usage.get("prompt_tokens", 0)
-        self.last_completion_tokens = usage.get("completion_tokens", 0)
-        self.last_total_tokens = usage.get("total_tokens", 0)
-
-    def should_compress(self, prompt_tokens: int = None) -> bool:
-        """Check if context exceeds the compression threshold."""
-        tokens = prompt_tokens if prompt_tokens is not None else self.last_prompt_tokens
-        return tokens >= self.threshold_tokens
-
-    def should_compress_preflight(self, messages: List[Dict[str, Any]]) -> bool:
-        """Quick pre-flight check using rough estimate (before API call)."""
-        rough_estimate = estimate_messages_tokens_rough(messages)
-        return rough_estimate >= self.threshold_tokens
-
-    def get_status(self) -> Dict[str, Any]:
-        """Get current compression status for display/logging."""
-        return {
-            "last_prompt_tokens": self.last_prompt_tokens,
-            "threshold_tokens": self.threshold_tokens,
-            "context_length": self.context_length,
-            "usage_percent": min(100, (self.last_prompt_tokens / self.context_length * 100)) if self.context_length else 0,
-            "compression_count": self.compression_count,
-        }
-
-    # ------------------------------------------------------------------
-    # Tool output pruning (cheap pre-pass, no LLM call)
-    # ------------------------------------------------------------------
-
-    def _prune_old_tool_results(
-        self, messages: List[Dict[str, Any]], protect_tail_count: int,
-    ) -> tuple[List[Dict[str, Any]], int]:
-        """Replace old tool result contents with a short placeholder.
-
-        Walks backward from the end, protecting the most recent
-        ``protect_tail_count`` messages. Older tool results get their
-        content replaced with a placeholder string.
-
-        Returns (pruned_messages, pruned_count).
-        """
-        if not messages:
-            return messages, 0
-
-        result = [m.copy() for m in messages]
-        pruned = 0
-        prune_boundary = len(result) - protect_tail_count
-
-        for i in range(prune_boundary):
-            msg = result[i]
-            if msg.get("role") != "tool":
-                continue
-            content = msg.get("content", "")
-            if not content or content == _PRUNED_TOOL_PLACEHOLDER:
-                continue
-            # Only prune if the content is substantial (>200 chars)
-            if len(content) > 200:
-                result[i] = {**msg, "content": _PRUNED_TOOL_PLACEHOLDER}
-                pruned += 1
-
-        return result, pruned
-
-    # ------------------------------------------------------------------
-    # Summarization
-    # ------------------------------------------------------------------
-
-    def _compute_summary_budget(self, turns_to_summarize: List[Dict[str, Any]]) -> int:
-        """Scale summary token budget with the amount of content being compressed.
-
-        The maximum scales with the model's context window (5% of context,
-        capped at ``_SUMMARY_TOKENS_CEILING``) so large-context models get
-        richer summaries instead of being hard-capped at 8K tokens.
-        """
-        content_tokens = estimate_messages_tokens_rough(turns_to_summarize)
-        budget = int(content_tokens * _SUMMARY_RATIO)
-        return max(_MIN_SUMMARY_TOKENS, min(budget, self.max_summary_tokens))
-
-    def _serialize_for_summary(self, turns: List[Dict[str, Any]]) -> str:
-        """Serialize conversation turns into labeled text for the summarizer.
-
-        Includes tool call arguments and result content (up to 3000 chars
-        per message) so the summarizer can preserve specific details like
-        file paths, commands, and outputs.
-        """
-        parts = []
-        for msg in turns:
-            role = msg.get("role", "unknown")
-            content = msg.get("content") or ""
-
-            # Tool results: keep more content than before (3000 chars)
-            if role == "tool":
-                tool_id = msg.get("tool_call_id", "")
-                if len(content) > 3000:
-                    content = content[:2000] + "\n...[truncated]...\n" + content[-800:]
-                parts.append(f"[TOOL RESULT {tool_id}]: {content}")
-                continue
-
-            # Assistant messages: include tool call names AND arguments
-            if role == "assistant":
-                if len(content) > 3000:
-                    content = content[:2000] + "\n...[truncated]...\n" + content[-800:]
-                tool_calls = msg.get("tool_calls", [])
-                if tool_calls:
-                    tc_parts = []
-                    for tc in tool_calls:
-                        if isinstance(tc, dict):
-                            fn = tc.get("function", {})
-                            name = fn.get("name", "?")
-                            args = fn.get("arguments", "")
-                            # Truncate long arguments but keep enough for context
-                            if len(args) > 500:
-                                args = args[:400] + "..."
-                            tc_parts.append(f"  {name}({args})")
-                        else:
-                            fn = getattr(tc, "function", None)
-                            name = getattr(fn, "name", "?") if fn else "?"
-                            tc_parts.append(f"  {name}(...)")
-                    content += "\n[Tool calls:\n" + "\n".join(tc_parts) + "\n]"
-                parts.append(f"[ASSISTANT]: {content}")
-                continue
-
-            # User and other roles
-            if len(content) > 3000:
-                content = content[:2000] + "\n...[truncated]...\n" + content[-800:]
-            parts.append(f"[{role.upper()}]: {content}")
-
-        return "\n\n".join(parts)
-
-    def _generate_summary(self, turns_to_summarize: List[Dict[str, Any]]) -> Optional[str]:
-        """Generate a structured summary of conversation turns.
-
-        Uses a structured template (Goal, Progress, Decisions, Files, Next Steps)
-        inspired by Pi-mono and OpenCode. When a previous summary exists,
-        generates an iterative update instead of summarizing from scratch.
-
-        Returns None if all attempts fail — the caller should drop
-        the middle turns without a summary rather than inject a useless
-        placeholder.
-        """
-        summary_budget = self._compute_summary_budget(turns_to_summarize)
-        content_to_summarize = self._serialize_for_summary(turns_to_summarize)
-
-        if self._previous_summary:
-            # Iterative update: preserve existing info, add new progress
-            prompt = f"""You are updating a context compaction summary. A previous compaction produced the summary below. New conversation turns have occurred since then and need to be incorporated.
-
-PREVIOUS SUMMARY:
-{self._previous_summary}
-
-NEW TURNS TO INCORPORATE:
-{content_to_summarize}
-
-Update the summary using this exact structure. PRESERVE all existing information that is still relevant. ADD new progress. Move items from "In Progress" to "Done" when completed. Remove information only if it is clearly obsolete.
-
-## Goal
-[What the user is trying to accomplish — preserve from previous summary, update if goal evolved]
-
-## Constraints & Preferences
-[User preferences, coding style, constraints, important decisions — accumulate across compactions]
-
-## Progress
-### Done
-[Completed work — include specific file paths, commands run, results obtained]
-### In Progress
-[Work currently underway]
-### Blocked
-[Any blockers or issues encountered]
-
-## Key Decisions
-[Important technical decisions and why they were made]
-
-## Relevant Files
-[Files read, modified, or created — with brief note on each. Accumulate across compactions.]
-
-## Next Steps
-[What needs to happen next to continue the work]
-
-## Critical Context
-[Any specific values, error messages, configuration details, or data that would be lost without explicit preservation]
-
-Target ~{summary_budget} tokens. Be specific — include file paths, command outputs, error messages, and concrete values rather than vague descriptions.
-
-Write only the summary body. Do not include any preamble or prefix."""
-        else:
-            # First compaction: summarize from scratch
-            prompt = f"""Create a structured handoff summary for a later assistant that will continue this conversation after earlier turns are compacted.
-
-TURNS TO SUMMARIZE:
-{content_to_summarize}
-
-Use this exact structure:
-
-## Goal
-[What the user is trying to accomplish]
-
-## Constraints & Preferences
-[User preferences, coding style, constraints, important decisions]
-
-## Progress
-### Done
-[Completed work — include specific file paths, commands run, results obtained]
-### In Progress
-[Work currently underway]
-### Blocked
-[Any blockers or issues encountered]
-
-## Key Decisions
-[Important technical decisions and why they were made]
-
-## Relevant Files
-[Files read, modified, or created — with brief note on each]
-
-## Next Steps
-[What needs to happen next to continue the work]
-
-## Critical Context
-[Any specific values, error messages, configuration details, or data that would be lost without explicit preservation]
-
-Target ~{summary_budget} tokens. Be specific — include file paths, command outputs, error messages, and concrete values rather than vague descriptions. The goal is to prevent the next assistant from repeating work or losing important details.
-
-Write only the summary body. Do not include any preamble or prefix."""
-
-        try:
-            call_kwargs = {
-                "task": "compression",
-                "messages": [{"role": "user", "content": prompt}],
-                "temperature": 0.3,
-                "max_tokens": summary_budget * 2,
-                # timeout resolved from auxiliary.compression.timeout config by call_llm
-            }
-            if self.summary_model:
-                call_kwargs["model"] = self.summary_model
-            response = call_llm(**call_kwargs)
-            content = response.choices[0].message.content
-            # Handle cases where content is not a string (e.g., dict from llama.cpp)
-            if not isinstance(content, str):
-                content = str(content) if content else ""
-            summary = content.strip()
-            # Store for iterative updates on next compaction
-            self._previous_summary = summary
-            return self._with_summary_prefix(summary)
-        except RuntimeError:
-            logging.warning("Context compression: no provider available for "
-                            "summary. Middle turns will be dropped without summary.")
-            return None
-        except Exception as e:
-            logging.warning("Failed to generate context summary: %s", e)
-            return None
-
-    @staticmethod
-    def _with_summary_prefix(summary: str) -> str:
-        """Normalize summary text to the current compaction handoff format."""
-        text = (summary or "").strip()
-        for prefix in (LEGACY_SUMMARY_PREFIX, SUMMARY_PREFIX):
-            if text.startswith(prefix):
-                text = text[len(prefix):].lstrip()
-                break
-        return f"{SUMMARY_PREFIX}\n{text}" if text else SUMMARY_PREFIX
-
-    # ------------------------------------------------------------------
-    # Tool-call / tool-result pair integrity helpers
-    # ------------------------------------------------------------------
-
-    @staticmethod
-    def _get_tool_call_id(tc) -> str:
-        """Extract the call ID from a tool_call entry (dict or SimpleNamespace)."""
-        if isinstance(tc, dict):
-            return tc.get("id", "")
-        return getattr(tc, "id", "") or ""
-
-    def _sanitize_tool_pairs(self, messages: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
-        """Fix orphaned tool_call / tool_result pairs after compression.
-
-        Two failure modes:
-        1. A tool *result* references a call_id whose assistant tool_call was
-           removed (summarized/truncated).  The API rejects this with
-           "No tool call found for function call output with call_id ...".
-        2. An assistant message has tool_calls whose results were dropped.
-           The API rejects this because every tool_call must be followed by
-           a tool result with the matching call_id.
-
-        This method removes orphaned results and inserts stub results for
-        orphaned calls so the message list is always well-formed.
-        """
-        surviving_call_ids: set = set()
-        for msg in messages:
-            if msg.get("role") == "assistant":
-                for tc in msg.get("tool_calls") or []:
-                    cid = self._get_tool_call_id(tc)
-                    if cid:
-                        surviving_call_ids.add(cid)
-
-        result_call_ids: set = set()
-        for msg in messages:
-            if msg.get("role") == "tool":
-                cid = msg.get("tool_call_id")
-                if cid:
-                    result_call_ids.add(cid)
-
-        # 1. Remove tool results whose call_id has no matching assistant tool_call
-        orphaned_results = result_call_ids - surviving_call_ids
-        if orphaned_results:
-            messages = [
-                m for m in messages
-                if not (m.get("role") == "tool" and m.get("tool_call_id") in orphaned_results)
-            ]
-            if not self.quiet_mode:
-                logger.info("Compression sanitizer: removed %d orphaned tool result(s)", len(orphaned_results))
-
-        # 2. Add stub results for assistant tool_calls whose results were dropped
-        missing_results = surviving_call_ids - result_call_ids
-        if missing_results:
-            patched: List[Dict[str, Any]] = []
-            for msg in messages:
-                patched.append(msg)
-                if msg.get("role") == "assistant":
-                    for tc in msg.get("tool_calls") or []:
-                        cid = self._get_tool_call_id(tc)
-                        if cid in missing_results:
-                            patched.append({
-                                "role": "tool",
-                                "content": "[Result from earlier conversation — see context summary above]",
-                                "tool_call_id": cid,
-                            })
-            messages = patched
-            if not self.quiet_mode:
-                logger.info("Compression sanitizer: added %d stub tool result(s)", len(missing_results))
-
-        return messages
-
-    def _align_boundary_forward(self, messages: List[Dict[str, Any]], idx: int) -> int:
-        """Push a compress-start boundary forward past any orphan tool results.
-
-        If ``messages[idx]`` is a tool result, slide forward until we hit a
-        non-tool message so we don't start the summarised region mid-group.
-        """
-        while idx < len(messages) and messages[idx].get("role") == "tool":
-            idx += 1
-        return idx
-
-    def _align_boundary_backward(self, messages: List[Dict[str, Any]], idx: int) -> int:
-        """Pull a compress-end boundary backward to avoid splitting a
-        tool_call / result group.
-
-        If the boundary falls in the middle of a tool-result group (i.e.
-        there are consecutive tool messages before ``idx``), walk backward
-        past all of them to find the parent assistant message.  If found,
-        move the boundary before the assistant so the entire
-        assistant + tool_results group is included in the summarised region
-        rather than being split (which causes silent data loss when
-        ``_sanitize_tool_pairs`` removes the orphaned tail results).
-        """
-        if idx <= 0 or idx >= len(messages):
-            return idx
-        # Walk backward past consecutive tool results
-        check = idx - 1
-        while check >= 0 and messages[check].get("role") == "tool":
-            check -= 1
-        # If we landed on the parent assistant with tool_calls, pull the
-        # boundary before it so the whole group gets summarised together.
-        if check >= 0 and messages[check].get("role") == "assistant" and messages[check].get("tool_calls"):
-            idx = check
-        return idx
-
-    # ------------------------------------------------------------------
-    # Tail protection by token budget
-    # ------------------------------------------------------------------
-
-    def _find_tail_cut_by_tokens(
-        self, messages: List[Dict[str, Any]], head_end: int,
-        token_budget: int | None = None,
-    ) -> int:
-        """Walk backward from the end of messages, accumulating tokens until
-        the budget is reached. Returns the index where the tail starts.
-
-        ``token_budget`` defaults to ``self.tail_token_budget`` which is
-        derived from ``summary_target_ratio * context_length``, so it
-        scales automatically with the model's context window.
-
-        Never cuts inside a tool_call/result group. Falls back to the old
-        ``protect_last_n`` if the budget would protect fewer messages.
-        """
-        if token_budget is None:
-            token_budget = self.tail_token_budget
-        n = len(messages)
-        min_tail = self.protect_last_n
-        accumulated = 0
-        cut_idx = n  # start from beyond the end
-
-        for i in range(n - 1, head_end - 1, -1):
-            msg = messages[i]
-            content = msg.get("content") or ""
-            msg_tokens = len(content) // _CHARS_PER_TOKEN + 10  # +10 for role/metadata
-            # Include tool call arguments in estimate
-            for tc in msg.get("tool_calls") or []:
-                if isinstance(tc, dict):
-                    args = tc.get("function", {}).get("arguments", "")
-                    msg_tokens += len(args) // _CHARS_PER_TOKEN
-            if accumulated + msg_tokens > token_budget and (n - i) >= min_tail:
-                break
-            accumulated += msg_tokens
-            cut_idx = i
-
-        # Ensure we protect at least protect_last_n messages
-        fallback_cut = n - min_tail
-        if cut_idx > fallback_cut:
-            cut_idx = fallback_cut
-
-        # If the token budget would protect everything (small conversations),
-        # fall back to the fixed protect_last_n approach so compression can
-        # still remove middle turns.
-        if cut_idx <= head_end:
-            cut_idx = fallback_cut
-
-        # Align to avoid splitting tool groups
-        cut_idx = self._align_boundary_backward(messages, cut_idx)
-
-        return max(cut_idx, head_end + 1)
-
-    # ------------------------------------------------------------------
-    # Main compression entry point
-    # ------------------------------------------------------------------
-
-    def compress(self, messages: List[Dict[str, Any]], current_tokens: int = None) -> List[Dict[str, Any]]:
-        """Compress conversation messages by summarizing middle turns.
-
-        Algorithm:
-          1. Prune old tool results (cheap pre-pass, no LLM call)
-          2. Protect head messages (system prompt + first exchange)
-          3. Find tail boundary by token budget (~20K tokens of recent context)
-          4. Summarize middle turns with structured LLM prompt
-          5. On re-compression, iteratively update the previous summary
-
-        After compression, orphaned tool_call / tool_result pairs are cleaned
-        up so the API never receives mismatched IDs.
-        """
-        n_messages = len(messages)
-        if n_messages <= self.protect_first_n + self.protect_last_n + 1:
-            if not self.quiet_mode:
-                logger.warning(
-                    "Cannot compress: only %d messages (need > %d)",
-                    n_messages,
-                    self.protect_first_n + self.protect_last_n + 1,
-                )
-            return messages
-
-        display_tokens = current_tokens if current_tokens else self.last_prompt_tokens or estimate_messages_tokens_rough(messages)
-
-        # Phase 1: Prune old tool results (cheap, no LLM call)
-        messages, pruned_count = self._prune_old_tool_results(
-            messages, protect_tail_count=self.protect_last_n * 3,
-        )
-        if pruned_count and not self.quiet_mode:
-            logger.info("Pre-compression: pruned %d old tool result(s)", pruned_count)
-
-        # Phase 2: Determine boundaries
-        compress_start = self.protect_first_n
-        compress_start = self._align_boundary_forward(messages, compress_start)
-
-        # Use token-budget tail protection instead of fixed message count
-        compress_end = self._find_tail_cut_by_tokens(messages, compress_start)
-
-        if compress_start >= compress_end:
-            return messages
-
-        turns_to_summarize = messages[compress_start:compress_end]
-
-        if not self.quiet_mode:
-            logger.info(
-                "Context compression triggered (%d tokens >= %d threshold)",
-                display_tokens,
-                self.threshold_tokens,
-            )
-            logger.info(
-                "Model context limit: %d tokens (%.0f%% = %d)",
-                self.context_length,
-                self.threshold_percent * 100,
-                self.threshold_tokens,
-            )
-            tail_msgs = n_messages - compress_end
-            logger.info(
-                "Summarizing turns %d-%d (%d turns), protecting %d head + %d tail messages",
-                compress_start + 1,
-                compress_end,
-                len(turns_to_summarize),
-                compress_start,
-                tail_msgs,
-            )
-
-        # Phase 3: Generate structured summary
-        summary = self._generate_summary(turns_to_summarize)
-
-        # Phase 4: Assemble compressed message list
-        compressed = []
-        for i in range(compress_start):
-            msg = messages[i].copy()
-            if i == 0 and msg.get("role") == "system" and self.compression_count == 0:
-                msg["content"] = (
-                    (msg.get("content") or "")
-                    + "\n\n[Note: Some earlier conversation turns have been compacted into a handoff summary to preserve context space. The current session state may still reflect earlier work, so build on that summary and state rather than re-doing work.]"
-                )
-            compressed.append(msg)
-
-        _merge_summary_into_tail = False
-        if summary:
-            last_head_role = messages[compress_start - 1].get("role", "user") if compress_start > 0 else "user"
-            first_tail_role = messages[compress_end].get("role", "user") if compress_end < n_messages else "user"
-            # Pick a role that avoids consecutive same-role with both neighbors.
-            # Priority: avoid colliding with head (already committed), then tail.
-            if last_head_role in ("assistant", "tool"):
-                summary_role = "user"
-            else:
-                summary_role = "assistant"
-            # If the chosen role collides with the tail AND flipping wouldn't
-            # collide with the head, flip it.
-            if summary_role == first_tail_role:
-                flipped = "assistant" if summary_role == "user" else "user"
-                if flipped != last_head_role:
-                    summary_role = flipped
-                else:
-                    # Both roles would create consecutive same-role messages
-                    # (e.g. head=assistant, tail=user — neither role works).
-                    # Merge the summary into the first tail message instead
-                    # of inserting a standalone message that breaks alternation.
-                    _merge_summary_into_tail = True
-            if not _merge_summary_into_tail:
-                compressed.append({"role": summary_role, "content": summary})
-        else:
-            if not self.quiet_mode:
-                logger.warning("No summary model available — middle turns dropped without summary")
-
-        for i in range(compress_end, n_messages):
-            msg = messages[i].copy()
-            if _merge_summary_into_tail and i == compress_end:
-                original = msg.get("content") or ""
-                msg["content"] = summary + "\n\n" + original
-                _merge_summary_into_tail = False
-            compressed.append(msg)
-
-        self.compression_count += 1
-
-        compressed = self._sanitize_tool_pairs(compressed)
-
-        if not self.quiet_mode:
-            new_estimate = estimate_messages_tokens_rough(compressed)
-            saved_estimate = display_tokens - new_estimate
-            logger.info(
-                "Compressed: %d -> %d messages (~%d tokens saved)",
-                n_messages,
-                len(compressed),
-                saved_estimate,
-            )
-            logger.info("Compression #%d complete", self.compression_count)
-
-        return compressed

agent/context_references.py

@@ -1,492 +0,0 @@
-from __future__ import annotations
-
-import asyncio
-import inspect
-import json
-import mimetypes
-import os
-import re
-import subprocess
-from dataclasses import dataclass, field
-from pathlib import Path
-from typing import Awaitable, Callable
-
-from agent.model_metadata import estimate_tokens_rough
-
-REFERENCE_PATTERN = re.compile(
-    r"(?<![\w/])@(?:(?P<simple>diff|staged)\b|(?P<kind>file|folder|git|url):(?P<value>\S+))"
-)
-TRAILING_PUNCTUATION = ",.;!?"
-_SENSITIVE_HOME_DIRS = (".ssh", ".aws", ".gnupg", ".kube", ".docker", ".azure", ".config/gh")
-_SENSITIVE_HERMES_DIRS = (Path("skills") / ".hub",)
-_SENSITIVE_HOME_FILES = (
-    Path(".ssh") / "authorized_keys",
-    Path(".ssh") / "id_rsa",
-    Path(".ssh") / "id_ed25519",
-    Path(".ssh") / "config",
-    Path(".bashrc"),
-    Path(".zshrc"),
-    Path(".profile"),
-    Path(".bash_profile"),
-    Path(".zprofile"),
-    Path(".netrc"),
-    Path(".pgpass"),
-    Path(".npmrc"),
-    Path(".pypirc"),
-)
-
-
-@dataclass(frozen=True)
-class ContextReference:
-    raw: str
-    kind: str
-    target: str
-    start: int
-    end: int
-    line_start: int | None = None
-    line_end: int | None = None
-
-
-@dataclass
-class ContextReferenceResult:
-    message: str
-    original_message: str
-    references: list[ContextReference] = field(default_factory=list)
-    warnings: list[str] = field(default_factory=list)
-    injected_tokens: int = 0
-    expanded: bool = False
-    blocked: bool = False
-
-
-def parse_context_references(message: str) -> list[ContextReference]:
-    refs: list[ContextReference] = []
-    if not message:
-        return refs
-
-    for match in REFERENCE_PATTERN.finditer(message):
-        simple = match.group("simple")
-        if simple:
-            refs.append(
-                ContextReference(
-                    raw=match.group(0),
-                    kind=simple,
-                    target="",
-                    start=match.start(),
-                    end=match.end(),
-                )
-            )
-            continue
-
-        kind = match.group("kind")
-        value = _strip_trailing_punctuation(match.group("value") or "")
-        line_start = None
-        line_end = None
-        target = value
-
-        if kind == "file":
-            range_match = re.match(r"^(?P<path>.+?):(?P<start>\d+)(?:-(?P<end>\d+))?$", value)
-            if range_match:
-                target = range_match.group("path")
-                line_start = int(range_match.group("start"))
-                line_end = int(range_match.group("end") or range_match.group("start"))
-
-        refs.append(
-            ContextReference(
-                raw=match.group(0),
-                kind=kind,
-                target=target,
-                start=match.start(),
-                end=match.end(),
-                line_start=line_start,
-                line_end=line_end,
-            )
-        )
-
-    return refs
-
-
-def preprocess_context_references(
-    message: str,
-    *,
-    cwd: str | Path,
-    context_length: int,
-    url_fetcher: Callable[[str], str | Awaitable[str]] | None = None,
-    allowed_root: str | Path | None = None,
-) -> ContextReferenceResult:
-    coro = preprocess_context_references_async(
-        message,
-        cwd=cwd,
-        context_length=context_length,
-        url_fetcher=url_fetcher,
-        allowed_root=allowed_root,
-    )
-    # Safe for both CLI (no loop) and gateway (loop already running).
-    try:
-        loop = asyncio.get_running_loop()
-    except RuntimeError:
-        loop = None
-    if loop and loop.is_running():
-        import concurrent.futures
-        with concurrent.futures.ThreadPoolExecutor(max_workers=1) as pool:
-            return pool.submit(asyncio.run, coro).result()
-    return asyncio.run(coro)
-
-
-async def preprocess_context_references_async(
-    message: str,
-    *,
-    cwd: str | Path,
-    context_length: int,
-    url_fetcher: Callable[[str], str | Awaitable[str]] | None = None,
-    allowed_root: str | Path | None = None,
-) -> ContextReferenceResult:
-    refs = parse_context_references(message)
-    if not refs:
-        return ContextReferenceResult(message=message, original_message=message)
-
-    cwd_path = Path(cwd).expanduser().resolve()
-    # Default to the current working directory so @ references cannot escape
-    # the active workspace unless a caller explicitly widens the root.
-    allowed_root_path = (
-        Path(allowed_root).expanduser().resolve() if allowed_root is not None else cwd_path
-    )
-    warnings: list[str] = []
-    blocks: list[str] = []
-    injected_tokens = 0
-
-    for ref in refs:
-        warning, block = await _expand_reference(
-            ref,
-            cwd_path,
-            url_fetcher=url_fetcher,
-            allowed_root=allowed_root_path,
-        )
-        if warning:
-            warnings.append(warning)
-        if block:
-            blocks.append(block)
-            injected_tokens += estimate_tokens_rough(block)
-
-    hard_limit = max(1, int(context_length * 0.50))
-    soft_limit = max(1, int(context_length * 0.25))
-    if injected_tokens > hard_limit:
-        warnings.append(
-            f"@ context injection refused: {injected_tokens} tokens exceeds the 50% hard limit ({hard_limit})."
-        )
-        return ContextReferenceResult(
-            message=message,
-            original_message=message,
-            references=refs,
-            warnings=warnings,
-            injected_tokens=injected_tokens,
-            expanded=False,
-            blocked=True,
-        )
-
-    if injected_tokens > soft_limit:
-        warnings.append(
-            f"@ context injection warning: {injected_tokens} tokens exceeds the 25% soft limit ({soft_limit})."
-        )
-
-    stripped = _remove_reference_tokens(message, refs)
-    final = stripped
-    if warnings:
-        final = f"{final}\n\n--- Context Warnings ---\n" + "\n".join(f"- {warning}" for warning in warnings)
-    if blocks:
-        final = f"{final}\n\n--- Attached Context ---\n\n" + "\n\n".join(blocks)
-
-    return ContextReferenceResult(
-        message=final.strip(),
-        original_message=message,
-        references=refs,
-        warnings=warnings,
-        injected_tokens=injected_tokens,
-        expanded=bool(blocks or warnings),
-        blocked=False,
-    )
-
-
-async def _expand_reference(
-    ref: ContextReference,
-    cwd: Path,
-    *,
-    url_fetcher: Callable[[str], str | Awaitable[str]] | None = None,
-    allowed_root: Path | None = None,
-) -> tuple[str | None, str | None]:
-    try:
-        if ref.kind == "file":
-            return _expand_file_reference(ref, cwd, allowed_root=allowed_root)
-        if ref.kind == "folder":
-            return _expand_folder_reference(ref, cwd, allowed_root=allowed_root)
-        if ref.kind == "diff":
-            return _expand_git_reference(ref, cwd, ["diff"], "git diff")
-        if ref.kind == "staged":
-            return _expand_git_reference(ref, cwd, ["diff", "--staged"], "git diff --staged")
-        if ref.kind == "git":
-            count = max(1, min(int(ref.target or "1"), 10))
-            return _expand_git_reference(ref, cwd, ["log", f"-{count}", "-p"], f"git log -{count} -p")
-        if ref.kind == "url":
-            content = await _fetch_url_content(ref.target, url_fetcher=url_fetcher)
-            if not content:
-                return f"{ref.raw}: no content extracted", None
-            return None, f"🌐 {ref.raw} ({estimate_tokens_rough(content)} tokens)\n{content}"
-    except Exception as exc:
-        return f"{ref.raw}: {exc}", None
-
-    return f"{ref.raw}: unsupported reference type", None
-
-
-def _expand_file_reference(
-    ref: ContextReference,
-    cwd: Path,
-    *,
-    allowed_root: Path | None = None,
-) -> tuple[str | None, str | None]:
-    path = _resolve_path(cwd, ref.target, allowed_root=allowed_root)
-    _ensure_reference_path_allowed(path)
-    if not path.exists():
-        return f"{ref.raw}: file not found", None
-    if not path.is_file():
-        return f"{ref.raw}: path is not a file", None
-    if _is_binary_file(path):
-        return f"{ref.raw}: binary files are not supported", None
-
-    text = path.read_text(encoding="utf-8")
-    if ref.line_start is not None:
-        lines = text.splitlines()
-        start_idx = max(ref.line_start - 1, 0)
-        end_idx = min(ref.line_end or ref.line_start, len(lines))
-        text = "\n".join(lines[start_idx:end_idx])
-
-    lang = _code_fence_language(path)
-    label = ref.raw
-    return None, f"📄 {label} ({estimate_tokens_rough(text)} tokens)\n```{lang}\n{text}\n```"
-
-
-def _expand_folder_reference(
-    ref: ContextReference,
-    cwd: Path,
-    *,
-    allowed_root: Path | None = None,
-) -> tuple[str | None, str | None]:
-    path = _resolve_path(cwd, ref.target, allowed_root=allowed_root)
-    _ensure_reference_path_allowed(path)
-    if not path.exists():
-        return f"{ref.raw}: folder not found", None
-    if not path.is_dir():
-        return f"{ref.raw}: path is not a folder", None
-
-    listing = _build_folder_listing(path, cwd)
-    return None, f"📁 {ref.raw} ({estimate_tokens_rough(listing)} tokens)\n{listing}"
-
-
-def _expand_git_reference(
-    ref: ContextReference,
-    cwd: Path,
-    args: list[str],
-    label: str,
-) -> tuple[str | None, str | None]:
-    try:
-        result = subprocess.run(
-            ["git", *args],
-            cwd=cwd,
-            capture_output=True,
-            text=True,
-            timeout=30,
-        )
-    except subprocess.TimeoutExpired:
-        return f"{ref.raw}: git command timed out (30s)", None
-    if result.returncode != 0:
-        stderr = (result.stderr or "").strip() or "git command failed"
-        return f"{ref.raw}: {stderr}", None
-    content = result.stdout.strip()
-    if not content:
-        content = "(no output)"
-    return None, f"🧾 {label} ({estimate_tokens_rough(content)} tokens)\n```diff\n{content}\n```"
-
-
-async def _fetch_url_content(
-    url: str,
-    *,
-    url_fetcher: Callable[[str], str | Awaitable[str]] | None = None,
-) -> str:
-    fetcher = url_fetcher or _default_url_fetcher
-    content = fetcher(url)
-    if inspect.isawaitable(content):
-        content = await content
-    return str(content or "").strip()
-
-
-async def _default_url_fetcher(url: str) -> str:
-    from tools.web_tools import web_extract_tool
-
-    raw = await web_extract_tool([url], format="markdown", use_llm_processing=True)
-    payload = json.loads(raw)
-    docs = payload.get("data", {}).get("documents", [])
-    if not docs:
-        return ""
-    doc = docs[0]
-    return str(doc.get("content") or doc.get("raw_content") or "").strip()
-
-
-def _resolve_path(cwd: Path, target: str, *, allowed_root: Path | None = None) -> Path:
-    path = Path(os.path.expanduser(target))
-    if not path.is_absolute():
-        path = cwd / path
-    resolved = path.resolve()
-    if allowed_root is not None:
-        try:
-            resolved.relative_to(allowed_root)
-        except ValueError as exc:
-            raise ValueError("path is outside the allowed workspace") from exc
-    return resolved
-
-
-def _ensure_reference_path_allowed(path: Path) -> None:
-    home = Path(os.path.expanduser("~")).resolve()
-    hermes_home = Path(
-        os.getenv("HERMES_HOME", str(home / ".hermes"))
-    ).expanduser().resolve()
-
-    blocked_exact = {home / rel for rel in _SENSITIVE_HOME_FILES}
-    blocked_exact.add(hermes_home / ".env")
-    blocked_dirs = [home / rel for rel in _SENSITIVE_HOME_DIRS]
-    blocked_dirs.extend(hermes_home / rel for rel in _SENSITIVE_HERMES_DIRS)
-
-    if path in blocked_exact:
-        raise ValueError("path is a sensitive credential file and cannot be attached")
-
-    for blocked_dir in blocked_dirs:
-        try:
-            path.relative_to(blocked_dir)
-        except ValueError:
-            continue
-        raise ValueError("path is a sensitive credential or internal Hermes path and cannot be attached")
-
-
-def _strip_trailing_punctuation(value: str) -> str:
-    stripped = value.rstrip(TRAILING_PUNCTUATION)
-    while stripped.endswith((")", "]", "}")):
-        closer = stripped[-1]
-        opener = {")": "(", "]": "[", "}": "{"}[closer]
-        if stripped.count(closer) > stripped.count(opener):
-            stripped = stripped[:-1]
-            continue
-        break
-    return stripped
-
-
-def _remove_reference_tokens(message: str, refs: list[ContextReference]) -> str:
-    pieces: list[str] = []
-    cursor = 0
-    for ref in refs:
-        pieces.append(message[cursor:ref.start])
-        cursor = ref.end
-    pieces.append(message[cursor:])
-    text = "".join(pieces)
-    text = re.sub(r"\s{2,}", " ", text)
-    text = re.sub(r"\s+([,.;:!?])", r"\1", text)
-    return text.strip()
-
-
-def _is_binary_file(path: Path) -> bool:
-    mime, _ = mimetypes.guess_type(path.name)
-    if mime and not mime.startswith("text/") and not any(
-        path.name.endswith(ext) for ext in (".py", ".md", ".txt", ".json", ".yaml", ".yml", ".toml", ".js", ".ts")
-    ):
-        return True
-    chunk = path.read_bytes()[:4096]
-    return b"\x00" in chunk
-
-
-def _build_folder_listing(path: Path, cwd: Path, limit: int = 200) -> str:
-    lines = [f"{path.relative_to(cwd)}/"]
-    entries = _iter_visible_entries(path, cwd, limit=limit)
-    for entry in entries:
-        rel = entry.relative_to(cwd)
-        indent = "  " * max(len(rel.parts) - len(path.relative_to(cwd).parts) - 1, 0)
-        if entry.is_dir():
-            lines.append(f"{indent}- {entry.name}/")
-        else:
-            meta = _file_metadata(entry)
-            lines.append(f"{indent}- {entry.name} ({meta})")
-    if len(entries) >= limit:
-        lines.append("- ...")
-    return "\n".join(lines)
-
-
-def _iter_visible_entries(path: Path, cwd: Path, limit: int) -> list[Path]:
-    rg_entries = _rg_files(path, cwd, limit=limit)
-    if rg_entries is not None:
-        output: list[Path] = []
-        seen_dirs: set[Path] = set()
-        for rel in rg_entries:
-            full = cwd / rel
-            for parent in full.parents:
-                if parent == cwd or parent in seen_dirs or path not in {parent, *parent.parents}:
-                    continue
-                seen_dirs.add(parent)
-                output.append(parent)
-            output.append(full)
-        return sorted({p for p in output if p.exists()}, key=lambda p: (not p.is_dir(), str(p)))
-
-    output = []
-    for root, dirs, files in os.walk(path):
-        dirs[:] = sorted(d for d in dirs if not d.startswith(".") and d != "__pycache__")
-        files = sorted(f for f in files if not f.startswith("."))
-        root_path = Path(root)
-        for d in dirs:
-            output.append(root_path / d)
-            if len(output) >= limit:
-                return output
-        for f in files:
-            output.append(root_path / f)
-            if len(output) >= limit:
-                return output
-    return output
-
-
-def _rg_files(path: Path, cwd: Path, limit: int) -> list[Path] | None:
-    try:
-        result = subprocess.run(
-            ["rg", "--files", str(path.relative_to(cwd))],
-            cwd=cwd,
-            capture_output=True,
-            text=True,
-            timeout=10,
-        )
-    except FileNotFoundError:
-        return None
-    except subprocess.TimeoutExpired:
-        return None
-    if result.returncode != 0:
-        return None
-    files = [Path(line.strip()) for line in result.stdout.splitlines() if line.strip()]
-    return files[:limit]
-
-
-def _file_metadata(path: Path) -> str:
-    if _is_binary_file(path):
-        return f"{path.stat().st_size} bytes"
-    try:
-        line_count = path.read_text(encoding="utf-8").count("\n") + 1
-    except Exception:
-        return f"{path.stat().st_size} bytes"
-    return f"{line_count} lines"
-
-
-def _code_fence_language(path: Path) -> str:
-    mapping = {
-        ".py": "python",
-        ".js": "javascript",
-        ".ts": "typescript",
-        ".tsx": "tsx",
-        ".jsx": "jsx",
-        ".json": "json",
-        ".md": "markdown",
-        ".sh": "bash",
-        ".yml": "yaml",
-        ".yaml": "yaml",
-        ".toml": "toml",
-    }
-    return mapping.get(path.suffix.lower(), "")

agent/copilot_acp_client.py

@@ -1,447 +0,0 @@
-"""OpenAI-compatible shim that forwards Hermes requests to `copilot --acp`.
-
-This adapter lets Hermes treat the GitHub Copilot ACP server as a chat-style
-backend. Each request starts a short-lived ACP session, sends the formatted
-conversation as a single prompt, collects text chunks, and converts the result
-back into the minimal shape Hermes expects from an OpenAI client.
-"""
-
-from __future__ import annotations
-
-import json
-import os
-import queue
-import shlex
-import subprocess
-import threading
-import time
-from collections import deque
-from pathlib import Path
-from types import SimpleNamespace
-from typing import Any
-
-ACP_MARKER_BASE_URL = "acp://copilot"
-_DEFAULT_TIMEOUT_SECONDS = 900.0
-
-
-def _resolve_command() -> str:
-    return (
-        os.getenv("HERMES_COPILOT_ACP_COMMAND", "").strip()
-        or os.getenv("COPILOT_CLI_PATH", "").strip()
-        or "copilot"
-    )
-
-
-def _resolve_args() -> list[str]:
-    raw = os.getenv("HERMES_COPILOT_ACP_ARGS", "").strip()
-    if not raw:
-        return ["--acp", "--stdio"]
-    return shlex.split(raw)
-
-
-def _jsonrpc_error(message_id: Any, code: int, message: str) -> dict[str, Any]:
-    return {
-        "jsonrpc": "2.0",
-        "id": message_id,
-        "error": {
-            "code": code,
-            "message": message,
-        },
-    }
-
-
-def _format_messages_as_prompt(messages: list[dict[str, Any]], model: str | None = None) -> str:
-    sections: list[str] = [
-        "You are being used as the active ACP agent backend for Hermes.",
-        "Use your own ACP capabilities and respond directly in natural language.",
-        "Do not emit OpenAI tool-call JSON.",
-    ]
-    if model:
-        sections.append(f"Hermes requested model hint: {model}")
-
-    transcript: list[str] = []
-    for message in messages:
-        if not isinstance(message, dict):
-            continue
-        role = str(message.get("role") or "unknown").strip().lower()
-        if role == "tool":
-            role = "tool"
-        elif role not in {"system", "user", "assistant"}:
-            role = "context"
-
-        content = message.get("content")
-        rendered = _render_message_content(content)
-        if not rendered:
-            continue
-
-        label = {
-            "system": "System",
-            "user": "User",
-            "assistant": "Assistant",
-            "tool": "Tool",
-            "context": "Context",
-        }.get(role, role.title())
-        transcript.append(f"{label}:\n{rendered}")
-
-    if transcript:
-        sections.append("Conversation transcript:\n\n" + "\n\n".join(transcript))
-
-    sections.append("Continue the conversation from the latest user request.")
-    return "\n\n".join(section.strip() for section in sections if section and section.strip())
-
-
-def _render_message_content(content: Any) -> str:
-    if content is None:
-        return ""
-    if isinstance(content, str):
-        return content.strip()
-    if isinstance(content, dict):
-        if "text" in content:
-            return str(content.get("text") or "").strip()
-        if "content" in content and isinstance(content.get("content"), str):
-            return str(content.get("content") or "").strip()
-        return json.dumps(content, ensure_ascii=True)
-    if isinstance(content, list):
-        parts: list[str] = []
-        for item in content:
-            if isinstance(item, str):
-                parts.append(item)
-            elif isinstance(item, dict):
-                text = item.get("text")
-                if isinstance(text, str) and text.strip():
-                    parts.append(text.strip())
-        return "\n".join(parts).strip()
-    return str(content).strip()
-
-
-def _ensure_path_within_cwd(path_text: str, cwd: str) -> Path:
-    candidate = Path(path_text)
-    if not candidate.is_absolute():
-        raise PermissionError("ACP file-system paths must be absolute.")
-    resolved = candidate.resolve()
-    root = Path(cwd).resolve()
-    try:
-        resolved.relative_to(root)
-    except ValueError as exc:
-        raise PermissionError(f"Path '{resolved}' is outside the session cwd '{root}'.") from exc
-    return resolved
-
-
-class _ACPChatCompletions:
-    def __init__(self, client: "CopilotACPClient"):
-        self._client = client
-
-    def create(self, **kwargs: Any) -> Any:
-        return self._client._create_chat_completion(**kwargs)
-
-
-class _ACPChatNamespace:
-    def __init__(self, client: "CopilotACPClient"):
-        self.completions = _ACPChatCompletions(client)
-
-
-class CopilotACPClient:
-    """Minimal OpenAI-client-compatible facade for Copilot ACP."""
-
-    def __init__(
-        self,
-        *,
-        api_key: str | None = None,
-        base_url: str | None = None,
-        default_headers: dict[str, str] | None = None,
-        acp_command: str | None = None,
-        acp_args: list[str] | None = None,
-        acp_cwd: str | None = None,
-        command: str | None = None,
-        args: list[str] | None = None,
-        **_: Any,
-    ):
-        self.api_key = api_key or "copilot-acp"
-        self.base_url = base_url or ACP_MARKER_BASE_URL
-        self._default_headers = dict(default_headers or {})
-        self._acp_command = acp_command or command or _resolve_command()
-        self._acp_args = list(acp_args or args or _resolve_args())
-        self._acp_cwd = str(Path(acp_cwd or os.getcwd()).resolve())
-        self.chat = _ACPChatNamespace(self)
-        self.is_closed = False
-        self._active_process: subprocess.Popen[str] | None = None
-        self._active_process_lock = threading.Lock()
-
-    def close(self) -> None:
-        proc: subprocess.Popen[str] | None
-        with self._active_process_lock:
-            proc = self._active_process
-            self._active_process = None
-        self.is_closed = True
-        if proc is None:
-            return
-        try:
-            proc.terminate()
-            proc.wait(timeout=2)
-        except Exception:
-            try:
-                proc.kill()
-            except Exception:
-                pass
-
-    def _create_chat_completion(
-        self,
-        *,
-        model: str | None = None,
-        messages: list[dict[str, Any]] | None = None,
-        timeout: float | None = None,
-        **_: Any,
-    ) -> Any:
-        prompt_text = _format_messages_as_prompt(messages or [], model=model)
-        response_text, reasoning_text = self._run_prompt(
-            prompt_text,
-            timeout_seconds=float(timeout or _DEFAULT_TIMEOUT_SECONDS),
-        )
-
-        usage = SimpleNamespace(
-            prompt_tokens=0,
-            completion_tokens=0,
-            total_tokens=0,
-            prompt_tokens_details=SimpleNamespace(cached_tokens=0),
-        )
-        assistant_message = SimpleNamespace(
-            content=response_text,
-            tool_calls=[],
-            reasoning=reasoning_text or None,
-            reasoning_content=reasoning_text or None,
-            reasoning_details=None,
-        )
-        choice = SimpleNamespace(message=assistant_message, finish_reason="stop")
-        return SimpleNamespace(
-            choices=[choice],
-            usage=usage,
-            model=model or "copilot-acp",
-        )
-
-    def _run_prompt(self, prompt_text: str, *, timeout_seconds: float) -> tuple[str, str]:
-        try:
-            proc = subprocess.Popen(
-                [self._acp_command] + self._acp_args,
-                stdin=subprocess.PIPE,
-                stdout=subprocess.PIPE,
-                stderr=subprocess.PIPE,
-                text=True,
-                bufsize=1,
-                cwd=self._acp_cwd,
-            )
-        except FileNotFoundError as exc:
-            raise RuntimeError(
-                f"Could not start Copilot ACP command '{self._acp_command}'. "
-                "Install GitHub Copilot CLI or set HERMES_COPILOT_ACP_COMMAND/COPILOT_CLI_PATH."
-            ) from exc
-
-        if proc.stdin is None or proc.stdout is None:
-            proc.kill()
-            raise RuntimeError("Copilot ACP process did not expose stdin/stdout pipes.")
-
-        self.is_closed = False
-        with self._active_process_lock:
-            self._active_process = proc
-
-        inbox: queue.Queue[dict[str, Any]] = queue.Queue()
-        stderr_tail: deque[str] = deque(maxlen=40)
-
-        def _stdout_reader() -> None:
-            for line in proc.stdout:
-                try:
-                    inbox.put(json.loads(line))
-                except Exception:
-                    inbox.put({"raw": line.rstrip("\n")})
-
-        def _stderr_reader() -> None:
-            if proc.stderr is None:
-                return
-            for line in proc.stderr:
-                stderr_tail.append(line.rstrip("\n"))
-
-        out_thread = threading.Thread(target=_stdout_reader, daemon=True)
-        err_thread = threading.Thread(target=_stderr_reader, daemon=True)
-        out_thread.start()
-        err_thread.start()
-
-        next_id = 0
-
-        def _request(method: str, params: dict[str, Any], *, text_parts: list[str] | None = None, reasoning_parts: list[str] | None = None) -> Any:
-            nonlocal next_id
-            next_id += 1
-            request_id = next_id
-            payload = {
-                "jsonrpc": "2.0",
-                "id": request_id,
-                "method": method,
-                "params": params,
-            }
-            proc.stdin.write(json.dumps(payload) + "\n")
-            proc.stdin.flush()
-
-            deadline = time.time() + timeout_seconds
-            while time.time() < deadline:
-                if proc.poll() is not None:
-                    break
-                try:
-                    msg = inbox.get(timeout=0.1)
-                except queue.Empty:
-                    continue
-
-                if self._handle_server_message(
-                    msg,
-                    process=proc,
-                    cwd=self._acp_cwd,
-                    text_parts=text_parts,
-                    reasoning_parts=reasoning_parts,
-                ):
-                    continue
-
-                if msg.get("id") != request_id:
-                    continue
-                if "error" in msg:
-                    err = msg.get("error") or {}
-                    raise RuntimeError(
-                        f"Copilot ACP {method} failed: {err.get('message') or err}"
-                    )
-                return msg.get("result")
-
-            stderr_text = "\n".join(stderr_tail).strip()
-            if proc.poll() is not None and stderr_text:
-                raise RuntimeError(f"Copilot ACP process exited early: {stderr_text}")
-            raise TimeoutError(f"Timed out waiting for Copilot ACP response to {method}.")
-
-        try:
-            _request(
-                "initialize",
-                {
-                    "protocolVersion": 1,
-                    "clientCapabilities": {
-                        "fs": {
-                            "readTextFile": True,
-                            "writeTextFile": True,
-                        }
-                    },
-                    "clientInfo": {
-                        "name": "hermes-agent",
-                        "title": "Hermes Agent",
-                        "version": "0.0.0",
-                    },
-                },
-            )
-            session = _request(
-                "session/new",
-                {
-                    "cwd": self._acp_cwd,
-                    "mcpServers": [],
-                },
-            ) or {}
-            session_id = str(session.get("sessionId") or "").strip()
-            if not session_id:
-                raise RuntimeError("Copilot ACP did not return a sessionId.")
-
-            text_parts: list[str] = []
-            reasoning_parts: list[str] = []
-            _request(
-                "session/prompt",
-                {
-                    "sessionId": session_id,
-                    "prompt": [
-                        {
-                            "type": "text",
-                            "text": prompt_text,
-                        }
-                    ],
-                },
-                text_parts=text_parts,
-                reasoning_parts=reasoning_parts,
-            )
-            return "".join(text_parts), "".join(reasoning_parts)
-        finally:
-            self.close()
-
-    def _handle_server_message(
-        self,
-        msg: dict[str, Any],
-        *,
-        process: subprocess.Popen[str],
-        cwd: str,
-        text_parts: list[str] | None,
-        reasoning_parts: list[str] | None,
-    ) -> bool:
-        method = msg.get("method")
-        if not isinstance(method, str):
-            return False
-
-        if method == "session/update":
-            params = msg.get("params") or {}
-            update = params.get("update") or {}
-            kind = str(update.get("sessionUpdate") or "").strip()
-            content = update.get("content") or {}
-            chunk_text = ""
-            if isinstance(content, dict):
-                chunk_text = str(content.get("text") or "")
-            if kind == "agent_message_chunk" and chunk_text and text_parts is not None:
-                text_parts.append(chunk_text)
-            elif kind == "agent_thought_chunk" and chunk_text and reasoning_parts is not None:
-                reasoning_parts.append(chunk_text)
-            return True
-
-        if process.stdin is None:
-            return True
-
-        message_id = msg.get("id")
-        params = msg.get("params") or {}
-
-        if method == "session/request_permission":
-            response = {
-                "jsonrpc": "2.0",
-                "id": message_id,
-                "result": {
-                    "outcome": {
-                        "outcome": "allow_once",
-                    }
-                },
-            }
-        elif method == "fs/read_text_file":
-            try:
-                path = _ensure_path_within_cwd(str(params.get("path") or ""), cwd)
-                content = path.read_text() if path.exists() else ""
-                line = params.get("line")
-                limit = params.get("limit")
-                if isinstance(line, int) and line > 1:
-                    lines = content.splitlines(keepends=True)
-                    start = line - 1
-                    end = start + limit if isinstance(limit, int) and limit > 0 else None
-                    content = "".join(lines[start:end])
-                response = {
-                    "jsonrpc": "2.0",
-                    "id": message_id,
-                    "result": {
-                        "content": content,
-                    },
-                }
-            except Exception as exc:
-                response = _jsonrpc_error(message_id, -32602, str(exc))
-        elif method == "fs/write_text_file":
-            try:
-                path = _ensure_path_within_cwd(str(params.get("path") or ""), cwd)
-                path.parent.mkdir(parents=True, exist_ok=True)
-                path.write_text(str(params.get("content") or ""))
-                response = {
-                    "jsonrpc": "2.0",
-                    "id": message_id,
-                    "result": None,
-                }
-            except Exception as exc:
-                response = _jsonrpc_error(message_id, -32602, str(exc))
-        else:
-            response = _jsonrpc_error(
-                message_id,
-                -32601,
-                f"ACP client method '{method}' is not supported by Hermes yet.",
-            )
-
-        process.stdin.write(json.dumps(response) + "\n")
-        process.stdin.flush()
-        return True

