fix(pet): detect terminals more broadly

incl. windows terminal and ghostty

.github/actions/detect-changes/action.yml

@@ -0,0 +1,62 @@
+name: Detect affected areas
+description: >-
+  Classify a PR's changed files into CI work lanes (python, frontend, site,
+  scan, deps, mcp_catalog) so the orchestrator can conditionally call only
+  the sub-workflows a PR can affect. Outputs are always "true" on push/dispatch
+  events and fail open (everything "true") when the diff cannot be computed.
+
+outputs:
+  python:
+    description: Run Python tests / ruff / ty / windows-footguns.
+    value: ${{ steps.classify.outputs.python }}
+  frontend:
+    description: Run the TypeScript typecheck matrix + desktop build.
+    value: ${{ steps.classify.outputs.frontend }}
+  docker_meta:
+    description: Docker setup and meta files have changed.
+    value: ${{ steps.classify.outputs.docker_meta }}
+  site:
+    description: Build the Docusaurus docs site.
+    value: ${{ steps.classify.outputs.site }}
+  scan:
+    description: Run the supply-chain critical-pattern scanner.
+    value: ${{ steps.classify.outputs.scan }}
+  deps:
+    description: Check pyproject.toml dependency upper bounds.
+    value: ${{ steps.classify.outputs.deps }}
+  mcp_catalog:
+    description: Require MCP catalog security review label.
+    value: ${{ steps.classify.outputs.mcp_catalog }}
+
+runs:
+  using: composite
+  steps:
+    - name: Classify changed files
+      id: classify
+      shell: bash
+      env:
+        GH_TOKEN: ${{ github.token }}
+        REPO: ${{ github.repository }}
+        EVENT_NAME: ${{ github.event_name }}
+        BASE_SHA: ${{ github.event.pull_request.base.sha }}
+        HEAD_SHA: ${{ github.event.pull_request.head.sha }}
+      run: |
+        set -euo pipefail
+
+        # Only pull_request events are gated. Other events (push, release,
+        # dispatch) leave CHANGED empty, so the classifier fails open and every
+        # lane runs. Post-merge / on-demand validation is never weakened.
+        if [ "$EVENT_NAME" = "pull_request" ]; then
+          # Use the compare endpoint with the pinned base/head SHAs from the
+          # event payload instead of the "current PR files" endpoint. The SHAs
+          # are frozen at trigger time, so the file list is deterministic even
+          # if the PR receives a new push between trigger and detect.
+          CHANGED="$(gh api \
+            --paginate \
+            "repos/${REPO}/compare/${BASE_SHA}...${HEAD_SHA}" \
+            --jq '.files[].filename' || true)"
+        fi
+
+        echo "Changed files:"
+        printf '%s\n' "${CHANGED:-(none)}"
+        printf '%s\n' "${CHANGED:-}" | python3 scripts/ci/classify_changes.py

.github/actions/retry/action.yml

@@ -0,0 +1,50 @@
+name: Retry a flaky command
+description: >-
+  Run a shell command, retrying on non-zero exit. For dependency installs
+  (npm ci, uv sync) whose only failures are transient network/toolchain
+  flakes — a node-gyp header fetch, a registry blip — so CI self-heals
+  instead of needing a manual re-run.
+
+inputs:
+  command:
+    description: Shell command to run (and retry).
+    required: true
+  attempts:
+    description: Max attempts before giving up.
+    default: "3"
+  delay:
+    description: Seconds to wait between attempts.
+    default: "10"
+  working-directory:
+    description: Directory to run in.
+    default: "."
+
+runs:
+  using: composite
+  steps:
+    - shell: bash
+      working-directory: ${{ inputs.working-directory }}
+      # command goes through env, never interpolated into the script body, so
+      # a command with quotes/specials can't break or inject into the runner.
+      env:
+        _CMD: ${{ inputs.command }}
+        _ATTEMPTS: ${{ inputs.attempts }}
+        _DELAY: ${{ inputs.delay }}
+      run: |
+        set -uo pipefail
+        n=0
+        while :; do
+          n=$((n + 1))
+          echo "::group::attempt $n/$_ATTEMPTS: $_CMD"
+          if bash -c "$_CMD"; then
+            echo "::endgroup::"
+            exit 0
+          fi
+          echo "::endgroup::"
+          if [ "$n" -ge "$_ATTEMPTS" ]; then
+            echo "::error::failed after $n attempts: $_CMD"
+            exit 1
+          fi
+          echo "::warning::attempt $n failed; retrying in ${_DELAY}s: $_CMD"
+          sleep "$_DELAY"
+        done

.github/workflows/build-windows-installer.yml

@@ -1,100 +0,0 @@
-name: Build Windows Installer
-
-on:
-  workflow_dispatch:
-
-permissions:
-  contents: read
-
-jobs:
-  # Gate: workflow_dispatch is already restricted to users with write access,
-  # but we want ADMIN-only. Explicitly check the triggering actor's repo
-  # permission via the API and fail fast for anyone below admin.
-  authorize:
-    name: Authorize (admins only)
-    runs-on: ubuntu-latest
-    timeout-minutes: 5
-    steps:
-      - name: Check actor is a repo admin
-        env:
-          GH_TOKEN: ${{ github.token }}
-          ACTOR: ${{ github.actor }}
-        run: |
-          set -euo pipefail
-          perm=$(gh api \
-            "repos/${{ github.repository }}/collaborators/${ACTOR}/permission" \
-            --jq '.permission')
-          echo "Actor '${ACTOR}' has permission: ${perm}"
-          if [ "${perm}" != "admin" ]; then
-            echo "::error::'${ACTOR}' is not a repo admin (permission=${perm}). Refusing to build/sign."
-            exit 1
-          fi
-          echo "Authorized: '${ACTOR}' is an admin."
-
-  build:
-    name: Hermes-Setup.exe
-    needs: authorize
-    runs-on: windows-latest
-    timeout-minutes: 30
-    permissions:
-      contents: read
-      # Required for OIDC auth to Azure (azure/login federated credentials).
-      id-token: write
-
-    steps:
-      - name: Checkout code
-        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
-
-      - name: Setup Node.js
-        uses: actions/setup-node@49933ea5288caeca8642d1e84afbd3f7d6820020  # v4
-        with:
-          node-version: 22
-          cache: npm
-
-      - name: Install npm dependencies
-        run: npm ci
-
-      - name: Setup Rust
-        uses: dtolnay/rust-toolchain@29eef336d9b2848a0b548edc03f92a220660cdb8  # stable
-
-      - name: Cache Rust targets
-        uses: Swatinem/rust-cache@e18b497796c12c097a38f9edb9d0641fb99eee32  # v2
-        with:
-          workspaces: apps/bootstrap-installer/src-tauri
-
-      - name: Build installer
-        run: npm run tauri:build
-        working-directory: apps/bootstrap-installer
-
-      - name: Azure login (OIDC)
-        uses: azure/login@a457da9ea143d694b1b9c7c869ebb04ebe844ef5  # v2
-        with:
-          client-id: ${{ secrets.AZURE_CLIENT_ID }}
-          tenant-id: ${{ secrets.AZURE_TENANT_ID }}
-          subscription-id: ${{ secrets.AZURE_SUBSCRIPTION_ID }}
-
-      - name: Sign Hermes-Setup.exe with Azure Artifact Signing
-        uses: azure/artifact-signing-action@c7ab2a863ab5f9a846ddb8265964877ef296ee82  # v2
-        with:
-          endpoint: ${{ vars.AZURE_SIGNING_ENDPOINT }}
-          signing-account-name: ${{ vars.AZURE_SIGNING_ACCOUNT_NAME }}
-          certificate-profile-name: ${{ vars.AZURE_SIGNING_CERTIFICATE_PROFILE }}
-          # Sign both the raw exe and the bundled NSIS installer.
-          files-folder: ${{ github.workspace }}\apps\bootstrap-installer\src-tauri\target\release
-          files-folder-filter: exe
-          files-folder-recurse: true
-          file-digest: SHA256
-          timestamp-rfc3161: http://timestamp.acs.microsoft.com
-          timestamp-digest: SHA256
-
-      - name: Upload NSIS installer
-        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a  # v7.0.1
-        with:
-          name: Hermes-Setup-installer
-          path: apps/bootstrap-installer/src-tauri/target/release/bundle/nsis/*.exe
-
-      - name: Upload raw exe
-        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a  # v7.0.1
-        with:
-          name: Hermes-Setup-exe
-          path: apps/bootstrap-installer/src-tauri/target/release/Hermes-Setup.exe

.github/workflows/ci.yml

@@ -0,0 +1,145 @@
+name: CI
+
+# Orchestrator workflow. Runs ``detect-changes`` once, then conditionally
+# calls the sub-workflows that a PR can actually affect. A final
+# ``all-checks-pass`` gate job aggregates results so branch protection only
+# needs to require a single check.
+#
+# Sub-workflows are triggered via ``workflow_call`` and keep their own job
+# definitions, matrices, and concurrency settings. They no longer have
+# ``push:`` / ``pull_request:`` triggers of their own — everything flows
+# through this file.
+
+on:
+  pull_request:
+  push:
+    branches: [main]
+
+permissions:
+  contents: read
+  pull-requests: write # needed by lint (PR comment) + supply-chain (PR comment)
+  actions: read # needed by osv-scanner (SARIF upload)
+  security-events: write # needed by osv-scanner (SARIF upload)
+
+concurrency:
+  group: ci-${{ github.ref }}
+  cancel-in-progress: ${{ github.event_name == 'pull_request' }}
+
+jobs:
+  # ─────────────────────────────────────────────────────────────────────
+  # detect: run the classifier once. Every downstream job reads its outputs
+  # to decide whether to run. On push/dispatch the classifier fails open
+  # (all lanes true) so post-merge validation is never weakened.
+  # ─────────────────────────────────────────────────────────────────────
+  detect:
+    runs-on: ubuntu-latest
+    outputs:
+      python: ${{ steps.classify.outputs.python }}
+      frontend: ${{ steps.classify.outputs.frontend }}
+      site: ${{ steps.classify.outputs.site }}
+      scan: ${{ steps.classify.outputs.scan }}
+      deps: ${{ steps.classify.outputs.deps }}
+      docker_meta: ${{ steps.classify.outputs.docker_meta }}
+      mcp_catalog: ${{ steps.classify.outputs.mcp_catalog }}
+      event_name: ${{ github.event_name }}
+    steps:
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+      - name: Detect affected areas
+        id: classify
+        uses: ./.github/actions/detect-changes
+
+  # ─────────────────────────────────────────────────────────────────────
+  # Lane-gated sub-workflows. Each runs in parallel after detect finishes.
+  # Skipped workflows (if condition is false) don't spin up runners.
+  # ─────────────────────────────────────────────────────────────────────
+  tests:
+    needs: detect
+    if: needs.detect.outputs.python == 'true'
+    uses: ./.github/workflows/tests.yml
+
+  lint:
+    needs: detect
+    if: needs.detect.outputs.python == 'true'
+    uses: ./.github/workflows/lint.yml
+    with:
+      event_name: ${{ needs.detect.outputs.event_name }}
+
+  typecheck:
+    needs: detect
+    if: needs.detect.outputs.frontend == 'true'
+    uses: ./.github/workflows/typecheck.yml
+
+  docs-site:
+    needs: detect
+    if: needs.detect.outputs.site == 'true'
+    uses: ./.github/workflows/docs-site-checks.yml
+
+  history-check:
+    needs: detect
+    if: needs.detect.outputs.event_name == 'pull_request'
+    uses: ./.github/workflows/history-check.yml
+
+  contributor-check:
+    needs: detect
+    if: needs.detect.outputs.python == 'true'
+    uses: ./.github/workflows/contributor-check.yml
+
+  uv-lockfile:
+    needs: detect
+    uses: ./.github/workflows/uv-lockfile-check.yml
+
+  docker-lint:
+    needs: detect
+    if: needs.detect.outputs.docker_meta == 'true'
+    uses: ./.github/workflows/docker-lint.yml
+
+  supply-chain:
+    needs: detect
+    if: needs.detect.outputs.event_name == 'pull_request' && (needs.detect.outputs.scan == 'true' || needs.detect.outputs.deps == 'true' || needs.detect.outputs.mcp_catalog == 'true')
+    uses: ./.github/workflows/supply-chain-audit.yml
+    with:
+      event_name: ${{ needs.detect.outputs.event_name }}
+      scan: ${{ needs.detect.outputs.scan == 'true' }}
+      deps: ${{ needs.detect.outputs.deps == 'true' }}
+      mcp_catalog: ${{ needs.detect.outputs.mcp_catalog == 'true' }}
+
+  osv-scanner:
+    needs: detect
+    uses: ./.github/workflows/osv-scanner.yml
+
+  # ─────────────────────────────────────────────────────────────────────
+  # Gate: runs after everything. ``if: always()`` ensures it reports a
+  # status even when some deps were skipped. Only actual ``failure``
+  # results cause it to fail; ``skipped`` is treated as success.
+  #
+  # Branch protection should require ONLY this check.
+  # ─────────────────────────────────────────────────────────────────────
+  all-checks-pass:
+    name: All required checks pass
+    needs:
+      - tests
+      - lint
+      - typecheck
+      - docs-site
+      - history-check
+      - contributor-check
+      - uv-lockfile
+      - docker-lint
+      - supply-chain
+      - osv-scanner
+    if: always()
+    runs-on: ubuntu-latest
+    steps:
+      - name: Evaluate job results
+        env:
+          RESULTS: ${{ toJSON(needs.*.result) }}
+        run: |
+          echo "$RESULTS" | python3 -c "
+          import json, sys
+          results = json.load(sys.stdin)
+          failed = [r for r in results if r == 'failure']
+          if failed:
+              print(f'::error::{len(failed)} job(s) failed')
+              sys.exit(1)
+          print('All checks passed (or were skipped)')
+          "

.github/workflows/contributor-check.yml

@@ -1,11 +1,8 @@
 name: Contributor Attribution Check
 
 on:
-  # No paths filter — the job must always run so the required check
-  # reports a status (path-gated workflows leave checks "pending" forever
-  # when no matching files change, which blocks merge).
-  pull_request:
-    branches: [main]
+  workflow_call:
+
 permissions:
   contents: read
 
@@ -17,21 +14,7 @@ jobs:
         with:
           fetch-depth: 0  # Full history needed for git log
 
-      - name: Check if relevant files changed
-        id: filter
-        run: |
-          BASE="${{ github.event.pull_request.base.sha }}"
-          HEAD="${{ github.event.pull_request.head.sha }}"
-          CHANGED=$(git diff --name-only "$BASE"..."$HEAD" -- '*.py' '**/*.py' '.github/workflows/contributor-check.yml' || true)
-          if [ -n "$CHANGED" ]; then
-            echo "run=true" >> "$GITHUB_OUTPUT"
-          else
-            echo "run=false" >> "$GITHUB_OUTPUT"
-            echo "No Python files changed, skipping attribution check."
-          fi
-
       - name: Check for unmapped contributor emails
-        if: steps.filter.outputs.run == 'true'
         run: |
           # Get the merge base between this PR and main
           MERGE_BASE=$(git merge-base origin/main HEAD)

.github/workflows/docker-lint.yml

@@ -11,19 +11,7 @@ name: Docker / shell lint
 # activate script doesn't exist at lint time.
 
 on:
-  push:
-    branches: [main]
-    paths:
-      - Dockerfile
-      - docker/**
-      - .hadolint.yaml
-      - .github/workflows/docker-lint.yml
-
-  # No paths filter — the job must always run so the required check
-  # reports a status (path-gated workflows leave checks "pending" forever
-  # when no matching files change, which blocks merge).
-  pull_request:
-    branches: [main]
+  workflow_call:
 
 permissions:
   contents: read

.github/workflows/docker-publish.yml

@@ -16,7 +16,6 @@ on:
   # reports a status (path-gated workflows leave checks "pending" forever
   # when no matching files change, which blocks merge).
   pull_request:
-    branches: [main]
 
   release:
     types: [published]
@@ -56,13 +55,21 @@ jobs:
       - name: Checkout code
         uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
 
+      # The image build + smoke test + integration tests run ONLY on
+      # push-to-main and release — never on PRs. They are the heaviest jobs
+      # in CI (~15-45 min) and a broken build surfaces on the main push (and
+      # is gated pre-merge by docker-lint + uv-lockfile-check). Every step
+      # below is skipped on PRs, so the job still reports green and the
+      # required check never hangs.
       - name: Set up Docker Buildx
+        if: github.event_name != 'pull_request'
         uses: docker/setup-buildx-action@8d2750c68a42422c14e847fe6c8ac0403b4cbd6f  # v3
 
       # Build once, load into the local daemon for smoke testing.  Cached
       # to gha with a per-arch scope; the push step below reuses every
       # layer from this build.
       - name: Build image (amd64, smoke test)
+        if: github.event_name != 'pull_request'
         uses: docker/build-push-action@bcafcacb16a39f128d818304e6c9c0c18556b85f  # v7.1.0
         with:
           context: .
@@ -76,6 +83,7 @@ jobs:
           cache-to: type=gha,mode=max,scope=docker-amd64
 
       - name: Smoke test image
+        if: github.event_name != 'pull_request'
         uses: ./.github/actions/hermes-smoke-test
         with:
           image: ${{ env.IMAGE_NAME }}:test
@@ -102,12 +110,15 @@ jobs:
       # cheapest path to coverage on every PR that touches docker code.
       # ---------------------------------------------------------------------
       - name: Install uv (for docker tests)
+        if: github.event_name != 'pull_request'
         uses: astral-sh/setup-uv@d4b2f3b6ecc6e67c4457f6d3e41ec42d3d0fcb86  # v5
 
       - name: Set up Python 3.11 (for docker tests)
+        if: github.event_name != 'pull_request'
         run: uv python install 3.11
 
       - name: Install Python dependencies (for docker tests)
+        if: github.event_name != 'pull_request'
         run: |
           uv venv .venv --python 3.11
           source .venv/bin/activate
@@ -118,6 +129,7 @@ jobs:
           uv pip install -e ".[dev]"
 
       - name: Run docker integration tests
+        if: github.event_name != 'pull_request'
         env:
           # Skip rebuild; use the image already loaded by the build step.
           HERMES_TEST_IMAGE: ${{ env.IMAGE_NAME }}:test
@@ -190,7 +202,9 @@ jobs:
       - name: Checkout code
         uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
 
+      # arm64 build runs only on push-to-main and release (see build-amd64).
       - name: Set up Docker Buildx
+        if: github.event_name != 'pull_request'
         uses: docker/setup-buildx-action@8d2750c68a42422c14e847fe6c8ac0403b4cbd6f  # v3
 
       # Log in to ghcr.io so the registry-backed build cache below can be
@@ -201,41 +215,21 @@ jobs:
       # crashed the build before the smoke test (the reason the gha cache
       # was removed from arm64 PRs in the first place).
       - name: Log in to ghcr.io (build cache)
+        if: github.event_name != 'pull_request'
         uses: docker/login-action@4907a6ddec9925e35a0a9e82d7399ccc52663121  # v4.1.0
         with:
           registry: ghcr.io
           username: ${{ github.actor }}
           password: ${{ secrets.GITHUB_TOKEN }}
 
-      # Build once, load into the local daemon for smoke testing.
-      #
-      # PR builds use the registry-backed cache READ-ONLY (cache-from only):
-      # they pull warm layers pushed by the most recent main build but never
-      # write, so rapid PR pushes don't race on cache writes or pollute the
-      # cache ref.  This restores warm-cache speed to arm64 PR builds (which
-      # were running fully uncached and were ~45% slower than amd64, making
-      # them the job most often cancelled on supersede).
+      # Build once, load into the local daemon for smoke testing, then push
+      # by digest below. Reads AND writes the registry-backed cache so the
+      # push reuses layers from this build and the next build starts warm.
       #
       # Registry cache (type=registry on ghcr.io) is used instead of the gha
       # cache that previously broke here: its credential is the job-lifetime
       # GITHUB_TOKEN, not a short-lived SAS token, so the cold-build-outlives-
       # token failure mode cannot recur.
-      - name: Build image (arm64, smoke test, cache read-only PR)
-        if: github.event_name == 'pull_request'
-        uses: docker/build-push-action@bcafcacb16a39f128d818304e6c9c0c18556b85f  # v7.1.0
-        with:
-          context: .
-          file: Dockerfile
-          load: true
-          platforms: linux/arm64
-          tags: ${{ env.IMAGE_NAME }}:test
-          build-args: |
-            HERMES_GIT_SHA=${{ github.sha }}
-          cache-from: type=registry,ref=ghcr.io/nousresearch/hermes-agent:buildcache-arm64
-
-      # Main/release builds read AND write the registry cache so the digest
-      # push below reuses layers from this smoke-test build, and so the next
-      # PR/main build starts warm.
       - name: Build image (arm64, smoke test, cached publish)
         if: github.event_name != 'pull_request'
         uses: docker/build-push-action@bcafcacb16a39f128d818304e6c9c0c18556b85f  # v7.1.0
@@ -251,6 +245,7 @@ jobs:
           cache-to: type=registry,ref=ghcr.io/nousresearch/hermes-agent:buildcache-arm64,mode=max
 
       - name: Smoke test image
+        if: github.event_name != 'pull_request'
         uses: ./.github/actions/hermes-smoke-test
         with:
           image: ${{ env.IMAGE_NAME }}:test

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

@@ -1,13 +1,7 @@
 name: Docs Site Checks
 
 on:
-  # No paths filter — the job must always run so the required check
-  # reports a status (path-gated workflows leave checks "pending" forever
-  # when no matching files change, which blocks merge).
-  pull_request:
-    branches: [main]
-
-  workflow_dispatch:
+  workflow_call:
 
 permissions:
   contents: read
@@ -25,15 +19,19 @@ jobs:
           cache-dependency-path: website/package-lock.json
 
       - name: Install website dependencies
-        run: npm ci
-        working-directory: website
+        uses: ./.github/actions/retry
+        with:
+          command: npm ci
+          working-directory: website
 
       - uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405 # v6.2.0
         with:
           python-version: "3.11"
 
       - name: Install ascii-guard
-        run: python -m pip install ascii-guard==2.3.0 pyyaml==6.0.3
+        uses: ./.github/actions/retry
+        with:
+          command: python -m pip install ascii-guard==2.3.0 pyyaml==6.0.3
 
       - name: Extract skill metadata for dashboard
         run: python3 website/scripts/extract-skills.py

.github/workflows/history-check.yml

@@ -14,11 +14,7 @@ name: History Check
 # the PR head and main to be non-empty.
 
 on:
-  # No paths filter — the job must always run so the required check
-  # reports a status (path-gated workflows leave checks "pending" forever
-  # when no matching files change, which blocks merge).
-  pull_request:
-    branches: [main]
+  workflow_call:
 
 permissions:
   contents: read

.github/workflows/lint.yml

@@ -9,18 +9,12 @@ name: Lint (ruff + ty)
 #      enforcement fails.
 
 on:
-  push:
-    branches: [main]
-    paths-ignore:
-      - "**/*.md"
-      - "docs/**"
-      - "website/**"
-
-  # No paths filter — the job must always run so the required check
-  # reports a status (path-gated workflows leave checks "pending" forever
-  # when no matching files change, which blocks merge).
-  pull_request:
-    branches: [main]
+  workflow_call:
+    inputs:
+      event_name:
+        description: The event name from the calling orchestrator (pull_request or push).
+        type: string
+        required: true
 
 permissions:
   contents: read
@@ -33,6 +27,7 @@ concurrency:
 jobs:
   lint-diff:
     name: ruff + ty diff
+    if: inputs.event_name == 'pull_request'
     runs-on: ubuntu-latest
     timeout-minutes: 10
     steps:
@@ -45,16 +40,16 @@ jobs:
         uses: astral-sh/setup-uv@d4b2f3b6ecc6e67c4457f6d3e41ec42d3d0fcb86 # v5
 
       - name: Install ruff + ty
-        run: |
-          uv tool install ruff
-          uv tool install ty
+        uses: ./.github/actions/retry
+        with:
+          command: uv tool install ruff && uv tool install ty
 
       - name: Determine base ref
         id: base
         run: |
           # For PRs, diff against the merge base with the target branch.
           # For pushes to main, diff against the previous commit on main.
-          if [ "${{ github.event_name }}" = "pull_request" ]; then
+          if [ "${{ inputs.event_name }}" = "pull_request" ]; then
             BASE_SHA=$(git merge-base "origin/${{ github.base_ref }}" HEAD)
             BASE_REF="origin/${{ github.base_ref }}"
           else
@@ -110,7 +105,7 @@ jobs:
             --base-ty   .lint-reports/base/ty.json \
             --head-ty   .lint-reports/head/ty.json \
             --base-ref  "${{ steps.base.outputs.ref }}" \
-            --head-ref  "${{ github.event_name == 'pull_request' && github.head_ref || github.ref_name }}" \
+            --head-ref  "${{ inputs.event_name == 'pull_request' && github.head_ref || github.ref_name }}" \
             --output    .lint-reports/summary.md
           cat .lint-reports/summary.md >> "$GITHUB_STEP_SUMMARY"
 
@@ -122,7 +117,7 @@ jobs:
           retention-days: 14
 
       - name: Post / update PR comment
-        if: github.event_name == 'pull_request' && github.event.pull_request.head.repo.full_name == github.repository
+        if: inputs.event_name == 'pull_request' && github.event.pull_request.head.repo.full_name == github.repository
         continue-on-error: true
         uses: actions/github-script@60a0d83039c74a4aee543508d2ffcb1c3799cdea # v7
         with:
@@ -172,7 +167,9 @@ jobs:
         uses: astral-sh/setup-uv@d4b2f3b6ecc6e67c4457f6d3e41ec42d3d0fcb86 # v5
 
       - name: Install ruff
-        run: uv tool install ruff
+        uses: ./.github/actions/retry
+        with:
+          command: uv tool install ruff
 
       - name: ruff check .
         # No --exit-zero, no || true. Exit code propagates to the job,

.github/workflows/osv-scanner.yml

@@ -1,8 +1,8 @@
 name: OSV-Scanner
 
 # Scans lockfiles (uv.lock, package-lock.json) against the OSV vulnerability
-# database. Runs on every PR that touches a lockfile and on a weekly schedule
-# against main.
+# database. Runs on every PR/push (via the ci.yml orchestrator's workflow_call)
+# and on a weekly schedule against main.
 #
 # This is detection-only — OSV-Scanner does NOT open PRs or modify pins.
 # It reports known CVEs in currently-pinned dependency versions so we can
@@ -10,9 +10,9 @@ name: OSV-Scanner
 # (full SHA / exact version) is preserved; only the notification signal
 # is added.
 #
-# Complements the existing supply-chain-audit.yml workflow (which scans
-# for malicious code patterns in PR diffs) by covering the orthogonal
-# "currently-pinned dep became known-vulnerable" case.
+# Complements the supply-chain-audit.yml workflow (which scans for malicious
+# code patterns in PR diffs) by covering the orthogonal "currently-pinned
+# dep became known-vulnerable" case.
 #
 # Uses Google's officially-recommended reusable workflow, pinned by SHA.
 # Findings land in the repo's Security tab (Code Scanning > OSV-Scanner).
@@ -20,19 +20,7 @@ name: OSV-Scanner
 # vulnerabilities in pinned deps that we may need to patch deliberately.
 
 on:
-  # No paths filter — the job must always run so the required check
-  # reports a status (path-gated workflows leave checks "pending" forever
-  # when no matching files change, which blocks merge).
-  pull_request:
-    branches: [main]
-  push:
-    branches: [main]
-    paths:
-      - "uv.lock"
-      - "pyproject.toml"
-      - "package.json"
-      - "package-lock.json"
-      - "website/package-lock.json"
+  workflow_call:
   schedule:
     # Weekly scan against main — catches CVEs published after merge for
     # deps that haven't changed since.