agent/credential_pool.py

@@ -1,848 +0,0 @@
-"""Persistent multi-credential pool for same-provider failover."""
-
-from __future__ import annotations
-
-import logging
-import random
-import threading
-import time
-import uuid
-import os
-from dataclasses import dataclass, fields, replace
-from typing import Any, Dict, List, Optional, Set, Tuple
-
-from hermes_constants import OPENROUTER_BASE_URL
-import hermes_cli.auth as auth_mod
-from hermes_cli.auth import (
-    ACCESS_TOKEN_REFRESH_SKEW_SECONDS,
-    CODEX_ACCESS_TOKEN_REFRESH_SKEW_SECONDS,
-    DEFAULT_AGENT_KEY_MIN_TTL_SECONDS,
-    PROVIDER_REGISTRY,
-    _agent_key_is_usable,
-    _codex_access_token_is_expiring,
-    _decode_jwt_claims,
-    _is_expiring,
-    _load_auth_store,
-    _load_provider_state,
-    read_credential_pool,
-    write_credential_pool,
-)
-
-logger = logging.getLogger(__name__)
-
-
-def _load_config_safe() -> Optional[dict]:
-    """Load config.yaml, returning None on any error."""
-    try:
-        from hermes_cli.config import load_config
-
-        return load_config()
-    except Exception:
-        return None
-
-
-# --- Status and type constants ---
-
-STATUS_OK = "ok"
-STATUS_EXHAUSTED = "exhausted"
-
-AUTH_TYPE_OAUTH = "oauth"
-AUTH_TYPE_API_KEY = "api_key"
-
-SOURCE_MANUAL = "manual"
-
-STRATEGY_FILL_FIRST = "fill_first"
-STRATEGY_ROUND_ROBIN = "round_robin"
-STRATEGY_RANDOM = "random"
-STRATEGY_LEAST_USED = "least_used"
-SUPPORTED_POOL_STRATEGIES = {
-    STRATEGY_FILL_FIRST,
-    STRATEGY_ROUND_ROBIN,
-    STRATEGY_RANDOM,
-    STRATEGY_LEAST_USED,
-}
-
-# Cooldown before retrying an exhausted credential.
-# 429 (rate-limited) cools down faster since quotas reset frequently.
-# 402 (billing/quota) and other codes use a longer default.
-EXHAUSTED_TTL_429_SECONDS = 60 * 60          # 1 hour
-EXHAUSTED_TTL_DEFAULT_SECONDS = 24 * 60 * 60 # 24 hours
-
-# Pool key prefix for custom OpenAI-compatible endpoints.
-# Custom endpoints all share provider='custom' but are keyed by their
-# custom_providers name: 'custom:<normalized_name>'.
-CUSTOM_POOL_PREFIX = "custom:"
-
-
-# Fields that are only round-tripped through JSON — never used for logic as attributes.
-_EXTRA_KEYS = frozenset({
-    "token_type", "scope", "client_id", "portal_base_url", "obtained_at",
-    "expires_in", "agent_key_id", "agent_key_expires_in", "agent_key_reused",
-    "agent_key_obtained_at", "tls",
-})
-
-
-@dataclass
-class PooledCredential:
-    provider: str
-    id: str
-    label: str
-    auth_type: str
-    priority: int
-    source: str
-    access_token: str
-    refresh_token: Optional[str] = None
-    last_status: Optional[str] = None
-    last_status_at: Optional[float] = None
-    last_error_code: Optional[int] = None
-    base_url: Optional[str] = None
-    expires_at: Optional[str] = None
-    expires_at_ms: Optional[int] = None
-    last_refresh: Optional[str] = None
-    inference_base_url: Optional[str] = None
-    agent_key: Optional[str] = None
-    agent_key_expires_at: Optional[str] = None
-    request_count: int = 0
-    extra: Dict[str, Any] = None  # type: ignore[assignment]
-
-    def __post_init__(self):
-        if self.extra is None:
-            self.extra = {}
-
-    def __getattr__(self, name: str):
-        if name in _EXTRA_KEYS:
-            return self.extra.get(name)
-        raise AttributeError(f"'{type(self).__name__}' object has no attribute {name!r}")
-
-    @classmethod
-    def from_dict(cls, provider: str, payload: Dict[str, Any]) -> "PooledCredential":
-        field_names = {f.name for f in fields(cls) if f.name != "provider"}
-        data = {k: payload.get(k) for k in field_names if k in payload}
-        extra = {k: payload[k] for k in _EXTRA_KEYS if k in payload and payload[k] is not None}
-        data["extra"] = extra
-        data.setdefault("id", uuid.uuid4().hex[:6])
-        data.setdefault("label", payload.get("source", provider))
-        data.setdefault("auth_type", AUTH_TYPE_API_KEY)
-        data.setdefault("priority", 0)
-        data.setdefault("source", SOURCE_MANUAL)
-        data.setdefault("access_token", "")
-        return cls(provider=provider, **data)
-
-    def to_dict(self) -> Dict[str, Any]:
-        _ALWAYS_EMIT = {"last_status", "last_status_at", "last_error_code"}
-        result: Dict[str, Any] = {}
-        for field_def in fields(self):
-            if field_def.name in ("provider", "extra"):
-                continue
-            value = getattr(self, field_def.name)
-            if value is not None or field_def.name in _ALWAYS_EMIT:
-                result[field_def.name] = value
-        for k, v in self.extra.items():
-            if v is not None:
-                result[k] = v
-        return result
-
-    @property
-    def runtime_api_key(self) -> str:
-        if self.provider == "nous":
-            return str(self.agent_key or self.access_token or "")
-        return str(self.access_token or "")
-
-    @property
-    def runtime_base_url(self) -> Optional[str]:
-        if self.provider == "nous":
-            return self.inference_base_url or self.base_url
-        return self.base_url
-
-
-def label_from_token(token: str, fallback: str) -> str:
-    claims = _decode_jwt_claims(token)
-    for key in ("email", "preferred_username", "upn"):
-        value = claims.get(key)
-        if isinstance(value, str) and value.strip():
-            return value.strip()
-    return fallback
-
-
-def _next_priority(entries: List[PooledCredential]) -> int:
-    return max((entry.priority for entry in entries), default=-1) + 1
-
-
-def _is_manual_source(source: str) -> bool:
-    normalized = (source or "").strip().lower()
-    return normalized == SOURCE_MANUAL or normalized.startswith(f"{SOURCE_MANUAL}:")
-
-
-def _exhausted_ttl(error_code: Optional[int]) -> int:
-    """Return cooldown seconds based on the HTTP status that caused exhaustion."""
-    if error_code == 429:
-        return EXHAUSTED_TTL_429_SECONDS
-    return EXHAUSTED_TTL_DEFAULT_SECONDS
-
-
-def _normalize_custom_pool_name(name: str) -> str:
-    """Normalize a custom provider name for use as a pool key suffix."""
-    return name.strip().lower().replace(" ", "-")
-
-
-def _iter_custom_providers(config: Optional[dict] = None):
-    """Yield (normalized_name, entry_dict) for each valid custom_providers entry."""
-    if config is None:
-        config = _load_config_safe()
-    if config is None:
-        return
-    custom_providers = config.get("custom_providers")
-    if not isinstance(custom_providers, list):
-        return
-    for entry in custom_providers:
-        if not isinstance(entry, dict):
-            continue
-        name = entry.get("name")
-        if not isinstance(name, str):
-            continue
-        yield _normalize_custom_pool_name(name), entry
-
-
-def get_custom_provider_pool_key(base_url: str) -> Optional[str]:
-    """Look up the custom_providers list in config.yaml and return 'custom:<name>' for a matching base_url.
-
-    Returns None if no match is found.
-    """
-    if not base_url:
-        return None
-    normalized_url = base_url.strip().rstrip("/")
-    for norm_name, entry in _iter_custom_providers():
-        entry_url = str(entry.get("base_url") or "").strip().rstrip("/")
-        if entry_url and entry_url == normalized_url:
-            return f"{CUSTOM_POOL_PREFIX}{norm_name}"
-    return None
-
-
-def list_custom_pool_providers() -> List[str]:
-    """Return all 'custom:*' pool keys that have entries in auth.json."""
-    pool_data = read_credential_pool(None)
-    return sorted(
-        key for key in pool_data
-        if key.startswith(CUSTOM_POOL_PREFIX)
-        and isinstance(pool_data.get(key), list)
-        and pool_data[key]
-    )
-
-
-def _get_custom_provider_config(pool_key: str) -> Optional[Dict[str, Any]]:
-    """Return the custom_providers config entry matching a pool key like 'custom:together.ai'."""
-    if not pool_key.startswith(CUSTOM_POOL_PREFIX):
-        return None
-    suffix = pool_key[len(CUSTOM_POOL_PREFIX):]
-    for norm_name, entry in _iter_custom_providers():
-        if norm_name == suffix:
-            return entry
-    return None
-
-
-def get_pool_strategy(provider: str) -> str:
-    """Return the configured selection strategy for a provider."""
-    config = _load_config_safe()
-    if config is None:
-        return STRATEGY_FILL_FIRST
-
-    strategies = config.get("credential_pool_strategies")
-    if not isinstance(strategies, dict):
-        return STRATEGY_FILL_FIRST
-
-    strategy = str(strategies.get(provider, "") or "").strip().lower()
-    if strategy in SUPPORTED_POOL_STRATEGIES:
-        return strategy
-    return STRATEGY_FILL_FIRST
-
-
-class CredentialPool:
-    def __init__(self, provider: str, entries: List[PooledCredential]):
-        self.provider = provider
-        self._entries = sorted(entries, key=lambda entry: entry.priority)
-        self._current_id: Optional[str] = None
-        self._strategy = get_pool_strategy(provider)
-        self._lock = threading.Lock()
-
-    def has_credentials(self) -> bool:
-        return bool(self._entries)
-
-    def has_available(self) -> bool:
-        """True if at least one entry is not currently in exhaustion cooldown."""
-        return bool(self._available_entries())
-
-    def entries(self) -> List[PooledCredential]:
-        return list(self._entries)
-
-    def current(self) -> Optional[PooledCredential]:
-        if not self._current_id:
-            return None
-        return next((entry for entry in self._entries if entry.id == self._current_id), None)
-
-    def _replace_entry(self, old: PooledCredential, new: PooledCredential) -> None:
-        """Swap an entry in-place by id, preserving sort order."""
-        for idx, entry in enumerate(self._entries):
-            if entry.id == old.id:
-                self._entries[idx] = new
-                return
-
-    def _persist(self) -> None:
-        write_credential_pool(
-            self.provider,
-            [entry.to_dict() for entry in self._entries],
-        )
-
-    def _mark_exhausted(self, entry: PooledCredential, status_code: Optional[int]) -> PooledCredential:
-        updated = replace(
-            entry,
-            last_status=STATUS_EXHAUSTED,
-            last_status_at=time.time(),
-            last_error_code=status_code,
-        )
-        self._replace_entry(entry, updated)
-        self._persist()
-        return updated
-
-    def _refresh_entry(self, entry: PooledCredential, *, force: bool) -> Optional[PooledCredential]:
-        if entry.auth_type != AUTH_TYPE_OAUTH or not entry.refresh_token:
-            if force:
-                self._mark_exhausted(entry, None)
-            return None
-
-        try:
-            if self.provider == "anthropic":
-                from agent.anthropic_adapter import refresh_anthropic_oauth_pure
-
-                refreshed = refresh_anthropic_oauth_pure(
-                    entry.refresh_token,
-                    use_json=entry.source.endswith("hermes_pkce"),
-                )
-                updated = replace(
-                    entry,
-                    access_token=refreshed["access_token"],
-                    refresh_token=refreshed["refresh_token"],
-                    expires_at_ms=refreshed["expires_at_ms"],
-                )
-            elif self.provider == "openai-codex":
-                refreshed = auth_mod.refresh_codex_oauth_pure(
-                    entry.access_token,
-                    entry.refresh_token,
-                )
-                updated = replace(
-                    entry,
-                    access_token=refreshed["access_token"],
-                    refresh_token=refreshed["refresh_token"],
-                    last_refresh=refreshed.get("last_refresh"),
-                )
-            elif self.provider == "nous":
-                nous_state = {
-                    "access_token": entry.access_token,
-                    "refresh_token": entry.refresh_token,
-                    "client_id": entry.client_id,
-                    "portal_base_url": entry.portal_base_url,
-                    "inference_base_url": entry.inference_base_url,
-                    "token_type": entry.token_type,
-                    "scope": entry.scope,
-                    "obtained_at": entry.obtained_at,
-                    "expires_at": entry.expires_at,
-                    "agent_key": entry.agent_key,
-                    "agent_key_expires_at": entry.agent_key_expires_at,
-                    "tls": entry.tls,
-                }
-                refreshed = auth_mod.refresh_nous_oauth_from_state(
-                    nous_state,
-                    min_key_ttl_seconds=DEFAULT_AGENT_KEY_MIN_TTL_SECONDS,
-                    force_refresh=force,
-                    force_mint=force,
-                )
-                # Apply returned fields: dataclass fields via replace, extras via dict update
-                field_updates = {}
-                extra_updates = dict(entry.extra)
-                _field_names = {f.name for f in fields(entry)}
-                for k, v in refreshed.items():
-                    if k in _field_names:
-                        field_updates[k] = v
-                    elif k in _EXTRA_KEYS:
-                        extra_updates[k] = v
-                updated = replace(entry, extra=extra_updates, **field_updates)
-            else:
-                return entry
-        except Exception as exc:
-            logger.debug("Credential refresh failed for %s/%s: %s", self.provider, entry.id, exc)
-            self._mark_exhausted(entry, None)
-            return None
-
-        updated = replace(updated, last_status=STATUS_OK, last_status_at=None, last_error_code=None)
-        self._replace_entry(entry, updated)
-        self._persist()
-        return updated
-
-    def _entry_needs_refresh(self, entry: PooledCredential) -> bool:
-        if entry.auth_type != AUTH_TYPE_OAUTH:
-            return False
-        if self.provider == "anthropic":
-            if entry.expires_at_ms is None:
-                return False
-            return int(entry.expires_at_ms) <= int(time.time() * 1000) + 120_000
-        if self.provider == "openai-codex":
-            return _codex_access_token_is_expiring(
-                entry.access_token,
-                CODEX_ACCESS_TOKEN_REFRESH_SKEW_SECONDS,
-            )
-        if self.provider == "nous":
-            # Nous refresh/mint can require network access and should happen when
-            # runtime credentials are actually resolved, not merely when the pool
-            # is enumerated for listing, migration, or selection.
-            return False
-        return False
-
-    def mark_used(self, entry_id: Optional[str] = None) -> None:
-        """Increment request_count for tracking. Used by least_used strategy."""
-        target_id = entry_id or self._current_id
-        if not target_id:
-            return
-        with self._lock:
-            for idx, entry in enumerate(self._entries):
-                if entry.id == target_id:
-                    self._entries[idx] = replace(entry, request_count=entry.request_count + 1)
-                    return
-
-    def select(self) -> Optional[PooledCredential]:
-        with self._lock:
-            return self._select_unlocked()
-
-    def _available_entries(self, *, clear_expired: bool = False, refresh: bool = False) -> List[PooledCredential]:
-        """Return entries not currently in exhaustion cooldown.
-
-        When *clear_expired* is True, entries whose cooldown has elapsed are
-        reset to STATUS_OK and persisted.  When *refresh* is True, entries
-        that need a token refresh are refreshed (skipped on failure).
-        """
-        now = time.time()
-        cleared_any = False
-        available: List[PooledCredential] = []
-        for entry in self._entries:
-            if entry.last_status == STATUS_EXHAUSTED:
-                ttl = _exhausted_ttl(entry.last_error_code)
-                if entry.last_status_at and now - entry.last_status_at < ttl:
-                    continue
-                if clear_expired:
-                    cleared = replace(entry, last_status=STATUS_OK, last_status_at=None, last_error_code=None)
-                    self._replace_entry(entry, cleared)
-                    entry = cleared
-                    cleared_any = True
-            if refresh and self._entry_needs_refresh(entry):
-                refreshed = self._refresh_entry(entry, force=False)
-                if refreshed is None:
-                    continue
-                entry = refreshed
-            available.append(entry)
-        if cleared_any:
-            self._persist()
-        return available
-
-    def _select_unlocked(self) -> Optional[PooledCredential]:
-        available = self._available_entries(clear_expired=True, refresh=True)
-        if not available:
-            self._current_id = None
-            return None
-
-        if self._strategy == STRATEGY_RANDOM:
-            entry = random.choice(available)
-            self._current_id = entry.id
-            return entry
-
-        if self._strategy == STRATEGY_LEAST_USED and len(available) > 1:
-            entry = min(available, key=lambda e: e.request_count)
-            self._current_id = entry.id
-            return entry
-
-        if self._strategy == STRATEGY_ROUND_ROBIN and len(available) > 1:
-            entry = available[0]
-            rotated = [candidate for candidate in self._entries if candidate.id != entry.id]
-            rotated.append(replace(entry, priority=len(self._entries) - 1))
-            self._entries = [replace(candidate, priority=idx) for idx, candidate in enumerate(rotated)]
-            self._persist()
-            self._current_id = entry.id
-            return self.current() or entry
-
-        entry = available[0]
-        self._current_id = entry.id
-        return entry
-
-    def peek(self) -> Optional[PooledCredential]:
-        current = self.current()
-        if current is not None:
-            return current
-        available = self._available_entries()
-        return available[0] if available else None
-
-    def mark_exhausted_and_rotate(self, *, status_code: Optional[int]) -> Optional[PooledCredential]:
-        with self._lock:
-            entry = self.current() or self._select_unlocked()
-            if entry is None:
-                return None
-            self._mark_exhausted(entry, status_code)
-            self._current_id = None
-            return self._select_unlocked()
-
-    def try_refresh_current(self) -> Optional[PooledCredential]:
-        with self._lock:
-            return self._try_refresh_current_unlocked()
-
-    def _try_refresh_current_unlocked(self) -> Optional[PooledCredential]:
-        entry = self.current()
-        if entry is None:
-            return None
-        refreshed = self._refresh_entry(entry, force=True)
-        if refreshed is not None:
-            self._current_id = refreshed.id
-        return refreshed
-
-    def reset_statuses(self) -> int:
-        count = 0
-        new_entries = []
-        for entry in self._entries:
-            if entry.last_status or entry.last_status_at or entry.last_error_code:
-                new_entries.append(replace(entry, last_status=None, last_status_at=None, last_error_code=None))
-                count += 1
-            else:
-                new_entries.append(entry)
-        if count:
-            self._entries = new_entries
-            self._persist()
-        return count
-
-    def remove_index(self, index: int) -> Optional[PooledCredential]:
-        if index < 1 or index > len(self._entries):
-            return None
-        removed = self._entries.pop(index - 1)
-        self._entries = [
-            replace(entry, priority=new_priority)
-            for new_priority, entry in enumerate(self._entries)
-        ]
-        self._persist()
-        if self._current_id == removed.id:
-            self._current_id = None
-        return removed
-
-    def add_entry(self, entry: PooledCredential) -> PooledCredential:
-        entry = replace(entry, priority=_next_priority(self._entries))
-        self._entries.append(entry)
-        self._persist()
-        return entry
-
-
-def _upsert_entry(entries: List[PooledCredential], provider: str, source: str, payload: Dict[str, Any]) -> bool:
-    existing_idx = None
-    for idx, entry in enumerate(entries):
-        if entry.source == source:
-            existing_idx = idx
-            break
-
-    if existing_idx is None:
-        payload.setdefault("id", uuid.uuid4().hex[:6])
-        payload.setdefault("priority", _next_priority(entries))
-        payload.setdefault("label", payload.get("label") or source)
-        entries.append(PooledCredential.from_dict(provider, payload))
-        return True
-
-    existing = entries[existing_idx]
-    field_updates = {}
-    extra_updates = {}
-    _field_names = {f.name for f in fields(existing)}
-    for key, value in payload.items():
-        if key in {"id", "priority"} or value is None:
-            continue
-        if key == "label" and existing.label:
-            continue
-        if key in _field_names:
-            if getattr(existing, key) != value:
-                field_updates[key] = value
-        elif key in _EXTRA_KEYS:
-            if existing.extra.get(key) != value:
-                extra_updates[key] = value
-    if field_updates or extra_updates:
-        if extra_updates:
-            field_updates["extra"] = {**existing.extra, **extra_updates}
-        entries[existing_idx] = replace(existing, **field_updates)
-        return True
-    return False
-
-
-def _normalize_pool_priorities(provider: str, entries: List[PooledCredential]) -> bool:
-    if provider != "anthropic":
-        return False
-
-    source_rank = {
-        "env:ANTHROPIC_TOKEN": 0,
-        "env:CLAUDE_CODE_OAUTH_TOKEN": 1,
-        "hermes_pkce": 2,
-        "claude_code": 3,
-        "env:ANTHROPIC_API_KEY": 4,
-    }
-    manual_entries = sorted(
-        (entry for entry in entries if _is_manual_source(entry.source)),
-        key=lambda entry: entry.priority,
-    )
-    seeded_entries = sorted(
-        (entry for entry in entries if not _is_manual_source(entry.source)),
-        key=lambda entry: (
-            source_rank.get(entry.source, len(source_rank)),
-            entry.priority,
-            entry.label,
-        ),
-    )
-
-    ordered = [*manual_entries, *seeded_entries]
-    id_to_idx = {entry.id: idx for idx, entry in enumerate(entries)}
-    changed = False
-    for new_priority, entry in enumerate(ordered):
-        if entry.priority != new_priority:
-            entries[id_to_idx[entry.id]] = replace(entry, priority=new_priority)
-            changed = True
-    return changed
-
-
-def _seed_from_singletons(provider: str, entries: List[PooledCredential]) -> Tuple[bool, Set[str]]:
-    changed = False
-    active_sources: Set[str] = set()
-    auth_store = _load_auth_store()
-
-    if provider == "anthropic":
-        from agent.anthropic_adapter import read_claude_code_credentials, read_hermes_oauth_credentials
-
-        for source_name, creds in (
-            ("hermes_pkce", read_hermes_oauth_credentials()),
-            ("claude_code", read_claude_code_credentials()),
-        ):
-            if creds and creds.get("accessToken"):
-                active_sources.add(source_name)
-                changed |= _upsert_entry(
-                    entries,
-                    provider,
-                    source_name,
-                    {
-                        "source": source_name,
-                        "auth_type": AUTH_TYPE_OAUTH,
-                        "access_token": creds.get("accessToken", ""),
-                        "refresh_token": creds.get("refreshToken"),
-                        "expires_at_ms": creds.get("expiresAt"),
-                        "label": label_from_token(creds.get("accessToken", ""), source_name),
-                    },
-                )
-
-    elif provider == "nous":
-        state = _load_provider_state(auth_store, "nous")
-        if state:
-            active_sources.add("device_code")
-            changed |= _upsert_entry(
-                entries,
-                provider,
-                "device_code",
-                {
-                    "source": "device_code",
-                    "auth_type": AUTH_TYPE_OAUTH,
-                    "access_token": state.get("access_token", ""),
-                    "refresh_token": state.get("refresh_token"),
-                    "expires_at": state.get("expires_at"),
-                    "token_type": state.get("token_type"),
-                    "scope": state.get("scope"),
-                    "client_id": state.get("client_id"),
-                    "portal_base_url": state.get("portal_base_url"),
-                    "inference_base_url": state.get("inference_base_url"),
-                    "agent_key": state.get("agent_key"),
-                    "agent_key_expires_at": state.get("agent_key_expires_at"),
-                    "tls": state.get("tls") if isinstance(state.get("tls"), dict) else None,
-                    "label": label_from_token(state.get("access_token", ""), "device_code"),
-                },
-            )
-
-    elif provider == "openai-codex":
-        state = _load_provider_state(auth_store, "openai-codex")
-        tokens = state.get("tokens") if isinstance(state, dict) else None
-        if isinstance(tokens, dict) and tokens.get("access_token"):
-            active_sources.add("device_code")
-            changed |= _upsert_entry(
-                entries,
-                provider,
-                "device_code",
-                {
-                    "source": "device_code",
-                    "auth_type": AUTH_TYPE_OAUTH,
-                    "access_token": tokens.get("access_token", ""),
-                    "refresh_token": tokens.get("refresh_token"),
-                    "base_url": "https://chatgpt.com/backend-api/codex",
-                    "last_refresh": state.get("last_refresh"),
-                    "label": label_from_token(tokens.get("access_token", ""), "device_code"),
-                },
-            )
-
-    return changed, active_sources
-
-
-def _seed_from_env(provider: str, entries: List[PooledCredential]) -> Tuple[bool, Set[str]]:
-    changed = False
-    active_sources: Set[str] = set()
-    if provider == "openrouter":
-        token = os.getenv("OPENROUTER_API_KEY", "").strip()
-        if token:
-            source = "env:OPENROUTER_API_KEY"
-            active_sources.add(source)
-            changed |= _upsert_entry(
-                entries,
-                provider,
-                source,
-                {
-                    "source": source,
-                    "auth_type": AUTH_TYPE_API_KEY,
-                    "access_token": token,
-                    "base_url": OPENROUTER_BASE_URL,
-                    "label": "OPENROUTER_API_KEY",
-                },
-            )
-        return changed, active_sources
-
-    pconfig = PROVIDER_REGISTRY.get(provider)
-    if not pconfig or pconfig.auth_type != AUTH_TYPE_API_KEY:
-        return changed, active_sources
-
-    env_url = ""
-    if pconfig.base_url_env_var:
-        env_url = os.getenv(pconfig.base_url_env_var, "").strip().rstrip("/")
-
-    env_vars = list(pconfig.api_key_env_vars)
-    if provider == "anthropic":
-        env_vars = [
-            "ANTHROPIC_TOKEN",
-            "CLAUDE_CODE_OAUTH_TOKEN",
-            "ANTHROPIC_API_KEY",
-        ]
-
-    for env_var in env_vars:
-        token = os.getenv(env_var, "").strip()
-        if not token:
-            continue
-        source = f"env:{env_var}"
-        active_sources.add(source)
-        auth_type = AUTH_TYPE_OAUTH if provider == "anthropic" and not token.startswith("sk-ant-api") else AUTH_TYPE_API_KEY
-        base_url = env_url or pconfig.inference_base_url
-        changed |= _upsert_entry(
-            entries,
-            provider,
-            source,
-            {
-                "source": source,
-                "auth_type": auth_type,
-                "access_token": token,
-                "base_url": base_url,
-                "label": env_var,
-            },
-        )
-    return changed, active_sources
-
-
-def _prune_stale_seeded_entries(entries: List[PooledCredential], active_sources: Set[str]) -> bool:
-    retained = [
-        entry
-        for entry in entries
-        if _is_manual_source(entry.source)
-        or entry.source in active_sources
-        or not (
-            entry.source.startswith("env:")
-            or entry.source in {"claude_code", "hermes_pkce"}
-        )
-    ]
-    if len(retained) == len(entries):
-        return False
-    entries[:] = retained
-    return True
-
-
-def _seed_custom_pool(pool_key: str, entries: List[PooledCredential]) -> Tuple[bool, Set[str]]:
-    """Seed a custom endpoint pool from custom_providers config and model config."""
-    changed = False
-    active_sources: Set[str] = set()
-
-    # Seed from the custom_providers config entry's api_key field
-    cp_config = _get_custom_provider_config(pool_key)
-    if cp_config:
-        api_key = str(cp_config.get("api_key") or "").strip()
-        base_url = str(cp_config.get("base_url") or "").strip().rstrip("/")
-        name = str(cp_config.get("name") or "").strip()
-        if api_key:
-            source = f"config:{name}"
-            active_sources.add(source)
-            changed |= _upsert_entry(
-                entries,
-                pool_key,
-                source,
-                {
-                    "source": source,
-                    "auth_type": AUTH_TYPE_API_KEY,
-                    "access_token": api_key,
-                    "base_url": base_url,
-                    "label": name or source,
-                },
-            )
-
-    # Seed from model.api_key if model.provider=='custom' and model.base_url matches
-    try:
-        config = _load_config_safe()
-        model_cfg = config.get("model") if config else None
-        if isinstance(model_cfg, dict):
-            model_provider = str(model_cfg.get("provider") or "").strip().lower()
-            model_base_url = str(model_cfg.get("base_url") or "").strip().rstrip("/")
-            model_api_key = ""
-            for k in ("api_key", "api"):
-                v = model_cfg.get(k)
-                if isinstance(v, str) and v.strip():
-                    model_api_key = v.strip()
-                    break
-            if model_provider == "custom" and model_base_url and model_api_key:
-                # Check if this model's base_url matches our custom provider
-                matched_key = get_custom_provider_pool_key(model_base_url)
-                if matched_key == pool_key:
-                    source = "model_config"
-                    active_sources.add(source)
-                    changed |= _upsert_entry(
-                        entries,
-                        pool_key,
-                        source,
-                        {
-                            "source": source,
-                            "auth_type": AUTH_TYPE_API_KEY,
-                            "access_token": model_api_key,
-                            "base_url": model_base_url,
-                            "label": "model_config",
-                        },
-                    )
-    except Exception:
-        pass
-
-    return changed, active_sources
-
-
-def load_pool(provider: str) -> CredentialPool:
-    provider = (provider or "").strip().lower()
-    raw_entries = read_credential_pool(provider)
-    entries = [PooledCredential.from_dict(provider, payload) for payload in raw_entries]
-
-    if provider.startswith(CUSTOM_POOL_PREFIX):
-        # Custom endpoint pool — seed from custom_providers config and model config
-        custom_changed, custom_sources = _seed_custom_pool(provider, entries)
-        changed = custom_changed
-        changed |= _prune_stale_seeded_entries(entries, custom_sources)
-    else:
-        singleton_changed, singleton_sources = _seed_from_singletons(provider, entries)
-        env_changed, env_sources = _seed_from_env(provider, entries)
-        changed = singleton_changed or env_changed
-        changed |= _prune_stale_seeded_entries(entries, singleton_sources | env_sources)
-        changed |= _normalize_pool_priorities(provider, entries)
-
-    if changed:
-        write_credential_pool(
-            provider,
-            [entry.to_dict() for entry in sorted(entries, key=lambda item: item.priority)],
-        )
-    return CredentialPool(provider, entries)

agent/insights.py

@@ -1,799 +0,0 @@
-"""
-Session Insights Engine for Hermes Agent.
-
-Analyzes historical session data from the SQLite state database to produce
-comprehensive usage insights — token consumption, cost estimates, tool usage
-patterns, activity trends, model/platform breakdowns, and session metrics.
-
-Inspired by Claude Code's /insights command, adapted for Hermes Agent's
-multi-platform architecture with additional cost estimation and platform
-breakdown capabilities.
-
-Usage:
-    from agent.insights import InsightsEngine
-    engine = InsightsEngine(db)
-    report = engine.generate(days=30)
-    print(engine.format_terminal(report))
-"""
-
-import json
-import time
-from collections import Counter, defaultdict
-from datetime import datetime
-from typing import Any, Dict, List
-
-from agent.usage_pricing import (
-    CanonicalUsage,
-    DEFAULT_PRICING,
-    estimate_usage_cost,
-    format_duration_compact,
-    get_pricing,
-    has_known_pricing,
-)
-
-_DEFAULT_PRICING = DEFAULT_PRICING
-
-
-def _has_known_pricing(model_name: str, provider: str = None, base_url: str = None) -> bool:
-    """Check if a model has known pricing (vs unknown/custom endpoint)."""
-    return has_known_pricing(model_name, provider=provider, base_url=base_url)
-
-
-def _get_pricing(model_name: str) -> Dict[str, float]:
-    """Look up pricing for a model. Uses fuzzy matching on model name.
-
-    Returns _DEFAULT_PRICING (zero cost) for unknown/custom models —
-    we can't assume costs for self-hosted endpoints, local inference, etc.
-    """
-    return get_pricing(model_name)
-
-
-def _estimate_cost(
-    session_or_model: Dict[str, Any] | str,
-    input_tokens: int = 0,
-    output_tokens: int = 0,
-    *,
-    cache_read_tokens: int = 0,
-    cache_write_tokens: int = 0,
-    provider: str = None,
-    base_url: str = None,
-) -> tuple[float, str]:
-    """Estimate the USD cost for a session row or a model/token tuple."""
-    if isinstance(session_or_model, dict):
-        session = session_or_model
-        model = session.get("model") or ""
-        usage = CanonicalUsage(
-            input_tokens=session.get("input_tokens") or 0,
-            output_tokens=session.get("output_tokens") or 0,
-            cache_read_tokens=session.get("cache_read_tokens") or 0,
-            cache_write_tokens=session.get("cache_write_tokens") or 0,
-        )
-        provider = session.get("billing_provider")
-        base_url = session.get("billing_base_url")
-    else:
-        model = session_or_model or ""
-        usage = CanonicalUsage(
-            input_tokens=input_tokens,
-            output_tokens=output_tokens,
-            cache_read_tokens=cache_read_tokens,
-            cache_write_tokens=cache_write_tokens,
-        )
-    result = estimate_usage_cost(
-        model,
-        usage,
-        provider=provider,
-        base_url=base_url,
-    )
-    return float(result.amount_usd or 0.0), result.status
-
-
-def _format_duration(seconds: float) -> str:
-    """Format seconds into a human-readable duration string."""
-    return format_duration_compact(seconds)
-
-
-def _bar_chart(values: List[int], max_width: int = 20) -> List[str]:
-    """Create simple horizontal bar chart strings from values."""
-    peak = max(values) if values else 1
-    if peak == 0:
-        return ["" for _ in values]
-    return ["█" * max(1, int(v / peak * max_width)) if v > 0 else "" for v in values]
-
-
-class InsightsEngine:
-    """
-    Analyzes session history and produces usage insights.
-
-    Works directly with a SessionDB instance (or raw sqlite3 connection)
-    to query session and message data.
-    """
-
-    def __init__(self, db):
-        """
-        Initialize with a SessionDB instance.
-
-        Args:
-            db: A SessionDB instance (from hermes_state.py)
-        """
-        self.db = db
-        self._conn = db._conn
-
-    def generate(self, days: int = 30, source: str = None) -> Dict[str, Any]:
-        """
-        Generate a complete insights report.
-
-        Args:
-            days: Number of days to look back (default: 30)
-            source: Optional filter by source platform
-
-        Returns:
-            Dict with all computed insights
-        """
-        cutoff = time.time() - (days * 86400)
-
-        # Gather raw data
-        sessions = self._get_sessions(cutoff, source)
-        tool_usage = self._get_tool_usage(cutoff, source)
-        message_stats = self._get_message_stats(cutoff, source)
-
-        if not sessions:
-            return {
-                "days": days,
-                "source_filter": source,
-                "empty": True,
-                "overview": {},
-                "models": [],
-                "platforms": [],
-                "tools": [],
-                "activity": {},
-                "top_sessions": [],
-            }
-
-        # Compute insights
-        overview = self._compute_overview(sessions, message_stats)
-        models = self._compute_model_breakdown(sessions)
-        platforms = self._compute_platform_breakdown(sessions)
-        tools = self._compute_tool_breakdown(tool_usage)
-        activity = self._compute_activity_patterns(sessions)
-        top_sessions = self._compute_top_sessions(sessions)
-
-        return {
-            "days": days,
-            "source_filter": source,
-            "empty": False,
-            "generated_at": time.time(),
-            "overview": overview,
-            "models": models,
-            "platforms": platforms,
-            "tools": tools,
-            "activity": activity,
-            "top_sessions": top_sessions,
-        }
-
-    # =========================================================================
-    # Data gathering (SQL queries)
-    # =========================================================================
-
-    # Columns we actually need (skip system_prompt, model_config blobs)
-    _SESSION_COLS = ("id, source, model, started_at, ended_at, "
-                     "message_count, tool_call_count, input_tokens, output_tokens, "
-                     "cache_read_tokens, cache_write_tokens, billing_provider, "
-                     "billing_base_url, billing_mode, estimated_cost_usd, "
-                     "actual_cost_usd, cost_status, cost_source")
-
-    # Pre-computed query strings — f-string evaluated once at class definition,
-    # not at runtime, so no user-controlled value can alter the query structure.
-    _GET_SESSIONS_WITH_SOURCE = (
-        f"SELECT {_SESSION_COLS} FROM sessions"
-        " WHERE started_at >= ? AND source = ?"
-        " ORDER BY started_at DESC"
-    )
-    _GET_SESSIONS_ALL = (
-        f"SELECT {_SESSION_COLS} FROM sessions"
-        " WHERE started_at >= ?"
-        " ORDER BY started_at DESC"
-    )
-
-    def _get_sessions(self, cutoff: float, source: str = None) -> List[Dict]:
-        """Fetch sessions within the time window."""
-        if source:
-            cursor = self._conn.execute(self._GET_SESSIONS_WITH_SOURCE, (cutoff, source))
-        else:
-            cursor = self._conn.execute(self._GET_SESSIONS_ALL, (cutoff,))
-        return [dict(row) for row in cursor.fetchall()]
-
-    def _get_tool_usage(self, cutoff: float, source: str = None) -> List[Dict]:
-        """Get tool call counts from messages.
-
-        Uses two sources:
-        1. tool_name column on 'tool' role messages (set by gateway)
-        2. tool_calls JSON on 'assistant' role messages (covers CLI where
-           tool_name is not populated on tool responses)
-        """
-        tool_counts = Counter()
-
-        # Source 1: explicit tool_name on tool response messages
-        if source:
-            cursor = self._conn.execute(
-                """SELECT m.tool_name, COUNT(*) as count
-                   FROM messages m
-                   JOIN sessions s ON s.id = m.session_id
-                   WHERE s.started_at >= ? AND s.source = ?
-                     AND m.role = 'tool' AND m.tool_name IS NOT NULL
-                   GROUP BY m.tool_name
-                   ORDER BY count DESC""",
-                (cutoff, source),
-            )
-        else:
-            cursor = self._conn.execute(
-                """SELECT m.tool_name, COUNT(*) as count
-                   FROM messages m
-                   JOIN sessions s ON s.id = m.session_id
-                   WHERE s.started_at >= ?
-                     AND m.role = 'tool' AND m.tool_name IS NOT NULL
-                   GROUP BY m.tool_name
-                   ORDER BY count DESC""",
-                (cutoff,),
-            )
-        for row in cursor.fetchall():
-            tool_counts[row["tool_name"]] += row["count"]
-
-        # Source 2: extract from tool_calls JSON on assistant messages
-        # (covers CLI sessions where tool_name is NULL on tool responses)
-        if source:
-            cursor2 = self._conn.execute(
-                """SELECT m.tool_calls
-                   FROM messages m
-                   JOIN sessions s ON s.id = m.session_id
-                   WHERE s.started_at >= ? AND s.source = ?
-                     AND m.role = 'assistant' AND m.tool_calls IS NOT NULL""",
-                (cutoff, source),
-            )
-        else:
-            cursor2 = self._conn.execute(
-                """SELECT m.tool_calls
-                   FROM messages m
-                   JOIN sessions s ON s.id = m.session_id
-                   WHERE s.started_at >= ?
-                     AND m.role = 'assistant' AND m.tool_calls IS NOT NULL""",
-                (cutoff,),
-            )
-
-        tool_calls_counts = Counter()
-        for row in cursor2.fetchall():
-            try:
-                calls = row["tool_calls"]
-                if isinstance(calls, str):
-                    calls = json.loads(calls)
-                if isinstance(calls, list):
-                    for call in calls:
-                        func = call.get("function", {}) if isinstance(call, dict) else {}
-                        name = func.get("name")
-                        if name:
-                            tool_calls_counts[name] += 1
-            except (json.JSONDecodeError, TypeError, AttributeError):
-                continue
-
-        # Merge: prefer tool_name source, supplement with tool_calls source
-        # for tools not already counted
-        if not tool_counts and tool_calls_counts:
-            # No tool_name data at all — use tool_calls exclusively
-            tool_counts = tool_calls_counts
-        elif tool_counts and tool_calls_counts:
-            # Both sources have data — use whichever has the higher count per tool
-            # (they may overlap, so take the max to avoid double-counting)
-            all_tools = set(tool_counts) | set(tool_calls_counts)
-            merged = Counter()
-            for tool in all_tools:
-                merged[tool] = max(tool_counts.get(tool, 0), tool_calls_counts.get(tool, 0))
-            tool_counts = merged
-
-        # Convert to the expected format
-        return [
-            {"tool_name": name, "count": count}
-            for name, count in tool_counts.most_common()
-        ]
-
-    def _get_message_stats(self, cutoff: float, source: str = None) -> Dict:
-        """Get aggregate message statistics."""
-        if source:
-            cursor = self._conn.execute(
-                """SELECT
-                     COUNT(*) as total_messages,
-                     SUM(CASE WHEN m.role = 'user' THEN 1 ELSE 0 END) as user_messages,
-                     SUM(CASE WHEN m.role = 'assistant' THEN 1 ELSE 0 END) as assistant_messages,
-                     SUM(CASE WHEN m.role = 'tool' THEN 1 ELSE 0 END) as tool_messages
-                   FROM messages m
-                   JOIN sessions s ON s.id = m.session_id
-                   WHERE s.started_at >= ? AND s.source = ?""",
-                (cutoff, source),
-            )
-        else:
-            cursor = self._conn.execute(
-                """SELECT
-                     COUNT(*) as total_messages,
-                     SUM(CASE WHEN m.role = 'user' THEN 1 ELSE 0 END) as user_messages,
-                     SUM(CASE WHEN m.role = 'assistant' THEN 1 ELSE 0 END) as assistant_messages,
-                     SUM(CASE WHEN m.role = 'tool' THEN 1 ELSE 0 END) as tool_messages
-                   FROM messages m
-                   JOIN sessions s ON s.id = m.session_id
-                   WHERE s.started_at >= ?""",
-                (cutoff,),
-            )
-        row = cursor.fetchone()
-        return dict(row) if row else {
-            "total_messages": 0, "user_messages": 0,
-            "assistant_messages": 0, "tool_messages": 0,
-        }
-
-    # =========================================================================
-    # Computation
-    # =========================================================================
-
-    def _compute_overview(self, sessions: List[Dict], message_stats: Dict) -> Dict:
-        """Compute high-level overview statistics."""
-        total_input = sum(s.get("input_tokens") or 0 for s in sessions)
-        total_output = sum(s.get("output_tokens") or 0 for s in sessions)
-        total_cache_read = sum(s.get("cache_read_tokens") or 0 for s in sessions)
-        total_cache_write = sum(s.get("cache_write_tokens") or 0 for s in sessions)
-        total_tokens = total_input + total_output + total_cache_read + total_cache_write
-        total_tool_calls = sum(s.get("tool_call_count") or 0 for s in sessions)
-        total_messages = sum(s.get("message_count") or 0 for s in sessions)
-
-        # Cost estimation (weighted by model)
-        total_cost = 0.0
-        actual_cost = 0.0
-        models_with_pricing = set()
-        models_without_pricing = set()
-        unknown_cost_sessions = 0
-        included_cost_sessions = 0
-        for s in sessions:
-            model = s.get("model") or ""
-            estimated, status = _estimate_cost(s)
-            total_cost += estimated
-            actual_cost += s.get("actual_cost_usd") or 0.0
-            display = model.split("/")[-1] if "/" in model else (model or "unknown")
-            if status == "included":
-                included_cost_sessions += 1
-            elif status == "unknown":
-                unknown_cost_sessions += 1
-            if _has_known_pricing(model, s.get("billing_provider"), s.get("billing_base_url")):
-                models_with_pricing.add(display)
-            else:
-                models_without_pricing.add(display)
-
-        # Session duration stats (guard against negative durations from clock drift)
-        durations = []
-        for s in sessions:
-            start = s.get("started_at")
-            end = s.get("ended_at")
-            if start and end and end > start:
-                durations.append(end - start)
-
-        total_hours = sum(durations) / 3600 if durations else 0
-        avg_duration = sum(durations) / len(durations) if durations else 0
-
-        # Earliest and latest session
-        started_timestamps = [s["started_at"] for s in sessions if s.get("started_at")]
-        date_range_start = min(started_timestamps) if started_timestamps else None
-        date_range_end = max(started_timestamps) if started_timestamps else None
-
-        return {
-            "total_sessions": len(sessions),
-            "total_messages": total_messages,
-            "total_tool_calls": total_tool_calls,
-            "total_input_tokens": total_input,
-            "total_output_tokens": total_output,
-            "total_cache_read_tokens": total_cache_read,
-            "total_cache_write_tokens": total_cache_write,
-            "total_tokens": total_tokens,
-            "estimated_cost": total_cost,
-            "actual_cost": actual_cost,
-            "total_hours": total_hours,
-            "avg_session_duration": avg_duration,
-            "avg_messages_per_session": total_messages / len(sessions) if sessions else 0,
-            "avg_tokens_per_session": total_tokens / len(sessions) if sessions else 0,
-            "user_messages": message_stats.get("user_messages") or 0,
-            "assistant_messages": message_stats.get("assistant_messages") or 0,
-            "tool_messages": message_stats.get("tool_messages") or 0,
-            "date_range_start": date_range_start,
-            "date_range_end": date_range_end,
-            "models_with_pricing": sorted(models_with_pricing),
-            "models_without_pricing": sorted(models_without_pricing),
-            "unknown_cost_sessions": unknown_cost_sessions,
-            "included_cost_sessions": included_cost_sessions,
-        }
-
-    def _compute_model_breakdown(self, sessions: List[Dict]) -> List[Dict]:
-        """Break down usage by model."""
-        model_data = defaultdict(lambda: {
-            "sessions": 0, "input_tokens": 0, "output_tokens": 0,
-            "cache_read_tokens": 0, "cache_write_tokens": 0,
-            "total_tokens": 0, "tool_calls": 0, "cost": 0.0,
-        })
-
-        for s in sessions:
-            model = s.get("model") or "unknown"
-            # Normalize: strip provider prefix for display
-            display_model = model.split("/")[-1] if "/" in model else model
-            d = model_data[display_model]
-            d["sessions"] += 1
-            inp = s.get("input_tokens") or 0
-            out = s.get("output_tokens") or 0
-            cache_read = s.get("cache_read_tokens") or 0
-            cache_write = s.get("cache_write_tokens") or 0
-            d["input_tokens"] += inp
-            d["output_tokens"] += out
-            d["cache_read_tokens"] += cache_read
-            d["cache_write_tokens"] += cache_write
-            d["total_tokens"] += inp + out + cache_read + cache_write
-            d["tool_calls"] += s.get("tool_call_count") or 0
-            estimate, status = _estimate_cost(s)
-            d["cost"] += estimate
-            d["has_pricing"] = _has_known_pricing(model, s.get("billing_provider"), s.get("billing_base_url"))
-            d["cost_status"] = status
-
-        result = [
-            {"model": model, **data}
-            for model, data in model_data.items()
-        ]
-        # Sort by tokens first, fall back to session count when tokens are 0
-        result.sort(key=lambda x: (x["total_tokens"], x["sessions"]), reverse=True)
-        return result
-
-    def _compute_platform_breakdown(self, sessions: List[Dict]) -> List[Dict]:
-        """Break down usage by platform/source."""
-        platform_data = defaultdict(lambda: {
-            "sessions": 0, "messages": 0, "input_tokens": 0,
-            "output_tokens": 0, "cache_read_tokens": 0,
-            "cache_write_tokens": 0, "total_tokens": 0, "tool_calls": 0,
-        })
-
-        for s in sessions:
-            source = s.get("source") or "unknown"
-            d = platform_data[source]
-            d["sessions"] += 1
-            d["messages"] += s.get("message_count") or 0
-            inp = s.get("input_tokens") or 0
-            out = s.get("output_tokens") or 0
-            cache_read = s.get("cache_read_tokens") or 0
-            cache_write = s.get("cache_write_tokens") or 0
-            d["input_tokens"] += inp
-            d["output_tokens"] += out
-            d["cache_read_tokens"] += cache_read
-            d["cache_write_tokens"] += cache_write
-            d["total_tokens"] += inp + out + cache_read + cache_write
-            d["tool_calls"] += s.get("tool_call_count") or 0
-
-        result = [
-            {"platform": platform, **data}
-            for platform, data in platform_data.items()
-        ]
-        result.sort(key=lambda x: x["sessions"], reverse=True)
-        return result
-
-    def _compute_tool_breakdown(self, tool_usage: List[Dict]) -> List[Dict]:
-        """Process tool usage data into a ranked list with percentages."""
-        total_calls = sum(t["count"] for t in tool_usage) if tool_usage else 0
-        result = []
-        for t in tool_usage:
-            pct = (t["count"] / total_calls * 100) if total_calls else 0
-            result.append({
-                "tool": t["tool_name"],
-                "count": t["count"],
-                "percentage": pct,
-            })
-        return result
-
-    def _compute_activity_patterns(self, sessions: List[Dict]) -> Dict:
-        """Analyze activity patterns by day of week and hour."""
-        day_counts = Counter()  # 0=Monday ... 6=Sunday
-        hour_counts = Counter()
-        daily_counts = Counter()  # date string -> count
-
-        for s in sessions:
-            ts = s.get("started_at")
-            if not ts:
-                continue
-            dt = datetime.fromtimestamp(ts)
-            day_counts[dt.weekday()] += 1
-            hour_counts[dt.hour] += 1
-            daily_counts[dt.strftime("%Y-%m-%d")] += 1
-
-        day_names = ["Mon", "Tue", "Wed", "Thu", "Fri", "Sat", "Sun"]
-        day_breakdown = [
-            {"day": day_names[i], "count": day_counts.get(i, 0)}
-            for i in range(7)
-        ]
-
-        hour_breakdown = [
-            {"hour": i, "count": hour_counts.get(i, 0)}
-            for i in range(24)
-        ]
-
-        # Busiest day and hour
-        busiest_day = max(day_breakdown, key=lambda x: x["count"]) if day_breakdown else None
-        busiest_hour = max(hour_breakdown, key=lambda x: x["count"]) if hour_breakdown else None
-
-        # Active days (days with at least one session)
-        active_days = len(daily_counts)
-
-        # Streak calculation
-        if daily_counts:
-            all_dates = sorted(daily_counts.keys())
-            current_streak = 1
-            max_streak = 1
-            for i in range(1, len(all_dates)):
-                d1 = datetime.strptime(all_dates[i - 1], "%Y-%m-%d")
-                d2 = datetime.strptime(all_dates[i], "%Y-%m-%d")
-                if (d2 - d1).days == 1:
-                    current_streak += 1
-                    max_streak = max(max_streak, current_streak)
-                else:
-                    current_streak = 1
-        else:
-            max_streak = 0
-
-        return {
-            "by_day": day_breakdown,
-            "by_hour": hour_breakdown,
-            "busiest_day": busiest_day,
-            "busiest_hour": busiest_hour,
-            "active_days": active_days,
-            "max_streak": max_streak,
-        }
-
-    def _compute_top_sessions(self, sessions: List[Dict]) -> List[Dict]:
-        """Find notable sessions (longest, most messages, most tokens)."""
-        top = []
-
-        # Longest by duration
-        sessions_with_duration = [
-            s for s in sessions
-            if s.get("started_at") and s.get("ended_at")
-        ]
-        if sessions_with_duration:
-            longest = max(
-                sessions_with_duration,
-                key=lambda s: (s["ended_at"] - s["started_at"]),
-            )
-            dur = longest["ended_at"] - longest["started_at"]
-            top.append({
-                "label": "Longest session",
-                "session_id": longest["id"][:16],
-                "value": _format_duration(dur),
-                "date": datetime.fromtimestamp(longest["started_at"]).strftime("%b %d"),
-            })
-
-        # Most messages
-        most_msgs = max(sessions, key=lambda s: s.get("message_count") or 0)
-        if (most_msgs.get("message_count") or 0) > 0:
-            top.append({
-                "label": "Most messages",
-                "session_id": most_msgs["id"][:16],
-                "value": f"{most_msgs['message_count']} msgs",
-                "date": datetime.fromtimestamp(most_msgs["started_at"]).strftime("%b %d") if most_msgs.get("started_at") else "?",
-            })
-
-        # Most tokens
-        most_tokens = max(
-            sessions,
-            key=lambda s: (s.get("input_tokens") or 0) + (s.get("output_tokens") or 0),
-        )
-        token_total = (most_tokens.get("input_tokens") or 0) + (most_tokens.get("output_tokens") or 0)
-        if token_total > 0:
-            top.append({
-                "label": "Most tokens",
-                "session_id": most_tokens["id"][:16],
-                "value": f"{token_total:,} tokens",
-                "date": datetime.fromtimestamp(most_tokens["started_at"]).strftime("%b %d") if most_tokens.get("started_at") else "?",
-            })
-
-        # Most tool calls
-        most_tools = max(sessions, key=lambda s: s.get("tool_call_count") or 0)
-        if (most_tools.get("tool_call_count") or 0) > 0:
-            top.append({
-                "label": "Most tool calls",
-                "session_id": most_tools["id"][:16],
-                "value": f"{most_tools['tool_call_count']} calls",
-                "date": datetime.fromtimestamp(most_tools["started_at"]).strftime("%b %d") if most_tools.get("started_at") else "?",
-            })
-
-        return top
-
-    # =========================================================================
-    # Formatting
-    # =========================================================================
-
-    def format_terminal(self, report: Dict) -> str:
-        """Format the insights report for terminal display (CLI)."""
-        if report.get("empty"):
-            days = report.get("days", 30)
-            src = f" (source: {report['source_filter']})" if report.get("source_filter") else ""
-            return f"  No sessions found in the last {days} days{src}."
-
-        lines = []
-        o = report["overview"]
-        days = report["days"]
-        src_filter = report.get("source_filter")
-
-        # Header
-        lines.append("")
-        lines.append("  ╔══════════════════════════════════════════════════════════╗")
-        lines.append("  ║                    📊 Hermes Insights                    ║")
-        period_label = f"Last {days} days"
-        if src_filter:
-            period_label += f" ({src_filter})"
-        padding = 58 - len(period_label) - 2
-        left_pad = padding // 2
-        right_pad = padding - left_pad
-        lines.append(f"  ║{' ' * left_pad} {period_label} {' ' * right_pad}║")
-        lines.append("  ╚══════════════════════════════════════════════════════════╝")
-        lines.append("")
-
-        # Date range
-        if o.get("date_range_start") and o.get("date_range_end"):
-            start_str = datetime.fromtimestamp(o["date_range_start"]).strftime("%b %d, %Y")
-            end_str = datetime.fromtimestamp(o["date_range_end"]).strftime("%b %d, %Y")
-            lines.append(f"  Period: {start_str} — {end_str}")
-            lines.append("")
-
-        # Overview
-        lines.append("  📋 Overview")
-        lines.append("  " + "─" * 56)
-        lines.append(f"  Sessions:          {o['total_sessions']:<12}  Messages:        {o['total_messages']:,}")
-        lines.append(f"  Tool calls:        {o['total_tool_calls']:<12,}  User messages:   {o['user_messages']:,}")
-        lines.append(f"  Input tokens:      {o['total_input_tokens']:<12,}  Output tokens:   {o['total_output_tokens']:,}")
-        cache_total = o.get("total_cache_read_tokens", 0) + o.get("total_cache_write_tokens", 0)
-        if cache_total > 0:
-            lines.append(f"  Cache read:        {o['total_cache_read_tokens']:<12,}  Cache write:     {o['total_cache_write_tokens']:,}")
-        cost_str = f"${o['estimated_cost']:.2f}"
-        if o.get("models_without_pricing"):
-            cost_str += " *"
-        lines.append(f"  Total tokens:      {o['total_tokens']:<12,}  Est. cost:       {cost_str}")
-        if o["total_hours"] > 0:
-            lines.append(f"  Active time:       ~{_format_duration(o['total_hours'] * 3600):<11}  Avg session:     ~{_format_duration(o['avg_session_duration'])}")
-        lines.append(f"  Avg msgs/session:  {o['avg_messages_per_session']:.1f}")
-        lines.append("")
-
-        # Model breakdown
-        if report["models"]:
-            lines.append("  🤖 Models Used")
-            lines.append("  " + "─" * 56)
-            lines.append(f"  {'Model':<30} {'Sessions':>8} {'Tokens':>12} {'Cost':>8}")
-            for m in report["models"]:
-                model_name = m["model"][:28]
-                if m.get("has_pricing"):
-                    cost_cell = f"${m['cost']:>6.2f}"
-                else:
-                    cost_cell = "     N/A"
-                lines.append(f"  {model_name:<30} {m['sessions']:>8} {m['total_tokens']:>12,} {cost_cell}")
-            if o.get("models_without_pricing"):
-                lines.append("  * Cost N/A for custom/self-hosted models")
-            lines.append("")
-
-        # Platform breakdown
-        if len(report["platforms"]) > 1 or (report["platforms"] and report["platforms"][0]["platform"] != "cli"):
-            lines.append("  📱 Platforms")
-            lines.append("  " + "─" * 56)
-            lines.append(f"  {'Platform':<14} {'Sessions':>8} {'Messages':>10} {'Tokens':>14}")
-            for p in report["platforms"]:
-                lines.append(f"  {p['platform']:<14} {p['sessions']:>8} {p['messages']:>10,} {p['total_tokens']:>14,}")
-            lines.append("")
-
-        # Tool usage
-        if report["tools"]:
-            lines.append("  🔧 Top Tools")
-            lines.append("  " + "─" * 56)
-            lines.append(f"  {'Tool':<28} {'Calls':>8} {'%':>8}")
-            for t in report["tools"][:15]:  # Top 15
-                lines.append(f"  {t['tool']:<28} {t['count']:>8,} {t['percentage']:>7.1f}%")
-            if len(report["tools"]) > 15:
-                lines.append(f"  ... and {len(report['tools']) - 15} more tools")
-            lines.append("")
-
-        # Activity patterns
-        act = report.get("activity", {})
-        if act.get("by_day"):
-            lines.append("  📅 Activity Patterns")
-            lines.append("  " + "─" * 56)
-
-            # Day of week chart
-            day_values = [d["count"] for d in act["by_day"]]
-            bars = _bar_chart(day_values, max_width=15)
-            for i, d in enumerate(act["by_day"]):
-                bar = bars[i]
-                lines.append(f"  {d['day']}  {bar:<15} {d['count']}")
-
-            lines.append("")
-
-            # Peak hours (show top 5 busiest hours)
-            busy_hours = sorted(act["by_hour"], key=lambda x: x["count"], reverse=True)
-            busy_hours = [h for h in busy_hours if h["count"] > 0][:5]
-            if busy_hours:
-                hour_strs = []
-                for h in busy_hours:
-                    hr = h["hour"]
-                    ampm = "AM" if hr < 12 else "PM"
-                    display_hr = hr % 12 or 12
-                    hour_strs.append(f"{display_hr}{ampm} ({h['count']})")
-                lines.append(f"  Peak hours: {', '.join(hour_strs)}")
-
-            if act.get("active_days"):
-                lines.append(f"  Active days: {act['active_days']}")
-            if act.get("max_streak") and act["max_streak"] > 1:
-                lines.append(f"  Best streak: {act['max_streak']} consecutive days")
-            lines.append("")
-
-        # Notable sessions
-        if report.get("top_sessions"):
-            lines.append("  🏆 Notable Sessions")
-            lines.append("  " + "─" * 56)
-            for ts in report["top_sessions"]:
-                lines.append(f"  {ts['label']:<20} {ts['value']:<18} ({ts['date']}, {ts['session_id']})")
-            lines.append("")
-
-        return "\n".join(lines)
-
-    def format_gateway(self, report: Dict) -> str:
-        """Format the insights report for gateway/messaging (shorter)."""
-        if report.get("empty"):
-            days = report.get("days", 30)
-            return f"No sessions found in the last {days} days."
-
-        lines = []
-        o = report["overview"]
-        days = report["days"]
-
-        lines.append(f"📊 **Hermes Insights** — Last {days} days\n")
-
-        # Overview
-        lines.append(f"**Sessions:** {o['total_sessions']} | **Messages:** {o['total_messages']:,} | **Tool calls:** {o['total_tool_calls']:,}")
-        cache_total = o.get("total_cache_read_tokens", 0) + o.get("total_cache_write_tokens", 0)
-        if cache_total > 0:
-            lines.append(f"**Tokens:** {o['total_tokens']:,} (in: {o['total_input_tokens']:,} / out: {o['total_output_tokens']:,} / cache: {cache_total:,})")
-        else:
-            lines.append(f"**Tokens:** {o['total_tokens']:,} (in: {o['total_input_tokens']:,} / out: {o['total_output_tokens']:,})")
-        cost_note = ""
-        if o.get("models_without_pricing"):
-            cost_note = " _(excludes custom/self-hosted models)_"
-        lines.append(f"**Est. cost:** ${o['estimated_cost']:.2f}{cost_note}")
-        if o["total_hours"] > 0:
-            lines.append(f"**Active time:** ~{_format_duration(o['total_hours'] * 3600)} | **Avg session:** ~{_format_duration(o['avg_session_duration'])}")
-        lines.append("")
-
-        # Models (top 5)
-        if report["models"]:
-            lines.append("**🤖 Models:**")
-            for m in report["models"][:5]:
-                cost_str = f"${m['cost']:.2f}" if m.get("has_pricing") else "N/A"
-                lines.append(f"  {m['model'][:25]} — {m['sessions']} sessions, {m['total_tokens']:,} tokens, {cost_str}")
-            lines.append("")
-
-        # Platforms (if multi-platform)
-        if len(report["platforms"]) > 1:
-            lines.append("**📱 Platforms:**")
-            for p in report["platforms"]:
-                lines.append(f"  {p['platform']} — {p['sessions']} sessions, {p['messages']:,} msgs")
-            lines.append("")
-
-        # Tools (top 8)
-        if report["tools"]:
-            lines.append("**🔧 Top Tools:**")
-            for t in report["tools"][:8]:
-                lines.append(f"  {t['tool']} — {t['count']:,} calls ({t['percentage']:.1f}%)")
-            lines.append("")
-
-        # Activity summary
-        act = report.get("activity", {})
-        if act.get("busiest_day") and act.get("busiest_hour"):
-            hr = act["busiest_hour"]["hour"]
-            ampm = "AM" if hr < 12 else "PM"
-            display_hr = hr % 12 or 12
-            lines.append(f"**📅 Busiest:** {act['busiest_day']['day']}s ({act['busiest_day']['count']} sessions), {display_hr}{ampm} ({act['busiest_hour']['count']} sessions)")
-            if act.get("active_days"):
-                lines.append(f"**Active days:** {act['active_days']}", )
-            if act.get("max_streak", 0) > 1:
-                lines.append(f"**Best streak:** {act['max_streak']} consecutive days")
-
-        return "\n".join(lines)

agent/memory_manager.py

@@ -1,335 +0,0 @@
-"""MemoryManager — orchestrates the built-in memory provider plus at most
-ONE external plugin memory provider.
-
-Single integration point in run_agent.py. Replaces scattered per-backend
-code with one manager that delegates to registered providers.
-
-The BuiltinMemoryProvider is always registered first and cannot be removed.
-Only ONE external (non-builtin) provider is allowed at a time — attempting
-to register a second external provider is rejected with a warning.  This
-prevents tool schema bloat and conflicting memory backends.
-
-Usage in run_agent.py:
-    self._memory_manager = MemoryManager()
-    self._memory_manager.add_provider(BuiltinMemoryProvider(...))
-    # Only ONE of these:
-    self._memory_manager.add_provider(plugin_provider)
-
-    # System prompt
-    prompt_parts.append(self._memory_manager.build_system_prompt())
-
-    # Pre-turn
-    context = self._memory_manager.prefetch_all(user_message)
-
-    # Post-turn
-    self._memory_manager.sync_all(user_msg, assistant_response)
-    self._memory_manager.queue_prefetch_all(user_msg)
-"""
-
-from __future__ import annotations
-
-import json
-import logging
-from typing import Any, Dict, List, Optional
-
-from agent.memory_provider import MemoryProvider
-
-logger = logging.getLogger(__name__)
-
-
-class MemoryManager:
-    """Orchestrates the built-in provider plus at most one external provider.
-
-    The builtin provider is always first. Only one non-builtin (external)
-    provider is allowed.  Failures in one provider never block the other.
-    """
-
-    def __init__(self) -> None:
-        self._providers: List[MemoryProvider] = []
-        self._tool_to_provider: Dict[str, MemoryProvider] = {}
-        self._has_external: bool = False  # True once a non-builtin provider is added
-
-    # -- Registration --------------------------------------------------------
-
-    def add_provider(self, provider: MemoryProvider) -> None:
-        """Register a memory provider.
-
-        Built-in provider (name ``"builtin"``) is always accepted.
-        Only **one** external (non-builtin) provider is allowed — a second
-        attempt is rejected with a warning.
-        """
-        is_builtin = provider.name == "builtin"
-
-        if not is_builtin:
-            if self._has_external:
-                existing = next(
-                    (p.name for p in self._providers if p.name != "builtin"), "unknown"
-                )
-                logger.warning(
-                    "Rejected memory provider '%s' — external provider '%s' is "
-                    "already registered. Only one external memory provider is "
-                    "allowed at a time. Configure which one via memory.provider "
-                    "in config.yaml.",
-                    provider.name, existing,
-                )
-                return
-            self._has_external = True
-
-        self._providers.append(provider)
-
-        # Index tool names → provider for routing
-        for schema in provider.get_tool_schemas():
-            tool_name = schema.get("name", "")
-            if tool_name and tool_name not in self._tool_to_provider:
-                self._tool_to_provider[tool_name] = provider
-            elif tool_name in self._tool_to_provider:
-                logger.warning(
-                    "Memory tool name conflict: '%s' already registered by %s, "
-                    "ignoring from %s",
-                    tool_name,
-                    self._tool_to_provider[tool_name].name,
-                    provider.name,
-                )
-
-        logger.info(
-            "Memory provider '%s' registered (%d tools)",
-            provider.name,
-            len(provider.get_tool_schemas()),
-        )
-
-    @property
-    def providers(self) -> List[MemoryProvider]:
-        """All registered providers in order."""
-        return list(self._providers)
-
-    @property
-    def provider_names(self) -> List[str]:
-        """Names of all registered providers."""
-        return [p.name for p in self._providers]
-
-    def get_provider(self, name: str) -> Optional[MemoryProvider]:
-        """Get a provider by name, or None if not registered."""
-        for p in self._providers:
-            if p.name == name:
-                return p
-        return None
-
-    # -- System prompt -------------------------------------------------------
-
-    def build_system_prompt(self) -> str:
-        """Collect system prompt blocks from all providers.
-
-        Returns combined text, or empty string if no providers contribute.
-        Each non-empty block is labeled with the provider name.
-        """
-        blocks = []
-        for provider in self._providers:
-            try:
-                block = provider.system_prompt_block()
-                if block and block.strip():
-                    blocks.append(block)
-            except Exception as e:
-                logger.warning(
-                    "Memory provider '%s' system_prompt_block() failed: %s",
-                    provider.name, e,
-                )
-        return "\n\n".join(blocks)
-
-    # -- Prefetch / recall ---------------------------------------------------
-
-    def prefetch_all(self, query: str, *, session_id: str = "") -> str:
-        """Collect prefetch context from all providers.
-
-        Returns merged context text labeled by provider. Empty providers
-        are skipped. Failures in one provider don't block others.
-        """
-        parts = []
-        for provider in self._providers:
-            try:
-                result = provider.prefetch(query, session_id=session_id)
-                if result and result.strip():
-                    parts.append(result)
-            except Exception as e:
-                logger.debug(
-                    "Memory provider '%s' prefetch failed (non-fatal): %s",
-                    provider.name, e,
-                )
-        return "\n\n".join(parts)
-
-    def queue_prefetch_all(self, query: str, *, session_id: str = "") -> None:
-        """Queue background prefetch on all providers for the next turn."""
-        for provider in self._providers:
-            try:
-                provider.queue_prefetch(query, session_id=session_id)
-            except Exception as e:
-                logger.debug(
-                    "Memory provider '%s' queue_prefetch failed (non-fatal): %s",
-                    provider.name, e,
-                )
-
-    # -- Sync ----------------------------------------------------------------
-
-    def sync_all(self, user_content: str, assistant_content: str, *, session_id: str = "") -> None:
-        """Sync a completed turn to all providers."""
-        for provider in self._providers:
-            try:
-                provider.sync_turn(user_content, assistant_content, session_id=session_id)
-            except Exception as e:
-                logger.warning(
-                    "Memory provider '%s' sync_turn failed: %s",
-                    provider.name, e,
-                )
-
-    # -- Tools ---------------------------------------------------------------
-
-    def get_all_tool_schemas(self) -> List[Dict[str, Any]]:
-        """Collect tool schemas from all providers."""
-        schemas = []
-        seen = set()
-        for provider in self._providers:
-            try:
-                for schema in provider.get_tool_schemas():
-                    name = schema.get("name", "")
-                    if name and name not in seen:
-                        schemas.append(schema)
-                        seen.add(name)
-            except Exception as e:
-                logger.warning(
-                    "Memory provider '%s' get_tool_schemas() failed: %s",
-                    provider.name, e,
-                )
-        return schemas
-
-    def get_all_tool_names(self) -> set:
-        """Return set of all tool names across all providers."""
-        return set(self._tool_to_provider.keys())
-
-    def has_tool(self, tool_name: str) -> bool:
-        """Check if any provider handles this tool."""
-        return tool_name in self._tool_to_provider
-
-    def handle_tool_call(
-        self, tool_name: str, args: Dict[str, Any], **kwargs
-    ) -> str:
-        """Route a tool call to the correct provider.
-
-        Returns JSON string result. Raises ValueError if no provider
-        handles the tool.
-        """
-        provider = self._tool_to_provider.get(tool_name)
-        if provider is None:
-            return json.dumps({"error": f"No memory provider handles tool '{tool_name}'"})
-        try:
-            return provider.handle_tool_call(tool_name, args, **kwargs)
-        except Exception as e:
-            logger.error(
-                "Memory provider '%s' handle_tool_call(%s) failed: %s",
-                provider.name, tool_name, e,
-            )
-            return json.dumps({"error": f"Memory tool '{tool_name}' failed: {e}"})
-
-    # -- Lifecycle hooks -----------------------------------------------------
-
-    def on_turn_start(self, turn_number: int, message: str, **kwargs) -> None:
-        """Notify all providers of a new turn.
-
-        kwargs may include: remaining_tokens, model, platform, tool_count.
-        """
-        for provider in self._providers:
-            try:
-                provider.on_turn_start(turn_number, message, **kwargs)
-            except Exception as e:
-                logger.debug(
-                    "Memory provider '%s' on_turn_start failed: %s",
-                    provider.name, e,
-                )
-
-    def on_session_end(self, messages: List[Dict[str, Any]]) -> None:
-        """Notify all providers of session end."""
-        for provider in self._providers:
-            try:
-                provider.on_session_end(messages)
-            except Exception as e:
-                logger.debug(
-                    "Memory provider '%s' on_session_end failed: %s",
-                    provider.name, e,
-                )
-
-    def on_pre_compress(self, messages: List[Dict[str, Any]]) -> str:
-        """Notify all providers before context compression.
-
-        Returns combined text from providers to include in the compression
-        summary prompt. Empty string if no provider contributes.
-        """
-        parts = []
-        for provider in self._providers:
-            try:
-                result = provider.on_pre_compress(messages)
-                if result and result.strip():
-                    parts.append(result)
-            except Exception as e:
-                logger.debug(
-                    "Memory provider '%s' on_pre_compress failed: %s",
-                    provider.name, e,
-                )
-        return "\n\n".join(parts)
-
-    def on_memory_write(self, action: str, target: str, content: str) -> None:
-        """Notify external providers when the built-in memory tool writes.
-
-        Skips the builtin provider itself (it's the source of the write).
-        """
-        for provider in self._providers:
-            if provider.name == "builtin":
-                continue
-            try:
-                provider.on_memory_write(action, target, content)
-            except Exception as e:
-                logger.debug(
-                    "Memory provider '%s' on_memory_write failed: %s",
-                    provider.name, e,
-                )
-
-    def on_delegation(self, task: str, result: str, *,
-                      child_session_id: str = "", **kwargs) -> None:
-        """Notify all providers that a subagent completed."""
-        for provider in self._providers:
-            try:
-                provider.on_delegation(
-                    task, result, child_session_id=child_session_id, **kwargs
-                )
-            except Exception as e:
-                logger.debug(
-                    "Memory provider '%s' on_delegation failed: %s",
-                    provider.name, e,
-                )
-
-    def shutdown_all(self) -> None:
-        """Shut down all providers (reverse order for clean teardown)."""
-        for provider in reversed(self._providers):
-            try:
-                provider.shutdown()
-            except Exception as e:
-                logger.warning(
-                    "Memory provider '%s' shutdown failed: %s",
-                    provider.name, e,
-                )
-
-    def initialize_all(self, session_id: str, **kwargs) -> None:
-        """Initialize all providers.
-
-        Automatically injects ``hermes_home`` into *kwargs* so that every
-        provider can resolve profile-scoped storage paths without importing
-        ``get_hermes_home()`` themselves.
-        """
-        if "hermes_home" not in kwargs:
-            from hermes_constants import get_hermes_home
-            kwargs["hermes_home"] = str(get_hermes_home())
-        for provider in self._providers:
-            try:
-                provider.initialize(session_id=session_id, **kwargs)
-            except Exception as e:
-                logger.warning(
-                    "Memory provider '%s' initialize failed: %s",
-                    provider.name, e,
-                )

agent/memory_provider.py

@@ -1,231 +0,0 @@
-"""Abstract base class for pluggable memory providers.
-
-Memory providers give the agent persistent recall across sessions. One
-external provider is active at a time alongside the always-on built-in
-memory (MEMORY.md / USER.md). The MemoryManager enforces this limit.
-
-Built-in memory is always active as the first provider and cannot be removed.
-External providers (Honcho, Hindsight, Mem0, etc.) are additive — they never
-disable the built-in store. Only one external provider runs at a time to
-prevent tool schema bloat and conflicting memory backends.
-
-Registration:
-  1. Built-in: BuiltinMemoryProvider — always present, not removable.
-  2. Plugins: Ship in plugins/memory/<name>/, activated by memory.provider config.
-
-Lifecycle (called by MemoryManager, wired in run_agent.py):
-  initialize()          — connect, create resources, warm up
-  system_prompt_block()  — static text for the system prompt
-  prefetch(query)        — background recall before each turn
-  sync_turn(user, asst)  — async write after each turn
-  get_tool_schemas()     — tool schemas to expose to the model
-  handle_tool_call()     — dispatch a tool call
-  shutdown()             — clean exit
-
-Optional hooks (override to opt in):
-  on_turn_start(turn, message, **kwargs) — per-turn tick with runtime context
-  on_session_end(messages)               — end-of-session extraction
-  on_pre_compress(messages) -> str       — extract before context compression
-  on_memory_write(action, target, content) — mirror built-in memory writes
-  on_delegation(task, result, **kwargs)  — parent-side observation of subagent work
-"""
-
-from __future__ import annotations
-
-import logging
-from abc import ABC, abstractmethod
-from typing import Any, Dict, List, Optional
-
-logger = logging.getLogger(__name__)
-
-
-class MemoryProvider(ABC):
-    """Abstract base class for memory providers."""
-
-    @property
-    @abstractmethod
-    def name(self) -> str:
-        """Short identifier for this provider (e.g. 'builtin', 'honcho', 'hindsight')."""
-
-    # -- Core lifecycle (implement these) ------------------------------------
-
-    @abstractmethod
-    def is_available(self) -> bool:
-        """Return True if this provider is configured, has credentials, and is ready.
-
-        Called during agent init to decide whether to activate the provider.
-        Should not make network calls — just check config and installed deps.
-        """
-
-    @abstractmethod
-    def initialize(self, session_id: str, **kwargs) -> None:
-        """Initialize for a session.
-
-        Called once at agent startup. May create resources (banks, tables),
-        establish connections, start background threads, etc.
-
-        kwargs always include:
-          - hermes_home (str): The active HERMES_HOME directory path. Use this
-            for profile-scoped storage instead of hardcoding ``~/.hermes``.
-          - platform (str): "cli", "telegram", "discord", "cron", etc.
-
-        kwargs may also include:
-          - agent_context (str): "primary", "subagent", "cron", or "flush".
-            Providers should skip writes for non-primary contexts (cron system
-            prompts would corrupt user representations).
-          - agent_identity (str): Profile name (e.g. "coder"). Use for
-            per-profile provider identity scoping.
-          - agent_workspace (str): Shared workspace name (e.g. "hermes").
-          - parent_session_id (str): For subagents, the parent's session_id.
-          - user_id (str): Platform user identifier (gateway sessions).
-        """
-
-    def system_prompt_block(self) -> str:
-        """Return text to include in the system prompt.
-
-        Called during system prompt assembly. Return empty string to skip.
-        This is for STATIC provider info (instructions, status). Prefetched
-        recall context is injected separately via prefetch().
-        """
-        return ""
-
-    def prefetch(self, query: str, *, session_id: str = "") -> str:
-        """Recall relevant context for the upcoming turn.
-
-        Called before each API call. Return formatted text to inject as
-        context, or empty string if nothing relevant. Implementations
-        should be fast — use background threads for the actual recall
-        and return cached results here.
-
-        session_id is provided for providers serving concurrent sessions
-        (gateway group chats, cached agents). Providers that don't need
-        per-session scoping can ignore it.
-        """
-        return ""
-
-    def queue_prefetch(self, query: str, *, session_id: str = "") -> None:
-        """Queue a background recall for the NEXT turn.
-
-        Called after each turn completes. The result will be consumed
-        by prefetch() on the next turn. Default is no-op — providers
-        that do background prefetching should override this.
-        """
-
-    def sync_turn(self, user_content: str, assistant_content: str, *, session_id: str = "") -> None:
-        """Persist a completed turn to the backend.
-
-        Called after each turn. Should be non-blocking — queue for
-        background processing if the backend has latency.
-        """
-
-    @abstractmethod
-    def get_tool_schemas(self) -> List[Dict[str, Any]]:
-        """Return tool schemas this provider exposes.
-
-        Each schema follows the OpenAI function calling format:
-        {"name": "...", "description": "...", "parameters": {...}}
-
-        Return empty list if this provider has no tools (context-only).
-        """
-
-    def handle_tool_call(self, tool_name: str, args: Dict[str, Any], **kwargs) -> str:
-        """Handle a tool call for one of this provider's tools.
-
-        Must return a JSON string (the tool result).
-        Only called for tool names returned by get_tool_schemas().
-        """
-        raise NotImplementedError(f"Provider {self.name} does not handle tool {tool_name}")
-
-    def shutdown(self) -> None:
-        """Clean shutdown — flush queues, close connections."""
-
-    # -- Optional hooks (override to opt in) ---------------------------------
-
-    def on_turn_start(self, turn_number: int, message: str, **kwargs) -> None:
-        """Called at the start of each turn with the user message.
-
-        Use for turn-counting, scope management, periodic maintenance.
-
-        kwargs may include: remaining_tokens, model, platform, tool_count.
-        Providers use what they need; extras are ignored.
-        """
-
-    def on_session_end(self, messages: List[Dict[str, Any]]) -> None:
-        """Called when a session ends (explicit exit or timeout).
-
-        Use for end-of-session fact extraction, summarization, etc.
-        messages is the full conversation history.
-
-        NOT called after every turn — only at actual session boundaries
-        (CLI exit, /reset, gateway session expiry).
-        """
-
-    def on_pre_compress(self, messages: List[Dict[str, Any]]) -> str:
-        """Called before context compression discards old messages.
-
-        Use to extract insights from messages about to be compressed.
-        messages is the list that will be summarized/discarded.
-
-        Return text to include in the compression summary prompt so the
-        compressor preserves provider-extracted insights. Return empty
-        string for no contribution (backwards-compatible default).
-        """
-        return ""
-
-    def on_delegation(self, task: str, result: str, *,
-                      child_session_id: str = "", **kwargs) -> None:
-        """Called on the PARENT agent when a subagent completes.
-
-        The parent's memory provider gets the task+result pair as an
-        observation of what was delegated and what came back. The subagent
-        itself has no provider session (skip_memory=True).
-
-        task: the delegation prompt
-        result: the subagent's final response
-        child_session_id: the subagent's session_id
-        """
-
-    def get_config_schema(self) -> List[Dict[str, Any]]:
-        """Return config fields this provider needs for setup.
-
-        Used by 'hermes memory setup' to walk the user through configuration.
-        Each field is a dict with:
-          key:         config key name (e.g. 'api_key', 'mode')
-          description: human-readable description
-          secret:      True if this should go to .env (default: False)
-          required:    True if required (default: False)
-          default:     default value (optional)
-          choices:     list of valid values (optional)
-          url:         URL where user can get this credential (optional)
-          env_var:     explicit env var name for secrets (default: auto-generated)
-
-        Return empty list if no config needed (e.g. local-only providers).
-        """
-        return []
-
-    def save_config(self, values: Dict[str, Any], hermes_home: str) -> None:
-        """Write non-secret config to the provider's native location.
-
-        Called by 'hermes memory setup' after collecting user inputs.
-        ``values`` contains only non-secret fields (secrets go to .env).
-        ``hermes_home`` is the active HERMES_HOME directory path.
-
-        Providers with native config files (JSON, YAML) should override
-        this to write to their expected location. Providers that use only
-        env vars can leave the default (no-op).
-
-        All new memory provider plugins MUST implement either:
-        - save_config() for native config file formats, OR
-        - use only env vars (in which case get_config_schema() fields
-          should all have ``env_var`` set and this method stays no-op).
-        """
-
-    def on_memory_write(self, action: str, target: str, content: str) -> None:
-        """Called when the built-in memory tool writes an entry.
-
-        action: 'add', 'replace', or 'remove'
-        target: 'memory' or 'user'
-        content: the entry content
-
-        Use to mirror built-in memory writes to your backend.
-        """

agent/model_metadata.py

@@ -1,931 +0,0 @@
-"""Model metadata, context lengths, and token estimation utilities.
-
-Pure utility functions with no AIAgent dependency. Used by ContextCompressor
-and run_agent.py for pre-flight context checks.
-"""
-
-import logging
-import os
-import re
-import time
-from pathlib import Path
-from typing import Any, Dict, List, Optional
-from urllib.parse import urlparse
-
-import requests
-import yaml
-
-from hermes_constants import OPENROUTER_MODELS_URL
-
-logger = logging.getLogger(__name__)
-
-# Provider names that can appear as a "provider:" prefix before a model ID.
-# Only these are stripped — Ollama-style "model:tag" colons (e.g. "qwen3.5:27b")
-# are preserved so the full model name reaches cache lookups and server queries.
-_PROVIDER_PREFIXES: frozenset[str] = frozenset({
-    "openrouter", "nous", "openai-codex", "copilot", "copilot-acp",
-    "zai", "kimi-coding", "minimax", "minimax-cn", "anthropic", "deepseek",
-    "opencode-zen", "opencode-go", "ai-gateway", "kilocode", "alibaba",
-    "custom", "local",
-    # Common aliases
-    "glm", "z-ai", "z.ai", "zhipu", "github", "github-copilot",
-    "github-models", "kimi", "moonshot", "claude", "deep-seek",
-    "opencode", "zen", "go", "vercel", "kilo", "dashscope", "aliyun", "qwen",
-})
-
-
-_OLLAMA_TAG_PATTERN = re.compile(
-    r"^(\d+\.?\d*b|latest|stable|q\d|fp?\d|instruct|chat|coder|vision|text)",
-    re.IGNORECASE,
-)
-
-
-def _strip_provider_prefix(model: str) -> str:
-    """Strip a recognised provider prefix from a model string.
-
-    ``"local:my-model"`` → ``"my-model"``
-    ``"qwen3.5:27b"``   → ``"qwen3.5:27b"``  (unchanged — not a provider prefix)
-    ``"qwen:0.5b"``     → ``"qwen:0.5b"``    (unchanged — Ollama model:tag)
-    ``"deepseek:latest"``→ ``"deepseek:latest"``(unchanged — Ollama model:tag)
-    """
-    if ":" not in model or model.startswith("http"):
-        return model
-    prefix, suffix = model.split(":", 1)
-    prefix_lower = prefix.strip().lower()
-    if prefix_lower in _PROVIDER_PREFIXES:
-        # Don't strip if suffix looks like an Ollama tag (e.g. "7b", "latest", "q4_0")
-        if _OLLAMA_TAG_PATTERN.match(suffix.strip()):
-            return model
-        return suffix
-    return model
-
-_model_metadata_cache: Dict[str, Dict[str, Any]] = {}
-_model_metadata_cache_time: float = 0
-_MODEL_CACHE_TTL = 3600
-_endpoint_model_metadata_cache: Dict[str, Dict[str, Dict[str, Any]]] = {}
-_endpoint_model_metadata_cache_time: Dict[str, float] = {}
-_ENDPOINT_MODEL_CACHE_TTL = 300
-
-# Descending tiers for context length probing when the model is unknown.
-# We start at 128K (a safe default for most modern models) and step down
-# on context-length errors until one works.
-CONTEXT_PROBE_TIERS = [
-    128_000,
-    64_000,
-    32_000,
-    16_000,
-    8_000,
-]
-
-# Default context length when no detection method succeeds.
-DEFAULT_FALLBACK_CONTEXT = CONTEXT_PROBE_TIERS[0]
-
-# Thin fallback defaults — only broad model family patterns.
-# These fire only when provider is unknown AND models.dev/OpenRouter/Anthropic
-# all miss. Replaced the previous 80+ entry dict.
-# For provider-specific context lengths, models.dev is the primary source.
-DEFAULT_CONTEXT_LENGTHS = {
-    # Anthropic Claude 4.6 (1M context) — bare IDs only to avoid
-    # fuzzy-match collisions (e.g. "anthropic/claude-sonnet-4" is a
-    # substring of "anthropic/claude-sonnet-4.6").
-    # OpenRouter-prefixed models resolve via OpenRouter live API or models.dev.
-    "claude-opus-4-6": 1000000,
-    "claude-sonnet-4-6": 1000000,
-    "claude-opus-4.6": 1000000,
-    "claude-sonnet-4.6": 1000000,
-    # Catch-all for older Claude models (must sort after specific entries)
-    "claude": 200000,
-    # OpenAI
-    "gpt-4.1": 1047576,
-    "gpt-5": 128000,
-    "gpt-4": 128000,
-    # Google
-    "gemini": 1048576,
-    # DeepSeek
-    "deepseek": 128000,
-    # Meta
-    "llama": 131072,
-    # Qwen
-    "qwen": 131072,
-    # MiniMax
-    "minimax": 204800,
-    # GLM
-    "glm": 202752,
-    # Kimi
-    "kimi": 262144,
-    # Hugging Face Inference Providers — model IDs use org/name format
-    "Qwen/Qwen3.5-397B-A17B": 131072,
-    "Qwen/Qwen3.5-35B-A3B": 131072,
-    "deepseek-ai/DeepSeek-V3.2": 65536,
-    "moonshotai/Kimi-K2.5": 262144,
-    "moonshotai/Kimi-K2-Thinking": 262144,
-    "MiniMaxAI/MiniMax-M2.5": 204800,
-    "XiaomiMiMo/MiMo-V2-Flash": 32768,
-    "zai-org/GLM-5": 202752,
-}
-
-_CONTEXT_LENGTH_KEYS = (
-    "context_length",
-    "context_window",
-    "max_context_length",
-    "max_position_embeddings",
-    "max_model_len",
-    "max_input_tokens",
-    "max_sequence_length",
-    "max_seq_len",
-    "n_ctx_train",
-    "n_ctx",
-)
-
-_MAX_COMPLETION_KEYS = (
-    "max_completion_tokens",
-    "max_output_tokens",
-    "max_tokens",
-)
-
-# Local server hostnames / address patterns
-_LOCAL_HOSTS = ("localhost", "127.0.0.1", "::1", "0.0.0.0")
-
-
-def _normalize_base_url(base_url: str) -> str:
-    return (base_url or "").strip().rstrip("/")
-
-
-def _is_openrouter_base_url(base_url: str) -> bool:
-    return "openrouter.ai" in _normalize_base_url(base_url).lower()
-
-
-def _is_custom_endpoint(base_url: str) -> bool:
-    normalized = _normalize_base_url(base_url)
-    return bool(normalized) and not _is_openrouter_base_url(normalized)
-
-
-_URL_TO_PROVIDER: Dict[str, str] = {
-    "api.openai.com": "openai",
-    "chatgpt.com": "openai",
-    "api.anthropic.com": "anthropic",
-    "api.z.ai": "zai",
-    "api.moonshot.ai": "kimi-coding",
-    "api.kimi.com": "kimi-coding",
-    "api.minimax": "minimax",
-    "dashscope.aliyuncs.com": "alibaba",
-    "dashscope-intl.aliyuncs.com": "alibaba",
-    "openrouter.ai": "openrouter",
-    "generativelanguage.googleapis.com": "google",
-    "inference-api.nousresearch.com": "nous",
-    "api.deepseek.com": "deepseek",
-    "api.githubcopilot.com": "copilot",
-    "models.github.ai": "copilot",
-    "api.fireworks.ai": "fireworks",
-}
-
-
-def _infer_provider_from_url(base_url: str) -> Optional[str]:
-    """Infer the models.dev provider name from a base URL.
-
-    This allows context length resolution via models.dev for custom endpoints
-    like DashScope (Alibaba), Z.AI, Kimi, etc. without requiring the user to
-    explicitly set the provider name in config.
-    """
-    normalized = _normalize_base_url(base_url)
-    if not normalized:
-        return None
-    parsed = urlparse(normalized if "://" in normalized else f"https://{normalized}")
-    host = parsed.netloc.lower() or parsed.path.lower()
-    for url_part, provider in _URL_TO_PROVIDER.items():
-        if url_part in host:
-            return provider
-    return None
-
-
-def _is_known_provider_base_url(base_url: str) -> bool:
-    return _infer_provider_from_url(base_url) is not None
-
-
-def is_local_endpoint(base_url: str) -> bool:
-    """Return True if base_url points to a local machine (localhost / RFC-1918 / WSL)."""
-    normalized = _normalize_base_url(base_url)
-    if not normalized:
-        return False
-    url = normalized if "://" in normalized else f"http://{normalized}"
-    try:
-        parsed = urlparse(url)
-        host = parsed.hostname or ""
-    except Exception:
-        return False
-    if host in _LOCAL_HOSTS:
-        return True
-    # RFC-1918 private ranges and link-local
-    import ipaddress
-    try:
-        addr = ipaddress.ip_address(host)
-        return addr.is_private or addr.is_loopback or addr.is_link_local
-    except ValueError:
-        pass
-    # Bare IP that looks like a private range (e.g. 172.26.x.x for WSL)
-    parts = host.split(".")
-    if len(parts) == 4:
-        try:
-            first, second = int(parts[0]), int(parts[1])
-            if first == 10:
-                return True
-            if first == 172 and 16 <= second <= 31:
-                return True
-            if first == 192 and second == 168:
-                return True
-        except ValueError:
-            pass
-    return False
-
-
-def detect_local_server_type(base_url: str) -> Optional[str]:
-    """Detect which local server is running at base_url by probing known endpoints.
-
-    Returns one of: "ollama", "lm-studio", "vllm", "llamacpp", or None.
-    """
-    import httpx
-
-    normalized = _normalize_base_url(base_url)
-    server_url = normalized
-    if server_url.endswith("/v1"):
-        server_url = server_url[:-3]
-
-    try:
-        with httpx.Client(timeout=2.0) as client:
-            # LM Studio exposes /api/v1/models — check first (most specific)
-            try:
-                r = client.get(f"{server_url}/api/v1/models")
-                if r.status_code == 200:
-                    return "lm-studio"
-            except Exception:
-                pass
-            # Ollama exposes /api/tags and responds with {"models": [...]}
-            # LM Studio returns {"error": "Unexpected endpoint"} with status 200
-            # on this path, so we must verify the response contains "models".
-            try:
-                r = client.get(f"{server_url}/api/tags")
-                if r.status_code == 200:
-                    try:
-                        data = r.json()
-                        if "models" in data:
-                            return "ollama"
-                    except Exception:
-                        pass
-            except Exception:
-                pass
-            # llama.cpp exposes /v1/props (older builds used /props without the /v1 prefix)
-            try:
-                r = client.get(f"{server_url}/v1/props")
-                if r.status_code != 200:
-                    r = client.get(f"{server_url}/props")  # fallback for older builds
-                if r.status_code == 200 and "default_generation_settings" in r.text:
-                    return "llamacpp"
-            except Exception:
-                pass
-            # vLLM: /version
-            try:
-                r = client.get(f"{server_url}/version")
-                if r.status_code == 200:
-                    data = r.json()
-                    if "version" in data:
-                        return "vllm"
-            except Exception:
-                pass
-    except Exception:
-        pass
-
-    return None
-
-
-def _iter_nested_dicts(value: Any):
-    if isinstance(value, dict):
-        yield value
-        for nested in value.values():
-            yield from _iter_nested_dicts(nested)
-    elif isinstance(value, list):
-        for item in value:
-            yield from _iter_nested_dicts(item)
-
-
-def _coerce_reasonable_int(value: Any, minimum: int = 1024, maximum: int = 10_000_000) -> Optional[int]:
-    try:
-        if isinstance(value, bool):
-            return None
-        if isinstance(value, str):
-            value = value.strip().replace(",", "")
-        result = int(value)
-    except (TypeError, ValueError):
-        return None
-    if minimum <= result <= maximum:
-        return result
-    return None
-
-
-def _extract_first_int(payload: Dict[str, Any], keys: tuple[str, ...]) -> Optional[int]:
-    keyset = {key.lower() for key in keys}
-    for mapping in _iter_nested_dicts(payload):
-        for key, value in mapping.items():
-            if str(key).lower() not in keyset:
-                continue
-            coerced = _coerce_reasonable_int(value)
-            if coerced is not None:
-                return coerced
-    return None
-
-
-def _extract_context_length(payload: Dict[str, Any]) -> Optional[int]:
-    return _extract_first_int(payload, _CONTEXT_LENGTH_KEYS)
-
-
-def _extract_max_completion_tokens(payload: Dict[str, Any]) -> Optional[int]:
-    return _extract_first_int(payload, _MAX_COMPLETION_KEYS)
-
-
-def _extract_pricing(payload: Dict[str, Any]) -> Dict[str, Any]:
-    alias_map = {
-        "prompt": ("prompt", "input", "input_cost_per_token", "prompt_token_cost"),
-        "completion": ("completion", "output", "output_cost_per_token", "completion_token_cost"),
-        "request": ("request", "request_cost"),
-        "cache_read": ("cache_read", "cached_prompt", "input_cache_read", "cache_read_cost_per_token"),
-        "cache_write": ("cache_write", "cache_creation", "input_cache_write", "cache_write_cost_per_token"),
-    }
-    for mapping in _iter_nested_dicts(payload):
-        normalized = {str(key).lower(): value for key, value in mapping.items()}
-        if not any(any(alias in normalized for alias in aliases) for aliases in alias_map.values()):
-            continue
-        pricing: Dict[str, Any] = {}
-        for target, aliases in alias_map.items():
-            for alias in aliases:
-                if alias in normalized and normalized[alias] not in (None, ""):
-                    pricing[target] = normalized[alias]
-                    break
-        if pricing:
-            return pricing
-    return {}
-
-
-def _add_model_aliases(cache: Dict[str, Dict[str, Any]], model_id: str, entry: Dict[str, Any]) -> None:
-    cache[model_id] = entry
-    if "/" in model_id:
-        bare_model = model_id.split("/", 1)[1]
-        cache.setdefault(bare_model, entry)
-
-
-def fetch_model_metadata(force_refresh: bool = False) -> Dict[str, Dict[str, Any]]:
-    """Fetch model metadata from OpenRouter (cached for 1 hour)."""
-    global _model_metadata_cache, _model_metadata_cache_time
-
-    if not force_refresh and _model_metadata_cache and (time.time() - _model_metadata_cache_time) < _MODEL_CACHE_TTL:
-        return _model_metadata_cache
-
-    try:
-        response = requests.get(OPENROUTER_MODELS_URL, timeout=10)
-        response.raise_for_status()
-        data = response.json()
-
-        cache = {}
-        for model in data.get("data", []):
-            model_id = model.get("id", "")
-            entry = {
-                "context_length": model.get("context_length", 128000),
-                "max_completion_tokens": model.get("top_provider", {}).get("max_completion_tokens", 4096),
-                "name": model.get("name", model_id),
-                "pricing": model.get("pricing", {}),
-            }
-            _add_model_aliases(cache, model_id, entry)
-            canonical = model.get("canonical_slug", "")
-            if canonical and canonical != model_id:
-                _add_model_aliases(cache, canonical, entry)
-
-        _model_metadata_cache = cache
-        _model_metadata_cache_time = time.time()
-        logger.debug("Fetched metadata for %s models from OpenRouter", len(cache))
-        return cache
-
-    except Exception as e:
-        logging.warning(f"Failed to fetch model metadata from OpenRouter: {e}")
-        return _model_metadata_cache or {}
-
-
-def fetch_endpoint_model_metadata(
-    base_url: str,
-    api_key: str = "",
-    force_refresh: bool = False,
-) -> Dict[str, Dict[str, Any]]:
-    """Fetch model metadata from an OpenAI-compatible ``/models`` endpoint.
-
-    This is used for explicit custom endpoints where hardcoded global model-name
-    defaults are unreliable. Results are cached in memory per base URL.
-    """
-    normalized = _normalize_base_url(base_url)
-    if not normalized or _is_openrouter_base_url(normalized):
-        return {}
-
-    if not force_refresh:
-        cached = _endpoint_model_metadata_cache.get(normalized)
-        cached_at = _endpoint_model_metadata_cache_time.get(normalized, 0)
-        if cached is not None and (time.time() - cached_at) < _ENDPOINT_MODEL_CACHE_TTL:
-            return cached
-
-    candidates = [normalized]
-    if normalized.endswith("/v1"):
-        alternate = normalized[:-3].rstrip("/")
-    else:
-        alternate = normalized + "/v1"
-    if alternate and alternate not in candidates:
-        candidates.append(alternate)
-
-    headers = {"Authorization": f"Bearer {api_key}"} if api_key else {}
-    last_error: Optional[Exception] = None
-
-    for candidate in candidates:
-        url = candidate.rstrip("/") + "/models"
-        try:
-            response = requests.get(url, headers=headers, timeout=10)
-            response.raise_for_status()
-            payload = response.json()
-            cache: Dict[str, Dict[str, Any]] = {}
-            for model in payload.get("data", []):
-                if not isinstance(model, dict):
-                    continue
-                model_id = model.get("id")
-                if not model_id:
-                    continue
-                entry: Dict[str, Any] = {"name": model.get("name", model_id)}
-                context_length = _extract_context_length(model)
-                if context_length is not None:
-                    entry["context_length"] = context_length
-                max_completion_tokens = _extract_max_completion_tokens(model)
-                if max_completion_tokens is not None:
-                    entry["max_completion_tokens"] = max_completion_tokens
-                pricing = _extract_pricing(model)
-                if pricing:
-                    entry["pricing"] = pricing
-                _add_model_aliases(cache, model_id, entry)
-
-            # If this is a llama.cpp server, query /props for actual allocated context
-            is_llamacpp = any(
-                m.get("owned_by") == "llamacpp"
-                for m in payload.get("data", []) if isinstance(m, dict)
-            )
-            if is_llamacpp:
-                try:
-                    # Try /v1/props first (current llama.cpp); fall back to /props for older builds
-                    base = candidate.rstrip("/").replace("/v1", "")
-                    props_resp = requests.get(base + "/v1/props", headers=headers, timeout=5)
-                    if not props_resp.ok:
-                        props_resp = requests.get(base + "/props", headers=headers, timeout=5)
-                    if props_resp.ok:
-                        props = props_resp.json()
-                        gen_settings = props.get("default_generation_settings", {})
-                        n_ctx = gen_settings.get("n_ctx")
-                        model_alias = props.get("model_alias", "")
-                        if n_ctx and model_alias and model_alias in cache:
-                            cache[model_alias]["context_length"] = n_ctx
-                except Exception:
-                    pass
-
-            _endpoint_model_metadata_cache[normalized] = cache
-            _endpoint_model_metadata_cache_time[normalized] = time.time()
-            return cache
-        except Exception as exc:
-            last_error = exc
-
-    if last_error:
-        logger.debug("Failed to fetch model metadata from %s/models: %s", normalized, last_error)
-    _endpoint_model_metadata_cache[normalized] = {}
-    _endpoint_model_metadata_cache_time[normalized] = time.time()
-    return {}
-
-
-def _get_context_cache_path() -> Path:
-    """Return path to the persistent context length cache file."""
-    hermes_home = Path(os.environ.get("HERMES_HOME", Path.home() / ".hermes"))
-    return hermes_home / "context_length_cache.yaml"
-
-
-def _load_context_cache() -> Dict[str, int]:
-    """Load the model+provider -> context_length cache from disk."""
-    path = _get_context_cache_path()
-    if not path.exists():
-        return {}
-    try:
-        with open(path) as f:
-            data = yaml.safe_load(f) or {}
-        return data.get("context_lengths", {})
-    except Exception as e:
-        logger.debug("Failed to load context length cache: %s", e)
-        return {}
-
-
-def save_context_length(model: str, base_url: str, length: int) -> None:
-    """Persist a discovered context length for a model+provider combo.
-
-    Cache key is ``model@base_url`` so the same model name served from
-    different providers can have different limits.
-    """
-    key = f"{model}@{base_url}"
-    cache = _load_context_cache()
-    if cache.get(key) == length:
-        return  # already stored
-    cache[key] = length
-    path = _get_context_cache_path()
-    try:
-        path.parent.mkdir(parents=True, exist_ok=True)
-        with open(path, "w") as f:
-            yaml.dump({"context_lengths": cache}, f, default_flow_style=False)
-        logger.info("Cached context length %s -> %s tokens", key, f"{length:,}")
-    except Exception as e:
-        logger.debug("Failed to save context length cache: %s", e)
-
-
-def get_cached_context_length(model: str, base_url: str) -> Optional[int]:
-    """Look up a previously discovered context length for model+provider."""
-    key = f"{model}@{base_url}"
-    cache = _load_context_cache()
-    return cache.get(key)
-
-
-def get_next_probe_tier(current_length: int) -> Optional[int]:
-    """Return the next lower probe tier, or None if already at minimum."""
-    for tier in CONTEXT_PROBE_TIERS:
-        if tier < current_length:
-            return tier
-    return None
-
-
-def parse_context_limit_from_error(error_msg: str) -> Optional[int]:
-    """Try to extract the actual context limit from an API error message.
-
-    Many providers include the limit in their error text, e.g.:
-      - "maximum context length is 32768 tokens"
-      - "context_length_exceeded: 131072"
-      - "Maximum context size 32768 exceeded"
-      - "model's max context length is 65536"
-    """
-    error_lower = error_msg.lower()
-    # Pattern: look for numbers near context-related keywords
-    patterns = [
-        r'(?:max(?:imum)?|limit)\s*(?:context\s*)?(?:length|size|window)?\s*(?:is|of|:)?\s*(\d{4,})',
-        r'context\s*(?:length|size|window)\s*(?:is|of|:)?\s*(\d{4,})',
-        r'(\d{4,})\s*(?:token)?\s*(?:context|limit)',
-        r'>\s*(\d{4,})\s*(?:max|limit|token)',  # "250000 tokens > 200000 maximum"
-        r'(\d{4,})\s*(?:max(?:imum)?)\b',  # "200000 maximum"
-    ]
-    for pattern in patterns:
-        match = re.search(pattern, error_lower)
-        if match:
-            limit = int(match.group(1))
-            # Sanity check: must be a reasonable context length
-            if 1024 <= limit <= 10_000_000:
-                return limit
-    return None
-
-
-def _model_id_matches(candidate_id: str, lookup_model: str) -> bool:
-    """Return True if *candidate_id* (from server) matches *lookup_model* (configured).
-
-    Supports two forms:
-    - Exact match:  "nvidia-nemotron-super-49b-v1" == "nvidia-nemotron-super-49b-v1"
-    - Slug match:   "nvidia/nvidia-nemotron-super-49b-v1" matches "nvidia-nemotron-super-49b-v1"
-                    (the part after the last "/" equals lookup_model)
-
-    This covers LM Studio's native API which stores models as "publisher/slug"
-    while users typically configure only the slug after the "local:" prefix.
-    """
-    if candidate_id == lookup_model:
-        return True
-    # Slug match: basename of candidate equals the lookup name
-    if "/" in candidate_id and candidate_id.rsplit("/", 1)[1] == lookup_model:
-        return True
-    return False
-
-
-def _query_local_context_length(model: str, base_url: str) -> Optional[int]:
-    """Query a local server for the model's context length."""
-    import httpx
-
-    # Strip recognised provider prefix (e.g., "local:model-name" → "model-name").
-    # Ollama "model:tag" colons (e.g. "qwen3.5:27b") are intentionally preserved.
-    model = _strip_provider_prefix(model)
-
-    # Strip /v1 suffix to get the server root
-    server_url = base_url.rstrip("/")
-    if server_url.endswith("/v1"):
-        server_url = server_url[:-3]
-
-    try:
-        server_type = detect_local_server_type(base_url)
-    except Exception:
-        server_type = None
-
-    try:
-        with httpx.Client(timeout=3.0) as client:
-            # Ollama: /api/show returns model details with context info
-            if server_type == "ollama":
-                resp = client.post(f"{server_url}/api/show", json={"name": model})
-                if resp.status_code == 200:
-                    data = resp.json()
-                    # Check model_info for context length
-                    model_info = data.get("model_info", {})
-                    for key, value in model_info.items():
-                        if "context_length" in key and isinstance(value, (int, float)):
-                            return int(value)
-                    # Check parameters string for num_ctx
-                    params = data.get("parameters", "")
-                    if "num_ctx" in params:
-                        for line in params.split("\n"):
-                            if "num_ctx" in line:
-                                parts = line.strip().split()
-                                if len(parts) >= 2:
-                                    try:
-                                        return int(parts[-1])
-                                    except ValueError:
-                                        pass
-
-            # LM Studio native API: /api/v1/models returns max_context_length.
-            # This is more reliable than the OpenAI-compat /v1/models which
-            # doesn't include context window information for LM Studio servers.
-            # Use _model_id_matches for fuzzy matching: LM Studio stores models as
-            # "publisher/slug" but users configure only "slug" after "local:" prefix.
-            if server_type == "lm-studio":
-                resp = client.get(f"{server_url}/api/v1/models")
-                if resp.status_code == 200:
-                    data = resp.json()
-                    for m in data.get("models", []):
-                        if _model_id_matches(m.get("key", ""), model) or _model_id_matches(m.get("id", ""), model):
-                            # Prefer loaded instance context (actual runtime value)
-                            for inst in m.get("loaded_instances", []):
-                                cfg = inst.get("config", {})
-                                ctx = cfg.get("context_length")
-                                if ctx and isinstance(ctx, (int, float)):
-                                    return int(ctx)
-                            # Fall back to max_context_length (theoretical model max)
-                            ctx = m.get("max_context_length") or m.get("context_length")
-                            if ctx and isinstance(ctx, (int, float)):
-                                return int(ctx)
-
-            # LM Studio / vLLM / llama.cpp: try /v1/models/{model}
-            resp = client.get(f"{server_url}/v1/models/{model}")
-            if resp.status_code == 200:
-                data = resp.json()
-                # vLLM returns max_model_len
-                ctx = data.get("max_model_len") or data.get("context_length") or data.get("max_tokens")
-                if ctx and isinstance(ctx, (int, float)):
-                    return int(ctx)
-
-            # Try /v1/models and find the model in the list.
-            # Use _model_id_matches to handle "publisher/slug" vs bare "slug".
-            resp = client.get(f"{server_url}/v1/models")
-            if resp.status_code == 200:
-                data = resp.json()
-                models_list = data.get("data", [])
-                for m in models_list:
-                    if _model_id_matches(m.get("id", ""), model):
-                        ctx = m.get("max_model_len") or m.get("context_length") or m.get("max_tokens")
-                        if ctx and isinstance(ctx, (int, float)):
-                            return int(ctx)
-    except Exception:
-        pass
-
-    return None
-
-
-def _normalize_model_version(model: str) -> str:
-    """Normalize version separators for matching.
-
-    Nous uses dashes: claude-opus-4-6, claude-sonnet-4-5
-    OpenRouter uses dots: claude-opus-4.6, claude-sonnet-4.5
-    Normalize both to dashes for comparison.
-    """
-    return model.replace(".", "-")
-
-
-def _query_anthropic_context_length(model: str, base_url: str, api_key: str) -> Optional[int]:
-    """Query Anthropic's /v1/models endpoint for context length.
-
-    Only works with regular ANTHROPIC_API_KEY (sk-ant-api*).
-    OAuth tokens (sk-ant-oat*) from Claude Code return 401.
-    """
-    if not api_key or api_key.startswith("sk-ant-oat"):
-        return None  # OAuth tokens can't access /v1/models
-    try:
-        base = base_url.rstrip("/")
-        if base.endswith("/v1"):
-            base = base[:-3]
-        url = f"{base}/v1/models?limit=1000"
-        headers = {
-            "x-api-key": api_key,
-            "anthropic-version": "2023-06-01",
-        }
-        resp = requests.get(url, headers=headers, timeout=10)
-        if resp.status_code != 200:
-            return None
-        data = resp.json()
-        for m in data.get("data", []):
-            if m.get("id") == model:
-                ctx = m.get("max_input_tokens")
-                if isinstance(ctx, int) and ctx > 0:
-                    return ctx
-    except Exception as e:
-        logger.debug("Anthropic /v1/models query failed: %s", e)
-    return None
-
-
-def _resolve_nous_context_length(model: str) -> Optional[int]:
-    """Resolve Nous Portal model context length via OpenRouter metadata.
-
-    Nous model IDs are bare (e.g. 'claude-opus-4-6') while OpenRouter uses
-    prefixed IDs (e.g. 'anthropic/claude-opus-4.6'). Try suffix matching
-    with version normalization (dot↔dash).
-    """
-    metadata = fetch_model_metadata()  # OpenRouter cache
-    # Exact match first
-    if model in metadata:
-        return metadata[model].get("context_length")
-
-    normalized = _normalize_model_version(model).lower()
-
-    for or_id, entry in metadata.items():
-        bare = or_id.split("/", 1)[1] if "/" in or_id else or_id
-        if bare.lower() == model.lower() or _normalize_model_version(bare).lower() == normalized:
-            return entry.get("context_length")
-
-    # Partial prefix match for cases like gemini-3-flash → gemini-3-flash-preview
-    # Require match to be at a word boundary (followed by -, :, or end of string)
-    model_lower = model.lower()
-    for or_id, entry in metadata.items():
-        bare = or_id.split("/", 1)[1] if "/" in or_id else or_id
-        for candidate, query in [(bare.lower(), model_lower), (_normalize_model_version(bare).lower(), normalized)]:
-            if candidate.startswith(query) and (
-                len(candidate) == len(query) or candidate[len(query)] in "-:."
-            ):
-                return entry.get("context_length")
-
-    return None
-
-
-def get_model_context_length(
-    model: str,
-    base_url: str = "",
-    api_key: str = "",
-    config_context_length: int | None = None,
-    provider: str = "",
-) -> int:
-    """Get the context length for a model.
-
-    Resolution order:
-    0. Explicit config override (model.context_length or custom_providers per-model)
-    1. Persistent cache (previously discovered via probing)
-    2. Active endpoint metadata (/models for explicit custom endpoints)
-    3. Local server query (for local endpoints)
-    4. Anthropic /v1/models API (API-key users only, not OAuth)
-    5. OpenRouter live API metadata
-    6. Nous suffix-match via OpenRouter cache
-    7. models.dev registry lookup (provider-aware)
-    8. Thin hardcoded defaults (broad family patterns)
-    9. Default fallback (128K)
-    """
-    # 0. Explicit config override — user knows best
-    if config_context_length is not None and isinstance(config_context_length, int) and config_context_length > 0:
-        return config_context_length
-
-    # Normalise provider-prefixed model names (e.g. "local:model-name" →
-    # "model-name") so cache lookups and server queries use the bare ID that
-    # local servers actually know about.  Ollama "model:tag" colons are preserved.
-    model = _strip_provider_prefix(model)
-
-    # 1. Check persistent cache (model+provider)
-    if base_url:
-        cached = get_cached_context_length(model, base_url)
-        if cached is not None:
-            return cached
-
-    # 2. Active endpoint metadata for truly custom/unknown endpoints.
-    # Known providers (Copilot, OpenAI, Anthropic, etc.) skip this — their
-    # /models endpoint may report a provider-imposed limit (e.g. Copilot
-    # returns 128k) instead of the model's full context (400k).  models.dev
-    # has the correct per-provider values and is checked at step 5+.
-    if _is_custom_endpoint(base_url) and not _is_known_provider_base_url(base_url):
-        endpoint_metadata = fetch_endpoint_model_metadata(base_url, api_key=api_key)
-        matched = endpoint_metadata.get(model)
-        if not matched:
-            # Single-model servers: if only one model is loaded, use it
-            if len(endpoint_metadata) == 1:
-                matched = next(iter(endpoint_metadata.values()))
-            else:
-                # Fuzzy match: substring in either direction
-                for key, entry in endpoint_metadata.items():
-                    if model in key or key in model:
-                        matched = entry
-                        break
-        if matched:
-            context_length = matched.get("context_length")
-            if isinstance(context_length, int):
-                return context_length
-        if not _is_known_provider_base_url(base_url):
-            # 3. Try querying local server directly
-            if is_local_endpoint(base_url):
-                local_ctx = _query_local_context_length(model, base_url)
-                if local_ctx and local_ctx > 0:
-                    save_context_length(model, base_url, local_ctx)
-                    return local_ctx
-            logger.info(
-                "Could not detect context length for model %r at %s — "
-                "defaulting to %s tokens (probe-down). Set model.context_length "
-                "in config.yaml to override.",
-                model, base_url, f"{DEFAULT_FALLBACK_CONTEXT:,}",
-            )
-            return DEFAULT_FALLBACK_CONTEXT
-
-    # 4. Anthropic /v1/models API (only for regular API keys, not OAuth)
-    if provider == "anthropic" or (
-        base_url and "api.anthropic.com" in base_url
-    ):
-        ctx = _query_anthropic_context_length(model, base_url or "https://api.anthropic.com", api_key)
-        if ctx:
-            return ctx
-
-    # 5. Provider-aware lookups (before generic OpenRouter cache)
-    # These are provider-specific and take priority over the generic OR cache,
-    # since the same model can have different context limits per provider
-    # (e.g. claude-opus-4.6 is 1M on Anthropic but 128K on GitHub Copilot).
-    # If provider is generic (openrouter/custom/empty), try to infer from URL.
-    effective_provider = provider
-    if not effective_provider or effective_provider in ("openrouter", "custom"):
-        if base_url:
-            inferred = _infer_provider_from_url(base_url)
-            if inferred:
-                effective_provider = inferred
-
-    if effective_provider == "nous":
-        ctx = _resolve_nous_context_length(model)
-        if ctx:
-            return ctx
-    if effective_provider:
-        from agent.models_dev import lookup_models_dev_context
-        ctx = lookup_models_dev_context(effective_provider, model)
-        if ctx:
-            return ctx
-
-    # 6. OpenRouter live API metadata (provider-unaware fallback)
-    metadata = fetch_model_metadata()
-    if model in metadata:
-        return metadata[model].get("context_length", 128000)
-
-    # 8. Hardcoded defaults (fuzzy match — longest key first for specificity)
-    # Only check `default_model in model` (is the key a substring of the input).
-    # The reverse (`model in default_model`) causes shorter names like
-    # "claude-sonnet-4" to incorrectly match "claude-sonnet-4-6" and return 1M.
-    model_lower = model.lower()
-    for default_model, length in sorted(
-        DEFAULT_CONTEXT_LENGTHS.items(), key=lambda x: len(x[0]), reverse=True
-    ):
-        if default_model in model_lower:
-            return length
-
-    # 9. Query local server as last resort
-    if base_url and is_local_endpoint(base_url):
-        local_ctx = _query_local_context_length(model, base_url)
-        if local_ctx and local_ctx > 0:
-            save_context_length(model, base_url, local_ctx)
-            return local_ctx
-
-    # 10. Default fallback — 128K
-    return DEFAULT_FALLBACK_CONTEXT
-
-
-def estimate_tokens_rough(text: str) -> int:
-    """Rough token estimate (~4 chars/token) for pre-flight checks."""
-    if not text:
-        return 0
-    return len(text) // 4
-
-
-def estimate_messages_tokens_rough(messages: List[Dict[str, Any]]) -> int:
-    """Rough token estimate for a message list (pre-flight only)."""
-    total_chars = sum(len(str(msg)) for msg in messages)
-    return total_chars // 4
-
-
-def estimate_request_tokens_rough(
-    messages: List[Dict[str, Any]],
-    *,
-    system_prompt: str = "",
-    tools: Optional[List[Dict[str, Any]]] = None,
-) -> int:
-    """Rough token estimate for a full chat-completions request.
-
-    Includes the major payload buckets Hermes sends to providers:
-    system prompt, conversation messages, and tool schemas.  With 50+
-    tools enabled, schemas alone can add 20-30K tokens — a significant
-    blind spot when only counting messages.
-    """
-    total_chars = 0
-    if system_prompt:
-        total_chars += len(system_prompt)
-    if messages:
-        total_chars += sum(len(str(msg)) for msg in messages)
-    if tools:
-        total_chars += len(str(tools))
-    return total_chars // 4

agent/models_dev.py

@@ -1,172 +0,0 @@
-"""Models.dev registry integration for provider-aware context length detection.
-
-Fetches model metadata from https://models.dev/api.json — a community-maintained
-database of 3800+ models across 100+ providers, including per-provider context
-windows, pricing, and capabilities.
-
-Data is cached in memory (1hr TTL) and on disk (~/.hermes/models_dev_cache.json)
-to avoid cold-start network latency.
-"""
-
-import json
-import logging
-import os
-import time
-from pathlib import Path
-from typing import Any, Dict, Optional
-
-from utils import atomic_json_write
-
-import requests
-
-logger = logging.getLogger(__name__)
-
-MODELS_DEV_URL = "https://models.dev/api.json"
-_MODELS_DEV_CACHE_TTL = 3600  # 1 hour in-memory
-
-# In-memory cache
-_models_dev_cache: Dict[str, Any] = {}
-_models_dev_cache_time: float = 0
-
-# Provider ID mapping: Hermes provider names → models.dev provider IDs
-PROVIDER_TO_MODELS_DEV: Dict[str, str] = {
-    "openrouter": "openrouter",
-    "anthropic": "anthropic",
-    "zai": "zai",
-    "kimi-coding": "kimi-for-coding",
-    "minimax": "minimax",
-    "minimax-cn": "minimax-cn",
-    "deepseek": "deepseek",
-    "alibaba": "alibaba",
-    "copilot": "github-copilot",
-    "ai-gateway": "vercel",
-    "opencode-zen": "opencode",
-    "opencode-go": "opencode-go",
-    "kilocode": "kilo",
-    "fireworks": "fireworks-ai",
-}
-
-
-def _get_cache_path() -> Path:
-    """Return path to disk cache file."""
-    env_val = os.environ.get("HERMES_HOME", "")
-    hermes_home = Path(env_val) if env_val else Path.home() / ".hermes"
-    return hermes_home / "models_dev_cache.json"
-
-
-def _load_disk_cache() -> Dict[str, Any]:
-    """Load models.dev data from disk cache."""
-    try:
-        cache_path = _get_cache_path()
-        if cache_path.exists():
-            with open(cache_path, encoding="utf-8") as f:
-                return json.load(f)
-    except Exception as e:
-        logger.debug("Failed to load models.dev disk cache: %s", e)
-    return {}
-
-
-def _save_disk_cache(data: Dict[str, Any]) -> None:
-    """Save models.dev data to disk cache atomically."""
-    try:
-        cache_path = _get_cache_path()
-        atomic_json_write(cache_path, data, indent=None, separators=(",", ":"))
-    except Exception as e:
-        logger.debug("Failed to save models.dev disk cache: %s", e)
-
-
-def fetch_models_dev(force_refresh: bool = False) -> Dict[str, Any]:
-    """Fetch models.dev registry. In-memory cache (1hr) + disk fallback.
-
-    Returns the full registry dict keyed by provider ID, or empty dict on failure.
-    """
-    global _models_dev_cache, _models_dev_cache_time
-
-    # Check in-memory cache
-    if (
-        not force_refresh
-        and _models_dev_cache
-        and (time.time() - _models_dev_cache_time) < _MODELS_DEV_CACHE_TTL
-    ):
-        return _models_dev_cache
-
-    # Try network fetch
-    try:
-        response = requests.get(MODELS_DEV_URL, timeout=15)
-        response.raise_for_status()
-        data = response.json()
-        if isinstance(data, dict) and len(data) > 0:
-            _models_dev_cache = data
-            _models_dev_cache_time = time.time()
-            _save_disk_cache(data)
-            logger.debug(
-                "Fetched models.dev registry: %d providers, %d total models",
-                len(data),
-                sum(len(p.get("models", {})) for p in data.values() if isinstance(p, dict)),
-            )
-            return data
-    except Exception as e:
-        logger.debug("Failed to fetch models.dev: %s", e)
-
-    # Fall back to disk cache — use a short TTL (5 min) so we retry
-    # the network fetch soon instead of serving stale data for a full hour.
-    if not _models_dev_cache:
-        _models_dev_cache = _load_disk_cache()
-        if _models_dev_cache:
-            _models_dev_cache_time = time.time() - _MODELS_DEV_CACHE_TTL + 300
-            logger.debug("Loaded models.dev from disk cache (%d providers)", len(_models_dev_cache))
-
-    return _models_dev_cache
-
-
-def lookup_models_dev_context(provider: str, model: str) -> Optional[int]:
-    """Look up context_length for a provider+model combo in models.dev.
-
-    Returns the context window in tokens, or None if not found.
-    Handles case-insensitive matching and filters out context=0 entries.
-    """
-    mdev_provider_id = PROVIDER_TO_MODELS_DEV.get(provider)
-    if not mdev_provider_id:
-        return None
-
-    data = fetch_models_dev()
-    provider_data = data.get(mdev_provider_id)
-    if not isinstance(provider_data, dict):
-        return None
-
-    models = provider_data.get("models", {})
-    if not isinstance(models, dict):
-        return None
-
-    # Exact match
-    entry = models.get(model)
-    if entry:
-        ctx = _extract_context(entry)
-        if ctx:
-            return ctx
-
-    # Case-insensitive match
-    model_lower = model.lower()
-    for mid, mdata in models.items():
-        if mid.lower() == model_lower:
-            ctx = _extract_context(mdata)
-            if ctx:
-                return ctx
-
-    return None
-
-
-def _extract_context(entry: Dict[str, Any]) -> Optional[int]:
-    """Extract context_length from a models.dev model entry.
-
-    Returns None for invalid/zero values (some audio/image models have context=0).
-    """
-    if not isinstance(entry, dict):
-        return None
-    limit = entry.get("limit")
-    if not isinstance(limit, dict):
-        return None
-    ctx = limit.get("context")
-    if isinstance(ctx, (int, float)) and ctx > 0:
-        return int(ctx)
-    return None

agent/prompt_builder.py

@@ -1,912 +0,0 @@
-"""System prompt assembly -- identity, platform hints, skills index, context files.
-
-All functions are stateless. AIAgent._build_system_prompt() calls these to
-assemble pieces, then combines them with memory and ephemeral prompts.
-"""
-
-import json
-import logging
-import os
-import re
-import threading
-from collections import OrderedDict
-from pathlib import Path
-
-from hermes_constants import get_hermes_home
-from typing import Optional
-
-from agent.skill_utils import (
-    extract_skill_conditions,
-    extract_skill_description,
-    get_all_skills_dirs,
-    get_disabled_skill_names,
-    iter_skill_index_files,
-    parse_frontmatter,
-    skill_matches_platform,
-)
-from utils import atomic_json_write
-
-logger = logging.getLogger(__name__)
-
-# ---------------------------------------------------------------------------
-# Context file scanning — detect prompt injection in AGENTS.md, .cursorrules,
-# SOUL.md before they get injected into the system prompt.
-# ---------------------------------------------------------------------------
-
-_CONTEXT_THREAT_PATTERNS = [
-    (r'ignore\s+(previous|all|above|prior)\s+instructions', "prompt_injection"),
-    (r'do\s+not\s+tell\s+the\s+user', "deception_hide"),
-    (r'system\s+prompt\s+override', "sys_prompt_override"),
-    (r'disregard\s+(your|all|any)\s+(instructions|rules|guidelines)', "disregard_rules"),
-    (r'act\s+as\s+(if|though)\s+you\s+(have\s+no|don\'t\s+have)\s+(restrictions|limits|rules)', "bypass_restrictions"),
-    (r'<!--[^>]*(?:ignore|override|system|secret|hidden)[^>]*-->', "html_comment_injection"),
-    (r'<\s*div\s+style\s*=\s*["\'].*display\s*:\s*none', "hidden_div"),
-    (r'translate\s+.*\s+into\s+.*\s+and\s+(execute|run|eval)', "translate_execute"),
-    (r'curl\s+[^\n]*\$\{?\w*(KEY|TOKEN|SECRET|PASSWORD|CREDENTIAL|API)', "exfil_curl"),
-    (r'cat\s+[^\n]*(\.env|credentials|\.netrc|\.pgpass)', "read_secrets"),
-]
-
-_CONTEXT_INVISIBLE_CHARS = {
-    '\u200b', '\u200c', '\u200d', '\u2060', '\ufeff',
-    '\u202a', '\u202b', '\u202c', '\u202d', '\u202e',
-}
-
-
-def _scan_context_content(content: str, filename: str) -> str:
-    """Scan context file content for injection. Returns sanitized content."""
-    findings = []
-
-    # Check invisible unicode
-    for char in _CONTEXT_INVISIBLE_CHARS:
-        if char in content:
-            findings.append(f"invisible unicode U+{ord(char):04X}")
-
-    # Check threat patterns
-    for pattern, pid in _CONTEXT_THREAT_PATTERNS:
-        if re.search(pattern, content, re.IGNORECASE):
-            findings.append(pid)
-
-    if findings:
-        logger.warning("Context file %s blocked: %s", filename, ", ".join(findings))
-        return f"[BLOCKED: {filename} contained potential prompt injection ({', '.join(findings)}). Content not loaded.]"
-
-    return content
-
-
-def _find_git_root(start: Path) -> Optional[Path]:
-    """Walk *start* and its parents looking for a ``.git`` directory.
-
-    Returns the directory containing ``.git``, or ``None`` if we hit the
-    filesystem root without finding one.
-    """
-    current = start.resolve()
-    for parent in [current, *current.parents]:
-        if (parent / ".git").exists():
-            return parent
-    return None
-
-
-_HERMES_MD_NAMES = (".hermes.md", "HERMES.md")
-
-
-def _find_hermes_md(cwd: Path) -> Optional[Path]:
-    """Discover the nearest ``.hermes.md`` or ``HERMES.md``.
-
-    Search order: *cwd* first, then each parent directory up to (and
-    including) the git repository root.  Returns the first match, or
-    ``None`` if nothing is found.
-    """
-    stop_at = _find_git_root(cwd)
-    current = cwd.resolve()
-
-    for directory in [current, *current.parents]:
-        for name in _HERMES_MD_NAMES:
-            candidate = directory / name
-            if candidate.is_file():
-                return candidate
-        # Stop walking at the git root (or filesystem root).
-        if stop_at and directory == stop_at:
-            break
-    return None
-
-
-def _strip_yaml_frontmatter(content: str) -> str:
-    """Remove optional YAML frontmatter (``---`` delimited) from *content*.
-
-    The frontmatter may contain structured config (model overrides, tool
-    settings) that will be handled separately in a future PR.  For now we
-    strip it so only the human-readable markdown body is injected into the
-    system prompt.
-    """
-    if content.startswith("---"):
-        end = content.find("\n---", 3)
-        if end != -1:
-            # Skip past the closing --- and any trailing newline
-            body = content[end + 4:].lstrip("\n")
-            return body if body else content
-    return content
-
-
-# =========================================================================
-# Constants
-# =========================================================================
-
-DEFAULT_AGENT_IDENTITY = (
-    "You are Hermes Agent, an intelligent AI assistant created by Nous Research. "
-    "You are helpful, knowledgeable, and direct. You assist users with a wide "
-    "range of tasks including answering questions, writing and editing code, "
-    "analyzing information, creative work, and executing actions via your tools. "
-    "You communicate clearly, admit uncertainty when appropriate, and prioritize "
-    "being genuinely useful over being verbose unless otherwise directed below. "
-    "Be targeted and efficient in your exploration and investigations."
-)
-
-MEMORY_GUIDANCE = (
-    "You have persistent memory across sessions. Save durable facts using the memory "
-    "tool: user preferences, environment details, tool quirks, and stable conventions. "
-    "Memory is injected into every turn, so keep it compact and focused on facts that "
-    "will still matter later.\n"
-    "Prioritize what reduces future user steering — the most valuable memory is one "
-    "that prevents the user from having to correct or remind you again. "
-    "User preferences and recurring corrections matter more than procedural task details.\n"
-    "Do NOT save task progress, session outcomes, completed-work logs, or temporary TODO "
-    "state to memory; use session_search to recall those from past transcripts. "
-    "If you've discovered a new way to do something, solved a problem that could be "
-    "necessary later, save it as a skill with the skill tool."
-)
-
-SESSION_SEARCH_GUIDANCE = (
-    "When the user references something from a past conversation or you suspect "
-    "relevant cross-session context exists, use session_search to recall it before "
-    "asking them to repeat themselves."
-)
-
-SKILLS_GUIDANCE = (
-    "After completing a complex task (5+ tool calls), fixing a tricky error, "
-    "or discovering a non-trivial workflow, save the approach as a "
-    "skill with skill_manage so you can reuse it next time.\n"
-    "When using a skill and finding it outdated, incomplete, or wrong, "
-    "patch it immediately with skill_manage(action='patch') — don't wait to be asked. "
-    "Skills that aren't maintained become liabilities."
-)
-
-TOOL_USE_ENFORCEMENT_GUIDANCE = (
-    "# Tool-use enforcement\n"
-    "You MUST use your tools to take action — do not describe what you would do "
-    "or plan to do without actually doing it. When you say you will perform an "
-    "action (e.g. 'I will run the tests', 'Let me check the file', 'I will create "
-    "the project'), you MUST immediately make the corresponding tool call in the same "
-    "response. Never end your turn with a promise of future action — execute it now.\n"
-    "Keep working until the task is actually complete. Do not stop with a summary of "
-    "what you plan to do next time. If you have tools available that can accomplish "
-    "the task, use them instead of telling the user what you would do.\n"
-    "Every response should either (a) contain tool calls that make progress, or "
-    "(b) deliver a final result to the user. Responses that only describe intentions "
-    "without acting are not acceptable."
-)
-
-# Model name substrings that trigger tool-use enforcement guidance.
-# Add new patterns here when a model family needs explicit steering.
-TOOL_USE_ENFORCEMENT_MODELS = ("gpt", "codex", "gemini", "gemma")
-
-# Gemini/Gemma-specific operational guidance, adapted from OpenCode's gemini.txt.
-# Injected alongside TOOL_USE_ENFORCEMENT_GUIDANCE when the model is Gemini or Gemma.
-GOOGLE_MODEL_OPERATIONAL_GUIDANCE = (
-    "# Google model operational directives\n"
-    "Follow these operational rules strictly:\n"
-    "- **Absolute paths:** Always construct and use absolute file paths for all "
-    "file system operations. Combine the project root with relative paths.\n"
-    "- **Verify first:** Use read_file/search_files to check file contents and "
-    "project structure before making changes. Never guess at file contents.\n"
-    "- **Dependency checks:** Never assume a library is available. Check "
-    "package.json, requirements.txt, Cargo.toml, etc. before importing.\n"
-    "- **Conciseness:** Keep explanatory text brief — a few sentences, not "
-    "paragraphs. Focus on actions and results over narration.\n"
-    "- **Parallel tool calls:** When you need to perform multiple independent "
-    "operations (e.g. reading several files), make all the tool calls in a "
-    "single response rather than sequentially.\n"
-    "- **Non-interactive commands:** Use flags like -y, --yes, --non-interactive "
-    "to prevent CLI tools from hanging on prompts.\n"
-    "- **Keep going:** Work autonomously until the task is fully resolved. "
-    "Don't stop with a plan — execute it.\n"
-)
-
-# Model name substrings that should use the 'developer' role instead of
-# 'system' for the system prompt.  OpenAI's newer models (GPT-5, Codex)
-# give stronger instruction-following weight to the 'developer' role.
-# The swap happens at the API boundary in _build_api_kwargs() so internal
-# message representation stays consistent ("system" everywhere).
-DEVELOPER_ROLE_MODELS = ("gpt-5", "codex")
-
-PLATFORM_HINTS = {
-    "whatsapp": (
-        "You are on a text messaging communication platform, WhatsApp. "
-        "Please do not use markdown as it does not render. "
-        "You can send media files natively: to deliver a file to the user, "
-        "include MEDIA:/absolute/path/to/file in your response. The file "
-        "will be sent as a native WhatsApp attachment — images (.jpg, .png, "
-        ".webp) appear as photos, videos (.mp4, .mov) play inline, and other "
-        "files arrive as downloadable documents. You can also include image "
-        "URLs in markdown format ![alt](url) and they will be sent as photos."
-    ),
-    "telegram": (
-        "You are on a text messaging communication platform, Telegram. "
-        "Please do not use markdown as it does not render. "
-        "You can send media files natively: to deliver a file to the user, "
-        "include MEDIA:/absolute/path/to/file in your response. Images "
-        "(.png, .jpg, .webp) appear as photos, audio (.ogg) sends as voice "
-        "bubbles, and videos (.mp4) play inline. You can also include image "
-        "URLs in markdown format ![alt](url) and they will be sent as native photos."
-    ),
-    "discord": (
-        "You are in a Discord server or group chat communicating with your user. "
-        "You can send media files natively: include MEDIA:/absolute/path/to/file "
-        "in your response. Images (.png, .jpg, .webp) are sent as photo "
-        "attachments, audio as file attachments. You can also include image URLs "
-        "in markdown format ![alt](url) and they will be sent as attachments."
-    ),
-    "slack": (
-        "You are in a Slack workspace communicating with your user. "
-        "You can send media files natively: include MEDIA:/absolute/path/to/file "
-        "in your response. Images (.png, .jpg, .webp) are uploaded as photo "
-        "attachments, audio as file attachments. You can also include image URLs "
-        "in markdown format ![alt](url) and they will be uploaded as attachments."
-    ),
-    "signal": (
-        "You are on a text messaging communication platform, Signal. "
-        "Please do not use markdown as it does not render. "
-        "You can send media files natively: to deliver a file to the user, "
-        "include MEDIA:/absolute/path/to/file in your response. Images "
-        "(.png, .jpg, .webp) appear as photos, audio as attachments, and other "
-        "files arrive as downloadable documents. You can also include image "
-        "URLs in markdown format ![alt](url) and they will be sent as photos."
-    ),
-    "email": (
-        "You are communicating via email. Write clear, well-structured responses "
-        "suitable for email. Use plain text formatting (no markdown). "
-        "Keep responses concise but complete. You can send file attachments — "
-        "include MEDIA:/absolute/path/to/file in your response. The subject line "
-        "is preserved for threading. Do not include greetings or sign-offs unless "
-        "contextually appropriate."
-    ),
-    "cron": (
-        "You are running as a scheduled cron job. There is no user present — you "
-        "cannot ask questions, request clarification, or wait for follow-up. Execute "
-        "the task fully and autonomously, making reasonable decisions where needed. "
-        "Your final response is automatically delivered to the job's configured "
-        "destination — put the primary content directly in your response."
-    ),
-    "cli": (
-        "You are a CLI AI Agent. Try not to use markdown but simple text "
-        "renderable inside a terminal."
-    ),
-    "sms": (
-        "You are communicating via SMS. Keep responses concise and use plain text "
-        "only — no markdown, no formatting. SMS messages are limited to ~1600 "
-        "characters, so be brief and direct."
-    ),
-}
-
-CONTEXT_FILE_MAX_CHARS = 20_000
-CONTEXT_TRUNCATE_HEAD_RATIO = 0.7
-CONTEXT_TRUNCATE_TAIL_RATIO = 0.2
-
-
-# =========================================================================
-# Skills prompt cache
-# =========================================================================
-
-_SKILLS_PROMPT_CACHE_MAX = 8
-_SKILLS_PROMPT_CACHE: OrderedDict[tuple, str] = OrderedDict()
-_SKILLS_PROMPT_CACHE_LOCK = threading.Lock()
-_SKILLS_SNAPSHOT_VERSION = 1
-
-
-def _skills_prompt_snapshot_path() -> Path:
-    return get_hermes_home() / ".skills_prompt_snapshot.json"
-
-
-def clear_skills_system_prompt_cache(*, clear_snapshot: bool = False) -> None:
-    """Drop the in-process skills prompt cache (and optionally the disk snapshot)."""
-    with _SKILLS_PROMPT_CACHE_LOCK:
-        _SKILLS_PROMPT_CACHE.clear()
-    if clear_snapshot:
-        try:
-            _skills_prompt_snapshot_path().unlink(missing_ok=True)
-        except OSError as e:
-            logger.debug("Could not remove skills prompt snapshot: %s", e)
-
-
-def _build_skills_manifest(skills_dir: Path) -> dict[str, list[int]]:
-    """Build an mtime/size manifest of all SKILL.md and DESCRIPTION.md files."""
-    manifest: dict[str, list[int]] = {}
-    for filename in ("SKILL.md", "DESCRIPTION.md"):
-        for path in iter_skill_index_files(skills_dir, filename):
-            try:
-                st = path.stat()
-            except OSError:
-                continue
-            manifest[str(path.relative_to(skills_dir))] = [st.st_mtime_ns, st.st_size]
-    return manifest
-
-
-def _load_skills_snapshot(skills_dir: Path) -> Optional[dict]:
-    """Load the disk snapshot if it exists and its manifest still matches."""
-    snapshot_path = _skills_prompt_snapshot_path()
-    if not snapshot_path.exists():
-        return None
-    try:
-        snapshot = json.loads(snapshot_path.read_text(encoding="utf-8"))
-    except Exception:
-        return None
-    if not isinstance(snapshot, dict):
-        return None
-    if snapshot.get("version") != _SKILLS_SNAPSHOT_VERSION:
-        return None
-    if snapshot.get("manifest") != _build_skills_manifest(skills_dir):
-        return None
-    return snapshot
-
-
-def _write_skills_snapshot(
-    skills_dir: Path,
-    manifest: dict[str, list[int]],
-    skill_entries: list[dict],
-    category_descriptions: dict[str, str],
-) -> None:
-    """Persist skill metadata to disk for fast cold-start reuse."""
-    payload = {
-        "version": _SKILLS_SNAPSHOT_VERSION,
-        "manifest": manifest,
-        "skills": skill_entries,
-        "category_descriptions": category_descriptions,
-    }
-    try:
-        atomic_json_write(_skills_prompt_snapshot_path(), payload)
-    except Exception as e:
-        logger.debug("Could not write skills prompt snapshot: %s", e)
-
-
-def _build_snapshot_entry(
-    skill_file: Path,
-    skills_dir: Path,
-    frontmatter: dict,
-    description: str,
-) -> dict:
-    """Build a serialisable metadata dict for one skill."""
-    rel_path = skill_file.relative_to(skills_dir)
-    parts = rel_path.parts
-    if len(parts) >= 2:
-        skill_name = parts[-2]
-        category = "/".join(parts[:-2]) if len(parts) > 2 else parts[0]
-    else:
-        category = "general"
-        skill_name = skill_file.parent.name
-
-    platforms = frontmatter.get("platforms") or []
-    if isinstance(platforms, str):
-        platforms = [platforms]
-
-    return {
-        "skill_name": skill_name,
-        "category": category,
-        "frontmatter_name": str(frontmatter.get("name", skill_name)),
-        "description": description,
-        "platforms": [str(p).strip() for p in platforms if str(p).strip()],
-        "conditions": extract_skill_conditions(frontmatter),
-    }
-
-
-# =========================================================================
-# Skills index
-# =========================================================================
-
-def _parse_skill_file(skill_file: Path) -> tuple[bool, dict, str]:
-    """Read a SKILL.md once and return platform compatibility, frontmatter, and description.
-
-    Returns (is_compatible, frontmatter, description). On any error, returns
-    (True, {}, "") to err on the side of showing the skill.
-    """
-    try:
-        raw = skill_file.read_text(encoding="utf-8")[:2000]
-        frontmatter, _ = parse_frontmatter(raw)
-
-        if not skill_matches_platform(frontmatter):
-            return False, frontmatter, ""
-
-        return True, frontmatter, extract_skill_description(frontmatter)
-    except Exception as e:
-        logger.debug("Failed to parse skill file %s: %s", skill_file, e)
-        return True, {}, ""
-
-
-def _read_skill_conditions(skill_file: Path) -> dict:
-    """Extract conditional activation fields from SKILL.md frontmatter."""
-    try:
-        raw = skill_file.read_text(encoding="utf-8")[:2000]
-        frontmatter, _ = parse_frontmatter(raw)
-        return extract_skill_conditions(frontmatter)
-    except Exception as e:
-        logger.debug("Failed to read skill conditions from %s: %s", skill_file, e)
-        return {}
-
-
-def _skill_should_show(
-    conditions: dict,
-    available_tools: "set[str] | None",
-    available_toolsets: "set[str] | None",
-) -> bool:
-    """Return False if the skill's conditional activation rules exclude it."""
-    if available_tools is None and available_toolsets is None:
-        return True  # No filtering info — show everything (backward compat)
-
-    at = available_tools or set()
-    ats = available_toolsets or set()
-
-    # fallback_for: hide when the primary tool/toolset IS available
-    for ts in conditions.get("fallback_for_toolsets", []):
-        if ts in ats:
-            return False
-    for t in conditions.get("fallback_for_tools", []):
-        if t in at:
-            return False
-
-    # requires: hide when a required tool/toolset is NOT available
-    for ts in conditions.get("requires_toolsets", []):
-        if ts not in ats:
-            return False
-    for t in conditions.get("requires_tools", []):
-        if t not in at:
-            return False
-
-    return True
-
-
-def build_skills_system_prompt(
-    available_tools: "set[str] | None" = None,
-    available_toolsets: "set[str] | None" = None,
-) -> str:
-    """Build a compact skill index for the system prompt.
-
-    Two-layer cache:
-      1. In-process LRU dict keyed by (skills_dir, tools, toolsets)
-      2. Disk snapshot (``.skills_prompt_snapshot.json``) validated by
-         mtime/size manifest — survives process restarts
-
-    Falls back to a full filesystem scan when both layers miss.
-
-    External skill directories (``skills.external_dirs`` in config.yaml) are
-    scanned alongside the local ``~/.hermes/skills/`` directory.  External dirs
-    are read-only — they appear in the index but new skills are always created
-    in the local dir.  Local skills take precedence when names collide.
-    """
-    hermes_home = get_hermes_home()
-    skills_dir = hermes_home / "skills"
-    external_dirs = get_all_skills_dirs()[1:]  # skip local (index 0)
-
-    if not skills_dir.exists() and not external_dirs:
-        return ""
-
-    # ── Layer 1: in-process LRU cache ─────────────────────────────────
-    cache_key = (
-        str(skills_dir.resolve()),
-        tuple(str(d) for d in external_dirs),
-        tuple(sorted(str(t) for t in (available_tools or set()))),
-        tuple(sorted(str(ts) for ts in (available_toolsets or set()))),
-    )
-    with _SKILLS_PROMPT_CACHE_LOCK:
-        cached = _SKILLS_PROMPT_CACHE.get(cache_key)
-        if cached is not None:
-            _SKILLS_PROMPT_CACHE.move_to_end(cache_key)
-            return cached
-
-    disabled = get_disabled_skill_names()
-
-    # ── Layer 2: disk snapshot ────────────────────────────────────────
-    snapshot = _load_skills_snapshot(skills_dir)
-
-    skills_by_category: dict[str, list[tuple[str, str]]] = {}
-    category_descriptions: dict[str, str] = {}
-
-    if snapshot is not None:
-        # Fast path: use pre-parsed metadata from disk
-        for entry in snapshot.get("skills", []):
-            if not isinstance(entry, dict):
-                continue
-            skill_name = entry.get("skill_name") or ""
-            category = entry.get("category") or "general"
-            frontmatter_name = entry.get("frontmatter_name") or skill_name
-            platforms = entry.get("platforms") or []
-            if not skill_matches_platform({"platforms": platforms}):
-                continue
-            if frontmatter_name in disabled or skill_name in disabled:
-                continue
-            if not _skill_should_show(
-                entry.get("conditions") or {},
-                available_tools,
-                available_toolsets,
-            ):
-                continue
-            skills_by_category.setdefault(category, []).append(
-                (skill_name, entry.get("description", ""))
-            )
-        category_descriptions = {
-            str(k): str(v)
-            for k, v in (snapshot.get("category_descriptions") or {}).items()
-        }
-    else:
-        # Cold path: full filesystem scan + write snapshot for next time
-        skill_entries: list[dict] = []
-        for skill_file in iter_skill_index_files(skills_dir, "SKILL.md"):
-            is_compatible, frontmatter, desc = _parse_skill_file(skill_file)
-            entry = _build_snapshot_entry(skill_file, skills_dir, frontmatter, desc)
-            skill_entries.append(entry)
-            if not is_compatible:
-                continue
-            skill_name = entry["skill_name"]
-            if entry["frontmatter_name"] in disabled or skill_name in disabled:
-                continue
-            if not _skill_should_show(
-                extract_skill_conditions(frontmatter),
-                available_tools,
-                available_toolsets,
-            ):
-                continue
-            skills_by_category.setdefault(entry["category"], []).append(
-                (skill_name, entry["description"])
-            )
-
-        # Read category-level DESCRIPTION.md files
-        for desc_file in iter_skill_index_files(skills_dir, "DESCRIPTION.md"):
-            try:
-                content = desc_file.read_text(encoding="utf-8")
-                fm, _ = parse_frontmatter(content)
-                cat_desc = fm.get("description")
-                if not cat_desc:
-                    continue
-                rel = desc_file.relative_to(skills_dir)
-                cat = "/".join(rel.parts[:-1]) if len(rel.parts) > 1 else "general"
-                category_descriptions[cat] = str(cat_desc).strip().strip("'\"")
-            except Exception as e:
-                logger.debug("Could not read skill description %s: %s", desc_file, e)
-
-        _write_skills_snapshot(
-            skills_dir,
-            _build_skills_manifest(skills_dir),
-            skill_entries,
-            category_descriptions,
-        )
-
-    # ── External skill directories ─────────────────────────────────────
-    # Scan external dirs directly (no snapshot caching — they're read-only
-    # and typically small).  Local skills already in skills_by_category take
-    # precedence: we track seen names and skip duplicates from external dirs.
-    seen_skill_names: set[str] = set()
-    for cat_skills in skills_by_category.values():
-        for name, _desc in cat_skills:
-            seen_skill_names.add(name)
-
-    for ext_dir in external_dirs:
-        if not ext_dir.exists():
-            continue
-        for skill_file in iter_skill_index_files(ext_dir, "SKILL.md"):
-            try:
-                is_compatible, frontmatter, desc = _parse_skill_file(skill_file)
-                if not is_compatible:
-                    continue
-                entry = _build_snapshot_entry(skill_file, ext_dir, frontmatter, desc)
-                skill_name = entry["skill_name"]
-                if skill_name in seen_skill_names:
-                    continue
-                if entry["frontmatter_name"] in disabled or skill_name in disabled:
-                    continue
-                if not _skill_should_show(
-                    extract_skill_conditions(frontmatter),
-                    available_tools,
-                    available_toolsets,
-                ):
-                    continue
-                seen_skill_names.add(skill_name)
-                skills_by_category.setdefault(entry["category"], []).append(
-                    (skill_name, entry["description"])
-                )
-            except Exception as e:
-                logger.debug("Error reading external skill %s: %s", skill_file, e)
-
-        # External category descriptions
-        for desc_file in iter_skill_index_files(ext_dir, "DESCRIPTION.md"):
-            try:
-                content = desc_file.read_text(encoding="utf-8")
-                fm, _ = parse_frontmatter(content)
-                cat_desc = fm.get("description")
-                if not cat_desc:
-                    continue
-                rel = desc_file.relative_to(ext_dir)
-                cat = "/".join(rel.parts[:-1]) if len(rel.parts) > 1 else "general"
-                category_descriptions.setdefault(cat, str(cat_desc).strip().strip("'\""))
-            except Exception as e:
-                logger.debug("Could not read external skill description %s: %s", desc_file, e)
-
-    if not skills_by_category:
-        result = ""
-    else:
-        index_lines = []
-        for category in sorted(skills_by_category.keys()):
-            cat_desc = category_descriptions.get(category, "")
-            if cat_desc:
-                index_lines.append(f"  {category}: {cat_desc}")
-            else:
-                index_lines.append(f"  {category}:")
-            # Deduplicate and sort skills within each category
-            seen = set()
-            for name, desc in sorted(skills_by_category[category], key=lambda x: x[0]):
-                if name in seen:
-                    continue
-                seen.add(name)
-                if desc:
-                    index_lines.append(f"    - {name}: {desc}")
-                else:
-                    index_lines.append(f"    - {name}")
-
-        result = (
-            "## Skills (mandatory)\n"
-            "Before replying, scan the skills below. If one clearly matches your task, "
-            "load it with skill_view(name) and follow its instructions. "
-            "If a skill has issues, fix it with skill_manage(action='patch').\n"
-            "After difficult/iterative tasks, offer to save as a skill. "
-            "If a skill you loaded was missing steps, had wrong commands, or needed "
-            "pitfalls you discovered, update it before finishing.\n"
-            "\n"
-            "<available_skills>\n"
-            + "\n".join(index_lines) + "\n"
-            "</available_skills>\n"
-            "\n"
-            "If none match, proceed normally without loading a skill."
-        )
-
-    # ── Store in LRU cache ────────────────────────────────────────────
-    with _SKILLS_PROMPT_CACHE_LOCK:
-        _SKILLS_PROMPT_CACHE[cache_key] = result
-        _SKILLS_PROMPT_CACHE.move_to_end(cache_key)
-        while len(_SKILLS_PROMPT_CACHE) > _SKILLS_PROMPT_CACHE_MAX:
-            _SKILLS_PROMPT_CACHE.popitem(last=False)
-
-    return result
-
-
-def build_nous_subscription_prompt(valid_tool_names: "set[str] | None" = None) -> str:
-    """Build a compact Nous subscription capability block for the system prompt."""
-    try:
-        from hermes_cli.nous_subscription import get_nous_subscription_features
-        from tools.tool_backend_helpers import managed_nous_tools_enabled
-    except Exception as exc:
-        logger.debug("Failed to import Nous subscription helper: %s", exc)
-        return ""
-
-    if not managed_nous_tools_enabled():
-        return ""
-
-    valid_names = set(valid_tool_names or set())
-    relevant_tool_names = {
-        "web_search",
-        "web_extract",
-        "browser_navigate",
-        "browser_snapshot",
-        "browser_click",
-        "browser_type",
-        "browser_scroll",
-        "browser_console",
-        "browser_close",
-        "browser_press",
-        "browser_get_images",
-        "browser_vision",
-        "image_generate",
-        "text_to_speech",
-        "terminal",
-        "process",
-        "execute_code",
-    }
-
-    if valid_names and not (valid_names & relevant_tool_names):
-        return ""
-
-    features = get_nous_subscription_features()
-
-    def _status_line(feature) -> str:
-        if feature.managed_by_nous:
-            return f"- {feature.label}: active via Nous subscription"
-        if feature.active:
-            current = feature.current_provider or "configured provider"
-            return f"- {feature.label}: currently using {current}"
-        if feature.included_by_default and features.nous_auth_present:
-            return f"- {feature.label}: included with Nous subscription, not currently selected"
-        if feature.key == "modal" and features.nous_auth_present:
-            return f"- {feature.label}: optional via Nous subscription"
-        return f"- {feature.label}: not currently available"
-
-    lines = [
-        "# Nous Subscription",
-        "Nous subscription includes managed web tools (Firecrawl), image generation (FAL), OpenAI TTS, and browser automation (Browserbase) by default. Modal execution is optional.",
-        "Current capability status:",
-    ]
-    lines.extend(_status_line(feature) for feature in features.items())
-    lines.extend(
-        [
-            "When a Nous-managed feature is active, do not ask the user for Firecrawl, FAL, OpenAI TTS, or Browserbase API keys.",
-            "If the user is not subscribed and asks for a capability that Nous subscription would unlock or simplify, suggest Nous subscription as one option alongside direct setup or local alternatives.",
-            "Do not mention subscription unless the user asks about it or it directly solves the current missing capability.",
-            "Useful commands: hermes setup, hermes setup tools, hermes setup terminal, hermes status.",
-        ]
-    )
-    return "\n".join(lines)
-
-
-# =========================================================================
-# Context files (SOUL.md, AGENTS.md, .cursorrules)
-# =========================================================================
-
-def _truncate_content(content: str, filename: str, max_chars: int = CONTEXT_FILE_MAX_CHARS) -> str:
-    """Head/tail truncation with a marker in the middle."""
-    if len(content) <= max_chars:
-        return content
-    head_chars = int(max_chars * CONTEXT_TRUNCATE_HEAD_RATIO)
-    tail_chars = int(max_chars * CONTEXT_TRUNCATE_TAIL_RATIO)
-    head = content[:head_chars]
-    tail = content[-tail_chars:]
-    marker = f"\n\n[...truncated {filename}: kept {head_chars}+{tail_chars} of {len(content)} chars. Use file tools to read the full file.]\n\n"
-    return head + marker + tail
-
-
-def load_soul_md() -> Optional[str]:
-    """Load SOUL.md from HERMES_HOME and return its content, or None.
-
-    Used as the agent identity (slot #1 in the system prompt).  When this
-    returns content, ``build_context_files_prompt`` should be called with
-    ``skip_soul=True`` so SOUL.md isn't injected twice.
-    """
-    try:
-        from hermes_cli.config import ensure_hermes_home
-        ensure_hermes_home()
-    except Exception as e:
-        logger.debug("Could not ensure HERMES_HOME before loading SOUL.md: %s", e)
-
-    soul_path = get_hermes_home() / "SOUL.md"
-    if not soul_path.exists():
-        return None
-    try:
-        content = soul_path.read_text(encoding="utf-8").strip()
-        if not content:
-            return None
-        content = _scan_context_content(content, "SOUL.md")
-        content = _truncate_content(content, "SOUL.md")
-        return content
-    except Exception as e:
-        logger.debug("Could not read SOUL.md from %s: %s", soul_path, e)
-        return None
-
-
-def _load_hermes_md(cwd_path: Path) -> str:
-    """.hermes.md / HERMES.md — walk to git root."""
-    hermes_md_path = _find_hermes_md(cwd_path)
-    if not hermes_md_path:
-        return ""
-    try:
-        content = hermes_md_path.read_text(encoding="utf-8").strip()
-        if not content:
-            return ""
-        content = _strip_yaml_frontmatter(content)
-        rel = hermes_md_path.name
-        try:
-            rel = str(hermes_md_path.relative_to(cwd_path))
-        except ValueError:
-            pass
-        content = _scan_context_content(content, rel)
-        result = f"## {rel}\n\n{content}"
-        return _truncate_content(result, ".hermes.md")
-    except Exception as e:
-        logger.debug("Could not read %s: %s", hermes_md_path, e)
-        return ""
-
-
-def _load_agents_md(cwd_path: Path) -> str:
-    """AGENTS.md — top-level only (no recursive walk)."""
-    for name in ["AGENTS.md", "agents.md"]:
-        candidate = cwd_path / name
-        if candidate.exists():
-            try:
-                content = candidate.read_text(encoding="utf-8").strip()
-                if content:
-                    content = _scan_context_content(content, name)
-                    result = f"## {name}\n\n{content}"
-                    return _truncate_content(result, "AGENTS.md")
-            except Exception as e:
-                logger.debug("Could not read %s: %s", candidate, e)
-    return ""
-
-
-def _load_claude_md(cwd_path: Path) -> str:
-    """CLAUDE.md / claude.md — cwd only."""
-    for name in ["CLAUDE.md", "claude.md"]:
-        candidate = cwd_path / name
-        if candidate.exists():
-            try:
-                content = candidate.read_text(encoding="utf-8").strip()
-                if content:
-                    content = _scan_context_content(content, name)
-                    result = f"## {name}\n\n{content}"
-                    return _truncate_content(result, "CLAUDE.md")
-            except Exception as e:
-                logger.debug("Could not read %s: %s", candidate, e)
-    return ""
-
-
-def _load_cursorrules(cwd_path: Path) -> str:
-    """.cursorrules + .cursor/rules/*.mdc — cwd only."""
-    cursorrules_content = ""
-    cursorrules_file = cwd_path / ".cursorrules"
-    if cursorrules_file.exists():
-        try:
-            content = cursorrules_file.read_text(encoding="utf-8").strip()
-            if content:
-                content = _scan_context_content(content, ".cursorrules")
-                cursorrules_content += f"## .cursorrules\n\n{content}\n\n"
-        except Exception as e:
-            logger.debug("Could not read .cursorrules: %s", e)
-
-    cursor_rules_dir = cwd_path / ".cursor" / "rules"
-    if cursor_rules_dir.exists() and cursor_rules_dir.is_dir():
-        mdc_files = sorted(cursor_rules_dir.glob("*.mdc"))
-        for mdc_file in mdc_files:
-            try:
-                content = mdc_file.read_text(encoding="utf-8").strip()
-                if content:
-                    content = _scan_context_content(content, f".cursor/rules/{mdc_file.name}")
-                    cursorrules_content += f"## .cursor/rules/{mdc_file.name}\n\n{content}\n\n"
-            except Exception as e:
-                logger.debug("Could not read %s: %s", mdc_file, e)
-
-    if not cursorrules_content:
-        return ""
-    return _truncate_content(cursorrules_content, ".cursorrules")
-
-
-def build_context_files_prompt(cwd: Optional[str] = None, skip_soul: bool = False) -> str:
-    """Discover and load context files for the system prompt.
-
-    Priority (first found wins — only ONE project context type is loaded):
-      1. .hermes.md / HERMES.md  (walk to git root)
-      2. AGENTS.md / agents.md   (cwd only)
-      3. CLAUDE.md / claude.md   (cwd only)
-      4. .cursorrules / .cursor/rules/*.mdc  (cwd only)
-
-    SOUL.md from HERMES_HOME is independent and always included when present.
-    Each context source is capped at 20,000 chars.
-
-    When *skip_soul* is True, SOUL.md is not included here (it was already
-    loaded via ``load_soul_md()`` for the identity slot).
-    """
-    if cwd is None:
-        cwd = os.getcwd()
-
-    cwd_path = Path(cwd).resolve()
-    sections = []
-
-    # Priority-based project context: first match wins
-    project_context = (
-        _load_hermes_md(cwd_path)
-        or _load_agents_md(cwd_path)
-        or _load_claude_md(cwd_path)
-        or _load_cursorrules(cwd_path)
-    )
-    if project_context:
-        sections.append(project_context)
-
-    # SOUL.md from HERMES_HOME only — skip when already loaded as identity
-    if not skip_soul:
-        soul_content = load_soul_md()
-        if soul_content:
-            sections.append(soul_content)
-
-    if not sections:
-        return ""
-    return "# Project Context\n\nThe following project context files have been loaded and should be followed:\n\n" + "\n".join(sections)