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

@@ -1,16 +1,5 @@
 name: Supply Chain Audit
 
-on:
-  # No paths filter — the jobs must always run so required checks
-  # report a status (path-gated workflows leave checks "pending" forever
-  # when no matching files change, which blocks merge).
-  pull_request:
-    types: [opened, synchronize, reopened]
-
-permissions:
-  pull-requests: write
-  contents: read
-
 # Narrow, high-signal scanner. Only fires on critical indicators of supply
 # chain attacks (e.g. the litellm-style payloads). Low-signal heuristics
 # (plain base64, plain exec/eval, dependency/Dockerfile/workflow edits,
@@ -19,56 +8,40 @@ permissions:
 # the scanner. Keep this file's checks ruthlessly narrow: if you find
 # yourself adding WARNING-tier patterns here again, make a separate
 # advisory-only workflow instead.
+#
+# Path-gating is handled centrally by the ``ci.yml`` orchestrator's
+# ``detect`` job. The orchestrator passes ``scan`` / ``deps`` /
+# ``mcp_catalog`` booleans as inputs; this workflow's jobs gate on those
+# inputs instead of re-computing the diff.
+
+on:
+  workflow_call:
+    inputs:
+      event_name:
+        description: The event name from the calling orchestrator.
+        type: string
+        required: true
+      scan:
+        description: Whether supply-chain-relevant files changed.
+        type: boolean
+        required: true
+      deps:
+        description: Whether pyproject.toml changed.
+        type: boolean
+        required: true
+      mcp_catalog:
+        description: Whether the MCP catalog / installer changed.
+        type: boolean
+        required: true
+
+permissions:
+  pull-requests: write
+  contents: read
 
 jobs:
-  # ── Path filter (shared by both scan and dep-bounds) ───────────────
-  changes:
-    runs-on: ubuntu-latest
-    outputs:
-      # True when any file the scanner cares about changed in this PR
-      scan: ${{ steps.filter.outputs.scan }}
-      # True when pyproject.toml changed in this PR
-      deps: ${{ steps.filter.outputs.deps }}
-      # True when the curated MCP catalog / bundled MCP manifests changed.
-      mcp_catalog: ${{ steps.filter.outputs.mcp_catalog }}
-    steps:
-      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
-        with:
-          fetch-depth: 0
-      - name: Check for relevant file changes
-        id: filter
-        run: |
-          BASE="${{ github.event.pull_request.base.sha }}"
-          HEAD="${{ github.event.pull_request.head.sha }}"
-          SCAN_FILES=$(git diff --name-only "$BASE"..."$HEAD" -- \
-            '*.py' '**/*.py' '*.pth' '**/*.pth' \
-            'setup.py' 'setup.cfg' \
-            'sitecustomize.py' 'usercustomize.py' '__init__.pth' \
-            'pyproject.toml' || true)
-          if [ -n "$SCAN_FILES" ]; then
-            echo "scan=true" >> "$GITHUB_OUTPUT"
-          else
-            echo "scan=false" >> "$GITHUB_OUTPUT"
-          fi
-          DEPS_FILES=$(git diff --name-only "$BASE"..."$HEAD" -- 'pyproject.toml' || true)
-          if [ -n "$DEPS_FILES" ]; then
-            echo "deps=true" >> "$GITHUB_OUTPUT"
-          else
-            echo "deps=false" >> "$GITHUB_OUTPUT"
-          fi
-          MCP_CATALOG_FILES=$(git diff --name-only "$BASE"..."$HEAD" -- \
-            'optional-mcps/**' \
-            'hermes_cli/mcp_catalog.py' || true)
-          if [ -n "$MCP_CATALOG_FILES" ]; then
-            echo "mcp_catalog=true" >> "$GITHUB_OUTPUT"
-          else
-            echo "mcp_catalog=false" >> "$GITHUB_OUTPUT"
-          fi
-
   scan:
     name: Scan PR for critical supply chain risks
-    needs: changes
-    if: needs.changes.outputs.scan == 'true'
+    if: inputs.scan
     runs-on: ubuntu-latest
     steps:
       - name: Checkout
@@ -111,7 +84,7 @@ jobs:
           fi
 
           # --- base64 decode + exec/eval on the same line (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)
+          B64_EXEC_HITS=$(echo "$DIFF" | grep -n '^+' | grep -iE 'base64\.(b64decode|decodebytes|urlsafe_b64decode)' | grep -iE 'exec\(|eval\(' | head -10 || true)
           if [ -n "$B64_EXEC_HITS" ]; then
             FINDINGS="${FINDINGS}
           ### 🚨 CRITICAL: base64 decode + exec/eval combo
@@ -125,7 +98,7 @@ jobs:
           fi
 
           # --- subprocess with encoded/obfuscated command argument ---
-          PROC_HITS=$(echo "$DIFF" | grep -n '^\+' | grep -E 'subprocess\.(Popen|call|run)\s*\(' | grep -iE 'base64|\\x[0-9a-f]{2}|chr\(' | head -10 || true)
+          PROC_HITS=$(echo "$DIFF" | grep -n '^+' | grep -E 'subprocess\.(Popen|call|run)\s*\(' | grep -iE 'base64|\\x[0-9a-f]{2}|chr\(' | head -10 || true)
           if [ -n "$PROC_HITS" ]; then
             FINDINGS="${FINDINGS}
           ### 🚨 CRITICAL: subprocess with encoded/obfuscated command
@@ -187,23 +160,9 @@ jobs:
           echo "::error::CRITICAL supply chain risk patterns detected in this PR. See the PR comment for details."
           exit 1
 
-  # Gate: reports success when scan was skipped (no relevant files changed).
-  # This ensures the required check always gets a status.
-  scan-gate:
-    name: Scan PR for critical supply chain risks
-    needs: changes
-    # always() so the gate still reports SUCCESS even if `changes` fails/is
-    # skipped — without it, a failed dependency would leave the required
-    # check unreported (i.e. "pending"), the exact failure mode this fixes.
-    if: always() && needs.changes.outputs.scan != 'true'
-    runs-on: ubuntu-latest
-    steps:
-      - run: echo "No supply-chain-relevant files changed, skipping scan."
-
   dep-bounds:
     name: Check PyPI dependency upper bounds
-    needs: changes
-    if: needs.changes.outputs.deps == 'true'
+    if: inputs.deps
     runs-on: ubuntu-latest
     steps:
       - name: Checkout
@@ -253,7 +212,7 @@ jobs:
           $(cat /tmp/unbounded.txt)
           \`\`\`
 
-          **Fix:** Add an upper bound, e.g. \`\"package>=1.2.0,<2\"\`
+          **Fix:** Add an upper bound, e.g. \`"package>=1.2.0,<2"\`
 
           ---
           *See PR #2810 and CONTRIBUTING.md for the full policy rationale.*"
@@ -266,23 +225,9 @@ jobs:
           echo "::error::PyPI dependencies without upper bounds detected. Add <next_major ceiling per CONTRIBUTING.md policy."
           exit 1
 
-  # Gate: reports success when dep-bounds was skipped (no pyproject.toml changed).
-  # This ensures the required check always gets a status.
-  dep-bounds-gate:
-    name: Check PyPI dependency upper bounds
-    needs: changes
-    # always() so the gate still reports SUCCESS even if `changes` fails/is
-    # skipped — without it, a failed dependency would leave the required
-    # check unreported (i.e. "pending"), the exact failure mode this fixes.
-    if: always() && needs.changes.outputs.deps != 'true'
-    runs-on: ubuntu-latest
-    steps:
-      - run: echo "No pyproject.toml changes, skipping dependency bounds check."
-
   mcp-catalog-review:
     name: MCP catalog security review
-    needs: changes
-    if: needs.changes.outputs.mcp_catalog == 'true'
+    if: inputs.mcp_catalog
     runs-on: ubuntu-latest
     steps:
       - name: Checkout
@@ -317,11 +262,3 @@ jobs:
           gh pr comment "$PR" --body "$BODY" || echo "::warning::Could not post PR comment (expected for fork PRs)"
           echo "::error::MCP catalog changes require the mcp-catalog-reviewed label."
           exit 1
-
-  mcp-catalog-review-gate:
-    name: MCP catalog security review
-    needs: changes
-    if: always() && needs.changes.outputs.mcp_catalog != 'true'
-    runs-on: ubuntu-latest
-    steps:
-      - run: echo "No MCP catalog changes, skipping MCP catalog security review."

.github/workflows/tests.yml

@@ -1,21 +1,12 @@
 name: Tests
 
 on:
-  push:
-    branches: [main]
-    paths-ignore:
-      - "**/*.md"
-      - "docs/**"
-  # No paths filter — the job must always run so the required check
-  # reports a status (path-gated workflows leave checks "pending" forever
-  # when no matching files change, which blocks merge).
-  pull_request:
-    branches: [main]
+  workflow_call:
 
 permissions:
   contents: read
 
-# Cancel in-progress runs for the same PR/branch
+# Cancel in-progress runs for the same ref
 concurrency:
   group: tests-${{ github.ref }}
   cancel-in-progress: true
@@ -49,7 +40,7 @@ jobs:
           RG_VERSION=15.1.0
           RG_SHA256=1c9297be4a084eea7ecaedf93eb03d058d6faae29bbc57ecdaf5063921491599
           RG_TARBALL=ripgrep-${RG_VERSION}-x86_64-unknown-linux-musl.tar.gz
-          curl -sSfL -o "$RG_TARBALL" \
+          curl -sSfL --retry 3 --retry-delay 5 -o "$RG_TARBALL" \
             "https://github.com/BurntSushi/ripgrep/releases/download/${RG_VERSION}/${RG_TARBALL}"
           echo "${RG_SHA256}  ${RG_TARBALL}" | sha256sum -c -
           tar -xzf "$RG_TARBALL"
@@ -78,7 +69,9 @@ jobs:
         # fails if the lock is out of sync with pyproject.toml), giving a
         # reproducible env. It also creates .venv itself, so no separate
         # `uv venv` step is needed.
-        run: uv sync --locked --python 3.11 --extra all --extra dev
+        uses: ./.github/actions/retry
+        with:
+          command: uv sync --locked --python 3.11 --extra all --extra dev
 
       - name: Minimize uv cache
         # Optimized for CI: prunes pre-built wheels that are cheap to
@@ -171,7 +164,7 @@ jobs:
           RG_VERSION=15.1.0
           RG_SHA256=1c9297be4a084eea7ecaedf93eb03d058d6faae29bbc57ecdaf5063921491599
           RG_TARBALL=ripgrep-${RG_VERSION}-x86_64-unknown-linux-musl.tar.gz
-          curl -sSfL -o "$RG_TARBALL" \
+          curl -sSfL --retry 3 --retry-delay 5 -o "$RG_TARBALL" \
             "https://github.com/BurntSushi/ripgrep/releases/download/${RG_VERSION}/${RG_TARBALL}"
           echo "${RG_SHA256}  ${RG_TARBALL}" | sha256sum -c -
           tar -xzf "$RG_TARBALL"
@@ -200,7 +193,9 @@ jobs:
         # fails if the lock is out of sync with pyproject.toml), giving a
         # reproducible env. It also creates .venv itself, so no separate
         # `uv venv` step is needed.
-        run: uv sync --locked --python 3.11 --extra all --extra dev
+        uses: ./.github/actions/retry
+        with:
+          command: uv sync --locked --python 3.11 --extra all --extra dev
 
       - name: Minimize uv cache
         # Optimized for CI: prunes pre-built wheels that are cheap to

.github/workflows/typecheck.yml

@@ -2,13 +2,7 @@
 name: Typecheck
 
 on:
-  push:
-    branches: [main]
-  # No paths filter — the job must always run so the required check
-  # reports a status (path-gated workflows leave checks "pending" forever
-  # when no matching files change, which blocks merge).
-  pull_request:
-    branches: [main]
+  workflow_call:
 
 jobs:
   typecheck:
@@ -24,7 +18,14 @@ jobs:
         with:
           node-version: 22
           cache: npm
-      - run: npm ci
+      # --ignore-scripts: typecheck only needs the TS sources + type defs, not
+      # native builds. Skipping install scripts drops node-pty's node-gyp
+      # header fetch — the transient flake that killed this job pre-`tsc` — and
+      # is faster. retry covers the remaining registry blips.
+      - 
+        uses: ./.github/actions/retry
+        with:
+          command: npm ci --ignore-scripts
       - run: npm run --prefix ${{ matrix.package }} typecheck
 
   # Production build of the desktop renderer. `typecheck` runs `tsc` only,
@@ -41,5 +42,10 @@ jobs:
         with:
           node-version: 22
           cache: npm
-      - run: npm ci
+      # Keep install scripts here: the production build may need node-pty's
+      # native binary. retry handles the transient install-time fetch flakes.
+      - 
+        uses: ./.github/actions/retry
+        with:
+          command: npm ci
       - run: npm run --prefix apps/desktop build

.github/workflows/uv-lockfile-check.yml

@@ -44,25 +44,14 @@ name: uv.lock check
 # the same way.  Better to catch it here than after merge.
 
 on:
-  push:
-    branches: [main]
-    paths:
-      - "pyproject.toml"
-      - "uv.lock"
-      - ".github/workflows/uv-lockfile-check.yml"
-
-  # No paths filter — the job must always run so the required check
-  # reports a status (path-gated workflows leave checks "pending" forever
-  # when no matching files change, which blocks merge).
-  pull_request:
-    branches: [main]
+  workflow_call:
 
 permissions:
   contents: read
 
 concurrency:
   group: uv-lockfile-check-${{ github.event.pull_request.number || github.ref }}
-  cancel-in-progress: ${{ github.event_name == 'pull_request' }}
+  cancel-in-progress: true
 
 jobs:
   check:

Dockerfile

@@ -290,6 +290,19 @@ ENV HERMES_TUI_DIR=/opt/hermes/ui-tui
 ENV HERMES_HOME=/opt/data
 ENV HERMES_WRITE_SAFE_ROOT=/opt/data
 ENV HERMES_DISABLE_LAZY_INSTALLS=1
+# The published image seals /opt/hermes (root-owned, read-only) so a runtime
+# lazy install can't mutate the agent's own venv and brick it. But opt-in
+# backends (Firecrawl web search, Exa, Feishu, …) keep their SDKs in
+# tools/lazy_deps.py — deliberately NOT baked into [all] (see pyproject.toml
+# policy 2026-05-12: one quarantined release must not break every install).
+# Redirect those lazy installs to a writable dir on the durable data volume.
+# lazy_deps appends this dir to the END of sys.path, so a package installed
+# here can only ADD modules — it can never shadow or downgrade a core module,
+# so the sealed-venv guarantee holds even with installs re-enabled. The dir
+# is seeded + chowned to the hermes user by docker/stage2-hook.sh and lives
+# on the /opt/data volume, so it persists across container recreates / image
+# updates (an ABI stamp invalidates it if a rebuild bumps the interpreter).
+ENV HERMES_LAZY_INSTALL_TARGET=/opt/data/lazy-packages
 
 # `docker exec` privilege-drop shim. When operators run
 # `docker exec <c> hermes ...` they default to root, and any file the

acp_adapter/entry.py

@@ -23,6 +23,11 @@ except ModuleNotFoundError:
     # new code but ``uv pip install -e .`` didn't finish.  Missing bootstrap
     # means UTF-8 stdio setup is skipped on Windows; POSIX is unaffected.
     pass
+else:
+    # Stop a ``utils/``/``proxy/``/``ui/`` package in the launch directory from
+    # shadowing Hermes's own modules — ``hermes acp`` can be started from any
+    # cwd, including a project that has same-named packages on its path.
+    hermes_bootstrap.harden_import_path()
 
 import argparse
 import asyncio

acp_adapter/tools.py

@@ -74,7 +74,7 @@ _POLISHED_TOOLS = {
     "kanban_create", "kanban_show", "kanban_comment", "kanban_complete",
     "kanban_block", "kanban_link", "kanban_heartbeat",
     "yb_query_group_info", "yb_query_group_members", "yb_search_sticker",
-    "yb_send_dm", "yb_send_sticker", "mixture_of_agents",
+    "yb_send_dm", "yb_send_sticker",
 }
 
 

agent/agent_init.py

@@ -106,7 +106,12 @@ def _custom_provider_extra_body_for_agent(
     base_url: str,
     custom_providers: List[Dict[str, Any]],
 ) -> Optional[Dict[str, Any]]:
-    if (provider or "").strip().lower() != "custom":
+    provider_norm = (provider or "").strip().lower()
+    if provider_norm == "custom":
+        provider_key_filter = ""
+    elif provider_norm.startswith("custom:"):
+        provider_key_filter = provider_norm.split(":", 1)[1].strip()
+    else:
         return None
 
     target_url = _normalized_custom_base_url(base_url)
@@ -117,6 +122,13 @@ def _custom_provider_extra_body_for_agent(
     for entry in custom_providers or []:
         if not isinstance(entry, dict):
             continue
+        if provider_key_filter:
+            entry_keys = {
+                str(entry.get("provider_key", "") or "").strip().lower(),
+                str(entry.get("name", "") or "").strip().lower(),
+            }
+            if provider_key_filter not in entry_keys:
+                continue
         if _normalized_custom_base_url(entry.get("base_url")) != target_url:
             continue
         extra_body = entry.get("extra_body")
@@ -707,6 +719,15 @@ def init_agent(
                     print("🔑 Using credentials: Microsoft Entra ID")
                 elif isinstance(effective_key, str) and len(effective_key) > 12:
                     print(f"🔑 Using token: {effective_key[:8]}...{effective_key[-4:]}")
+    elif agent.provider == "moa":
+        from agent.moa_loop import MoAClient
+        agent.api_mode = "chat_completions"
+        agent.client = MoAClient(agent.model or "default")
+        agent._client_kwargs = {}
+        agent.api_key = api_key or "moa-virtual-provider"
+        agent.base_url = base_url or "moa://local"
+        if not agent.quiet_mode:
+            print(f"🤖 AI Agent initialized with MoA preset: {agent.model}")
     elif agent.api_mode == "bedrock_converse":
         # AWS Bedrock — uses boto3 directly, no OpenAI client needed.
         # Region is extracted from the base_url or defaults to us-east-1.
@@ -1506,6 +1527,7 @@ def init_agent(
     # 3. Check general plugin system (user-installed plugins)
     # 4. Fall back to built-in ContextCompressor
     _selected_engine = None
+    _copy_failed = False
     _engine_name = "compressor"  # default
     try:
         _ctx_cfg = _agent_cfg.get("context", {}) if isinstance(_agent_cfg, dict) else {}
@@ -1523,15 +1545,35 @@ def init_agent(
 
         # Try general plugin system as fallback
         if _selected_engine is None:
+            _candidate = None
             try:
                 from hermes_cli.plugins import get_plugin_context_engine
                 _candidate = get_plugin_context_engine()
-                if _candidate and _candidate.name == _engine_name:
-                    _selected_engine = _candidate
             except Exception:
-                pass
+                _candidate = None
+            if _candidate is not None and _candidate.name == _engine_name:
+                # Deep-copy the shared plugin singleton so a child agent's
+                # update_model() can't mutate the parent's compressor (#42449).
+                # Copy can fail for engines holding uncopyable state (locks, DB
+                # connections, clients); in that case fall back to the built-in
+                # compressor with an ACCURATE message rather than silently
+                # mislabelling it "not found".
+                import copy
+                try:
+                    _selected_engine = copy.deepcopy(_candidate)
+                except Exception as _copy_err:
+                    _copy_failed = True
+                    _ra().logger.warning(
+                        "Context engine '%s' could not be safely copied for this "
+                        "agent (%s) — falling back to built-in compressor. Plugin "
+                        "engines that hold uncopyable state (locks, DB connections) "
+                        "should implement __deepcopy__ to copy only mutable budget "
+                        "state.",
+                        _engine_name, _copy_err,
+                    )
+                    _selected_engine = None
 
-        if _selected_engine is None:
+        if _selected_engine is None and not _copy_failed:
             _ra().logger.warning(
                 "Context engine '%s' not found — falling back to built-in compressor",
                 _engine_name,
@@ -1621,16 +1663,27 @@ def init_agent(
             for t in agent.tools
             if isinstance(t, dict)
         }
-        for _schema in agent.context_compressor.get_tool_schemas():
-            _tname = _schema.get("name", "")
-            if _tname and _tname in _existing_tool_names:
+        from agent.memory_manager import normalize_tool_schema as _normalize_tool_schema
+        for _raw_schema in agent.context_compressor.get_tool_schemas():
+            _schema = _normalize_tool_schema(_raw_schema)
+            if _schema is None:
+                # A schema with no resolvable name (e.g. an already-wrapped
+                # entry) would append a nameless tool that strict providers
+                # 400 on, disabling the whole toolset (#47707). Skip it.
+                _ra().logger.warning(
+                    "Context engine returned a tool schema with no resolvable "
+                    "name; skipping to avoid poisoning the request (%r)",
+                    _raw_schema,
+                )
+                continue
+            _tname = _schema["name"]
+            if _tname in _existing_tool_names:
                 continue  # already registered via plugin/cache path
             _wrapped = {"type": "function", "function": _schema}
             agent.tools.append(_wrapped)
-            if _tname:
-                agent.valid_tool_names.add(_tname)
-                agent._context_engine_tool_names.add(_tname)
-                _existing_tool_names.add(_tname)
+            agent.valid_tool_names.add(_tname)
+            agent._context_engine_tool_names.add(_tname)
+            _existing_tool_names.add(_tname)
 
     # Notify context engine of session start
     if hasattr(agent, "context_compressor") and agent.context_compressor:

agent/agent_runtime_helpers.py

@@ -1697,6 +1697,27 @@ def switch_model(agent, new_model, new_provider, api_key='', base_url='', api_mo
         old_model, old_provider, new_model, new_provider,
     )
 
+    # ── Persist billing route to session DB ──
+    # The agent's _session_db / session_id may not be set in all contexts
+    # (tests, bare agents without a session DB, etc.).  This ensures the
+    # dashboard Model cards show the actual provider after a mid-session
+    # /model switch instead of the stale session-creation provider.
+    # See #48248 for the full bug description.
+    _session_db = getattr(agent, "_session_db", None)
+    _session_id = getattr(agent, "session_id", None)
+    if _session_db is not None and _session_id:
+        try:
+            _session_db.update_session_billing_route(
+                _session_id,
+                provider=agent.provider,
+                base_url=agent.base_url,
+                billing_mode=getattr(agent, "api_mode", None),
+            )
+        except Exception:
+            logger.warning(
+                "Failed to persist billing route after model switch",
+                exc_info=True,
+            )
 
 
 def invoke_tool(agent, function_name: str, function_args: dict, effective_task_id: str,

agent/anthropic_adapter.py

@@ -1297,7 +1297,15 @@ def run_oauth_setup_token() -> Optional[str]:
 # Stores credentials in ~/.hermes/.anthropic_oauth.json (our own file).
 
 _OAUTH_CLIENT_ID = "9d1c250a-e61b-44d9-88ed-5944d1962f5e"
-_OAUTH_TOKEN_URL = "https://console.anthropic.com/v1/oauth/token"
+# Anthropic migrated the OAuth token endpoint to platform.claude.com;
+# console.anthropic.com now 404s. Callers should iterate _OAUTH_TOKEN_URLS
+# (new host first, console fallback). _OAUTH_TOKEN_URL is kept as the primary
+# for backward compatibility with existing imports and now points at the live host.
+_OAUTH_TOKEN_URLS = [
+    "https://platform.claude.com/v1/oauth/token",
+    "https://console.anthropic.com/v1/oauth/token",
+]
+_OAUTH_TOKEN_URL = _OAUTH_TOKEN_URLS[0]
 _OAUTH_REDIRECT_URI = "https://console.anthropic.com/oauth/code/callback"
 _OAUTH_SCOPES = "org:create_api_key user:profile user:inference"
 _HERMES_OAUTH_FILE = get_hermes_home() / ".anthropic_oauth.json"
@@ -1395,18 +1403,34 @@ def run_hermes_oauth_login_pure() -> Optional[Dict[str, Any]]:
             "code_verifier": verifier,
         }).encode()
 
-        req = urllib.request.Request(
-            _OAUTH_TOKEN_URL,
-            data=exchange_data,
-            headers={
-                "Content-Type": "application/json",
-                "User-Agent": f"claude-cli/{_get_claude_code_version()} (external, cli)",
-            },
-            method="POST",
-        )
+        # Anthropic migrated the OAuth token endpoint to platform.claude.com;
+        # console.anthropic.com now 404s. Try the new host first, then fall
+        # back to console for older deployments (mirrors the refresh path).
+        result = None
+        last_error = None
+        for endpoint in _OAUTH_TOKEN_URLS:
+            req = urllib.request.Request(
+                endpoint,
+                data=exchange_data,
+                headers={
+                    "Content-Type": "application/json",
+                    "User-Agent": f"claude-cli/{_get_claude_code_version()} (external, cli)",
+                },
+                method="POST",
+            )
+            try:
+                with urllib.request.urlopen(req, timeout=15) as resp:
+                    result = json.loads(resp.read().decode())
+                break
+            except Exception as exc:
+                last_error = exc
+                logger.debug("Anthropic token exchange failed at %s: %s", endpoint, exc)
+                continue
 
-        with urllib.request.urlopen(req, timeout=15) as resp:
-            result = json.loads(resp.read().decode())
+        if result is None:
+            raise last_error if last_error is not None else ValueError(
+                "Anthropic token exchange failed"
+            )
     except Exception as e:
         print(f"Token exchange failed: {e}")
         return None

agent/auxiliary_client.py

@@ -101,6 +101,7 @@ class _OpenAIProxy:
 OpenAI = _OpenAIProxy()  # module-level name, resolves lazily on call/isinstance
 
 from agent.credential_pool import load_pool
+from agent.model_metadata import MINIMUM_CONTEXT_LENGTH, get_model_context_length
 from hermes_cli.config import get_hermes_home
 from hermes_constants import OPENROUTER_BASE_URL
 from utils import base_url_host_matches, base_url_hostname, env_float, model_forces_max_completion_tokens, normalize_proxy_env_vars
@@ -665,6 +666,28 @@ def _pool_runtime_base_url(entry: Any, fallback: str = "") -> str:
     return str(url or "").strip().rstrip("/")
 
 
+# Hostnames (lowercase, exact) that the auxiliary Anthropic path is allowed to
+# be pointed at via config.yaml model.base_url. Anything else falls back to the
+# Anthropic default — operators routing main-session traffic through a
+# non-Anthropic host (e.g. OpenRouter, OpenAI) with provider=anthropic in config
+# must NOT have that foreign host leak into the auxiliary client. See #52608.
+_ANTHROPIC_COMPATIBLE_HOSTS = frozenset({
+    "api.anthropic.com",
+})
+
+
+def _is_anthropic_compatible_host(url: str) -> bool:
+    """Return True if ``url``'s hostname is an Anthropic endpoint we trust for aux calls."""
+    if not url:
+        return False
+    try:
+        from urllib.parse import urlparse
+        host = (urlparse(url).hostname or "").strip().lower().rstrip(".")
+        return host in _ANTHROPIC_COMPATIBLE_HOSTS
+    except Exception:
+        return False
+
+
 def _nous_min_key_ttl_seconds() -> int:
     try:
         return max(60, int(os.getenv("HERMES_NOUS_MIN_KEY_TTL_SECONDS", "1800")))
@@ -2255,9 +2278,16 @@ def _try_anthropic(explicit_api_key: str = None) -> Tuple[Optional[Any], Optiona
     if not token:
         return None, None
 
-    # Allow base URL override from config.yaml model.base_url, but only
-    # when the configured provider is anthropic — otherwise a non-Anthropic
-    # base_url (e.g. Codex endpoint) would leak into Anthropic requests.
+    # Allow base URL override from config.yaml model.base_url, but only when:
+    #   1. the configured provider is anthropic (otherwise a non-Anthropic
+    #      base_url, e.g. Codex endpoint, would leak into Anthropic requests), AND
+    #   2. the override URL actually points at an Anthropic-compatible endpoint.
+    # Without gate (2), operators who route main-session traffic through a
+    # non-Anthropic provider that accepts Anthropic-format requests (e.g.
+    # OpenRouter at openrouter.ai/api/v1, with provider=anthropic in config.yaml)
+    # would have every auxiliary side-channel call (memory extractors,
+    # reflection, vision, title generation) 401 from the foreign host —
+    # see issue #52608.
     base_url = _pool_runtime_base_url(entry, _ANTHROPIC_DEFAULT_BASE_URL) if pool_present else _ANTHROPIC_DEFAULT_BASE_URL
     try:
         from hermes_cli.config import load_config
@@ -2267,7 +2297,7 @@ def _try_anthropic(explicit_api_key: str = None) -> Tuple[Optional[Any], Optiona
             cfg_provider = str(model_cfg.get("provider") or "").strip().lower()
             if cfg_provider == "anthropic":
                 cfg_base_url = (model_cfg.get("base_url") or "").strip().rstrip("/")
-                if cfg_base_url:
+                if cfg_base_url and _is_anthropic_compatible_host(cfg_base_url):
                     base_url = cfg_base_url
     except Exception:
         pass
@@ -2470,7 +2500,7 @@ def _is_payment_error(exc: Exception) -> bool:
     # but sometimes wrap them in 429 or other codes.
     # Daily quota exhaustion from Bedrock, Vertex AI, and similar providers
     # uses different language but is semantically identical to credit exhaustion.
-    if status in {402, 404, 429, None}:
+    if status in {402, 403, 404, 429, None}:
         if any(kw in err_lower for kw in (
             "credits", "insufficient funds",
             "can only afford", "billing",
@@ -2479,6 +2509,8 @@ def _is_payment_error(exc: Exception) -> bool:
             "balance_depleted", "no usable credits",
             "model_not_supported_on_free_tier",
             "not available on the free tier",
+            "requires a subscription", "upgrade for access",
+            "upgrade for higher limits", "reached your session usage limit",
             # Daily / monthly / weekly quota exhaustion keywords
             "quota exceeded", "quota_exceeded",
             "too many tokens per day", "daily limit",
@@ -2697,6 +2729,79 @@ def _is_model_not_found_error(exc: Exception) -> bool:
     ))
 