agent/prompt_caching.py

@@ -1,72 +0,0 @@
-"""Anthropic prompt caching (system_and_3 strategy).
-
-Reduces input token costs by ~75% on multi-turn conversations by caching
-the conversation prefix. Uses 4 cache_control breakpoints (Anthropic max):
-  1. System prompt (stable across all turns)
-  2-4. Last 3 non-system messages (rolling window)
-
-Pure functions -- no class state, no AIAgent dependency.
-"""
-
-import copy
-from typing import Any, Dict, List
-
-
-def _apply_cache_marker(msg: dict, cache_marker: dict, native_anthropic: bool = False) -> None:
-    """Add cache_control to a single message, handling all format variations."""
-    role = msg.get("role", "")
-    content = msg.get("content")
-
-    if role == "tool":
-        if native_anthropic:
-            msg["cache_control"] = cache_marker
-        return
-
-    if content is None or content == "":
-        msg["cache_control"] = cache_marker
-        return
-
-    if isinstance(content, str):
-        msg["content"] = [
-            {"type": "text", "text": content, "cache_control": cache_marker}
-        ]
-        return
-
-    if isinstance(content, list) and content:
-        last = content[-1]
-        if isinstance(last, dict):
-            last["cache_control"] = cache_marker
-
-
-def apply_anthropic_cache_control(
-    api_messages: List[Dict[str, Any]],
-    cache_ttl: str = "5m",
-    native_anthropic: bool = False,
-) -> List[Dict[str, Any]]:
-    """Apply system_and_3 caching strategy to messages for Anthropic models.
-
-    Places up to 4 cache_control breakpoints: system prompt + last 3 non-system messages.
-
-    Returns:
-        Deep copy of messages with cache_control breakpoints injected.
-    """
-    messages = copy.deepcopy(api_messages)
-    if not messages:
-        return messages
-
-    marker = {"type": "ephemeral"}
-    if cache_ttl == "1h":
-        marker["ttl"] = "1h"
-
-    breakpoints_used = 0
-
-    if messages[0].get("role") == "system":
-        _apply_cache_marker(messages[0], marker, native_anthropic=native_anthropic)
-        breakpoints_used += 1
-
-    remaining = 4 - breakpoints_used
-    non_sys = [i for i in range(len(messages)) if messages[i].get("role") != "system"]
-    for idx in non_sys[-remaining:]:
-        _apply_cache_marker(messages[idx], marker, native_anthropic=native_anthropic)
-
-    return messages

agent/redact.py

@@ -1,176 +0,0 @@
-"""Regex-based secret redaction for logs and tool output.
-
-Applies pattern matching to mask API keys, tokens, and credentials
-before they reach log files, verbose output, or gateway logs.
-
-Short tokens (< 18 chars) are fully masked. Longer tokens preserve
-the first 6 and last 4 characters for debuggability.
-"""
-
-import logging
-import os
-import re
-
-logger = logging.getLogger(__name__)
-
-# Snapshot at import time so runtime env mutations (e.g. LLM-generated
-# `export HERMES_REDACT_SECRETS=false`) cannot disable redaction mid-session.
-_REDACT_ENABLED = os.getenv("HERMES_REDACT_SECRETS", "").lower() not in ("0", "false", "no", "off")
-
-# Known API key prefixes -- match the prefix + contiguous token chars
-_PREFIX_PATTERNS = [
-    r"sk-[A-Za-z0-9_-]{10,}",           # OpenAI / OpenRouter / Anthropic (sk-ant-*)
-    r"ghp_[A-Za-z0-9]{10,}",            # GitHub PAT (classic)
-    r"github_pat_[A-Za-z0-9_]{10,}",    # GitHub PAT (fine-grained)
-    r"gho_[A-Za-z0-9]{10,}",            # GitHub OAuth access token
-    r"ghu_[A-Za-z0-9]{10,}",            # GitHub user-to-server token
-    r"ghs_[A-Za-z0-9]{10,}",            # GitHub server-to-server token
-    r"ghr_[A-Za-z0-9]{10,}",            # GitHub refresh token
-    r"xox[baprs]-[A-Za-z0-9-]{10,}",    # Slack tokens
-    r"AIza[A-Za-z0-9_-]{30,}",          # Google API keys
-    r"pplx-[A-Za-z0-9]{10,}",           # Perplexity
-    r"fal_[A-Za-z0-9_-]{10,}",          # Fal.ai
-    r"fc-[A-Za-z0-9]{10,}",             # Firecrawl
-    r"bb_live_[A-Za-z0-9_-]{10,}",      # BrowserBase
-    r"gAAAA[A-Za-z0-9_=-]{20,}",        # Codex encrypted tokens
-    r"AKIA[A-Z0-9]{16}",                # AWS Access Key ID
-    r"sk_live_[A-Za-z0-9]{10,}",        # Stripe secret key (live)
-    r"sk_test_[A-Za-z0-9]{10,}",        # Stripe secret key (test)
-    r"rk_live_[A-Za-z0-9]{10,}",        # Stripe restricted key
-    r"SG\.[A-Za-z0-9_-]{10,}",          # SendGrid API key
-    r"hf_[A-Za-z0-9]{10,}",             # HuggingFace token
-    r"r8_[A-Za-z0-9]{10,}",             # Replicate API token
-    r"npm_[A-Za-z0-9]{10,}",            # npm access token
-    r"pypi-[A-Za-z0-9_-]{10,}",         # PyPI API token
-    r"dop_v1_[A-Za-z0-9]{10,}",         # DigitalOcean PAT
-    r"doo_v1_[A-Za-z0-9]{10,}",         # DigitalOcean OAuth
-    r"am_[A-Za-z0-9_-]{10,}",           # AgentMail API key
-    r"sk_[A-Za-z0-9_]{10,}",            # ElevenLabs TTS key (sk_ underscore, not sk- dash)
-    r"tvly-[A-Za-z0-9]{10,}",           # Tavily search API key
-    r"exa_[A-Za-z0-9]{10,}",            # Exa search API key
-]
-
-# ENV assignment patterns: KEY=value where KEY contains a secret-like name
-_SECRET_ENV_NAMES = r"(?:API_?KEY|TOKEN|SECRET|PASSWORD|PASSWD|CREDENTIAL|AUTH)"
-_ENV_ASSIGN_RE = re.compile(
-    rf"([A-Z_]*{_SECRET_ENV_NAMES}[A-Z_]*)\s*=\s*(['\"]?)(\S+)\2",
-    re.IGNORECASE,
-)
-
-# JSON field patterns: "apiKey": "value", "token": "value", etc.
-_JSON_KEY_NAMES = r"(?:api_?[Kk]ey|token|secret|password|access_token|refresh_token|auth_token|bearer|secret_value|raw_secret|secret_input|key_material)"
-_JSON_FIELD_RE = re.compile(
-    rf'("{_JSON_KEY_NAMES}")\s*:\s*"([^"]+)"',
-    re.IGNORECASE,
-)
-
-# Authorization headers
-_AUTH_HEADER_RE = re.compile(
-    r"(Authorization:\s*Bearer\s+)(\S+)",
-    re.IGNORECASE,
-)
-
-# Telegram bot tokens: bot<digits>:<token> or <digits>:<token>,
-# where token part is restricted to [-A-Za-z0-9_] and length >= 30
-_TELEGRAM_RE = re.compile(
-    r"(bot)?(\d{8,}):([-A-Za-z0-9_]{30,})",
-)
-
-# Private key blocks: -----BEGIN RSA PRIVATE KEY----- ... -----END RSA PRIVATE KEY-----
-_PRIVATE_KEY_RE = re.compile(
-    r"-----BEGIN[A-Z ]*PRIVATE KEY-----[\s\S]*?-----END[A-Z ]*PRIVATE KEY-----"
-)
-
-# Database connection strings: protocol://user:PASSWORD@host
-# Catches postgres, mysql, mongodb, redis, amqp URLs and redacts the password
-_DB_CONNSTR_RE = re.compile(
-    r"((?:postgres(?:ql)?|mysql|mongodb(?:\+srv)?|redis|amqp)://[^:]+:)([^@]+)(@)",
-    re.IGNORECASE,
-)
-
-# E.164 phone numbers: +<country><number>, 7-15 digits
-# Negative lookahead prevents matching hex strings or identifiers
-_SIGNAL_PHONE_RE = re.compile(r"(\+[1-9]\d{6,14})(?![A-Za-z0-9])")
-
-# Compile known prefix patterns into one alternation
-_PREFIX_RE = re.compile(
-    r"(?<![A-Za-z0-9_-])(" + "|".join(_PREFIX_PATTERNS) + r")(?![A-Za-z0-9_-])"
-)
-
-
-def _mask_token(token: str) -> str:
-    """Mask a token, preserving prefix for long tokens."""
-    if len(token) < 18:
-        return "***"
-    return f"{token[:6]}...{token[-4:]}"
-
-
-def redact_sensitive_text(text: str) -> str:
-    """Apply all redaction patterns to a block of text.
-
-    Safe to call on any string -- non-matching text passes through unchanged.
-    Disabled when security.redact_secrets is false in config.yaml.
-    """
-    if text is None:
-        return None
-    if not isinstance(text, str):
-        text = str(text)
-    if not text:
-        return text
-    if not _REDACT_ENABLED:
-        return text
-
-    # Known prefixes (sk-, ghp_, etc.)
-    text = _PREFIX_RE.sub(lambda m: _mask_token(m.group(1)), text)
-
-    # ENV assignments: OPENAI_API_KEY=sk-abc...
-    def _redact_env(m):
-        name, quote, value = m.group(1), m.group(2), m.group(3)
-        return f"{name}={quote}{_mask_token(value)}{quote}"
-    text = _ENV_ASSIGN_RE.sub(_redact_env, text)
-
-    # JSON fields: "apiKey": "value"
-    def _redact_json(m):
-        key, value = m.group(1), m.group(2)
-        return f'{key}: "{_mask_token(value)}"'
-    text = _JSON_FIELD_RE.sub(_redact_json, text)
-
-    # Authorization headers
-    text = _AUTH_HEADER_RE.sub(
-        lambda m: m.group(1) + _mask_token(m.group(2)),
-        text,
-    )
-
-    # Telegram bot tokens
-    def _redact_telegram(m):
-        prefix = m.group(1) or ""
-        digits = m.group(2)
-        return f"{prefix}{digits}:***"
-    text = _TELEGRAM_RE.sub(_redact_telegram, text)
-
-    # Private key blocks
-    text = _PRIVATE_KEY_RE.sub("[REDACTED PRIVATE KEY]", text)
-
-    # Database connection string passwords
-    text = _DB_CONNSTR_RE.sub(lambda m: f"{m.group(1)}***{m.group(3)}", text)
-
-    # E.164 phone numbers (Signal, WhatsApp)
-    def _redact_phone(m):
-        phone = m.group(1)
-        if len(phone) <= 8:
-            return phone[:2] + "****" + phone[-2:]
-        return phone[:4] + "****" + phone[-4:]
-    text = _SIGNAL_PHONE_RE.sub(_redact_phone, text)
-
-    return text
-
-
-class RedactingFormatter(logging.Formatter):
-    """Log formatter that redacts secrets from all log messages."""
-
-    def __init__(self, fmt=None, datefmt=None, style='%', **kwargs):
-        super().__init__(fmt, datefmt, style, **kwargs)
-
-    def format(self, record: logging.LogRecord) -> str:
-        original = super().format(record)
-        return redact_sensitive_text(original)

agent/skill_commands.py

@@ -1,297 +0,0 @@
-"""Shared slash command helpers for skills and built-in prompt-style modes.
-
-Shared between CLI (cli.py) and gateway (gateway/run.py) so both surfaces
-can invoke skills via /skill-name commands and prompt-only built-ins like
-/plan.
-"""
-
-import json
-import logging
-import re
-from datetime import datetime
-from pathlib import Path
-from typing import Any, Dict, Optional
-
-logger = logging.getLogger(__name__)
-
-_skill_commands: Dict[str, Dict[str, Any]] = {}
-_PLAN_SLUG_RE = re.compile(r"[^a-z0-9]+")
-
-
-def build_plan_path(
-    user_instruction: str = "",
-    *,
-    now: datetime | None = None,
-) -> Path:
-    """Return the default workspace-relative markdown path for a /plan invocation.
-
-    Relative paths are intentional: file tools are task/backend-aware and resolve
-    them against the active working directory for local, docker, ssh, modal,
-    daytona, and similar terminal backends. That keeps the plan with the active
-    workspace instead of the Hermes host's global home directory.
-    """
-    slug_source = (user_instruction or "").strip().splitlines()[0] if user_instruction else ""
-    slug = _PLAN_SLUG_RE.sub("-", slug_source.lower()).strip("-")
-    if slug:
-        slug = "-".join(part for part in slug.split("-")[:8] if part)[:48].strip("-")
-    slug = slug or "conversation-plan"
-    timestamp = (now or datetime.now()).strftime("%Y-%m-%d_%H%M%S")
-    return Path(".hermes") / "plans" / f"{timestamp}-{slug}.md"
-
-
-def _load_skill_payload(skill_identifier: str, task_id: str | None = None) -> tuple[dict[str, Any], Path | None, str] | None:
-    """Load a skill by name/path and return (loaded_payload, skill_dir, display_name)."""
-    raw_identifier = (skill_identifier or "").strip()
-    if not raw_identifier:
-        return None
-
-    try:
-        from tools.skills_tool import SKILLS_DIR, skill_view
-
-        identifier_path = Path(raw_identifier).expanduser()
-        if identifier_path.is_absolute():
-            try:
-                normalized = str(identifier_path.resolve().relative_to(SKILLS_DIR.resolve()))
-            except Exception:
-                normalized = raw_identifier
-        else:
-            normalized = raw_identifier.lstrip("/")
-
-        loaded_skill = json.loads(skill_view(normalized, task_id=task_id))
-    except Exception:
-        return None
-
-    if not loaded_skill.get("success"):
-        return None
-
-    skill_name = str(loaded_skill.get("name") or normalized)
-    skill_path = str(loaded_skill.get("path") or "")
-    skill_dir = None
-    if skill_path:
-        try:
-            skill_dir = SKILLS_DIR / Path(skill_path).parent
-        except Exception:
-            skill_dir = None
-
-    return loaded_skill, skill_dir, skill_name
-
-
-def _build_skill_message(
-    loaded_skill: dict[str, Any],
-    skill_dir: Path | None,
-    activation_note: str,
-    user_instruction: str = "",
-    runtime_note: str = "",
-) -> str:
-    """Format a loaded skill into a user/system message payload."""
-    from tools.skills_tool import SKILLS_DIR
-
-    content = str(loaded_skill.get("content") or "")
-
-    parts = [activation_note, "", content.strip()]
-
-    if loaded_skill.get("setup_skipped"):
-        parts.extend(
-            [
-                "",
-                "[Skill setup note: Required environment setup was skipped. Continue loading the skill and explain any reduced functionality if it matters.]",
-            ]
-        )
-    elif loaded_skill.get("gateway_setup_hint"):
-        parts.extend(
-            [
-                "",
-                f"[Skill setup note: {loaded_skill['gateway_setup_hint']}]",
-            ]
-        )
-    elif loaded_skill.get("setup_needed") and loaded_skill.get("setup_note"):
-        parts.extend(
-            [
-                "",
-                f"[Skill setup note: {loaded_skill['setup_note']}]",
-            ]
-        )
-
-    supporting = []
-    linked_files = loaded_skill.get("linked_files") or {}
-    for entries in linked_files.values():
-        if isinstance(entries, list):
-            supporting.extend(entries)
-
-    if not supporting and skill_dir:
-        for subdir in ("references", "templates", "scripts", "assets"):
-            subdir_path = skill_dir / subdir
-            if subdir_path.exists():
-                for f in sorted(subdir_path.rglob("*")):
-                    if f.is_file():
-                        rel = str(f.relative_to(skill_dir))
-                        supporting.append(rel)
-
-    if supporting and skill_dir:
-        try:
-            skill_view_target = str(skill_dir.relative_to(SKILLS_DIR))
-        except ValueError:
-            # Skill is from an external dir — use the skill name instead
-            skill_view_target = skill_dir.name
-        parts.append("")
-        parts.append("[This skill has supporting files you can load with the skill_view tool:]")
-        for sf in supporting:
-            parts.append(f"- {sf}")
-        parts.append(
-            f'\nTo view any of these, use: skill_view(name="{skill_view_target}", file_path="<path>")'
-        )
-
-    if user_instruction:
-        parts.append("")
-        parts.append(f"The user has provided the following instruction alongside the skill invocation: {user_instruction}")
-
-    if runtime_note:
-        parts.append("")
-        parts.append(f"[Runtime note: {runtime_note}]")
-
-    return "\n".join(parts)
-
-
-def scan_skill_commands() -> Dict[str, Dict[str, Any]]:
-    """Scan ~/.hermes/skills/ and return a mapping of /command -> skill info.
-
-    Returns:
-        Dict mapping "/skill-name" to {name, description, skill_md_path, skill_dir}.
-    """
-    global _skill_commands
-    _skill_commands = {}
-    try:
-        from tools.skills_tool import SKILLS_DIR, _parse_frontmatter, skill_matches_platform, _get_disabled_skill_names
-        from agent.skill_utils import get_external_skills_dirs
-        disabled = _get_disabled_skill_names()
-        seen_names: set = set()
-
-        # Scan local dir first, then external dirs
-        dirs_to_scan = []
-        if SKILLS_DIR.exists():
-            dirs_to_scan.append(SKILLS_DIR)
-        dirs_to_scan.extend(get_external_skills_dirs())
-
-        for scan_dir in dirs_to_scan:
-            for skill_md in scan_dir.rglob("SKILL.md"):
-                if any(part in ('.git', '.github', '.hub') for part in skill_md.parts):
-                    continue
-                try:
-                    content = skill_md.read_text(encoding='utf-8')
-                    frontmatter, body = _parse_frontmatter(content)
-                    # Skip skills incompatible with the current OS platform
-                    if not skill_matches_platform(frontmatter):
-                        continue
-                    name = frontmatter.get('name', skill_md.parent.name)
-                    if name in seen_names:
-                        continue
-                    # Respect user's disabled skills config
-                    if name in disabled:
-                        continue
-                    description = frontmatter.get('description', '')
-                    if not description:
-                        for line in body.strip().split('\n'):
-                            line = line.strip()
-                            if line and not line.startswith('#'):
-                                description = line[:80]
-                                break
-                    seen_names.add(name)
-                    cmd_name = name.lower().replace(' ', '-').replace('_', '-')
-                    _skill_commands[f"/{cmd_name}"] = {
-                        "name": name,
-                        "description": description or f"Invoke the {name} skill",
-                        "skill_md_path": str(skill_md),
-                        "skill_dir": str(skill_md.parent),
-                    }
-                except Exception:
-                    continue
-    except Exception:
-        pass
-    return _skill_commands
-
-
-def get_skill_commands() -> Dict[str, Dict[str, Any]]:
-    """Return the current skill commands mapping (scan first if empty)."""
-    if not _skill_commands:
-        scan_skill_commands()
-    return _skill_commands
-
-
-def build_skill_invocation_message(
-    cmd_key: str,
-    user_instruction: str = "",
-    task_id: str | None = None,
-    runtime_note: str = "",
-) -> Optional[str]:
-    """Build the user message content for a skill slash command invocation.
-
-    Args:
-        cmd_key: The command key including leading slash (e.g., "/gif-search").
-        user_instruction: Optional text the user typed after the command.
-
-    Returns:
-        The formatted message string, or None if the skill wasn't found.
-    """
-    commands = get_skill_commands()
-    skill_info = commands.get(cmd_key)
-    if not skill_info:
-        return None
-
-    loaded = _load_skill_payload(skill_info["skill_dir"], task_id=task_id)
-    if not loaded:
-        return f"[Failed to load skill: {skill_info['name']}]"
-
-    loaded_skill, skill_dir, skill_name = loaded
-    activation_note = (
-        f'[SYSTEM: The user has invoked the "{skill_name}" skill, indicating they want '
-        "you to follow its instructions. The full skill content is loaded below.]"
-    )
-    return _build_skill_message(
-        loaded_skill,
-        skill_dir,
-        activation_note,
-        user_instruction=user_instruction,
-        runtime_note=runtime_note,
-    )
-
-
-def build_preloaded_skills_prompt(
-    skill_identifiers: list[str],
-    task_id: str | None = None,
-) -> tuple[str, list[str], list[str]]:
-    """Load one or more skills for session-wide CLI preloading.
-
-    Returns (prompt_text, loaded_skill_names, missing_identifiers).
-    """
-    prompt_parts: list[str] = []
-    loaded_names: list[str] = []
-    missing: list[str] = []
-
-    seen: set[str] = set()
-    for raw_identifier in skill_identifiers:
-        identifier = (raw_identifier or "").strip()
-        if not identifier or identifier in seen:
-            continue
-        seen.add(identifier)
-
-        loaded = _load_skill_payload(identifier, task_id=task_id)
-        if not loaded:
-            missing.append(identifier)
-            continue
-
-        loaded_skill, skill_dir, skill_name = loaded
-        activation_note = (
-            f'[SYSTEM: The user launched this CLI session with the "{skill_name}" skill '
-            "preloaded. Treat its instructions as active guidance for the duration of this "
-            "session unless the user overrides them.]"
-        )
-        prompt_parts.append(
-            _build_skill_message(
-                loaded_skill,
-                skill_dir,
-                activation_note,
-            )
-        )
-        loaded_names.append(skill_name)
-
-    return "\n\n".join(prompt_parts), loaded_names, missing

agent/skill_utils.py

@@ -1,276 +0,0 @@
-"""Lightweight skill metadata utilities shared by prompt_builder and skills_tool.
-
-This module intentionally avoids importing the tool registry, CLI config, or any
-heavy dependency chain.  It is safe to import at module level without triggering
-tool registration or provider resolution.
-"""
-
-import logging
-import os
-import re
-import sys
-from pathlib import Path
-from typing import Any, Dict, List, Optional, Set, Tuple
-
-from hermes_constants import get_hermes_home
-
-logger = logging.getLogger(__name__)
-
-# ── Platform mapping ──────────────────────────────────────────────────────
-
-PLATFORM_MAP = {
-    "macos": "darwin",
-    "linux": "linux",
-    "windows": "win32",
-}
-
-EXCLUDED_SKILL_DIRS = frozenset((".git", ".github", ".hub"))
-
-# ── Lazy YAML loader ─────────────────────────────────────────────────────
-
-_yaml_load_fn = None
-
-
-def yaml_load(content: str):
-    """Parse YAML with lazy import and CSafeLoader preference."""
-    global _yaml_load_fn
-    if _yaml_load_fn is None:
-        import yaml
-
-        loader = getattr(yaml, "CSafeLoader", None) or yaml.SafeLoader
-
-        def _load(value: str):
-            return yaml.load(value, Loader=loader)
-
-        _yaml_load_fn = _load
-    return _yaml_load_fn(content)
-
-
-# ── Frontmatter parsing ──────────────────────────────────────────────────
-
-
-def parse_frontmatter(content: str) -> Tuple[Dict[str, Any], str]:
-    """Parse YAML frontmatter from a markdown string.
-
-    Uses yaml with CSafeLoader for full YAML support (nested metadata, lists)
-    with a fallback to simple key:value splitting for robustness.
-
-    Returns:
-        (frontmatter_dict, remaining_body)
-    """
-    frontmatter: Dict[str, Any] = {}
-    body = content
-
-    if not content.startswith("---"):
-        return frontmatter, body
-
-    end_match = re.search(r"\n---\s*\n", content[3:])
-    if not end_match:
-        return frontmatter, body
-
-    yaml_content = content[3 : end_match.start() + 3]
-    body = content[end_match.end() + 3 :]
-
-    try:
-        parsed = yaml_load(yaml_content)
-        if isinstance(parsed, dict):
-            frontmatter = parsed
-    except Exception:
-        # Fallback: simple key:value parsing for malformed YAML
-        for line in yaml_content.strip().split("\n"):
-            if ":" not in line:
-                continue
-            key, value = line.split(":", 1)
-            frontmatter[key.strip()] = value.strip()
-
-    return frontmatter, body
-
-
-# ── Platform matching ─────────────────────────────────────────────────────
-
-
-def skill_matches_platform(frontmatter: Dict[str, Any]) -> bool:
-    """Return True when the skill is compatible with the current OS.
-
-    Skills declare platform requirements via a top-level ``platforms`` list
-    in their YAML frontmatter::
-
-        platforms: [macos]          # macOS only
-        platforms: [macos, linux]   # macOS and Linux
-
-    If the field is absent or empty the skill is compatible with **all**
-    platforms (backward-compatible default).
-    """
-    platforms = frontmatter.get("platforms")
-    if not platforms:
-        return True
-    if not isinstance(platforms, list):
-        platforms = [platforms]
-    current = sys.platform
-    for platform in platforms:
-        normalized = str(platform).lower().strip()
-        mapped = PLATFORM_MAP.get(normalized, normalized)
-        if current.startswith(mapped):
-            return True
-    return False
-
-
-# ── Disabled skills ───────────────────────────────────────────────────────
-
-
-def get_disabled_skill_names() -> Set[str]:
-    """Read disabled skill names from config.yaml.
-
-    Resolves platform from ``HERMES_PLATFORM`` env var, falls back to
-    the global disabled list.  Reads the config file directly (no CLI
-    config imports) to stay lightweight.
-    """
-    config_path = get_hermes_home() / "config.yaml"
-    if not config_path.exists():
-        return set()
-    try:
-        parsed = yaml_load(config_path.read_text(encoding="utf-8"))
-    except Exception as e:
-        logger.debug("Could not read skill config %s: %s", config_path, e)
-        return set()
-    if not isinstance(parsed, dict):
-        return set()
-
-    skills_cfg = parsed.get("skills")
-    if not isinstance(skills_cfg, dict):
-        return set()
-
-    resolved_platform = os.getenv("HERMES_PLATFORM")
-    if resolved_platform:
-        platform_disabled = (skills_cfg.get("platform_disabled") or {}).get(
-            resolved_platform
-        )
-        if platform_disabled is not None:
-            return _normalize_string_set(platform_disabled)
-    return _normalize_string_set(skills_cfg.get("disabled"))
-
-
-def _normalize_string_set(values) -> Set[str]:
-    if values is None:
-        return set()
-    if isinstance(values, str):
-        values = [values]
-    return {str(v).strip() for v in values if str(v).strip()}
-
-
-# ── External skills directories ──────────────────────────────────────────
-
-
-def get_external_skills_dirs() -> List[Path]:
-    """Read ``skills.external_dirs`` from config.yaml and return validated paths.
-
-    Each entry is expanded (``~`` and ``${VAR}``) and resolved to an absolute
-    path.  Only directories that actually exist are returned.  Duplicates and
-    paths that resolve to the local ``~/.hermes/skills/`` are silently skipped.
-    """
-    config_path = get_hermes_home() / "config.yaml"
-    if not config_path.exists():
-        return []
-    try:
-        parsed = yaml_load(config_path.read_text(encoding="utf-8"))
-    except Exception:
-        return []
-    if not isinstance(parsed, dict):
-        return []
-
-    skills_cfg = parsed.get("skills")
-    if not isinstance(skills_cfg, dict):
-        return []
-
-    raw_dirs = skills_cfg.get("external_dirs")
-    if not raw_dirs:
-        return []
-    if isinstance(raw_dirs, str):
-        raw_dirs = [raw_dirs]
-    if not isinstance(raw_dirs, list):
-        return []
-
-    local_skills = (get_hermes_home() / "skills").resolve()
-    seen: Set[Path] = set()
-    result: List[Path] = []
-
-    for entry in raw_dirs:
-        entry = str(entry).strip()
-        if not entry:
-            continue
-        # Expand ~ and environment variables
-        expanded = os.path.expanduser(os.path.expandvars(entry))
-        p = Path(expanded).resolve()
-        if p == local_skills:
-            continue
-        if p in seen:
-            continue
-        if p.is_dir():
-            seen.add(p)
-            result.append(p)
-        else:
-            logger.debug("External skills dir does not exist, skipping: %s", p)
-
-    return result
-
-
-def get_all_skills_dirs() -> List[Path]:
-    """Return all skill directories: local ``~/.hermes/skills/`` first, then external.
-
-    The local dir is always first (and always included even if it doesn't exist
-    yet — callers handle that).  External dirs follow in config order.
-    """
-    dirs = [get_hermes_home() / "skills"]
-    dirs.extend(get_external_skills_dirs())
-    return dirs
-
-
-# ── Condition extraction ──────────────────────────────────────────────────
-
-
-def extract_skill_conditions(frontmatter: Dict[str, Any]) -> Dict[str, List]:
-    """Extract conditional activation fields from parsed frontmatter."""
-    metadata = frontmatter.get("metadata")
-    # Handle cases where metadata is not a dict (e.g., a string from malformed YAML)
-    if not isinstance(metadata, dict):
-        metadata = {}
-    hermes = metadata.get("hermes") or {}
-    if not isinstance(hermes, dict):
-        hermes = {}
-    return {
-        "fallback_for_toolsets": hermes.get("fallback_for_toolsets", []),
-        "requires_toolsets": hermes.get("requires_toolsets", []),
-        "fallback_for_tools": hermes.get("fallback_for_tools", []),
-        "requires_tools": hermes.get("requires_tools", []),
-    }
-
-
-# ── Description extraction ────────────────────────────────────────────────
-
-
-def extract_skill_description(frontmatter: Dict[str, Any]) -> str:
-    """Extract a truncated description from parsed frontmatter."""
-    raw_desc = frontmatter.get("description", "")
-    if not raw_desc:
-        return ""
-    desc = str(raw_desc).strip().strip("'\"")
-    if len(desc) > 60:
-        return desc[:57] + "..."
-    return desc
-
-
-# ── File iteration ────────────────────────────────────────────────────────
-
-
-def iter_skill_index_files(skills_dir: Path, filename: str):
-    """Walk skills_dir yielding sorted paths matching *filename*.
-
-    Excludes ``.git``, ``.github``, ``.hub`` directories.
-    """
-    matches = []
-    for root, dirs, files in os.walk(skills_dir):
-        dirs[:] = [d for d in dirs if d not in EXCLUDED_SKILL_DIRS]
-        if filename in files:
-            matches.append(Path(root) / filename)
-    for path in sorted(matches, key=lambda p: str(p.relative_to(skills_dir))):
-        yield path

agent/smart_model_routing.py

@@ -1,194 +0,0 @@
-"""Helpers for optional cheap-vs-strong model routing."""
-
-from __future__ import annotations
-
-import os
-import re
-from typing import Any, Dict, Optional
-
-from utils import is_truthy_value
-
-_COMPLEX_KEYWORDS = {
-    "debug",
-    "debugging",
-    "implement",
-    "implementation",
-    "refactor",
-    "patch",
-    "traceback",
-    "stacktrace",
-    "exception",
-    "error",
-    "analyze",
-    "analysis",
-    "investigate",
-    "architecture",
-    "design",
-    "compare",
-    "benchmark",
-    "optimize",
-    "optimise",
-    "review",
-    "terminal",
-    "shell",
-    "tool",
-    "tools",
-    "pytest",
-    "test",
-    "tests",
-    "plan",
-    "planning",
-    "delegate",
-    "subagent",
-    "cron",
-    "docker",
-    "kubernetes",
-}
-
-_URL_RE = re.compile(r"https?://|www\.", re.IGNORECASE)
-
-
-def _coerce_bool(value: Any, default: bool = False) -> bool:
-    return is_truthy_value(value, default=default)
-
-
-def _coerce_int(value: Any, default: int) -> int:
-    try:
-        return int(value)
-    except (TypeError, ValueError):
-        return default
-
-
-def choose_cheap_model_route(user_message: str, routing_config: Optional[Dict[str, Any]]) -> Optional[Dict[str, Any]]:
-    """Return the configured cheap-model route when a message looks simple.
-
-    Conservative by design: if the message has signs of code/tool/debugging/
-    long-form work, keep the primary model.
-    """
-    cfg = routing_config or {}
-    if not _coerce_bool(cfg.get("enabled"), False):
-        return None
-
-    cheap_model = cfg.get("cheap_model") or {}
-    if not isinstance(cheap_model, dict):
-        return None
-    provider = str(cheap_model.get("provider") or "").strip().lower()
-    model = str(cheap_model.get("model") or "").strip()
-    if not provider or not model:
-        return None
-
-    text = (user_message or "").strip()
-    if not text:
-        return None
-
-    max_chars = _coerce_int(cfg.get("max_simple_chars"), 160)
-    max_words = _coerce_int(cfg.get("max_simple_words"), 28)
-
-    if len(text) > max_chars:
-        return None
-    if len(text.split()) > max_words:
-        return None
-    if text.count("\n") > 1:
-        return None
-    if "```" in text or "`" in text:
-        return None
-    if _URL_RE.search(text):
-        return None
-
-    lowered = text.lower()
-    words = {token.strip(".,:;!?()[]{}\"'`") for token in lowered.split()}
-    if words & _COMPLEX_KEYWORDS:
-        return None
-
-    route = dict(cheap_model)
-    route["provider"] = provider
-    route["model"] = model
-    route["routing_reason"] = "simple_turn"
-    return route
-
-
-def resolve_turn_route(user_message: str, routing_config: Optional[Dict[str, Any]], primary: Dict[str, Any]) -> Dict[str, Any]:
-    """Resolve the effective model/runtime for one turn.
-
-    Returns a dict with model/runtime/signature/label fields.
-    """
-    route = choose_cheap_model_route(user_message, routing_config)
-    if not route:
-        return {
-            "model": primary.get("model"),
-            "runtime": {
-                "api_key": primary.get("api_key"),
-                "base_url": primary.get("base_url"),
-                "provider": primary.get("provider"),
-                "api_mode": primary.get("api_mode"),
-                "command": primary.get("command"),
-                "args": list(primary.get("args") or []),
-                "credential_pool": primary.get("credential_pool"),
-            },
-            "label": None,
-            "signature": (
-                primary.get("model"),
-                primary.get("provider"),
-                primary.get("base_url"),
-                primary.get("api_mode"),
-                primary.get("command"),
-                tuple(primary.get("args") or ()),
-            ),
-        }
-
-    from hermes_cli.runtime_provider import resolve_runtime_provider
-
-    explicit_api_key = None
-    api_key_env = str(route.get("api_key_env") or "").strip()
-    if api_key_env:
-        explicit_api_key = os.getenv(api_key_env) or None
-
-    try:
-        runtime = resolve_runtime_provider(
-            requested=route.get("provider"),
-            explicit_api_key=explicit_api_key,
-            explicit_base_url=route.get("base_url"),
-        )
-    except Exception:
-        return {
-            "model": primary.get("model"),
-            "runtime": {
-                "api_key": primary.get("api_key"),
-                "base_url": primary.get("base_url"),
-                "provider": primary.get("provider"),
-                "api_mode": primary.get("api_mode"),
-                "command": primary.get("command"),
-                "args": list(primary.get("args") or []),
-                "credential_pool": primary.get("credential_pool"),
-            },
-            "label": None,
-            "signature": (
-                primary.get("model"),
-                primary.get("provider"),
-                primary.get("base_url"),
-                primary.get("api_mode"),
-                primary.get("command"),
-                tuple(primary.get("args") or ()),
-            ),
-        }
-
-    return {
-        "model": route.get("model"),
-        "runtime": {
-            "api_key": runtime.get("api_key"),
-            "base_url": runtime.get("base_url"),
-            "provider": runtime.get("provider"),
-            "api_mode": runtime.get("api_mode"),
-            "command": runtime.get("command"),
-            "args": list(runtime.get("args") or []),
-        },
-        "label": f"smart route → {route.get('model')} ({runtime.get('provider')})",
-        "signature": (
-            route.get("model"),
-            runtime.get("provider"),
-            runtime.get("base_url"),
-            runtime.get("api_mode"),
-            runtime.get("command"),
-            tuple(runtime.get("args") or ()),
-        ),
-    }

agent/title_generator.py

@@ -1,125 +0,0 @@
-"""Auto-generate short session titles from the first user/assistant exchange.
-
-Runs asynchronously after the first response is delivered so it never
-adds latency to the user-facing reply.
-"""
-
-import logging
-import threading
-from typing import Optional
-
-from agent.auxiliary_client import call_llm
-
-logger = logging.getLogger(__name__)
-
-_TITLE_PROMPT = (
-    "Generate a short, descriptive title (3-7 words) for a conversation that starts with the "
-    "following exchange. The title should capture the main topic or intent. "
-    "Return ONLY the title text, nothing else. No quotes, no punctuation at the end, no prefixes."
-)
-
-
-def generate_title(user_message: str, assistant_response: str, timeout: float = 30.0) -> Optional[str]:
-    """Generate a session title from the first exchange.
-
-    Uses the auxiliary LLM client (cheapest/fastest available model).
-    Returns the title string or None on failure.
-    """
-    # Truncate long messages to keep the request small
-    user_snippet = user_message[:500] if user_message else ""
-    assistant_snippet = assistant_response[:500] if assistant_response else ""
-
-    messages = [
-        {"role": "system", "content": _TITLE_PROMPT},
-        {"role": "user", "content": f"User: {user_snippet}\n\nAssistant: {assistant_snippet}"},
-    ]
-
-    try:
-        response = call_llm(
-            task="compression",  # reuse compression task config (cheap/fast model)
-            messages=messages,
-            max_tokens=30,
-            temperature=0.3,
-            timeout=timeout,
-        )
-        title = (response.choices[0].message.content or "").strip()
-        # Clean up: remove quotes, trailing punctuation, prefixes like "Title: "
-        title = title.strip('"\'')
-        if title.lower().startswith("title:"):
-            title = title[6:].strip()
-        # Enforce reasonable length
-        if len(title) > 80:
-            title = title[:77] + "..."
-        return title if title else None
-    except Exception as e:
-        logger.debug("Title generation failed: %s", e)
-        return None
-
-
-def auto_title_session(
-    session_db,
-    session_id: str,
-    user_message: str,
-    assistant_response: str,
-) -> None:
-    """Generate and set a session title if one doesn't already exist.
-
-    Called in a background thread after the first exchange completes.
-    Silently skips if:
-    - session_db is None
-    - session already has a title (user-set or previously auto-generated)
-    - title generation fails
-    """
-    if not session_db or not session_id:
-        return
-
-    # Check if title already exists (user may have set one via /title before first response)
-    try:
-        existing = session_db.get_session_title(session_id)
-        if existing:
-            return
-    except Exception:
-        return
-
-    title = generate_title(user_message, assistant_response)
-    if not title:
-        return
-
-    try:
-        session_db.set_session_title(session_id, title)
-        logger.debug("Auto-generated session title: %s", title)
-    except Exception as e:
-        logger.debug("Failed to set auto-generated title: %s", e)
-
-
-def maybe_auto_title(
-    session_db,
-    session_id: str,
-    user_message: str,
-    assistant_response: str,
-    conversation_history: list,
-) -> None:
-    """Fire-and-forget title generation after the first exchange.
-
-    Only generates a title when:
-    - This appears to be the first user→assistant exchange
-    - No title is already set
-    """
-    if not session_db or not session_id or not user_message or not assistant_response:
-        return
-
-    # Count user messages in history to detect first exchange.
-    # conversation_history includes the exchange that just happened,
-    # so for a first exchange we expect exactly 1 user message
-    # (or 2 counting system). Be generous: generate on first 2 exchanges.
-    user_msg_count = sum(1 for m in (conversation_history or []) if m.get("role") == "user")
-    if user_msg_count > 2:
-        return
-
-    thread = threading.Thread(
-        target=auto_title_session,
-        args=(session_db, session_id, user_message, assistant_response),
-        daemon=True,
-        name="auto-title",
-    )
-    thread.start()

agent/trajectory.py

@@ -1,56 +0,0 @@
-"""Trajectory saving utilities and static helpers.
-
-_convert_to_trajectory_format stays as an AIAgent method (batch_runner.py
-calls agent._convert_to_trajectory_format). Only the static helpers and
-the file-write logic live here.
-"""
-
-import json
-import logging
-from datetime import datetime
-from typing import Any, Dict, List
-
-logger = logging.getLogger(__name__)
-
-
-def convert_scratchpad_to_think(content: str) -> str:
-    """Convert <REASONING_SCRATCHPAD> tags to <think> tags."""
-    if not content or "<REASONING_SCRATCHPAD>" not in content:
-        return content
-    return content.replace("<REASONING_SCRATCHPAD>", "<think>").replace("</REASONING_SCRATCHPAD>", "</think>")
-
-
-def has_incomplete_scratchpad(content: str) -> bool:
-    """Check if content has an opening <REASONING_SCRATCHPAD> without a closing tag."""
-    if not content:
-        return False
-    return "<REASONING_SCRATCHPAD>" in content and "</REASONING_SCRATCHPAD>" not in content
-
-
-def save_trajectory(trajectory: List[Dict[str, Any]], model: str,
-                    completed: bool, filename: str = None):
-    """Append a trajectory entry to a JSONL file.
-
-    Args:
-        trajectory: The ShareGPT-format conversation list.
-        model: Model name for metadata.
-        completed: Whether the conversation completed successfully.
-        filename: Override output filename. Defaults to trajectory_samples.jsonl
-                  or failed_trajectories.jsonl based on ``completed``.
-    """
-    if filename is None:
-        filename = "trajectory_samples.jsonl" if completed else "failed_trajectories.jsonl"
-
-    entry = {
-        "conversations": trajectory,
-        "timestamp": datetime.now().isoformat(),
-        "model": model,
-        "completed": completed,
-    }
-
-    try:
-        with open(filename, "a", encoding="utf-8") as f:
-            f.write(json.dumps(entry, ensure_ascii=False) + "\n")
-        logger.info("Trajectory saved to %s", filename)
-    except Exception as e:
-        logger.warning("Failed to save trajectory: %s", e)

agent/usage_pricing.py

@@ -1,656 +0,0 @@
-from __future__ import annotations
-
-from dataclasses import dataclass
-from datetime import datetime, timezone
-from decimal import Decimal
-from typing import Any, Dict, Literal, Optional
-
-from agent.model_metadata import fetch_endpoint_model_metadata, fetch_model_metadata
-
-DEFAULT_PRICING = {"input": 0.0, "output": 0.0}
-
-_ZERO = Decimal("0")
-_ONE_MILLION = Decimal("1000000")
-
-CostStatus = Literal["actual", "estimated", "included", "unknown"]
-CostSource = Literal[
-    "provider_cost_api",
-    "provider_generation_api",
-    "provider_models_api",
-    "official_docs_snapshot",
-    "user_override",
-    "custom_contract",
-    "none",
-]
-
-
-@dataclass(frozen=True)
-class CanonicalUsage:
-    input_tokens: int = 0
-    output_tokens: int = 0
-    cache_read_tokens: int = 0
-    cache_write_tokens: int = 0
-    reasoning_tokens: int = 0
-    request_count: int = 1
-    raw_usage: Optional[dict[str, Any]] = None
-
-    @property
-    def prompt_tokens(self) -> int:
-        return self.input_tokens + self.cache_read_tokens + self.cache_write_tokens
-
-    @property
-    def total_tokens(self) -> int:
-        return self.prompt_tokens + self.output_tokens
-
-
-@dataclass(frozen=True)
-class BillingRoute:
-    provider: str
-    model: str
-    base_url: str = ""
-    billing_mode: str = "unknown"
-
-
-@dataclass(frozen=True)
-class PricingEntry:
-    input_cost_per_million: Optional[Decimal] = None
-    output_cost_per_million: Optional[Decimal] = None
-    cache_read_cost_per_million: Optional[Decimal] = None
-    cache_write_cost_per_million: Optional[Decimal] = None
-    request_cost: Optional[Decimal] = None
-    source: CostSource = "none"
-    source_url: Optional[str] = None
-    pricing_version: Optional[str] = None
-    fetched_at: Optional[datetime] = None
-
-
-@dataclass(frozen=True)
-class CostResult:
-    amount_usd: Optional[Decimal]
-    status: CostStatus
-    source: CostSource
-    label: str
-    fetched_at: Optional[datetime] = None
-    pricing_version: Optional[str] = None
-    notes: tuple[str, ...] = ()
-
-
-_UTC_NOW = lambda: datetime.now(timezone.utc)
-
-
-# Official docs snapshot entries. Models whose published pricing and cache
-# semantics are stable enough to encode exactly.
-_OFFICIAL_DOCS_PRICING: Dict[tuple[str, str], PricingEntry] = {
-    (
-        "anthropic",
-        "claude-opus-4-20250514",
-    ): PricingEntry(
-        input_cost_per_million=Decimal("15.00"),
-        output_cost_per_million=Decimal("75.00"),
-        cache_read_cost_per_million=Decimal("1.50"),
-        cache_write_cost_per_million=Decimal("18.75"),
-        source="official_docs_snapshot",
-        source_url="https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching",
-        pricing_version="anthropic-prompt-caching-2026-03-16",
-    ),
-    (
-        "anthropic",
-        "claude-sonnet-4-20250514",
-    ): PricingEntry(
-        input_cost_per_million=Decimal("3.00"),
-        output_cost_per_million=Decimal("15.00"),
-        cache_read_cost_per_million=Decimal("0.30"),
-        cache_write_cost_per_million=Decimal("3.75"),
-        source="official_docs_snapshot",
-        source_url="https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching",
-        pricing_version="anthropic-prompt-caching-2026-03-16",
-    ),
-    # OpenAI
-    (
-        "openai",
-        "gpt-4o",
-    ): PricingEntry(
-        input_cost_per_million=Decimal("2.50"),
-        output_cost_per_million=Decimal("10.00"),
-        cache_read_cost_per_million=Decimal("1.25"),
-        source="official_docs_snapshot",
-        source_url="https://openai.com/api/pricing/",
-        pricing_version="openai-pricing-2026-03-16",
-    ),
-    (
-        "openai",
-        "gpt-4o-mini",
-    ): PricingEntry(
-        input_cost_per_million=Decimal("0.15"),
-        output_cost_per_million=Decimal("0.60"),
-        cache_read_cost_per_million=Decimal("0.075"),
-        source="official_docs_snapshot",
-        source_url="https://openai.com/api/pricing/",
-        pricing_version="openai-pricing-2026-03-16",
-    ),
-    (
-        "openai",
-        "gpt-4.1",
-    ): PricingEntry(
-        input_cost_per_million=Decimal("2.00"),
-        output_cost_per_million=Decimal("8.00"),
-        cache_read_cost_per_million=Decimal("0.50"),
-        source="official_docs_snapshot",
-        source_url="https://openai.com/api/pricing/",
-        pricing_version="openai-pricing-2026-03-16",
-    ),
-    (
-        "openai",
-        "gpt-4.1-mini",
-    ): PricingEntry(
-        input_cost_per_million=Decimal("0.40"),
-        output_cost_per_million=Decimal("1.60"),
-        cache_read_cost_per_million=Decimal("0.10"),
-        source="official_docs_snapshot",
-        source_url="https://openai.com/api/pricing/",
-        pricing_version="openai-pricing-2026-03-16",
-    ),
-    (
-        "openai",
-        "gpt-4.1-nano",
-    ): PricingEntry(
-        input_cost_per_million=Decimal("0.10"),
-        output_cost_per_million=Decimal("0.40"),
-        cache_read_cost_per_million=Decimal("0.025"),
-        source="official_docs_snapshot",
-        source_url="https://openai.com/api/pricing/",
-        pricing_version="openai-pricing-2026-03-16",
-    ),
-    (
-        "openai",
-        "o3",
-    ): PricingEntry(
-        input_cost_per_million=Decimal("10.00"),
-        output_cost_per_million=Decimal("40.00"),
-        cache_read_cost_per_million=Decimal("2.50"),
-        source="official_docs_snapshot",
-        source_url="https://openai.com/api/pricing/",
-        pricing_version="openai-pricing-2026-03-16",
-    ),
-    (
-        "openai",
-        "o3-mini",
-    ): PricingEntry(
-        input_cost_per_million=Decimal("1.10"),
-        output_cost_per_million=Decimal("4.40"),
-        cache_read_cost_per_million=Decimal("0.55"),
-        source="official_docs_snapshot",
-        source_url="https://openai.com/api/pricing/",
-        pricing_version="openai-pricing-2026-03-16",
-    ),
-    # Anthropic older models (pre-4.6 generation)
-    (
-        "anthropic",
-        "claude-3-5-sonnet-20241022",
-    ): PricingEntry(
-        input_cost_per_million=Decimal("3.00"),
-        output_cost_per_million=Decimal("15.00"),
-        cache_read_cost_per_million=Decimal("0.30"),
-        cache_write_cost_per_million=Decimal("3.75"),
-        source="official_docs_snapshot",
-        source_url="https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching",
-        pricing_version="anthropic-pricing-2026-03-16",
-    ),
-    (
-        "anthropic",
-        "claude-3-5-haiku-20241022",
-    ): PricingEntry(
-        input_cost_per_million=Decimal("0.80"),
-        output_cost_per_million=Decimal("4.00"),
-        cache_read_cost_per_million=Decimal("0.08"),
-        cache_write_cost_per_million=Decimal("1.00"),
-        source="official_docs_snapshot",
-        source_url="https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching",
-        pricing_version="anthropic-pricing-2026-03-16",
-    ),
-    (
-        "anthropic",
-        "claude-3-opus-20240229",
-    ): PricingEntry(
-        input_cost_per_million=Decimal("15.00"),
-        output_cost_per_million=Decimal("75.00"),
-        cache_read_cost_per_million=Decimal("1.50"),
-        cache_write_cost_per_million=Decimal("18.75"),
-        source="official_docs_snapshot",
-        source_url="https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching",
-        pricing_version="anthropic-pricing-2026-03-16",
-    ),
-    (
-        "anthropic",
-        "claude-3-haiku-20240307",
-    ): PricingEntry(
-        input_cost_per_million=Decimal("0.25"),
-        output_cost_per_million=Decimal("1.25"),
-        cache_read_cost_per_million=Decimal("0.03"),
-        cache_write_cost_per_million=Decimal("0.30"),
-        source="official_docs_snapshot",
-        source_url="https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching",
-        pricing_version="anthropic-pricing-2026-03-16",
-    ),
-    # DeepSeek
-    (
-        "deepseek",
-        "deepseek-chat",
-    ): PricingEntry(
-        input_cost_per_million=Decimal("0.14"),
-        output_cost_per_million=Decimal("0.28"),
-        source="official_docs_snapshot",
-        source_url="https://api-docs.deepseek.com/quick_start/pricing",
-        pricing_version="deepseek-pricing-2026-03-16",
-    ),
-    (
-        "deepseek",
-        "deepseek-reasoner",
-    ): PricingEntry(
-        input_cost_per_million=Decimal("0.55"),
-        output_cost_per_million=Decimal("2.19"),
-        source="official_docs_snapshot",
-        source_url="https://api-docs.deepseek.com/quick_start/pricing",
-        pricing_version="deepseek-pricing-2026-03-16",
-    ),
-    # Google Gemini
-    (
-        "google",
-        "gemini-2.5-pro",
-    ): PricingEntry(
-        input_cost_per_million=Decimal("1.25"),
-        output_cost_per_million=Decimal("10.00"),
-        source="official_docs_snapshot",
-        source_url="https://ai.google.dev/pricing",
-        pricing_version="google-pricing-2026-03-16",
-    ),
-    (
-        "google",
-        "gemini-2.5-flash",
-    ): PricingEntry(
-        input_cost_per_million=Decimal("0.15"),
-        output_cost_per_million=Decimal("0.60"),
-        source="official_docs_snapshot",
-        source_url="https://ai.google.dev/pricing",
-        pricing_version="google-pricing-2026-03-16",
-    ),
-    (
-        "google",
-        "gemini-2.0-flash",
-    ): PricingEntry(
-        input_cost_per_million=Decimal("0.10"),
-        output_cost_per_million=Decimal("0.40"),
-        source="official_docs_snapshot",
-        source_url="https://ai.google.dev/pricing",
-        pricing_version="google-pricing-2026-03-16",
-    ),
-}
-
-
-def _to_decimal(value: Any) -> Optional[Decimal]:
-    if value is None:
-        return None
-    try:
-        return Decimal(str(value))
-    except Exception:
-        return None
-
-
-def _to_int(value: Any) -> int:
-    try:
-        return int(value or 0)
-    except Exception:
-        return 0
-
-
-def resolve_billing_route(
-    model_name: str,
-    provider: Optional[str] = None,
-    base_url: Optional[str] = None,
-) -> BillingRoute:
-    provider_name = (provider or "").strip().lower()
-    base = (base_url or "").strip().lower()
-    model = (model_name or "").strip()
-    if not provider_name and "/" in model:
-        inferred_provider, bare_model = model.split("/", 1)
-        if inferred_provider in {"anthropic", "openai", "google"}:
-            provider_name = inferred_provider
-            model = bare_model
-
-    if provider_name == "openai-codex":
-        return BillingRoute(provider="openai-codex", model=model, base_url=base_url or "", billing_mode="subscription_included")
-    if provider_name == "openrouter" or "openrouter.ai" in base:
-        return BillingRoute(provider="openrouter", model=model, base_url=base_url or "", billing_mode="official_models_api")
-    if provider_name == "anthropic":
-        return BillingRoute(provider="anthropic", model=model.split("/")[-1], base_url=base_url or "", billing_mode="official_docs_snapshot")
-    if provider_name == "openai":
-        return BillingRoute(provider="openai", model=model.split("/")[-1], base_url=base_url or "", billing_mode="official_docs_snapshot")
-    if provider_name in {"custom", "local"} or (base and "localhost" in base):
-        return BillingRoute(provider=provider_name or "custom", model=model, base_url=base_url or "", billing_mode="unknown")
-    return BillingRoute(provider=provider_name or "unknown", model=model.split("/")[-1] if model else "", base_url=base_url or "", billing_mode="unknown")
-
-
-def _lookup_official_docs_pricing(route: BillingRoute) -> Optional[PricingEntry]:
-    return _OFFICIAL_DOCS_PRICING.get((route.provider, route.model.lower()))
-
-
-def _openrouter_pricing_entry(route: BillingRoute) -> Optional[PricingEntry]:
-    return _pricing_entry_from_metadata(
-        fetch_model_metadata(),
-        route.model,
-        source_url="https://openrouter.ai/docs/api/api-reference/models/get-models",
-        pricing_version="openrouter-models-api",
-    )
-
-
-def _pricing_entry_from_metadata(
-    metadata: Dict[str, Dict[str, Any]],
-    model_id: str,
-    *,
-    source_url: str,
-    pricing_version: str,
-) -> Optional[PricingEntry]:
-    if model_id not in metadata:
-        return None
-    pricing = metadata[model_id].get("pricing") or {}
-    prompt = _to_decimal(pricing.get("prompt"))
-    completion = _to_decimal(pricing.get("completion"))
-    request = _to_decimal(pricing.get("request"))
-    cache_read = _to_decimal(
-        pricing.get("cache_read")
-        or pricing.get("cached_prompt")
-        or pricing.get("input_cache_read")
-    )
-    cache_write = _to_decimal(
-        pricing.get("cache_write")
-        or pricing.get("cache_creation")
-        or pricing.get("input_cache_write")
-    )
-    if prompt is None and completion is None and request is None:
-        return None
-
-    def _per_token_to_per_million(value: Optional[Decimal]) -> Optional[Decimal]:
-        if value is None:
-            return None
-        return value * _ONE_MILLION
-
-    return PricingEntry(
-        input_cost_per_million=_per_token_to_per_million(prompt),
-        output_cost_per_million=_per_token_to_per_million(completion),
-        cache_read_cost_per_million=_per_token_to_per_million(cache_read),
-        cache_write_cost_per_million=_per_token_to_per_million(cache_write),
-        request_cost=request,
-        source="provider_models_api",
-        source_url=source_url,
-        pricing_version=pricing_version,
-        fetched_at=_UTC_NOW(),
-    )
-
-
-def get_pricing_entry(
-    model_name: str,
-    provider: Optional[str] = None,
-    base_url: Optional[str] = None,
-    api_key: Optional[str] = None,
-) -> Optional[PricingEntry]:
-    route = resolve_billing_route(model_name, provider=provider, base_url=base_url)
-    if route.billing_mode == "subscription_included":
-        return PricingEntry(
-            input_cost_per_million=_ZERO,
-            output_cost_per_million=_ZERO,
-            cache_read_cost_per_million=_ZERO,
-            cache_write_cost_per_million=_ZERO,
-            source="none",
-            pricing_version="included-route",
-        )
-    if route.provider == "openrouter":
-        return _openrouter_pricing_entry(route)
-    if route.base_url:
-        entry = _pricing_entry_from_metadata(
-            fetch_endpoint_model_metadata(route.base_url, api_key=api_key or ""),
-            route.model,
-            source_url=f"{route.base_url.rstrip('/')}/models",
-            pricing_version="openai-compatible-models-api",
-        )
-        if entry:
-            return entry
-    return _lookup_official_docs_pricing(route)
-
-
-def normalize_usage(
-    response_usage: Any,
-    *,
-    provider: Optional[str] = None,
-    api_mode: Optional[str] = None,
-) -> CanonicalUsage:
-    """Normalize raw API response usage into canonical token buckets.
-
-    Handles three API shapes:
-    - Anthropic: input_tokens/output_tokens/cache_read_input_tokens/cache_creation_input_tokens
-    - Codex Responses: input_tokens includes cache tokens; input_tokens_details.cached_tokens separates them
-    - OpenAI Chat Completions: prompt_tokens includes cache tokens; prompt_tokens_details.cached_tokens separates them
-
-    In both Codex and OpenAI modes, input_tokens is derived by subtracting cache
-    tokens from the total — the API contract is that input/prompt totals include
-    cached tokens and the details object breaks them out.
-    """
-    if not response_usage:
-        return CanonicalUsage()
-
-    provider_name = (provider or "").strip().lower()
-    mode = (api_mode or "").strip().lower()
-
-    if mode == "anthropic_messages" or provider_name == "anthropic":
-        input_tokens = _to_int(getattr(response_usage, "input_tokens", 0))
-        output_tokens = _to_int(getattr(response_usage, "output_tokens", 0))
-        cache_read_tokens = _to_int(getattr(response_usage, "cache_read_input_tokens", 0))
-        cache_write_tokens = _to_int(getattr(response_usage, "cache_creation_input_tokens", 0))
-    elif mode == "codex_responses":
-        input_total = _to_int(getattr(response_usage, "input_tokens", 0))
-        output_tokens = _to_int(getattr(response_usage, "output_tokens", 0))
-        details = getattr(response_usage, "input_tokens_details", None)
-        cache_read_tokens = _to_int(getattr(details, "cached_tokens", 0) if details else 0)
-        cache_write_tokens = _to_int(
-            getattr(details, "cache_creation_tokens", 0) if details else 0
-        )
-        input_tokens = max(0, input_total - cache_read_tokens - cache_write_tokens)
-    else:
-        prompt_total = _to_int(getattr(response_usage, "prompt_tokens", 0))
-        output_tokens = _to_int(getattr(response_usage, "completion_tokens", 0))
-        details = getattr(response_usage, "prompt_tokens_details", None)
-        cache_read_tokens = _to_int(getattr(details, "cached_tokens", 0) if details else 0)
-        cache_write_tokens = _to_int(
-            getattr(details, "cache_write_tokens", 0) if details else 0
-        )
-        input_tokens = max(0, prompt_total - cache_read_tokens - cache_write_tokens)
-
-    reasoning_tokens = 0
-    output_details = getattr(response_usage, "output_tokens_details", None)
-    if output_details:
-        reasoning_tokens = _to_int(getattr(output_details, "reasoning_tokens", 0))
-
-    return CanonicalUsage(
-        input_tokens=input_tokens,
-        output_tokens=output_tokens,
-        cache_read_tokens=cache_read_tokens,
-        cache_write_tokens=cache_write_tokens,
-        reasoning_tokens=reasoning_tokens,
-    )
-
-
-def estimate_usage_cost(
-    model_name: str,
-    usage: CanonicalUsage,
-    *,
-    provider: Optional[str] = None,
-    base_url: Optional[str] = None,
-    api_key: Optional[str] = None,
-) -> CostResult:
-    route = resolve_billing_route(model_name, provider=provider, base_url=base_url)
-    if route.billing_mode == "subscription_included":
-        return CostResult(
-            amount_usd=_ZERO,
-            status="included",
-            source="none",
-            label="included",
-            pricing_version="included-route",
-        )
-
-    entry = get_pricing_entry(model_name, provider=provider, base_url=base_url, api_key=api_key)
-    if not entry:
-        return CostResult(amount_usd=None, status="unknown", source="none", label="n/a")
-
-    notes: list[str] = []
-    amount = _ZERO
-
-    if usage.input_tokens and entry.input_cost_per_million is None:
-        return CostResult(amount_usd=None, status="unknown", source=entry.source, label="n/a")
-    if usage.output_tokens and entry.output_cost_per_million is None:
-        return CostResult(amount_usd=None, status="unknown", source=entry.source, label="n/a")
-    if usage.cache_read_tokens:
-        if entry.cache_read_cost_per_million is None:
-            return CostResult(
-                amount_usd=None,
-                status="unknown",
-                source=entry.source,
-                label="n/a",
-                notes=("cache-read pricing unavailable for route",),
-            )
-    if usage.cache_write_tokens:
-        if entry.cache_write_cost_per_million is None:
-            return CostResult(
-                amount_usd=None,
-                status="unknown",
-                source=entry.source,
-                label="n/a",
-                notes=("cache-write pricing unavailable for route",),
-            )
-
-    if entry.input_cost_per_million is not None:
-        amount += Decimal(usage.input_tokens) * entry.input_cost_per_million / _ONE_MILLION
-    if entry.output_cost_per_million is not None:
-        amount += Decimal(usage.output_tokens) * entry.output_cost_per_million / _ONE_MILLION
-    if entry.cache_read_cost_per_million is not None:
-        amount += Decimal(usage.cache_read_tokens) * entry.cache_read_cost_per_million / _ONE_MILLION
-    if entry.cache_write_cost_per_million is not None:
-        amount += Decimal(usage.cache_write_tokens) * entry.cache_write_cost_per_million / _ONE_MILLION
-    if entry.request_cost is not None and usage.request_count:
-        amount += Decimal(usage.request_count) * entry.request_cost
-
-    status: CostStatus = "estimated"
-    label = f"~${amount:.2f}"
-    if entry.source == "none" and amount == _ZERO:
-        status = "included"
-        label = "included"
-
-    if route.provider == "openrouter":
-        notes.append("OpenRouter cost is estimated from the models API until reconciled.")
-
-    return CostResult(
-        amount_usd=amount,
-        status=status,
-        source=entry.source,
-        label=label,
-        fetched_at=entry.fetched_at,
-        pricing_version=entry.pricing_version,
-        notes=tuple(notes),
-    )
-
-
-def has_known_pricing(
-    model_name: str,
-    provider: Optional[str] = None,
-    base_url: Optional[str] = None,
-    api_key: Optional[str] = None,
-) -> bool:
-    """Check whether we have pricing data for this model+route.
-
-    Uses direct lookup instead of routing through the full estimation
-    pipeline — avoids creating dummy usage objects just to check status.
-    """
-    route = resolve_billing_route(model_name, provider=provider, base_url=base_url)
-    if route.billing_mode == "subscription_included":
-        return True
-    entry = get_pricing_entry(model_name, provider=provider, base_url=base_url, api_key=api_key)
-    return entry is not None
-
-
-def get_pricing(
-    model_name: str,
-    provider: Optional[str] = None,
-    base_url: Optional[str] = None,
-    api_key: Optional[str] = None,
-) -> Dict[str, float]:
-    """Backward-compatible thin wrapper for legacy callers.
-
-    Returns only non-cache input/output fields when a pricing entry exists.
-    Unknown routes return zeroes.
-    """
-    entry = get_pricing_entry(model_name, provider=provider, base_url=base_url, api_key=api_key)
-    if not entry:
-        return {"input": 0.0, "output": 0.0}
-    return {
-        "input": float(entry.input_cost_per_million or _ZERO),
-        "output": float(entry.output_cost_per_million or _ZERO),
-    }
-
-
-def estimate_cost_usd(
-    model: str,
-    input_tokens: int,
-    output_tokens: int,
-    *,
-    provider: Optional[str] = None,
-    base_url: Optional[str] = None,
-    api_key: Optional[str] = None,
-) -> float:
-    """Backward-compatible helper for legacy callers.
-
-    This uses non-cached input/output only. New code should call
-    `estimate_usage_cost()` with canonical usage buckets.
-    """
-    result = estimate_usage_cost(
-        model,
-        CanonicalUsage(input_tokens=input_tokens, output_tokens=output_tokens),
-        provider=provider,
-        base_url=base_url,
-        api_key=api_key,
-    )
-    return float(result.amount_usd or _ZERO)
-
-
-def format_duration_compact(seconds: float) -> str:
-    if seconds < 60:
-        return f"{seconds:.0f}s"
-    minutes = seconds / 60
-    if minutes < 60:
-        return f"{minutes:.0f}m"
-    hours = minutes / 60
-    if hours < 24:
-        remaining_min = int(minutes % 60)
-        return f"{int(hours)}h {remaining_min}m" if remaining_min else f"{int(hours)}h"
-    days = hours / 24
-    return f"{days:.1f}d"
-
-
-def format_token_count_compact(value: int) -> str:
-    abs_value = abs(int(value))
-    if abs_value < 1_000:
-        return str(int(value))
-
-    sign = "-" if value < 0 else ""
-    units = ((1_000_000_000, "B"), (1_000_000, "M"), (1_000, "K"))
-    for threshold, suffix in units:
-        if abs_value >= threshold:
-            scaled = abs_value / threshold
-            if scaled < 10:
-                text = f"{scaled:.2f}"
-            elif scaled < 100:
-                text = f"{scaled:.1f}"
-            else:
-                text = f"{scaled:.0f}"
-            if "." in text:
-                text = text.rstrip("0").rstrip(".")
-            return f"{sign}{text}{suffix}"
-
-    return f"{value:,}"

atropos/Dockerfile

@@ -0,0 +1,41 @@
+# Dockerfile for atropos-agent sandbox server
+# Runs inside Nomad containers to handle tool execution
+# Includes bubblewrap for namespace-based slot isolation
+
+FROM python:3.11-slim
+
+# Install system dependencies
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    # Bubblewrap for namespace isolation
+    bubblewrap \
+    # `script` for PTY allocation (used for stable tmux+asciinema startup)
+    util-linux \
+    # Git for SWE-style tasks (cloning repos)
+    git \
+    # tmux for stateful terminal sessions (Phase 4.7+)
+    tmux \
+    # Common tools agents might need
+    curl \
+    wget \
+    jq \
+    # Cleanup
+    && rm -rf /var/lib/apt/lists/*
+
+# Install Python dependencies (sandbox server + optional terminal recording)
+RUN pip install --no-cache-dir aiohttp asciinema
+
+# Copy the sandbox server
+COPY sandbox_server.py /app/sandbox_server.py
+
+WORKDIR /app
+
+# Create data directory for slot workspaces
+RUN mkdir -p /data
+
+# Verify bubblewrap is installed and working
+RUN bwrap --version
+
+EXPOSE 8080
+
+# Default command - can be overridden by Nomad job spec
+CMD ["python", "sandbox_server.py", "--port", "8080", "--slots", "10", "--data-dir", "/data"]

atropos/__init__.py

@@ -0,0 +1,47 @@
+"""
+Atropos integration for Hermes-Agent.
+
+This package is intentionally optional: Hermes-Agent should work without Atropos.
+If you import anything from `atropos.*` without having `atroposlib` installed,
+we raise a clear error with install instructions.
+
+Install (recommended, from repo checkout):
+  uv sync --extra atropos
+
+Or (pip / editable):
+  pip install -e '.[atropos]'
+"""
+
+from __future__ import annotations
+
+
+def _require_atroposlib() -> None:
+    try:
+        import atroposlib  # noqa: F401
+    except ModuleNotFoundError as exc:  # pragma: no cover
+        raise ModuleNotFoundError(
+            "Hermes-Agent Atropos integration requires `atroposlib`, but it is not installed.\n"
+            "Install it with:\n"
+            "  uv sync --extra atropos\n"
+            "or:\n"
+            "  pip install -e '.[atropos]'\n"
+        ) from exc
+
+
+_require_atroposlib()
+
+# Re-export the most commonly used pieces for convenience.
+# Agent imports are eager (always available).
+from .agent import AgentConfig, AgentResult, AgentStep, AtroposAgent, SequenceData  # noqa: E402
+
+# Env imports are lazy to avoid pulling in deleted atropos.tools dependencies.
+# Use: from atropos.envs import AgentEnv, AgentEnvConfig  (if needed)
+
+__all__ = [
+    "AtroposAgent",
+    "AgentConfig",
+    "AgentResult",
+    "AgentStep",
+    "SequenceData",
+]
+

atropos/agent/__init__.py

@@ -0,0 +1,15 @@
+"""
+Agent abstractions for atropos-agent.
+
+Provides the core AtroposAgent class for running ReACT-style agent loops.
+"""
+
+from .atropos_agent import AgentConfig, AgentResult, AgentStep, AtroposAgent, SequenceData
+
+__all__ = [
+    "AtroposAgent",
+    "AgentConfig",
+    "AgentResult",
+    "AgentStep",
+    "SequenceData",
+]

atropos/agent/atropos_agent.py

@@ -0,0 +1,850 @@
+"""
+ReACT-style agent implementation for atropos-agent.
+
+This module provides the core AtroposAgent class that implements a basic
+Reason-Act-Observe loop with tool calling capabilities.
+
+Uses ManagedServer from atroposlib for automatic token/logprob tracking,
+making trajectories ready for RL training.
+
+The agent uses Hermes-style XML tags for tool calls:
+- <think>...</think> for reasoning
+- <tool_call>{"name": "...", "arguments": {...}}</tool_call> for actions
+- <tool_response>...</tool_response> for observations
+"""
+
+import asyncio
+import os
+import json
+import time
+from contextlib import asynccontextmanager
+from dataclasses import dataclass, field
+from uuid import uuid4
+from typing import Any, AsyncGenerator, Awaitable, Callable, Dict, List, Optional, Union
+
+from dotenv import load_dotenv
+import httpx
+
+from ..tools import ToolCall, ToolRegistry, ToolResult
+from atroposlib.envs.server_handling.managed_server import ManagedServer
+
+load_dotenv()
+
+
+# Default system prompt with tool calling instructions.
+AGENT_SYSTEM_PROMPT = """You are a deep thinking AI. You MUST enclose your internal reasoning inside <think>...</think> tags.
+
+You are a function calling AI model.
+
+You are provided with function signatures within <tools></tools> XML tags.
+You must call one or more functions to assist with the user query. Don't make assumptions about what values to plug into functions.
+You can ONLY respond without a tool call if you are totally certain you have the final answer to the user's question or task
+After calling & executing a function, you will be provided with function results within <tool_response></tool_response> XML tags.
+
+Here are the available tools:
+<tools>
+{tools_json}
+</tools>
+
+Use the following JSON schema for each tool call you will make:
+{"title": "FunctionCall", "type": "object", "properties": {"name": {"title": "Name", "type": "string"}, "arguments": {"title": "Arguments", "type": "object"}}, "required": ["name", "arguments"]}
+
+## REQUIRED TOOL FORMAT
+
+When you decide to call a tool, your assistant message MUST be:
+1) exactly one <think>...</think> block, followed by
+2) one or more <tool_call>...</tool_call> blocks,
+and NOTHING else in that message.
+
+If you need to explain anything, put it inside <think>. Do NOT write natural language outside <think> or <tool_call>.
+
+For each function call return a JSON object with function name and arguments within <tool_call></tool_call> XML tags as follows:
+<tool_call>
+{"name": "<function-name>", "arguments": {"arg1": "value1"}}
+</tool_call>
+
+Each <tool_call> must be on its own and contain ONLY the JSON object (no extra text).
+The JSON inside <tool_call> MUST be valid JSON with double quotes.
+
+Do NOT output <tool_response> in an assistant message.
+
+After you receive tool results, you may either call more tools (same required format) or provide the final answer.
+When providing the final answer, do NOT include any <tool_call> blocks.
+
+## TERMINAL TOOL NOTES
+
+- Commands execute under POSIX `/bin/sh` (not bash).
+- Each tool call runs in a fresh shell: environment changes (like `cd` or venv activation) do not persist across tool calls.
+- Avoid bash-only features like `source`, `[[ ... ]]`, or process substitution.
+- Prefer explicit venv usage:
+  - `python -m venv .venv && . .venv/bin/activate && python -m pip install -e .` (POSIX `.` activation), or
+  - `.venv/bin/python -m pip install -e .` (no activation required).
+
+## ICL (examples)
+
+User: Show the current directory.
+Assistant:
+<think>I should run pwd.</think>
+<tool_call>
+{"name": "terminal", "arguments": {"command": "pwd"}}
+</tool_call>
+User: <tool_response>{"success": true, "output": "/tmp\\n"}</tool_response>
+Assistant: /tmp
+
+User: List files, then count them.
+Assistant:
+<think>I should count files.</think>
+<tool_call>
+{"name": "terminal", "arguments": {"command": "ls -1 | wc -l"}}
+</tool_call>
+User: <tool_response>{"success": true, "output": "3\\n"}</tool_response>
+Assistant: 3
+
+User: Run pwd, then print ok (two tool calls).
+Assistant:
+<think>I should run two commands.</think>
+<tool_call>
+{"name": "terminal", "arguments": {"command": "pwd"}}
+</tool_call>
+<tool_call>
+{"name": "terminal", "arguments": {"command": "echo ok"}}
+</tool_call>
+User: <tool_response>{"success": true, "output": "/tmp\\n"}</tool_response>
+User: <tool_response>{"success": true, "output": "ok\\n"}</tool_response>
+Assistant: ok
+"""
+
+
+@dataclass
+class AgentConfig:
+    """Configuration for the AtroposAgent."""
+    
+    # Generation parameters
+    temperature: Optional[float] = 0.7
+    # Default to "let the backend decide" (important for tool-tag completions that may be longer).
+    max_tokens: Optional[int] = None
+    
+    # Agent behavior
+    max_steps: int = 50
+    system_prompt: Optional[str] = None
+    tool_delay_s: float = 0.0
+    
+    # Working directory for tools
+    working_dir: Optional[str] = None
+
+
+@dataclass
+class SequenceData:
+    """Token/logprob data from a single completion."""
+    
+    full_text: str
+    tokens: List[int]
+    masked_tokens: List[int]  # -100 for prompt, actual IDs for completion
+    logprobs: List[float]  # 1.0 for prompt, actual values for completion
+    metadata: Optional[Dict[str, Any]] = None
+    
+    @classmethod
+    def from_sequence_node(cls, node) -> "SequenceData":
+        """Create from a ManagedServer SequenceNode."""
+        return cls(
+            full_text=node.full_text,
+            tokens=node.tokens,
+            masked_tokens=node.masked_tokens,
+            logprobs=node.logprobs,
+            metadata=getattr(node, "metadata", None),
+        )
+
+
+@dataclass
+class AgentStep:
+    """A single step in the agent's trajectory."""
+    
+    step_number: int
+    assistant_message: str
+    tool_calls: List[ToolCall] = field(default_factory=list)
+    tool_results: List[ToolResult] = field(default_factory=list)
+    sequence_data: Optional[SequenceData] = None  # Token data from this step
+    
+    @property
+    def has_tool_calls(self) -> bool:
+        return len(self.tool_calls) > 0
+
+
+@dataclass
+class AgentResult:
+    """Result of running an agent trajectory."""
+    
+    success: bool
+    final_response: str
+    steps: List[AgentStep] = field(default_factory=list)
+    total_tokens: int = 0
+    error: Optional[str] = None
+    metadata: Dict[str, Any] = field(default_factory=dict)
+    
+    # Full trajectory token data for RL training
+    trajectory_data: Optional[SequenceData] = None
+    
+    @property
+    def num_steps(self) -> int:
+        return len(self.steps)
+    
+    @property
+    def total_tool_calls(self) -> int:
+        return sum(len(step.tool_calls) for step in self.steps)
+    
+    def to_messages(self) -> List[Dict[str, str]]:
+        """Convert trajectory to messages format for logging."""
+        messages = []
+        for step in self.steps:
+            messages.append({"role": "assistant", "content": step.assistant_message})
+            if step.tool_results:
+                # Combine all tool responses
+                responses = "\n".join(r.to_xml() for r in step.tool_results)
+                messages.append({"role": "user", "content": responses})
+        return messages
+    
+    def to_scored_data(self, score: float) -> Optional[Dict[str, Any]]:
+        """
+        Convert to format suitable for ScoredDataGroup.
+        
+        Args:
+            score: The score for this trajectory
+            
+        Returns:
+            Dict with tokens, masks, scores suitable for training, or None if no data
+        """
+        if self.trajectory_data is None:
+            return None
+        
+        return {
+            "tokens": self.trajectory_data.tokens,
+            "masks": self.trajectory_data.masked_tokens,
+            "scores": score,
+            "logprobs": self.trajectory_data.logprobs,
+        }
+
+
+class AtroposAgent:
+    """
+    A ReACT-style agent that uses LLMs with tool calling.
+    
+    This implementation wraps ManagedServer for automatic token/logprob tracking,
+    making trajectories ready for RL training.
+    
+    Example:
+        # `server` may be an Atropos `ServerManager` (recommended) or a single `APIServer`.
+        # In practice, environments usually construct this via `BaseEnv`.
+        server = ...
+        tools = ToolRegistry()
+        tools.register(BashTool())
+        
+        agent = AtroposAgent(server=server, tools=tools)
+        result = await agent.run("List the files in the current directory")
+        
+        # Access token data for training
+        if result.trajectory_data:
+            print(f"Tokens: {result.trajectory_data.tokens}")
+            print(f"Masked: {result.trajectory_data.masked_tokens}")
+    """
+    
+    def __init__(
+        self,
+        server,  # ServerManager or APIServer
+        tools: Optional[ToolRegistry] = None,
+        config: Optional[AgentConfig] = None,
+        tokenizer: Optional[Any] = None,
+        execute_tool: Optional[Callable[[ToolCall], Awaitable[ToolResult]]] = None,
+    ):
+        self.server = server
+        self.tools = tools or ToolRegistry()
+        self.config = config or AgentConfig()
+        self.tokenizer = tokenizer or getattr(server, "tokenizer", None)
+        self.execute_tool = execute_tool or self.tools.execute
+
+    @asynccontextmanager
+    async def _managed(self) -> AsyncGenerator[Any, None]:
+        """
+        Yield a ManagedServer-like object.
+
+        - If `self.server` is a ServerManager, use its `managed_server()` context manager.
+        - If `self.server` is a single APIServer, wrap it in `ManagedServer` directly.
+        """
+        if os.getenv("ATROPOS_BYPASS_MANAGED_SERVER") == "1":
+            yield _DirectChatCompletionClient(server=self.server)
+            return
+        if hasattr(self.server, "managed_server"):
+            async with self.server.managed_server(tokenizer=self.tokenizer) as managed:
+                yield managed
+        else:
+            managed = ManagedServer(server=self.server, tokenizer=self.tokenizer)
+            try:
+                yield managed
+            finally:
+                managed.reset()
+    
+    def _build_system_prompt(self) -> str:
+        """Build the system prompt with tool descriptions."""
+        if self.config.system_prompt:
+            return self.config.system_prompt
+
+        tools_json = self.tools.get_prompt_tool_definitions_json()
+        # Avoid `str.format()` here because the prompt contains many literal `{}` braces
+        # in JSON examples; we only want to substitute the single `{tools_json}` token.
+        return AGENT_SYSTEM_PROMPT.replace("{tools_json}", tools_json)
+
+    def _infer_server_model_for_debug(self) -> Optional[str]:
+        """
+        Best-effort inference of the configured model name for debug payload saving.
+
+        ManagedServer/server_manager typically injects `model` internally, so `chat_kwargs`
+        may not contain it. For replaying saved payloads via curl, it's useful to persist it.
+        """
+        servers = getattr(self.server, "servers", None)
+        if isinstance(servers, list) and servers:
+            s0 = servers[0]
+            cfg = getattr(s0, "config", None)
+            model = getattr(cfg, "model_name", None) or getattr(s0, "model_name", None)
+            if isinstance(model, str) and model:
+                return model
+        model = getattr(self.server, "model_name", None) or getattr(self.server, "model", None)
+        if isinstance(model, str) and model:
+            return model
+        return None
+
+    def _infer_server_base_url_for_debug(self) -> Optional[str]:
+        """
+        Best-effort inference of the configured base_url for debug logging.
+
+        This is helpful when diagnosing hangs / retries at the transport layer.
+        """
+        servers = getattr(self.server, "servers", None)
+        if isinstance(servers, list) and servers:
+            s0 = servers[0]
+            cfg = getattr(s0, "config", None)
+            base_url = getattr(cfg, "base_url", None) or getattr(s0, "base_url", None)
+            if isinstance(base_url, str) and base_url:
+                return base_url
+        base_url = getattr(self.server, "base_url", None)
+        if isinstance(base_url, str) and base_url:
+            return base_url
+        return None
+
+    def _extract_response_metadata(self, response: Any) -> Dict[str, Any]:
+        """
+        Extract lightweight, JSON-serializable metadata from an OpenAI-style response.
+
+        This is useful for debugging training runs, especially when ManagedServer state
+        tracking is unavailable (e.g. OpenAI-compatible chat endpoints).
+        """
+        meta: Dict[str, Any] = {}
+        try:
+            rid = getattr(response, "id", None)
+            if isinstance(rid, str) and rid:
+                meta["id"] = rid
+            model = getattr(response, "model", None)
+            if isinstance(model, str) and model:
+                meta["model"] = model
+            created = getattr(response, "created", None)
+            if isinstance(created, int):
+                meta["created"] = created
+            system_fingerprint = getattr(response, "system_fingerprint", None)
+            if isinstance(system_fingerprint, str) and system_fingerprint:
+                meta["system_fingerprint"] = system_fingerprint
+
+            choices = getattr(response, "choices", None)
+            if isinstance(choices, list) and choices:
+                fr = getattr(choices[0], "finish_reason", None)
+                if isinstance(fr, str) and fr:
+                    meta["finish_reason"] = fr
+
+            usage = getattr(response, "usage", None)
+            if usage is not None:
+                if hasattr(usage, "model_dump"):
+                    meta["usage"] = usage.model_dump()
+                elif isinstance(usage, dict):
+                    meta["usage"] = usage
+        except Exception:
+            pass
+        return meta
+
+    def _debug_dump_request(self, *, step_num: int, chat_kwargs: Dict[str, Any]) -> None:
+        if os.getenv("ATROPOS_DEBUG_AGENT_REQUEST") != "1":
+            return
+        try:
+            # Avoid dumping megabytes by default; messages can be huge.
+            meta = {
+                "step": step_num,
+                "base_url": self._infer_server_base_url_for_debug(),
+                "model": chat_kwargs.get("model") or self._infer_server_model_for_debug(),
+                "chat_kwargs_keys": sorted(list(chat_kwargs.keys())),
+                "n": chat_kwargs.get("n"),
+                "max_tokens": chat_kwargs.get("max_tokens"),
+                "temperature": chat_kwargs.get("temperature"),
+                "num_messages": len(chat_kwargs.get("messages") or []),
+            }
+            print("\n=== ATROPOS_DEBUG_AGENT_REQUEST ===", flush=True)
+            print(meta, flush=True)
+
+            if os.getenv("ATROPOS_DEBUG_AGENT_REQUEST_FULL") == "1":
+                payload = dict(chat_kwargs)
+                # Make the payload more legible and less huge.
+                try:
+                    dumped = json.dumps(payload, ensure_ascii=False, indent=2)
+                except Exception:
+                    dumped = repr(payload)
+                print("\n=== ATROPOS_DEBUG_AGENT_REQUEST_FULL ===", flush=True)
+                print(dumped[:200_000], flush=True)
+
+            # Optional: save the FULL request payload to disk (no truncation).
+            save_dir = os.getenv("ATROPOS_DEBUG_AGENT_REQUEST_SAVE_DIR")
+            if save_dir:
+                os.makedirs(save_dir, exist_ok=True)
+                payload: Dict[str, Any] = dict(chat_kwargs)
+                if "model" not in payload:
+                    model = self._infer_server_model_for_debug()
+                    if model:
+                        payload["model"] = model
+                # Use a unique filename so parallel trajectories don't clobber each other.
+                fname = os.path.join(
+                    save_dir,
+                    f"atropos_agent_request_step{step_num}_{int(time.time()*1000)}_{os.getpid()}_{uuid4().hex}.json",
+                )
+                with open(fname, "w", encoding="utf-8") as f:
+                    json.dump(payload, f, ensure_ascii=False, indent=2)
+                print(f"[AtroposAgent] saved request payload: {fname}", flush=True)
+        except Exception:
+            return
+
+    def _debug_dump_response(self, *, step_num: int, response: Any) -> None:
+        if os.getenv("ATROPOS_DEBUG_AGENT_RESPONSE") != "1":
+            return
+        print("\n=== ATROPOS_DEBUG_AGENT_RESPONSE ===", flush=True)
+        print({"step": step_num, "type": type(response).__name__}, flush=True)
+        try:
+            dumped = response.model_dump()  # openai pydantic model
+        except Exception:
+            dumped = getattr(response, "__dict__", {"repr": repr(response)})
+        # Keep the dump bounded; we only need enough to see the assistant message content.
+        text = str(dumped)
+        print(text[:200_000], flush=True)
+
+    async def _chat_completion_with_debug(
+        self, *, managed: Any, step_num: int, chat_kwargs: Dict[str, Any]
+    ) -> Any:
+        """
+        Call `managed.chat_completion()` with optional timeout + richer failure logging.
+
+        Debug env vars:
+        - `ATROPOS_AGENT_CHAT_TIMEOUT_S`: if set, wraps the await in `asyncio.wait_for`.
+        - `ATROPOS_DEBUG_AGENT_WAIT_EVERY_S`: if set, prints a heartbeat while waiting.
+        """
+        # Hard guardrail: never allow a single chat completion to block for too long.
+        # This is essential for RL data-gen stability; long hangs should be treated as failures (score=0).
+        timeout_s_raw = os.getenv("ATROPOS_AGENT_CHAT_TIMEOUT_S")
+        timeout_s_default = 240.0
+        timeout_s = float(timeout_s_raw) if timeout_s_raw else timeout_s_default
+        timeout_s = min(timeout_s, 240.0)
+
+        wait_every_raw = os.getenv("ATROPOS_DEBUG_AGENT_WAIT_EVERY_S")
+        wait_every_s = float(wait_every_raw) if wait_every_raw else None
+
+        async def _await_call() -> Any:
+            if not wait_every_s or wait_every_s <= 0:
+                return await managed.chat_completion(**chat_kwargs)
+
+            # Heartbeat mode: wait in chunks without cancelling the underlying request.
+            # NOTE: do NOT use `asyncio.wait_for(task, timeout=...)` here, because a timeout
+            # will cancel the task and surface as `CancelledError` on the next loop.
+            task = asyncio.create_task(managed.chat_completion(**chat_kwargs))
+            t0 = time.perf_counter()
+            try:
+                while True:
+                    done, _pending = await asyncio.wait({task}, timeout=wait_every_s)
+                    if task in done:
+                        return task.result()
+
+                    waited = time.perf_counter() - t0
+                    print(
+                        f"[AtroposAgent] step={step_num} still waiting for chat_completion... ({waited:.1f}s)",
+                        flush=True,
+                    )
+            except asyncio.CancelledError:
+                task.cancel()
+                raise
+
+        try:
+            return await asyncio.wait_for(_await_call(), timeout=timeout_s)
+        except asyncio.TimeoutError as e:
+            print("\n=== ATROPOS_DEBUG_AGENT_CHAT_TIMEOUT ===", flush=True)
+            print({"step": step_num, "timeout_s": timeout_s}, flush=True)
+            raise RuntimeError(f"chat_completion timed out after {timeout_s:.1f}s") from e
+        except asyncio.CancelledError:
+            # Treat cancellation as a hard failure rather than crashing the whole env run.
+            # (Atropos/BaseEnv may cancel tasks during shutdown or retries.)
+            raise RuntimeError("chat_completion cancelled") from None
+        except Exception as e:
+            detail: Dict[str, Any] = {
+                "step": step_num,
+                "exc_type": type(e).__name__,
+                "exc_str": str(e),
+            }
+            if isinstance(e, httpx.HTTPStatusError):
+                try:
+                    detail["status_code"] = e.response.status_code
+                    detail["response_text"] = e.response.text[:20_000]
+                except Exception:
+                    pass
+            elif isinstance(e, httpx.RequestError):
+                detail["request"] = repr(getattr(e, "request", None))
+
+            print("\n=== ATROPOS_DEBUG_AGENT_CHAT_FAILURE ===", flush=True)
+            print(detail, flush=True)
+            raise
+
+    async def run(
+        self,
+        task: str,
+        initial_messages: Optional[List[Dict[str, str]]] = None,
+    ) -> AgentResult:
+        """
+        Run the agent on a task using ManagedServer for token tracking.
+        
+        Args:
+            task: The task/prompt for the agent
+            initial_messages: Optional additional context messages
+            
+        Returns:
+            AgentResult with the trajectory, final response, and token data
+        """
+        messages = [
+            {"role": "system", "content": self._build_system_prompt()},
+        ]
+        
+        if initial_messages:
+            messages.extend(initial_messages)
+        
+        messages.append({"role": "user", "content": task})
+        
+        steps = []
+        final_response = ""
+        final_node = None
+        final_prompt_messages: Optional[List[Dict[str, str]]] = None
+        last_node = None
+        last_prompt_messages: Optional[List[Dict[str, str]]] = None
+        last_response_text: str = ""
+        
+        # Use ManagedServer for automatic token tracking
+        async with self._managed() as managed:
+            for step_num in range(self.config.max_steps):
+                # ReACT loop iteration here, just call -> tools -> observe until done (no tools called)
+                try:
+                    # Keep a copy of the prompt messages used for this completion.
+                    # Useful for reconstructing tokens/masks when state tracking is unavailable.
+                    prompt_messages = list(messages)
+                    chat_kwargs: Dict[str, Any] = {"messages": messages, "n": 1}
+                    if self.config.max_tokens is not None:
+                        chat_kwargs["max_tokens"] = self.config.max_tokens
+                    if self.config.temperature is not None:
+                        chat_kwargs["temperature"] = self.config.temperature
+
+                    t_req = time.perf_counter()
+                    print(
+                        f"[AtroposAgent] step={step_num+1} chat_completion start "
+                        f"(messages={len(messages)}, max_tokens={self.config.max_tokens}, temp={self.config.temperature})",
+                        flush=True,
+                    )
+                    self._debug_dump_request(step_num=step_num + 1, chat_kwargs=chat_kwargs)
+                    response = await self._chat_completion_with_debug(
+                        managed=managed, step_num=step_num + 1, chat_kwargs=chat_kwargs
+                    )
+                    self._debug_dump_response(step_num=step_num + 1, response=response)
+                    response_meta = self._extract_response_metadata(response)
+                    print(
+                        f"[AtroposAgent] step={step_num+1} chat_completion done in {time.perf_counter() - t_req:.2f}s",
+                        flush=True,
+                    )
+                    
+                    current_node = None
+                    if hasattr(managed, "get_state"):
+                        state = managed.get_state()
+                        nodes = state.get("nodes", [])
+                        current_node = nodes[-1] if nodes else None
+                    
+                except Exception as e:
+                    return AgentResult(
+                        success=False,
+                        final_response="",
+                        steps=steps,
+                        error=f"Generation error: {str(e)}",
+                    )
+                
+                msg = response.choices[0].message
+                # Some OpenAI-compatible servers populate `message.reasoning` and leave `content=""`.
+                response_text = (msg.content or "") or (getattr(msg, "reasoning", None) or "")
+                tool_calls = ToolCall.parse_from_text(response_text)
+                last_node = current_node
+                last_prompt_messages = prompt_messages
+                last_response_text = response_text
+
+                step_sequence_data = SequenceData.from_sequence_node(current_node) if current_node else None
+                if step_sequence_data is None:
+                    if response_meta:
+                        # We still want metadata for debugging even if token/logprob state tracking is unavailable.
+                        step_sequence_data = SequenceData(
+                            full_text=response_text,
+                            tokens=[],
+                            masked_tokens=[],
+                            logprobs=[],
+                            metadata=response_meta,
+                        )
+                else:
+                    merged = dict(response_meta)
+                    node_meta = step_sequence_data.metadata
+                    if isinstance(node_meta, dict):
+                        merged.update(node_meta)
+                    step_sequence_data.metadata = merged or step_sequence_data.metadata
+                
+                step = AgentStep(
+                    step_number=step_num + 1,
+                    assistant_message=response_text,
+                    tool_calls=tool_calls,
+                    sequence_data=step_sequence_data,
+                )
+                
+                if not tool_calls:
+                    steps.append(step)
+                    final_response = response_text
+                    final_node = current_node
+                    final_prompt_messages = prompt_messages
+                    break
+                
+                messages.append({"role": "assistant", "content": response_text})
+                
+                tool_responses = []
+                for call in tool_calls:
+                    result = await self.execute_tool(call)
+                    step.tool_results.append(result)
+                    tool_responses.append(result.to_xml())
+                    if self.config.tool_delay_s > 0:
+                        await asyncio.sleep(self.config.tool_delay_s)
+                
+                steps.append(step)
+            
+                responses_text = "\n".join(tool_responses)
+                # Tool observations are represented as user content with Hermes-style tags.
+                # This is compatible with most OpenAI-compatible chat APIs and ensures
+                # tokenizers/chat templates include tool outputs during training.
+                messages.append({"role": "user", "content": responses_text})
+            
+            else:
+                # Reached max steps without completing
+                # Return a failure result but include the last observed completion so callers can
+                # record the trajectory (score=0) without triggering retries.
+                final_response = last_response_text or final_response
+                final_node = last_node
+                final_prompt_messages = last_prompt_messages
+                trajectory_data = None
+                if final_node:
+                    trajectory_data = SequenceData.from_sequence_node(final_node)
+                elif final_prompt_messages is not None and self.tokenizer is not None:
+                    if hasattr(self.tokenizer, "apply_chat_template"):
+                        prompt_text = self.tokenizer.apply_chat_template(
+                            final_prompt_messages, tokenize=False, add_generation_prompt=True
+                        )
+                        prompt_tokens = self.tokenizer.encode(prompt_text, add_special_tokens=False)
+                    else:
+                        prompt_text = "\n".join([f"{m['role']}: {m['content']}" for m in final_prompt_messages])
+                        prompt_tokens = self.tokenizer.encode(prompt_text, add_special_tokens=True)
+                    output_tokens = self.tokenizer.encode(final_response, add_special_tokens=False)
+                    tokens = prompt_tokens + output_tokens
+                    masked_tokens = ([-100] * len(prompt_tokens)) + output_tokens
+                    logprobs = ([1.0] * len(prompt_tokens)) + ([0.0] * len(output_tokens))
+                    trajectory_data = SequenceData(
+                        full_text=f"{prompt_text}{final_response}",
+                        tokens=tokens,
+                        masked_tokens=masked_tokens,
+                        logprobs=logprobs,
+                    )
+                # Preserve response metadata (if any) even on failure trajectories.
+                try:
+                    if trajectory_data is not None and steps:
+                        last_step = steps[-1]
+                        if last_step.sequence_data and isinstance(last_step.sequence_data.metadata, dict):
+                            trajectory_data.metadata = dict(last_step.sequence_data.metadata)
+                except Exception:
+                    pass
+                return AgentResult(
+                    success=False,
+                    final_response=final_response,
+                    steps=steps,
+                    error=f"Reached maximum steps ({self.config.max_steps})",
+                    trajectory_data=trajectory_data,
+                )
+        
+        # Build result with trajectory data
+        trajectory_data = None
+        if final_node:
+            trajectory_data = SequenceData.from_sequence_node(final_node)
+        elif final_prompt_messages is not None and self.tokenizer is not None:
+            if hasattr(self.tokenizer, "apply_chat_template"):
+                prompt_text = self.tokenizer.apply_chat_template(
+                    final_prompt_messages, tokenize=False, add_generation_prompt=True
+                )
+                prompt_tokens = self.tokenizer.encode(prompt_text, add_special_tokens=False)
+            else:
+                prompt_text = "\n".join([f"{m['role']}: {m['content']}" for m in final_prompt_messages])
+                prompt_tokens = self.tokenizer.encode(prompt_text, add_special_tokens=True)
+            output_tokens = self.tokenizer.encode(final_response, add_special_tokens=False)
+            tokens = prompt_tokens + output_tokens
+            masked_tokens = ([-100] * len(prompt_tokens)) + output_tokens
+            logprobs = ([1.0] * len(prompt_tokens)) + ([0.0] * len(output_tokens))
+            trajectory_data = SequenceData(
+                full_text=f"{prompt_text}{final_response}",
+                tokens=tokens,
+                masked_tokens=masked_tokens,
+                logprobs=logprobs,
+            )
+
+        # Ensure trajectory_data carries the most recent metadata we observed (if any).
+        try:
+            if trajectory_data is not None and steps:
+                last_step = steps[-1]
+                if last_step.sequence_data and isinstance(last_step.sequence_data.metadata, dict):
+                    trajectory_data.metadata = dict(last_step.sequence_data.metadata)
+        except Exception:
+            pass
+        
+        return AgentResult(
+            success=True,
+            final_response=final_response,
+            steps=steps,
+            trajectory_data=trajectory_data,
+        )
+    
+    async def run_single_turn(
+        self,
+        messages: List[Dict[str, str]],
+        execute_tools: bool = True,
+    ) -> tuple[str, List[ToolResult], Optional[SequenceData]]:
+        """
+        Run a single turn of the agent (one LLM call + tool execution).
+        
+        This is useful for integration with BaseEnv where you want more
+        control over the loop.
+        
+        Args:
+            messages: The conversation history
+            execute_tools: Whether to execute parsed tool calls
+            
+        Returns:
+            Tuple of (response_text, tool_results, sequence_data)
+        """
+        async with self._managed() as managed:
+            chat_kwargs: Dict[str, Any] = {"messages": messages, "n": 1}
+            if self.config.max_tokens is not None:
+                chat_kwargs["max_tokens"] = self.config.max_tokens
+            if self.config.temperature is not None:
+                chat_kwargs["temperature"] = self.config.temperature
+
+            self._debug_dump_request(step_num=1, chat_kwargs=chat_kwargs)
+            response = await self._chat_completion_with_debug(managed=managed, step_num=1, chat_kwargs=chat_kwargs)
+            self._debug_dump_response(step_num=1, response=response)
+            
+            current_node = None
+            if hasattr(managed, "get_state"):
+                state = managed.get_state()
+                nodes = state.get("nodes", [])
+                current_node = nodes[-1] if nodes else None
+        
+        msg = response.choices[0].message
+        response_text = (msg.content or "") or (getattr(msg, "reasoning", None) or "")
+        tool_results = []
+        
+        if execute_tools:
+            tool_calls = ToolCall.parse_from_text(response_text)
+            for call in tool_calls:
+                result = await self.execute_tool(call)
+                tool_results.append(result)
+        
+        sequence_data = SequenceData.from_sequence_node(current_node) if current_node else None
+        
+        return response_text, tool_results, sequence_data
+
+
+class _DirectChatCompletionClient:
+    """
+    Minimal stand-in for ManagedServer that calls the OpenAI-compatible endpoint directly.
+
+    This is for isolating issues where `ManagedServer.chat_completion()` hangs or misbehaves.
+    It intentionally does NOT do token/logprob tracking.
+    """
+
+    def __init__(self, server: Any):
+        self._server = server
+
+    def _server_config(self) -> tuple[str, str, str]:
+        # ServerManager case: first configured server.
+        servers = getattr(self._server, "servers", None)
+        if isinstance(servers, list) and servers:
+            s0 = servers[0]
+            cfg = getattr(s0, "config", None)
+            base_url = getattr(cfg, "base_url", None) or getattr(s0, "base_url", None)
+            api_key = getattr(cfg, "api_key", None) or getattr(s0, "api_key", None)
+            model = getattr(cfg, "model_name", None) or getattr(s0, "model_name", None)
+            if isinstance(base_url, str) and isinstance(api_key, str) and isinstance(model, str):
+                return base_url.rstrip("/"), api_key, model
+
+        # APIServer-like fallback.
+        base_url = getattr(self._server, "base_url", None)
+        api_key = getattr(self._server, "api_key", None)
+        model = getattr(self._server, "model_name", None) or getattr(self._server, "model", None)
+        if isinstance(base_url, str) and isinstance(api_key, str) and isinstance(model, str):
+            return base_url.rstrip("/"), api_key, model
+
+        raise RuntimeError("Unable to resolve server base_url/api_key/model for direct chat completion")
+
+    async def chat_completion(self, *, messages: List[Dict[str, str]], n: int = 1, **kwargs: Any) -> Any:
+        base_url, api_key, model = self._server_config()
+        url = f"{base_url}/chat/completions"
+
+        payload: Dict[str, Any] = {
+            "model": model,
+            "messages": messages,
+            "n": n,
+        }
+        # Pass through common generation kwargs.
+        for k in ("max_tokens", "temperature", "top_p", "presence_penalty", "frequency_penalty", "stop"):
+            if k in kwargs and kwargs[k] is not None:
+                payload[k] = kwargs[k]
+
+        timeout_s = float(os.getenv("ATROPOS_DIRECT_REQUEST_TIMEOUT_S") or "120")
+        print(f"[AtroposAgent] DIRECT chat_completion POST {url} (timeout={timeout_s}s)", flush=True)
+        async with httpx.AsyncClient(timeout=timeout_s) as client:
+            resp = await client.post(
+                url,
+                headers={"Authorization": f"Bearer {api_key}", "Content-Type": "application/json"},
+                json=payload,
+            )
+            resp.raise_for_status()
+            data = resp.json()
+
+        # Return a very small object compatible with the code paths that read
+        # `response.choices[0].message.content`.
+        class _Msg:
+            def __init__(self, d: Dict[str, Any]):
+                self.content = d.get("content")
+                self.reasoning = d.get("reasoning")
+
+        class _Choice:
+            def __init__(self, d: Dict[str, Any]):
+                self.message = _Msg(d.get("message") or {})
+
+        class _Resp:
+            def __init__(self, d: Dict[str, Any]):
+                self._d = d
+                self.choices = [_Choice(c) for c in (d.get("choices") or [])]
+
+            def model_dump(self) -> Dict[str, Any]:
+                return self._d
+
+        return _Resp(data)

atropos/api/__init__.py

@@ -0,0 +1,6 @@
+"""
+FastAPI services for atropos-agent.
+
+- tool_executor_server: queued/batched sandbox tool execution (Phase 4)
+"""
+

atropos/api/tool_executor_server.py

@@ -0,0 +1,254 @@
+"""
+Tool Executor API (Phase 4)
+
+This service provides a queued, batched execution layer on top of a ToolBackend.
+It mirrors the stateful FastAPI + app.state pattern used in:
+  atropos/atroposlib/api/server.py
+
+Run (dev):
+  uv run uvicorn atropos_agent.api.tool_executor_server:app --host 0.0.0.0 --port 9001
+"""
+
+from __future__ import annotations
+
+import os
+from typing import Any, Dict, Optional
+from pathlib import Path
+
+from fastapi import FastAPI, Header, HTTPException, status
+from pydantic import BaseModel, Field
+
+from ..backends.nomad_backend import NomadBackendConfig, NomadToolBackend
+from ..tools import ToolRegistry, build_tool_registry
+from ..tools.base import (
+    ArtifactArchiveRequestPayload,
+    ArtifactArchiveResponsePayload,
+    ArtifactListRequestPayload,
+    ArtifactListResponsePayload,
+    ArtifactReadRequestPayload,
+    ArtifactReadResponsePayload,
+    ToolExecutorExecuteRequest,
+    ToolExecutorReleaseRequest,
+    ToolResultPayload,
+)
+from ..tools.tool_executor import ToolExecutor, ToolExecutorConfig
+
+
+class ToolExecutorServerConfig(BaseModel):
+    nomad_address: str = Field(default="http://localhost:4646")
+    job_id: str = Field(default="atropos-sandbox-tool-executor")
+    image: str = Field(default="atropos-sandbox:local")
+    slots_per_container: int = Field(default=10)
+    min_containers: int = Field(default=1)
+    max_containers: int = Field(default=10)
+    privileged: bool = Field(default=False)
+    acquire_timeout_s: float = Field(default=30.0)
+
+    batch_window_ms: int = Field(default=20)
+    max_batch_size: int = Field(default=200)
+    allow_network: bool = Field(default=True)
+
+    tool_server_url: Optional[str] = Field(default=None)
+    tool_server_token: Optional[str] = Field(default=None)
+
+    token: Optional[str] = Field(default=None, description="Bearer token required for requests (optional in dev).")
+
+    purge_job_on_shutdown: bool = Field(default=True)
+
+    @classmethod
+    def from_env(cls) -> "ToolExecutorServerConfig":
+        # In dev, prefer loading secrets/config from the repo-local `.env` (not committed).
+        try:
+            from dotenv import load_dotenv  # type: ignore
+        except Exception:  # pragma: no cover
+            load_dotenv = None  # type: ignore[assignment]
+        if load_dotenv is not None:
+            env_path = Path(__file__).resolve().parents[2] / ".env"
+            if env_path.exists():
+                load_dotenv(dotenv_path=env_path)
+
+        def _get_bool(name: str, default: bool) -> bool:
+            raw = os.getenv(name)
+            if raw is None:
+                return default
+            return raw.strip().lower() in {"1", "true", "yes", "y", "on"}
+
+        return cls(
+            nomad_address=os.getenv("TOOL_EXECUTOR_NOMAD_ADDRESS", "http://localhost:4646"),
+            job_id=os.getenv("TOOL_EXECUTOR_JOB_ID", "atropos-sandbox-tool-executor"),
+            image=os.getenv("TOOL_EXECUTOR_IMAGE", "atropos-sandbox:local"),
+            slots_per_container=int(os.getenv("TOOL_EXECUTOR_SLOTS", "10")),
+            min_containers=int(os.getenv("TOOL_EXECUTOR_MIN_CONTAINERS", "1")),
+            max_containers=int(os.getenv("TOOL_EXECUTOR_MAX_CONTAINERS", "10")),
+            privileged=_get_bool("TOOL_EXECUTOR_PRIVILEGED", False),
+            acquire_timeout_s=float(os.getenv("TOOL_EXECUTOR_ACQUIRE_TIMEOUT_S", "30.0")),
+            batch_window_ms=int(os.getenv("TOOL_EXECUTOR_BATCH_WINDOW_MS", "20")),
+            max_batch_size=int(os.getenv("TOOL_EXECUTOR_MAX_BATCH_SIZE", "200")),
+            allow_network=_get_bool("TOOL_EXECUTOR_ALLOW_NETWORK", True),
+            tool_server_url=os.getenv("TOOL_EXECUTOR_TOOL_SERVER_URL") or None,
+            tool_server_token=os.getenv("TOOL_EXECUTOR_TOOL_SERVER_TOKEN") or None,
+            token=os.getenv("TOOL_EXECUTOR_TOKEN") or None,
+            purge_job_on_shutdown=_get_bool("TOOL_EXECUTOR_PURGE_JOB_ON_SHUTDOWN", True),
+        )
+
+
+app = FastAPI(title="Atropos-Agent Tool Executor")
+
+
+@app.get("/")
+async def root() -> Dict[str, str]:
+    return {"message": "Atropos-Agent Tool Executor"}
+
+
+def _check_auth(cfg: ToolExecutorServerConfig, authorization: Optional[str]) -> None:
+    if not cfg.token:
+        return
+    if not authorization:
+        raise HTTPException(status_code=status.HTTP_401_UNAUTHORIZED, detail="Missing Authorization header")
+    if not authorization.lower().startswith("bearer "):
+        raise HTTPException(status_code=status.HTTP_401_UNAUTHORIZED, detail="Invalid Authorization header")
+    token = authorization.split(" ", 1)[1].strip()
+    if token != cfg.token:
+        raise HTTPException(status_code=status.HTTP_403_FORBIDDEN, detail="Invalid token")
+
+
+@app.on_event("startup")
+async def _startup() -> None:
+    cfg = ToolExecutorServerConfig.from_env()
+
+    # Default to Atropos "full" tool surface: sandbox + external (if tool_server_url provided).
+    tools: ToolRegistry = build_tool_registry(
+        enabled_toolsets=["full"],
+        disabled_toolsets=None,
+        tool_server_url=cfg.tool_server_url,
+    )
+
+    backend = NomadToolBackend(
+        NomadBackendConfig(
+            nomad_address=cfg.nomad_address,
+            sandbox_job_id=cfg.job_id,
+            sandbox_image=cfg.image,
+            slots_per_container=cfg.slots_per_container,
+            min_containers=cfg.min_containers,
+            max_containers=cfg.max_containers,
+            privileged=cfg.privileged,
+            acquire_timeout_s=cfg.acquire_timeout_s,
+            purge_job_on_start=False,
+        )
+    )
+    await backend.start()
+
+    executor = ToolExecutor(
+        backend=backend,
+        tools=tools,
+        config=ToolExecutorConfig(
+            batch_window_ms=cfg.batch_window_ms,
+            max_batch_size=cfg.max_batch_size,
+            allow_network=cfg.allow_network,
+            tool_server_url=cfg.tool_server_url,
+            tool_server_token=cfg.tool_server_token,
+        ),
+    )
+    await executor.start()
+
+    app.state.cfg = cfg
+    app.state.backend = backend
+    app.state.executor = executor
+
+
+@app.on_event("shutdown")
+async def _shutdown() -> None:
+    executor: Optional[ToolExecutor] = getattr(app.state, "executor", None)
+    backend: Optional[NomadToolBackend] = getattr(app.state, "backend", None)
+    cfg: Optional[ToolExecutorServerConfig] = getattr(app.state, "cfg", None)
+
+    if executor is not None:
+        await executor.close()
+
+    if backend is not None:
+        await backend.stop(purge=bool(cfg.purge_job_on_shutdown) if cfg else False)
+
+
+@app.get("/health")
+async def health() -> Dict[str, Any]:
+    return {"status": "ok"}
+
+
+@app.get("/status")
+async def status_endpoint() -> Dict[str, Any]:
+    executor: ToolExecutor = app.state.executor
+    backend: NomadToolBackend = app.state.backend
+
+    return {
+        "queue_size": executor.queue_size(),
+        "total_requests": executor.total_requests,
+        "total_errors": executor.total_errors,
+        "pool": backend.get_stats(),
+    }
+
+
+@app.post("/execute", response_model=ToolResultPayload)
+async def execute_tool(
+    req: ToolExecutorExecuteRequest,
+    authorization: Optional[str] = Header(default=None),
+    status_code: int = status.HTTP_200_OK,  # noqa: B008
+) -> ToolResultPayload:
+    cfg: ToolExecutorServerConfig = app.state.cfg
+    _check_auth(cfg, authorization)
+
+    executor: ToolExecutor = app.state.executor
+    result = await executor.execute(
+        trajectory_id=req.trajectory_id,
+        call=req.tool.to_tool_call(),
+        timeout_s=req.timeout_s,
+    )
+    return ToolResultPayload.from_tool_result(result)
+
+
+@app.post("/release")
+async def release_trajectory(
+    req: ToolExecutorReleaseRequest,
+    authorization: Optional[str] = Header(default=None),
+) -> Dict[str, Any]:
+    cfg: ToolExecutorServerConfig = app.state.cfg
+    _check_auth(cfg, authorization)
+
+    executor: ToolExecutor = app.state.executor
+    await executor.release_trajectory(req.trajectory_id, reset_workspace=req.reset_workspace)
+    return {"status": "ok"}
+
+
+@app.post("/artifacts/read", response_model=ArtifactReadResponsePayload)
+async def artifacts_read(
+    req: ArtifactReadRequestPayload,
+    authorization: Optional[str] = Header(default=None),
+) -> ArtifactReadResponsePayload:
+    cfg: ToolExecutorServerConfig = app.state.cfg
+    _check_auth(cfg, authorization)
+
+    executor: ToolExecutor = app.state.executor
+    return await executor.read_artifact(req)
+
+
+@app.post("/artifacts/list", response_model=ArtifactListResponsePayload)
+async def artifacts_list(
+    req: ArtifactListRequestPayload,
+    authorization: Optional[str] = Header(default=None),
+) -> ArtifactListResponsePayload:
+    cfg: ToolExecutorServerConfig = app.state.cfg
+    _check_auth(cfg, authorization)
+
+    executor: ToolExecutor = app.state.executor
+    return await executor.list_artifacts(req)
+
+
+@app.post("/artifacts/archive", response_model=ArtifactArchiveResponsePayload)
+async def artifacts_archive(
+    req: ArtifactArchiveRequestPayload,
+    authorization: Optional[str] = Header(default=None),
+) -> ArtifactArchiveResponsePayload:
+    cfg: ToolExecutorServerConfig = app.state.cfg
+    _check_auth(cfg, authorization)
+
+    executor: ToolExecutor = app.state.executor
+    return await executor.archive_artifacts(req)

atropos/api/tool_server.py

@@ -0,0 +1,140 @@
+"""
+External ToolServer (Phase 4.5+).
+
+This server executes tools that must NOT run inside the sandbox, typically
+because they require credentials or access to external services.
+
+Run (dev):
+  uv run uvicorn atropos_agent.api.tool_server:app --host 0.0.0.0 --port 9002
+"""
+
+from __future__ import annotations
+
+import asyncio
+import os
+import inspect
+from typing import Any, Dict, List, Optional
+from pathlib import Path
+
+from fastapi import FastAPI, Header, HTTPException, status
+from pydantic import BaseModel, Field
+
+from ..tools import ToolRegistry, build_tool_registry
+from ..tools.base import ToolResultPayload, ToolServerExecuteRequest
+
+
+class ToolServerConfig(BaseModel):
+    token: Optional[str] = Field(
+        default=None,
+        description="Bearer token required for requests (optional in dev).",
+    )
+    max_concurrency: int = Field(default=16, ge=1, description="Max concurrent tool executions.")
+
+    @classmethod
+    def from_env(cls) -> "ToolServerConfig":
+        # In dev, prefer loading secrets from the repo-local `.env` (not committed).
+        try:
+            from dotenv import load_dotenv  # type: ignore
+        except Exception:  # pragma: no cover
+            load_dotenv = None  # type: ignore[assignment]
+        if load_dotenv is not None:
+            env_path = Path(__file__).resolve().parents[2] / ".env"
+            if env_path.exists():
+                load_dotenv(dotenv_path=env_path)
+
+        token = os.getenv("TOOL_SERVER_TOKEN") or None
+        max_concurrency = int(os.getenv("TOOL_SERVER_MAX_CONCURRENCY", "16"))
+        return cls(token=token, max_concurrency=max_concurrency)
+
+
+app = FastAPI(title="Atropos-Agent Tool Server")
+
+
+@app.get("/")
+async def root() -> Dict[str, str]:
+    return {"message": "Atropos-Agent Tool Server"}
+
+
+@app.on_event("startup")
+async def _startup() -> None:
+    cfg = ToolServerConfig.from_env()
+
+    # External-only registry. It will only include tools that are enabled by toolsets and
+    # whose Hermes requirements/keys are satisfied in this process.
+    tools: ToolRegistry = build_tool_registry(
+        enabled_toolsets=["all"],
+        disabled_toolsets=["terminal", "sandbox", "filesystem", "terminal_stateful", "default"],
+        tool_server_url="enabled",
+    )
+
+    app.state.cfg = cfg
+    app.state.tools = tools
+    app.state.semaphore = asyncio.Semaphore(cfg.max_concurrency)
+
+
+@app.get("/health")
+async def health() -> Dict[str, Any]:
+    return {"status": "ok"}
+
+
+@app.get("/tools")
+async def list_tools() -> Dict[str, Any]:
+    tools: ToolRegistry = app.state.tools
+    return {"tools": [s.to_dict() for s in tools.get_schemas()]}
+
+
+def _check_auth(cfg: ToolServerConfig, authorization: Optional[str]) -> None:
+    if not cfg.token:
+        return
+    if not authorization:
+        raise HTTPException(status_code=status.HTTP_401_UNAUTHORIZED, detail="Missing Authorization header")
+    if not authorization.lower().startswith("bearer "):
+        raise HTTPException(status_code=status.HTTP_401_UNAUTHORIZED, detail="Invalid Authorization header")
+    token = authorization.split(" ", 1)[1].strip()
+    if token != cfg.token:
+        raise HTTPException(status_code=status.HTTP_403_FORBIDDEN, detail="Invalid token")
+
+
+@app.post("/execute", response_model=ToolResultPayload)
+async def execute_tool(
+    req: ToolServerExecuteRequest,
+    authorization: Optional[str] = Header(default=None),
+) -> ToolResultPayload:
+    cfg: ToolServerConfig = app.state.cfg
+    _check_auth(cfg, authorization)
+
+    tools: ToolRegistry = app.state.tools
+    sem: asyncio.Semaphore = app.state.semaphore
+
+    tool = tools.get(req.tool.name)
+    if tool is None:
+        return ToolResultPayload(
+            success=False,
+            error=f"Unknown tool: {req.tool.name}",
+            uniq_id=req.tool.uniq_id,
+        )
+
+    async with sem:
+        try:
+            kwargs = dict(req.tool.arguments)
+            sig = inspect.signature(tool.execute).parameters
+            # Some tools can benefit from extra context.
+            if req.trajectory_id and "trajectory_id" in sig:
+                kwargs["trajectory_id"] = req.trajectory_id
+            if req.slot_id and "slot_id" in sig:
+                kwargs["slot_id"] = req.slot_id
+            if req.container_addr and "container_addr" in sig:
+                kwargs["container_addr"] = req.container_addr
+            if "task_id" in sig:
+                kwargs["task_id"] = req.trajectory_id
+            result = await tool.execute(**kwargs)
+        except Exception as e:
+            return ToolResultPayload(
+                success=False,
+                error=f"Tool execution error: {e}",
+                uniq_id=req.tool.uniq_id,
+            )
+
+    if result.uniq_id is None:
+        result.uniq_id = req.tool.uniq_id
+    return ToolResultPayload.from_tool_result(result)

atropos/backends/__init__.py

@@ -0,0 +1,27 @@
+from __future__ import annotations
+
+from typing import Any
+
+from .base import ToolBackend
+from .modal_backend import ModalSandboxConfig, ModalToolBackend
+from .nomad_backend import NomadBackendConfig, NomadToolBackend
+
+
+def create_tool_backend(cfg: Any) -> ToolBackend:
+    mode = str(getattr(cfg, "tool_pool_mode", "nomad")).strip().lower()
+    if mode == "nomad":
+        return NomadToolBackend(NomadBackendConfig.from_agent_env_config(cfg))
+    if mode == "modal":
+        return ModalToolBackend(ModalSandboxConfig.from_agent_env_config(cfg))
+    raise ValueError(f"Unknown tool_pool_mode: {mode}")
+
+
+__all__ = [
+    "ToolBackend",
+    "create_tool_backend",
+    "NomadBackendConfig",
+    "NomadToolBackend",
+    "ModalSandboxConfig",
+    "ModalToolBackend",
+]
+