 
+def _is_model_incompatible_error(exc: Exception) -> bool:
+    """Detect "this route cannot serve this model" 400s (capability mismatch).
+
+    Distinct from :func:`_is_model_not_found_error` (the model does not exist
+    anywhere): here the model name is valid but the *current provider/account*
+    is structurally unable to run it. The canonical case is a configured
+    fallback that cannot run the main model — e.g. an ``openai-codex`` /
+    ChatGPT-account fallback asked to compress a ``glm-5.2`` conversation::
+
+        Error code: 400 - {'detail': "The 'glm-5.2' model is not supported
+        when using Codex with a ChatGPT account."}
+
+    The candidate authenticates fine and builds a client, so the auth and
+    payment predicates don't fire and the call would otherwise raise and
+    abort the whole auxiliary task (commonly compression — which then drops
+    middle turns and churns the session, destroying the prompt cache).
+    Treating it as a fallback-worthy capability error lets the chain skip the
+    incapable route and continue to the next candidate, mirroring the
+    context-window feasibility screen (#52392).
+
+    Billing/quota 400s belong to :func:`_is_payment_error`; "model does not
+    exist" 400s belong to :func:`_is_model_not_found_error`. This predicate
+    explicitly excludes both so the three don't overlap.
+    """
+    status = getattr(exc, "status_code", None)
+    if status not in {400, None}:
+        return False
+    err_lower = str(exc).lower()
+    # Not-found 400s ("invalid model ID", "model does not exist") are owned by
+    # _is_model_not_found_error. Billing/free-tier 400s are owned by the
+    # payment path — key on the billing keywords directly here rather than
+    # calling _is_payment_error(), because that predicate is status-gated
+    # ({402,403,404,429,None}) and would not recognise a 400-coded billing
+    # body, letting it leak into this capability bucket.
+    if _is_model_not_found_error(exc):
+        return False
+    if any(kw in err_lower for kw in (
+        "credits", "insufficient funds", "billing", "out of funds",
+        "balance_depleted", "no usable credits", "payment required",
+        "free tier", "free-tier", "not available on the free tier",
+        "model_not_supported_on_free_tier", "quota",
+    )):
+        return False
+    return any(kw in err_lower for kw in (
+        "is not supported when using",   # codex/ChatGPT-account model gating
+        "model is not supported",
+        "not supported with this",
+        "not supported for this account",
+        "model_not_supported",
+        "does not support this model",
+        "unsupported model",
+    ))
+
+
+def _is_invalid_aux_response_error(exc: Exception) -> bool:
+    """Detect provider responses that authenticated but cannot serve aux shape.
+
+    Some OpenAI-compatible routes return HTTP 200 with an empty/malformed
+    ChatCompletion instead of a normal provider error.  That is still a
+    provider/model capability failure for auxiliary tasks: downstream callers
+    need ``choices[0].message`` and should be able to continue through the
+    same fallback path as explicit model-incompatibility errors.
+    """
+    if not isinstance(exc, RuntimeError):
+        return False
+    msg = str(exc).lower()
+    return (
+        "auxiliary " in msg
+        and "llm returned invalid response" in msg
+        and "choices[0].message" in msg
+    )
+
+
 def _evict_cached_clients(provider: str) -> None:
     """Drop cached auxiliary clients for a provider so fresh creds are used."""
     normalized = _normalize_aux_provider(provider)
@@ -3147,6 +3252,88 @@ def _try_main_agent_model_fallback(
     return client, resolved_model or main_model, label
 
 
+# ── Context-window screening for runtime fallback chains (issue #52392) ──
+#
+# When the runtime auxiliary fallback chain selects a candidate that is
+# reachable but has a context window smaller than the compression task
+# requires, the call errors out instead of continuing to the next, viable
+# candidate. The startup feasibility check in
+# ``agent.conversation_compression.check_compression_model_feasibility``
+# already filters too-small auxiliary models at startup, but the runtime
+# fallback chain (``_try_configured_fallback_chain`` and
+# ``_try_main_fallback_chain``) does not apply the same filter, so
+# compression can stop at the first alive door even if the room behind it
+# is too small.
+#
+# The helpers below screen each candidate by its effective context window
+# before it is returned. ``None`` results from ``get_model_context_length``
+# are passed through (we cannot prove a model is too small, so we do not
+# block it). This preserves the existing fallback surface for
+# unrecognised/custom models while closing the gap on the well-known ones.
+
+def _task_minimum_context_length(task: Optional[str]) -> Optional[int]:
+    """Return the minimum context length required for an auxiliary task.
+
+    Only ``compression`` carries an explicit minimum today (the same
+    ``MINIMUM_CONTEXT_LENGTH`` (64K) floor that
+    ``check_compression_model_feasibility`` already enforces at startup).
+    Other tasks (``vision``, ``title_generation``, ``web_extract``,
+    ``skills_hub``, ``mcp``, ``session_search``) return ``None`` — they
+    have no per-task context floor and the runtime chain must remain
+    permissive for them.
+
+    Returns ``None`` for an empty/``None`` task name so the helper is a
+    safe no-op when called from generic sites.
+    """
+    if not task:
+        return None
+    if task == "compression":
+        return MINIMUM_CONTEXT_LENGTH
+    return None
+
+
+def _candidate_context_window(
+    provider: str,
+    model: str,
+    base_url: str = "",
+    api_key: str = "",
+) -> Optional[int]:
+    """Resolve the effective context window for a fallback candidate.
+
+    Thin wrapper around :func:`agent.model_metadata.get_model_context_length`
+    that swallows probe failures (returns ``None``). Callers treat
+    ``None`` as "unknown — pass through" so the existing fallback
+    surface is preserved when the context-length resolver chain cannot
+    determine a value (custom endpoints, models not in the registry,
+    offline endpoints).
+
+    Best-effort, never raises — the runtime fallback chain must keep
+    moving even if the resolver hits a probe error.
+    """
+    if not model:
+        return None
+    try:
+        ctx = get_model_context_length(
+            model,
+            base_url=base_url,
+            api_key=api_key,
+            provider=provider,
+        )
+    except Exception as exc:
+        logger.debug(
+            "Auxiliary fallback: could not resolve context window for %s/%s: %s",
+            provider, model, exc,
+        )
+        return None
+    # ``get_model_context_length`` returns an int (with a 256K default
+    # fallback when nothing else matches). We still propagate ``None`` if
+    # a future change returns ``Optional[int]`` — being explicit is
+    # cheap and the test suite covers both shapes.
+    if isinstance(ctx, int) and ctx > 0:
+        return ctx
+    return None
+
+
 def _try_configured_fallback_chain(
     task: str,
     failed_provider: str,
@@ -3171,6 +3358,7 @@ def _try_configured_fallback_chain(
 
     skip = failed_provider.lower().strip()
     tried = []
+    min_ctx = _task_minimum_context_length(task)
 
     for i, entry in enumerate(chain):
         if not isinstance(entry, dict):
@@ -3188,6 +3376,20 @@ def _try_configured_fallback_chain(
             fb_client, resolved_model = None, None
 
         if fb_client is not None:
+            if min_ctx is not None and resolved_model:
+                fb_ctx = _candidate_context_window(
+                    fb_provider,
+                    resolved_model,
+                    base_url=str(entry.get("base_url") or ""),
+                    api_key=_fallback_entry_api_key(entry) or "",
+                )
+                if fb_ctx is not None and fb_ctx < min_ctx:
+                    logger.info(
+                        "Auxiliary %s: skipping %s (%s context=%d < min=%d), continuing chain",
+                        task, label, resolved_model, fb_ctx, min_ctx,
+                    )
+                    tried.append(f"{label} (context too small: {fb_ctx}<{min_ctx})")
+                    continue
             logger.info(
                 "Auxiliary %s: %s on %s — configured fallback to %s (%s)",
                 task, reason, failed_provider, label, resolved_model or fb_model or "default",
@@ -3203,6 +3405,28 @@ def _try_configured_fallback_chain(
     return None, None, ""
 
 
+def _try_configured_fallback_for_unavailable_client(
+    task: Optional[str],
+    failed_provider: str,
+) -> Tuple[Optional[Any], Optional[str], str]:
+    """Try task fallback_chain when an explicit aux provider cannot build.
+
+    This covers the "no client" case before any request is sent: missing
+    raw env key, unavailable OAuth/pool credentials, or provider resolver
+    returning ``(None, None)``.  It deliberately stops at the configured
+    per-task fallback chain; the main-agent model remains the last-resort
+    runtime fallback for request-time capacity errors.
+    """
+    explicit = (failed_provider or "").strip().lower()
+    if not task or not explicit or explicit in {"auto"}:
+        return None, None, ""
+    return _try_configured_fallback_chain(
+        task,
+        explicit,
+        reason="provider unavailable",
+    )
+
+
 def _fallback_entry_api_key(entry: Dict[str, Any]) -> Optional[str]:
     """Resolve inline or env-backed API key from a fallback-chain entry."""
     explicit = str(entry.get("api_key") or "").strip()
@@ -3261,6 +3485,7 @@ def _try_main_fallback_chain(
     main_norm = (_read_main_provider() or "").strip().lower()
     skip = {p for p in (failed_norm, main_norm, "auto") if p}
     tried: List[str] = []
+    min_ctx = _task_minimum_context_length(task)
 
     for i, entry in enumerate(chain):
         if not isinstance(entry, dict):
@@ -3284,6 +3509,20 @@ def _try_main_fallback_chain(
             logger.debug("Auxiliary %s: main fallback %s failed to resolve: %s", task or "call", label, exc)
             fb_client, resolved_model = None, None
         if fb_client is not None:
+            if min_ctx is not None:
+                fb_ctx = _candidate_context_window(
+                    fb_provider,
+                    resolved_model or fb_model,
+                    base_url=str(entry.get("base_url") or ""),
+                    api_key=_fallback_entry_api_key(entry) or "",
+                )
+                if fb_ctx is not None and fb_ctx < min_ctx:
+                    logger.info(
+                        "Auxiliary %s: skipping %s (context=%d < min=%d), continuing chain",
+                        task or "call", label, fb_ctx, min_ctx,
+                    )
+                    tried.append(f"{label} (context too small: {fb_ctx}<{min_ctx})")
+                    continue
             logger.info(
                 "Auxiliary %s: %s on %s — main fallback chain to %s (%s)",
                 task or "call", reason, failed_provider or "auto", label,
@@ -5254,6 +5493,9 @@ def _validate_llm_response(response: Any, task: str = None) -> Any:
         if not choices or not hasattr(choices[0], "message"):
             raise AttributeError("missing choices[0].message")
     except (AttributeError, TypeError, IndexError) as exc:
+        recovered = _recover_aux_response_message(response)
+        if recovered is not None:
+            return recovered
         response_type = type(response).__name__
         response_preview = str(response)[:120]
         raise RuntimeError(
@@ -5265,6 +5507,64 @@ def _validate_llm_response(response: Any, task: str = None) -> Any:
     return response
 
 
+def _recover_aux_response_message(response: Any) -> Optional[Any]:
+    """Synthesize chat-completions shape from Responses-style text fields.
+
+    Auxiliary callers consume ``choices[0].message``.  Some compatible
+    endpoints return text outside ``choices`` (for example ``output_text`` or
+    ``output`` items).  Preserve that response before declaring it malformed.
+    """
+    text = _extract_aux_response_text(response)
+    if not text:
+        return None
+
+    choice = SimpleNamespace(
+        message=SimpleNamespace(content=text),
+        finish_reason=getattr(response, "finish_reason", None) or "stop",
+    )
+    try:
+        response.choices = [choice]
+        return response
+    except Exception:
+        return SimpleNamespace(
+            id=getattr(response, "id", ""),
+            model=getattr(response, "model", ""),
+            object=getattr(response, "object", "chat.completion"),
+            choices=[choice],
+            usage=getattr(response, "usage", None),
+        )
+
+
+def _extract_aux_response_text(response: Any) -> str:
+    output_text = _obj_get(response, "output_text")
+    if isinstance(output_text, str) and output_text.strip():
+        return output_text.strip()
+
+    output = _obj_get(response, "output")
+    if not isinstance(output, list):
+        return ""
+
+    parts: List[str] = []
+    for item in output:
+        item_type = _obj_get(item, "type")
+        if item_type and item_type != "message":
+            continue
+        for part in (_obj_get(item, "content") or []):
+            part_type = _obj_get(part, "type")
+            if part_type in {"output_text", "text", None}:
+                text = _obj_get(part, "text")
+                if isinstance(text, str) and text.strip():
+                    parts.append(text.strip())
+    return "\n".join(parts).strip()
+
+
+def _obj_get(obj: Any, key: str, default: Any = None) -> Any:
+    value = getattr(obj, key, default)
+    if value is default and isinstance(obj, dict):
+        value = obj.get(key, default)
+    return value
+
+
 def call_llm(
     task: str = None,
     *,
@@ -5344,21 +5644,30 @@ def call_llm(
         )
         if client is None:
             # When the user explicitly chose a non-OpenRouter provider but no
-            # credentials were found, fail fast instead of silently routing
-            # through OpenRouter (which causes confusing 404s).
+            # credentials were found, honor the task fallback_chain before
+            # raising.  Missing raw env keys are recoverable for auxiliary
+            # tasks because fallback entries may use OAuth / credential-pool
+            # auth (for example openai-codex).
             _explicit = (resolved_provider or "").strip().lower()
             if _explicit and _explicit not in {"auto", "openrouter", "custom"}:
-                raise RuntimeError(
-                    f"Provider '{_explicit}' is set in config.yaml but no API key "
-                    f"was found. Set the {_explicit.upper()}_API_KEY environment "
-                    f"variable, or switch to a different provider with `hermes model`."
+                fb_client, fb_model, fb_label = _try_configured_fallback_for_unavailable_client(
+                    task, _explicit,
                 )
+                if fb_client is not None:
+                    client, final_model = fb_client, fb_model
+                    resolved_provider = fb_label or resolved_provider
+                else:
+                    raise RuntimeError(
+                        f"Provider '{_explicit}' is set in config.yaml but no API key "
+                        f"was found. Set the {_explicit.upper()}_API_KEY environment "
+                        f"variable, or switch to a different provider with `hermes model`."
+                    )
             # For auto/custom with no credentials, try the full auto chain
             # rather than hardcoding OpenRouter (which may be depleted).
             # Pass model=None so each provider uses its own default —
             # resolved_model may be an OpenRouter-format slug that doesn't
             # work on other providers.
-            if not resolved_base_url:
+            if client is None and not resolved_base_url:
                 logger.info("Auxiliary %s: provider %s unavailable, trying auto-detection chain",
                             task or "call", resolved_provider)
                 client, final_model = _get_cached_client("auto", main_runtime=main_runtime, task=task)
@@ -5657,6 +5966,8 @@ def call_llm(
             _is_payment_error(first_err)
             or _is_connection_error(first_err)
             or _is_rate_limit_error(first_err)
+            or _is_model_incompatible_error(first_err)
+            or _is_invalid_aux_response_error(first_err)
         )
         # Respect explicit provider choice for transient errors (auth, request
         # validation, etc.) but allow fallback when the provider clearly cannot
@@ -5667,7 +5978,20 @@ def call_llm(
         is_auto = resolved_provider in {"auto", "", None}
         # Capacity errors bypass the explicit-provider gate: the provider
         # literally cannot serve this request regardless of user intent.
-        is_capacity_error = _is_payment_error(first_err) or _is_connection_error(first_err)
+        # Rate limits are included: after retries are exhausted, a 429 means
+        # the provider cannot serve this request — fall back. See #52228.
+        # Model-incompatibility 400s are also a hard capability mismatch (the
+        # route cannot run this model at all — e.g. a codex/ChatGPT-account
+        # fallback asked to compress a glm-5.2 conversation), so they bypass
+        # the explicit-provider gate and continue to the next candidate
+        # instead of aborting the auxiliary task and churning the session.
+        is_capacity_error = (
+            _is_payment_error(first_err)
+            or _is_connection_error(first_err)
+            or _is_rate_limit_error(first_err)
+            or _is_model_incompatible_error(first_err)
+            or _is_invalid_aux_response_error(first_err)
+        )
         if should_fallback and (is_auto or is_capacity_error):
             if _is_payment_error(first_err):
                 reason = "payment error"
@@ -5680,6 +6004,10 @@ def call_llm(
                 )
             elif _is_rate_limit_error(first_err):
                 reason = "rate limit"
+            elif _is_model_incompatible_error(first_err):
+                reason = "model incompatible with route"
+            elif _is_invalid_aux_response_error(first_err):
+                reason = "invalid provider response"
             else:
                 reason = "connection error"
             logger.info("Auxiliary %s: %s on %s (%s), trying fallback",
@@ -5854,12 +6182,21 @@ async def async_call_llm(
         if client is None:
             _explicit = (resolved_provider or "").strip().lower()
             if _explicit and _explicit not in {"auto", "openrouter", "custom"}:
-                raise RuntimeError(
-                    f"Provider '{_explicit}' is set in config.yaml but no API key "
-                    f"was found. Set the {_explicit.upper()}_API_KEY environment "
-                    f"variable, or switch to a different provider with `hermes model`."
+                fb_client, fb_model, fb_label = _try_configured_fallback_for_unavailable_client(
+                    task, _explicit,
                 )
-            if not resolved_base_url:
+                if fb_client is not None:
+                    client, final_model = _to_async_client(
+                        fb_client, fb_model or "", is_vision=(task == "vision")
+                    )
+                    resolved_provider = fb_label or resolved_provider
+                else:
+                    raise RuntimeError(
+                        f"Provider '{_explicit}' is set in config.yaml but no API key "
+                        f"was found. Set the {_explicit.upper()}_API_KEY environment "
+                        f"variable, or switch to a different provider with `hermes model`."
+                    )
+            if client is None and not resolved_base_url:
                 logger.info("Auxiliary %s: provider %s unavailable, trying auto-detection chain",
                             task or "call", resolved_provider)
                 client, final_model = _get_cached_client("auto", async_mode=True, main_runtime=main_runtime, task=task)
@@ -6109,12 +6446,24 @@ async def async_call_llm(
             _is_payment_error(first_err)
             or _is_connection_error(first_err)
             or _is_rate_limit_error(first_err)
+            or _is_model_incompatible_error(first_err)
+            or _is_invalid_aux_response_error(first_err)
         )
-        # Capacity errors (payment/quota/connection) bypass the explicit-provider
-        # gate — the provider cannot serve the request regardless of user intent.
+        # Capacity errors (payment/quota/connection/rate-limit) bypass the
+        # explicit-provider gate — the provider cannot serve the request
+        # regardless of user intent. Rate limits are included: after retries
+        # are exhausted, a 429 means the provider is at capacity. See #52228.
         # See #26803: daily token quota must fall back like a 402 credit error.
+        # Model-incompatibility 400s (route cannot run this model at all)
+        # bypass the gate too — see the sync call_llm() path for rationale.
         is_auto = resolved_provider in {"auto", "", None}
-        is_capacity_error = _is_payment_error(first_err) or _is_connection_error(first_err)
+        is_capacity_error = (
+            _is_payment_error(first_err)
+            or _is_connection_error(first_err)
+            or _is_rate_limit_error(first_err)
+            or _is_model_incompatible_error(first_err)
+            or _is_invalid_aux_response_error(first_err)
+        )
         if should_fallback and (is_auto or is_capacity_error):
             if _is_payment_error(first_err):
                 reason = "payment error"
@@ -6123,6 +6472,10 @@ async def async_call_llm(
                 )
             elif _is_rate_limit_error(first_err):
                 reason = "rate limit"
+            elif _is_model_incompatible_error(first_err):
+                reason = "model incompatible with route"
+            elif _is_invalid_aux_response_error(first_err):
+                reason = "invalid provider response"
             else:
                 reason = "connection error"
             logger.info("Auxiliary %s (async): %s on %s (%s), trying fallback",

agent/background_review.py

@@ -27,6 +27,131 @@ from typing import Any, Dict, List, Optional
 logger = logging.getLogger(__name__)
 
 
+# ---------------------------------------------------------------------------
+# Background-review aux-model selector + routed digest.
+#
+# The review fork runs on the MAIN model by default ("auto"), replaying the
+# full conversation — already warm in the prompt cache, so cheap cache reads.
+# Optimal and unchanged. A user can route the review to a different, cheaper
+# model via auxiliary.background_review.{provider,model}. A different model
+# cannot reuse the parent's cache (different key), so the fork is cold
+# regardless — replaying the full transcript would just cold-write it. So when
+# (and only when) routed to a different model, we replay a compact DIGEST to
+# minimise cold-written tokens. Same model -> full replay; different model ->
+# digest. That's the whole policy.
+# ---------------------------------------------------------------------------
+
+
+def _resolve_review_runtime(agent: Any) -> Dict[str, Any]:
+    """Resolve provider/model/credentials for the review fork.
+
+    Default (auto / unset / same as parent): inherit the parent's live runtime
+    (with codex_app_server -> codex_responses downgrade). ``routed`` is False —
+    the fork uses the main model and the warm cache, exactly as before. When
+    ``auxiliary.background_review.{provider,model}`` names a concrete model
+    different from the parent's, resolve that runtime and set ``routed=True``.
+    """
+    parent_runtime = agent._current_main_runtime()
+    parent_api_mode = parent_runtime.get("api_mode") or None
+    if parent_api_mode == "codex_app_server":
+        parent_api_mode = "codex_responses"
+    parent = {
+        "provider": agent.provider,
+        "model": agent.model,
+        "api_key": parent_runtime.get("api_key") or None,
+        "base_url": parent_runtime.get("base_url") or None,
+        "api_mode": parent_api_mode,
+        "routed": False,
+    }
+    try:
+        from hermes_cli.config import load_config
+        cfg = load_config()
+    except Exception:
+        return parent
+    aux = cfg.get("auxiliary", {}) if isinstance(cfg.get("auxiliary"), dict) else {}
+    task = aux.get("background_review", {}) if isinstance(aux.get("background_review"), dict) else {}
+    task_provider = (str(task.get("provider", "")).strip() or None)
+    task_model = (str(task.get("model", "")).strip() or None)
+    task_base_url = (str(task.get("base_url", "")).strip() or None)
+    task_api_key = (str(task.get("api_key", "")).strip() or None)
+    if not (task_provider and task_provider != "auto" and task_model):
+        return parent
+    if task_provider == (agent.provider or "") and task_model == (agent.model or ""):
+        return parent  # same model/provider as parent -> not routed
+    try:
+        from hermes_cli.runtime_provider import resolve_runtime_provider
+        rp = resolve_runtime_provider(
+            requested=task_provider,
+            target_model=task_model,
+            explicit_api_key=task_api_key,
+            explicit_base_url=task_base_url,
+        )
+        return {
+            "provider": rp.get("provider") or task_provider,
+            "model": task_model,
+            "api_key": rp.get("api_key"),
+            "base_url": rp.get("base_url"),
+            "api_mode": rp.get("api_mode"),
+            "routed": True,
+        }
+    except Exception as e:
+        logger.debug("background-review aux routing failed (%s); using main model", e)
+        return parent
+
+
+def _msg_text(m: Dict) -> str:
+    c = m.get("content")
+    if isinstance(c, str):
+        return c.strip()
+    if isinstance(c, list):
+        return " ".join(b.get("text", "") for b in c if isinstance(b, dict)).strip()
+    return ""
+
+
+def _digest_history(messages_snapshot: List[Dict], tail: int = 24) -> List[Dict]:
+    """Compact replay for the routed (different-model) path only.
+
+    Keeps the recent ``tail`` messages verbatim, collapses older turns into one
+    synthetic user-role digest, preserving role alternation. Used ONLY when
+    routed to a different model (cache cold regardless, so fewer cold-written
+    tokens is a pure win). Never on the main-model path (full replay stays warm).
+    """
+    msgs = list(messages_snapshot or [])
+    if len(msgs) <= tail:
+        return msgs
+    keep = msgs[-tail:]
+    while keep and isinstance(keep[0], dict) and keep[0].get("role") == "tool":
+        tail += 1
+        if len(msgs) <= tail:
+            return msgs
+        keep = msgs[-tail:]
+    old = msgs[:-len(keep)]
+    lines: List[str] = []
+    for m in old:
+        if not isinstance(m, dict):
+            continue
+        role = m.get("role")
+        text = _msg_text(m).replace("\n", " ")
+        if role == "user" and text:
+            lines.append(f"USER: {text[:300]}")
+        elif role == "assistant":
+            tcs = m.get("tool_calls") or []
+            if tcs:
+                names = [(tc.get("function") or {}).get("name", "?") for tc in tcs if isinstance(tc, dict)]
+                lines.append(f"ASSISTANT[tools: {', '.join(names)}]")
+            if text:
+                lines.append(f"ASSISTANT: {text[:200]}")
+    digest = {
+        "role": "user",
+        "content": (
+            "[Earlier conversation digest — older turns summarised to bound the "
+            "review's cold-write cost on the routed aux model. Recent turns "
+            "follow verbatim below.]\n" + "\n".join(lines)
+        ),
+    }
+    return [digest] + keep
+
+
 # Review-prompt strings — used by ``spawn_background_review_thread`` to build
 # the user-message that the forked review agent receives.  AIAgent exposes
 # them as class attributes (``_MEMORY_REVIEW_PROMPT`` etc.) for back-compat;
@@ -488,18 +613,13 @@ def _run_review_in_thread(
             # creds, or credential-pool setups where the resolver can't
             # reconstruct auth from scratch -- producing the spurious
             # "No LLM provider configured" warning at end of turn.
-            _parent_runtime = agent._current_main_runtime()
-            _parent_api_mode = _parent_runtime.get("api_mode") or None
-            # The review fork needs to call agent-loop tools (memory,
-            # skill_manage). Those tools require Hermes' own dispatch,
-            # which the codex_app_server runtime bypasses entirely
-            # (it runs the turn inside codex's subprocess). So when
-            # the parent is on codex_app_server, downgrade the review
-            # fork to codex_responses — same auth/credentials, but
-            # talks to the OpenAI Responses API directly so Hermes
-            # owns the loop and the agent-loop tools dispatch.
-            if _parent_api_mode == "codex_app_server":
-                _parent_api_mode = "codex_responses"
+            # _resolve_review_runtime() returns the parent's live runtime by
+            # default (routed=False; main model, warm cache), or — when the user
+            # set auxiliary.background_review.{provider,model} to a different
+            # model — that model's runtime (routed=True). The codex_app_server
+            # -> codex_responses downgrade is applied inside the resolver.
+            _rt = _resolve_review_runtime(agent)
+            _routed = bool(_rt.get("routed"))
             # skip_memory=True keeps the review fork from
             # touching external memory plugins (honcho, mem0,
             # supermemory, etc.).  Without it, the fork's
@@ -519,14 +639,14 @@ def _run_review_in_thread(
             # in the request body — Anthropic's cache key includes it.
             # (The runtime whitelist below still restricts dispatch.)
             review_agent = AIAgent(
-                model=agent.model,
+                model=_rt.get("model") or agent.model,
                 max_iterations=16,
                 quiet_mode=True,
                 platform=agent.platform,
-                provider=agent.provider,
-                api_mode=_parent_api_mode,
-                base_url=_parent_runtime.get("base_url") or None,
-                api_key=_parent_runtime.get("api_key") or None,
+                provider=_rt.get("provider") or agent.provider,
+                api_mode=_rt.get("api_mode"),
+                base_url=_rt.get("base_url") or None,
+                api_key=_rt.get("api_key") or None,
                 credential_pool=getattr(agent, "_credential_pool", None),
                 parent_session_id=agent.session_id,
                 enabled_toolsets=getattr(agent, "enabled_toolsets", None),
@@ -565,15 +685,20 @@ def _run_review_in_thread(
             # issue #25322 and PR #17276 for the full analysis +
             # measured impact (~26% end-to-end cost reduction on
             # Sonnet 4.5).
-            review_agent._cached_system_prompt = agent._cached_system_prompt
-            # Defensive: pin session_start + session_id to the
-            # parent's so any code path that re-renders parts of
-            # the system prompt (compression, plugin hooks) still
-            # produces byte-identical output. The cached-prompt
-            # assignment above already short-circuits the normal
-            # rebuild path, but these pins guarantee parity even
-            # if a future code path bypasses the cache.
-            review_agent.session_start = agent.session_start
+            # Share the parent's warm cached system prompt ONLY when the review
+            # runs on the SAME model (not routed). When routed to a different
+            # model the parent's cached prompt is for the wrong model/cache key
+            # and would miss anyway, so let the routed fork build its own.
+            if not _routed:
+                review_agent._cached_system_prompt = agent._cached_system_prompt
+                # Defensive: pin session_start + session_id to the
+                # parent's so any code path that re-renders parts of
+                # the system prompt (compression, plugin hooks) still
+                # produces byte-identical output. The cached-prompt
+                # assignment above already short-circuits the normal
+                # rebuild path, but these pins guarantee parity even
+                # if a future code path bypasses the cache.
+                review_agent.session_start = agent.session_start
             review_agent.session_id = agent.session_id
             # The fork shares the parent's live session_id (pinned above for
             # prefix-cache parity). It is single-lifecycle and calls close()
@@ -615,6 +740,13 @@ def _run_review_in_thread(
                 ),
             )
             try:
+                # Routed to a different model -> replay a digest (cache is cold
+                # on that model anyway, so minimise cold-written tokens). Same
+                # model -> replay the full snapshot (warm cache reads).
+                _review_history = (
+                    _digest_history(messages_snapshot) if _routed
+                    else messages_snapshot
+                )
                 review_agent.run_conversation(
                     user_message=(
                         prompt
@@ -622,7 +754,7 @@ def _run_review_in_thread(
                         "management tools. Other tools will be denied "
                         "at runtime — do not attempt them."
                     ),
-                    conversation_history=messages_snapshot,
+                    conversation_history=_review_history,
                 )
             finally:
                 clear_thread_tool_whitelist()

agent/chat_completion_helpers.py

@@ -2561,6 +2561,17 @@ def interruptible_streaming_api_call(agent, api_kwargs: dict, *, on_first_delta=
             _stream_stale_timeout = max(_stream_stale_timeout_base, 240.0)
         else:
             _stream_stale_timeout = _stream_stale_timeout_base
+        # Reasoning-model floor: known reasoning models (Nemotron 3 Ultra,
+        # OpenAI o1/o3, Anthropic Opus 4.x thinking, DeepSeek R1, Qwen QwQ,
+        # xAI Grok reasoning, etc.) routinely exceed the default 180s chat-
+        # model threshold during their thinking phase.  The cloud gateway
+        # upstream kills the socket first, surfacing as BrokenPipeError.
+        # Raises the floor only — never overrides explicit user config
+        # (handled by get_provider_stale_timeout above).
+        from agent.reasoning_timeouts import get_reasoning_stale_timeout_floor
+        _reasoning_floor = get_reasoning_stale_timeout_floor(api_kwargs.get("model"))
+        if _reasoning_floor is not None:
+            _stream_stale_timeout = max(_stream_stale_timeout, _reasoning_floor)
 
     t = threading.Thread(target=_call, daemon=True)
     t.start()

agent/coding_context.py

@@ -83,6 +83,59 @@ _PROJECT_MARKERS = (
 # Agent-instruction files surfaced separately from manifests in the snapshot.
 _CONTEXT_FILES = ("AGENTS.md", "CLAUDE.md", ".cursorrules")
 