atropos/backends/base.py

@@ -0,0 +1,89 @@
+"""
+Backend interfaces for AgentEnv tool execution.
+
+The goal of this module is to decouple ToolExecutor / AgentEnv from any single
+execution backend (Nomad/Docker today; Modal later).
+"""
+
+from __future__ import annotations
+
+from typing import Any, Dict, List, Optional, Protocol, Tuple
+
+from ..slots.executor import ExecutionResult
+from ..slots.slot import Slot
+
+
+class ToolBackend(Protocol):
+    """
+    Minimal interface required by ToolExecutor.
+
+    Backends provide:
+    - lifecycle (start/stop)
+    - slot acquisition/release (workspace affinity)
+    - batched tool execution across slots
+    - optional artifact helpers (for env verification / demos)
+    """
+
+    @property
+    def default_timeout_s(self) -> Optional[float]:
+        """Default sandbox execution timeout in seconds (if any)."""
+
+    async def start(self) -> None:
+        """Start the backend (provision workers/containers, health checks, etc)."""
+
+    async def stop(self, *, purge: bool = False) -> None:
+        """Stop the backend and optionally purge remote resources."""
+
+    async def acquire(self, trajectory_id: Optional[str] = None) -> Slot:
+        """Acquire a slot for a trajectory (workspace affinity)."""
+
+    async def release(self, slot: Slot, *, reset_workspace: bool = False) -> None:
+        """Release a slot back to the pool."""
+
+    async def execute_batch(
+        self,
+        requests: List[Tuple[Slot, str, Dict[str, Any]]],
+        *,
+        timeout_s: Optional[float] = None,
+    ) -> List[ExecutionResult]:
+        """Execute a batch of sandbox tool calls and return results in order."""
+
+    # ---------------------------------------------------------------------
+    # Optional artifact helpers (supported by the Nomad sandbox-server today)
+    # ---------------------------------------------------------------------
+
+    async def read_artifact(
+        self,
+        slot: Slot,
+        path: str,
+        *,
+        encoding: str = "text",
+        max_bytes: Optional[int] = None,
+        include_sha256: bool = False,
+        timeout_s: Optional[float] = None,
+    ) -> Dict[str, Any]:
+        raise NotImplementedError
+
+    async def list_artifacts(
+        self,
+        slot: Slot,
+        path: str = ".",
+        *,
+        recursive: bool = False,
+        max_entries: Optional[int] = None,
+        timeout_s: Optional[float] = None,
+    ) -> Dict[str, Any]:
+        raise NotImplementedError
+
+    async def archive_artifacts(
+        self,
+        slot: Slot,
+        path: str = ".",
+        *,
+        archive_format: str = "tar.gz",
+        max_bytes: Optional[int] = None,
+        max_entries: Optional[int] = None,
+        timeout_s: Optional[float] = None,
+    ) -> Dict[str, Any]:
+        raise NotImplementedError
+

atropos/backends/nomad_backend.py

@@ -0,0 +1,156 @@
+"""
+Nomad/Docker tool backend.
+
+This backend is the current default for AgentEnv: it provisions a Nomad job
+running `sandbox_server.py` and multiplexes stateless slots inside each container.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple
+
+from ..slots import Slot, SlotPool, SlotPoolConfig
+from ..slots.executor import ExecutionResult
+from .base import ToolBackend
+
+
+@dataclass(frozen=True)
+class NomadBackendConfig:
+    nomad_address: str
+    sandbox_job_id: str
+    sandbox_image: str
+    slots_per_container: int
+    min_containers: int
+    max_containers: int
+    privileged: bool
+    acquire_timeout_s: float
+    purge_job_on_start: bool
+    # Driver selection: "docker" or "singularity"
+    driver: str = "docker"
+    # Path to .sif file for singularity driver (required if driver="singularity")
+    singularity_image: Optional[str] = None
+
+    @classmethod
+    def from_agent_env_config(cls, cfg: Any) -> "NomadBackendConfig":
+        return cls(
+            nomad_address=str(getattr(cfg, "nomad_address")),
+            sandbox_job_id=str(getattr(cfg, "sandbox_job_id")),
+            sandbox_image=str(getattr(cfg, "sandbox_image")),
+            slots_per_container=int(getattr(cfg, "slots_per_container")),
+            min_containers=int(getattr(cfg, "min_containers")),
+            max_containers=int(getattr(cfg, "max_containers")),
+            privileged=bool(getattr(cfg, "privileged")),
+            acquire_timeout_s=float(getattr(cfg, "acquire_timeout_s")),
+            purge_job_on_start=bool(getattr(cfg, "purge_job_on_start", False)),
+            driver=str(getattr(cfg, "driver", "docker")),
+            singularity_image=getattr(cfg, "singularity_image", None),
+        )
+
+
+class NomadToolBackend(ToolBackend):
+    def __init__(self, config: NomadBackendConfig):
+        self.config = config
+        self.pool = SlotPool(
+            SlotPoolConfig(
+                nomad_address=config.nomad_address,
+                job_id=config.sandbox_job_id,
+                image=config.sandbox_image,
+                slots_per_container=config.slots_per_container,
+                min_containers=config.min_containers,
+                max_containers=config.max_containers,
+                privileged=config.privileged,
+                acquire_timeout=config.acquire_timeout_s,
+                purge_job_on_start=bool(config.purge_job_on_start),
+                driver=config.driver,
+                singularity_image=config.singularity_image,
+            )
+        )
+
+    @property
+    def default_timeout_s(self) -> Optional[float]:
+        t = getattr(self.pool.executor, "timeout", None)
+        total = getattr(t, "total", None)
+        try:
+            return float(total) if total is not None else None
+        except Exception:
+            return None
+
+    async def start(self) -> None:
+        await self.pool.start()
+
+    async def stop(self, *, purge: bool = False) -> None:
+        await self.pool.stop(purge_job=purge)
+
+    async def acquire(self, trajectory_id: Optional[str] = None) -> Slot:
+        return await self.pool.acquire(trajectory_id)
+
+    async def release(self, slot: Slot, *, reset_workspace: bool = False) -> None:
+        await self.pool.release(slot, reset_workspace=reset_workspace)
+
+    async def execute_batch(
+        self,
+        requests: List[Tuple[Slot, str, Dict[str, Any]]],
+        *,
+        timeout_s: Optional[float] = None,
+    ) -> List[ExecutionResult]:
+        return await self.pool.execute_batch(requests, timeout=timeout_s)
+
+    async def read_artifact(
+        self,
+        slot: Slot,
+        path: str,
+        *,
+        encoding: str = "text",
+        max_bytes: Optional[int] = None,
+        include_sha256: bool = False,
+        timeout_s: Optional[float] = None,
+    ) -> Dict[str, Any]:
+        return await self.pool.executor.read_artifact(
+            slot,
+            path,
+            encoding=encoding,
+            max_bytes=max_bytes,
+            include_sha256=include_sha256,
+            timeout=timeout_s,
+        )
+
+    async def list_artifacts(
+        self,
+        slot: Slot,
+        path: str = ".",
+        *,
+        recursive: bool = False,
+        max_entries: Optional[int] = None,
+        timeout_s: Optional[float] = None,
+    ) -> Dict[str, Any]:
+        return await self.pool.executor.list_artifacts(
+            slot,
+            path,
+            recursive=recursive,
+            max_entries=max_entries,
+            timeout=timeout_s,
+        )
+
+    async def archive_artifacts(
+        self,
+        slot: Slot,
+        path: str = ".",
+        *,
+        archive_format: str = "tar.gz",
+        max_bytes: Optional[int] = None,
+        max_entries: Optional[int] = None,
+        timeout_s: Optional[float] = None,
+    ) -> Dict[str, Any]:
+        return await self.pool.executor.archive_artifacts(
+            slot,
+            path,
+            archive_format=archive_format,
+            max_bytes=max_bytes,
+            max_entries=max_entries,
+            timeout=timeout_s,
+        )
+
+    def get_stats(self) -> Dict[str, Any]:
+        return self.pool.get_stats()
+

atropos/envs/__init__.py

@@ -0,0 +1,18 @@
+"""
+Environment implementations for atropos-agent.
+
+NOTE: AgentEnv is the OLD environment system, replaced by
+environments/hermes_base_env.py (HermesAgentBaseEnv).
+Import is lazy to avoid pulling in deleted dependencies.
+"""
+
+
+def __getattr__(name):
+    """Lazy import to avoid breaking when old dependencies are removed."""
+    if name in ("AgentEnv", "AgentEnvConfig"):
+        from .agent_env import AgentEnv, AgentEnvConfig
+        return {"AgentEnv": AgentEnv, "AgentEnvConfig": AgentEnvConfig}[name]
+    raise AttributeError(f"module {__name__!r} has no attribute {name!r}")
+
+
+__all__ = ["AgentEnv", "AgentEnvConfig"]

atropos/envs/agent_env.py

@@ -0,0 +1,537 @@
+"""
+AgentEnv - Atropos BaseEnv extension for agent/tool-call workloads.
+
+AgentEnv is responsible for starting the sandbox tool execution backend and
+providing helpers for running agent trajectories with queued/batched tool calls.
+"""
+
+from __future__ import annotations
+import os
+import asyncio
+import time
+import uuid
+from abc import ABC, abstractmethod
+from typing import Any, Awaitable, Callable, Dict, Generic, List, Optional, Tuple, TypeVar
+
+from pydantic import Field
+
+from atroposlib.envs.base import APIServerConfig, BaseEnv, BaseEnvConfig, Item, ScoredDataGroup, ScoredDataItem
+from atroposlib.envs.server_handling.server_baseline import AsyncSemWithAdaptiveWeight
+
+from ..agent import AgentConfig, AgentResult, AtroposAgent
+from ..backends import ToolBackend, create_tool_backend
+from ..tools import ToolRegistry, build_tool_registry
+from ..tools.tool_executor import ToolExecutor, ToolExecutorConfig
+
+# Main BaseEnv child classes. Child class THESE to get agent+tooling functionality easily.
+
+class AgentEnvConfig(BaseEnvConfig):
+    tool_pool_mode: str = Field(default="nomad", description="Tool execution backend ('nomad' or 'modal')")
+
+    allow_network: bool = Field(
+        default=True,
+        description="Whether sandbox bash commands may access the network (env policy).",
+    )
+    require_sandbox: bool = Field(
+        default=False,
+        description="Fail closed if bubblewrap sandboxing is unavailable/unusable for stateless sandbox tools.",
+    )
+    require_stateful_sandbox: bool = Field(
+        default=False,
+        description="Fail closed if bubblewrap/PID isolation is unavailable for stateful terminal tools (tmux).",
+    )
+    tool_batch_window_ms: int = Field(default=20, description="ToolExecutor batching window (ms)")
+    tool_max_batch_size: int = Field(default=200, description="ToolExecutor maximum batch size")
+
+    # nomad mode settings. TODO: Add Modal support, split this into own config
+    nomad_address: str = Field(default="http://localhost:4646", description="Nomad API address")
+    sandbox_job_id: str = Field(default="atropos-sandbox-agent-env", description="Nomad job id for sandbox containers")
+    sandbox_image: str = Field(default="atropos-sandbox:local", description="Docker image for sandbox containers")
+    slots_per_container: int = Field(default=10, description="Nomad mode: slots per container")
+    min_containers: int = Field(default=1, description="Nomad mode: minimum containers")
+    max_containers: int = Field(default=10, description="Nomad mode: maximum containers")
+    privileged: bool = Field(default=False, description="Nomad mode: run container privileged")
+    acquire_timeout_s: float = Field(default=30.0, description="Slot acquisition timeout (seconds)")
+    purge_job_on_start: bool = Field(
+        default=False,
+        description=(
+            "Nomad mode: stop/purge the sandbox job on startup. This is helpful in local dev and training runs "
+            "to recover from previous crashes that leave the job in a restart backoff state."
+        ),
+    )
+    purge_job_on_shutdown: bool = Field(default=True, description="Nomad mode: stop/purge job on shutdown")
+    
+    # Nomad driver selection (docker or singularity)
+    driver: str = Field(
+        default="docker",
+        description="Nomad task driver: 'docker' (default) or 'singularity' (for HPC without sudo Docker)",
+    )
+    singularity_image: Optional[str] = Field(
+        default=None,
+        description="Path to .sif file for Singularity driver (required if driver='singularity')",
+    )
+
+    # Modal mode settings
+    modal_app_name: str = Field(default="atropos-sandbox", description="Modal app name prefix")
+    modal_image: str = Field(default="python:3.11", description="Modal: container image")
+    modal_gpu: Optional[str] = Field(default=None, description="Modal: GPU type (None, 'T4', 'A10G', 'A100', 'H100')")
+    modal_cpu: float = Field(default=1.0, description="Modal: CPU cores")
+    modal_memory: int = Field(default=2048, description="Modal: memory in MB")
+    modal_slots_per_sandbox: int = Field(default=10, description="Modal: slots per sandbox")
+    modal_min_sandboxes: int = Field(default=1, description="Modal: minimum sandboxes")
+    modal_max_sandboxes: int = Field(default=5, description="Modal: maximum sandboxes")
+    modal_idle_timeout: int = Field(default=120, description="Modal: server-side idle timeout (seconds)")
+    modal_max_lifetime: int = Field(default=3600, description="Modal: max sandbox lifetime (seconds)")
+    modal_acquire_timeout: float = Field(default=60.0, description="Modal: slot acquisition timeout (seconds)")
+    modal_execution_timeout: float = Field(default=30.0, description="Modal: default command execution timeout (seconds)")
+    modal_secrets: str = Field(default="", description="Modal: comma-separated list of Modal Secret names")
+    modal_env_vars: str = Field(default="", description="Modal: semicolon-separated KEY=VALUE pairs for env vars")
+    modal_workspace_base: str = Field(default="/data", description="Modal: workspace base directory in sandbox")
+
+    # basic agent defaults
+    agent_max_steps: int = Field(default=50, description="Max ReACT steps per trajectory")
+    agent_temperature: float = Field(default=0.7, description="Sampling temperature")
+    agent_max_tokens: Optional[int] = Field(
+        default=None,
+        description="Max tokens per model response (default: let backend decide)",
+    )
+    agent_tool_delay_s: float = Field(default=0.0, description="Delay between tool calls (seconds)")
+
+    # tool selection
+    enabled_toolsets: List[str] = Field(
+        default_factory=lambda: ["default"],
+        description="Toolsets to enable (Hermes-style grouping).",
+    )
+    disabled_toolsets: List[str] = Field(
+        default_factory=list,
+        description="Toolsets to disable (applied after enabled_toolsets).",
+    )
+
+    # external ToolServer routing (Phase 4.5+)
+    tool_server_url: Optional[str] = Field(
+        default=None,
+        description="Base URL for external ToolServer (enables external tools).",
+    )
+    tool_server_token: Optional[str] = Field(
+        default=None,
+        description="Bearer token for ToolServer auth (optional in dev).",
+    )
+
+AgentEnvConfigT = TypeVar("AgentEnvConfigT", bound="AgentEnvConfig")
+
+
+class AgentEnv(BaseEnv, ABC, Generic[AgentEnvConfigT]):
+    env_config_cls = AgentEnvConfig
+
+    def __init__(
+        self,
+        config: AgentEnvConfigT,
+        server_configs: List[APIServerConfig],
+        slurm: bool = False,
+        testing: bool = False,
+    ):
+        super().__init__(config, server_configs, slurm, testing)
+        self.config: AgentEnvConfigT = config
+
+        self.tools: ToolRegistry = self.build_tools()
+
+        self._backend: Optional[ToolBackend] = None
+        self._tool_executor: Optional[ToolExecutor] = None
+        self._tool_server_inprocess: bool = False
+        self._trajectory_workspace_meta: Dict[str, Dict[str, Any]] = {}
+
+    def build_tools(self) -> ToolRegistry:
+        """Wraps original Hermes-Agent ToolRegistry for atropos AgentEnv use.
+        See Hermes-Agent docs for toolsets and available tools etc.
+        """
+        return build_tool_registry(
+            enabled_toolsets=self.config.enabled_toolsets or ["default"],
+            disabled_toolsets=self.config.disabled_toolsets or None,
+            tool_server_url=self.config.tool_server_url,
+        )
+
+    @abstractmethod
+    def build_task(self, item: Item) -> str:
+        """Return the user-facing task string for the agent."""
+
+    @abstractmethod
+    async def score_trajectory(self, item: Item, final_response: str) -> float:
+        """Return a scalar score for this trajectory."""
+
+    async def setup_trajectory_workspace(
+        self,
+        item: Item,
+        *,
+        trajectory_id: str,
+        exec_tool: Callable[["ToolCall"], Awaitable["ToolResult"]],
+    ) -> Dict[str, Any]:
+        """
+        Optional hook: prepare the sandbox workspace before the agent starts.
+
+        Examples:
+        - clone a repo and checkout a commit
+        - write fixture files (e.g. images) for external-tool demos
+        - pre-install dependencies
+
+        Default: no-op.
+        """
+        _ = (item, trajectory_id, exec_tool)
+        return {}
+
+    async def verify_and_score_trajectory(
+        self,
+        item: Item,
+        final_response: str,
+        *,
+        trajectory_id: str,
+        exec_tool: Callable[["ToolCall"], Awaitable["ToolResult"]],
+        agent_result: Optional[AgentResult] = None,
+        workspace_meta: Optional[Dict[str, Any]] = None,
+    ) -> tuple[float, Dict[str, Any]]:
+        """
+        Optional hook: run in-sandbox verification before scoring.
+
+        Many agent envs need to execute verification inside the same trajectory
+        workspace (e.g. pytest) before releasing/resetting the slot.
+
+        Default: calls `score_trajectory()` and returns empty metadata.
+        """
+        _ = (trajectory_id, exec_tool, agent_result, workspace_meta)  # default ignores in-workspace verification
+        score = await self.score_trajectory(item, final_response)
+        return score, {}
+
+    def build_agent_config(self, item: Item) -> AgentConfig:  # noqa: ARG002
+        return AgentConfig(
+            max_steps=self.config.agent_max_steps,
+            temperature=self.config.agent_temperature,
+            max_tokens=self.config.agent_max_tokens,
+            tool_delay_s=self.config.agent_tool_delay_s,
+        )
+
+    async def setup(self) -> None:
+        print(f"[AgentEnv] setup(): starting tool backend ({self.config.tool_pool_mode})", flush=True)
+        await self._start_tool_backend()
+        print("[AgentEnv] setup(): configuring server concurrency", flush=True)
+        self._configure_server_concurrency()
+        print("[AgentEnv] setup(): running env-specific setup_agent_env()", flush=True)
+        await self.setup_agent_env()
+        print("[AgentEnv] setup(): done", flush=True)
+
+    def _configure_server_concurrency(self) -> None:
+        """
+        Ensure the LLM server concurrency isn't accidentally capped below `group_size`.
+
+        In `BaseEnv process` mode, groups are collected concurrently and if the underlying
+        ServerManager/OpenAIServer semaphore is left at 1, we serialize inference even
+        when `--env.group_size` is > 1.
+        """
+        desired = int(getattr(self.config, "group_size", 1) or 1)
+        if desired <= 1:
+            return
+
+        servers = getattr(self.server, "servers", None)
+        if not isinstance(servers, list) or not servers:
+            return
+
+        for s in servers:
+            sem = getattr(s, "sem", None)
+            eval_sem = getattr(s, "eval_sem", None)
+            # Only increase; never shrink.
+            if sem is not None and getattr(sem, "max_val", 0) < desired:
+                s.sem = AsyncSemWithAdaptiveWeight(desired)
+                if hasattr(s, "config") and hasattr(s.config, "num_max_requests_at_once"):
+                    s.config.num_max_requests_at_once = desired
+            if eval_sem is not None and getattr(eval_sem, "max_val", 0) < desired:
+                s.eval_sem = AsyncSemWithAdaptiveWeight(desired)
+                if hasattr(s, "config") and hasattr(s.config, "num_requests_for_eval"):
+                    s.config.num_requests_for_eval = desired
+
+    @abstractmethod
+    async def setup_agent_env(self) -> None:
+        """Subclass hook for env-specific setup."""
+
+    async def evaluate(self, *args, **kwargs):  # noqa: ARG002
+        """
+        Default eval hook (no-op).
+
+        Atropos BaseEnv requires an `evaluate()` implementation. Many agent envs
+        won't have a meaningful evaluation path during early PoC work; they can
+        override this when needed.
+        """
+        return {}
+
+    async def env_manager(self):
+        try:
+            return await super().env_manager()
+        finally:
+            await self.shutdown_tool_backend()
+
+    async def process_manager(self):
+        try:
+            return await super().process_manager()
+        finally:
+            await self.shutdown_tool_backend()
+
+    async def _start_tool_backend(self) -> None:
+        if self._tool_executor is not None:
+            return
+
+        tool_server_url = self.config.tool_server_url
+        tool_server_client = None
+        if tool_server_url == "inprocess":
+            import httpx
+            from ..api.tool_server import app as tool_server_app
+
+            await tool_server_app.router.startup()
+            tool_server_client = httpx.AsyncClient(
+                transport=httpx.ASGITransport(app=tool_server_app),
+                base_url="http://toolserver",
+            )
+            tool_server_url = "http://toolserver"
+            self._tool_server_inprocess = True
+
+        backend = create_tool_backend(self.config)
+        await backend.start()
+
+        executor = ToolExecutor(
+            backend=backend,
+            tools=self.tools,
+            config=ToolExecutorConfig(
+                batch_window_ms=self.config.tool_batch_window_ms,
+                max_batch_size=self.config.tool_max_batch_size,
+                allow_network=self.config.allow_network,
+                require_sandbox=self.config.require_sandbox,
+                require_stateful_sandbox=self.config.require_stateful_sandbox,
+                tool_server_url=tool_server_url,
+                tool_server_token=self.config.tool_server_token,
+            ),
+        )
+        await executor.start()
+        if tool_server_client is not None:
+            executor._tool_server_client = tool_server_client  # type: ignore[attr-defined]
+
+        self._backend = backend
+        self._tool_executor = executor
+
+    async def shutdown_tool_backend(self) -> None:
+        executor = self._tool_executor
+        backend = self._backend
+        inprocess_tool_server = self._tool_server_inprocess
+        self._tool_executor = None
+        self._backend = None
+        self._tool_server_inprocess = False
+
+        if executor is not None:
+            await executor.close()
+        if backend is not None:
+            await backend.stop(purge=bool(self.config.purge_job_on_shutdown))
+        if inprocess_tool_server:
+            from ..api.tool_server import app as tool_server_app
+
+            await tool_server_app.router.shutdown()
+
+    async def collect_trajectory(
+        self, item: Item
+    ) -> Tuple[Optional[ScoredDataItem], List[Item]]:
+        if self._tool_executor is None:
+            raise RuntimeError("Tool backend not started")
+
+        trajectory_id = str(uuid.uuid4())
+        t0 = time.perf_counter()
+        print(f"[AgentEnv] collect_trajectory(): tid={trajectory_id} start", flush=True)
+        task = self.build_task(item)
+        agent_config = self.build_agent_config(item)
+        if os.getenv("ATROPOS_DEBUG_PRINT_TASK") == "1":
+            print(f"Starting trajectory {trajectory_id} with task: {task}", flush=True)
+        else:
+            # Avoid printing the full task prompt by default (can be huge/noisy).
+            one_line = " ".join(str(task).splitlines()).strip()
+            preview = one_line[:240] + ("…" if len(one_line) > 240 else "")
+            print(f"Starting trajectory {trajectory_id} (task preview): {preview}", flush=True)
+
+        async def _exec(call):
+            return await self._tool_executor.execute(trajectory_id, call)
+
+        agent = AtroposAgent(
+            server=self.server,
+            tokenizer=self.tokenizer,
+            tools=self.tools,
+            config=agent_config,
+            execute_tool=_exec,
+        )
+
+        try:
+            print(f"[AgentEnv] tid={trajectory_id} setup_trajectory_workspace() start", flush=True)
+            workspace_meta = await self.setup_trajectory_workspace(item, trajectory_id=trajectory_id, exec_tool=_exec)
+            if not isinstance(workspace_meta, dict):
+                workspace_meta = {}
+            self._trajectory_workspace_meta[trajectory_id] = workspace_meta
+            print(
+                f"[AgentEnv] tid={trajectory_id} setup_trajectory_workspace() done in {time.perf_counter() - t0:.2f}s",
+                flush=True,
+            )
+
+            print(f"[AgentEnv] tid={trajectory_id} agent.run() start", flush=True)
+            result = await agent.run(task)
+            print(
+                f"[AgentEnv] tid={trajectory_id} agent.run() done in {time.perf_counter() - t0:.2f}s "
+                f"success={result.success} tool_calls={result.total_tool_calls}",
+                flush=True,
+            )
+            if not result.success or result.trajectory_data is None:
+                # Do not trigger BaseEnv retries for agent failures.
+                # Record the trajectory with score 0.0 so training/eval can see the failure mode.
+                messages = [{"role": "system", "content": agent._build_system_prompt()}]  # noqa: SLF001
+                messages.append({"role": "user", "content": task})
+                for step in result.steps:
+                    messages.append({"role": "assistant", "content": step.assistant_message})
+                    if step.tool_results:
+                        tool_text = "\n".join(r.to_xml() for r in step.tool_results)
+                        messages.append({"role": "user", "content": tool_text})
+
+                scored: ScoredDataItem = {
+                    "tokens": (result.trajectory_data.tokens if result.trajectory_data else []),
+                    "masks": (result.trajectory_data.masked_tokens if result.trajectory_data else []),
+                    "scores": 0.0,
+                }
+                if result.trajectory_data is not None:
+                    scored["inference_logprobs"] = result.trajectory_data.logprobs  # type: ignore[typeddict-unknown-key]
+                    if getattr(result.trajectory_data, "metadata", None):
+                        scored["overrides"] = {"managed_metadata": result.trajectory_data.metadata}
+                if self.config.include_messages:
+                    # Record a final failure marker as a user-side tool_response-like block so it survives templates.
+                    import json
+
+                    err = result.error or "agent_failed"
+                    messages.append(
+                        {
+                            "role": "user",
+                            "content": f"<tool_response>{json.dumps({'success': False, 'error': err})}</tool_response>",
+                        }
+                    )
+                    scored["messages"] = messages
+                return scored, []
+
+            print(f"[AgentEnv] tid={trajectory_id} verify_and_score_trajectory() start", flush=True)
+            score, score_metadata = await self.verify_and_score_trajectory(
+                item,
+                result.final_response,
+                trajectory_id=trajectory_id,
+                exec_tool=_exec,
+                agent_result=result,
+                workspace_meta=workspace_meta,
+            )
+            print(
+                f"[AgentEnv] tid={trajectory_id} verify_and_score_trajectory() done in {time.perf_counter() - t0:.2f}s "
+                f"score={score}",
+                flush=True,
+            )
+
+            messages = [{"role": "system", "content": agent._build_system_prompt()}]  # noqa: SLF001
+            messages.append({"role": "user", "content": task})
+            for step in result.steps:
+                messages.append({"role": "assistant", "content": step.assistant_message})
+                if step.tool_results:
+                    tool_text = "\n".join(r.to_xml() for r in step.tool_results)
+                    messages.append({"role": "user", "content": tool_text})
+
+            # Optional: allow env verification to attach additional messages (e.g. install logs).
+            if self.config.include_messages and isinstance(score_metadata, dict):
+                extra = score_metadata.get("verification_messages")
+                if isinstance(extra, list):
+                    for m in extra:
+                        if isinstance(m, dict) and isinstance(m.get("role"), str) and isinstance(m.get("content"), str):
+                            messages.append({"role": m["role"], "content": m["content"]})
+
+            scored: ScoredDataItem = {
+                "tokens": result.trajectory_data.tokens,
+                "masks": result.trajectory_data.masked_tokens,
+                "scores": score,
+            }
+            # Atroposlib expects policy logprobs at the *group* level under `inference_logprobs`.
+            # We stash per-item values here and lift them into the group in `collect_trajectories()`.
+            scored["inference_logprobs"] = result.trajectory_data.logprobs  # type: ignore[typeddict-unknown-key]
+            if getattr(result.trajectory_data, "metadata", None):
+                scored["overrides"] = {"managed_metadata": result.trajectory_data.metadata}
+            if self.config.include_messages:
+                scored["messages"] = messages
+
+            return scored, []
+        finally:
+            self._trajectory_workspace_meta.pop(trajectory_id, None)
+            print(f"[AgentEnv] tid={trajectory_id} release_trajectory(reset_workspace=True)", flush=True)
+            await self._tool_executor.release_trajectory(trajectory_id, reset_workspace=True)
+            print(f"[AgentEnv] collect_trajectory(): tid={trajectory_id} done in {time.perf_counter() - t0:.2f}s", flush=True)
+
+    async def collect_trajectories(
+        self, item: Item
+    ) -> Tuple[Optional[ScoredDataGroup], List[Item]]:
+        tasks = [self.collect_trajectory(item) for _ in range(self.config.group_size)]
+        results = await asyncio.gather(*tasks)
+
+        backlog: List[Item] = []
+        items: List[ScoredDataItem] = []
+        for scored, b in results:
+            backlog.extend(b)
+            if scored is not None:
+                items.append(scored)
+
+        if len(items) != self.config.group_size:
+            return None, backlog
+
+        group: ScoredDataGroup = ScoredDataGroup(
+            tokens=[],
+            masks=[],
+            scores=[],
+            advantages=[],
+            ref_logprobs=[],
+            messages=[] if self.config.include_messages else None,
+            inference_logprobs=[],
+            group_overrides={},
+            overrides=[],
+            images=[],
+            generation_params=None,
+        )
+
+        for it in items:
+            group["tokens"].append(it["tokens"])
+            group["masks"].append(it["masks"])
+            group["scores"].append(it["scores"])
+            # policy logprobs (for PPO/GRPO training) if present
+            lp = it.get("inference_logprobs")  # type: ignore[typeddict-item]
+            if lp is not None:
+                group["inference_logprobs"].append(lp)
+            group["overrides"].append(it.get("overrides") or {})  # type: ignore[typeddict-item]
+            if group.get("messages") is not None and it.get("messages") is not None:
+                group["messages"].append(it["messages"])
+
+        return group, backlog
+
+    async def run_agent(self, task: str, *, trajectory_id: Optional[str] = None) -> Tuple[str, Dict[str, Any]]:
+        """
+        Run the AtroposAgent on a single task and return (final_response, debug).
+
+        This is a helper intended for simple environments and tests.
+        """
+        if self._tool_executor is None:
+            raise RuntimeError("Tool backend not started")
+
+        tid = trajectory_id or str(uuid.uuid4())
+
+        async def _exec(call):
+            return await self._tool_executor.execute(tid, call)
+
+        agent = AtroposAgent(
+            server=self.server,
+            tokenizer=self.tokenizer,
+            tools=self.tools,
+            config=AgentConfig(
+                max_steps=self.config.agent_max_steps,
+                temperature=self.config.agent_temperature,
+                max_tokens=self.config.agent_max_tokens,
+            ),
+            execute_tool=_exec,
+        )
+        result = await agent.run(task)
+        await self._tool_executor.release_trajectory(tid, reset_workspace=True)
+        return result.final_response, {"success": result.success, "error": result.error, "tool_calls": result.total_tool_calls}

atropos/envs/hermes_compat_test_env.py

@@ -0,0 +1,171 @@
+"""
+Hermes-Agent + Atropos (Nomad sandbox) compatibility smoke environment.
+
+This environment is intended to validate, end-to-end:
+  BaseEnv.process -> AgentEnv -> ToolExecutor (batched) -> Nomad SlotPool -> sandbox_server
+
+It forces the model to use a sandbox tool by asking it to run a command that
+generates a high-entropy token inside the sandbox, then repeat it exactly.
+
+Run (process mode):
+  uv run python -m atropos.envs.hermes_compat_test_env process --env.use_wandb false --env.total_steps 2 --env.group_size 1
+"""
+
+from __future__ import annotations
+
+import os
+from typing import Any, Dict, List, Tuple
+
+from dotenv import load_dotenv
+from pydantic import Field
+
+from atroposlib.envs.base import APIServerConfig, Item
+
+from ..agent import AgentConfig, AgentResult
+from ..tools import ToolCall
+from .agent_env import AgentEnv, AgentEnvConfig
+
+load_dotenv()
+
+
+def _forced_tool_item() -> Item:
+    # Use double quotes in the shell command and show JSON escaping explicitly.
+    # This avoids invalid JSON escapes like `\\'` (not valid JSON) that some models produce.
+    cmd = 'python -c "import secrets; print(secrets.token_hex(16))"'
+    return {
+        "command": cmd,
+        "prompt": (
+            "You are acting as an agent inside a sandboxed environment.\n"
+            "You MUST use the terminal tool to execute commands.\n"
+            "Run this exact command:\n"
+            f"{cmd}\n"
+            "When you call the tool, use valid JSON inside <tool_call>. Example:\n"
+            '<tool_call>{"name": "terminal", "arguments": {"command": '
+            '"python -c \\\\"import secrets; print(secrets.token_hex(16))\\\\""}}'
+            "</tool_call>\n"
+            "Then respond with EXACTLY what it printed (the hex token) and nothing else.\n"
+            "Do not guess. Do not explain."
+        ),
+    }
+
+
+class HermesCompatTestEnvConfig(AgentEnvConfig):
+    server_base_url: str = Field(
+        default="http://127.0.0.1:8080",
+        description="Base URL for an OpenAI-compatible chat server (without /v1).",
+    )
+    server_model: str = Field(default="hermes-4-36b", description="Model name")
+    tokenizer_name: str = Field(default="NousResearch/Hermes-4.3-36B", description="Tokenizer name for RL tokenization")
+
+
+class HermesCompatTestEnv(AgentEnv[HermesCompatTestEnvConfig]):
+    name = "hermes_compat_test_env"
+    env_config_cls = HermesCompatTestEnvConfig
+
+    def __init__(
+        self,
+        config: HermesCompatTestEnvConfig,
+        server_configs: List[APIServerConfig],
+        slurm: bool = False,
+        testing: bool = False,
+    ):
+        super().__init__(config, server_configs, slurm, testing)
+        self._iter = 0
+
+    @classmethod
+    def config_init(cls) -> Tuple[HermesCompatTestEnvConfig, List[APIServerConfig]]:
+        base_url = (
+            os.getenv("ATROPOS_SERVER_BASE_URL")
+            or os.getenv("OPENAI_BASE_URL")
+            or os.getenv("LLM_BASE_URL")
+            or "http://127.0.0.1:8080"
+        )
+        model = os.getenv("ATROPOS_SERVER_MODEL") or os.getenv("LLM_MODEL") or "hermes-4-36b"
+        api_key = os.getenv("ATROPOS_SERVER_API_KEY") or os.getenv("NOUS_API_KEY") or os.getenv("OPENAI_API_KEY") or "local"
+
+        env_config = HermesCompatTestEnvConfig(
+            tokenizer_name=os.getenv("ATROPOS_TOKENIZER_NAME") or "NousResearch/Hermes-4.3-36B",
+            group_size=1,
+            use_wandb=False,
+            include_messages=True,
+            ensure_scores_are_not_same=False,
+            total_steps=2,
+            batch_size=1,
+            server_base_url=base_url,
+            server_model=model,
+            # Tooling: sandbox-only terminal.
+            enabled_toolsets=["terminal"],
+            disabled_toolsets=[],
+            # Default to Nomad sandboxing; users can override via --env.* args.
+            sandbox_image=os.getenv("ATROPOS_SANDBOX_IMAGE") or "atropos-sandbox:local",
+            # In local dev it's common for a previous crash to leave the job in backoff.
+            purge_job_on_start=True,
+            purge_job_on_shutdown=True,
+        )
+
+        server_configs = [
+            APIServerConfig(
+                model_name=model,
+                base_url=f"{base_url.rstrip('/')}/v1",
+                api_key=api_key,
+                num_max_requests_at_once=1,
+                num_requests_for_eval=1,
+                timeout=120,
+            )
+        ]
+        return env_config, server_configs
+
+    async def setup_agent_env(self) -> None:
+        return None
+
+    async def get_next_item(self) -> Item:
+        self._iter += 1
+        return _forced_tool_item()
+
+    def build_task(self, item: Item) -> str:
+        return str(item.get("prompt") or "")
+
+    def build_agent_config(self, item: Item) -> AgentConfig:  # noqa: ARG002
+        # Avoid imposing max_tokens by default; tool-tag responses can be long for some models.
+        return AgentConfig(
+            max_steps=min(8, int(self.config.agent_max_steps)),
+            temperature=0.2,
+            max_tokens=None,
+        )
+
+    async def score_trajectory(self, item: Item, final_response: str) -> float:
+        # Scoring happens in verify_and_score_trajectory so we can inspect tool results.
+        _ = (item, final_response)
+        return 0.0
+
+    async def verify_and_score_trajectory(
+        self,
+        item: Item,
+        final_response: str,
+        *,
+        trajectory_id: str,  # noqa: ARG002
+        exec_tool,  # noqa: ARG002
+        agent_result: AgentResult | None = None,
+        workspace_meta: Dict[str, Any] | None = None,  # noqa: ARG002
+    ) -> tuple[float, Dict[str, Any]]:
+        if agent_result is None:
+            return 0.0, {"error": "Missing agent_result"}
+
+        observed: str = ""
+        tool_ok = False
+        for step in agent_result.steps:
+            for res in step.tool_results:
+                if not res.success:
+                    return 0.0, {"error": res.error, "output": res.output}
+                out = (res.output or "").strip()
+                if out:
+                    observed = out.splitlines()[-1].strip()
+                    tool_ok = True
+
+        final = (final_response or "").strip()
+        score = 1.0 if tool_ok and agent_result.total_tool_calls > 0 and observed and final == observed else 0.0
+        return score, {"observed": observed, "tool_calls": agent_result.total_tool_calls, "command": item.get("command")}
+
+
+if __name__ == "__main__":
+    HermesCompatTestEnv.cli()

atropos/envs/sandbox_terminal_smoke_env.py

@@ -0,0 +1,172 @@
+"""
+Nomad sandbox terminal smoke environment (training-oriented).
+
+Validates, end-to-end:
+  BaseEnv.process -> AgentEnv -> ToolExecutor (batched) -> Nomad SlotPool -> sandbox_server
+
+It forces the model to use a sandbox tool by asking it to run a command that
+generates a high-entropy token inside the sandbox, then repeat it exactly.
+
+Run (process mode):
+  uv run python -m atropos.envs.sandbox_terminal_smoke_env process --env.use_wandb false --env.total_steps 2 --env.group_size 1
+"""
+
+from __future__ import annotations
+
+import os
+from typing import Any, Dict, List, Tuple
+
+from dotenv import load_dotenv
+from pydantic import Field
+
+from atroposlib.envs.base import APIServerConfig, Item
+
+from ..agent import AgentConfig, AgentResult
+from ..tools import ToolCall
+from .agent_env import AgentEnv, AgentEnvConfig
+
+load_dotenv()
+
+STRICT_TOOLCALL_SYSTEM_PROMPT = None
+
+
+def _forced_tool_item() -> Item:
+    # Use double quotes in the shell command and show JSON escaping explicitly.
+    # This avoids invalid JSON escapes like `\\'` (not valid JSON) that some models produce.
+    cmd = 'python -c "import secrets; print(secrets.token_hex(16))"'
+    return {
+        "command": cmd,
+        "prompt": (
+            "You MUST use the terminal tool.\n"
+            "Run this exact command:\n"
+            f"{cmd}\n"
+            "When you call the tool, use valid JSON inside <tool_call>. Example:\n"
+            '<tool_call>{"name": "terminal", "arguments": {"command": '
+            '"python -c \\\\"import secrets; print(secrets.token_hex(16))\\\\""}}'
+            "</tool_call>\n"
+            "Then respond with EXACTLY what it printed (the hex token) and nothing else.\n"
+            "Do not guess. Do not explain."
+        ),
+    }
+
+
+class SandboxTerminalSmokeEnvConfig(AgentEnvConfig):
+    server_base_url: str = Field(
+        default="http://127.0.0.1:8080",
+        description="Base URL for an OpenAI-compatible chat server (without /v1).",
+    )
+    server_model: str = Field(default="hermes-4-36b", description="Model name")
+    tokenizer_name: str = Field(default="NousResearch/Hermes-4.3-36B", description="Tokenizer name for RL tokenization")
+
+
+class SandboxTerminalSmokeEnv(AgentEnv[SandboxTerminalSmokeEnvConfig]):
+    name = "sandbox_terminal_smoke_env"
+    env_config_cls = SandboxTerminalSmokeEnvConfig
+
+    def __init__(
+        self,
+        config: SandboxTerminalSmokeEnvConfig,
+        server_configs: List[APIServerConfig],
+        slurm: bool = False,
+        testing: bool = False,
+    ):
+        super().__init__(config, server_configs, slurm, testing)
+        self._iter = 0
+
+    @classmethod
+    def config_init(cls) -> Tuple[SandboxTerminalSmokeEnvConfig, List[APIServerConfig]]:
+        base_url = (
+            os.getenv("ATROPOS_SERVER_BASE_URL")
+            or os.getenv("OPENAI_BASE_URL")
+            or os.getenv("LLM_BASE_URL")
+            or "http://127.0.0.1:8080"
+        )
+        model = os.getenv("ATROPOS_SERVER_MODEL") or os.getenv("LLM_MODEL") or "hermes-4-36b"
+        api_key = os.getenv("ATROPOS_SERVER_API_KEY") or os.getenv("NOUS_API_KEY") or os.getenv("OPENAI_API_KEY") or "local"
+
+        env_config = SandboxTerminalSmokeEnvConfig(
+            tokenizer_name=os.getenv("ATROPOS_TOKENIZER_NAME") or "NousResearch/Hermes-4.3-36B",
+            group_size=1,
+            use_wandb=False,
+            include_messages=True,
+            ensure_scores_are_not_same=False,
+            total_steps=2,
+            batch_size=1,
+            server_base_url=base_url,
+            server_model=model,
+            # Tooling: sandbox-only terminal.
+            enabled_toolsets=["terminal"],
+            disabled_toolsets=[],
+            # Default to Nomad sandboxing; users can override via --env.* args.
+            sandbox_image=os.getenv("ATROPOS_SANDBOX_IMAGE") or "atropos-sandbox:local",
+            purge_job_on_start=True,
+            purge_job_on_shutdown=True,
+        )
+
+        server_configs = [
+            APIServerConfig(
+                model_name=model,
+                base_url=f"{base_url.rstrip('/')}/v1",
+                api_key=api_key,
+                num_max_requests_at_once=1,
+                num_requests_for_eval=1,
+                timeout=120,
+            )
+        ]
+        return env_config, server_configs
+
+    async def setup_agent_env(self) -> None:
+        return None
+
+    async def get_next_item(self) -> Item:
+        self._iter += 1
+        return _forced_tool_item()
+
+    def build_task(self, item: Item) -> str:
+        return str(item.get("prompt") or "")
+
+    def build_agent_config(self, item: Item) -> AgentConfig:  # noqa: ARG002
+        # Avoid imposing max_tokens by default; tool-tag responses can be long for some models.
+        return AgentConfig(
+            max_steps=min(8, int(self.config.agent_max_steps)),
+            temperature=0.2,
+            max_tokens=None,
+            system_prompt=STRICT_TOOLCALL_SYSTEM_PROMPT,
+        )
+
+    async def score_trajectory(self, item: Item, final_response: str) -> float:
+        # Scoring happens in verify_and_score_trajectory so we can inspect tool results.
+        _ = (item, final_response)
+        return 0.0
+
+    async def verify_and_score_trajectory(
+        self,
+        item: Item,
+        final_response: str,
+        *,
+        trajectory_id: str,  # noqa: ARG002
+        exec_tool,  # noqa: ARG002
+        agent_result: AgentResult | None = None,
+        workspace_meta: Dict[str, Any] | None = None,  # noqa: ARG002
+    ) -> tuple[float, Dict[str, Any]]:
+        if agent_result is None:
+            return 0.0, {"error": "Missing agent_result"}
+
+        observed: str = ""
+        tool_ok = False
+        for step in agent_result.steps:
+            for res in step.tool_results:
+                if not res.success:
+                    return 0.0, {"error": res.error, "output": res.output}
+                out = (res.output or "").strip()
+                if out:
+                    observed = out.splitlines()[-1].strip()
+                    tool_ok = True
+
+        final = (final_response or "").strip()
+        score = 1.0 if tool_ok and agent_result.total_tool_calls > 0 and observed and final == observed else 0.0
+        return score, {"observed": observed, "tool_calls": agent_result.total_tool_calls, "command": item.get("command")}
+
+
+if __name__ == "__main__":
+    SandboxTerminalSmokeEnv.cli()

atropos/envs/swe_smith_oracle_env.py

@@ -0,0 +1,418 @@
+"""
+SWE-smith-oracle environment.
+
+This environment is intentionally minimal:
+- prepares a sandbox workspace by cloning a public GitHub repo at `base_commit`
+- runs an AtroposAgent tool loop to apply a fix
+- verifies by running pytest nodeids from the dataset (reward = pass/fail)
+- Python only (no multi-language support currently, need to properly bauild & add to dropbox)
+- TODO: Get the other nonpython sandboxes up and running, then add a config knob to switch between them per row
+- oh and add to dockerhub
+
+Dataset: NousResearch/SWE-smith-oracle (train; does NOT use SWE-bench eval set).
+"""
+
+from __future__ import annotations
+
+import os
+import random
+import time
+from typing import Any, Dict, List, Optional, Tuple
+
+from pydantic import Field
+
+from atroposlib.envs.base import APIServerConfig, Item
+
+from ..agent import AgentConfig
+from ..tools import ToolCall
+from .agent_env import AgentEnv, AgentEnvConfig
+
+
+class SweSmithOracleEnvConfig(AgentEnvConfig):
+    dataset_name: str = Field(default="NousResearch/SWE-smith-oracle")
+    dataset_split: str = Field(default="train")
+    max_items: int = Field(default=0, description="0 = no limit")
+    shuffle: bool = Field(default=True)
+    seed: int = Field(default=0)
+
+    python_only: bool = Field(default=True, description="Filter to Python-evaluable rows")
+    score_include_fail_to_pass: bool = Field(
+        default=True,
+        description=(
+            "If true (default), score tests on PASS_TO_PASS ∪ FAIL_TO_PASS. "
+            "Disable to only run PASS_TO_PASS (faster but weaker signal)."
+        ),
+    )
+
+    prompt_mode: str = Field(
+        default="problem_statement",
+        description="Task prompt content: 'problem_statement' (fast) or 'problem_statement+text' (slower, includes dataset 'text').",
+    )
+
+    repo_base_url: str = Field(default="https://github.com", description="Base URL for repo cloning")
+    install_timeout_s: float = Field(default=600.0)
+    test_timeout_s: float = Field(default=600.0)
+
+    tokenizer_name: str = Field(default="NousResearch/Hermes-4.3-36B", description="Tokenizer name for RL tokenization")
+
+
+class SweSmithOracleEnv(AgentEnv[SweSmithOracleEnvConfig]):
+    """
+    SWE-smith-oracle AgentEnv.
+
+    This is designed for benchmarking multiplexed slot execution vs naive container-per-trajectory.
+    """
+
+    name = "swe_smith_oracle_env"
+    env_config_cls = SweSmithOracleEnvConfig
+
+    def __init__(
+        self,
+        config: SweSmithOracleEnvConfig,
+        server_configs: List[APIServerConfig],
+        slurm: bool = False,
+        testing: bool = False,
+    ):
+        super().__init__(config, server_configs, slurm, testing)
+        self._dataset = None
+        self._indices: List[int] = []
+        self._cursor = 0
+
+    @classmethod
+    def config_init(cls) -> Tuple[SweSmithOracleEnvConfig, List[APIServerConfig]]:
+        # Defaults for running the env via CLI in offline `process` mode.
+        # Override via env vars or `--env.*` flags as needed.
+        base_url_raw = (
+            os.getenv("ATROPOS_SERVER_BASE_URL")
+            or os.getenv("OPENAI_BASE_URL")
+            or os.getenv("LLM_BASE_URL")
+            or "http://127.0.0.1:8080"
+        )
+        base_url = base_url_raw.rstrip("/")
+        if not base_url.endswith("/v1"):
+            base_url = f"{base_url}/v1"
+        model = os.getenv("ATROPOS_SERVER_MODEL") or os.getenv("LLM_MODEL") or "hermes-4-36b"
+        api_key = os.getenv("ATROPOS_SERVER_API_KEY") or os.getenv("NOUS_API_KEY") or os.getenv("OPENAI_API_KEY") or "local"
+
+        env_config = SweSmithOracleEnvConfig(
+            tokenizer_name=os.getenv("ATROPOS_TOKENIZER_NAME") or "NousResearch/Hermes-4.3-36B",
+            group_size=1,
+            use_wandb=False,
+            rollout_server_url="http://localhost:8000",
+            total_steps=1,
+            batch_size=1,
+            steps_per_eval=1,
+            max_token_length=8192,
+            inference_weight=1.0,
+            wandb_name="swe_smith_oracle",
+            enabled_toolsets=["terminal"],
+            disabled_toolsets=[],
+            sandbox_image=os.getenv("ATROPOS_SANDBOX_IMAGE") or "atropos-sandbox:local",
+            purge_job_on_start=True,
+            purge_job_on_shutdown=True,
+        )
+
+        server_configs = [
+            APIServerConfig(
+                model_name=model,
+                base_url=base_url,
+                api_key=api_key,
+                num_max_requests_at_once=1,
+                num_requests_for_eval=1,
+                timeout=int(os.getenv("ATROPOS_SERVER_TIMEOUT_S") or "300"),
+            ),
+        ]
+
+        return env_config, server_configs
+
+    async def setup_agent_env(self) -> None:
+        from datasets import load_dataset
+
+        t0 = time.perf_counter()
+        print(
+            f"[SweSmithOracleEnv] loading dataset {self.config.dataset_name}:{self.config.dataset_split} "
+            f"(python_only={self.config.python_only}, max_items={self.config.max_items or 'all'})",
+            flush=True,
+        )
+        ds = load_dataset(self.config.dataset_name, split=self.config.dataset_split)
+        self._dataset = ds
+
+        indices: List[int] = []
+        for idx in range(len(ds)):
+            row = ds[idx]
+            if self.config.python_only and not self._is_python_row(row):
+                continue
+            indices.append(idx)
+
+        if self.config.shuffle:
+            rnd = random.Random(self.config.seed)
+            rnd.shuffle(indices)
+
+        if self.config.max_items and self.config.max_items > 0:
+            indices = indices[: self.config.max_items]
+
+        self._indices = indices
+        self._cursor = 0
+
+        print(
+            f"[SweSmithOracleEnv] loaded {len(self._indices)} items from {self.config.dataset_name}:{self.config.dataset_split} "
+            f"in {time.perf_counter() - t0:.2f}s",
+            flush=True,
+        )
+
+    def _is_python_row(self, row: Dict[str, Any]) -> bool:
+        nodeids = row.get("PASS_TO_PASS")
+        if not isinstance(nodeids, list) or not nodeids:
+            return False
+        for nid in nodeids:
+            if not isinstance(nid, str) or ".py::" not in nid:
+                return False
+        return True
+
+    async def get_next_item(self) -> Item:
+        print(f"[SweSmithOracleEnv] get_next_item() cursor={self._cursor}/{len(self._indices)}", flush=True)
+        if not self._dataset or not self._indices:
+            raise RuntimeError("Dataset not initialized (did setup() run?)")
+        if self._cursor >= len(self._indices):
+            self._cursor = 0
+        idx = self._indices[self._cursor]
+        self._cursor += 1
+        return dict(self._dataset[idx])
+
+    def _repo_name(self, item: Item) -> str:
+        repo = item.get("repo") or ""
+        if isinstance(repo, str) and "/" in repo:
+            return repo.split("/")[-1]
+        return "repo"
+
+    def build_task(self, item: Item) -> str:
+        repo = item.get("repo") or ""
+        base_commit = item.get("base_commit") or ""
+        problem = str(item.get("problem_statement") or "")
+        context = str(item.get("text") or "")
+
+        nodeids = self._tests_for_item(item)
+        tests_list = "\n".join(f"- {t}" for t in nodeids)
+
+        repo_dir = self._repo_name(item)
+
+        tests_block = (
+            "Run these tests to verify:\n"
+            f"{tests_list}\n\n"
+            "When done, briefly describe what you changed and confirm tests pass."
+        )
+
+        prompt_mode = (self.config.prompt_mode or "problem_statement").strip().lower()
+        if prompt_mode not in {"problem_statement", "problem_statement+text"}:
+            raise ValueError(
+                f"Invalid prompt_mode={self.config.prompt_mode!r}. "
+                "Expected 'problem_statement' or 'problem_statement+text'."
+            )
+
+        context_block = ""
+        if prompt_mode == "problem_statement+text" and context:
+            # Note: We intentionally do NOT truncate/cap here. This mode is for debugging / richer prompts and can be slow.
+            context_block = f"\nAdditional context:\n{context}\n"
+
+        return (
+            "You are a senior software engineer. Fix the repository so the specified tests pass.\n\n"
+            f"Repository: {repo} (checked out at base_commit={base_commit})\n"
+            f"Workspace path: ./{repo_dir}\n\n"
+            "Constraints:\n"
+            "- You MUST use the terminal tool to inspect, edit, and verify the repository. Do not respond with a patch file.\n"
+            f"- Start by inspecting the repo (e.g. `ls`, `cd ./{repo_dir}`, `git status`).\n"
+            "- Use a workspace-local virtualenv (e.g. inside the repo at ./.venv) to avoid cross-run contamination.\n"
+            "- Use non-interactive commands only.\n\n"
+            "- Terminal commands run under POSIX /bin/sh and each tool call runs in a fresh shell (no persisted env vars).\n"
+            "  Avoid bash-only `source`; prefer `. .venv/bin/activate` or `.venv/bin/python ...`.\n\n"
+            "Problem statement:\n"
+            f"{problem}\n\n"
+            f"{context_block}\n"
+            f"{tests_block}"
+        )
+
+    def build_agent_config(self, item: Item) -> AgentConfig:  # noqa: ARG002
+        # SWE tasks are longer than the simple test env.
+        return AgentConfig(
+            max_steps=self.config.agent_max_steps,
+            temperature=self.config.agent_temperature,
+            max_tokens=self.config.agent_max_tokens,
+            tool_delay_s=self.config.agent_tool_delay_s,
+        )
+
+    async def setup_trajectory_workspace(self, item: Item, *, trajectory_id: str, exec_tool) -> Dict[str, Any]:
+        t0 = time.perf_counter()
+        repo = item.get("repo")
+        base_commit = item.get("base_commit")
+        instance_id = item.get("instance_id") or item.get("id") or item.get("problem_id")
+        if not isinstance(repo, str) or not isinstance(base_commit, str):
+            raise RuntimeError("Invalid dataset row: missing repo/base_commit")
+
+        repo_dir = self._repo_name(item)
+        clone_url = f"{self.config.repo_base_url.rstrip('/')}/{repo}.git"
+        print(
+            f"[SweSmithOracleEnv] tid={trajectory_id} setup_trajectory_workspace(): "
+            f"repo={repo} base_commit={base_commit} instance_id={instance_id} dir=./{repo_dir}",
+            flush=True,
+        )
+
+        # Repo setup strategy:
+        # - Maintain a shared, per-container bare repo cache under /data/repo_cache
+        # - For each trajectory, create an isolated git worktree under the slot workspace
+        # This avoids cloning/fetching full repos per trajectory and is crucial for multiplexing.
+
+        def _repo_cache_slug(repo_name: str) -> str:
+            return repo_name.replace("/", "__")
+
+        repo_slug = _repo_cache_slug(repo)
+        cache_root = "/data/repo_cache"
+        bare_repo = f"{cache_root}/{repo_slug}.git"
+        lock_file = f"{cache_root}/.locks/{repo_slug}.lock"
+
+        # Use flock to serialize operations that mutate the shared bare repo (fetch/worktree).
+        # util-linux (flock) is included in the sandbox image.
+        worktree_cmd = (
+            "set -e; "
+            f"rm -rf {repo_dir}; "
+            f"mkdir -p {cache_root}/.locks; "
+            f": > {lock_file}; "
+            f"flock -x {lock_file} sh -lc '"
+            f"set -e; "
+            "export GIT_TERMINAL_PROMPT=0; "
+            "export GIT_LFS_SKIP_SMUDGE=1; "
+            f"if [ ! -d \"{bare_repo}\" ]; then "
+            f"  git init --bare \"{bare_repo}\"; "
+            f"  git -C \"{bare_repo}\" remote add origin \"{clone_url}\"; "
+            "fi; "
+            f"git -C \"{bare_repo}\" remote set-url origin \"{clone_url}\"; "
+            f"git -C \"{bare_repo}\" worktree prune || true; "
+            f"if ! git -C \"{bare_repo}\" cat-file -e \"{base_commit}^{{commit}}\" 2>/dev/null; then "
+            f"  git -C \"{bare_repo}\" fetch --depth 1 origin \"{base_commit}\" || true; "
+            "fi; "
+            f"if ! git -C \"{bare_repo}\" cat-file -e \"{base_commit}^{{commit}}\" 2>/dev/null; then "
+            f"  git -C \"{bare_repo}\" fetch --prune origin; "
+            "fi; "
+            f"git --git-dir=\"{bare_repo}\" worktree add --detach \"{repo_dir}\" \"{base_commit}\"; "
+            "'"
+        )
+
+        print(f"[SweSmithOracleEnv] tid={trajectory_id} preparing worktree from repo cache", flush=True)
+        res = await exec_tool(
+            ToolCall(
+                name="terminal",
+                arguments={"command": worktree_cmd, "timeout": self.config.install_timeout_s},
+            )
+        )
+        if not res.success:
+            raise RuntimeError(
+                "git worktree setup failed "
+                f"(repo={repo}, base_commit={base_commit}, instance_id={instance_id}): {res.error}\n{res.output}"
+            )
+
+        print(
+            f"[SweSmithOracleEnv] tid={trajectory_id} setup_trajectory_workspace(): worktree ready in {time.perf_counter() - t0:.2f}s",
+            flush=True,
+        )
+        return {"repo_dir": repo_dir, "base_commit": base_commit}
+
+    def _tests_for_item(self, item: Item) -> List[str]:
+        tests: List[str] = []
+        if self.config.score_include_fail_to_pass:
+            for key in ("PASS_TO_PASS", "FAIL_TO_PASS"):
+                nodeids = item.get(key)
+                if isinstance(nodeids, list):
+                    tests.extend([n for n in nodeids if isinstance(n, str)])
+        else:
+            nodeids = item.get("PASS_TO_PASS")
+            if isinstance(nodeids, list):
+                tests.extend([n for n in nodeids if isinstance(n, str)])
+        # Stable order for reproducibility.
+        return sorted(dict.fromkeys(tests))
+
+    def _chunk_nodeids(self, nodeids: List[str], max_per_chunk: int = 50) -> List[List[str]]:
+        chunks: List[List[str]] = []
+        for i in range(0, len(nodeids), max_per_chunk):
+            chunks.append(nodeids[i : i + max_per_chunk])
+        return chunks
+
+    async def verify_and_score_trajectory(
+        self,
+        item: Item,
+        final_response: str,  # noqa: ARG002
+        *,
+        trajectory_id: str,
+        exec_tool,
+        agent_result=None,
+        workspace_meta: Optional[Dict[str, Any]] = None,
+    ) -> tuple[float, Dict[str, Any]]:
+        _ = trajectory_id
+        repo_dir = self._repo_name(item)
+
+        # Training correctness: do not reward trajectories that never actually used tools.
+        if agent_result is not None and getattr(agent_result, "total_tool_calls", 0) <= 0:
+            print(
+                f"[SweSmithOracleEnv] tid={trajectory_id} verify (dataset_tests): no tool calls; score=0.0",
+                flush=True,
+            )
+            return 0.0, {
+                "verification_mode": "dataset_tests",
+                "error": "No tool calls were made by the agent",
+            }
+
+        nodeids = self._tests_for_item(item)
+        if not nodeids:
+            return 0.0, {"error": "No tests provided"}
+
+        print(f"[SweSmithOracleEnv] tid={trajectory_id} verify (dataset_tests): ensuring venv + deps", flush=True)
+        setup_cmd = (
+            f"cd {repo_dir} && "
+            "python -m venv .venv && "
+            ". .venv/bin/activate && "
+            "python -m pip install -U pip setuptools wheel && "
+            "python -m pip install -e . && "
+            "python -m pip install pytest"
+        )
+        setup_res = await exec_tool(
+            ToolCall(name="terminal", arguments={"command": setup_cmd, "timeout": self.config.install_timeout_s})
+        )
+        verification_messages = [{"role": "user", "content": setup_res.to_xml()}]
+        if not setup_res.success:
+            return 0.0, {
+                "verification_mode": "dataset_tests",
+                "phase": "install",
+                "error": setup_res.error,
+                "output": setup_res.output,
+                "verification_messages": verification_messages,
+            }
+
+        chunks = self._chunk_nodeids(nodeids, max_per_chunk=50)
+        for chunk_idx, chunk in enumerate(chunks):
+            joined = " ".join(chunk)
+            cmd = f"cd {repo_dir} && . .venv/bin/activate && python -m pytest -q {joined}"
+            res = await exec_tool(
+                ToolCall(
+                    name="terminal",
+                    arguments={"command": cmd, "timeout": self.config.test_timeout_s},
+                )
+            )
+            verification_messages.append({"role": "user", "content": res.to_xml()})
+            if not res.success:
+                return 0.0, {
+                    "verification_mode": "dataset_tests",
+                    "phase": "pytest",
+                    "failed_chunk": chunk_idx,
+                    "error": res.error,
+                    "output": res.output,
+                    "verification_messages": verification_messages,
+                }
+
+        return 1.0, {"verification_mode": "dataset_tests", "passed": True, "verification_messages": verification_messages}
+
+    async def score_trajectory(self, item: Item, final_response: str) -> float:
+        # Not used; scoring happens in verify_and_score_trajectory.
+        _ = (item, final_response)
+        return 0.0
+
+
+if __name__ == "__main__":
+    SweSmithOracleEnv.cli()

atropos/envs/test_env.py

@@ -0,0 +1,217 @@
+"""
+Simple test environment for validating the atropos-agent setup.
+
+This environment uses a local OpenAI-compatible server for LLM testing to verify:
+- BaseEnv extension works correctly
+- API communication via OpenAI-compatible endpoint
+- Basic trajectory collection
+
+This is a minimal environment for testing, not production use.
+"""
+
+import os
+from typing import Dict, List, Optional, Tuple
+
+from dotenv import load_dotenv
+from pydantic import Field
+
+from atroposlib.envs.base import (
+    APIServerConfig,
+    Item,
+)
+
+from ..agent import AgentConfig
+from .agent_env import AgentEnv, AgentEnvConfig
+
+# Load environment variables from .env file
+load_dotenv()
+
+
+# Simple test prompts for validation
+TEST_PROMPTS = [
+    {
+        "prompt": "What is 2 + 2? Answer with just the number.",
+        "expected": "4",
+    },
+    {
+        "prompt": "What is the capital of France? Answer with just the city name.",
+        "expected": "Paris",
+    },
+    {
+        "prompt": "What color is the sky on a clear day? Answer with just the color.",
+        "expected": "Blue",
+    },
+    {
+        "prompt": "How many days are in a week? Answer with just the number.",
+        "expected": "7",
+    },
+    {
+        "prompt": "What is 10 * 5? Answer with just the number.",
+        "expected": "50",
+    },
+]
+
+SYSTEM_PROMPT = (
+    "You are a helpful assistant. Answer questions concisely and directly. "
+    "When asked for a simple answer, provide just that answer without explanation."
+)
+
+
+class SimpleTestEnvConfig(AgentEnvConfig):
+    """Configuration for the simple test environment."""
+
+    server_base_url: str = Field(
+        default="http://127.0.0.1:8080",
+        description="Base URL for an OpenAI-compatible server (without /v1)",
+    )
+    server_model: str = Field(
+        default="hermes-4-36b",
+        description="Model name",
+    )
+    tokenizer_name: str = Field(default="NousResearch/Hermes-4.3-36B", description="Tokenizer name for RL tokenization")
+
+
+class SimpleTestEnv(AgentEnv[SimpleTestEnvConfig]):
+    """
+    A simple test environment to validate the atropos-agent setup.
+    
+    Uses a local OpenAI-compatible LLM endpoint with basic question-answering tasks.
+    Scoring is based on whether the response contains the expected answer.
+    """
+
+    name = "simple_test_env"
+    env_config_cls = SimpleTestEnvConfig
+
+    def __init__(
+        self,
+        config: SimpleTestEnvConfig,
+        server_configs: List[APIServerConfig],
+        slurm: bool = False,
+        testing: bool = False,
+    ):
+        super().__init__(config, server_configs, slurm, testing)
+        self.iter = 0
+        self.test_prompts = TEST_PROMPTS
+        self.percent_correct_buffer: List[float] = []
+
+    @classmethod
+    def config_init(cls) -> Tuple[SimpleTestEnvConfig, List[APIServerConfig]]:
+        """
+        Initialize configuration with local server settings from environment variables.
+        """
+        base_url = (
+            os.getenv("ATROPOS_SERVER_BASE_URL")
+            or os.getenv("OPENAI_BASE_URL")
+            or os.getenv("LLM_BASE_URL")
+            or "http://127.0.0.1:8080"
+        )
+        model = os.getenv("ATROPOS_SERVER_MODEL") or os.getenv("LLM_MODEL") or "hermes-4-36b"
+        api_key = os.getenv("ATROPOS_SERVER_API_KEY") or os.getenv("NOUS_API_KEY") or os.getenv("OPENAI_API_KEY") or "local"
+
+        env_config = SimpleTestEnvConfig(
+            tokenizer_name=os.getenv("ATROPOS_TOKENIZER_NAME") or "NousResearch/Hermes-4.3-36B",
+            group_size=4,
+            use_wandb=False,  # Disable wandb for simple testing
+            rollout_server_url="http://localhost:8000",
+            total_steps=10,
+            batch_size=16,
+            steps_per_eval=5,
+            max_token_length=2048,
+            inference_weight=1.0,
+            wandb_name="simple_test",
+            server_base_url=base_url,
+            server_model=model,
+        )
+
+        # OpenAI-compatible servers typically expose chat completions at /v1.
+        server_configs = [
+            APIServerConfig(
+                model_name=model,
+                base_url=f"{base_url}/v1",
+                api_key=api_key,
+                num_max_requests_at_once=4,
+                num_requests_for_eval=8,
+                timeout=120,  # Local models may be slower
+            ),
+        ]
+
+        return env_config, server_configs
+
+    async def setup_agent_env(self):
+        """Setup the environment - load test data."""
+        print(f"SimpleTestEnv setup complete. {len(self.test_prompts)} test prompts loaded.")
+        print(f"Using server at: {self.config.server_base_url}")
+        print(f"Model: {self.config.server_model}")
+
+    async def get_next_item(self) -> Item:
+        """Get the next test prompt."""
+        item = self.test_prompts[self.iter % len(self.test_prompts)]
+        self.iter += 1
+        return item
+
+    def build_task(self, item: Item) -> str:
+        return item["prompt"]
+
+    def build_agent_config(self, item: Item) -> AgentConfig:  # noqa: ARG002
+        return AgentConfig(
+            max_steps=5,
+            temperature=0.7,
+            max_tokens=256,
+            system_prompt=SYSTEM_PROMPT,
+        )
+
+    async def score_trajectory(self, item: Item, final_response: str) -> float:
+        expected = item["expected"].lower()
+        response_lower = (final_response or "").lower()
+        score = 1.0 if expected in response_lower else 0.0
+        self.percent_correct_buffer.append(score)
+        return score
+
+    async def evaluate(self, *args, **kwargs):
+        """
+        Simple evaluation - run through all test prompts once.
+        """
+        correct = 0
+        total = len(self.test_prompts)
+
+        for item in self.test_prompts:
+            messages = [
+                {"role": "system", "content": SYSTEM_PROMPT},
+                {"role": "user", "content": item["prompt"]},
+            ]
+
+            response = await self.server.chat_completion(
+                messages=messages,
+                n=1,
+                max_tokens=256,
+                temperature=0.0,  # Greedy for eval
+                split="eval",
+            )
+
+            response_text = response.choices[0].message.content or ""
+            expected = item["expected"].lower()
+
+            if expected in response_text.lower():
+                correct += 1
+
+        accuracy = correct / total
+        print(f"Evaluation: {correct}/{total} = {accuracy:.2%} accuracy")
+        return {"eval_accuracy": accuracy}
+
+    async def wandb_log(self, wandb_metrics: Optional[Dict] = None):
+        """Log metrics (simplified for testing)."""
+        if wandb_metrics is None:
+            wandb_metrics = {}
+
+        if self.percent_correct_buffer:
+            avg_correct = sum(self.percent_correct_buffer) / len(self.percent_correct_buffer)
+            wandb_metrics["train/percent_correct"] = avg_correct
+            print(f"Train accuracy: {avg_correct:.2%}")
+            self.percent_correct_buffer = []
+
+        await super().wandb_log(wandb_metrics)
+
+
+if __name__ == "__main__":
+    # Allow running as CLI
+    SimpleTestEnv.cli()

atropos/envs/toolserver_smoke_env.py

@@ -0,0 +1,165 @@
+"""
+ToolServer routing smoke environment.
+
+Validates that:
+  - sandbox tools run through Nomad SlotPool (terminal -> bash in sandbox)
+  - external tools run through ToolServer (skills_list)
+
+This env uses ToolServer in-process by default (`tool_server_url="inprocess"`),
+so it is self-contained for local testing.
+
+Run:
+  uv run python -m atropos.envs.toolserver_smoke_env process --env.use_wandb false --env.total_steps 1 --env.group_size 1
+"""
+
+from __future__ import annotations
+
+import os
+from typing import Any, Dict, List, Tuple
+
+from dotenv import load_dotenv
+from pydantic import Field
+
+from atroposlib.envs.base import APIServerConfig, Item
+
+from ..agent import AgentConfig, AgentResult
+from .agent_env import AgentEnv, AgentEnvConfig
+
+load_dotenv()
+
+
+class ToolServerSmokeEnvConfig(AgentEnvConfig):
+    server_base_url: str = Field(
+        default="http://127.0.0.1:8080",
+        description="Base URL for an OpenAI-compatible chat server (without /v1).",
+    )
+    server_model: str = Field(default="hermes-4-36b", description="Model name")
+    tokenizer_name: str = Field(default="NousResearch/Hermes-4.3-36B", description="Tokenizer name for RL tokenization")
+
+
+class ToolServerSmokeEnv(AgentEnv[ToolServerSmokeEnvConfig]):
+    name = "toolserver_smoke_env"
+    env_config_cls = ToolServerSmokeEnvConfig
+
+    def __init__(
+        self,
+        config: ToolServerSmokeEnvConfig,
+        server_configs: List[APIServerConfig],
+        slurm: bool = False,
+        testing: bool = False,
+    ):
+        super().__init__(config, server_configs, slurm, testing)
+        self._iter = 0
+
+    @classmethod
+    def config_init(cls) -> Tuple[ToolServerSmokeEnvConfig, List[APIServerConfig]]:
+        base_url = (
+            os.getenv("ATROPOS_SERVER_BASE_URL")
+            or os.getenv("OPENAI_BASE_URL")
+            or os.getenv("LLM_BASE_URL")
+            or "http://127.0.0.1:8080"
+        )
+        model = os.getenv("ATROPOS_SERVER_MODEL") or os.getenv("LLM_MODEL") or "hermes-4-36b"
+        api_key = os.getenv("ATROPOS_SERVER_API_KEY") or os.getenv("NOUS_API_KEY") or os.getenv("OPENAI_API_KEY") or "local"
+
+        env_config = ToolServerSmokeEnvConfig(
+            tokenizer_name=os.getenv("ATROPOS_TOKENIZER_NAME") or "NousResearch/Hermes-4.3-36B",
+            group_size=1,
+            use_wandb=False,
+            include_messages=True,
+            ensure_scores_are_not_same=False,
+            total_steps=1,
+            batch_size=1,
+            server_base_url=base_url,
+            server_model=model,
+            enabled_toolsets=["terminal", "skills"],
+            disabled_toolsets=[],
+            # Self-contained ToolServer for local smoke.
+            tool_server_url="inprocess",
+            sandbox_image=os.getenv("ATROPOS_SANDBOX_IMAGE") or "atropos-sandbox:local",
+            purge_job_on_start=True,
+            purge_job_on_shutdown=True,
+        )
+
+        server_configs = [
+            APIServerConfig(
+                model_name=model,
+                base_url=f"{base_url.rstrip('/')}/v1",
+                api_key=api_key,
+                num_max_requests_at_once=1,
+                num_requests_for_eval=1,
+                timeout=120,
+            )
+        ]
+        return env_config, server_configs
+
+    async def setup_agent_env(self) -> None:
+        return None
+
+    async def get_next_item(self) -> Item:
+        self._iter += 1
+        return {
+            "prompt": (
+                "You MUST call exactly one tool per assistant message.\n"
+                "\n"
+                "Step 1) Call the skills_list tool (no arguments), then stop.\n"
+                "Step 2) After you receive the tool response, call the terminal tool to run:\n"
+                "python -c \"print('ok')\"\n"
+                "Step 3) After you receive the terminal tool response, answer with just: ok\n"
+                "\n"
+                "Tool call format requirements:\n"
+                "- Every tool call MUST be a complete XML block with a closing tag.\n"
+                "- Do NOT emit a second <tool_call> in the same assistant message.\n"
+                "\n"
+                "Example:\n"
+                "<tool_call>{\"name\": \"skills_list\", \"arguments\": {}}</tool_call>\n"
+                "Do not include anything else in your final answer."
+            )
+        }
+
+    def build_task(self, item: Item) -> str:
+        return str(item.get("prompt") or "")
+
+    def build_agent_config(self, item: Item) -> AgentConfig:  # noqa: ARG002
+        return AgentConfig(
+            max_steps=min(10, int(self.config.agent_max_steps)),
+            temperature=0.2,
+            max_tokens=None,
+        )
+
+    async def score_trajectory(self, item: Item, final_response: str) -> float:
+        _ = (item, final_response)
+        return 0.0
+
+    async def verify_and_score_trajectory(
+        self,
+        item: Item,
+        final_response: str,
+        *,
+        trajectory_id: str,  # noqa: ARG002
+        exec_tool,  # noqa: ARG002
+        agent_result: AgentResult | None = None,
+        workspace_meta: Dict[str, Any] | None = None,  # noqa: ARG002
+    ) -> tuple[float, Dict[str, Any]]:
+        if agent_result is None:
+            return 0.0, {"error": "Missing agent_result"}
+
+        called = {c.name for s in agent_result.steps for c in s.tool_calls}
+        need = {"skills_list", "terminal"}
+        if not need.issubset(called):
+            return 0.0, {"error": f"Missing tool calls: {sorted(need - called)}", "called": sorted(called)}
+
+        terminal_ok = False
+        for step in agent_result.steps:
+            for call, res in zip(step.tool_calls, step.tool_results):
+                if call.name != "terminal":
+                    continue
+                if res.success and (res.output or "").strip().splitlines()[-1].strip() == "ok":
+                    terminal_ok = True
+
+        score = 1.0 if terminal_ok and (final_response or "").strip() == "ok" else 0.0
+        return score, {"called": sorted(called), "final": (final_response or "").strip()}
+
+
+if __name__ == "__main__":
+    ToolServerSmokeEnv.cli()

atropos/nomad/__init__.py

@@ -0,0 +1,11 @@
+"""
+Nomad integration for atropos-agent.
+
+Provides:
+- NomadClient: Client for Nomad HTTP API
+- Job templates for sandbox containers
+"""
+
+from .client import NomadClient
+
+__all__ = ["NomadClient"]

atropos/nomad/client.py

@@ -0,0 +1,500 @@
+"""
+Nomad API Client for atropos-agent.
+
+Provides a simple async client for interacting with the Nomad HTTP API:
+- Submit/stop jobs
+- Query allocations
+- Get allocation addresses
+- Scale jobs up/down
+"""
+
+import asyncio
+import json
+import os
+from dataclasses import dataclass, field
+from enum import Enum
+from pathlib import Path
+from typing import Any, Dict, List, Optional
+
+import aiohttp
+
+
+class AllocationStatus(Enum):
+    """Nomad allocation status."""
+    PENDING = "pending"
+    RUNNING = "running"
+    COMPLETE = "complete"
+    FAILED = "failed"
+    LOST = "lost"
+
+
+@dataclass
+class Allocation:
+    """Information about a Nomad allocation."""
+    id: str
+    job_id: str
+    task_group: str
+    node_id: str
+    status: AllocationStatus
+    # Network info for reaching the allocation
+    address: Optional[str] = None
+    port: Optional[int] = None
+    
+    @property
+    def http_address(self) -> Optional[str]:
+        """Get full HTTP address for the allocation."""
+        if self.address and self.port:
+            return f"http://{self.address}:{self.port}"
+        return None
+
+
+@dataclass
+class JobStatus:
+    """Status of a Nomad job."""
+    id: str
+    name: str
+    status: str
+    allocations: List[Allocation] = field(default_factory=list)
+    count: int = 0  # Number of task groups
+
+
+class NomadClient:
+    """
+    Async client for Nomad HTTP API.
+    
+    Usage:
+        client = NomadClient(address="http://localhost:4646")
+        
+        # Submit a job
+        await client.submit_job(job_spec)
+        
+        # Get allocations
+        allocs = await client.get_job_allocations("sandbox-python")
+        
+        # Scale job
+        await client.scale_job("sandbox-python", count=5)
+    """
+    
+    def __init__(
+        self,
+        address: str = "http://localhost:4646",
+        token: Optional[str] = None,
+        timeout: float = 30.0,
+    ):
+        self.address = address.rstrip("/")
+        self.token = token or os.environ.get("NOMAD_TOKEN")
+        self.timeout = aiohttp.ClientTimeout(total=timeout)
+        self._session: Optional[aiohttp.ClientSession] = None
+    
+    async def _get_session(self) -> aiohttp.ClientSession:
+        """Get or create HTTP session."""
+        if self._session is None or self._session.closed:
+            headers = {}
+            if self.token:
+                headers["X-Nomad-Token"] = self.token
+            self._session = aiohttp.ClientSession(
+                timeout=self.timeout,
+                headers=headers,
+            )
+        return self._session
+    
+    async def close(self):
+        """Close the HTTP session."""
+        if self._session and not self._session.closed:
+            await self._session.close()
+    
+    async def __aenter__(self):
+        return self
+    
+    async def __aexit__(self, exc_type, exc_val, exc_tb):
+        await self.close()
+    
+    async def _request(
+        self,
+        method: str,
+        path: str,
+        data: Optional[Dict[str, Any]] = None,
+    ) -> Dict[str, Any]:
+        """Make an HTTP request to Nomad API."""
+        session = await self._get_session()
+        url = f"{self.address}{path}"
+        
+        try:
+            async with session.request(method, url, json=data) as response:
+                if response.status == 404:
+                    return {"error": "not_found", "status": 404}
+                
+                text = await response.text()
+                if not text:
+                    return {"status": response.status}
+                
+                try:
+                    result = json.loads(text)
+                except json.JSONDecodeError:
+                    return {"text": text, "status": response.status}
+                
+                if response.status >= 400:
+                    return {"error": result, "status": response.status}
+                
+                return result if isinstance(result, dict) else {"data": result, "status": response.status}
+                
+        except aiohttp.ClientError as e:
+            return {"error": str(e), "status": 0}
+    
+    # Job Operations
+    
+    async def submit_job(self, job_spec: Dict[str, Any]) -> Dict[str, Any]:
+        """
+        Submit a job to Nomad.
+        
+        Args:
+            job_spec: Job specification dict (HCL converted to JSON)
+            
+        Returns:
+            Response with EvalID if successful
+        """
+        return await self._request("POST", "/v1/jobs", {"Job": job_spec})
+    
+    async def stop_job(self, job_id: str, purge: bool = False) -> Dict[str, Any]:
+        """
+        Stop (and optionally purge) a job.
+        
+        Args:
+            job_id: Job identifier
+            purge: If True, completely remove the job
+        """
+        path = f"/v1/job/{job_id}"
+        if purge:
+            path += "?purge=true"
+        return await self._request("DELETE", path)
+    
+    async def get_job(self, job_id: str) -> Optional[Dict[str, Any]]:
+        """Get job details."""
+        result = await self._request("GET", f"/v1/job/{job_id}")
+        if "error" in result and result.get("status") == 404:
+            return None
+        return result
+    
+    async def get_job_status(self, job_id: str) -> Optional[JobStatus]:
+        """Get job status with allocations."""
+        job = await self.get_job(job_id)
+        if not job:
+            return None
+        
+        allocs = await self.get_job_allocations(job_id)
+        
+        # Get count from task groups
+        count = 0
+        task_groups = job.get("TaskGroups", [])
+        for tg in task_groups:
+            count += tg.get("Count", 1)
+        
+        return JobStatus(
+            id=job_id,
+            name=job.get("Name", job_id),
+            status=job.get("Status", "unknown"),
+            allocations=allocs,
+            count=count,
+        )
+    
+    # Allocation Operations
+    
+    async def get_job_allocations(self, job_id: str) -> List[Allocation]:
+        """Get all allocations for a job."""
+        result = await self._request("GET", f"/v1/job/{job_id}/allocations")
+        
+        if "error" in result:
+            return []
+        
+        allocs_data = result.get("data", result) if isinstance(result, dict) else result
+        if not isinstance(allocs_data, list):
+            return []
+        
+        allocations = []
+        for alloc_data in allocs_data:
+            # Parse allocation info
+            alloc_id = alloc_data.get("ID", "")
+            status_str = alloc_data.get("ClientStatus", "unknown")
+            
+            try:
+                status = AllocationStatus(status_str)
+            except ValueError:
+                status = AllocationStatus.PENDING
+            
+            # Get network info - need to fetch detailed allocation for this
+            address = None
+            port = None
+            
+            # First try the summary data
+            resources = alloc_data.get("AllocatedResources") or {}
+            shared = resources.get("Shared") or {}
+            networks = shared.get("Networks") or []
+            
+            # If no networks in summary, fetch detailed allocation
+            if not networks and alloc_id:
+                detailed = await self.get_allocation(alloc_id)
+                if detailed:
+                    resources = detailed.get("AllocatedResources") or {}
+                    shared = resources.get("Shared") or {}
+                    networks = shared.get("Networks") or []
+            
+            if networks:
+                network = networks[0]
+                address = network.get("IP")
+                # Look for dynamic ports OR reserved ports (Singularity/raw_exec uses reserved)
+                dyn_ports = network.get("DynamicPorts") or []
+                reserved_ports = network.get("ReservedPorts") or []
+                for dp in dyn_ports + reserved_ports:
+                    if dp.get("Label") == "http":
+                        port = dp.get("Value")
+                        break
+            
+            allocations.append(Allocation(
+                id=alloc_id,
+                job_id=job_id,
+                task_group=alloc_data.get("TaskGroup", ""),
+                node_id=alloc_data.get("NodeID", ""),
+                status=status,
+                address=address,
+                port=port,
+            ))
+        
+        return allocations
+    
+    async def get_allocation(self, alloc_id: str) -> Optional[Dict[str, Any]]:
+        """Get detailed allocation info."""
+        result = await self._request("GET", f"/v1/allocation/{alloc_id}")
+        if "error" in result and result.get("status") == 404:
+            return None
+        return result
+    
+    # Scaling Operations
+    
+    async def scale_job(self, job_id: str, count: int, task_group: str = "sandbox") -> Dict[str, Any]:
+        """
+        Scale a job's task group to specified count.
+        
+        Args:
+            job_id: Job identifier
+            count: Desired number of allocations
+            task_group: Name of task group to scale
+        """
+        payload = {
+            "Count": count,
+            "Target": {
+                "Group": task_group,
+            },
+        }
+        return await self._request("POST", f"/v1/job/{job_id}/scale", payload)
+    
+    async def get_job_scale_status(self, job_id: str) -> Dict[str, int]:
+        """
+        Get current scale status for a job.
+        
+        Returns:
+            Dict mapping task group name to count
+        """
+        result = await self._request("GET", f"/v1/job/{job_id}/scale")
+        
+        if "error" in result:
+            return {}
+        
+        task_groups = result.get("TaskGroups", {})
+        return {
+            name: info.get("Running", 0)
+            for name, info in task_groups.items()
+        }
+    
+    # Health Check
+    
+    async def is_healthy(self) -> bool:
+        """Check if Nomad is reachable and healthy."""
+        try:
+            result = await self._request("GET", "/v1/status/leader")
+            return "error" not in result
+        except Exception:
+            return False
+    
+    async def get_leader(self) -> Optional[str]:
+        """Get current Nomad leader address."""
+        result = await self._request("GET", "/v1/status/leader")
+        if isinstance(result, dict) and "data" in result:
+            return result["data"]
+        return None
+
+
+def load_job_template(
+    template_name: str = "sandbox",
+    **kwargs,
+) -> Dict[str, Any]:
+    """
+    Load and configure a job template.
+    
+    Args:
+        template_name: Name of template (e.g., "sandbox")
+        **kwargs: Template variables to substitute
+        
+    Returns:
+        Job specification dict ready for Nomad API
+    """
+    # Default job template for sandbox container
+    if template_name == "sandbox":
+        return create_sandbox_job(**kwargs)
+    else:
+        raise ValueError(f"Unknown template: {template_name}")
+
+
+def create_sandbox_job(
+    job_id: str = "atropos-sandbox",
+    image: str = "atropos-sandbox:local",  # Use :local tag to avoid registry pull
+    count: int = 1,
+    slots_per_container: int = 10,
+    privileged: bool = False,
+    cpu: int = 500,
+    memory: int = 512,
+    port: int = 8080,
+    datacenter: str = "dc1",
+    driver: str = "docker",  # "docker" or "singularity"
+    singularity_image: str = None,  # Path to .sif file for singularity driver
+) -> Dict[str, Any]:
+    """
+    Create a sandbox job specification.
+    
+    This job runs the sandbox_server.py inside a container,
+    with the specified number of slots for agent workspaces.
+    
+    Args:
+        job_id: Unique job identifier
+        image: Docker image to use (for docker driver)
+        count: Number of container instances
+        slots_per_container: Number of slots per container
+        privileged: Run container in privileged mode (recommended for bubblewrap)
+        cpu: CPU allocation in MHz
+        memory: Memory allocation in MB
+        port: HTTP port for sandbox server
+        datacenter: Nomad datacenter
+        driver: Container driver - "docker" or "singularity"
+        singularity_image: Path to .sif file (required if driver="singularity")
+        
+    Returns:
+        Job specification dict
+    """
+    # Build task config based on driver
+    if driver == "singularity":
+        if not singularity_image:
+            raise ValueError("singularity_image path required when driver='singularity'")
+        
+        # Use raw_exec driver to run apptainer via shell for variable expansion
+        # The container binds the allocation directory for workspace persistence
+        # For raw_exec, we use static port since Nomad's dynamic port mapping doesn't
+        # work the same as Docker - the process runs directly on the host.
+        shell_cmd = (
+            f'apptainer run '
+            f'--bind "$NOMAD_ALLOC_DIR/data:/data" '
+            f'--pwd /app '
+            f'--env PYTHONUNBUFFERED=1 '
+            f'{singularity_image} '
+            f'python sandbox_server.py '
+            f'--port {port} '
+            f'--slots {slots_per_container} '
+            f'--data-dir /data'
+        )
+        task_config = {
+            "command": "/bin/sh",
+            "args": ["-c", shell_cmd],
+        }
+        task_driver = "raw_exec"
+    else:
+        # Docker driver (default)
+        task_config = {
+            "image": image,
+            "force_pull": False,  # Use local image, don't try to pull
+            "ports": ["http"],
+            "privileged": privileged,
+            "command": "python",
+            "args": [
+                "sandbox_server.py",
+                "--port", str(port),
+                "--slots", str(slots_per_container),
+                "--data-dir", "/data",
+            ],
+            # Note: On Linux, you can mount persistent storage:
+            # "volumes": ["${NOMAD_ALLOC_DIR}/data:/data"],
+            # On macOS/Docker Desktop, skip volumes for PoC
+            # (container /data is ephemeral but works for testing)
+        }
+        task_driver = "docker"
+    
+    # For Singularity/raw_exec, use static ports since the process runs directly on host.
+    # For Docker, use dynamic ports with port mapping.
+    if driver == "singularity":
+        network_config = {
+            "Mode": "host",
+            "ReservedPorts": [
+                {
+                    "Label": "http",
+                    "Value": port,
+                }
+            ],
+        }
+    else:
+        network_config = {
+            "Mode": "host",
+            "DynamicPorts": [
+                {
+                    "Label": "http",
+                    "To": port,
+                }
+            ],
+        }
+    
+    return {
+        "ID": job_id,
+        "Name": job_id,
+        "Type": "service",
+        "Datacenters": [datacenter],
+        "TaskGroups": [
+            {
+                "Name": "sandbox",
+                "Count": count,
+                # Speed up deployments and avoid Consul checks. Without this, Nomad may
+                # keep an "active deployment" around for the default MinHealthyTime,
+                # which blocks immediate scaling under load.
+                "Update": {
+                    "HealthCheck": "task_states",
+                    "MinHealthyTime": 0,
+                },
+                "Networks": [network_config],
+                "Tasks": [
+                    {
+                        "Name": "sandbox-server",
+                        "Driver": task_driver,
+                        "Config": task_config,
+                        "Env": {
+                            "PYTHONUNBUFFERED": "1",
+                            "NOMAD_ALLOC_DIR": "${NOMAD_ALLOC_DIR}",
+                        },
+                        "Resources": {
+                            "CPU": cpu,
+                            "MemoryMB": memory,
+                        },
+                        # Note: Services with Checks require Consul, which we skip for the PoC
+                    }
+                ],
+                "RestartPolicy": {
+                    "Attempts": 3,
+                    "Interval": 300_000_000_000,  # 5 minutes
+                    "Delay": 10_000_000_000,     # 10 seconds
+                    "Mode": "delay",
+                },
+                "ReschedulePolicy": {
+                    "Attempts": 5,
+                    "Interval": 3600_000_000_000,  # 1 hour
+                    "Delay": 30_000_000_000,      # 30 seconds
+                    "DelayFunction": "exponential",
+                    "MaxDelay": 300_000_000_000,  # 5 minutes
+                    "Unlimited": False,
+                },
+            }
+        ],
+    }

atropos/slots/__init__.py

@@ -0,0 +1,20 @@
+"""
+Slot-based multiplexing for atropos-agent.
+
+Provides:
+- Slot: Isolated workspace for a single trajectory
+- SlotPool: Manages slots across Nomad allocations  
+- SandboxExecutor: Executes tools in sandbox containers
+"""
+
+from .executor import SandboxExecutor
+from .pool import SlotPool, SlotPoolConfig
+from .slot import Slot, SlotState
+
+__all__ = [
+    "Slot",
+    "SlotState",
+    "SlotPool",
+    "SlotPoolConfig",
+    "SandboxExecutor",
+]

atropos/slots/executor.py

@@ -0,0 +1,457 @@
+"""
+SandboxExecutor - HTTP client for sandbox container communication.
+
+Sends tool execution requests to sandbox_server.py running inside Nomad containers.
+Supports single and batch execution for efficiency.
+"""
+
+import asyncio
+import uuid
+from dataclasses import dataclass, field
+from typing import Any, Dict, List, Optional, Tuple
+
+import aiohttp
+
+from .slot import Slot, SlotState
+from ..tools.base import ToolCall, ToolResult
+
+
+@dataclass
+class ExecutionRequest:
+    """Request to execute a tool in a slot."""
+    slot: Slot
+    tool_name: str
+    args: Dict[str, Any]
+    execution_id: str = field(default_factory=lambda: str(uuid.uuid4()))
+    timeout: float = 30.0
+
+
+@dataclass
+class ExecutionResult:
+    """Result from sandbox execution."""
+    success: bool
+    output: str = ""
+    error: str = ""
+    execution_id: str = ""
+    slot_id: str = ""
+    metadata: Dict[str, Any] = field(default_factory=dict)
+    
+    def to_tool_result(self) -> ToolResult:
+        """Convert to ToolResult for agent consumption."""
+        return ToolResult(
+            success=self.success,
+            output=self.output,
+            error=self.error,
+            metadata=self.metadata,
+            uniq_id=self.execution_id,
+        )
+
+
+class SandboxExecutor:
+    """
+    HTTP client for executing tools in sandbox containers.
+    
+    Communicates with sandbox_server.py running inside Nomad allocations.
+    Supports both single execution and batched parallel execution.
+    
+    Usage:
+        executor = SandboxExecutor()
+        
+        # Single execution
+        result = await executor.execute(slot, "bash", {"command": "ls"})
+        
+        # Batch execution
+        results = await executor.execute_batch([
+            (slot1, "bash", {"command": "ls"}),
+            (slot2, "write_file", {"path": "test.txt", "content": "hello"}),
+        ])
+    """
+    
+    def __init__(
+        self,
+        timeout: float = 30.0,
+        max_retries: int = 3,
+        retry_delay: float = 1.0,
+    ):
+        self.timeout = aiohttp.ClientTimeout(total=timeout)
+        self.max_retries = max_retries
+        self.retry_delay = retry_delay
+        self._session: Optional[aiohttp.ClientSession] = None
+    
+    async def _get_session(self) -> aiohttp.ClientSession:
+        """Get or create HTTP session."""
+        if self._session is None or self._session.closed:
+            self._session = aiohttp.ClientSession(timeout=self.timeout)
+        return self._session
+    
+    async def close(self):
+        """Close HTTP session."""
+        if self._session and not self._session.closed:
+            await self._session.close()
+    
+    async def __aenter__(self):
+        return self
+    
+    async def __aexit__(self, exc_type, exc_val, exc_tb):
+        await self.close()
+    
+    async def execute(
+        self,
+        slot: Slot,
+        tool_name: str,
+        args: Dict[str, Any],
+        timeout: Optional[float] = None,
+    ) -> ExecutionResult:
+        """
+        Execute a tool in a slot's workspace.
+        
+        Args:
+            slot: Slot to execute in
+            tool_name: Name of tool (bash, read_file, write_file)
+            args: Tool arguments
+            timeout: Optional timeout override
+            
+        Returns:
+            ExecutionResult with output or error
+        """
+        execution_id = str(uuid.uuid4())
+        exec_timeout = timeout or self.timeout.total or 30.0
+        
+        # Mark slot as executing
+        original_state = slot.state
+        try:
+            if slot.state == SlotState.ACQUIRED:
+                slot.start_execution(execution_id)
+            
+            result = await self._send_execute_request(
+                container_addr=slot.container_addr,
+                slot_id=slot.slot_id,
+                tool_name=tool_name,
+                args=args,
+                execution_id=execution_id,
+                timeout=exec_timeout,
+            )
+            result.slot_id = slot.slot_id
+            return result
+            
+        finally:
+            # Restore slot state
+            if slot.state == SlotState.EXECUTING:
+                slot.end_execution()
+    
+    async def _send_execute_request(
+        self,
+        container_addr: str,
+        slot_id: str,
+        tool_name: str,
+        args: Dict[str, Any],
+        execution_id: str,
+        timeout: float,
+    ) -> ExecutionResult:
+        """Send execution request to sandbox server with retry logic."""
+        session = await self._get_session()
+        url = f"{container_addr}/execute"
+        
+        payload = {
+            "slot_id": slot_id,
+            "tool": tool_name,
+            "args": args,
+            "execution_id": execution_id,
+            "timeout": timeout,
+        }
+        
+        last_error = None
+        for attempt in range(self.max_retries):
+            try:
+                async with session.post(url, json=payload) as response:
+                    data = await response.json()
+                    
+                    return ExecutionResult(
+                        success=data.get("success", False),
+                        output=data.get("output", ""),
+                        error=data.get("error", ""),
+                        execution_id=data.get("execution_id", execution_id),
+                        metadata=data.get("metadata", {}),
+                    )
+                    
+            except aiohttp.ClientError as e:
+                last_error = str(e)
+                if attempt < self.max_retries - 1:
+                    await asyncio.sleep(self.retry_delay * (attempt + 1))
+                continue
+            except asyncio.TimeoutError:
+                last_error = f"Request timed out after {timeout}s"
+                break
+            except Exception as e:
+                last_error = str(e)
+                break
+        
+        return ExecutionResult(
+            success=False,
+            error=f"Failed after {self.max_retries} attempts: {last_error}",
+            execution_id=execution_id,
+        )
+    
+    async def execute_batch(
+        self,
+        requests: List[Tuple[Slot, str, Dict[str, Any]]],
+        timeout: Optional[float] = None,
+    ) -> List[ExecutionResult]:
+        """
+        Execute multiple tools in parallel across slots.
+        
+        This is the key optimization - we batch tool calls to maximize
+        container utilization while agents are waiting for LLM responses.
+        
+        Args:
+            requests: List of (slot, tool_name, args) tuples
+            timeout: Optional timeout override
+            
+        Returns:
+            List of ExecutionResults in same order as requests
+        """
+        if not requests:
+            return []
+        
+        # Group requests by container address for batch API
+        by_container: Dict[str, List[Tuple[int, Slot, str, Dict[str, Any], str]]] = {}
+        
+        for idx, (slot, tool_name, args) in enumerate(requests):
+            execution_id = str(uuid.uuid4())
+            container = slot.container_addr
+            
+            if container not in by_container:
+                by_container[container] = []
+            by_container[container].append((idx, slot, tool_name, args, execution_id))
+            
+            # Mark slots as executing
+            if slot.state == SlotState.ACQUIRED:
+                slot.start_execution(execution_id)
+        
+        # Execute batches in parallel
+        exec_timeout = timeout or self.timeout.total or 30.0
+        batch_tasks = []
+        
+        for container_addr, batch_requests in by_container.items():
+            task = self._send_batch_request(
+                container_addr=container_addr,
+                batch_requests=batch_requests,
+                timeout=exec_timeout,
+            )
+            batch_tasks.append(task)
+        
+        # Gather all batch results
+        batch_results = await asyncio.gather(*batch_tasks, return_exceptions=True)
+        
+        # Collect results in original order
+        results: List[Optional[ExecutionResult]] = [None] * len(requests)
+        
+        for batch_result in batch_results:
+            if isinstance(batch_result, Exception):
+                # Mark all in this batch as failed
+                continue
+            
+            for idx, result in batch_result:
+                results[idx] = result
+        
+        # Fill in any missing results
+        for idx, result in enumerate(results):
+            if result is None:
+                slot, tool_name, args = requests[idx]
+                results[idx] = ExecutionResult(
+                    success=False,
+                    error="Batch execution failed",
+                    slot_id=slot.slot_id,
+                )
+        
+        # End execution on all slots
+        for slot, _, _ in requests:
+            if slot.state == SlotState.EXECUTING:
+                slot.end_execution()
+        
+        return results  # type: ignore
+    
+    async def _send_batch_request(
+        self,
+        container_addr: str,
+        batch_requests: List[Tuple[int, Slot, str, Dict[str, Any], str]],
+        timeout: float,
+    ) -> List[Tuple[int, ExecutionResult]]:
+        """Send batch execution request to a single container."""
+        session = await self._get_session()
+        url = f"{container_addr}/batch"
+        
+        # Build batch payload
+        payload = [
+            {
+                "slot_id": slot.slot_id,
+                "tool": tool_name,
+                "args": args,
+                "execution_id": execution_id,
+                "timeout": timeout,
+            }
+            for _, slot, tool_name, args, execution_id in batch_requests
+        ]
+        
+        try:
+            async with session.post(url, json=payload) as response:
+                data = await response.json()
+                
+                if not isinstance(data, list):
+                    raise ValueError(f"Expected list response, got {type(data)}")
+                
+                results = []
+                for i, (idx, slot, _, _, execution_id) in enumerate(batch_requests):
+                    if i < len(data):
+                        item = data[i]
+                        result = ExecutionResult(
+                            success=item.get("success", False),
+                            output=item.get("output", ""),
+                            error=item.get("error", ""),
+                            execution_id=item.get("execution_id", execution_id),
+                            slot_id=slot.slot_id,
+                            metadata=item.get("metadata", {}),
+                        )
+                    else:
+                        result = ExecutionResult(
+                            success=False,
+                            error="Missing result in batch response",
+                            execution_id=execution_id,
+                            slot_id=slot.slot_id,
+                        )
+                    results.append((idx, result))
+                
+                return results
+                
+        except Exception as e:
+            # Return error for all requests in batch
+            return [
+                (idx, ExecutionResult(
+                    success=False,
+                    error=str(e),
+                    execution_id=execution_id,
+                    slot_id=slot.slot_id,
+                ))
+                for idx, slot, _, _, execution_id in batch_requests
+            ]
+    
+    async def reset_slot(self, slot: Slot) -> ExecutionResult:
+        """
+        Reset a slot's workspace (delete all files).
+        
+        Useful when reusing a slot for a new trajectory.
+        """
+        session = await self._get_session()
+        url = f"{slot.container_addr}/reset"
+        
+        try:
+            async with session.post(url, json={"slot_id": slot.slot_id}) as response:
+                data = await response.json()
+                return ExecutionResult(
+                    success=data.get("success", False),
+                    output=data.get("output", ""),
+                    error=data.get("error", ""),
+                    slot_id=slot.slot_id,
+                )
+        except Exception as e:
+            return ExecutionResult(
+                success=False,
+                error=str(e),
+                slot_id=slot.slot_id,
+            )
+    
+    async def health_check(self, container_addr: str) -> bool:
+        """Check if a sandbox container is healthy."""
+        session = await self._get_session()
+        url = f"{container_addr}/health"
+        
+        try:
+            async with session.get(url) as response:
+                data = await response.json()
+                return data.get("status") == "ok"
+        except Exception:
+            return False
+    
+    async def get_container_status(
+        self, 
+        container_addr: str
+    ) -> Optional[Dict[str, Any]]:
+        """Get status info from a sandbox container."""
+        session = await self._get_session()
+        url = f"{container_addr}/health"
+        
+        try:
+            async with session.get(url) as response:
+                return await response.json()
+        except Exception:
+            return None
+
+    # -------------------------------------------------------------------------
+    # Artifact helpers (optional)
+    # -------------------------------------------------------------------------
+
+    async def _post_json(
+        self,
+        url: str,
+        payload: Dict[str, Any],
+        timeout: Optional[float] = None,
+    ) -> Dict[str, Any]:
+        session = await self._get_session()
+        try:
+            async with session.post(url, json=payload, timeout=timeout) as response:
+                data = await response.json()
+                if isinstance(data, dict):
+                    data.setdefault("http_status", response.status)
+                    return data
+                return {"success": False, "error": f"Unexpected response type: {type(data)}", "http_status": response.status}
+        except Exception as e:
+            return {"success": False, "error": str(e)}
+
+    async def read_artifact(
+        self,
+        slot: Slot,
+        path: str,
+        *,
+        encoding: str = "text",
+        max_bytes: Optional[int] = None,
+        include_sha256: bool = False,
+        timeout: Optional[float] = None,
+    ) -> Dict[str, Any]:
+        url = f"{slot.container_addr}/artifacts/read"
+        payload: Dict[str, Any] = {"slot_id": slot.slot_id, "path": path, "encoding": encoding, "include_sha256": include_sha256}
+        if max_bytes is not None:
+            payload["max_bytes"] = max_bytes
+        return await self._post_json(url, payload, timeout=timeout)
+
+    async def list_artifacts(
+        self,
+        slot: Slot,
+        path: str = ".",
+        *,
+        recursive: bool = False,
+        max_entries: Optional[int] = None,
+        timeout: Optional[float] = None,
+    ) -> Dict[str, Any]:
+        url = f"{slot.container_addr}/artifacts/list"
+        payload: Dict[str, Any] = {"slot_id": slot.slot_id, "path": path, "recursive": recursive}
+        if max_entries is not None:
+            payload["max_entries"] = max_entries
+        return await self._post_json(url, payload, timeout=timeout)
+
+    async def archive_artifacts(
+        self,
+        slot: Slot,
+        path: str = ".",
+        *,
+        archive_format: str = "tar.gz",
+        max_bytes: Optional[int] = None,
+        max_entries: Optional[int] = None,
+        timeout: Optional[float] = None,
+    ) -> Dict[str, Any]:
+        url = f"{slot.container_addr}/artifacts/archive"
+        payload: Dict[str, Any] = {"slot_id": slot.slot_id, "path": path, "format": archive_format}
+        if max_bytes is not None:
+            payload["max_bytes"] = max_bytes
+        if max_entries is not None:
+            payload["max_entries"] = max_entries
+        return await self._post_json(url, payload, timeout=timeout)

atropos/slots/pool.py

@@ -0,0 +1,659 @@
+"""
+SlotPool - Manages slots across Nomad allocations.
+
+The SlotPool is the core abstraction for slot-based multiplexing:
+- Tracks available/acquired slots across containers
+- Handles slot acquisition and release
+- Auto-scales Nomad job count based on demand
+- Provides batched tool execution
+"""
+
+import asyncio
+import logging
+import os
+import subprocess
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Any, Dict, List, Optional, Tuple
+
+from ..nomad.client import (
+    Allocation,
+    AllocationStatus,
+    NomadClient,
+    create_sandbox_job,
+)
+from .executor import ExecutionResult, SandboxExecutor
+from .slot import Slot, SlotState, create_slots_for_allocation
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class SlotPoolConfig:
+    """Configuration for SlotPool."""
+    
+    # Nomad settings
+    nomad_address: str = "http://localhost:4646"
+    job_id: str = "atropos-sandbox"
+    datacenter: str = "dc1"
+    
+    # Container settings
+    image: str = "atropos-sandbox:local"  # Use :local tag to avoid registry pull
+    slots_per_container: int = 10
+    privileged: bool = False
+    cpu: int = 500  # MHz
+    memory: int = 512  # MB
+    
+    # Driver selection: "docker" or "singularity"
+    driver: str = "docker"
+    # Path to .sif file for singularity driver (required if driver="singularity")
+    singularity_image: Optional[str] = None
+    
+    # Scaling settings
+    min_containers: int = 1
+    max_containers: int = 10
+    
+    # Timeouts
+    acquire_timeout: float = 30.0  # Seconds between acquire polls (also triggers scale-up attempts)
+    health_check_interval: float = 30.0  # Seconds between health checks
+    scale_cooldown: float = 60.0  # Seconds between scale operations
+
+    # Job lifecycle
+    purge_job_on_start: bool = False  # Purge any pre-existing job before starting (local dev/training friendly)
+
+    # Local Docker image convenience (macOS/Nomad dev mode)
+    auto_build_local_image: bool = True  # If image endswith :local and is missing, build it from the bundled Dockerfile.
+    dockerfile_path: Optional[str] = None  # Override Dockerfile path (default: Hermes-Agent/atropos/Dockerfile).
+    docker_build_context: Optional[str] = None  # Override build context (default: Hermes-Agent/atropos).
+
+
+class SlotPool:
+    """
+    Manages a pool of slots across Nomad allocations.
+    
+    The SlotPool:
+    - Deploys sandbox containers to Nomad
+    - Tracks slots across all running containers
+    - Handles slot acquisition/release
+    - Auto-scales based on demand
+    - Provides batched execution via SandboxExecutor
+    
+    Usage:
+        config = SlotPoolConfig(
+            nomad_address="http://localhost:4646",
+            job_id="my-sandbox",
+            slots_per_container=10,
+        )
+        
+        pool = SlotPool(config)
+        await pool.start()
+        
+        # Acquire a slot
+        slot = await pool.acquire()
+        
+        # Execute tool
+        result = await pool.execute(slot, "bash", {"command": "ls"})
+        
+        # Release slot
+        await pool.release(slot)
+        
+        # Shutdown
+        await pool.stop()
+    """
+    
+    def __init__(self, config: Optional[SlotPoolConfig] = None):
+        self.config = config or SlotPoolConfig()
+        
+        # Nomad client
+        self.nomad = NomadClient(address=self.config.nomad_address)
+        
+        # Sandbox executor for tool execution
+        self.executor = SandboxExecutor()
+        
+        # Slot tracking
+        self._slots: Dict[str, Slot] = {}  # slot_key -> Slot
+        self._available_queue: asyncio.Queue[str] = asyncio.Queue()
+        self._lock = asyncio.Lock()
+        self._scale_lock = asyncio.Lock()
+        
+        # State
+        self._started = False
+        self._health_task: Optional[asyncio.Task] = None
+        self._scale_task: Optional[asyncio.Task] = None
+        self._last_scale_time = 0.0
+
+    def _default_dockerfile_path(self) -> Path:
+        # Hermes-Agent/atropos/Dockerfile lives next to this module in source checkouts.
+        return Path(__file__).resolve().parents[1] / "Dockerfile"
+
+    def _default_build_context(self) -> Path:
+        return Path(__file__).resolve().parents[1]
+
+    def _docker_image_exists(self, image: str) -> bool:
+        try:
+            proc = subprocess.run(
+                ["docker", "image", "inspect", image],
+                stdout=subprocess.DEVNULL,
+                stderr=subprocess.DEVNULL,
+                check=False,
+                env={**os.environ, "DOCKER_CLI_HINTS": "false"},
+            )
+            return proc.returncode == 0
+        except FileNotFoundError:
+            return False
+
+    def _try_build_local_image(self, image: str) -> None:
+        dockerfile = Path(self.config.dockerfile_path) if self.config.dockerfile_path else self._default_dockerfile_path()
+        context = Path(self.config.docker_build_context) if self.config.docker_build_context else self._default_build_context()
+
+        if not dockerfile.exists():
+            raise RuntimeError(
+                f"Sandbox Dockerfile not found at {dockerfile}. "
+                "Build the sandbox image manually or set --env.purge_job_on_start false and provide a non-local image."
+            )
+        if not context.exists():
+            raise RuntimeError(f"Docker build context not found at {context}")
+
+        # Prefer buildx+--load to ensure the image ends up in the local daemon (required by Nomad's docker driver).
+        buildx_cmd = [
+            "docker",
+            "buildx",
+            "build",
+            "--load",
+            "-t",
+            image,
+            "-f",
+            str(dockerfile),
+            str(context),
+        ]
+        proc = subprocess.run(buildx_cmd, check=False, env={**os.environ, "DOCKER_CLI_HINTS": "false"})
+        if proc.returncode == 0:
+            return
+
+        # Fallback to classic docker build if buildx isn't available.
+        build_cmd = ["docker", "build", "-t", image, "-f", str(dockerfile), str(context)]
+        proc2 = subprocess.run(build_cmd, check=False, env={**os.environ, "DOCKER_CLI_HINTS": "false"})
+        if proc2.returncode != 0:
+            raise RuntimeError(
+                f"Failed to build local sandbox image {image}. "
+                f"Tried: {' '.join(buildx_cmd)} and {' '.join(build_cmd)}"
+            )
+
+    def _ensure_local_image(self) -> None:
+        image = (self.config.image or "").strip()
+        if not image.endswith(":local"):
+            return
+        if not self.config.auto_build_local_image:
+            return
+
+        if self._docker_image_exists(image):
+            return
+
+        logger.info(f"Local sandbox image {image} not found; building it now...")
+        self._try_build_local_image(image)
+
+    def _slot_key(self, alloc_id: str, slot_id: str) -> str:
+        """Generate unique key for a slot."""
+        return f"{alloc_id}:{slot_id}"
+    
+    @property
+    def total_slots(self) -> int:
+        """Total number of slots in pool."""
+        return len(self._slots)
+    
+    @property
+    def available_slots(self) -> int:
+        """Number of available slots."""
+        return sum(1 for s in self._slots.values() if s.is_available)
+    
+    @property
+    def acquired_slots(self) -> int:
+        """Number of acquired slots."""
+        return sum(1 for s in self._slots.values() if s.is_acquired)
+    
+    async def start(self) -> None:
+        """
+        Start the slot pool.
+        
+        - Checks if Nomad is healthy
+        - Deploys sandbox job if not running
+        - Discovers existing allocations
+        - Starts health check background task
+        """
+        if self._started:
+            return
+        
+        logger.info(f"Starting SlotPool (job_id={self.config.job_id})")
+
+        try:
+            # Make sure local sandbox images exist before Nomad tries to pull them.
+            # This is a common footgun in macOS dev mode with :local tags.
+            self._ensure_local_image()
+
+            # Check Nomad health
+            if not await self.nomad.is_healthy():
+                raise RuntimeError(f"Nomad is not reachable at {self.config.nomad_address}")
+
+            if self.config.purge_job_on_start:
+                logger.info(f"Purging any existing Nomad job: {self.config.job_id}")
+                await self.nomad.stop_job(self.config.job_id, purge=True)
+
+            # Check if job exists (after optional purge)
+            job = await self.nomad.get_job(self.config.job_id)
+
+            if job is None:
+                # Deploy new job
+                logger.info(f"Deploying sandbox job: {self.config.job_id} (driver={self.config.driver})")
+                job_spec = create_sandbox_job(
+                    job_id=self.config.job_id,
+                    image=self.config.image,
+                    count=self.config.min_containers,
+                    slots_per_container=self.config.slots_per_container,
+                    privileged=self.config.privileged,
+                    cpu=self.config.cpu,
+                    memory=self.config.memory,
+                    datacenter=self.config.datacenter,
+                    driver=self.config.driver,
+                    singularity_image=self.config.singularity_image,
+                )
+                result = await self.nomad.submit_job(job_spec)
+                if "error" in result:
+                    raise RuntimeError(f"Failed to submit job: {result}")
+
+            # Wait for allocations to be running (even if the job already existed).
+            await self._wait_for_healthy_allocations(self.config.min_containers)
+
+            # Discover existing allocations and slots
+            await self._refresh_slots()
+
+            # Start health check task
+            self._health_task = asyncio.create_task(self._health_check_loop())
+
+            self._started = True
+            logger.info(f"SlotPool started: {self.total_slots} slots available")
+        except Exception:
+            # Ensure aiohttp sessions are not leaked if we fail to start.
+            await self.stop(purge_job=False)
+            raise
+    
+    async def stop(self, purge_job: bool = False) -> None:
+        """
+        Stop the slot pool.
+        
+        Args:
+            purge_job: If True, also stop the Nomad job
+        """
+        logger.info("Stopping SlotPool")
+
+        # Cancel health check task
+        if self._health_task:
+            self._health_task.cancel()
+            try:
+                await self._health_task
+            except asyncio.CancelledError:
+                pass
+            finally:
+                self._health_task = None
+
+        if self._scale_task:
+            self._scale_task.cancel()
+            try:
+                await self._scale_task
+            except asyncio.CancelledError:
+                pass
+            finally:
+                self._scale_task = None
+
+        # Optionally stop the job (do this even if start() never completed).
+        if purge_job:
+            logger.info(f"Stopping Nomad job: {self.config.job_id}")
+            await self.nomad.stop_job(self.config.job_id, purge=True)
+
+        # Close connections
+        await self.executor.close()
+        await self.nomad.close()
+
+        self._started = False
+        self._slots.clear()
+
+        # Clear the queue
+        while not self._available_queue.empty():
+            try:
+                self._available_queue.get_nowait()
+            except asyncio.QueueEmpty:
+                break
+    
+    async def acquire(self, trajectory_id: Optional[str] = None) -> Slot:
+        """
+        Acquire an available slot.
+        
+        If no slots are available, waits up to acquire_timeout seconds.
+        If still no slots, attempts to scale up.
+        
+        Args:
+            trajectory_id: Optional ID of trajectory acquiring the slot
+            
+        Returns:
+            Acquired Slot
+            
+        Raises:
+            asyncio.TimeoutError: If no slot becomes available
+        """
+        if not self._started:
+            raise RuntimeError("SlotPool not started")
+
+        while True:
+            try:
+                # Try to get an available slot
+                slot_key = await asyncio.wait_for(
+                    self._available_queue.get(),
+                    timeout=self.config.acquire_timeout,
+                )
+            except asyncio.TimeoutError:
+                # Try to scale up, but keep waiting even if scaling isn't possible.
+                # In practice, slots may become available shortly (e.g. contention),
+                # and scaling may be temporarily blocked by Nomad deployments.
+                await self._try_scale_up()
+                continue
+
+            slot = self._slots.get(slot_key)
+            if slot is None:
+                # Slot was removed; discard stale queue entry and retry.
+                continue
+
+            try:
+                slot.acquire(trajectory_id)
+            except RuntimeError:
+                # Slot isn't actually available (e.g. duplicate queue entry); retry.
+                continue
+
+            logger.debug(f"Acquired slot {slot.slot_id} (alloc={slot.alloc_id[:8]})")
+            return slot
+    
+    async def release(self, slot: Slot, reset_workspace: bool = False) -> None:
+        """
+        Release a slot back to the pool.
+        
+        Args:
+            slot: Slot to release
+            reset_workspace: If True, clear the workspace files
+        """
+        slot_key = self._slot_key(slot.alloc_id, slot.slot_id)
+        
+        if slot_key not in self._slots:
+            logger.warning(f"Releasing unknown slot: {slot_key}")
+            return
+        
+        # Optionally reset workspace
+        if reset_workspace:
+            await self.executor.reset_slot(slot)
+        
+        slot.release()
+        await self._available_queue.put(slot_key)
+        
+        logger.debug(f"Released slot {slot.slot_id}")
+    
+    async def execute(
+        self,
+        slot: Slot,
+        tool_name: str,
+        args: Dict[str, Any],
+        timeout: Optional[float] = None,
+    ) -> ExecutionResult:
+        """
+        Execute a tool in a slot's workspace.
+        
+        Args:
+            slot: Slot to execute in
+            tool_name: Name of tool (bash, read_file, write_file)
+            args: Tool arguments
+            timeout: Optional timeout override
+            
+        Returns:
+            ExecutionResult
+        """
+        return await self.executor.execute(slot, tool_name, args, timeout)
+    
+    async def execute_batch(
+        self,
+        requests: List[Tuple[Slot, str, Dict[str, Any]]],
+        timeout: Optional[float] = None,
+    ) -> List[ExecutionResult]:
+        """
+        Execute multiple tools in parallel.
+        
+        This is the key optimization - batch execution across multiple slots
+        maximizes container utilization.
+        
+        Args:
+            requests: List of (slot, tool_name, args) tuples
+            timeout: Optional timeout override
+            
+        Returns:
+            List of ExecutionResults in same order
+        """
+        return await self.executor.execute_batch(requests, timeout)
+    
+    async def _refresh_slots(self) -> None:
+        """Refresh slot inventory from Nomad allocations."""
+        async with self._lock:
+            allocs = await self.nomad.get_job_allocations(self.config.job_id)
+            
+            # Track which slots we've seen
+            seen_keys = set()
+            
+            for alloc in allocs:
+                if alloc.status != AllocationStatus.RUNNING:
+                    continue
+                
+                if not alloc.http_address:
+                    continue
+                
+                # Check container health
+                healthy = await self.executor.health_check(alloc.http_address)
+                if not healthy:
+                    continue
+                
+                # Create slots for this allocation
+                for i in range(self.config.slots_per_container):
+                    slot_id = f"slot_{i}"
+                    slot_key = self._slot_key(alloc.id, slot_id)
+                    seen_keys.add(slot_key)
+                    
+                    if slot_key not in self._slots:
+                        # New slot
+                        slot = Slot(
+                            slot_id=slot_id,
+                            alloc_id=alloc.id,
+                            container_addr=alloc.http_address,
+                        )
+                        self._slots[slot_key] = slot
+                        await self._available_queue.put(slot_key)
+                        logger.debug(f"Added slot: {slot_key}")
+            
+            # Remove slots from dead allocations
+            for slot_key in list(self._slots.keys()):
+                if slot_key not in seen_keys:
+                    slot = self._slots.pop(slot_key)
+                    logger.debug(f"Removed slot: {slot_key}")
+    
+    async def _wait_for_healthy_allocations(
+        self, 
+        min_count: int, 
+        timeout: float = 120.0
+    ) -> None:
+        """Wait for allocations to become healthy."""
+        import time
+        start = time.time()
+
+        def _summarize_alloc_detail(detail: Dict[str, Any]) -> str:
+            task_states = detail.get("TaskStates") or {}
+            parts: List[str] = []
+            if isinstance(task_states, dict):
+                for task_name, st in task_states.items():
+                    events = (st or {}).get("Events") or []
+                    if isinstance(events, list) and events:
+                        # Include a few recent events; the latest can be a generic restart message
+                        # while the true root cause is slightly earlier (e.g. image pull failure).
+                        recent = events[-3:]
+                        msgs: List[str] = []
+                        for ev in recent:
+                            desc = ev.get("DisplayMessage") or ev.get("Message") or ev.get("Type") or ""
+                            if desc:
+                                msgs.append(desc)
+                        if msgs:
+                            parts.append(f"{task_name}: " + " | ".join(msgs))
+            return "; ".join(parts)
+
+        def _alloc_events_lower(detail: Dict[str, Any]) -> str:
+            task_states = detail.get("TaskStates") or {}
+            texts: List[str] = []
+            if isinstance(task_states, dict):
+                for _task_name, st in task_states.items():
+                    events = (st or {}).get("Events") or []
+                    if isinstance(events, list):
+                        for ev in events[-10:]:
+                            desc = ev.get("DisplayMessage") or ev.get("Message") or ev.get("Type") or ""
+                            if desc:
+                                texts.append(desc)
+            return " ".join(texts).lower()
+        
+        while time.time() - start < timeout:
+            allocs = await self.nomad.get_job_allocations(self.config.job_id)
+            
+            healthy_count = 0
+            for alloc in allocs:
+                if alloc.status == AllocationStatus.RUNNING and alloc.http_address:
+                    if await self.executor.health_check(alloc.http_address):
+                        healthy_count += 1
+
+                # Fast-fail on obvious driver/image errors to avoid waiting out the full timeout.
+                if alloc.id:
+                    detail = await self.nomad.get_allocation(alloc.id)
+                    if isinstance(detail, dict):
+                        summary = _summarize_alloc_detail(detail)
+                        lowered = _alloc_events_lower(detail) or summary.lower()
+                        if "failed to pull" in lowered or "pull access denied" in lowered:
+                            raise RuntimeError(
+                                "Nomad allocation failed to start due to a Docker image pull error. "
+                                f"Allocation {alloc.id[:8]}: {summary}\n"
+                                "If you're using a local image tag (e.g. `atropos-sandbox:local`) on macOS, "
+                                "make sure the image is loaded into Docker, e.g.:\n"
+                                "  docker buildx build --load -t atropos-sandbox:local -f Hermes-Agent/atropos/Dockerfile Hermes-Agent/atropos"
+                            )
+                        if "exceeded allowed attempts" in lowered:
+                            raise RuntimeError(
+                                "Nomad allocation is crash-looping and has entered restart backoff. "
+                                f"Allocation {alloc.id[:8]}: {summary}\n"
+                                "Inspect logs with:\n"
+                                f"  nomad alloc logs -stderr -task sandbox-server {alloc.id}\n"
+                                "Common causes include: missing local Docker image tag, container entrypoint error, "
+                                "or sandbox-server startup failure."
+                            )
+            
+            if healthy_count >= min_count:
+                return
+            
+            await asyncio.sleep(2.0)
+
+        # Timed out: include allocation status detail to help debugging.
+        allocs = await self.nomad.get_job_allocations(self.config.job_id)
+        alloc_lines: List[str] = []
+        for alloc in allocs[:10]:
+            addr = alloc.http_address or "-"
+            line = f"{alloc.id[:8]} status={alloc.status.value} http={addr}"
+            detail = await self.nomad.get_allocation(alloc.id)
+            if isinstance(detail, dict):
+                summary = _summarize_alloc_detail(detail)
+                if summary:
+                    line += f" detail={summary}"
+            alloc_lines.append(line)
+
+        hint = (
+            "Timed out waiting for healthy sandbox allocations.\n"
+            f"Job: {self.config.job_id}, desired_healthy: {min_count}\n"
+            "Allocations:\n  - " + "\n  - ".join(alloc_lines)
+        )
+        raise RuntimeError(hint)
+    
+    async def _try_scale_up(self) -> bool:
+        """Attempt to scale up the job."""
+        import time
+
+        async with self._scale_lock:
+            # Check cooldown
+            if time.time() - self._last_scale_time < self.config.scale_cooldown:
+                return False
+
+            # Check max containers
+            status = await self.nomad.get_job_status(self.config.job_id)
+            if status is None:
+                return False
+
+            current_count = status.count
+            if current_count >= self.config.max_containers:
+                logger.warning(f"Cannot scale up: already at max ({self.config.max_containers})")
+                return False
+
+            # Scale up
+            new_count = min(current_count + 1, self.config.max_containers)
+            logger.info(f"Scaling up from {current_count} to {new_count} containers")
+
+            scale_resp = await self.nomad.scale_job(
+                self.config.job_id,
+                count=new_count,
+                task_group="sandbox",
+            )
+
+            # Nomad may return non-JSON errors (e.g. plain text) with a status field.
+            if isinstance(scale_resp, dict) and scale_resp.get("status", 200) >= 400:
+                logger.warning(f"Scale request rejected: {scale_resp}")
+                self._last_scale_time = time.time()
+                return False
+
+            self._last_scale_time = time.time()
+
+            # Wait for new allocation in the background so contended acquires can still
+            # make progress (e.g. by grabbing slots released by other trajectories).
+            if self._scale_task is None or self._scale_task.done():
+                self._scale_task = asyncio.create_task(self._wait_for_scale(new_count))
+
+            return True
+
+    async def _wait_for_scale(self, desired_count: int) -> None:
+        try:
+            await self._wait_for_healthy_allocations(desired_count, timeout=60.0)
+            await self._refresh_slots()
+        except asyncio.CancelledError:
+            raise
+        except Exception as e:
+            logger.error(f"Failed to scale up: {e}")
+    
+    async def _health_check_loop(self) -> None:
+        """Background task to monitor container health."""
+        while True:
+            try:
+                await asyncio.sleep(self.config.health_check_interval)
+                await self._refresh_slots()
+            except asyncio.CancelledError:
+                break
+            except Exception as e:
+                logger.error(f"Health check error: {e}")
+    
+    def get_stats(self) -> Dict[str, Any]:
+        """Get pool statistics."""
+        slots_by_state = {}
+        for slot in self._slots.values():
+            state = slot.state.value
+            slots_by_state[state] = slots_by_state.get(state, 0) + 1
+
+        container_count = len({s.alloc_id for s in self._slots.values()}) if self._slots else 0
+        
+        return {
+            "total_slots": self.total_slots,
+            "available_slots": self.available_slots,
+            "acquired_slots": self.acquired_slots,
+            "containers": container_count,
+            "slots_by_state": slots_by_state,
+            "started": self._started,
+        }

atropos/slots/slot.py

@@ -0,0 +1,159 @@
+"""
+Slot abstraction for atropos-agent.
+
+A Slot represents an isolated workspace for a single agent trajectory.
+Slots are hosted on Nomad allocations and provide workspace isolation
+via filesystem directories.
+"""
+
+from dataclasses import dataclass, field
+from enum import Enum
+from typing import Any, Dict, Optional
+import uuid
+
+
+class SlotState(Enum):
+    """State of a slot in the pool."""
+    AVAILABLE = "available"      # Ready to be acquired
+    ACQUIRED = "acquired"        # Assigned to a trajectory
+    EXECUTING = "executing"      # Currently executing a tool
+    RELEASING = "releasing"      # Being released back to pool
+    ERROR = "error"              # In error state
+
+
+@dataclass
+class Slot:
+    """
+    An isolated workspace for a single agent trajectory.
+    
+    Slots are the unit of scheduling - each trajectory runs in its own slot,
+    with an isolated workspace directory. Multiple slots share a container.
+    
+    Attributes:
+        slot_id: Unique identifier for this slot (e.g., "slot_0")
+        alloc_id: Nomad allocation ID hosting this slot
+        container_addr: HTTP address of the sandbox server (e.g., "http://10.0.0.1:8080")
+        workspace_dir: Path to workspace in container (e.g., "/data/slot_0")
+        state: Current state of the slot
+        trajectory_id: ID of trajectory currently using this slot (if acquired)
+        metadata: Additional metadata
+    """
+    slot_id: str
+    alloc_id: str
+    container_addr: str
+    workspace_dir: str = ""
+    state: SlotState = SlotState.AVAILABLE
+    trajectory_id: Optional[str] = None
+    metadata: Dict[str, Any] = field(default_factory=dict)
+    
+    def __post_init__(self):
+        """Set default workspace_dir if not provided."""
+        if not self.workspace_dir:
+            self.workspace_dir = f"/data/{self.slot_id}"
+    
+    @property
+    def is_available(self) -> bool:
+        """Check if slot is available for acquisition."""
+        return self.state == SlotState.AVAILABLE
+    
+    @property
+    def is_acquired(self) -> bool:
+        """Check if slot is currently acquired."""
+        return self.state in (SlotState.ACQUIRED, SlotState.EXECUTING)
+    
+    def acquire(self, trajectory_id: Optional[str] = None) -> None:
+        """
+        Mark slot as acquired by a trajectory.
+        
+        Args:
+            trajectory_id: Optional ID of acquiring trajectory
+        """
+        if not self.is_available:
+            raise RuntimeError(f"Cannot acquire slot {self.slot_id}: state is {self.state}")
+        
+        self.state = SlotState.ACQUIRED
+        self.trajectory_id = trajectory_id or str(uuid.uuid4())
+    
+    def start_execution(self, execution_id: Optional[str] = None) -> None:
+        """Mark slot as executing."""
+        if self.state != SlotState.ACQUIRED:
+            raise RuntimeError(f"Cannot start execution on slot {self.slot_id}: state is {self.state}")
+        
+        self.state = SlotState.EXECUTING
+        if execution_id:
+            self.metadata["current_execution_id"] = execution_id
+    
+    def end_execution(self) -> None:
+        """Mark execution as complete, return to acquired state."""
+        if self.state != SlotState.EXECUTING:
+            raise RuntimeError(f"Cannot end execution on slot {self.slot_id}: state is {self.state}")
+        
+        self.state = SlotState.ACQUIRED
+        self.metadata.pop("current_execution_id", None)
+    
+    def release(self) -> None:
+        """Release slot back to available state."""
+        self.state = SlotState.AVAILABLE
+        self.trajectory_id = None
+        self.metadata.pop("current_execution_id", None)
+    
+    def mark_error(self, error: str) -> None:
+        """Mark slot as in error state."""
+        self.state = SlotState.ERROR
+        self.metadata["error"] = error
+    
+    def to_dict(self) -> Dict[str, Any]:
+        """Convert to dictionary for serialization."""
+        return {
+            "slot_id": self.slot_id,
+            "alloc_id": self.alloc_id,
+            "container_addr": self.container_addr,
+            "workspace_dir": self.workspace_dir,
+            "state": self.state.value,
+            "trajectory_id": self.trajectory_id,
+            "metadata": self.metadata,
+        }
+    
+    @classmethod
+    def from_dict(cls, data: Dict[str, Any]) -> "Slot":
+        """Create from dictionary."""
+        return cls(
+            slot_id=data["slot_id"],
+            alloc_id=data["alloc_id"],
+            container_addr=data["container_addr"],
+            workspace_dir=data.get("workspace_dir", ""),
+            state=SlotState(data.get("state", "available")),
+            trajectory_id=data.get("trajectory_id"),
+            metadata=data.get("metadata", {}),
+        )
+    
+    def __repr__(self) -> str:
+        return f"Slot({self.slot_id}, state={self.state.value}, alloc={self.alloc_id[:8]}...)"
+
+
+def create_slots_for_allocation(
+    alloc_id: str,
+    container_addr: str,
+    num_slots: int = 10,
+) -> list["Slot"]:
+    """
+    Create slots for a Nomad allocation.
+    
+    Args:
+        alloc_id: Nomad allocation ID
+        container_addr: HTTP address of sandbox server
+        num_slots: Number of slots to create
+        
+    Returns:
+        List of Slot objects
+    """
+    slots = []
+    for i in range(num_slots):
+        slot_id = f"slot_{i}"
+        slots.append(Slot(
+            slot_id=slot_id,
+            alloc_id=alloc_id,
+            container_addr=container_addr,
+            workspace_dir=f"/data/{slot_id}",
+        ))
+    return slots