+# Source-file extensions that make a git repo a *code* workspace even with no
+# manifest. Without this, `git init` on a notes/writing/research folder (a huge
+# non-coding use case) would flip the whole session into the coding posture just
+# for having a `.git`. A manifest still wins on its own (see `_PROJECT_MARKERS`).
+_CODE_EXTENSIONS = frozenset({
+    ".py", ".pyi", ".ipynb", ".js", ".jsx", ".ts", ".tsx", ".mjs", ".cjs",
+    ".go", ".rs", ".java", ".kt", ".kts", ".scala", ".rb", ".php", ".c", ".h",
+    ".cc", ".cpp", ".hpp", ".cs", ".swift", ".m", ".mm", ".dart", ".ex", ".exs",
+    ".lua", ".sh", ".bash", ".zsh", ".sql", ".vue", ".svelte", ".r", ".jl",
+    ".hs", ".clj", ".erl", ".pl",
+})
+
+# Dirs never worth scanning for the code check (deps/build/vcs/venv noise).
+_CODE_SCAN_SKIP_DIRS = frozenset({
+    ".git", "node_modules", "venv", ".venv", "__pycache__", "dist", "build",
+    "target", ".next", ".turbo", "vendor",
+})
+
+# Bounded sweep: a code workspace reveals itself in the first handful of entries.
+_CODE_SCAN_MAX_ENTRIES = 500
+
+
+def _has_code_files(root: Path) -> bool:
+    """Cheap, bounded check for source files in a repo's top two levels.
+
+    Lets a git repo of loose scripts (no manifest) still read as a code
+    workspace while a bare notes/writing repo does not. Scans the root and its
+    immediate subdirectories only, capped at ``_CODE_SCAN_MAX_ENTRIES`` stats —
+    a handful of readdirs at session start, not a full walk.
+    """
+    seen = 0
+    stack = [(root, True)]
+    while stack:
+        directory, is_root = stack.pop()
+        try:
+            with os.scandir(directory) as entries:
+                for entry in entries:
+                    seen += 1
+                    if seen > _CODE_SCAN_MAX_ENTRIES:
+                        return False
+                    name = entry.name
+                    try:
+                        if entry.is_file():
+                            if os.path.splitext(name)[1].lower() in _CODE_EXTENSIONS:
+                                return True
+                        elif is_root and entry.is_dir() and name not in _CODE_SCAN_SKIP_DIRS and not name.startswith("."):
+                            stack.append((Path(entry.path), False))
+                    except OSError:
+                        continue
+        except OSError:
+            continue
+    return False
+
 # Lockfile → package manager, checked in priority order.
 _PY_LOCKFILES = (("uv.lock", "uv"), ("poetry.lock", "poetry"), ("Pipfile.lock", "pipenv"))
 _JS_LOCKFILES = (
@@ -368,10 +421,16 @@ def _detect_profile_name(mode: str, platform: str, cwd_str: str) -> str:
     if platform and platform.strip().lower() not in INTERACTIVE_CODING_PLATFORMS:
         return GENERAL_PROFILE.name
     cwd = Path(cwd_str)
+    # A recognized project root (manifest / AGENTS.md / .cursorrules) is a code
+    # workspace on its own — cheap stat checks, no scan.
+    if _marker_root(cwd) is not None:
+        return CODING_PROFILE.name
     git_root = _git_root(cwd)
     if git_root is not None and git_root == _home():
         git_root = None  # dotfiles repo at $HOME — not a code workspace
-    if git_root is not None or _marker_root(cwd) is not None:
+    # A bare git repo only counts when it actually holds code, so `git init` on a
+    # notes/writing/research folder stays in the general posture.
+    if git_root is not None and _has_code_files(git_root):
         return CODING_PROFILE.name
     return GENERAL_PROFILE.name
 
@@ -635,25 +694,32 @@ def _read_small(path: Path) -> str:
         return ""
 
 
-def _project_facts(root: Path) -> list[str]:
-    """Detected project facts for the workspace snapshot.
+@dataclass(frozen=True)
+class ProjectFacts:
+    """Structured project facts — the model's verify loop, detected once.
 
-    The point is to hand the model its *verify loop* up front — which manifest,
-    which package manager, and the exact test/lint/build commands — instead of
-    making it rediscover them every session. Cheap: stat calls plus reads of a
-    couple of small files; built once at prompt-build time (cache-safe).
+    The same data that feeds the workspace snapshot, exposed structurally so
+    non-prompt consumers (e.g. the desktop verify UI) read it instead of
+    re-detecting and drifting from the prompt.
     """
-    facts: list[str] = []
 
+    manifests: list[str]
+    package_managers: list[str]
+    verify_commands: list[str]
+    context_files: list[str]
+
+
+def detect_project_facts(root: Path) -> ProjectFacts:
+    """Detect manifests, package manager(s), verify commands, and context files.
+
+    Cheap: stat calls plus reads of a couple of small files. The single source
+    of truth for both the prompt snapshot (:func:`_project_facts`) and the
+    gateway's ``project.facts`` — so the UI never re-sniffs verify commands.
+    """
     manifests = [m for m in _PROJECT_MARKERS if m not in _CONTEXT_FILES and (root / m).is_file()]
-    package_managers = [
-        pm for lock, pm in (*_PY_LOCKFILES, *_JS_LOCKFILES) if (root / lock).is_file()
-    ]
-    if manifests:
-        line = f"- Project: {', '.join(manifests[:6])}"
-        if package_managers:
-            line += f" ({'/'.join(dict.fromkeys(package_managers))})"
-        facts.append(line)
+    package_managers = list(
+        dict.fromkeys(pm for lock, pm in (*_PY_LOCKFILES, *_JS_LOCKFILES) if (root / lock).is_file())
+    )
 
     verify: list[str] = []
     if (root / "scripts" / "run_tests.sh").is_file():
@@ -673,17 +739,61 @@ def _project_facts(root: Path) -> list[str]:
             f"make {name}" for name in _VERIFY_TARGETS
             if re.search(rf"^{re.escape(name)}\s*:", makefile, re.MULTILINE)
         )
-    if verify:
-        deduped = list(dict.fromkeys(verify))[:_MAX_VERIFY_COMMANDS]
-        facts.append(f"- Verify: {'; '.join(deduped)}")
 
-    context_files = [c for c in _CONTEXT_FILES if (root / c).is_file()]
-    if context_files:
-        facts.append(f"- Context files: {', '.join(context_files)}")
+    return ProjectFacts(
+        manifests=manifests,
+        package_managers=package_managers,
+        verify_commands=list(dict.fromkeys(verify))[:_MAX_VERIFY_COMMANDS],
+        context_files=[c for c in _CONTEXT_FILES if (root / c).is_file()],
+    )
+
+
+def _project_facts(root: Path) -> list[str]:
+    """Render :func:`detect_project_facts` as workspace-snapshot lines.
+
+    Hands the model its *verify loop* up front — which manifest, which package
+    manager, and the exact test/lint/build commands — instead of making it
+    rediscover them every session. Built once at prompt-build time; the string
+    output must stay byte-stable to preserve the prompt cache.
+    """
+    f = detect_project_facts(root)
+    facts: list[str] = []
+
+    if f.manifests:
+        line = f"- Project: {', '.join(f.manifests[:6])}"
+        if f.package_managers:
+            line += f" ({'/'.join(f.package_managers)})"
+        facts.append(line)
+    if f.verify_commands:
+        facts.append(f"- Verify: {'; '.join(f.verify_commands)}")
+    if f.context_files:
+        facts.append(f"- Context files: {', '.join(f.context_files)}")
 
     return facts
 
 
+def project_facts_for(cwd: Optional[str | Path] = None) -> Optional[dict[str, Any]]:
+    """Structured project facts for ``cwd`` — ``None`` outside a workspace.
+
+    Same detection the system-prompt snapshot uses (git root, else marker root),
+    exposed for non-prompt consumers (the desktop verify UI) so they never
+    re-derive "are we coding?" or duplicate the verify-command sniffing.
+    """
+    resolved = _resolve_cwd(cwd)
+    root = _git_root(resolved) or _marker_root(resolved)
+    if root is None:
+        return None
+
+    f = detect_project_facts(root)
+    return {
+        "root": str(root),
+        "manifests": f.manifests,
+        "packageManagers": f.package_managers,
+        "verifyCommands": f.verify_commands,
+        "contextFiles": f.context_files,
+    }
+
+
 def build_coding_workspace_block(cwd: Optional[str | Path] = None) -> str:
     """Workspace snapshot for the system prompt (empty outside a workspace).
 

agent/context_compressor.py

@@ -890,7 +890,15 @@ class ContextCompressor(ContextEngine):
         # This is independent of the abort_on_summary_failure config flag:
         # rotating on a broken credential is never the right behavior.
         self._last_summary_auth_failure: bool = False
-        # When a user-configured summary model fails and we recover by
+        # Set when summary generation ultimately fails due to a transient
+        # network/connection error (httpx/httpcore connection drop, premature
+        # stream close, etc.) — distinct from auth failures but treated the
+        # same way by compress(): ABORT and preserve the session unchanged
+        # rather than destroy the middle window for a deterministic
+        # "summary unavailable" marker. Retrying once the network recovers is
+        # strictly better than discarding context for a transient blip
+        # (#29559, #25585). Independent of abort_on_summary_failure.
+        self._last_summary_network_failure: bool = False
         # retrying on the main model, record the failure so gateway /
         # CLI callers can still warn the user even though compression
         # succeeded.  Silent recovery would hide the broken config.
@@ -1687,6 +1695,7 @@ This compaction should PRIORITISE preserving all information related to the focu
             self._summary_model_fallen_back = False
             self._last_summary_error = None
             self._last_summary_auth_failure = False
+            self._last_summary_network_failure = False
             return self._with_summary_prefix(summary)
         except Exception as e:
             # ``call_llm`` raises ``RuntimeError`` for two very different cases:
@@ -1819,6 +1828,15 @@ This compaction should PRIORITISE preserving all information related to the focu
             if len(err_text) > 220:
                 err_text = err_text[:217].rstrip() + "..."
             self._last_summary_error = err_text
+            # A terminal connection/network failure (we reach this branch only
+            # after any main-model fallback has already been tried or is
+            # unavailable). Flag it so compress() ABORTS and preserves the
+            # session unchanged instead of destroying the middle window for a
+            # placeholder marker — retrying once the network recovers is
+            # strictly better than dropping context (#29559, #25585). Mirrors
+            # the auth-failure carve-out; independent of abort_on_summary_failure.
+            if _is_streaming_closed:
+                self._last_summary_network_failure = True
             logger.warning(
                 "Failed to generate context summary: %s. "
                 "Further summary attempts paused for %d seconds.",
@@ -2382,6 +2400,7 @@ This compaction should PRIORITISE preserving all information related to the focu
         self._last_aux_model_failure_model = None
         self._last_compress_aborted = False
         self._last_summary_auth_failure = False
+        self._last_summary_network_failure = False
 
         # Manual /compress (force=True) bypasses the failure cooldown so the
         # user can retry immediately after an auto-compress abort.  Without
@@ -2498,15 +2517,21 @@ This compaction should PRIORITISE preserving all information related to the focu
         #           surface a warning.
         # Default is False (historical behavior).
         #
-        # EXCEPTION — auth failures always abort. A 401/403 from the summary
-        # call means the credential or endpoint is broken (invalid/blocked
-        # key, or a token pointed at the wrong inference host). Rotating into
+        # EXCEPTION — auth AND transient network failures always abort. A
+        # 401/403 from the summary call means the credential or endpoint is
+        # broken (invalid/blocked key, or a token pointed at the wrong
+        # inference host). A connection/stream-close error means the network
+        # blipped at the compaction moment (#29559). In BOTH cases rotating into
         # a child session with a placeholder summary on a broken credential
         # strands the user on a degraded session for zero benefit — every
         # subsequent call fails the same way. So when the failure was an auth
         # error we abort regardless of abort_on_summary_failure, preserving
         # the conversation unchanged until the credential is fixed.
-        if not summary and (self.abort_on_summary_failure or self._last_summary_auth_failure):
+        if not summary and (
+            self.abort_on_summary_failure
+            or self._last_summary_auth_failure
+            or self._last_summary_network_failure
+        ):
             n_skipped = compress_end - compress_start
             self._last_summary_dropped_count = 0  # nothing actually dropped
             self._last_summary_fallback_used = False
@@ -2521,6 +2546,15 @@ This compaction should PRIORITISE preserving all information related to the focu
                         "with /compress or start fresh with /new.",
                         n_skipped,
                     )
+                elif self._last_summary_network_failure:
+                    logger.warning(
+                        "Summary generation failed with a network/connection "
+                        "error — aborting compression. %d message(s) preserved "
+                        "unchanged; the session was NOT rotated. This is "
+                        "transient: retry with /compress once connectivity "
+                        "recovers, or continue the conversation as-is.",
+                        n_skipped,
+                    )
                 else:
                     logger.warning(
                         "Summary generation failed — aborting compression "

agent/conversation_compression.py

@@ -90,6 +90,7 @@ def check_compression_model_feasibility(agent: Any) -> None:
     try:
         from agent.auxiliary_client import (
             _resolve_task_provider_model,
+            _try_configured_fallback_for_unavailable_client,
             get_text_auxiliary_client,
         )
         from agent.model_metadata import (
@@ -97,10 +98,6 @@ def check_compression_model_feasibility(agent: Any) -> None:
             get_model_context_length,
         )
 
-        client, aux_model = get_text_auxiliary_client(
-            "compression",
-            main_runtime=agent._current_main_runtime(),
-        )
         # Best-effort aux provider label for the warning message. The
         # configured provider may be "auto", in which case we fall back
         # to the client's base_url hostname so the user can still tell
@@ -109,6 +106,19 @@ def check_compression_model_feasibility(agent: Any) -> None:
             _aux_cfg_provider, _, _, _, _ = _resolve_task_provider_model("compression")
         except Exception:
             _aux_cfg_provider = ""
+        client, aux_model = get_text_auxiliary_client(
+            "compression",
+            main_runtime=agent._current_main_runtime(),
+        )
+        if client is None or not aux_model:
+            fb_client, fb_model, fb_label = _try_configured_fallback_for_unavailable_client(
+                "compression",
+                _aux_cfg_provider,
+            )
+            if fb_client is not None and fb_model:
+                client, aux_model = fb_client, fb_model
+                if "(" in fb_label and fb_label.endswith(")"):
+                    _aux_cfg_provider = fb_label.rsplit("(", 1)[1][:-1]
         if client is None or not aux_model:
             if _aux_cfg_provider and _aux_cfg_provider != "auto":
                 msg = (
@@ -805,10 +815,11 @@ def try_shrink_image_parts_in_messages(
     Pillow couldn't help (caller should surface the original error).
 
     Strategy: look for ``image_url`` / ``input_image`` parts carrying a
-    ``data:image/...;base64,...`` payload.  For each one whose encoded
-    size exceeds 4 MB (a safe target that slides under Anthropic's 5 MB
-    ceiling with header overhead) or whose longest side exceeds
-    ``max_dimension``, write the base64 to a tempfile, call
+    ``data:image/...;base64,...`` payload, plus Anthropic-native
+    ``{"type": "image", "source": {"type": "base64", ...}}`` blocks.
+    For each one whose encoded size exceeds 4 MB (a safe target that slides
+    under Anthropic's 5 MB ceiling with header overhead) or whose longest side
+    exceeds ``max_dimension``, write the base64 to a tempfile, call
     ``vision_tools._resize_image_for_vision`` to produce a smaller data
     URL, and substitute it in place.
 
@@ -964,6 +975,28 @@ def try_shrink_image_parts_in_messages(
             logger.warning("image-shrink recovery: re-encode failed — %s", exc)
             return None, triggered_by is not None
 
+    def _source_to_data_url(source: Any) -> Optional[str]:
+        if not isinstance(source, dict) or source.get("type") != "base64":
+            return None
+        data = source.get("data")
+        if not isinstance(data, str) or not data:
+            return None
+        media_type = str(source.get("media_type") or "image/jpeg").strip()
+        if not media_type.startswith("image/"):
+            media_type = "image/jpeg"
+        return f"data:{media_type};base64,{data}"
+
+    def _write_data_url_to_source(source: dict, data_url: str) -> None:
+        header, _, data = data_url.partition(",")
+        media_type = "image/jpeg"
+        if header.startswith("data:"):
+            candidate = header[len("data:"):].split(";", 1)[0].strip()
+            if candidate.startswith("image/"):
+                media_type = candidate
+        source["type"] = "base64"
+        source["media_type"] = media_type
+        source["data"] = data
+
     for msg in api_messages:
         if not isinstance(msg, dict):
             continue
@@ -974,6 +1007,16 @@ def try_shrink_image_parts_in_messages(
             if not isinstance(part, dict):
                 continue
             ptype = part.get("type")
+            if ptype == "image":
+                source = part.get("source")
+                url = _source_to_data_url(source)
+                resized, unshrinkable = _shrink_data_url(url or "")
+                if resized and isinstance(source, dict):
+                    _write_data_url_to_source(source, resized)
+                    changed_count += 1
+                elif unshrinkable:
+                    unshrinkable_oversized += 1
+                continue
             if ptype not in {"image_url", "input_image"}:
                 continue
             image_value = part.get("image_url")

agent/conversation_loop.py

@@ -35,6 +35,7 @@ from agent.turn_context import build_turn_context
 from agent.turn_retry_state import TurnRetryState
 from agent.memory_manager import build_memory_context_block
 from agent.message_sanitization import (
+    close_interrupted_tool_sequence,
     _repair_tool_call_arguments,
     _sanitize_messages_non_ascii,
     _sanitize_messages_surrogates,
@@ -55,7 +56,7 @@ from agent.model_metadata import (
 )
 from agent.process_bootstrap import _install_safe_stdio
 from agent.prompt_caching import apply_anthropic_cache_control
-from agent.retry_utils import jittered_backoff
+from agent.retry_utils import adaptive_rate_limit_backoff, jittered_backoff
 from agent.trajectory import has_incomplete_scratchpad
 from agent.usage_pricing import estimate_usage_cost, normalize_usage
 from hermes_constants import PARTIAL_STREAM_STUB_ID
@@ -501,6 +502,7 @@ def run_conversation(
     stream_callback: Optional[callable] = None,
     persist_user_message: Optional[str] = None,
     persist_user_timestamp: Optional[float] = None,
+    moa_config: Optional[dict[str, Any]] = None,
 ) -> Dict[str, Any]:
     """
     Run a complete conversation with tool calling until completion.
@@ -523,6 +525,19 @@ def run_conversation(
     Returns:
         Dict: Complete conversation result with final response and message history
     """
+    if moa_config is None:
+        try:
+            from hermes_cli.moa_config import decode_moa_turn
+
+            _decoded_message, _decoded_moa_config = decode_moa_turn(user_message)
+            if _decoded_moa_config is not None:
+                user_message = _decoded_message
+                moa_config = _decoded_moa_config
+                if persist_user_message is None:
+                    persist_user_message = _decoded_message
+        except Exception:
+            pass
+
     # ── Per-turn setup (the prologue) ──
     # All once-per-turn setup — stdio guarding, retry-counter resets, user
     # message sanitization, todo/nudge hydration, system-prompt restore-or-
@@ -801,6 +816,29 @@ def run_conversation(
         if effective_system:
             api_messages = [{"role": "system", "content": effective_system}] + api_messages
 
+        if moa_config:
+            try:
+                from agent.moa_loop import aggregate_moa_context
+
+                _moa_context = aggregate_moa_context(
+                    user_prompt=original_user_message if isinstance(original_user_message, str) else str(original_user_message),
+                    api_messages=api_messages,
+                    reference_models=moa_config.get("reference_models") or [],
+                    aggregator=moa_config.get("aggregator") or {},
+                    temperature=float(moa_config.get("reference_temperature", 0.6) or 0.6),
+                    aggregator_temperature=float(moa_config.get("aggregator_temperature", 0.4) or 0.4),
+                    max_tokens=int(moa_config.get("max_tokens", 4096) or 4096),
+                )
+                if _moa_context:
+                    for _msg in reversed(api_messages):
+                        if _msg.get("role") == "user":
+                            _base = _msg.get("content", "")
+                            if isinstance(_base, str):
+                                _msg["content"] = _base + "\n\n" + _moa_context
+                            break
+            except Exception as _moa_exc:
+                logger.warning("MoA context aggregation failed: %s", _moa_exc)
+
         # Inject ephemeral prefill messages right after the system prompt
         # but before conversation history. Same API-call-time-only pattern.
         if agent.prefill_messages:
@@ -1122,7 +1160,7 @@ def run_conversation(
                 # stream.  Mirror the ACP exclusion used for Responses
                 # API upgrade (lines ~1083-1085).
                 elif (
-                    agent.provider == "copilot-acp"
+                    agent.provider in {"copilot-acp", "moa"}
                     or str(agent.base_url or "").lower().startswith("acp://copilot")
                     or str(agent.base_url or "").lower().startswith("acp+tcp://")
                 ):
@@ -1396,10 +1434,12 @@ def run_conversation(
                     while time.time() < sleep_end:
                         if agent._interrupt_requested:
                             agent._vprint(f"{agent.log_prefix}⚡ Interrupt detected during retry wait, aborting.", force=True)
+                            _interrupt_text = f"Operation interrupted during retry ({_failure_hint}, attempt {retry_count}/{max_retries})."
+                            close_interrupted_tool_sequence(messages, _interrupt_text)
                             agent._persist_session(messages, conversation_history)
                             agent.clear_interrupt()
                             return {
-                                "final_response": f"Operation interrupted during retry ({_failure_hint}, attempt {retry_count}/{max_retries}).",
+                                "final_response": _interrupt_text,
                                 "messages": messages,
                                 "api_calls": api_call_count,
                                 "completed": False,
@@ -1971,9 +2011,21 @@ def run_conversation(
                     agent.thinking_callback("")
                 api_elapsed = time.time() - api_start_time
                 agent._vprint(f"{agent.log_prefix}⚡ Interrupted during API call.", force=True)
-                agent._persist_session(messages, conversation_history)
                 interrupted = True
-                final_response = f"{INTERRUPT_WAITING_FOR_MODEL_PREFIX}{api_elapsed:.1f}s elapsed)."
+                # Preserve any assistant text already streamed to the user
+                # before the stop landed. Dropping it leaves history with no
+                # record of the half-finished reply on screen, so the next turn
+                # the model "forgets" what it just said — exactly what users hit
+                # when they stop to redirect mid-response.
+                _partial = agent._strip_think_blocks(
+                    getattr(agent, "_current_streamed_assistant_text", "") or ""
+                ).strip()
+                if _partial:
+                    messages.append({"role": "assistant", "content": _partial})
+                    final_response = _partial
+                else:
+                    final_response = f"{INTERRUPT_WAITING_FOR_MODEL_PREFIX}{api_elapsed:.1f}s elapsed)."
+                agent._persist_session(messages, conversation_history)
                 break
 
             except Exception as api_error:
@@ -2663,10 +2715,12 @@ def run_conversation(
                 # Check for interrupt before deciding to retry
                 if agent._interrupt_requested:
                     agent._vprint(f"{agent.log_prefix}⚡ Interrupt detected during error handling, aborting retries.", force=True)
+                    _interrupt_text = f"Operation interrupted: handling API error ({error_type}: {agent._clean_error_message(str(api_error))})."
+                    close_interrupted_tool_sequence(messages, _interrupt_text)
                     agent._persist_session(messages, conversation_history)
                     agent.clear_interrupt()
                     return {
-                        "final_response": f"Operation interrupted: handling API error ({error_type}: {agent._clean_error_message(str(api_error))}).",
+                        "final_response": _interrupt_text,
                         "messages": messages,
                         "api_calls": api_call_count,
                         "completed": False,
@@ -3485,6 +3539,65 @@ def run_conversation(
                             force=True,
                         )
 
+                    # Detect thinking-timeout pattern: a known reasoning model
+                    # hit a transport-layer error before the first content
+                    # token arrived.  Distinct from _is_stream_drop above
+                    # (which fires for large file-write stream drops) and
+                    # from any classifier reason that's not a transport
+                    # timeout.  Reuses the reasoning-model allowlist from
+                    # agent/reasoning_timeouts.py (Fixes #52217) so the
+                    # trigger is consistent with what the per-model
+                    # stale-timeout floor covers.  After the classifier
+                    # override at agent/error_classifier.py:720-738 (this
+                    # PR), transport disconnects on reasoning models route
+                    # to FailoverReason.timeout rather than
+                    # context_overflow, so this branch actually fires.
+                    # Detection and message text live in
+                    # agent.thinking_timeout_guidance so they're
+                    # unit-testable without driving the full retry loop.
+                    # (Part 2 of Fixes #52310.)
+                    from agent.thinking_timeout_guidance import (
+                        is_thinking_timeout,
+                    )
+                    _is_thinking_timeout = is_thinking_timeout(
+                        classified,
+                        _model,
+                        error_msg,
+                    )
+                    if _is_thinking_timeout:
+                        agent._vprint(
+                            f"{agent.log_prefix}   💡 The model's thinking "
+                            f"phase exceeded the upstream proxy's idle "
+                            f"timeout before the first content token "
+                            f"arrived. This is a known issue with "
+                            f"reasoning models behind cloud gateways "
+                            f"(NVIDIA NIM, OpenAI, Anthropic, DeepSeek).",
+                            force=True,
+                        )
+                        agent._vprint(
+                            f"{agent.log_prefix}      Workarounds in priority order:",
+                            force=True,
+                        )
+                        agent._vprint(
+                            f"{agent.log_prefix}      1. Set "
+                            f"`providers.{_provider}.models.{_model}.stale_timeout_seconds: 900` "
+                            f"in `~/.hermes/config.yaml` to extend the per-call "
+                            f"timeout. (Hermes's built-in floor is 600s for "
+                            f"known reasoning models — if you still see this "
+                            f"after raising, the upstream cap is even shorter.)",
+                            force=True,
+                        )
+                        agent._vprint(
+                            f"{agent.log_prefix}      2. Lower `reasoning_budget` or set "
+                            f"`reasoning_effort: medium` on this model if the provider supports it.",
+                            force=True,
+                        )
+                        agent._vprint(
+                            f"{agent.log_prefix}      3. Use a smaller / faster reasoning "
+                            f"model if the task doesn't require deep thinking.",
+                            force=True,
+                        )
+
                     logger.error(
                         "%sAPI call failed after %s retries. %s | provider=%s model=%s msgs=%s tokens=~%s",
                         agent.log_prefix, max_retries, _final_summary,
@@ -3501,7 +3614,22 @@ def run_conversation(
                             _final_response += f"\n\n{_billing_guidance}"
                     else:
                         _final_response = f"API call failed after {max_retries} retries: {_final_summary}"
-                    if _is_stream_drop:
+                    if _is_thinking_timeout:
+                        # Thinking-timeout guidance overrides the generic
+                        # stream-drop guidance — the latter is wrong for
+                        # this case (it suggests splitting large file
+                        # writes, which isn't what happened).  See the
+                        # reasoning-model override at
+                        # agent/error_classifier.py:720-738 and the
+                        # detection block above for context.
+                        from agent.thinking_timeout_guidance import (
+                            build_thinking_timeout_guidance,
+                        )
+                        _final_response += build_thinking_timeout_guidance(
+                            provider=_provider,
+                            model=_model,
+                        )
+                    elif _is_stream_drop:
                         _final_response += (
                             "\n\nThe provider's stream connection keeps "
                             "dropping — this often happens when generating "
@@ -3537,16 +3665,38 @@ def run_conversation(
                             except (TypeError, ValueError):
                                 pass
                 wait_time = _retry_after if _retry_after else jittered_backoff(retry_count, base_delay=2.0, max_delay=60.0)
+                _backoff_policy = None
+                if is_rate_limited and not _retry_after:
+                    wait_time, _backoff_policy = adaptive_rate_limit_backoff(
+                        retry_count,
+                        base_url=str(_base),
+                        model=_model,
+                        error=api_error,
+                        default_wait=wait_time,
+                    )
                 if is_rate_limited:
-                    agent._buffer_status(f"⏱️ Rate limited. Waiting {wait_time:.1f}s (attempt {retry_count + 1}/{max_retries})...")
+                    _policy_note = ""
+                    if _backoff_policy == "zai_coding_overload_long":
+                        _policy_note = " (Z.AI Coding overload adaptive long backoff)"
+                    elif _backoff_policy == "zai_coding_overload_short":
+                        _policy_note = " (Z.AI Coding overload short retry)"
+                    _rate_limit_status = f"⏱️ Rate limited. Waiting {wait_time:.1f}s (attempt {retry_count + 1}/{max_retries}){_policy_note}..."
+                    # Normal retries are buffered to avoid noisy transient chatter. Long
+                    # Z.AI Coding waits are different: they can last minutes, so surface
+                    # progress immediately instead of making the TUI look frozen.
+                    if _backoff_policy == "zai_coding_overload_long":
+                        agent._emit_status(_rate_limit_status)
+                    else:
+                        agent._buffer_status(_rate_limit_status)
                 else:
                     agent._buffer_status(f"⏳ Retrying in {wait_time:.1f}s (attempt {retry_count}/{max_retries})...")
                 logger.warning(
-                    "Retrying API call in %ss (attempt %s/%s) %s error=%s",
+                    "Retrying API call in %ss (attempt %s/%s) %s policy=%s error=%s",
                     wait_time,
                     retry_count,
                     max_retries,
                     agent._client_log_context(),
+                    _backoff_policy or "default",
                     api_error,
                 )
                 # Sleep in small increments so we can respond to interrupts quickly
@@ -3556,10 +3706,12 @@ def run_conversation(
                 while time.time() < sleep_end:
                     if agent._interrupt_requested:
                         agent._vprint(f"{agent.log_prefix}⚡ Interrupt detected during retry wait, aborting.", force=True)
+                        _interrupt_text = f"Operation interrupted: retrying API call after error (retry {retry_count}/{max_retries})."
+                        close_interrupted_tool_sequence(messages, _interrupt_text)
                         agent._persist_session(messages, conversation_history)
                         agent.clear_interrupt()
                         return {
-                            "final_response": f"Operation interrupted: retrying API call after error (retry {retry_count}/{max_retries}).",
+                            "final_response": _interrupt_text,
                             "messages": messages,
                             "api_calls": api_call_count,
                             "completed": False,
@@ -4050,6 +4202,19 @@ def run_conversation(
 
                 messages.append(assistant_msg)
                 agent._emit_interim_assistant_message(assistant_msg)
+                try:
+                    # Persist the assistant tool-call turn before any tool
+                    # side effects run. If a destructive tool restarts or
+                    # terminates Hermes mid-turn, resume logic still sees the
+                    # exact tool-call block that already executed.
+                    agent._flush_messages_to_session_db(messages, conversation_history)
+                except Exception as exc:
+                    logger.warning(
+                        "Incremental tool-call persistence failed before execution "
+                        "(session=%s): %s",
+                        agent.session_id or "none",
+                        exc,
+                    )
 
                 # Close any open streaming display (response box, reasoning
                 # box) before tool execution begins.  Intermediate turns may
@@ -4479,9 +4644,10 @@ def run_conversation(
                 final_msg = agent._build_assistant_message(assistant_message, finish_reason)
 
                 # Pop thinking-only prefill and empty-response retry
-                # scaffolding before appending the final response.  These
-                # internal turns are only for the next API retry and should
-                # not become durable transcript context.
+                # scaffolding before appending either a final response or a
+                # verification-stop follow-up. These internal turns are only
+                # for the next API retry and should not become durable
+                # transcript context.
                 while (
                     messages
                     and isinstance(messages[-1], dict)
@@ -4493,6 +4659,44 @@ def run_conversation(
                 ):
                     messages.pop()
 
+                try:
+                    from agent.verification_stop import (
+                        build_verify_on_stop_nudge,
+                        verify_on_stop_enabled,
+                    )
+
+                    if verify_on_stop_enabled():
+                        _verify_nudge = build_verify_on_stop_nudge(
+                            session_id=getattr(agent, "session_id", None),
+                            changed_paths=getattr(agent, "_turn_file_mutation_paths", set()),
+                            attempts=getattr(agent, "_verification_stop_nudges", 0),
+                        )
+                    else:
+                        _verify_nudge = None
+                except Exception:
+                    logger.debug("verification stop-loop check failed", exc_info=True)
+                    _verify_nudge = None
+
+                if _verify_nudge:
+                    agent._verification_stop_nudges = (
+                        getattr(agent, "_verification_stop_nudges", 0) + 1
+                    )
+                    final_msg["finish_reason"] = "verification_required"
+                    messages.append(final_msg)
+                    # Keep the attempted final answer in model history so the
+                    # synthetic user nudge preserves role alternation, but do
+                    # not surface it to the user as an interim answer. The
+                    # whole point of this guard is to prevent premature
+                    # "done" claims before checks run.
+                    messages.append({
+                        "role": "user",
+                        "content": _verify_nudge,
+                        "_verification_stop_synthetic": True,
+                    })
+                    agent._session_messages = messages
+                    agent._emit_status("↻ Verification required before finishing")
+                    continue
+
                 messages.append(final_msg)
                 
                 _turn_exit_reason = f"text_response(finish_reason={finish_reason})"

agent/credential_pool.py

@@ -11,6 +11,7 @@ import uuid
 import re
 from dataclasses import dataclass, fields, replace
 from datetime import datetime, timezone
+from pathlib import Path
 from typing import Any, Dict, List, Optional, Set, Tuple
 
 from hermes_constants import OPENROUTER_BASE_URL
@@ -447,6 +448,63 @@ def get_pool_strategy(provider: str) -> str:
 DEFAULT_MAX_CONCURRENT_PER_CREDENTIAL = 1
 
 
+def _write_through_provider_state_to_global_root(
+    provider_id: str, state: Dict[str, Any]
+) -> None:
+    """Persist a rotated OAuth ``state`` into the global-root auth.json.
+
+    Best-effort write-through for the multi-profile rotation hazard
+    (#48415 / #43589): nous, openai-codex, and xai-oauth rotate the
+    refresh_token on refresh, so when a profile pool refresh rotates a grant
+    it resolved from the root fallback, the rotated chain must land back in
+    root. Otherwise root keeps a now-revoked refresh token and every other
+    profile reading the stale root grant dies with ``refresh_token_reused`` /
+    ``invalid_grant`` once its access token expires.
+
+    Only updates ``providers.<provider_id>`` in the root store; never touches
+    the profile store (the caller already saved that). Swallows all errors — a
+    failed write-through degrades to the pre-existing behavior (root stale), it
+    must never break the profile's own successful save. Mirrors
+    ``hermes_cli.auth._write_through_xai_oauth_to_global_root`` (which covers
+    the non-pool xAI refresh path) for the credential-pool refresh path.
+    """
+    try:
+        global_path = auth_mod._global_auth_file_path()
+    except Exception:
+        return
+    if global_path is None:
+        # Classic mode (profile == root); the profile save already hit root.
+        return
+    # Seat belt: under pytest, refuse to write the real user's
+    # ~/.hermes/auth.json even when HERMES_HOME points at a profile path
+    # (mirrors the read-side guard in _load_global_auth_store). Uses the
+    # unmodified HOME env, not Path.home() which fixtures may monkeypatch.
+    if os.environ.get("PYTEST_CURRENT_TEST"):
+        real_home_env = os.environ.get("HOME", "")
+        if real_home_env:
+            real_root = Path(real_home_env) / ".hermes" / "auth.json"
+            try:
+                if global_path.resolve(strict=False) == real_root.resolve(strict=False):
+                    return
+            except Exception:
+                return
+    try:
+        if global_path.exists():
+            global_store = _load_auth_store(global_path)
+        else:
+            global_store = {}
+        if not isinstance(global_store, dict):
+            return
+        _store_provider_state(global_store, provider_id, dict(state), set_active=False)
+        auth_mod._save_auth_store(global_store, global_path)
+    except Exception as exc:  # pragma: no cover - best effort
+        logger.debug(
+            "%s pool refresh: write-through to global root failed: %s",
+            provider_id,
+            exc,
+        )
+
+
 class CredentialPool:
     def __init__(self, provider: str, entries: List[PooledCredential]):
         self.provider = provider
@@ -800,6 +858,28 @@ class CredentialPool:
         try:
             with _auth_store_lock():
                 auth_store = _load_auth_store()
+                # Decide BEFORE writing whether this profile is reading the
+                # grant from the global root (no own providers.<id> block) vs.
+                # genuinely shadowing it. A pool refresh rotates single-use
+                # OAuth refresh tokens, so a profile that resolved the grant
+                # from root MUST write the rotated chain back to root too —
+                # otherwise root keeps a revoked refresh token and every other
+                # profile reading the stale root grant dies with
+                # refresh_token_reused / invalid_grant once its access token
+                # expires. This mirrors the xAI write-through in
+                # hermes_cli.auth._save_xai_oauth_tokens (#43589); the pool
+                # refresh path is the Codex/xAI analog reported in #48415.
+                _wt_provider_id = {
+                    "nous": "nous",
+                    "openai-codex": "openai-codex",
+                    "xai-oauth": "xai-oauth",
+                }.get(self.provider)
+                write_through_to_root = bool(_wt_provider_id) and not (
+                    isinstance(auth_store.get("providers"), dict)
+                    and isinstance(
+                        auth_store["providers"].get(_wt_provider_id), dict
+                    )
+                )
                 if self.provider == "nous":
                     state = _load_provider_state(auth_store, "nous")
                     if state is None:
@@ -855,6 +935,10 @@ class CredentialPool:
                     return
 
                 _save_auth_store(auth_store)
+                if write_through_to_root and _wt_provider_id:
+                    _write_through_provider_state_to_global_root(
+                        _wt_provider_id, state
+                    )
         except Exception as exc:
             logger.debug("Failed to sync %s pool entry back to auth store: %s", self.provider, exc)
 

agent/curator.py

@@ -377,8 +377,10 @@ CURATOR_REVIEW_PROMPT = (
     "bodies + `references/`, `templates/`, and `scripts/` subfiles for "
     "session-specific detail — not one-session-one-skill micro-entries.\n\n"
     "Hard rules — do not violate:\n"
-    "1. DO NOT touch bundled or hub-installed skills. The candidate list "
-    "below is already filtered to agent-created skills only.\n"
+    "1. DO NOT touch bundled, hub-installed, or external-dir skills "
+    "(`skills.external_dirs`). The candidate list below is already filtered "
+    "to local curator-managed skills only; external skills are externally "
+    "owned and read-only to this background curator.\n"
     "2. DO NOT delete any skill. Archiving (moving the skill's directory "
     "into ~/.hermes/skills/.archive/) is the maximum destructive action. "
     "Archives are recoverable; deletion is not.\n"
@@ -469,8 +471,9 @@ CURATOR_REVIEW_PROMPT = (
     "skill, or `absorbed_into=\"\"` when you're truly pruning with no "
     "forwarding target. This drives cron-job skill-reference migration — "
     "guessing from your YAML summary after the fact is fragile.\n"
-    "  - terminal                       — mv a sibling into the archive "
-    "OR move its content into a support subfile\n\n"
+    "  - terminal                       — move LOCAL candidate content into "
+    "a support subfile when package integrity requires it; never mv, cp, rm, "
+    "patch, or rewrite bundled, hub-installed, or external-dir skills\n\n"
     "'keep' is a legitimate decision ONLY when the skill is already a "
     "class-level umbrella and none of the proposed merges would improve "
     "discoverability. 'This is narrow but distinct from its siblings' "
@@ -1843,6 +1846,14 @@ def _run_llm_review(prompt: str) -> Dict[str, Any]:
         # Disable recursive nudges — the curator must never spawn its own review.
         review_agent._memory_nudge_interval = 0
         review_agent._skill_nudge_interval = 0
+        # Tag this fork as autonomous background curation so skill_manage's
+        # background-review write guard fires. Without this the fork inherits
+        # the default "assistant_tool" origin, is_background_review() is False,
+        # and the external/bundled/hub-installed skill_manage guards never
+        # trigger during the curation pass they exist to protect against.
+        # turn_context.py binds this onto the write-origin ContextVar at turn
+        # start (see agent/turn_context.py).
+        review_agent._memory_write_origin = "background_review"
 
         # Redirect the forked agent's stdout/stderr to /dev/null while it
         # runs so its tool-call chatter doesn't pollute the foreground

agent/display.py

@@ -6,6 +6,7 @@ Used by AIAgent._execute_tool_calls for CLI feedback.
 
 import logging
 import os
+import re
 import sys
 import threading
 import time
@@ -15,6 +16,7 @@ from pathlib import Path
 from typing import Any
 
 from utils import safe_json_loads
+from agent.redact import redact_sensitive_text
 from agent.tool_result_classification import file_mutation_result_landed
 
 # ANSI escape codes for coloring tool failure indicators
@@ -177,6 +179,223 @@ def _truncate_preview(text: str, max_len: int | None) -> str:
     return text
 
 
+_SHELL_SILENT_HEADS = {"cd", "pushd", "popd", "export", "set", "unset", "source", ".", "true", "false", ":"}
+_SHELL_PIPE_TAIL_HEADS = {"head", "tail", "wc", "sort", "uniq"}
+
+
+def _shell_basename(head: str) -> str:
+    return head.rsplit("/", 1)[-1] if head else ""
+
+
+def _split_shell_words(segment: str) -> list[str]:
+    words: list[str] = []
+    buf: list[str] = []
+    quote: str | None = None
+
+    for i, ch in enumerate(segment):
+        if quote:
+            buf.append(ch)
+            if ch == quote and (i == 0 or segment[i - 1] != "\\"):
+                quote = None
+            continue
+
+        if ch in {"'", '"'}:
+            quote = ch
+            buf.append(ch)
+            continue
+
+        if ch.isspace():
+            if buf:
+                words.append("".join(buf))
+                buf = []
+            continue
+
+        buf.append(ch)
+
+    if buf:
+        words.append("".join(buf))
+
+    return words
+
+
+def _strip_shell_pipe_tail(segment: str) -> str:
+    words = _split_shell_words(segment)
+    out: list[str] = []
+
+    for i, word in enumerate(words):
+        if word == "|" and _shell_basename(words[i + 1] if i + 1 < len(words) else "") in _SHELL_PIPE_TAIL_HEADS:
+            break
+        out.append(word)
+
+    return " ".join(out).strip()
+
+
+def _split_shell_compound(command: str) -> list[str]:
+    segments: list[str] = []
+    buf: list[str] = []
+    quote: str | None = None
+    i = 0
+
+    while i < len(command):
+        ch = command[i]
+
+        if quote:
+            buf.append(ch)
+            if ch == quote and (i == 0 or command[i - 1] != "\\"):
+                quote = None
+            i += 1
+            continue
+
+        if ch in {"'", '"'}:
+            quote = ch
+            buf.append(ch)
+            i += 1
+            continue
+
+        op_len = 2 if command.startswith("&&", i) or command.startswith("||", i) else 1 if ch in {";", "\n"} else 0
+        if op_len:
+            segment = _strip_shell_pipe_tail("".join(buf).strip())
+            if segment:
+                segments.append(segment)
+            buf = []
+            i += op_len
+            continue
+
+        buf.append(ch)
+        i += 1
+
+    segment = _strip_shell_pipe_tail("".join(buf).strip())
+    if segment:
+        segments.append(segment)
+
+    return segments
+
+
+def _shell_head_word(segment: str) -> str:
+    words = _split_shell_words(segment)
+    index = 0
+    while index < len(words) and re.match(r"^[A-Za-z_]\w*=", words[index]):
+        index += 1
+    return _shell_basename(words[index] if index < len(words) else "")
+
+
+def _clean_shell_segment(segment: str) -> str:
+    words = _split_shell_words(segment)
+    out: list[str] = []
+    i = 0
+    while i < len(words):
+        word = words[i]
+        if re.match(r"^\d*(?:>>?|<)$", word):
+            i += 2
+            continue
+        if re.match(r"^\d*(?:>&|<&)\d+$", word) or re.match(r"^\d*>&\d+$", word):
+            i += 1
+            continue
+        out.append(word)
+        i += 1
+    return " ".join(out).strip()
+
+
+def _is_shell_boundary_echo(segment: str) -> bool:
+    words = _split_shell_words(segment)
+    if _shell_basename(words[0] if words else "") != "echo":
+        return False
+    rest = " ".join(words[1:])
+    return bool(re.search(r"-{2,}|_exit=|(?:^|\s|=)\$[?{]|PIPESTATUS", rest))
+
+
+def summarize_shell_command(command: str) -> str:
+    """Compact shell wrapper/plumbing for display while preserving raw command elsewhere."""
+    original = _oneline(command)
+    if not original:
+        return ""
+
+    segments = _split_shell_compound(original)
+    if len(segments) <= 1:
+        return _clean_shell_segment(segments[0] if segments else original) or original
+
+    core: list[str] = []
+    for segment in segments:
+        cleaned = _clean_shell_segment(segment)
+        head = _shell_head_word(cleaned)
+        if cleaned and head not in _SHELL_SILENT_HEADS and not _is_shell_boundary_echo(cleaned):
+            core.append(cleaned)
+
+    if not core:
+        return original
+    if len(core) == 1:
+        return core[0]
+
+    count = len(core) - 1
+    return f"{core[0]} + {count} {'command' if count == 1 else 'commands'}"
+
+
+def _read_file_line_label(args: dict) -> str:
+    offset = args.get("offset")
+    limit = args.get("limit")
+    if not isinstance(offset, int) or offset <= 0:
+        return ""
+    if not isinstance(limit, int) or limit <= 1:
+        return f"L{offset}"
+    return f"L{offset}-{offset + limit - 1}"
+
+
+def redact_browser_typed_text_for_display(value: Any, typed_text: Any) -> Any:
+    """Apply secret redaction to browser_type text in display-facing payloads.
+
+    Backends sometimes echo the attempted input in error strings or fallback
+    metadata.  When the raw typed value contains a recognizable secret (API
+    key, token, JWT, etc.) the redacted form differs from the raw value, so we
+    replace every occurrence of the raw value with its redacted form before a
+    browser_type result reaches logs, callbacks, the model, or chat history.
+
+    Normal typed text (search queries, addresses, form fields) matches no
+    secret pattern, so it passes through unchanged and stays readable.
+
+    Redaction is forced here regardless of the global ``security.redact_secrets``
+    preference: a typed credential leaking into chat history is a security
+    boundary, not mere log hygiene.
+    """
+    if typed_text is None:
+        return value
+    needle = str(typed_text)
+    if needle == "":
+        return value
+    redacted = redact_sensitive_text(needle, force=True)
+    if redacted == needle:
+        # Nothing secret-looking in the typed text; leave payload untouched.
+        return value
+    if isinstance(value, str):
+        return value.replace(needle, redacted)
+    if isinstance(value, dict):
+        return {
+            key: redact_browser_typed_text_for_display(item, typed_text)
+            for key, item in value.items()
+        }
+    if isinstance(value, list):
+        return [redact_browser_typed_text_for_display(item, typed_text) for item in value]
+    if isinstance(value, tuple):
+        return tuple(redact_browser_typed_text_for_display(item, typed_text) for item in value)
+    return value
+
+
+def redact_tool_args_for_display(tool_name: str, args: dict | None) -> dict | None:
+    """Return a copy of tool args safe for logs/progress UI.
+
+    For ``browser_type`` the ``text`` argument is run through the same
+    secret-pattern redactor used for logs.  Recognizable credentials (API
+    keys, tokens) are masked before the value reaches tool progress
+    notifications; normal typed text is left intact for debuggability.
+    """
+    if not isinstance(args, dict):
+        return args
+    if tool_name == "browser_type" and isinstance(args.get("text"), str):
+        safe_args = dict(args)
+        safe_args["text"] = redact_sensitive_text(args["text"], force=True)
+        return safe_args
+    return args
+
+
 def _delegate_task_goal_parts(tasks: Any, *, per_goal_len: int) -> tuple[int, list[str]]:
     if not isinstance(tasks, list):
         return 0, []
@@ -200,13 +419,14 @@ def build_tool_preview(tool_name: str, args: dict, max_len: int | None = None) -
         max_len = _tool_preview_max_len
     if not args:
         return None
+    args = redact_tool_args_for_display(tool_name, args) or args
     primary_args = {
         "terminal": "command", "web_search": "query", "web_extract": "urls",
         "read_file": "path", "write_file": "path", "patch": "path",
         "search_files": "pattern", "browser_navigate": "url",
         "browser_click": "ref", "browser_type": "text",
         "image_generate": "prompt", "text_to_speech": "text",
-        "vision_analyze": "question", "mixture_of_agents": "user_prompt",
+        "vision_analyze": "question",
         "skill_view": "name", "skills_list": "category",
         "cronjob": "action",
         "execute_code": "code", "delegate_task": "goal",
@@ -253,6 +473,23 @@ def build_tool_preview(tool_name: str, args: dict, max_len: int | None = None) -
         else:
             return f"planning {len(todos_arg)} task(s)"
 
+    if tool_name in {"terminal", "execute_code"}:
+        key = "code" if tool_name == "execute_code" else "command"
+        command = args.get(key)
+        if command is None:
+            return None
+        preview = summarize_shell_command(str(command))
+        return _truncate_preview(preview, max_len) if preview else None
+
+    if tool_name == "read_file":
+        path = args.get("path") or args.get("file") or args.get("filepath")
+        if path is None:
+            return None
+        label = Path(str(path).replace("\\", "/")).name or str(path)
+        line_label = _read_file_line_label(args)
+        preview = f"{label} {line_label}".strip()
+        return _truncate_preview(preview, max_len) if preview else None
+
     if tool_name == "session_search":
         query = _oneline(args.get("query", ""))
         return f"recall: \"{query[:25]}{'...' if len(query) > 25 else ''}\""
@@ -906,6 +1143,7 @@ def get_cute_tool_message(
     When *result* is provided the line is checked for failure indicators.
     Failed tool calls get a red prefix and an informational suffix.
     """
+    args = redact_tool_args_for_display(tool_name, args) or args
     dur = f"{duration:.1f}s"
     is_failure, failure_suffix = _detect_tool_failure(tool_name, result)
     skin_prefix = get_skin_tool_prefix()
@@ -943,7 +1181,7 @@ def get_cute_tool_message(
             return _wrap(f"┊ 📄 fetch     {_trunc(domain, 35)}{extra}  {dur}")
         return _wrap(f"┊ 📄 fetch     pages  {dur}")
     if tool_name == "terminal":
-        return _wrap(f"┊ 💻 $         {_trunc(args.get('command', ''), 42)}  {dur}")
+        return _wrap(f"┊ 💻 $         {_trunc(build_tool_preview(tool_name, args) or args.get('command', ''), 42)}  {dur}")
     if tool_name == "process":
         action = args.get("action", "?")
         sid = args.get("session_id", "")[:12]
@@ -951,7 +1189,7 @@ def get_cute_tool_message(
                   "wait": f"wait {sid}", "kill": f"kill {sid}", "write": f"write {sid}", "submit": f"submit {sid}"}
         return _wrap(f"┊ ⚙️  proc      {labels.get(action, f'{action} {sid}')}  {dur}")
     if tool_name == "read_file":
-        return _wrap(f"┊ 📖 read      {_path(args.get('path', ''))}  {dur}")
+        return _wrap(f"┊ 📖 read      {_trunc(build_tool_preview(tool_name, args) or args.get('path', ''), 42)}  {dur}")
     if tool_name == "write_file":
         return _wrap(f"┊ ✍️  write     {_path(args.get('path', ''))}  {dur}")
     if tool_name == "patch":
@@ -1037,8 +1275,6 @@ def get_cute_tool_message(
         return _wrap(f"┊ 🔊 speak     {_trunc(args.get('text', ''), 30)}  {dur}")
     if tool_name == "vision_analyze":
         return _wrap(f"┊ 👁️  vision    {_trunc(args.get('question', ''), 30)}  {dur}")
-    if tool_name == "mixture_of_agents":
-        return _wrap(f"┊ 🧠 reason    {_trunc(args.get('user_prompt', ''), 30)}  {dur}")
     if tool_name == "send_message":
         return _wrap(f"┊ 📨 send      {args.get('target', '?')}: \"{_trunc(args.get('message', ''), 25)}\"  {dur}")
     if tool_name == "cronjob":

agent/error_classifier.py

@@ -717,6 +717,26 @@ def classify_api_error(
 
     is_disconnect = any(p in error_msg for p in _SERVER_DISCONNECT_PATTERNS)
     if is_disconnect and not status_code:
+        # Reasoning-model override: a transport disconnect on a reasoning
+        # model is much more likely the upstream proxy idle-killing a
+        # long thinking stream than a true context overflow — even on
+        # large sessions.  The default disconnect+large-session routing
+        # below would otherwise send the user into the compression
+        # branch (should_compress=True) and silently delete
+        # conversation history on a phantom context-length error.
+        # Reasoning models have multi-minute thinking phases that
+        # routinely exceed the cloud gateway's idle window (NVIDIA
+        # NIM ~120s — first-party repro at NVIDIA/NemoClaw#4846;
+        # OpenAI worker / Anthropic stream-idle similar).  The
+        # per-reasoning-model stale-timeout floor in
+        # agent/reasoning_timeouts.py raises the stale-detector
+        # threshold to tolerate long thinking, so a true
+        # transport-layer failure here is recoverable via the retry
+        # path — not via context compression.  Reclassify as timeout.
+        # (Part 1 of Fixes #52310.)
+        from agent.reasoning_timeouts import get_reasoning_stale_timeout_floor
+        if get_reasoning_stale_timeout_floor(model) is not None:
+            return _result(FailoverReason.timeout, retryable=True)
         # Absolute token/message-count thresholds are only a proxy for smaller
         # context windows.  Large-context sessions can have hundreds of
         # messages while still being far below their actual token budget.

agent/learn_prompt.py

@@ -0,0 +1,136 @@
+#!/usr/bin/env python3
+"""``/learn`` — build the standards-guided prompt that turns whatever the user
+described into a reusable skill.
+
+``/learn`` is open-ended. The user can point it at anything they can describe:
+a directory of code, an API doc URL, a workflow they just walked the agent
+through in this conversation, or pasted notes. This module builds ONE prompt
+that instructs the live agent to:
+
+  1. Gather the sources the user named, using the tools it already has
+     (``read_file`` / ``search_files`` for dirs, ``web_extract`` for URLs, the
+     current conversation for "what I just did", the user's text for pasted
+     material).
+  2. Author a single ``SKILL.md`` via ``skill_manage`` that follows the Hermes
+     skill-authoring standards (description <=60 chars, the modern section
+     order, Hermes-tool framing, no invented commands).
+
+There is no separate distillation engine and no model-tool footprint: the
+agent does the work with its existing toolset, so this works identically on
+local, Docker, and remote terminal backends. Every surface (CLI ``/learn``,
+gateway ``/learn``, the dashboard "Learn a skill" panel) calls
+:func:`build_learn_prompt` and feeds the result to the agent as a normal turn.
+"""
+
+from __future__ import annotations
+
+# The house-style rules, distilled from AGENTS.md "Skill authoring standards
+# (HARDLINE)" and the hermes-agent-dev new-skill salvage reference. Embedded in
+# the prompt so the agent authors skills the way a maintainer would by hand.
+_AUTHORING_STANDARDS = """\
+Follow the Hermes skill-authoring standards exactly. These are the same
+HARDLINE rules a maintainer enforces in review:
+
+Frontmatter:
+- name: lowercase-hyphenated, <=64 chars, no spaces.
+- description: ONE sentence, **<=60 characters**, ends with a period. State the
+  capability, not the implementation. No marketing words (powerful,
+  comprehensive, seamless, advanced, robust). Do NOT repeat the skill name. If
+  the description contains a colon, wrap the whole value in double quotes.
+  This is the most-violated rule and it is NOT cosmetic: the system-prompt
+  skill index truncates the description to 60 chars and loads it every
+  session, so anything past char 60 is silently cut and never routes. After
+  you write the description, COUNT the characters; if it is over 60, cut it
+  down before saving — do not ship a sentence and hope.
+    Good (<=60): `Search arXiv papers by keyword, author, or ID.`
+    Bad (123):   `A comprehensive skill that lets the agent search arXiv for
+                  academic papers using keywords, authors, and categories.`
+- version: 0.1.0
+- author: always the literal value `Hermes`. NEVER fill it from the host
+  environment — the OS/login username (e.g. the `user=` line in your
+  environment hints), git config, or any identity you can probe must not be
+  written. Skills get shared and published, so an environment-derived name is
+  a privacy leak the user never opted into; the skill names itself as Hermes.
+- platforms: declare `[macos]`, `[linux]`, and/or `[windows]` IF the skill
+  uses OS-bound primitives (osascript/apt/systemctl => the matching OS; /proc,
+  os.setsid, signal.SIGKILL => linux; fcntl/termios => POSIX). Prefer fixing it
+  cross-platform first (tempfile.gettempdir(), pathlib.Path, psutil); gate only
+  when the dependency is genuinely platform-bound. Omit the field for portable
+  skills.
+- metadata.hermes.tags: a few Capitalized, Relevant, Tags.
+
+Body section order (omit a section only if it genuinely has no content):
+1. "# <Human Title>" then a 2-3 sentence intro: what it does, what it does NOT
+   do, and the key dependency stance (e.g. "stdlib only").
+2. "## When to Use" — bullet list of concrete trigger phrases.
+3. "## Prerequisites" — exact env vars, install steps, credentials.
+4. "## How to Run" — the canonical invocation, framed through Hermes tools.
+5. "## Quick Reference" — a flat command/endpoint list, no narration.
+6. "## Procedure" — numbered steps with copy-paste-exact commands.
+7. "## Pitfalls" — known limits, rate limits, things that look broken but aren't.
+8. "## Verification" — a single command/check that proves the skill worked.
+
+Hermes-tool framing (this is what makes it a skill, not shell docs):
+- Frame running scripts as "invoke through the `terminal` tool".
+- Reference Hermes tools by name in backticks: `terminal`, `read_file`,
+  `write_file`, `search_files`, `patch`, `web_extract`, `web_search`,
+  `vision_analyze`, `browser_navigate`, `delegate_task`, `image_generate`,
+  `text_to_speech`, `cronjob`, `memory`, `skill_view`, `execute_code`.
+- Do NOT name shell utilities the agent already has wrapped: say `read_file`
+  not cat/head/tail, `search_files` not grep/rg/find/ls, `patch` not sed/awk,
+  `web_extract` not curl-to-scrape, `write_file` not echo>file or heredocs.
+- Third-party CLIs (ffmpeg, gh, an SDK) are fine inside a script file, but the
+  prose still frames them as "invoke through the `terminal` tool". If the
+  skill needs an MCP server, name it and document its setup in Prerequisites.
+
+Quality bar:
+- Prefer exact commands, endpoint URLs, function signatures, and config keys
+  that appear VERBATIM in the source. NEVER invent flags, paths, or APIs — if
+  you didn't see it in the source, don't write it.
+- Keep it tight and scannable: ~100 lines for a simple skill, ~200 for a
+  complex one. Don't re-paste the source docs.
+- Don't write a router/index/hub skill that only points at other skills.
+- Larger scripts/parsers belong in a `scripts/` file (add via
+  `skill_manage` write_file), referenced from SKILL.md by relative path — not
+  inlined for the agent to re-type every run. References go in `references/`,
+  templates in `templates/`."""
+
+
+def build_learn_prompt(user_request: str) -> str:
+    """Build the agent prompt for an open-ended ``/learn`` request.
+
+    Args:
+        user_request: the free-text the user gave after ``/learn`` — a
+            description of the workflow, paths, URLs, or "what I just did".
+
+    Returns:
+        A complete instruction the agent runs as a normal turn. The agent
+        gathers the described sources with its existing tools and authors the
+        skill via ``skill_manage``.
+    """
+    req = (user_request or "").strip()
+    if not req:
+        req = (
+            "the workflow we just went through in this conversation — review "
+            "the steps taken and distill them into a reusable skill"
+        )
+
+    return (
+        "[/learn] The user wants you to learn a reusable skill from the "
+        "source(s) they described below, and save it.\n\n"
+        f"WHAT TO LEARN FROM:\n{req}\n\n"
+        "Do this:\n"
+        "1. Gather the material. Resolve whatever the user named using the "
+        "tools you already have — `read_file`/`search_files` for local files "
+        "or directories, `web_extract` for URLs, the current conversation "
+        "history if they referred to something you just did, and the text "
+        "they pasted as-is. If the request is ambiguous about scope, make a "
+        "reasonable choice and note it; do not stall.\n"
+        "2. Author ONE SKILL.md and save it with the `skill_manage` tool "
+        "(action=\"create\"). Pick a sensible category. If the procedure needs "
+        "a non-trivial script, add it under the skill's `scripts/` with "
+        "`skill_manage` write_file and reference it by relative path.\n\n"
+        f"{_AUTHORING_STANDARDS}\n\n"
+        "When done, tell the user the skill name, its category, and a "
+        "one-line summary of what it captured."
+    )

agent/memory_manager.py

@@ -46,6 +46,39 @@ logger = logging.getLogger(__name__)
 _SYNC_DRAIN_TIMEOUT_S = 5.0
 
 
+def normalize_tool_schema(schema: Any) -> Optional[Dict[str, Any]]:
+    """Return a function-tool dict with a resolvable top-level ``name``.
+
+    Context engines and memory providers expose tool schemas via
+    ``get_tool_schemas()``. The expected shape is a bare function schema
+    (``{"name": ..., "description": ..., "parameters": ...}``) which callers
+    wrap as ``{"type": "function", "function": schema}``.
+
+    Some providers instead return an entry that is *already* in OpenAI tool
+    form (``{"type": "function", "function": {"name": ...}}``). Wrapping that
+    a second time produces ``{"type": "function", "function": {"type":
+    "function", "function": {...}}}`` whose ``function`` has no top-level
+    ``name``. Strict providers (e.g. DeepSeek) reject the *entire* request
+    with ``tools[N].function: missing field name`` (HTTP 400), so one bad
+    schema disables the whole toolset and breaks every turn (#47707).
+
+    This helper normalizes both shapes to the bare function schema and
+    returns ``None`` for anything without a resolvable name, so callers can
+    skip-with-warning rather than appending a nameless tool.
+    """
+    if not isinstance(schema, dict):
+        return None
+    # Unwrap an already-wrapped OpenAI tool entry.
+    if schema.get("type") == "function" and isinstance(schema.get("function"), dict):
+        schema = schema["function"]
+        if not isinstance(schema, dict):
+            return None
+    name = schema.get("name", "")
+    if not name or not isinstance(name, str):
+        return None
+    return schema
+
+
 def memory_provider_tools_enabled(enabled_toolsets: Optional[List[str]]) -> bool:
     """Return whether external memory-provider tools should be exposed."""
     if enabled_toolsets is None:
@@ -92,11 +125,17 @@ def inject_memory_provider_tools(agent: Any) -> int:
         agent.valid_tool_names = valid_tool_names
 
     added = 0
-    for schema in get_schemas():
-        if not isinstance(schema, dict):
+    for raw_schema in get_schemas():
+        schema = normalize_tool_schema(raw_schema)
+        if schema is None:
+            logger.warning(
+                "Memory provider returned a tool schema with no resolvable "
+                "name; skipping to avoid poisoning the request (%r)",
+                raw_schema,
+            )
             continue
-        tool_name = schema.get("name", "")
-        if not tool_name or tool_name in existing_tool_names:
+        tool_name = schema["name"]
+        if tool_name in existing_tool_names:
             continue
         tools.append({"type": "function", "function": schema})
         valid_tool_names.add(tool_name)
@@ -370,8 +409,11 @@ class MemoryManager:
         _core_tool_names = set(_HERMES_CORE_TOOLS)
 
         # Index tool names → provider for routing
-        for schema in provider.get_tool_schemas():
-            tool_name = schema.get("name", "")
+        for raw_schema in provider.get_tool_schemas():
+            schema = normalize_tool_schema(raw_schema)
+            if schema is None:
+                continue
+            tool_name = schema["name"]
             if tool_name in _core_tool_names:
                 logger.warning(
                     "Memory provider '%s' tool '%s' shadows a reserved core "
@@ -658,11 +700,19 @@ class MemoryManager:
         seen = set()
         for provider in self._providers:
             try:
-                for schema in provider.get_tool_schemas():
-                    name = schema.get("name", "")
+                for raw_schema in provider.get_tool_schemas():
+                    schema = normalize_tool_schema(raw_schema)
+                    if schema is None:
+                        logger.warning(
+                            "Memory provider '%s' returned a tool schema with "
+                            "no resolvable name; skipping (%r)",
+                            provider.name, raw_schema,
+                        )
+                        continue
+                    name = schema["name"]
                     if name in _core_tool_names:
                         continue
-                    if name and name not in seen:
+                    if name not in seen:
                         schemas.append(schema)
                         seen.add(name)
             except Exception as e:

agent/message_sanitization.py

@@ -279,6 +279,38 @@ def _repair_tool_call_arguments(raw_args: str, tool_name: str = "?") -> str:
     return "{}"
 
 
+def close_interrupted_tool_sequence(messages: list, final_response: Any = None) -> bool:
+    """Append a synthetic assistant turn when an interrupted tail is a tool result.
+
+    A turn cut short by ``/stop`` can leave the transcript ending on a raw
+    ``tool`` message (a tool finished, or its execution was cancelled, but the
+    model never streamed a closing assistant turn). Persisting that tail means
+    the next user message lands as ``… tool → user`` — a role-alternation
+    violation that strict providers (Gemini, Claude) react to by hallucinating
+    a continuation of the user's message and ignoring prior context, which
+    reads to the user as "lost context" (#48879).
+
+    ``finalize_turn`` closes this on the happy interrupt path, but the
+    retry/backoff/error interrupt aborts in ``conversation_loop`` ``return``
+    early and never reach it — this shared helper closes the sequence on all of
+    them. ``final_response`` is usually empty on an interrupt, so an explicit
+    placeholder is used rather than an empty-content assistant turn.
+
+    Mutates ``messages`` in place. Returns True if a closing turn was appended.
+    """
+    if not messages:
+        return False
+    last = messages[-1]
+    if not isinstance(last, dict) or last.get("role") != "tool":
+        return False
+    text = final_response if isinstance(final_response, str) else ""
+    messages.append({
+        "role": "assistant",
+        "content": text.strip() or "Operation interrupted.",
+    })
+    return True
+
+
 def _strip_non_ascii(text: str) -> str:
     """Remove non-ASCII characters, replacing with closest ASCII equivalent or removing.
 
@@ -431,6 +463,7 @@ def _sanitize_structure_non_ascii(payload: Any) -> bool:
 
 __all__ = [
     "_SURROGATE_RE",
+    "close_interrupted_tool_sequence",
     "_sanitize_surrogates",
     "_sanitize_structure_surrogates",
     "_sanitize_messages_surrogates",

agent/moa_loop.py

@@ -0,0 +1,306 @@
+"""Mixture-of-Agents runtime helpers for /moa turns.
+
+The slash command is deliberately not a model tool. It marks one user turn as
+MoA-enabled; the normal Hermes agent loop still owns tool calling and turn
+termination, while this module gathers reference-model context before each model
+iteration.
+"""
+
+from __future__ import annotations
+
+import logging
+from concurrent.futures import ThreadPoolExecutor
+from typing import Any
+
+from agent.auxiliary_client import call_llm
+from agent.transports import get_transport
+
+logger = logging.getLogger(__name__)
+
+# Upper bound on concurrent reference-model calls. References are independent
+# advisory calls (no tools, no inter-dependence), so we fan them out the same
+# way delegate_task runs a batch: all in flight at once, results collected when
+# every reference finishes. Presets rarely list more than a handful of
+# references; this cap just protects against a pathologically large preset
+# opening dozens of sockets at once.
+_MAX_REFERENCE_WORKERS = 8
+
+
+def _slot_label(slot: dict[str, str]) -> str:
+    return f"{slot.get('provider', '').strip()}:{slot.get('model', '').strip()}"
+
+
+def _run_reference(
+    slot: dict[str, str],
+    ref_messages: list[dict[str, Any]],
+    *,
+    temperature: float,
+    max_tokens: int,
+) -> tuple[str, str]:
+    """Call one reference model and return ``(label, text)``.
+
+    Never raises: a failed reference becomes a labelled note so the aggregator
+    can still act with partial context. Designed to run inside a thread pool —
+    ``call_llm`` is synchronous/blocking, so threads (not asyncio) are the right
+    concurrency primitive, mirroring ``delegate_task``'s batch fan-out.
+    """
+    label = _slot_label(slot)
+    try:
+        response = call_llm(
+            task="moa_reference",
+            provider=slot["provider"],
+            model=slot["model"],
+            messages=ref_messages,
+            temperature=temperature,
+            max_tokens=max_tokens,
+        )
+        return label, _extract_text(response) or "(empty response)"
+    except Exception as exc:
+        logger.warning("MoA reference model %s failed: %s", label, exc)
+        return label, f"[failed: {exc}]"
+
+
+def _run_references_parallel(
+    reference_models: list[dict[str, str]],
+    ref_messages: list[dict[str, Any]],
+    *,
+    temperature: float,
+    max_tokens: int,
+) -> list[tuple[str, str]]:
+    """Fan out all reference models in parallel, returning outputs in order.
+
+    Like ``delegate_task``'s batch mode, every reference is dispatched at once
+    and we block until all of them finish before handing the joined results to
+    the aggregator. Output order matches ``reference_models`` so the
+    ``Reference {idx}`` labelling stays stable. MoA presets that reference
+    another MoA preset are skipped here (recursion guard) with a labelled note.
+    """
+    if not reference_models:
+        return []
+
+    results: list[tuple[str, str] | None] = [None] * len(reference_models)
+    futures = {}
+    workers = min(_MAX_REFERENCE_WORKERS, len(reference_models))
+    with ThreadPoolExecutor(max_workers=workers) as executor:
+        for idx, slot in enumerate(reference_models):
+            if slot.get("provider") == "moa":
+                results[idx] = (
+                    _slot_label(slot),
+                    "[skipped: MoA presets cannot recursively reference MoA]",
+                )
+                continue
+            futures[
+                executor.submit(
+                    _run_reference,
+                    slot,
+                    ref_messages,
+                    temperature=temperature,
+                    max_tokens=max_tokens,
+                )
+            ] = idx
+        # Collect every reference before returning — the aggregator needs the
+        # complete set, so there is no early-exit / first-completed path here.
+        for future, idx in futures.items():
+            results[idx] = future.result()
+
+    return [r for r in results if r is not None]
+
+
+def _reference_messages(messages: list[dict[str, Any]]) -> list[dict[str, Any]]:
+    """Build an advisory-safe view of the conversation for reference models.
+
+    Reference calls are advisory: they never call tools and never emit the
+    ``tool_calls`` the main model did. Replaying the full transcript verbatim
+    (a) re-bills the ~8K-token Hermes system prompt per reference per
+    iteration and (b) risks 400s from strict providers (Mistral, Fireworks)
+    that reject orphan ``tool`` messages or ``tool_calls`` the reference never
+    produced. We keep only the user/assistant *text* turns, dropping the
+    system prompt, any ``tool``-role messages, and any ``tool_calls`` payloads.
+    """
+    trimmed: list[dict[str, Any]] = []
+    for msg in messages:
+        role = msg.get("role")
+        if role not in ("user", "assistant"):
+            # Drop system prompt and tool-result messages.
+            continue
+        content = msg.get("content")
+        if not isinstance(content, str):
+            # Skip non-text (multimodal/tool-call-only) assistant turns.
+            if not content:
+                continue
+        text = content if isinstance(content, str) else ""
+        if role == "assistant" and not text.strip():
+            # Assistant turn that was purely tool calls — nothing advisory.
+            continue
+        trimmed.append({"role": role, "content": text})
+    if not trimmed:
+        # Degenerate case (e.g. first turn was stripped): fall back to a
+        # minimal user turn so the reference still has something to answer.
+        for msg in reversed(messages):
+            if msg.get("role") == "user" and isinstance(msg.get("content"), str):
+                return [{"role": "user", "content": msg["content"]}]
+    return trimmed
+
+
+
+def _extract_text(response: Any) -> str:
+    try:
+        transport = get_transport("chat_completions")
+        if transport is None:
+            raise RuntimeError("chat_completions transport unavailable")
+        normalized = transport.normalize_response(response)
+        text = (normalized.content or "").strip()
+        if text:
+            return text
+    except Exception:
+        pass
+    try:
+        content = response.choices[0].message.content
+        return (content or "").strip()
+    except Exception:
+        return ""
+
+
+def aggregate_moa_context(
+    *,
+    user_prompt: str,
+    api_messages: list[dict[str, Any]],
+    reference_models: list[dict[str, str]],
+    aggregator: dict[str, str],
+    temperature: float = 0.6,
+    aggregator_temperature: float = 0.4,
+    max_tokens: int = 4096,
+) -> str:
+    """Run configured reference models and synthesize their advice.
+
+    Failures are returned as model-specific notes instead of aborting the normal
+    agent loop; the main model can still act with partial context.
+    """
+    reference_outputs: list[tuple[str, str]] = []
+    ref_messages = _reference_messages(api_messages)
+    reference_outputs = _run_references_parallel(
+        reference_models,
+        ref_messages,
+        temperature=temperature,
+        max_tokens=max_tokens,
+    )
+
+    joined = "\n\n".join(
+        f"Reference {idx} — {label}:\n{text}"
+        for idx, (label, text) in enumerate(reference_outputs, start=1)
+    )
+    synth_prompt = (
+        "You are the aggregator in a Mixture of Agents process. Synthesize the "
+        "reference responses into concise, actionable guidance for the main "
+        "Hermes agent. Focus on next steps, tool-use strategy, risks, and any "
+        "disagreements. Do not answer the user directly unless that is all that "
+        "is needed; produce context the main agent should use in its normal loop.\n\n"
+        f"Original user prompt:\n{user_prompt}\n\n"
+        f"Reference responses:\n{joined}"
+    )
+
+    agg_label = _slot_label(aggregator)
+    try:
+        response = call_llm(
+            task="moa_aggregator",
+            provider=aggregator["provider"],
+            model=aggregator["model"],
+            messages=[{"role": "user", "content": synth_prompt}],
+            temperature=aggregator_temperature,
+            max_tokens=max_tokens,
+        )
+        synthesis = _extract_text(response)
+    except Exception as exc:
+        logger.warning("MoA aggregator model %s failed: %s", agg_label, exc)
+        synthesis = ""
+
+    if not synthesis:
+        synthesis = joined
+
+    return (
+        "[Mixture of Agents context — use this as private guidance for the "
+        "normal Hermes agent loop. You may call tools, continue reasoning, or "
+        "finish normally.]\n"
+        f"Aggregator: {agg_label}\n"
+        f"References: {', '.join(_slot_label(slot) for slot in reference_models)}\n\n"
+        f"{synthesis.strip()}"
+    )
+
+
+class MoAChatCompletions:
+    """OpenAI-chat-compatible facade where the aggregator is the acting model."""
+
+    def __init__(self, preset_name: str):
+        self.preset_name = preset_name or "default"
+
+    def create(self, **api_kwargs: Any) -> Any:
+        from hermes_cli.config import load_config
+        from hermes_cli.moa_config import resolve_moa_preset
+
+        preset = resolve_moa_preset(load_config().get("moa") or {}, self.preset_name)
+        messages = list(api_kwargs.get("messages") or [])
+        reference_models = preset.get("reference_models") or []
+        aggregator = preset.get("aggregator") or {}
+        max_tokens = int(preset.get("max_tokens", api_kwargs.get("max_tokens") or 4096) or 4096)
+        temperature = float(preset.get("reference_temperature", 0.6) or 0.6)
+        aggregator_temperature = float(preset.get("aggregator_temperature", api_kwargs.get("temperature") or 0.4) or 0.4)
+
+        # When the preset is disabled, skip the reference fan-out and let the
+        # configured aggregator act alone — it is the preset's acting model, so
+        # a disabled MoA preset is simply "use the aggregator directly."
+        if not preset.get("enabled", True):
+            reference_models = []
+
+        reference_outputs: list[tuple[str, str]] = []
+        ref_messages = _reference_messages(messages)
+        reference_outputs = _run_references_parallel(
+            reference_models,
+            ref_messages,
+            temperature=temperature,
+            max_tokens=max_tokens,
+        )
+
+        agg_messages = [dict(m) for m in messages]
+        if reference_outputs:
+            joined = "\n\n".join(
+                f"Reference {idx} — {label}:\n{text}"
+                for idx, (label, text) in enumerate(reference_outputs, start=1)
+            )
+            guidance = (
+                "[Mixture of Agents reference context]\n"
+                f"Preset: {self.preset_name}\n"
+                f"Aggregator/acting model: {_slot_label(aggregator)}\n"
+                f"References: {', '.join(label for label, _ in reference_outputs)}\n\n"
+                "Use the reference responses below as private context. You are the aggregator and acting model: "
+                "answer the user directly or call tools as needed.\n\n"
+                f"{joined}"
+            )
+            for msg in reversed(agg_messages):
+                if msg.get("role") == "user" and isinstance(msg.get("content"), str):
+                    msg["content"] = msg["content"] + "\n\n" + guidance
+                    break
+            else:
+                agg_messages.append({"role": "user", "content": guidance})
+
+        if aggregator.get("provider") == "moa":
+            raise RuntimeError("MoA aggregator cannot be another MoA preset")
+        agg_kwargs = dict(api_kwargs)
+        agg_kwargs["messages"] = agg_messages
+        agg_kwargs["model"] = aggregator.get("model")
+        agg_kwargs["temperature"] = aggregator_temperature
+        return call_llm(
+            task="moa_aggregator",
+            provider=aggregator.get("provider"),
+            model=aggregator.get("model"),
+            messages=agg_messages,
+            temperature=aggregator_temperature,
+            max_tokens=agg_kwargs.get("max_tokens"),
+            tools=agg_kwargs.get("tools"),
+            extra_body=agg_kwargs.get("extra_body"),
+        )
+
+
+class MoAClient:
+    def __init__(self, preset_name: str):
+        self.chat = type("_MoAChat", (), {})()
+        self.chat.completions = MoAChatCompletions(preset_name)

agent/oneshot.py

@@ -0,0 +1,158 @@
+"""Shared one-off LLM requests for non-conversational helpers.
+
+A "one-shot" is a single, stateless model call that runs *outside* any
+conversation: it never touches a session's history, never breaks prompt
+caching, and returns plain text. UI surfaces use it for small generative
+chores — a commit message from a diff, a rename suggestion, a summary —
+where spinning up an agent turn would be wrong (it would pollute the thread)
+and hand-rolling an LLM call at every call site would be worse.
+
+Two ways to call it:
+
+  * ``run_oneshot(instructions=..., user_input=...)`` — caller supplies the
+    full prompt.
+  * ``run_oneshot(template="commit_message", variables={...})`` — caller
+    names a registered template and passes its variables; the template owns
+    the prompt engineering so it stays consistent across CLI/TUI/desktop.
+
+Model selection rides the same auxiliary plumbing as title generation
+(:func:`agent.auxiliary_client.call_llm`): pass ``main_runtime`` to inherit
+the live session's provider/model, otherwise the configured ``task`` (default
+``title_generation``) resolves a cheap/fast backend.
+"""
+
+import logging
+from typing import Any, Callable, Dict, Optional, Tuple
+
+from agent.auxiliary_client import call_llm, extract_content_or_reasoning
+
+logger = logging.getLogger(__name__)
+
+# A template turns a variables dict into a (instructions, user_input) pair.
+# Templates are plain callables (not str.format) so diff/code payloads with
+# literal "{" / "}" pass through untouched.
+PromptTemplate = Callable[[Dict[str, Any]], Tuple[str, str]]
+
+
+def _truncate(text: str, limit: int) -> str:
+    text = text or ""
+    if len(text) <= limit:
+        return text
+    return text[:limit].rstrip() + "\n…(truncated)"
+
+
+_COMMIT_INSTRUCTIONS = (
+    "You write git commit messages. Given a diff of staged changes, write ONE "
+    "concise Conventional Commits message describing what the change does and why.\n"
+    "Rules:\n"
+    "- Subject line: type(scope): summary — imperative mood, lower-case, no "
+    "trailing period, ≤ 72 characters. Types: feat, fix, refactor, perf, docs, "
+    "test, build, chore, style, ci.\n"
+    "- Omit the scope if it isn't obvious.\n"
+    "- Add a short body (wrapped at ~72 cols) ONLY when the change needs "
+    "explanation; skip it for small/obvious changes.\n"
+    "- Describe the actual change, never restate the diff line-by-line.\n"
+    "- Return ONLY the commit message text — no quotes, no markdown fences, no "
+    "preamble."
+)
+
+
+def _commit_message_template(variables: Dict[str, Any]) -> Tuple[str, str]:
+    diff = _truncate(str(variables.get("diff") or ""), 12000)
+    recent = _truncate(str(variables.get("recent_commits") or ""), 1500)
+
+    parts = []
+    if recent.strip():
+        parts.append(
+            "Recent commit subjects from this repo (match their style/conventions):\n"
+            f"{recent}"
+        )
+    parts.append("Diff to describe:\n" + (diff or "(no textual diff available)"))
+
+    # "Regenerate" must yield something new even on models that decode greedily
+    # / pin temperature server-side. A trailing nonce isn't enough, so we hand
+    # back the previous message and require a genuinely different one.
+    avoid = _truncate(str(variables.get("avoid") or "").strip(), 1000)
+    if avoid:
+        parts.append(
+            "You already proposed the message below and the user wants a "
+            "different one. Write a NEW message with different wording (and, if "
+            "reasonable, a different emphasis or scope framing) — do not repeat "
+            f"it:\n{avoid}"
+        )
+
+    return _COMMIT_INSTRUCTIONS, "\n\n".join(parts)
+
+
+# Registry of named templates. Add an entry here to give a new surface a
+# consistent, reusable prompt without teaching every caller the prompt text.
+PROMPT_TEMPLATES: Dict[str, PromptTemplate] = {
+    "commit_message": _commit_message_template,
+}
+
+
+def render_template(name: str, variables: Optional[Dict[str, Any]] = None) -> Tuple[str, str]:
+    """Resolve a registered template into (instructions, user_input).
+
+    Raises KeyError if the template name is unknown so callers fail loudly
+    instead of silently sending an empty prompt.
+    """
+    template = PROMPT_TEMPLATES.get(name)
+    if template is None:
+        raise KeyError(f"unknown one-shot template: {name}")
+    return template(variables or {})
+
+
+def run_oneshot(
+    *,
+    instructions: str = "",
+    user_input: str = "",
+    template: Optional[str] = None,
+    variables: Optional[Dict[str, Any]] = None,
+    task: str = "title_generation",
+    max_tokens: int = 1024,
+    temperature: Optional[float] = 0.3,
+    timeout: float = 60.0,
+    main_runtime: Optional[Dict[str, Any]] = None,
+) -> str:
+    """Run a single stateless LLM request and return its text.
+
+    Provide either a registered ``template`` (+ ``variables``) or an explicit
+    ``instructions`` / ``user_input`` pair. Returns the model's text answer,
+    stripped of surrounding whitespace and any wrapping code fence.
+
+    Raises RuntimeError when no LLM provider is configured (surfaced from
+    :func:`call_llm`) and KeyError for an unknown template name.
+    """
+    if template:
+        instructions, user_input = render_template(template, variables)
+
+    if not (instructions or "").strip() and not (user_input or "").strip():
+        raise ValueError("run_oneshot requires a template or instructions/user_input")
+
+    messages = []
+    if (instructions or "").strip():
+        messages.append({"role": "system", "content": instructions})
+    messages.append({"role": "user", "content": user_input or ""})
+
+    response = call_llm(
+        task=task,
+        messages=messages,
+        max_tokens=max_tokens,
+        temperature=temperature,
+        timeout=timeout,
+        main_runtime=main_runtime,
+    )
+
+    text = (extract_content_or_reasoning(response) or "").strip()
+    return _strip_code_fence(text)
+
+
+def _strip_code_fence(text: str) -> str:
+    """Drop a single wrapping ``` fence the model may have added."""
+    if not text.startswith("```"):
+        return text
+    lines = text.splitlines()
+    if len(lines) >= 2 and lines[0].startswith("```") and lines[-1].strip() == "```":
+        return "\n".join(lines[1:-1]).strip()
+    return text

agent/pet/__init__.py

@@ -0,0 +1,51 @@
+"""Petdex pet engine — shared core for the CLI, TUI, and desktop surfaces.
+
+Petdex (https://github.com/crafter-station/petdex) is a public gallery of
+animated sprite "pets" for coding agents.  Each pet is a ``pet.json`` plus a
+``spritesheet.{webp,png}`` of 192×208 px cells. Current Codex/petdex sheets use
+an 8-column × 9-row atlas; older Hermes/petdex sheets used an 8-row atlas.
+Hermes infers the row taxonomy from the sheet and maps agent activity onto
+idle/run/review/failed/wave/jump.
+
+This package is the **single source of truth** for the feature so the base
+CLI (Python) and TUI (Ink, via ``tui_gateway``) never duplicate the hard
+parts:
+
+- :mod:`agent.pet.constants` — frame geometry + the :class:`PetState` enum.
+- :mod:`agent.pet.state`     — map agent activity → a :class:`PetState`.
+- :mod:`agent.pet.manifest`  — fetch the public petdex manifest.
+- :mod:`agent.pet.store`     — install / list / resolve pets on disk
+                               (profile-aware via ``get_hermes_home()``).
+- :mod:`agent.pet.render`    — decode a spritesheet and encode frames for a
+                               terminal (kitty / iTerm2 / sixel graphics
+                               protocols, with a Unicode half-block
+                               fallback).
+
+Rendering in the Electron desktop is necessarily TypeScript (canvas), but it
+reuses the same on-disk store and the same state semantics.
+
+The whole feature is a *display* concern: it adds no model tool, mutates no
+system prompt or toolset, and therefore has zero effect on prompt caching.
+"""
+
+from agent.pet.constants import (
+    DEFAULT_SCALE,
+    FRAME_H,
+    FRAME_W,
+    FRAMES_PER_STATE,
+    LOOP_MS,
+    STATE_ROWS,
+    PetState,
+)
+from agent.pet.state import derive_pet_state
+
+__all__ = [
+    "DEFAULT_SCALE",
+    "FRAME_H",
+    "FRAME_W",
+    "FRAMES_PER_STATE",
+    "LOOP_MS",
+    "STATE_ROWS",
+    "PetState",
+    "derive_pet_state",
+]

agent/pet/constants.py

@@ -0,0 +1,167 @@
+"""Pet sprite geometry + animation-state taxonomy.
+
+These values are the common petdex/Codex pet geometry. The real ``pet.json``
+usually only carries ``id``/``displayName``/``description``/``spritesheetPath``;
+row taxonomy is inferred from the atlas shape so Hermes can render both legacy
+8-row sheets and current 9-row Codex sheets.
+"""
+
+from __future__ import annotations
+
+from enum import Enum
+
+# Frame geometry (pixels). Current Codex/petdex spritesheets are 8 columns x 9
+# rows (1536x1872), while older Hermes/petdex sheets used 9 columns x 8 rows
+# (1728x1664). Renderers derive both row taxonomy and real column count from the
+# concrete sheet, so either shape works.
+FRAME_W = 192
+FRAME_H = 208
+
+# Frames consumed per animation state (the petdex web app uses CSS
+# ``steps(6)``).  A sheet may physically contain more columns; we only step
+# through the first ``FRAMES_PER_STATE``.
+FRAMES_PER_STATE = 6
+
+# Full-loop duration for one state, milliseconds (petdex default).
+LOOP_MS = 1100
+
+# Default on-screen scale relative to native frame size.  ``display.pet.scale``
+# is the single master scalar: the desktop canvas multiplies its native pixels
+# by it and every terminal surface derives its half-block/kitty column width
+# from it (see :func:`cols_for_scale`), so one number shrinks all three
+# interfaces together.  (petdex's own clients render at 0.7; we default smaller
+# so the kitty/GUI mascot stays a glanceable corner sprite.  The half-block
+# fallback can't shrink as far — see ``UNICODE_MIN_COLS`` — and clamps to its
+# legibility floor instead.)
+DEFAULT_SCALE = 0.33
+
+# User-settable scale bounds (``/pet scale``, desktop slider).  Floor keeps the
+# pet clickable/visible; ceiling stops a fat-fingered value from filling the
+# screen.  The unicode fallback additionally clamps to ``UNICODE_MIN_COLS``.
+MIN_SCALE = 0.1
+MAX_SCALE = 3.0
+
+
+def clamp_scale(scale: float) -> float:
+    """Clamp *scale* to ``[MIN_SCALE, MAX_SCALE]`` (the single validation point)."""
+    return max(MIN_SCALE, min(MAX_SCALE, scale))
+
+# Terminal cells one native frame spans at ``scale == 1.0``.  A cell is ~8px
+# wide, a frame is ``FRAME_W`` (192) px → 24 cells.  This mirrors the kitty
+# graphics placement (``scaled_px // 8``) so at full scale every renderer agrees.
+BASE_UNICODE_COLS = FRAME_W // 8
+
+# Legibility floor for the half-block fallback.  A half-block cell samples the
+# sprite at only 1 horizontal + 2 vertical taps, so below this width a 192×208
+# pet collapses into an unreadable blob *regardless* of scale.  kitty/GUI draw
+# true pixels and have no such floor — that's why the same ``scale: 0.33`` is
+# crisp there but mush in half-blocks.  ``scale`` shrinks the unicode pet down
+# TO this floor (and grows it above), instead of past it into noise.
+UNICODE_MIN_COLS = 16
+
+
+def cols_for_scale(scale: float) -> int:
+    """Half-block width implied by *scale*, clamped to the legibility floor.
+
+    Above the floor it tracks the kitty cell box (``scaled_px // 8``) so the two
+    renderers converge at larger sizes; below it the floor keeps the sprite
+    readable rather than letting it devolve into a blob.
+    """
+    return max(UNICODE_MIN_COLS, round(BASE_UNICODE_COLS * (scale or DEFAULT_SCALE)))
+
+
+def resolve_cols(scale: float, unicode_cols: int = 0) -> int:
+    """Resolve terminal width: explicit *unicode_cols* override, else from *scale*."""
+    return int(unicode_cols) if unicode_cols and int(unicode_cols) > 0 else cols_for_scale(scale)
+
+
+class PetState(str, Enum):
+    """Animation state a pet can be shown in.
+
+    These are Hermes' activity state names. They are not always identical to the
+    source atlas row names: Codex-format pets use rows like ``jumping`` /
+    ``running`` while the UI keeps the shorter ``jump`` / ``run`` names.
+    """
+
+    IDLE = "idle"
+    WAVE = "wave"
+    RUN = "run"
+    FAILED = "failed"
+    REVIEW = "review"
+    JUMP = "jump"
+    WAITING = "waiting"
+
+
+# Legacy Hermes/petdex row order (top -> bottom) used by the older 8-row,
+# 9-column atlas shape.
+LEGACY_STATE_ROWS: list[str] = [
+    PetState.IDLE.value,
+    PetState.WAVE.value,
+    PetState.RUN.value,
+    PetState.FAILED.value,
+    PetState.REVIEW.value,
+    PetState.JUMP.value,
+    "extra1",
+    "extra2",
+]
+
+# Current Petdex row order (top -> bottom) used by 1536x1872 atlases:
+# 8 columns x 9 rows of 192x208 cells.
+CODEX_STATE_ROWS: list[str] = [
+    PetState.IDLE.value,
+    "running-right",
+    "running-left",
+    "waving",
+    "jumping",
+    PetState.FAILED.value,
+    PetState.WAITING.value,
+    "running",
+    PetState.REVIEW.value,
+]
+
+# Default/fallback for callers without a sheet. Prefer the current 9-row Codex
+# format because generated pets and the public Codex pet contract use it.
+STATE_ROWS: list[str] = CODEX_STATE_ROWS
+
+# Canonical Hermes activity names -> accepted row-name aliases in descending
+# preference. This keeps our internal state names stable (`wave`/`jump`/`run`)
+# while matching Petdex's current `waving`/`jumping`/`running` taxonomy.
+STATE_ALIASES: dict[str, tuple[str, ...]] = {
+    PetState.IDLE.value: (PetState.IDLE.value,),
+    PetState.WAVE.value: (PetState.WAVE.value, "waving"),
+    PetState.JUMP.value: (PetState.JUMP.value, "jumping"),
+    PetState.RUN.value: (PetState.RUN.value, "running"),
+    PetState.FAILED.value: (PetState.FAILED.value,),
+    PetState.REVIEW.value: (PetState.REVIEW.value,),
+    PetState.WAITING.value: (PetState.WAITING.value,),
+}
+
+
+def state_aliases_for(state: "PetState | str") -> tuple[str, ...]:
+    """Return accepted row-name aliases for *state* (always non-empty)."""
+    value = state.value if isinstance(state, PetState) else str(state)
+    aliases = STATE_ALIASES.get(value)
+    return aliases if aliases else (value,)
+
+
+def state_rows_for_grid(row_count: int | None) -> list[str]:
+    """Return the row taxonomy for a spritesheet with *row_count* rows."""
+    try:
+        rows = int(row_count or 0)
+    except (TypeError, ValueError):
+        rows = 0
+
+    if rows >= len(CODEX_STATE_ROWS):
+        return CODEX_STATE_ROWS
+    return LEGACY_STATE_ROWS
+
+
+def state_row_index(state: "PetState | str", row_count: int | None = None) -> int:
+    """Return the spritesheet row index for *state* (clamped, never raises)."""
+    rows = state_rows_for_grid(row_count)
+    for name in state_aliases_for(state):
+        try:
+            return rows.index(name)
+        except ValueError:
+            continue
+    return 0  # fall back to the idle row

agent/pet/generate/__init__.py

@@ -0,0 +1,29 @@
+"""Pet generation — base-draft → hatch pipeline.
+
+Public surface used by the gateway RPCs, the CLI ``hermes pets generate``
+command, and tests:
+
+- :func:`generate_base_drafts` / :func:`hatch_pet` — the two-step flow.
+- :class:`HatchResult`, :class:`GenerationError`.
+- :mod:`atlas` — deterministic frame extraction + atlas composition/validation.
+
+Image generation is delegated to the active reference-capable
+:class:`~agent.image_gen_provider.ImageGenProvider` (OpenAI gpt-image-2 or Krea);
+atlas assembly is fully deterministic so it's testable without any API calls.
+"""
+
+from __future__ import annotations
+
+from agent.pet.generate.imagegen import GenerationError
+from agent.pet.generate.orchestrate import (
+    HatchResult,
+    generate_base_drafts,
+    hatch_pet,
+)
+
+__all__ = [
+    "GenerationError",
+    "HatchResult",
+    "generate_base_drafts",
+    "hatch_pet",
+]

agent/pet/generate/imagegen.py

@@ -0,0 +1,251 @@
+"""Thin image-generation layer for pet sprites.
+
+Wraps the active :class:`~agent.image_gen_provider.ImageGenProvider` with the
+two things sprite generation needs that the agent-facing ``image_generate`` tool
+doesn't expose: **N variants** (loop) and **reference-image grounding** (so each
+animation row stays the same character as the chosen base).
+
+Reference grounding only works on providers that support it — currently OpenAI
+``gpt-image-2`` (image edits) and Krea (style references). We resolve to one of
+those and surface a clear, actionable error otherwise rather than silently
+producing an ungrounded, drifting pet.
+"""
+
+from __future__ import annotations
+
+import logging
+import os
+from dataclasses import dataclass
+from pathlib import Path
+
+logger = logging.getLogger(__name__)
+
+# Providers that can ground generation on a reference image, in preference order
+# (Nous Portal → OpenAI → OpenRouter → …). OpenRouter/Nous run a quality-first
+# model chain and may fall back depending on account access and endpoint behavior,
+# so fidelity can vary by configured backend + model availability.
+_REF_CAPABLE = ("nous", "openai", "openai-codex", "openrouter", "krea")
+
+# Friendly display label per reference-capable provider, surfaced in the desktop
+# pet-gen picker.
+_PROVIDER_LABELS: dict[str, str] = {
+    "nous": "Nous Portal",
+    "openrouter": "OpenRouter",
+    "openai": "OpenAI",
+    "openai-codex": "OpenAI (Codex)",
+    "krea": "Krea",
+}
+
+
+def _forced_provider_from_env() -> str | None:
+    """Optional QA override to force a pet-gen backend.
+
+    `HERMES_PET_IMAGE_PROVIDER=<name>` (e.g. `openrouter`) bypasses the normal
+    active/default provider resolution for pet generation only. Unknown values are
+    ignored so existing users are unaffected.
+    """
+    forced = os.environ.get("HERMES_PET_IMAGE_PROVIDER", "").strip().lower()
+    return forced if forced in _REF_CAPABLE else None
+
+
+class GenerationError(RuntimeError):
+    """Raised on any image-generation failure (no provider, API error, IO)."""
+
+
+@dataclass(frozen=True)
+class SpriteProvider:
+    """Resolved provider plus whether it can take reference images."""
+
+    name: str
+    provider: object
+    supports_references: bool
+
+
+def _discover() -> None:
+    try:
+        from hermes_cli.plugins import _ensure_plugins_discovered
+
+        _ensure_plugins_discovered()
+    except Exception as exc:  # noqa: BLE001 - discovery is best-effort
+        logger.debug("image-gen plugin discovery failed: %s", exc)
+
+
+def resolve_provider(*, require_references: bool = True, prefer: str | None = None) -> SpriteProvider:
+    """Pick the image provider to use for sprite work.
+
+    Preference: an explicit *prefer* choice (the desktop pet-gen picker) when it's
+    reference-capable and configured, then the configured/active provider when
+    it's reference-capable, else the first available reference-capable provider.
+    With *require_references* off we fall back to any available provider (used for
+    prompt-only base drafts).
+    """
+    _discover()
+    from agent.image_gen_registry import get_active_provider, get_provider
+
+    # QA override: force one provider for pet-gen iteration regardless of the
+    # globally active image_gen backend.
+    forced = _forced_provider_from_env()
+    if forced:
+        chosen = get_provider(forced)
+        if chosen is not None and chosen.is_available():
+            return SpriteProvider(name=forced, provider=chosen, supports_references=True)
+
+    # An explicit user pick wins when it's reference-capable and has credentials;
+    # otherwise we ignore it and fall through to the normal resolution.
+    if prefer:
+        chosen = get_provider(prefer)
+        if prefer in _REF_CAPABLE and chosen is not None and chosen.is_available():
+            return SpriteProvider(name=prefer, provider=chosen, supports_references=True)
+
+    # Configured / active provider first.
+    active = None
+    try:
+        active = get_active_provider()
+    except Exception:  # noqa: BLE001
+        active = None
+    if active is not None:
+        name = getattr(active, "name", "")
+        if name in _REF_CAPABLE and active.is_available():
+            return SpriteProvider(name=name, provider=active, supports_references=True)
+
+    # Any available reference-capable provider.
+    for name in _REF_CAPABLE:
+        provider = get_provider(name)
+        if provider is not None and provider.is_available():
+            return SpriteProvider(name=name, provider=provider, supports_references=True)
+
+    if not require_references and active is not None and active.is_available():
+        return SpriteProvider(
+            name=getattr(active, "name", "unknown"), provider=active, supports_references=False
+        )
+
+    raise GenerationError(
+        "Pet generation needs an image backend that supports reference images. "
+        "Open `hermes tools` → Image Generation and configure Nous Portal, "
+        "OpenRouter, or OpenAI (gpt-image-2) with an API key."
+    )
+
+
+def list_sprite_providers() -> list[dict]:
+    """The reference-capable providers available to pick for pet generation.
+
+    Returns ``[{name, label, default}]`` for every ref-capable provider the user
+    actually has credentials for, in preference order, marking the one
+    :func:`resolve_provider` would choose with no explicit preference. Empty when
+    none is configured (the picker hides itself). Best-effort: discovery hiccups
+    yield an empty list.
+    """
+    _discover()
+    from agent.image_gen_registry import get_provider
+
+    try:
+        default_name = resolve_provider(require_references=True).name
+    except GenerationError:
+        default_name = ""
+
+    out: list[dict] = []
+    for name in _REF_CAPABLE:
+        provider = get_provider(name)
+        if provider is None or not provider.is_available():
+            continue
+        out.append(
+            {
+                "name": name,
+                "label": _PROVIDER_LABELS.get(name, name),
+                "default": name == default_name,
+            }
+        )
+    return out
+
+
+def _save_local(image_ref: str, *, prefix: str) -> Path:
+    """Return a local path for *image_ref*, downloading it if it's a URL."""
+    if image_ref.startswith(("http://", "https://")):
+        from agent.image_gen_provider import save_url_image
+
+        return Path(save_url_image(image_ref, prefix=prefix))
+    return Path(image_ref)
+
+
+def _rejected_background(error: str) -> bool:
+    """True when a provider error is specifically about the ``background`` param.
+
+    Transparent backgrounds are a per-model capability (e.g. some gpt-image tiers
+    reject ``background=transparent`` outright). We detect that one rejection so
+    we can retry without the flag rather than failing the whole pet — our chroma
+    key pass makes the result transparent regardless.
+    """
+    lowered = (error or "").lower()
+    return "background" in lowered and ("not supported" in lowered or "transparent" in lowered)
+
+
+def generate(
+    prompt: str,
+    *,
+    n: int = 1,
+    reference_images: list[Path] | None = None,
+    provider: SpriteProvider | None = None,
+    prefix: str = "pet_gen",
+    aspect_ratio: str = "square",
+) -> list[Path]:
+    """Generate *n* sprite images and return their local paths.
+
+    *reference_images* grounds the output on a base image (required for rows).
+    *aspect_ratio* picks the canvas: ``"square"`` for single-character base
+    drafts, ``"landscape"`` for multi-frame row strips (the wider 1536px canvas
+    gives every frame real horizontal room so winged poses don't have to be
+    shrunk to avoid touching their neighbors).
+    We *ask* for a transparent background, but fall back to an opaque generation
+    (cleaned up downstream by the chroma-key pass) on models that reject the
+    flag. Raises :class:`GenerationError` if nothing usable comes back.
+    """
+    sprite = provider or resolve_provider(require_references=bool(reference_images))
+    if reference_images and not sprite.supports_references:
+        raise GenerationError(
+            f"image backend '{sprite.name}' cannot use reference images; "
+            "configure OpenAI gpt-image-2 or Krea for pet generation"
+        )
+
+    refs = [str(p) for p in (reference_images or [])]
+
+    def _run(extra: dict) -> tuple[Path | None, str]:
+        kwargs: dict = {"aspect_ratio": aspect_ratio, **extra}
+        if refs:
+            # Providers disagree on the ref kwarg name: our OpenRouter/Nous
+            # backends read ``reference_images``, OpenAI's gpt-image-2 reads
+            # ``reference_image_urls``. Send both; each ignores the other.
+            kwargs["reference_images"] = refs
+            kwargs["reference_image_urls"] = refs
+        try:
+            result = sprite.provider.generate(prompt, **kwargs)
+        except Exception as exc:  # noqa: BLE001 - normalize provider crashes
+            logger.debug("provider.generate crashed: %s", exc)
+            return None, str(exc)
+        if not isinstance(result, dict) or not result.get("success"):
+            return None, (result or {}).get("error", "unknown error") if isinstance(result, dict) else "no result"
+        image_ref = result.get("image")
+        if not image_ref:
+            return None, "provider returned no image"
+        try:
+            return _save_local(str(image_ref), prefix=prefix), ""
+        except Exception as exc:  # noqa: BLE001
+            return None, f"could not save generated image: {exc}"
+
+    out: list[Path] = []
+    last_error = ""
+    allow_transparent = True
+    for _ in range(max(1, n)):
+        path, err = _run({"background": "transparent"} if allow_transparent else {})
+        # Model doesn't support the transparent flag → drop it for this and every
+        # remaining variant (no point re-probing a capability we just disproved).
+        if path is None and allow_transparent and _rejected_background(err):
+            allow_transparent = False
+            path, err = _run({})
+        if path is not None:
+            out.append(path)
+        else:
+            last_error = err
+
+    if not out:
+        raise GenerationError(last_error or "image generation produced no output")
+    return out

agent/pet/generate/orchestrate.py

@@ -0,0 +1,358 @@
+"""Pet generation orchestration — the base-draft → hatch flow.
+
+Two steps, mirroring the UX across every surface:
+
+1. :func:`generate_base_drafts` — a handful of prompt-only "what should this pet
+   look like" variants. Cheap; the user picks one (or retries for a fresh set).
+2. :func:`hatch_pet` — takes the chosen base and generates one grounded row
+   strip per Hermes state, slices each into frames, composes the atlas, validates
+   it, and writes the pet into the store.
+
+Splitting it this way bounds cost (4 cheap base calls per round; the ~6 row
+calls happen once, on the pet you actually keep) and gives each UI a natural
+preview/loading point.
+"""
+
+from __future__ import annotations
+
+import logging
+import time
+from concurrent.futures import ThreadPoolExecutor, as_completed
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Callable
+
+from agent.pet.generate import atlas, imagegen, prompts
+from agent.pet.generate.imagegen import GenerationError, SpriteProvider
+
+logger = logging.getLogger(__name__)
+
+# (event, detail) — e.g. ("row", "idle"), ("compose", ""), ("save", "<slug>").
+ProgressFn = Callable[[str, str], None]
+
+# Image generations are independent network calls, so we fan them out instead of
+# blocking on each in turn — a hatch is ~8 row calls that would otherwise run
+# back-to-back and routinely blow past the client's RPC timeout. Capped so we
+# don't hammer the provider's rate limit (one cold call can still be slow).
+_MAX_PARALLEL_GENERATIONS = 4
+# How many times to (re)generate a single row before accepting a best-effort
+# slice. Early attempts demand clean per-pose gutters; the last is lenient so a
+# stubborn row still yields frames instead of dropping out entirely.
+_ROW_GEN_ATTEMPTS = 3
+_MIN_FILLED_STATES = 6
+_REQUIRED_STATES = frozenset({"idle", "running-right", "waving"})
+
+
+@dataclass(frozen=True)
+class HatchResult:
+    """Outcome of a successful :func:`hatch_pet`."""
+
+    slug: str
+    display_name: str
+    spritesheet: Path
+    states: list[str]
+    validation: dict
+
+
+def _harden_transparency(path: Path) -> Path:
+    """Key out any solid backdrop the provider painted; save as an RGBA PNG.
+
+    ``background=transparent`` is requested on every call, but image models honor
+    it inconsistently — some still paint a flat (often near-white) backdrop. We
+    run the same chroma-key pass the row extractor uses so every base draft the
+    user picks between (and the reference the rows are grounded on) is a clean
+    cutout. Best-effort: a decode failure leaves the original untouched.
+    """
+    from PIL import Image
+
+    try:
+        with Image.open(path) as opened:
+            keyed = atlas.remove_background(opened.convert("RGBA"))
+        # Zero the RGB of any leftover semi-transparent edge pixels so a keyed
+        # draft has no colored halo when composited on the dark UI.
+        keyed = atlas._clear_transparent_rgb(keyed)
+        out = path.with_suffix(".png")
+        keyed.save(out, format="PNG")
+        return out
+    except Exception as exc:  # noqa: BLE001 - cosmetic; fall back to the raw image
+        logger.debug("base draft transparency hardening failed for %s: %s", path, exc)
+        return path
+
+
+def generate_base_drafts(
+    concept: str,
+    *,
+    n: int = 4,
+    style: str = "auto",
+    reference_images: list[Path] | None = None,
+    provider: SpriteProvider | None = None,
+    on_draft: Callable[[int, Path], None] | None = None,
+    is_cancelled: Callable[[], bool] | None = None,
+) -> list[Path]:
+    """Generate *n* candidate base looks for *concept*; returns image paths.
+
+    Each draft is hardened to a transparent cutout (see :func:`_harden_transparency`).
+    Drafts are generated concurrently and *on_draft(index, path)* fires as each
+    one finishes (not at the end) so callers can stream previews to the UI
+    instead of leaving it blank until the whole batch is done.
+
+    *is_cancelled*, when supplied, is polled cooperatively: a draft that hasn't
+    started yet is skipped, and once it trips we stop staging/streaming further
+    drafts and cancel any queued work (already-in-flight provider calls can't be
+    hard-killed, but their results are dropped).
+    """
+    # A user reference image (e.g. their own pet) grounds every draft, so it
+    # needs a reference-capable provider — same requirement as the row passes.
+    refs = reference_images or None
+    sprite = provider or imagegen.resolve_provider(require_references=bool(refs))
+    cancelled = is_cancelled or (lambda: False)
+
+    # Each draft is its own one-shot generation, run concurrently so the user
+    # waits for one image, not N. A single draft failing must not sink the set.
+    # Each gets a distinct variation nudge so the options aren't near-duplicates.
+    logger.info("pet generate: drafting %d base looks for %r (style=%s)", n, concept, style)
+
+    def _one(index: int) -> tuple[int, Path | None, str | None]:
+        if cancelled():
+            return index, None, None
+        t0 = time.monotonic()
+        variation = prompts.BASE_VARIATIONS[index % len(prompts.BASE_VARIATIONS)]
+        prompt = prompts.build_base_prompt(concept, style=style, variation=variation)
+        try:
+            out = imagegen.generate(prompt, n=1, reference_images=refs, provider=sprite, prefix="pet_base")
+        except Exception as exc:  # noqa: BLE001 - tolerate a single failed draft
+            logger.warning("pet generate: draft %d failed after %.1fs: %s", index, time.monotonic() - t0, exc)
+            return index, None, str(exc)
+        if not out:
+            logger.warning("pet generate: draft %d produced no image", index)
+            return index, None, "the image provider returned no image"
+        logger.info("pet generate: draft %d ready in %.1fs", index, time.monotonic() - t0)
+        return index, _harden_transparency(out[0]), None
+
+    workers = max(1, min(n, _MAX_PARALLEL_GENERATIONS))
+    results: dict[int, Path] = {}
+    errors: list[str] = []
+    with ThreadPoolExecutor(max_workers=workers) as pool:
+        futures = [pool.submit(_one, i) for i in range(n)]
+        # as_completed runs in *this* (the caller's) thread, so on_draft — and any
+        # gateway event it emits — inherits the request's bound transport, unlike
+        # the worker threads above.
+        for fut in as_completed(futures):
+            if cancelled():
+                logger.info("pet generate: cancelled — dropping remaining drafts")
+                for pending in futures:
+                    pending.cancel()
+                break
+            index, path, err = fut.result()
+            if path is None:
+                if err:
+                    errors.append(err)
+                continue
+            results[index] = path
+            if on_draft is not None:
+                try:
+                    on_draft(index, path)
+                except Exception as exc:  # noqa: BLE001 - progress is best-effort
+                    logger.debug("on_draft callback failed: %s", exc)
+
+    drafts = [results[i] for i in sorted(results)]
+    if not drafts and not cancelled():
+        # Surface *why* — every draft failed for a reason (a content-policy refusal
+        # on a name like "minion", a provider/auth error, …); the most common one
+        # is the representative cause. Far more useful than "no usable drafts".
+        raise GenerationError(_drafts_failed_reason(errors))
+    return drafts
+
+
+def _drafts_failed_reason(errors: list[str]) -> str:
+    """The representative reason a draft round produced nothing, humanized."""
+    if not errors:
+        return "image generation produced no usable drafts"
+    from collections import Counter
+
+    return _humanize_image_error(Counter(errors).most_common(1)[0][0])
+
+
+def _humanize_image_error(error: str) -> str:
+    """Turn a raw provider error into a friendly, actionable sentence.
+
+    The big one is moderation: image models refuse trademarked characters and
+    real people (e.g. "minion"), which reads as an opaque 400 otherwise.
+    """
+    low = error.lower()
+    if any(s in low for s in ("moderation_blocked", "safety system", "content policy", "content_policy")):
+        return (
+            "The image provider blocked this prompt — its safety filter rejects "
+            "trademarked characters and real people. Try an original description."
+        )
+    if any(s in low for s in ("api key", "unauthorized", "401", "auth")):
+        return "The image provider rejected the request — check your API key in Settings → Providers."
+    if "rate limit" in low or "429" in low:
+        return "The image provider is rate-limiting — wait a moment and try again."
+    # Otherwise the first line, trimmed of the noisy provider envelope.
+    return error.splitlines()[0].strip()[:200]
+
+
+def hatch_pet(
+    *,
+    base_image: str | Path,
+    slug: str,
+    display_name: str = "",
+    description: str = "",
+    concept: str = "",
+    style: str = "auto",
+    on_progress: ProgressFn | None = None,
+    provider: SpriteProvider | None = None,
+    is_cancelled: Callable[[], bool] | None = None,
+) -> HatchResult:
+    """Turn an approved base image into a full, installed Hermes pet.
+
+    Generates a grounded row strip per state, extracts frames, composes +
+    validates the atlas, and registers it. The idle row falls back to the base
+    look so the pet always renders. Raises :class:`GenerationError` on failure.
+
+    *is_cancelled*, when supplied, is polled cooperatively: rows that haven't
+    started are skipped, queued rows are cancelled, and once every row is done we
+    abort (raising :class:`GenerationError`) before composing/saving so a stopped
+    hatch never writes a half-built pet.
+    """
+    base = Path(base_image)
+    if not base.is_file():
+        raise GenerationError(f"base image not found: {base}")
+
+    sprite = provider or imagegen.resolve_provider(require_references=True)
+    progress = on_progress or (lambda *_: None)
+    cancelled = is_cancelled or (lambda: False)
+    label = concept or display_name or slug
+
+    frames_by_state: dict[str, list] = {}
+    total_rows = len(atlas.ROW_SPECS)
+    logger.info("pet hatch %r: generating %d animation rows", slug, total_rows)
+
+    # Generate every state's row strip concurrently — they're independent
+    # grounded calls, so the hatch waits for the slowest row, not their sum. A
+    # single row failing is tolerated (idle is guaranteed below).
+    def _gen_row(spec: tuple[str, int, int]) -> tuple[str, list | None]:
+        state, _row, count = spec
+        if cancelled():
+            return state, None
+        t0 = time.monotonic()
+        last_exc: Exception | None = None
+        # Self-healing: a model occasionally returns a row whose poses are touching
+        # (no clean gutters), which slices badly. We retry such rolls; only the
+        # final attempt falls back to lenient ``auto`` slicing so a stubborn row
+        # still yields *something* rather than dropping the whole row.
+        for attempt in range(_ROW_GEN_ATTEMPTS):
+            if cancelled():
+                return state, None
+            strict = attempt < _ROW_GEN_ATTEMPTS - 1
+            try:
+                strips = imagegen.generate(
+                    prompts.build_row_prompt(state, count, label, style=style),
+                    n=1,
+                    reference_images=[base],
+                    provider=sprite,
+                    prefix=f"pet_row_{state}",
+                    # Wider canvas → each frame gets real horizontal room, so winged
+                    # poses keep a full, healthy size and still leave clean gutters.
+                    aspect_ratio="landscape",
+                )
+                # ``components`` requires clean per-pose gutters (raises otherwise),
+                # so a touching roll is rejected and regenerated; the last attempt
+                # uses ``auto`` (equal-slot fallback, never raises). Raw (fit=False)
+                # so normalize_cells registers the whole pet at once.
+                method = "components" if strict else "auto"
+                frames = atlas.extract_strip_frames(strips[0], count, method=method, fit=False)
+                logger.info(
+                    "pet hatch %r: row %r ready in %.1fs (attempt %d)",
+                    slug, state, time.monotonic() - t0, attempt + 1,
+                )
+                return state, frames
+            except Exception as exc:  # noqa: BLE001 - retried; one bad row is tolerated
+                last_exc = exc
+                logger.warning(
+                    "pet hatch %r: row %r attempt %d/%d failed: %s",
+                    slug, state, attempt + 1, _ROW_GEN_ATTEMPTS, exc,
+                )
+        logger.warning(
+            "pet hatch %r: row %r gave up after %.1fs: %s",
+            slug, state, time.monotonic() - t0, last_exc,
+        )
+        return state, None
+
+    # running-left is derived by mirroring running-right (guaranteed-consistent
+    # and one fewer generation), so we don't generate it directly.
+    generated_specs = [spec for spec in atlas.ROW_SPECS if spec[0] != "running-left"]
+
+    workers = max(1, min(len(generated_specs), _MAX_PARALLEL_GENERATIONS))
+    done = 0
+    with ThreadPoolExecutor(max_workers=workers) as pool:
+        futures = [pool.submit(_gen_row, spec) for spec in generated_specs]
+        # as_completed runs on the caller (request) thread, so progress events
+        # emitted here inherit the request transport — unlike the worker threads.
+        for fut in as_completed(futures):
+            if cancelled():
+                logger.info("pet hatch %r: cancelled — dropping remaining rows", slug)
+                for pending in futures:
+                    pending.cancel()
+                break
+            state, frames = fut.result()
+            done += 1
+            progress("row", f"{state}:{done}:{total_rows}")
+            if frames:
+                frames_by_state[state] = frames
+
+    if cancelled():
+        raise GenerationError("hatch cancelled")
+
+    # Derive running-left from the approved running-right row (per-frame mirror,
+    # preserving order/timing). Missing running-right is rejected below; a pet
+    # without its canonical walk cycle is a failed hatch, not a shippable mascot.
+    right = frames_by_state.get("running-right")
+    if right:
+        done += 1
+        progress("row", f"running-left:{done}:{total_rows}")
+        frames_by_state["running-left"] = atlas.mirror_frames(right)
+        logger.info("pet hatch %r: row 'running-left' mirrored from running-right", slug)
+    else:
+        logger.warning("pet hatch %r: no running-right to mirror; left walk left empty", slug)
+
+    # Idle is the resting state the renderer falls back to — guarantee it.
+    if not frames_by_state.get("idle"):
+        progress("row", "idle-fallback")
+        frames_by_state["idle"] = [atlas.single_frame(base, fit=False)]
+
+    progress("compose", "")
+    logger.info("pet hatch %r: composing atlas from %d states", slug, len(frames_by_state))
+    # One shared scale + baseline across every state so the pet never slides or
+    # pulses size between frames; compose just packs the normalized cells.
+    sheet = atlas.compose_atlas(atlas.normalize_cells(frames_by_state))
+    validation = atlas.validate_atlas(sheet)
+    if not validation["ok"]:
+        raise GenerationError("; ".join(validation["errors"]) or "atlas validation failed")
+    filled_states = set(validation["filled_states"])
+    missing_required = sorted(_REQUIRED_STATES - filled_states)
+    if missing_required:
+        raise GenerationError(f"missing required animation row(s): {', '.join(missing_required)}")
+    if len(filled_states) < _MIN_FILLED_STATES:
+        raise GenerationError(
+            f"only {len(filled_states)}/{len(atlas.ROW_SPECS)} animation rows were usable; regenerate"
+        )
+
+    from agent.pet import store
+
+    progress("save", slug)
+    logger.info("pet hatch %r: saving pet", slug)
+    pet = store.register_local_pet(
+        sheet,
+        slug=slug,
+        display_name=display_name or slug,
+        description=description,
+    )
+    return HatchResult(
+        slug=pet.slug,
+        display_name=pet.display_name,
+        spritesheet=pet.spritesheet,
+        states=validation["filled_states"],
+        validation=validation,
+    )

agent/pet/generate/prompts.py

@@ -0,0 +1,183 @@
+"""Prompt builders for pet generation.
+
+Two prompt shapes: a *base* prompt (prompt-only, produces the canonical look the
+user picks between) and per-*state* *row* prompts (grounded on the chosen base,
+produce one horizontal strip of N poses). Prompts stay concise and
+sprite-production oriented; the identity lock and "one transparent row" framing
+matter more than flowery description.
+
+We generate the full petdex/Codex nine-state set (see
+:data:`agent.pet.generate.atlas.ROW_SPECS`) so a hatched pet is a valid
+``petdex submit`` spritesheet.
+"""
+
+from __future__ import annotations
+
+# What each petdex/Codex state should depict (kept short — these go straight into
+# the row prompt). Phrased to avoid the common sprite-gen failure modes (detached
+# effects, motion lines, shadows). Critical distinction: ``running`` is the
+# *working* state (in place), while ``running-right`` / ``running-left`` are the
+# actual directional walk/run cycles.
+STATE_ACTIONS: dict[str, str] = {
+    "idle": "a calm idle loop: subtle breathing, a tiny blink or gentle bob, no big gestures",
+    "running-right": (
+        "a sideways walk/run locomotion cycle moving to the RIGHT: the character "
+        "faces and travels right with clear directional steps, a smooth gait loop"
+    ),
+    "running-left": (
+        "a sideways walk/run locomotion cycle moving to the LEFT: the character "
+        "faces and travels left with clear directional steps (the mirror of the "
+        "right-facing run)"
+    ),
+    "waving": "a friendly greeting: raising a paw/hand/limb to wave, clear up-and-down gesture",
+    "jumping": "a happy celebration jump: anticipation, lift off the ground, peak, and land",
+    "failed": "a sad or deflated reaction: slumped, dejected, small frown — readable but not noisy",
+    "waiting": (
+        "an expectant 'waiting on you' pose: looking up/out as if asking for input "
+        "or approval — distinct from idle and review"
+    ),
+    "running": (
+        "focused active work, staying IN PLACE (NOT walking or foot-running): "
+        "leaning in, concentrating, busy 'thinking / processing / typing' energy"
+    ),
+    "review": "careful inspection: a focused lean, head tilt, studying something intently",
+}
+
+_STYLE_HINTS: dict[str, str] = {
+    # Default to the popular petdex look: crisp 16-bit PIXEL ART, not the smooth
+    # 2D illustration (let alone 3D render) gpt-image reaches for by default.
+    "auto": (
+        " Style: crisp 16-bit PIXEL-ART game sprite — visible square pixels, a small "
+        "limited palette, clean dark outline, flat cel shading, chunky chibi "
+        "proportions, like a classic SNES/JRPG party member or a petdex.dev mascot. "
+        "Absolutely NOT 3D-rendered, NOT a smooth painted or vector illustration, "
+        "NOT photorealistic — no soft gradients, no realistic lighting, no figurine look."
+    ),
+    "pixel": " Render in clean 16-bit pixel-art style with visible square pixels and a limited palette.",
+    "plush": " Render as a soft plush toy.",
+    "clay": " Render as a claymation / soft 3D clay figure.",
+    "sticker": " Render as a glossy die-cut sticker.",
+    "flat-vector": " Render in flat vector mascot style.",
+    "3d-toy": " Render as a glossy 3D toy.",
+    "painterly": " Render in a soft painterly style.",
+}
+
+_BACKGROUND = (
+    "Center the character on a SINGLE flat, uniform, high-contrast chroma-key "
+    "background — pure hot magenta #FF00FF (only if magenta appears on the "
+    "character, use pure green #00FF00 instead). The background is ONE continuous "
+    "even color that completely surrounds the character with NO gradient, "
+    "vignette, texture, pattern, scenery, shadow, ground line, frame, border, "
+    "panel, comic cell, gutter line, grid, or divider of any kind, so it keys out "
+    "cleanly. The background color must not appear anywhere on the character. "
+    "No text, no labels, no speech bubbles, no UI."
+)
+
+
+def style_hint(style: str | None) -> str:
+    return _STYLE_HINTS.get((style or "auto").strip().lower(), "")
+
+
+# Row strips are generated on the wider landscape canvas (see imagegen.generate /
+# orchestrate). The extra width is what lets each pose stay a healthy size AND
+# leave a real gutter — used here only to cite concrete pixel numbers.
+_ASSUMED_STRIP_WIDTH = 1536
+
+
+def _spacing_spec(frame_count: int) -> tuple[int, int]:
+    """(per-pose width px, gap px) for a row of *frame_count* poses.
+
+    Pixel counts alone don't hold — the model fills each slot edge-to-edge with
+    the full wingspan, so neighbors touch even when bodies are spaced. The lever
+    that works is proportional containment on a wide canvas: give each pose its
+    own equal cell and keep the ENTIRE silhouette (wings/tail/halo included)
+    inside it. On the 1536px landscape strip ~70% occupancy still leaves a
+    generous gutter, so the pet stays a normal, good-looking size — no shrinking.
+    """
+    slots = max(1, frame_count)
+    slot_w = _ASSUMED_STRIP_WIDTH / slots
+    pose_px = round(slot_w * 0.7)
+    gap_px = max(48, round(slot_w * 0.3))
+    return pose_px, gap_px
+
+
+# Per-draft nudges so the 4 base options are actually distinct — gpt-image returns
+# near-duplicates for a single prompt. We vary the *look* (palette, build,
+# expression, accents), NOT the pose, so the chosen base still grounds clean,
+# consistent animation rows.
+BASE_VARIATIONS: tuple[str, ...] = (
+    "",
+    "a distinctly different colour palette and markings",
+    "a heavier, broader silhouette with sturdier proportions",
+    "a different facial structure and expression matching the concept tone, with unique accent/accessory details",
+    "a leaner, taller build and an alternate colour scheme",
+    "bolder, more saturated colours and a stronger expression matching the concept tone",
+)
+
+
+def build_base_prompt(concept: str, *, style: str | None = "auto", variation: str = "") -> str:
+    """The base look: a single, clean, centered full-body mascot.
+
+    *variation* differentiates one draft from the next (see :data:`BASE_VARIATIONS`).
+    """
+    concept = (concept or "a distinctive mascot creature").strip()
+    nudge = f" Make this design distinct: {variation}." if variation else ""
+    return (
+        f"A stylized mascot pet character: {concept}. "
+        "Honor the requested tone and mood exactly (cute, eerie, scary, menacing, whimsical, etc.) "
+        "while staying non-graphic. "
+        "Compact, whole-body silhouette that reads clearly at small size, "
+        "clear readable facial features, simple consistent palette. "
+        # A neutral, symmetric, at-rest stance makes the cleanest identity anchor
+        "Neutral front-facing standing pose, upright and symmetric, arms/limbs "
+        "relaxed at the sides, feet together on the ground, any cape/accessories "
+        "hanging straight and still."
+        f"{nudge} "
+        f"{_BACKGROUND}{style_hint(style)}"
+    )
+
+
+def build_row_prompt(state: str, frame_count: int, concept: str, *, style: str | None = "auto") -> str:
+    """A row strip: *frame_count* poses of the SAME character, left→right.
+
+    The attached base image is the identity source of truth; the prompt locks
+    species, palette, face, and props to it.
+    """
+    action = STATE_ACTIONS.get(state, "a simple idle pose")
+    concept = (concept or "the mascot").strip()
+    pose_px, gap_px = _spacing_spec(frame_count)
+    return (
+        f"Using the attached reference image as the exact same character "
+        f"(same species, face, colors, markings, proportions, and props), "
+        "preserving the same emotional tone/mood (e.g., scary stays scary, cute stays cute), "
+        f"draw a single WIDE horizontal strip of {frame_count} animation frames showing {action}. "
+        f"LAYOUT: arrange {frame_count} poses in ONE horizontal row at equal spacing, "
+        "each pose centered in its own imaginary equal region. Draw NO panel borders, "
+        "NO comic cells, NO boxes, NO vertical divider/gutter lines, NO grid, NO frame "
+        "outlines between poses — the backdrop is one unbroken flat field behind all of them. "
+        "Fill the WHOLE strip with the SAME single flat chroma-key color as the attached "
+        "reference image's background (identical hue in every frame, no per-pose color shifts). "
+        f"SPACING (critical): draw each pose at a consistent, healthy, clearly "
+        f"visible size (roughly {pose_px}px wide on a {_ASSUMED_STRIP_WIDTH}px "
+        f"strip) — do NOT shrink it tiny — but keep its ENTIRE silhouette "
+        f"(wings, tail, halo, horns, cape, every appendage) fully INSIDE its own "
+        f"cell. Leave at least {gap_px}px of empty chroma-key background between "
+        f"neighboring silhouettes at their closest point (wingtip to wingtip), and "
+        f"the same empty margin before the first pose and after the last. If a wing, "
+        f"cape, or tail would reach into a neighbor, FOLD or angle it inward rather "
+        f"than letting it cross the gap. Silhouettes must NEVER touch, overlap, "
+        f"share a shadow, share a ground line, share motion trails, or merge into "
+        f"one connected shape. "
+        # Registration: a clean sprite sheet keeps the character locked in place
+        # so only the action moves — this is what stops the loop sliding/pulsing.
+        "REGISTRATION (critical): the character is the SAME height and SAME width "
+        "in every frame, drawn at the SAME scale, centered over the SAME point, "
+        "with all feet aligned to the SAME invisible horizontal baseline across the "
+        "whole strip — this baseline is conceptual ONLY: draw NO ground line, floor, "
+        "platform, horizon, or contact shadow beneath the feet. Keep the body's center, size, and stance fixed frame to "
+        "frame — ONLY the limbs/features the action needs may move. Capes, cloaks, "
+        "bags, and scarves stay in the SAME place and shape every frame (no "
+        "swinging, flowing, or drifting) unless the action itself requires it. No "
+        "pose is cropped at the strip edges. "
+        f"{_BACKGROUND}{style_hint(style)}"
+    )

agent/pet/manifest.py

@@ -0,0 +1,165 @@
+"""Fetch the public petdex manifest.
+
+``https://petdex.dev/api/manifest`` 307-redirects to a JSON document on R2:
+
+    {
+      "generatedAt": "...",
+      "total": 2926,
+      "pets": [
+        {"slug": "boba", "displayName": "Boba", "kind": "creature",
+         "submittedBy": "railly",
+         "spritesheetUrl": "https://assets.petdex.dev/.../spritesheet.webp",
+         "petJsonUrl": "https://assets.petdex.dev/.../pet.json",
+         "zipUrl": "https://assets.petdex.dev/.../boba.zip"},
+        ...
+      ]
+    }
+
+Read-only and unauthenticated; no credentials involved.
+"""
+
+from __future__ import annotations
+
+import logging
+import threading
+import time
+from dataclasses import dataclass
+
+logger = logging.getLogger(__name__)
+
+MANIFEST_URL = "https://petdex.dev/api/manifest"
+
+_DEFAULT_TIMEOUT = 10.0
+
+# In-process cache for the (large, slow, identical-per-call) manifest. The list
+# is a static CDN object that barely changes, yet a single session can ask for
+# it many times — every gallery open, plus a full re-fetch per install/select
+# (``find_entry``). A short TTL collapses those into one network hit without
+# going stale for long. Cleared by :func:`clear_cache` (tests).
+_MANIFEST_TTL = 300.0
+_cache: tuple[float, list[ManifestEntry]] | None = None
+
+_prefetch_lock = threading.Lock()
+_prefetching = False
+
+
+def clear_cache() -> None:
+    """Drop the cached manifest (forces the next fetch to hit the network)."""
+    global _cache
+    _cache = None
+
+
+def _cache_is_warm() -> bool:
+    return _cache is not None and time.monotonic() - _cache[0] < _MANIFEST_TTL
+
+
+def prefetch(*, timeout: float = _DEFAULT_TIMEOUT) -> None:
+    """Warm the manifest cache in a daemon thread — idempotent, never blocks.
+
+    The desktop picker calls this when it loads the (instant) local-only gallery
+    so the full petdex catalog is usually cached by the time it's requested,
+    without ever holding up the user's own pets on a network round-trip.
+    """
+    global _prefetching
+
+    if _cache_is_warm():
+        return
+
+    with _prefetch_lock:
+        if _prefetching:
+            return
+        _prefetching = True
+
+    def _run() -> None:
+        global _prefetching
+        try:
+            fetch_manifest(timeout=timeout)
+        except Exception as exc:  # noqa: BLE001 - best-effort warm
+            logger.debug("petdex manifest prefetch failed: %s", exc)
+        finally:
+            _prefetching = False
+
+    threading.Thread(target=_run, name="petdex-prefetch", daemon=True).start()
+
+
+@dataclass(frozen=True)
+class ManifestEntry:
+    """A single pet's row in the manifest."""
+
+    slug: str
+    display_name: str
+    kind: str
+    submitted_by: str
+    spritesheet_url: str
+    pet_json_url: str
+    zip_url: str
+
+    @classmethod
+    def from_dict(cls, data: dict) -> "ManifestEntry":
+        return cls(
+            slug=str(data.get("slug", "")).strip(),
+            display_name=str(data.get("displayName", "") or data.get("slug", "")),
+            kind=str(data.get("kind", "") or "pet"),
+            submitted_by=str(data.get("submittedBy", "") or ""),
+            spritesheet_url=str(data.get("spritesheetUrl", "") or ""),
+            pet_json_url=str(data.get("petJsonUrl", "") or ""),
+            zip_url=str(data.get("zipUrl", "") or ""),
+        )
+
+
+class ManifestError(RuntimeError):
+    """Raised when the manifest can't be fetched or parsed."""
+
+
+def fetch_manifest(*, timeout: float = _DEFAULT_TIMEOUT, force: bool = False) -> list[ManifestEntry]:
+    """Return every approved pet from the public manifest.
+
+    Cached in-process for ``_MANIFEST_TTL`` seconds (pass ``force=True`` to
+    bypass). Follows the 307 redirect to R2.  Raises :class:`ManifestError` on
+    any network/parse failure so callers can surface a clean message.
+    """
+    global _cache
+
+    if not force and _cache is not None and time.monotonic() - _cache[0] < _MANIFEST_TTL:
+        return _cache[1]
+
+    try:
+        import httpx
+    except ImportError as exc:  # pragma: no cover - httpx is a core dep
+        raise ManifestError("httpx is required to fetch the petdex manifest") from exc
+
+    try:
+        resp = httpx.get(
+            MANIFEST_URL,
+            timeout=timeout,
+            follow_redirects=True,
+            headers={"User-Agent": "hermes-agent-petdex"},
+        )
+        resp.raise_for_status()
+        payload = resp.json()
+    except Exception as exc:  # noqa: BLE001 - normalize to one error type
+        raise ManifestError(f"could not fetch petdex manifest: {exc}") from exc
+
+    pets = payload.get("pets") if isinstance(payload, dict) else None
+    if not isinstance(pets, list):
+        raise ManifestError("petdex manifest had no 'pets' array")
+
+    entries: list[ManifestEntry] = []
+    for raw in pets:
+        if not isinstance(raw, dict):
+            continue
+        entry = ManifestEntry.from_dict(raw)
+        if entry.slug and entry.spritesheet_url:
+            entries.append(entry)
+
+    _cache = (time.monotonic(), entries)
+    return entries
+
+
+def find_entry(slug: str, *, timeout: float = _DEFAULT_TIMEOUT) -> ManifestEntry | None:
+    """Return the manifest entry for *slug*, or ``None`` if not listed."""
+    slug = slug.strip().lower()
+    for entry in fetch_manifest(timeout=timeout):
+        if entry.slug.lower() == slug:
+            return entry
+    return None

agent/pet/render.py

@@ -0,0 +1,575 @@
+"""Decode a pet spritesheet and encode frames for a terminal.
+
+Shared by the base CLI (writes the escape bytes to its own stdout) and the
+TUI (``tui_gateway`` ships the encoded bytes to Ink, which writes them) so the
+decode + capability-detection + protocol-encoding logic exists exactly once.
+
+Supported output modes, in fidelity order:
+
+- ``kitty``   — the kitty graphics protocol (kitty, Ghostty, WezTerm).
+- ``iterm``   — iTerm2 inline images (iTerm2, WezTerm).
+- ``sixel``   — DEC sixel (xterm -ti vt340, foot, mlterm, WezTerm, …).
+- ``unicode`` — 24-bit half-block downscale; works in any truecolor terminal.
+
+Frame decoding requires Pillow (a core Hermes dependency).  If Pillow or the
+spritesheet is unavailable the renderer degrades to ``unicode`` text or an
+empty string rather than raising.
+"""
+
+from __future__ import annotations
+from agent.pet.terminal import detect_terminal_graphics
+
+import base64
+import io
+import logging
+import os
+import sys
+from functools import lru_cache
+from pathlib import Path
+
+from agent.pet.constants import (
+    DEFAULT_SCALE,
+    FRAME_H,
+    FRAME_W,
+    FRAMES_PER_STATE,
+    PetState,
+    state_row_index,
+)
+
+logger = logging.getLogger(__name__)
+
+# Public render-mode names accepted by ``display.pet.render_mode``.
+RENDER_MODES = ("auto", "kitty", "iterm", "sixel", "unicode", "off")
+
+
+# ─────────────────────────────────────────────────────────────────────────
+# Terminal capability detection
+# ─────────────────────────────────────────────────────────────────────────
+
+
+def resolve_mode(configured: str | None, *, stream=None) -> str:
+    """Resolve the effective render mode from config + the environment.
+
+    ``configured`` is ``display.pet.render_mode`` (``auto`` → detect).  Returns
+    ``off`` when not attached to a TTY (no point emitting graphics into a pipe
+    or logfile).
+    """
+    mode = (configured or "auto").strip().lower()
+    if mode not in RENDER_MODES:
+        mode = "auto"
+    if mode == "off":
+        return "off"
+
+    stream = stream or sys.stdout
+    try:
+        if not (hasattr(stream, "isatty") and stream.isatty()):
+            return "off"
+    except (ValueError, OSError):
+        return "off"
+
+    if mode == "auto":
+        return detect_terminal_graphics()
+    return mode
+
+
+# ─────────────────────────────────────────────────────────────────────────
+# Frame decoding
+# ─────────────────────────────────────────────────────────────────────────
+
+def _open_sheet(path: Path):
+    from PIL import Image
+
+    img = Image.open(path)
+    return img.convert("RGBA")
+
+
+# Max alpha at/below which a frame counts as blank padding.  petdex sheets are
+# left-packed: a state with fewer real frames than ``FRAMES_PER_STATE`` fills
+# the trailing columns with fully transparent cells.  Animating into one flashes
+# the pet blank, so we stop the row at the first such gap.
+_BLANK_ALPHA = 8
+
+
+def _frame_is_blank(frame) -> bool:
+    """True if *frame* has no meaningfully opaque pixel (transparent padding)."""
+    return frame.getchannel("A").getextrema()[1] <= _BLANK_ALPHA
+
+
+@lru_cache(maxsize=16)
+def _raw_frames(
+    sheet_path: str,
+    state_value: str,
+    frame_w: int,
+    frame_h: int,
+    frames_per_state: int,
+) -> tuple:
+    """Cropped, padding-trimmed RGBA frames for one state row (unscaled).
+
+    Steps across the row until the first blank column so pets with ragged
+    per-state frame counts never animate into empty padding.  Cached; returns
+    ``()`` on any decode failure.
+    """
+    try:
+        sheet = _open_sheet(Path(sheet_path))
+        cols = max(1, sheet.width // frame_w)
+        rows = max(1, sheet.height // frame_h)
+        row = state_row_index(state_value, rows)
+        top = row * frame_h
+        # Clamp the row to the sheet (some pets ship fewer rows than the 8 the
+        # taxonomy reserves).
+        if top + frame_h > sheet.height:
+            top = max(0, sheet.height - frame_h)
+
+        frames = []
+        for i in range(min(frames_per_state, cols)):
+            left = i * frame_w
+            frame = sheet.crop((left, top, left + frame_w, top + frame_h))
+            if _frame_is_blank(frame):
+                break  # trailing transparent padding — real frames end here
+            frames.append(frame)
+        return tuple(frames)
+    except Exception as exc:  # noqa: BLE001 - cosmetic feature, never fatal
+        logger.debug("pet frame decode failed (%s, %s): %s", sheet_path, state_value, exc)
+        return ()
+
+
+@lru_cache(maxsize=8)
+def _frames_for(
+    sheet_path: str,
+    state_value: str,
+    frame_w: int,
+    frame_h: int,
+    frames_per_state: int,
+    scale_w: int,
+    scale_h: int,
+):
+    """Return padding-trimmed RGBA frames for one state row, scaled.
+
+    Thin scaling layer over :func:`_raw_frames`; both are cached so repeated
+    frame requests during animation are free.
+    """
+    raw = _raw_frames(sheet_path, state_value, frame_w, frame_h, frames_per_state)
+    if not raw or (scale_w, scale_h) == (frame_w, frame_h):
+        return list(raw)
+    from PIL import Image
+
+    return [f.resize((scale_w, scale_h), Image.LANCZOS) for f in raw]
+
+
+def state_frame_counts(
+    sheet_path: str | Path,
+    *,
+    frame_w: int = FRAME_W,
+    frame_h: int = FRAME_H,
+    frames_per_state: int = FRAMES_PER_STATE,
+) -> dict[str, int]:
+    """Map each driven :class:`PetState` → its real (padding-trimmed) frame count.
+
+    The single source of truth for "how many frames does this state actually
+    have?".  The CLI/TUI consume the trimmed frame lists directly; the gateway
+    ships this map to the desktop canvas, which steps its own loop.
+    """
+    return {
+        state.value: len(
+            _raw_frames(str(sheet_path), state.value, frame_w, frame_h, frames_per_state)
+        )
+        for state in PetState
+    }
+
+
+# ─────────────────────────────────────────────────────────────────────────
+# Encoders
+# ─────────────────────────────────────────────────────────────────────────
+
+def _png_bytes(frame) -> bytes:
+    buf = io.BytesIO()
+    frame.save(buf, format="PNG