atropos/terminal/__init__.py

@@ -0,0 +1,2 @@
+"""Terminal helpers for stateful sandbox interactions."""
+

atropos/terminal/asciinema_stream.py

@@ -0,0 +1,115 @@
+from __future__ import annotations
+
+import json
+from typing import Any
+
+import pyte
+
+
+class AsciinemaStreamDecoder:
+    def __init__(self, *, default_width: int = 80, default_height: int = 24) -> None:
+        self._default_width = max(1, int(default_width))
+        self._default_height = max(1, int(default_height))
+        self._buffer = ""
+        self._has_header = False
+        self.width = self._default_width
+        self.height = self._default_height
+        self._screen = pyte.Screen(self.width, self.height)
+        self._stream = pyte.Stream(self._screen)
+
+    def reset(self) -> None:
+        self._buffer = ""
+        self._has_header = False
+        self.width = self._default_width
+        self.height = self._default_height
+        self._screen = pyte.Screen(self.width, self.height)
+        self._stream = pyte.Stream(self._screen)
+
+    def feed(self, chunk: str | bytes) -> None:
+        if not chunk:
+            return
+        if isinstance(chunk, bytes):
+            chunk = chunk.decode("utf-8", errors="replace")
+        self._buffer += chunk
+        while True:
+            line, sep, rest = self._buffer.partition("\n")
+            if not sep:
+                break
+            self._buffer = rest
+            line = line.strip()
+            if not line:
+                continue
+            parsed = self._parse_json_line(line)
+            if parsed is None:
+                continue
+            if not self._has_header:
+                if isinstance(parsed, dict):
+                    self._init_from_header(parsed)
+                    continue
+                if isinstance(parsed, list):
+                    self._has_header = True
+                    self._apply_event(parsed)
+                    continue
+                continue
+            if isinstance(parsed, list):
+                self._apply_event(parsed)
+
+    def render(self) -> str:
+        return "\n".join(self._screen.display)
+
+    def _parse_json_line(self, line: str) -> Any | None:
+        try:
+            return json.loads(line)
+        except json.JSONDecodeError:
+            return None
+
+    def _init_from_header(self, header: dict[str, Any]) -> None:
+        width = _coerce_int(
+            header.get("width") or header.get("columns") or header.get("cols"),
+            self._default_width,
+        )
+        height = _coerce_int(
+            header.get("height") or header.get("rows") or header.get("lines"),
+            self._default_height,
+        )
+        self.width = max(1, width)
+        self.height = max(1, height)
+        self._screen = pyte.Screen(self.width, self.height)
+        self._stream = pyte.Stream(self._screen)
+        self._has_header = True
+
+    def _apply_event(self, event: list[Any]) -> None:
+        if len(event) < 2:
+            return
+        event_type = event[1]
+        payload = event[2] if len(event) > 2 else ""
+        if event_type == "o":
+            if isinstance(payload, str):
+                self._stream.feed(payload)
+        elif event_type == "r":
+            width, height = _parse_resize(payload)
+            if width and height:
+                self.width = width
+                self.height = height
+                self._screen.resize(width, height)
+
+
+def _coerce_int(value: Any, default: int) -> int:
+    try:
+        return int(value)
+    except (TypeError, ValueError):
+        return int(default)
+
+
+def _parse_resize(payload: Any) -> tuple[int, int]:
+    if isinstance(payload, str) and "x" in payload:
+        left, right = payload.lower().split("x", 1)
+        return _coerce_int(left, 0), _coerce_int(right, 0)
+    if isinstance(payload, dict):
+        width = _coerce_int(payload.get("width") or payload.get("columns") or payload.get("cols"), 0)
+        height = _coerce_int(payload.get("height") or payload.get("rows") or payload.get("lines"), 0)
+        return width, height
+    if isinstance(payload, list) and len(payload) >= 2:
+        return _coerce_int(payload[0], 0), _coerce_int(payload[1], 0)
+    return 0, 0
+

atropos/tools/__init__.py

@@ -0,0 +1,31 @@
+"""
+Tool abstractions for atropos-agent.
+
+Provides base Tool class, ToolCall/ToolResult types, and specialized tools.
+
+Kept modules:
+- base.py: ToolSchema, ToolCall, ToolResult, Tool ABC, ToolRegistry
+- tool_executor.py: Batched execution queue with slot routing
+- terminal_stateful_tool.py: Persistent terminal sessions
+- tmux_tool.py: Tmux-based streaming terminal
+
+Removed (replaced by hermes-agent equivalents):
+- build_registry.py → model_tools.py + toolsets.py
+- sandbox_stubs.py → atropos/backends/ execute() methods
+- hermes_external_tools.py → environments/agent_loop.py handle_function_call()
+- toolset_resolver.py → toolsets.py
+"""
+
+from .base import Tool, ToolCall, ToolRegistry, ToolResult, ToolSchema
+from .terminal_stateful_tool import TerminalStatefulTool
+from .tmux_tool import TmuxTool
+
+__all__ = [
+    "Tool",
+    "ToolCall",
+    "ToolRegistry",
+    "ToolResult",
+    "ToolSchema",
+    "TerminalStatefulTool",
+    "TmuxTool",
+]

atropos/tools/base.py

@@ -0,0 +1,423 @@
+"""
+Base Tool abstraction for atropos-agent.
+
+Tools follow a simple pattern:
+1. Define schema (name, description, parameters)
+2. Implement execute() method
+3. Return ToolResult with output/error
+
+Tool calls use Hermes-style XML tags:
+<tool_call>{"name": "bash", "arguments": {"command": "ls"}}</tool_call>
+"""
+
+import json
+import re
+import uuid
+from abc import ABC, abstractmethod
+from dataclasses import dataclass, field
+from typing import Any, Dict, List, Literal, Optional
+
+from pydantic import BaseModel, Field
+
+
+@dataclass
+class ToolSchema:
+    """JSON Schema for a tool's parameters."""
+    
+    name: str
+    description: str
+    parameters: Dict[str, Any] = field(default_factory=dict)
+    required: List[str] = field(default_factory=list)
+    external: bool = False  # Whether the tool must be executed via an external ToolServer (secret proxy) and not inside the sandbox.
+    
+    def to_dict(self) -> Dict[str, Any]:
+        """Convert to OpenAI-compatible function schema."""
+        return {
+            "type": "function",
+            "function": {
+                "name": self.name,
+                "description": self.description,
+                "parameters": {
+                    "type": "object",
+                    "properties": self.parameters,
+                    "required": self.required,
+                },
+            },
+        }
+    
+    def to_prompt_description(self) -> str:
+        """Convert to human-readable description for system prompt."""
+        params_desc = []
+        for name, spec in self.parameters.items():
+            req = "(required)" if name in self.required else "(optional)"
+            desc = spec.get("description", "")
+            param_type = spec.get("type", "string")
+            params_desc.append(f"  - {name} ({param_type}) {req}: {desc}")
+        
+        params_str = "\n".join(params_desc) if params_desc else "  (no parameters)"
+        return f"**{self.name}**: {self.description}\nParameters:\n{params_str}"
+
+
+@dataclass
+class ToolCall:
+    """A parsed tool call from model output."""
+    
+    name: str
+    arguments: Dict[str, Any]
+    raw_text: str = ""  # Original XML/JSON text
+    uniq_id: str = field(default_factory=lambda: str(uuid.uuid4()))  # Unique tool-call id for traceability/reconstruction.
+    
+    @classmethod
+    def parse_from_text(cls, text: str) -> List["ToolCall"]:
+        """
+        Extract tool calls from text using Hermes-style XML tags.
+        
+        Supported formats (STRICT: requires well-formed closing tags):
+        - Hermes JSON wrapper:
+          <tool_call>{"name": "...", "arguments": {...}}</tool_call>
+        - GLM/llama.cpp style:
+          <tool_call>terminal{"command":"ls -la"}</tool_call>
+        """
+        calls: List["ToolCall"] = []
+
+        if not text:
+            return calls
+
+        def _append_from_payload(*, name: str, arguments: Dict[str, Any], raw: str, uniq_id: Optional[str] = None) -> None:
+            if not isinstance(name, str) or not name:
+                return
+            if not isinstance(arguments, dict):
+                return
+            calls.append(
+                cls(
+                    name=name,
+                    arguments=arguments,
+                    raw_text=raw,
+                    uniq_id=uniq_id or str(uuid.uuid4()),
+                )
+            )
+
+        # STRICT parsing: only accept well-formed <tool_call>...</tool_call> blocks.
+        pattern = r"<tool_call>\s*(.*?)\s*</tool_call>"
+        for inner in re.findall(pattern, text, re.DOTALL):
+            cleaned = (inner or "").strip()
+            if not cleaned:
+                continue
+
+            # Hermes JSON wrapper.
+            if cleaned.startswith("{"):
+                try:
+                    data = json.loads(cleaned)
+                except json.JSONDecodeError:
+                    continue
+                uniq_id = data.get("uniq_id") or data.get("id") or None
+                _append_from_payload(
+                    name=data.get("name", ""),
+                    arguments=data.get("arguments", {}),
+                    raw=inner,
+                    uniq_id=uniq_id,
+                )
+                continue
+
+            # GLM/llama.cpp style: terminal{...}
+            m = re.match(r"^\s*([A-Za-z0-9_.:\\-]+)\s*(\{.*\})\s*$", cleaned, re.DOTALL)
+            if not m:
+                continue
+            name = m.group(1)
+            args_text = m.group(2)
+            try:
+                args = json.loads(args_text)
+            except json.JSONDecodeError:
+                continue
+            _append_from_payload(name=name, arguments=args, raw=inner)
+
+        return calls
+    
+    @classmethod
+    def has_tool_call(cls, text: str) -> bool:
+        """Check if text contains any tool calls."""
+        return bool(re.search(r"<tool_call>", text))
+
+
+@dataclass
+class ToolResult:
+    """Result from executing a tool."""
+    
+    success: bool
+    output: str = ""
+    error: str = ""
+    metadata: Dict[str, Any] = field(default_factory=dict)
+    uniq_id: Optional[str] = None  # Should match ToolCall.uniq_id for async execution tracking.
+    
+    def to_xml(self) -> str:
+        """Format as XML for including in conversation."""
+        data = {
+            "success": self.success,
+            "output": self.output,
+        }
+        if self.uniq_id:
+            data["uniq_id"] = self.uniq_id
+        if self.error:
+            data["error"] = self.error
+        if self.metadata:
+            data["metadata"] = self.metadata
+        return f"<tool_response>{json.dumps(data)}</tool_response>"
+    
+    def to_dict(self) -> Dict[str, Any]:
+        """Convert to dictionary."""
+        return {
+            "success": self.success,
+            "output": self.output,
+            "error": self.error,
+            "metadata": self.metadata,
+            "uniq_id": self.uniq_id,
+        }
+
+
+class Tool(ABC):
+    """
+    Abstract base class for tools.
+    
+    Subclasses must implement:
+    - schema: ToolSchema describing the tool
+    - execute(): async method that performs the tool action
+    """
+    
+    @property
+    @abstractmethod
+    def schema(self) -> ToolSchema:
+        """Return the tool's schema."""
+        pass
+    
+    @property
+    def name(self) -> str:
+        """Tool name (from schema)."""
+        return self.schema.name
+    
+    @abstractmethod
+    async def execute(self, **kwargs) -> ToolResult:
+        """
+        Execute the tool with given arguments.
+        
+        Args:
+            **kwargs: Tool-specific arguments
+            
+        Returns:
+            ToolResult with success/failure and output
+        """
+        pass
+    
+    def is_available(self) -> tuple[bool, str | None]:
+        """
+        Return whether this tool should be exposed/executable in the current process.
+
+        Tools that depend on optional binaries/services/env vars can override this
+        to avoid advertising a tool that will fail at runtime.
+        """
+        return True, None
+
+    async def __call__(self, **kwargs) -> ToolResult:
+        """Allow calling tool instance directly."""
+        return await self.execute(**kwargs)
+
+# Note: This is only wrapping declarations for the external ToolServer (for execution on external process tools), and tools preinstalled in envs
+class ToolRegistry:
+    """Registry of available tools."""
+    
+    def __init__(self):
+        self._tools: Dict[str, Tool] = {}
+    
+    def register(self, tool: Tool) -> None:
+        """Register a tool."""
+        self._tools[tool.name] = tool
+    
+    def get(self, name: str) -> Optional[Tool]:
+        """Get a tool by name."""
+        return self._tools.get(name)
+    
+    def list_tools(self) -> List[Tool]:
+        """List all registered tools."""
+        return list(self._tools.values())
+    
+    def get_schemas(self) -> List[ToolSchema]:
+        """Get schemas for all registered tools."""
+        return [tool.schema for tool in self._tools.values()]
+    
+    def get_prompt_description(self) -> str:
+        """Generate tool descriptions for system prompt."""
+        descriptions = [tool.schema.to_prompt_description() for tool in self._tools.values()]
+        return "\n\n".join(descriptions)
+
+    def get_prompt_tool_definitions_json(self) -> str:
+        """
+        Return a Hermes-style JSON list of tool definitions for use inside a `<tools>...</tools>` block.
+
+        Hermes trajectories historically use a simplified schema list:
+          [{"name": ..., "description": ..., "parameters": {...}, "required": null}, ...]
+        """
+        formatted: List[Dict[str, Any]] = []
+        for tool in self._tools.values():
+            fn = tool.schema.to_dict().get("function", {})
+            formatted.append(
+                {
+                    "name": fn.get("name", tool.name),
+                    "description": fn.get("description", ""),
+                    "parameters": fn.get("parameters", {}),
+                    # Keep parity with Hermes saved trajectories (required is typically null there).
+                    "required": None,
+                }
+            )
+        return json.dumps(formatted, ensure_ascii=False)
+    
+    async def execute(self, call: ToolCall) -> ToolResult:
+        """Execute a tool call."""
+        tool = self.get(call.name)
+        if tool is None:
+            return ToolResult(
+                success=False,
+                error=f"Unknown tool: {call.name}",
+                uniq_id=call.uniq_id,
+            )
+        
+        try:
+            result = await tool.execute(**call.arguments)
+            if result.uniq_id is None:
+                result.uniq_id = call.uniq_id
+            return result
+        except Exception as e:
+            return ToolResult(
+                success=False,
+                error=f"Tool execution error: {str(e)}",
+                uniq_id=call.uniq_id,
+            )
+
+
+# =============================================================================
+# FastAPI / transport models
+# =============================================================================
+
+
+class ToolCallPayload(BaseModel):
+    name: str
+    arguments: Dict[str, Any] = Field(default_factory=dict)
+    uniq_id: str
+
+    @classmethod
+    def from_tool_call(cls, call: ToolCall) -> "ToolCallPayload":
+        return cls(name=call.name, arguments=call.arguments, uniq_id=call.uniq_id)
+
+    def to_tool_call(self) -> ToolCall:
+        return ToolCall(name=self.name, arguments=self.arguments, uniq_id=self.uniq_id)
+
+
+class ToolResultPayload(BaseModel):
+    success: bool
+    output: str = ""
+    error: str = ""
+    metadata: Dict[str, Any] = Field(default_factory=dict)
+    uniq_id: Optional[str] = None
+
+    @classmethod
+    def from_tool_result(cls, result: ToolResult) -> "ToolResultPayload":
+        return cls(
+            success=result.success,
+            output=result.output,
+            error=result.error,
+            metadata=result.metadata,
+            uniq_id=result.uniq_id,
+        )
+
+    def to_tool_result(self) -> ToolResult:
+        return ToolResult(
+            success=self.success,
+            output=self.output,
+            error=self.error,
+            metadata=self.metadata,
+            uniq_id=self.uniq_id,
+        )
+
+
+class ToolExecutorExecuteRequest(BaseModel):
+    trajectory_id: str
+    tool: ToolCallPayload
+    timeout_s: Optional[float] = None
+
+
+class ToolExecutorReleaseRequest(BaseModel):
+    trajectory_id: str
+    reset_workspace: bool = False
+
+
+class ToolServerExecuteRequest(BaseModel):
+    trajectory_id: Optional[str] = None
+    tool: ToolCallPayload
+    timeout_s: Optional[float] = None
+    # Optional sandbox context for tools that need workspace artifacts.
+    # This is set by ToolExecutor and is NOT model-controlled.
+    slot_id: Optional[str] = None
+    container_addr: Optional[str] = None
+
+
+# =============================================================================
+# Artifact transport models
+# =============================================================================
+
+
+class ArtifactReadRequestPayload(BaseModel):
+    trajectory_id: str
+    path: str
+    encoding: Literal["text", "base64"] = "text"
+    max_bytes: Optional[int] = None
+    include_sha256: bool = False
+
+
+class ArtifactReadResponsePayload(BaseModel):
+    success: bool
+    content: str = ""
+    error: str = ""
+    encoding: str = "text"
+    truncated: bool = False
+    bytes: int = 0
+    file_size: Optional[int] = None
+    path: str = ""
+    mime: Optional[str] = None
+    sha256: Optional[str] = None
+
+
+class ArtifactListRequestPayload(BaseModel):
+    trajectory_id: str
+    path: str = "."
+    recursive: bool = False
+    max_entries: Optional[int] = None
+
+
+class ArtifactListEntryPayload(BaseModel):
+    path: str
+    is_dir: bool
+    size: int
+    mtime: float
+
+
+class ArtifactListResponsePayload(BaseModel):
+    success: bool
+    entries: List[ArtifactListEntryPayload] = Field(default_factory=list)
+    truncated: bool = False
+    error: str = ""
+
+
+class ArtifactArchiveRequestPayload(BaseModel):
+    trajectory_id: str
+    path: str = "."
+    format: Literal["tar.gz", "tgz"] = "tar.gz"
+    max_bytes: Optional[int] = None
+    max_entries: Optional[int] = None
+
+
+class ArtifactArchiveResponsePayload(BaseModel):
+    success: bool
+    content: str = ""
+    error: str = ""
+    encoding: str = "base64"
+    format: str = "tar.gz"
+    bytes: int = 0
+    entry_count: int = 0

atropos/tools/terminal_stateful_tool.py

@@ -0,0 +1,45 @@
+"""
+Stateful terminal tool schema.
+
+This is a sandbox tool that routes to the sandbox server as `bash_stateful`
+via ToolExecutor mapping. It exists to expose an explicit, opt-in terminal
+primitive suitable for stateful workflows (e.g. tmux sessions / TUIs).
+"""
+
+from __future__ import annotations
+
+from typing import Optional
+
+from .base import Tool, ToolResult, ToolSchema
+
+
+class TerminalStatefulTool(Tool):
+    @property
+    def schema(self) -> ToolSchema:
+        return ToolSchema(
+            name="terminal_stateful",
+            description=(
+                "Execute a command in the sandbox, allowing stateful/background processes to persist "
+                "across tool calls within the same trajectory slot (e.g. tmux sessions). "
+                "Use sparingly; output is still non-interactive."
+            ),
+            parameters={
+                "command": {"type": "string", "description": "The command to execute"},
+                "timeout": {
+                    "type": "integer",
+                    "description": "Command timeout in seconds (optional).",
+                    "minimum": 1,
+                },
+            },
+            required=["command"],
+        )
+
+    def is_available(self) -> tuple[bool, str | None]:
+        return True, None
+
+    async def execute(self, command: str, timeout: Optional[int] = None) -> ToolResult:
+        _ = (command, timeout)
+        return ToolResult(
+            success=False,
+            error="terminal_stateful must be executed via ToolExecutor inside the sandbox",
+        )

atropos/tools/tmux_tool.py

@@ -0,0 +1,89 @@
+"""
+tmux tool schema (sandbox).
+
+This is a sandbox tool that provides basic tmux session control suitable for
+TUI-style terminal interactions:
+- send keys (arrow keys, enter, etc.)
+- capture the current screen buffer
+
+Execution is routed by ToolExecutor to the sandbox server's `tmux` backend.
+"""
+
+from __future__ import annotations
+
+from typing import Any, Dict, Optional
+
+from .base import Tool, ToolResult, ToolSchema
+
+
+class TmuxTool(Tool):
+    @property
+    def schema(self) -> ToolSchema:
+        return ToolSchema(
+            name="tmux",
+            description=(
+                "Control a per-trajectory tmux session inside the sandbox (stateful terminal). "
+                "Use this for TUI-style interactions: send keys and capture the current screen."
+            ),
+            parameters={
+                "action": {
+                    "type": "string",
+                    "description": "Action to perform: start | send_keys | stream | stop.",
+                    "enum": ["start", "send_keys", "stream", "stop", "capture"],
+                },
+                "keys": {
+                    "description": "Keys to send (string or list of strings) when action=send_keys.",
+                },
+                "block": {
+                    "type": "boolean",
+                    "description": "If true, wait for shell command completion (only valid at a shell prompt).",
+                    "default": False,
+                },
+                "min_wait_s": {
+                    "type": "number",
+                    "description": "For non-blocking send_keys, sleep this long after sending keys (seconds).",
+                    "default": 0.0,
+                },
+                "max_wait_s": {
+                    "type": "number",
+                    "description": "For blocking send_keys, max time to wait for completion (seconds).",
+                },
+                "capture_entire": {
+                    "type": "boolean",
+                    "description": "Deprecated. Streaming is preferred.",
+                    "default": False,
+                },
+                "max_bytes": {
+                    "type": "integer",
+                    "description": "Max bytes to return per stream call.",
+                },
+                "reset": {
+                    "type": "boolean",
+                    "description": "If true, reset stream offset to the beginning of the asciinema recording.",
+                    "default": False,
+                },
+                "pane_width": {
+                    "type": "integer",
+                    "description": "Pane width for action=start (columns).",
+                    "minimum": 20,
+                },
+                "pane_height": {
+                    "type": "integer",
+                    "description": "Pane height for action=start (rows).",
+                    "minimum": 10,
+                },
+            },
+            required=["action"],
+        )
+
+    def is_available(self) -> tuple[bool, str | None]:
+        return True, None
+
+    async def execute(self, **kwargs: Dict[str, Any]) -> ToolResult:
+        # This tool is intended to be executed via ToolExecutor -> sandbox server.
+        # We keep a safe fallback for non-sandbox contexts.
+        action = str(kwargs.get("action") or "").strip()
+        return ToolResult(
+            success=False,
+            error=f"tmux tool must be executed in the sandbox (got action={action!r})",
+        )

atropos/tools/tool_executor.py

@@ -0,0 +1,500 @@
+"""
+ToolExecutor - queued, batched tool dispatch for multiplexed agent trajectories.
+
+This component is responsible for:
+- Maintaining trajectory -> Slot affinity (workspace continuity)
+- Batching sandbox tool calls across trajectories to maximize container utilization
+- Routing external tools (ToolSchema.external=True) to a ToolServer (Phase 4.5)
+
+For now, only sandbox tools are executed:
+- bash
+- read_file
+- write_file
+"""
+
+from __future__ import annotations
+
+import asyncio
+import time
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional
+
+import httpx
+
+from .base import (
+    ArtifactArchiveRequestPayload,
+    ArtifactArchiveResponsePayload,
+    ArtifactListRequestPayload,
+    ArtifactListResponsePayload,
+    ArtifactReadRequestPayload,
+    ArtifactReadResponsePayload,
+    ToolCall,
+    ToolCallPayload,
+    ToolRegistry,
+    ToolResult,
+    ToolResultPayload,
+    ToolServerExecuteRequest,
+)
+from ..backends.base import ToolBackend
+from ..slots import Slot
+
+
+@dataclass
+class ToolExecutorConfig:
+    batch_window_ms: int = 20
+    max_batch_size: int = 200
+    allow_network: bool = True
+    require_sandbox: bool = False
+    require_stateful_sandbox: bool = False
+    tool_server_url: Optional[str] = None
+    tool_server_token: Optional[str] = None
+
+
+@dataclass
+class _QueuedToolRequest:
+    trajectory_id: str
+    call: ToolCall
+    timeout_s: Optional[float]
+    future: asyncio.Future
+
+
+class ToolExecutor:
+    def __init__(
+        self,
+        backend: ToolBackend,
+        tools: ToolRegistry,
+        config: Optional[ToolExecutorConfig] = None,
+    ) -> None:
+        self.backend = backend
+        self.tools = tools
+        self.config = config or ToolExecutorConfig()
+
+        self._queue: asyncio.Queue[Optional[_QueuedToolRequest]] = asyncio.Queue()
+        self._task: Optional[asyncio.Task] = None
+        self._stopping = asyncio.Event()
+
+        self._slots_lock = asyncio.Lock()
+        self._slot_by_trajectory: Dict[str, Slot] = {}
+
+        self._tool_server_client: Optional[httpx.AsyncClient] = None
+        self._tool_server_lock = asyncio.Lock()
+
+        # lightweight stats for status endpoints
+        self.total_requests: int = 0
+        self.total_errors: int = 0
+        self.latencies_s: List[float] = []
+
+    async def start(self) -> None:
+        if self._task is None:
+            self._task = asyncio.create_task(self._run_loop())
+
+    def queue_size(self) -> int:
+        return self._queue.qsize()
+
+    async def close(self) -> None:
+        self._stopping.set()
+        await self._queue.put(None)
+        if self._task:
+            await self._task
+            self._task = None
+
+        client = self._tool_server_client
+        self._tool_server_client = None
+        if client is not None:
+            await client.aclose()
+
+        # Best-effort release any remaining slots.
+        async with self._slots_lock:
+            slots = list(self._slot_by_trajectory.items())
+            self._slot_by_trajectory.clear()
+
+        for _, slot in slots:
+            try:
+                await self.backend.release(slot, reset_workspace=False)
+            except Exception:
+                pass
+
+    async def execute(
+        self,
+        trajectory_id: str,
+        call: ToolCall,
+        timeout_s: Optional[float] = None,
+    ) -> ToolResult:
+        if self._task is None:
+            raise RuntimeError("ToolExecutor not started (call start() first)")
+
+        # Allow tool args to suggest a timeout (Hermes-compatible terminal tool),
+        # but never let the model choose "infinite" timeouts.
+        if timeout_s is None:
+            raw_timeout = call.arguments.get("timeout")
+            if isinstance(raw_timeout, (int, float)):
+                timeout_s = float(raw_timeout)
+        if timeout_s is not None:
+            timeout_s = max(1.0, min(float(timeout_s), 600.0))
+
+        loop = asyncio.get_running_loop()
+        fut: asyncio.Future = loop.create_future()
+        started = time.perf_counter()
+        await self._queue.put(_QueuedToolRequest(trajectory_id=trajectory_id, call=call, timeout_s=timeout_s, future=fut))
+        try:
+            result: ToolResult = await fut
+            return result
+        finally:
+            self.latencies_s.append(time.perf_counter() - started)
+
+    async def release_trajectory(self, trajectory_id: str, reset_workspace: bool = False) -> None:
+        async with self._slots_lock:
+            slot = self._slot_by_trajectory.pop(trajectory_id, None)
+
+        if slot is not None:
+            await self.backend.release(slot, reset_workspace=reset_workspace)
+
+    async def _get_slot_if_present(self, trajectory_id: str) -> Optional[Slot]:
+        async with self._slots_lock:
+            return self._slot_by_trajectory.get(trajectory_id)
+
+    # ---------------------------------------------------------------------
+    # Artifact helpers (optional)
+    # ---------------------------------------------------------------------
+
+    async def read_artifact(self, req: ArtifactReadRequestPayload) -> ArtifactReadResponsePayload:
+        slot = await self._get_slot_if_present(req.trajectory_id)
+        if slot is None:
+            return ArtifactReadResponsePayload(success=False, error="No active slot for trajectory (run a sandbox tool first)")
+        data = await self.backend.read_artifact(
+            slot,
+            req.path,
+            encoding=req.encoding,
+            max_bytes=req.max_bytes,
+            include_sha256=req.include_sha256,
+        )
+        if isinstance(data, dict):
+            data = dict(data)
+            data.pop("http_status", None)
+        try:
+            return ArtifactReadResponsePayload(**(data or {}))
+        except Exception as e:
+            return ArtifactReadResponsePayload(success=False, error=f"Invalid artifact read response: {e}")
+
+    async def list_artifacts(self, req: ArtifactListRequestPayload) -> ArtifactListResponsePayload:
+        slot = await self._get_slot_if_present(req.trajectory_id)
+        if slot is None:
+            return ArtifactListResponsePayload(success=False, error="No active slot for trajectory (run a sandbox tool first)")
+        data = await self.backend.list_artifacts(
+            slot,
+            req.path,
+            recursive=req.recursive,
+            max_entries=req.max_entries,
+        )
+        if isinstance(data, dict):
+            data = dict(data)
+            data.pop("http_status", None)
+        try:
+            return ArtifactListResponsePayload(**(data or {}))
+        except Exception as e:
+            return ArtifactListResponsePayload(success=False, error=f"Invalid artifact list response: {e}")
+
+    async def archive_artifacts(self, req: ArtifactArchiveRequestPayload) -> ArtifactArchiveResponsePayload:
+        slot = await self._get_slot_if_present(req.trajectory_id)
+        if slot is None:
+            return ArtifactArchiveResponsePayload(success=False, error="No active slot for trajectory (run a sandbox tool first)")
+        data = await self.backend.archive_artifacts(
+            slot,
+            req.path,
+            archive_format=req.format,
+            max_bytes=req.max_bytes,
+            max_entries=req.max_entries,
+        )
+        if isinstance(data, dict):
+            data = dict(data)
+            data.pop("http_status", None)
+        try:
+            return ArtifactArchiveResponsePayload(**(data or {}))
+        except Exception as e:
+            return ArtifactArchiveResponsePayload(success=False, error=f"Invalid artifact archive response: {e}")
+
+    async def _get_or_acquire_slot(self, trajectory_id: str) -> Slot:
+        async with self._slots_lock:
+            existing = self._slot_by_trajectory.get(trajectory_id)
+            if existing is not None:
+                return existing
+
+        slot = await self.backend.acquire(trajectory_id)
+
+        async with self._slots_lock:
+            existing = self._slot_by_trajectory.get(trajectory_id)
+            if existing is not None:
+                # Another coroutine won the race; return its slot.
+                await self.backend.release(slot, reset_workspace=False)
+                return existing
+            self._slot_by_trajectory[trajectory_id] = slot
+            return slot
+
+    async def _run_loop(self) -> None:
+        pending: List[_QueuedToolRequest] = []
+        deadline: Optional[float] = None
+
+        batch_window_s = max(0.0, self.config.batch_window_ms / 1000.0)
+        max_batch = max(1, self.config.max_batch_size)
+
+        while True:
+            if self._stopping.is_set() and self._queue.empty() and not pending:
+                break
+
+            timeout = None
+            if pending and deadline is not None:
+                timeout = max(0.0, deadline - time.perf_counter())
+
+            try:
+                item = await asyncio.wait_for(self._queue.get(), timeout=timeout)
+                if item is None:
+                    continue
+                pending.append(item)
+                if len(pending) == 1:
+                    deadline = time.perf_counter() + batch_window_s
+                if len(pending) < max_batch:
+                    continue
+            except asyncio.TimeoutError:
+                # batch window elapsed
+                pass
+
+            if not pending:
+                deadline = None
+                continue
+
+            batch = pending
+            pending = []
+            deadline = None
+
+            await self._execute_batch(batch)
+
+    async def _get_tool_server_client(self) -> httpx.AsyncClient:
+        url = self.config.tool_server_url
+        if not url:
+            raise RuntimeError("ToolServer not configured")
+
+        if self._tool_server_client is not None:
+            return self._tool_server_client
+
+        async with self._tool_server_lock:
+            if self._tool_server_client is None:
+                self._tool_server_client = httpx.AsyncClient(base_url=url.rstrip("/"))
+            return self._tool_server_client
+
+    def _tool_server_headers(self) -> Dict[str, str]:
+        token = self.config.tool_server_token
+        if not token:
+            return {}
+        return {"Authorization": f"Bearer {token}"}
+
+    async def _execute_external(self, req: _QueuedToolRequest) -> ToolResult:
+        client = await self._get_tool_server_client()
+        slot_id: Optional[str] = None
+        container_addr: Optional[str] = None
+        slot = await self._get_slot_if_present(req.trajectory_id)
+        if slot is not None:
+            slot_id = slot.slot_id
+            container_addr = slot.container_addr
+
+        payload = ToolServerExecuteRequest(
+            trajectory_id=req.trajectory_id,
+            tool=ToolCallPayload.from_tool_call(req.call),
+            timeout_s=req.timeout_s,
+            slot_id=slot_id,
+            container_addr=container_addr,
+        )
+
+        try:
+            resp = await client.post(
+                "/execute",
+                json=payload.model_dump(),
+                headers=self._tool_server_headers(),
+                timeout=req.timeout_s,
+            )
+            resp.raise_for_status()
+            data = resp.json()
+            parsed = ToolResultPayload(**data)
+            result = parsed.to_tool_result()
+            if result.uniq_id is None:
+                result.uniq_id = req.call.uniq_id
+            return result
+        except Exception as e:
+            return ToolResult(
+                success=False,
+                error=f"External tool failed: {e}",
+                uniq_id=req.call.uniq_id,
+            )
+
+    async def _execute_batch(self, batch: List[_QueuedToolRequest]) -> None:
+        # Resolve tool schemas once per request and separate sandbox/external/unknown.
+        sandbox_items: List[_QueuedToolRequest] = []
+        external_items: List[_QueuedToolRequest] = []
+        unknown_items: List[_QueuedToolRequest] = []
+
+        for it in batch:
+            tool = self.tools.get(it.call.name)
+            if tool is None:
+                unknown_items.append(it)
+                continue
+
+            schema = tool.schema
+            if not schema.external:
+                sandbox_items.append(it)
+            else:
+                external_items.append(it)
+
+        for it in unknown_items:
+            self.total_requests += 1
+            self.total_errors += 1
+            if not it.future.done():
+                it.future.set_result(
+                    ToolResult(
+                        success=False,
+                        error=f"Unknown tool: {it.call.name}",
+                        uniq_id=it.call.uniq_id,
+                    )
+                )
+
+        if external_items:
+            if not self.config.tool_server_url:
+                for it in external_items:
+                    self.total_requests += 1
+                    self.total_errors += 1
+                    if not it.future.done():
+                        it.future.set_result(
+                            ToolResult(
+                                success=False,
+                                error=f"External tool not available (ToolServer not configured): {it.call.name}",
+                                uniq_id=it.call.uniq_id,
+                            )
+                        )
+            else:
+                results = await asyncio.gather(*[self._execute_external(it) for it in external_items])
+                for it, res in zip(external_items, results):
+                    self.total_requests += 1
+                    if not getattr(res, "success", False):
+                        self.total_errors += 1
+                    if not it.future.done():
+                        it.future.set_result(res)
+
+        if not sandbox_items:
+            return
+
+        # Acquire slots for the distinct trajectories in this batch.
+        try:
+            traj_ids = list({it.trajectory_id for it in sandbox_items})
+            slots = await asyncio.gather(*[self._get_or_acquire_slot(tid) for tid in traj_ids])
+            slot_by_traj = dict(zip(traj_ids, slots))
+        except Exception as e:
+            for it in sandbox_items:
+                self.total_requests += 1
+                self.total_errors += 1
+                if not it.future.done():
+                    it.future.set_result(
+                        ToolResult(
+                            success=False,
+                            error=f"Failed to acquire slot: {e}",
+                            uniq_id=it.call.uniq_id,
+                        )
+                    )
+            return
+
+        # Group by timeout so we don't accidentally make short timeouts wait on long ones.
+        by_timeout: Dict[float, List[_QueuedToolRequest]] = {}
+        default_timeout = self.backend.default_timeout_s
+
+        for it in sandbox_items:
+            t = it.timeout_s
+            if t is None:
+                t = default_timeout
+            if t is None:
+                t = 30.0
+            by_timeout.setdefault(float(t), []).append(it)
+
+        for timeout_s, items in by_timeout.items():
+            requests = []
+            dispatched: List[_QueuedToolRequest] = []
+            for it in items:
+                slot = slot_by_traj[it.trajectory_id]
+                tool_name = it.call.name
+                args = dict(it.call.arguments)
+
+                # Hermes compatibility: treat `terminal` as an alias of sandbox `bash`.
+                if tool_name == "terminal":
+                    if args.get("background"):
+                        self.total_requests += 1
+                        self.total_errors += 1
+                        if not it.future.done():
+                            it.future.set_result(
+                                ToolResult(
+                                    success=False,
+                                    error="terminal background execution is not supported in sandbox",
+                                    uniq_id=it.call.uniq_id,
+                                )
+                            )
+                        continue
+                    tool_name = "bash"
+                    # `timeout` is handled at the ToolExecutor level, not passed to the sandbox tool args.
+                    args.pop("timeout", None)
+                elif tool_name == "terminal_stateful":
+                    tool_name = "bash_stateful"
+                    args.pop("timeout", None)
+                elif tool_name == "tmux":
+                    # `tmux` is a sandbox tool backed by the stateful session manager.
+                    # Network policy is env-controlled.
+                    args.pop("allow_network", None)
+
+                if tool_name == "bash":
+                    # Network policy is set by the environment/executor, not by the model.
+                    args.pop("allow_network", None)
+                    args.pop("require_sandbox", None)
+                    args["allow_network"] = bool(self.config.allow_network)
+                    args["require_sandbox"] = bool(self.config.require_sandbox)
+                    # `timeout` is handled at the ToolExecutor level, not passed to the sandbox tool args.
+                    args.pop("timeout", None)
+                elif tool_name == "bash_stateful":
+                    # Network policy is set by the environment/executor, not by the model.
+                    args.pop("allow_network", None)
+                    args.pop("require_sandbox", None)
+                    args.pop("require_stateful_sandbox", None)
+                    args["allow_network"] = bool(self.config.allow_network)
+                    args["require_stateful_sandbox"] = bool(self.config.require_stateful_sandbox)
+                    args.pop("timeout", None)
+                elif tool_name == "tmux":
+                    # Network policy applies to the underlying stateful session.
+                    args.pop("allow_network", None)
+                    args.pop("require_sandbox", None)
+                    args.pop("require_stateful_sandbox", None)
+                    args["allow_network"] = bool(self.config.allow_network)
+                    args["require_stateful_sandbox"] = bool(self.config.require_stateful_sandbox)
+
+                requests.append((slot, tool_name, args))
+                dispatched.append(it)
+
+            results = None
+            try:
+                if not dispatched:
+                    continue
+                results = await self.backend.execute_batch(requests, timeout_s=timeout_s)
+            except Exception as e:
+                for it in items:
+                    self.total_requests += 1
+                    self.total_errors += 1
+                    if not it.future.done():
+                        it.future.set_result(
+                            ToolResult(
+                                success=False,
+                                error=f"Batch execution failed: {e}",
+                                uniq_id=it.call.uniq_id,
+                            )
+                        )
+                continue
+
+            for it, res in zip(dispatched, results):
+                self.total_requests += 1
+                if not getattr(res, "success", False):
+                    self.total_errors += 1
+                tool_result = res.to_tool_result()
+                tool_result.uniq_id = it.call.uniq_id
+                if not it.future.done():
+                    it.future.set_result(tool_result)

atropos_compatible_agent.py

@@ -0,0 +1,415 @@
+#!/usr/bin/env python3
+"""
+Atropos-compatible Hermes agent runner.
+
+This is a minimal subclass of Hermes-Agent's `AIAgent` that swaps the OpenAI
+function-calling backend for Atroposlib's `ManagedServer`/`ServerManager` backend
+and uses Hermes-style XML tool tags:
+
+- <tool_call>{"name": "...", "arguments": {...}}</tool_call>
+- <tool_response>{...}</tool_response>
+
+Tool observations are appended as `role="user"` messages containing one or more
+`<tool_response>` blocks so they survive common chat templates during tokenization.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import json
+import re
+import time
+import warnings
+import os
+from contextlib import asynccontextmanager
+from typing import Any, AsyncGenerator, Dict, List, Optional, Tuple
+
+from model_tools import cleanup_vm, handle_function_call
+from run_agent import AIAgent
+
+_TOOL_CALL_RE = re.compile(r"<tool_call>\\s*(.*?)\\s*</tool_call>", re.DOTALL)
+
+
+ATROPOS_TOOL_SYSTEM_PROMPT = """You are a helpful AI assistant with access to tools.
+
+## Available Tools
+<tools>
+{tool_descriptions}
+</tools>
+
+## How to Use Tools
+To call a tool, output:
+<tool_call>{{"name": "tool_name", "arguments": {{"arg1": "value1"}}}}</tool_call>
+
+You may include optional reasoning in <think>...</think> before tool calls.
+
+After each tool call, you will receive tool results as:
+<tool_response>{{...}}</tool_response>
+
+Continue until finished, then provide a final response with no <tool_call> blocks.
+"""
+
+
+class AtroposAIAgent(AIAgent):
+    """
+    Hermes `AIAgent` variant that uses Atroposlib ServerManager/ManagedServer.
+
+    Notes:
+    - The default Hermes `AIAgent` remains unchanged; this class is opt-in.
+    - The underlying server must expose `managed_server(tokenizer=...)` OR be a single
+      APIServer-compatible object usable by Atroposlib's `ManagedServer`.
+    """
+
+    def __init__(
+        self,
+        *,
+        server: Any,
+        tokenizer: Any = None,
+        model: str = "local",
+        max_iterations: int = 10,
+        tool_delay: float = 0.0,
+        enabled_toolsets: Optional[List[str]] = None,
+        disabled_toolsets: Optional[List[str]] = None,
+        save_trajectories: bool = False,
+        verbose_logging: bool = False,
+        quiet_mode: bool = False,
+        ephemeral_system_prompt: Optional[str] = None,
+        log_prefix_chars: int = 100,
+        log_prefix: str = "",
+        session_id: Optional[str] = None,
+        temperature: Optional[float] = None,
+        max_tokens: Optional[int] = None,
+    ):
+        # Call parent init mainly to reuse tool selection + trajectory saving utilities.
+        super().__init__(
+            base_url="http://unused",
+            api_key="dummy-key",
+            model=model,
+            max_iterations=max_iterations,
+            tool_delay=tool_delay,
+            enabled_toolsets=enabled_toolsets,
+            disabled_toolsets=disabled_toolsets,
+            save_trajectories=save_trajectories,
+            verbose_logging=verbose_logging,
+            quiet_mode=quiet_mode,
+            ephemeral_system_prompt=ephemeral_system_prompt,
+            log_prefix_chars=log_prefix_chars,
+            log_prefix=log_prefix,
+            session_id=session_id,
+        )
+
+        self.server = server
+        self.tokenizer = tokenizer
+        self.temperature = temperature
+        self.max_tokens = max_tokens
+
+    @asynccontextmanager
+    async def _managed(self) -> AsyncGenerator[Any, None]:
+        if hasattr(self.server, "managed_server"):
+            with warnings.catch_warnings():
+                warnings.filterwarnings(
+                    "ignore",
+                    message=r"Using OpenAIServer with managed_server does not allow for state tracking",
+                    category=UserWarning,
+                )
+                async with self.server.managed_server(tokenizer=self.tokenizer) as managed:
+                    yield managed
+            return
+
+        # Fall back to directly wrapping a single server object.
+        from atroposlib.envs.server_handling.managed_server import ManagedServer
+
+        managed = ManagedServer(server=self.server, tokenizer=self.tokenizer)
+        try:
+            yield managed
+        finally:
+            managed.reset()
+
+    def _tool_descriptions_text(self) -> str:
+        if not self.tools:
+            return "(no tools available)"
+
+        parts: List[str] = []
+        for tool in self.tools:
+            fn = (tool or {}).get("function", {})
+            name = fn.get("name", "")
+            desc = (fn.get("description") or "").strip()
+            if not name:
+                continue
+            if desc:
+                parts.append(f"- {name}: {desc}")
+            else:
+                parts.append(f"- {name}")
+        return "\n".join(parts) if parts else "(no tools available)"
+
+    def _build_system_prompt(self, system_message: Optional[str]) -> Optional[str]:
+        tool_prompt = ATROPOS_TOOL_SYSTEM_PROMPT.format(
+            tool_descriptions=self._tool_descriptions_text()
+        )
+
+        parts: List[str] = []
+        if system_message:
+            parts.append(system_message)
+        if self.ephemeral_system_prompt:
+            parts.append(self.ephemeral_system_prompt)
+        parts.append(tool_prompt)
+
+        return "\n\n".join(parts)
+
+    def _parse_tool_calls(self, content: str) -> Tuple[List[Tuple[str, Dict[str, Any]]], List[str]]:
+        """
+        Returns:
+          (calls, errors)
+        """
+        calls: List[Tuple[str, Dict[str, Any]]] = []
+        errors: List[str] = []
+
+        for raw in _TOOL_CALL_RE.findall(content or ""):
+            try:
+                payload = json.loads(raw)
+            except json.JSONDecodeError as exc:
+                errors.append(f"Invalid JSON inside <tool_call>: {exc}")
+                continue
+
+            name = payload.get("name")
+            args = payload.get("arguments", {})
+            if not isinstance(name, str) or not name:
+                errors.append("Tool call missing 'name' string")
+                continue
+            if not isinstance(args, dict):
+                errors.append("Tool call 'arguments' must be an object")
+                continue
+
+            calls.append((name, args))
+
+        return calls, errors
+
+    async def run_conversation_async(
+        self,
+        user_message: str,
+        system_message: Optional[str] = None,
+        conversation_history: Optional[List[Dict[str, Any]]] = None,
+        task_id: Optional[str] = None,
+    ) -> Dict[str, Any]:
+        import uuid
+
+        effective_task_id = task_id or str(uuid.uuid4())
+
+        messages: List[Dict[str, Any]] = conversation_history.copy() if conversation_history else []
+        messages.append({"role": "user", "content": user_message})
+
+        active_system_prompt = self._build_system_prompt(system_message)
+
+        api_call_count = 0
+        final_response: Optional[str] = None
+        managed_state: Optional[Dict[str, Any]] = None
+        completed = False
+
+        try:
+            async with self._managed() as managed:
+                while api_call_count < self.max_iterations:
+                    api_call_count += 1
+
+                    api_messages = messages.copy()
+                    if active_system_prompt:
+                        api_messages = [{"role": "system", "content": active_system_prompt}] + api_messages
+
+                    chat_kwargs: Dict[str, Any] = {"messages": api_messages, "n": 1}
+                    if self.max_tokens is not None:
+                        chat_kwargs["max_tokens"] = self.max_tokens
+                    if self.temperature is not None:
+                        chat_kwargs["temperature"] = self.temperature
+
+                    # Prefer OpenAI tool calling when supported by the backend:
+                    # - Many providers normalize Hermes-style <tool_call> tags into tool_calls when `tools` is provided.
+                    # - ManagedServer (atroposlib) does prompt->completion conversion and does not support `tools`.
+                    #   Only pass `tools` when we're calling an OpenAI-compatible chat endpoint directly.
+                    tool_schemas = self.tools if self.tools else None
+                    managed_cls = type(managed).__name__
+                    if tool_schemas and managed_cls != "ManagedServer":
+                        chat_kwargs["tools"] = tool_schemas
+
+                    if os.getenv("HERMES_DEBUG_ATROPOS_REQUEST") == "1":
+                        meta = {
+                            "managed_type": managed_cls,
+                            "model": getattr(getattr(managed, "config", None), "model_name", self.model),
+                            "base_url": getattr(getattr(managed, "config", None), "base_url", None),
+                            "kwargs": chat_kwargs,
+                        }
+                        # Avoid dumping megabytes of data accidentally.
+                        # (Messages can be large; this is still "full" but bounded.)
+                        print("\n=== HERMES_DEBUG_ATROPOS_REQUEST ===", flush=True)
+                        print(json.dumps(meta, ensure_ascii=False, indent=2)[:200_000], flush=True)
+
+                    response = await managed.chat_completion(**chat_kwargs)
+
+                    if os.getenv("HERMES_DEBUG_ATROPOS_RESPONSE") == "1":
+                        try:
+                            dumped = response.model_dump()  # openai pydantic model
+                        except Exception:
+                            dumped = getattr(response, "__dict__", {"repr": repr(response)})
+                        print("\n=== HERMES_DEBUG_ATROPOS_RESPONSE: ChatCompletion (raw) ===", flush=True)
+                        print(json.dumps(dumped, ensure_ascii=False, indent=2), flush=True)
+
+                    if hasattr(managed, "get_state"):
+                        managed_state = managed.get_state()
+
+                    msg = response.choices[0].message
+                    assistant_content = (msg.content or "")
+                    msg_reasoning = getattr(msg, "reasoning", None)
+
+                    # Use tool_calls if the backend provides them (preferred).
+                    structured_tool_calls = getattr(msg, "tool_calls", None)
+
+                    # If the backend emits content="" but includes useful text in reasoning,
+                    # use it for parsing *only if needed* (e.g. tool tags).
+                    if assistant_content == "" and isinstance(msg_reasoning, str) and msg_reasoning:
+                        if os.getenv("HERMES_DEBUG_ATROPOS_RESPONSE") == "1":
+                            print("\n=== HERMES_DEBUG_ATROPOS_RESPONSE: message.reasoning present (content empty) ===", flush=True)
+                            print(msg_reasoning, flush=True)
+
+                    assistant_msg: Dict[str, Any] = {"role": "assistant", "content": assistant_content}
+                    if structured_tool_calls:
+                        # Preserve tool_calls so the next request is consistent with OpenAI protocol.
+                        try:
+                            assistant_msg["tool_calls"] = [
+                                {
+                                    "id": tc.id,
+                                    "type": tc.type,
+                                    "function": {"name": tc.function.name, "arguments": tc.function.arguments},
+                                }
+                                for tc in structured_tool_calls
+                            ]
+                        except Exception:
+                            # Best-effort; keep conversation moving.
+                            pass
+                    messages.append(assistant_msg)
+
+                    # Mode A: OpenAI tool calling (preferred when supported)
+                    if structured_tool_calls:
+                        for tc in structured_tool_calls:
+                            tool_start = time.time()
+                            try:
+                                tool_args = json.loads(tc.function.arguments or "{}")
+                            except Exception:
+                                tool_args = {}
+                            tool_result = handle_function_call(tc.function.name, tool_args, effective_task_id)
+                            tool_duration = time.time() - tool_start
+
+                            # Keep the raw tool result as tool content (OpenAI protocol expects role=tool).
+                            messages.append(
+                                {
+                                    "role": "tool",
+                                    "tool_call_id": tc.id,
+                                    "content": tool_result,
+                                }
+                            )
+
+                            if self.tool_delay and self.tool_delay > 0:
+                                await asyncio.sleep(self.tool_delay)
+
+                        # Continue loop after tool execution.
+                        continue
+
+                    # Mode B: Hermes XML tool tags in assistant text (fallback).
+                    parse_source = assistant_content or (msg_reasoning or "")
+                    tool_calls, parse_errors = self._parse_tool_calls(parse_source)
+
+                    if parse_errors and not tool_calls:
+                        # Ask the model to retry with valid tool JSON.
+                        err_text = "; ".join(parse_errors[:3])
+                        messages.append(
+                            {
+                                "role": "user",
+                                "content": (
+                                    f"<tool_response>{json.dumps({'error': err_text}, ensure_ascii=False)}</tool_response>\n"
+                                    "The previous <tool_call> blocks were invalid. Please output valid JSON inside <tool_call>."
+                                ),
+                            }
+                        )
+                        continue
+
+                    if not tool_calls:
+                        # No tool calls: treat as final answer.
+                        final_response = (assistant_content or "").strip()
+                        completed = True
+                        break
+
+                    tool_responses: List[str] = []
+                    for tool_name, tool_args in tool_calls:
+                        tool_start = time.time()
+                        tool_result = handle_function_call(tool_name, tool_args, effective_task_id)
+                        tool_duration = time.time() - tool_start
+
+                        try:
+                            parsed = json.loads(tool_result)
+                            payload: Any = parsed
+                        except Exception:
+                            payload = tool_result
+
+                        tool_payload = {
+                            "name": tool_name,
+                            "duration_s": round(tool_duration, 3),
+                            "result": payload,
+                        }
+                        tool_responses.append(
+                            f"<tool_response>{json.dumps(tool_payload, ensure_ascii=False)}</tool_response>"
+                        )
+
+                        if self.tool_delay and self.tool_delay > 0:
+                            await asyncio.sleep(self.tool_delay)
+
+                    messages.append({"role": "user", "content": "\n".join(tool_responses)})
+
+                if final_response is None:
+                    final_response = "I've reached the maximum number of iterations."
+
+        finally:
+            try:
+                cleanup_vm(effective_task_id)
+            except Exception:
+                pass
+
+        # Save trajectory using Hermes formatting (optional).
+        self._save_trajectory(messages, user_message, completed=completed)
+
+        return {
+            "final_response": final_response,
+            "messages": messages,
+            "api_calls": api_call_count,
+            "completed": completed,
+            "managed_state": managed_state,
+            "system_prompt": active_system_prompt,
+            "task_id": effective_task_id,
+        }
+
+    def run_conversation(self, *args: Any, **kwargs: Any) -> Dict[str, Any]:
+        """
+        Sync wrapper for convenience.
+
+        If called from within a running event loop (e.g. prompt_toolkit), this
+        runs the async conversation in a dedicated thread to avoid nested loops.
+        """
+        try:
+            asyncio.get_running_loop()
+        except RuntimeError:
+            return asyncio.run(self.run_conversation_async(*args, **kwargs))
+
+        import queue
+        import threading
+
+        out: "queue.Queue[object]" = queue.Queue(maxsize=1)
+
+        def runner() -> None:
+            try:
+                out.put(asyncio.run(self.run_conversation_async(*args, **kwargs)))
+            except BaseException as exc:  # noqa: BLE001
+                out.put(exc)
+
+        thread = threading.Thread(target=runner, daemon=True)
+        thread.start()
+
+        result = out.get()
+        if isinstance(result, BaseException):
+            raise result
+        return result  # type: ignore[return-value]

batch_runner.py

@@ -27,14 +27,16 @@ import time
 from pathlib import Path
 from typing import List, Dict, Any, Optional, Tuple
 from datetime import datetime
-from multiprocessing import Pool, Lock
+from multiprocessing import Pool, Manager, Lock
 import traceback
+
 from rich.progress import Progress, SpinnerColumn, BarColumn, TextColumn, TimeRemainingColumn, MofNCompleteColumn
 from rich.console import Console
 import fire
 
 from run_agent import AIAgent
 from toolset_distributions import (
+    get_distribution, 
     list_distributions, 
     sample_toolsets_from_distribution,
     validate_distribution
@@ -128,7 +130,6 @@ def _extract_tool_stats(messages: List[Dict[str, Any]]) -> Dict[str, Dict[str, i
         # Track tool calls from assistant messages
         if msg["role"] == "assistant" and "tool_calls" in msg and msg["tool_calls"]:
             for tool_call in msg["tool_calls"]:
-                if not tool_call or not isinstance(tool_call, dict): continue
                 tool_name = tool_call["function"]["name"]
                 tool_call_id = tool_call["id"]
                 
@@ -172,7 +173,7 @@ def _extract_tool_stats(messages: List[Dict[str, Any]]) -> Dict[str, Dict[str, i
                     if content_json.get("success") is False:
                         is_success = False
                         
-            except (json.JSONDecodeError, ValueError, TypeError):
+            except:
                 # If not JSON, check if content is empty or explicitly states an error
                 # Note: We avoid simple substring matching to prevent false positives
                 if not content:
@@ -239,7 +240,7 @@ def _process_single_prompt(
     
     Args:
         prompt_index (int): Index of prompt in dataset
-        prompt_data (Dict): Prompt data containing 'prompt' field and optional 'image' field
+        prompt_data (Dict): Prompt data containing 'prompt' field
         batch_num (int): Batch number
         config (Dict): Configuration dict with agent parameters
         
@@ -247,58 +248,6 @@ def _process_single_prompt(
         Dict: Result containing trajectory, stats, and metadata
     """
     prompt = prompt_data["prompt"]
-    task_id = f"task_{prompt_index}"
-    
-    # Per-prompt container image override: if the dataset row has an 'image' field,
-    # register it for this task's sandbox. Works with Docker, Modal, Singularity, and Daytona.
-    container_image = prompt_data.get("image") or prompt_data.get("docker_image")
-    if container_image:
-        # Verify the image is accessible before spending tokens on the agent loop.
-        # For Docker: check local cache, then try pulling.
-        # For Modal: skip local check (Modal pulls server-side).
-        env_type = os.getenv("TERMINAL_ENV", "local")
-        if env_type == "docker":
-            import subprocess as _sp
-            try:
-                probe = _sp.run(
-                    ["docker", "image", "inspect", container_image],
-                    capture_output=True, timeout=10,
-                )
-                if probe.returncode != 0:
-                    if config.get("verbose"):
-                        print(f"   Prompt {prompt_index}: Pulling docker image {container_image}...", flush=True)
-                    pull = _sp.run(
-                        ["docker", "pull", container_image],
-                        capture_output=True, text=True, timeout=600,
-                    )
-                    if pull.returncode != 0:
-                        return {
-                            "success": False,
-                            "prompt_index": prompt_index,
-                            "error": f"Docker image not available: {container_image}\n{pull.stderr[:500]}",
-                            "trajectory": None,
-                            "tool_stats": {},
-                            "toolsets_used": [],
-                            "metadata": {"batch_num": batch_num, "timestamp": datetime.now().isoformat()},
-                        }
-            except FileNotFoundError:
-                pass  # Docker CLI not installed — skip check (e.g., Modal backend)
-            except Exception as img_err:
-                if config.get("verbose"):
-                    print(f"   Prompt {prompt_index}: Docker image check failed: {img_err}", flush=True)
-
-        from tools.terminal_tool import register_task_env_overrides
-        overrides = {
-            "docker_image": container_image,
-            "modal_image": container_image,
-            "singularity_image": f"docker://{container_image}",
-            "daytona_image": container_image,
-        }
-        if prompt_data.get("cwd"):
-            overrides["cwd"] = prompt_data["cwd"]
-        register_task_env_overrides(task_id, overrides)
-        if config.get("verbose"):
-            print(f"   Prompt {prompt_index}: Using container image {container_image}")
     
     try:
         # Sample toolsets from distribution for this prompt
@@ -327,12 +276,10 @@ def _process_single_prompt(
             max_tokens=config.get("max_tokens"),
             reasoning_config=config.get("reasoning_config"),
             prefill_messages=config.get("prefill_messages"),
-            skip_context_files=True,  # Don't pollute trajectories with SOUL.md/AGENTS.md
-            skip_memory=True,  # Don't use persistent memory in batch runs
         )
 
         # Run the agent with task_id to ensure each task gets its own isolated VM
-        result = agent.run_conversation(prompt, task_id=task_id)
+        result = agent.run_conversation(prompt, task_id=f"task_{prompt_index}")
         
         # Extract tool usage statistics
         tool_stats = _extract_tool_stats(result["messages"])
@@ -607,7 +554,7 @@ class BatchRunner:
         # Create batches
         self.batches = self._create_batches()
         
-        print("📊 Batch Runner Initialized")
+        print(f"📊 Batch Runner Initialized")
         print(f"   Dataset: {self.dataset_file} ({len(self.dataset)} prompts)")
         print(f"   Batch size: {self.batch_size}")
         print(f"   Total batches: {len(self.batches)}")
@@ -701,13 +648,14 @@ class BatchRunner:
             lock (Lock): Optional lock for thread-safe access
         """
         checkpoint_data["last_updated"] = datetime.now().isoformat()
-
-        from utils import atomic_json_write
+        
         if lock:
             with lock:
-                atomic_json_write(self.checkpoint_file, checkpoint_data)
+                with open(self.checkpoint_file, 'w', encoding='utf-8') as f:
+                    json.dump(checkpoint_data, f, indent=2, ensure_ascii=False)
         else:
-            atomic_json_write(self.checkpoint_file, checkpoint_data)
+            with open(self.checkpoint_file, 'w', encoding='utf-8') as f:
+                json.dump(checkpoint_data, f, indent=2, ensure_ascii=False)
     
     def _scan_completed_prompts_by_content(self) -> set:
         """
@@ -827,20 +775,18 @@ class BatchRunner:
             print("=" * 70)
             print(f"   Original dataset size:     {len(self.dataset):,} prompts")
             print(f"   Already completed:         {len(skipped_indices):,} prompts")
-            print("   ─────────────────────────────────────────")
+            print(f"   ─────────────────────────────────────────")
             print(f"   🎯 RESUMING WITH:          {len(filtered_entries):,} prompts")
             print(f"   New batches created:       {len(batches_to_process)}")
             print("=" * 70 + "\n")
         
-        # Load existing checkpoint (so resume doesn't clobber prior progress)
-        checkpoint_data = self._load_checkpoint()
-        if checkpoint_data.get("run_name") != self.run_name:
-            checkpoint_data = {
-                "run_name": self.run_name,
-                "completed_prompts": [],
-                "batch_stats": {},
-                "last_updated": None
-            }
+        # Initialize checkpoint data (needed for saving at the end)
+        checkpoint_data = {
+            "run_name": self.run_name,
+            "completed_prompts": [],
+            "batch_stats": {},
+            "last_updated": None
+        }
         
         # Prepare configuration for workers
         config = {
@@ -862,7 +808,7 @@ class BatchRunner:
         }
         
         # For backward compatibility, still track by index (but this is secondary to content matching)
-        completed_prompts_set = set(checkpoint_data.get("completed_prompts", []))
+        completed_prompts_set = set()
         
         # Aggregate statistics across all batches
         total_tool_stats = {}
@@ -871,9 +817,6 @@ class BatchRunner:
         
         print(f"\n🔧 Initializing {self.num_workers} worker processes...")
         
-        # Checkpoint writes happen in the parent process; keep a lock for safety.
-        checkpoint_lock = Lock()
-
         # Process batches in parallel
         with Pool(processes=self.num_workers) as pool:
             # Create tasks for each batch
@@ -889,7 +832,7 @@ class BatchRunner:
             ]
             
             print(f"✅ Created {len(tasks)} batch tasks")
-            print("🚀 Starting parallel batch processing...\n")
+            print(f"🚀 Starting parallel batch processing...\n")
             
             # Use rich Progress for better visual tracking with persistent bottom bar
             # redirect_stdout/stderr lets rich manage all output so progress bar stays clean
@@ -919,28 +862,6 @@ class BatchRunner:
                     for result in pool.imap_unordered(_process_batch_worker, tasks):
                         results.append(result)
                         progress.update(task, advance=1)
-
-                        # Incremental checkpoint update (so resume works after crash)
-                        try:
-                            batch_num = result.get('batch_num')
-                            completed = result.get('completed_prompts', []) or []
-                            completed_prompts_set.update(completed)
-
-                            if isinstance(batch_num, int):
-                                checkpoint_data.setdefault('batch_stats', {})[str(batch_num)] = {
-                                    'processed': result.get('processed', 0),
-                                    'skipped': result.get('skipped', 0),
-                                    'discarded_no_reasoning': result.get('discarded_no_reasoning', 0),
-                                }
-
-                            checkpoint_data['completed_prompts'] = sorted(completed_prompts_set)
-                            self._save_checkpoint(checkpoint_data, lock=checkpoint_lock)
-                        except Exception as ckpt_err:
-                            # Don't fail the run if checkpoint write fails
-                            print(f"⚠️  Warning: Failed to save incremental checkpoint: {ckpt_err}")
-                except Exception as e:
-                    logger.error("Batch worker failed: %s", e, exc_info=True)
-                    raise
                 finally:
                     root_logger.setLevel(original_level)
         
@@ -969,12 +890,9 @@ class BatchRunner:
             for key in total_reasoning_stats:
                 total_reasoning_stats[key] += batch_result.get("reasoning_stats", {}).get(key, 0)
         
-        # Save final checkpoint (best-effort; incremental writes already happened)
-        try:
-            checkpoint_data["completed_prompts"] = all_completed_prompts
-            self._save_checkpoint(checkpoint_data, lock=checkpoint_lock)
-        except Exception as ckpt_err:
-            print(f"⚠️  Warning: Failed to save final checkpoint: {ckpt_err}")
+        # Save final checkpoint
+        checkpoint_data["completed_prompts"] = all_completed_prompts
+        self._save_checkpoint(checkpoint_data)
         
         # Calculate success rates
         for tool_name in total_tool_stats:
@@ -1058,7 +976,7 @@ class BatchRunner:
         print(f"✅ Total trajectories in merged file: {total_entries - filtered_entries}")
         print(f"✅ Total batch files merged: {batch_files_found}")
         print(f"⏱️  Total duration: {round(time.time() - start_time, 2)}s")
-        print("\n📈 Tool Usage Statistics:")
+        print(f"\n📈 Tool Usage Statistics:")
         print("-" * 70)
         
         if total_tool_stats:
@@ -1085,7 +1003,7 @@ class BatchRunner:
         # Print reasoning coverage stats
         total_discarded = sum(r.get("discarded_no_reasoning", 0) for r in results)
         
-        print("\n🧠 Reasoning Coverage:")
+        print(f"\n🧠 Reasoning Coverage:")
         print("-" * 70)
         total_turns = total_reasoning_stats["total_assistant_turns"]
         with_reasoning = total_reasoning_stats["turns_with_reasoning"]
@@ -1102,8 +1020,8 @@ class BatchRunner:
             print(f"   🚫 Samples discarded (zero reasoning): {total_discarded:,}")
         
         print(f"\n💾 Results saved to: {self.output_dir}")
-        print("   - Trajectories: trajectories.jsonl (combined)")
-        print("   - Individual batches: batch_*.jsonl (for debugging)")
+        print(f"   - Trajectories: trajectories.jsonl (combined)")
+        print(f"   - Individual batches: batch_*.jsonl (for debugging)")
         print(f"   - Statistics: {self.stats_file.name}")
         print(f"   - Checkpoint: {self.checkpoint_file.name}")
 
@@ -1113,7 +1031,7 @@ def main(
     batch_size: int = None,
     run_name: str = None,
     distribution: str = "default",
-    model: str = "anthropic/claude-sonnet-4.6",
+    model: str = "anthropic/claude-sonnet-4-20250514",
     api_key: str = None,
     base_url: str = "https://openrouter.ai/api/v1",
     max_turns: int = 10,
@@ -1156,7 +1074,7 @@ def main(
         providers_order (str): Comma-separated list of OpenRouter providers to try in order (e.g. "anthropic,openai,google")
         provider_sort (str): Sort providers by "price", "throughput", or "latency" (OpenRouter only)
         max_tokens (int): Maximum tokens for model responses (optional, uses model default if not set)
-        reasoning_effort (str): OpenRouter reasoning effort level: "xhigh", "high", "medium", "low", "minimal", "none" (default: "medium")
+        reasoning_effort (str): OpenRouter reasoning effort level: "xhigh", "high", "medium", "low", "minimal", "none" (default: "xhigh")
         reasoning_disabled (bool): Completely disable reasoning/thinking tokens (default: False)
         prefill_messages_file (str): Path to JSON file containing prefill messages (list of {role, content} dicts)
         max_samples (int): Only process the first N samples from the dataset (optional, processes all if not set)
@@ -1217,7 +1135,7 @@ def main(
     providers_order_list = [p.strip() for p in providers_order.split(",")] if providers_order else None
     
     # Build reasoning_config from CLI flags
-    # --reasoning_disabled takes priority, then --reasoning_effort, then default (medium)
+    # --reasoning_disabled takes priority, then --reasoning_effort, then default (xhigh)
     reasoning_config = None
     if reasoning_disabled:
         # Completely disable reasoning/thinking tokens
@@ -1239,7 +1157,7 @@ def main(
             with open(prefill_messages_file, 'r', encoding='utf-8') as f:
                 prefill_messages = json.load(f)
             if not isinstance(prefill_messages, list):
-                print("❌ Error: prefill_messages_file must contain a JSON array of messages")
+                print(f"❌ Error: prefill_messages_file must contain a JSON array of messages")
                 return
             print(f"💬 Loaded {len(prefill_messages)} prefill messages from {prefill_messages_file}")
         except Exception as e: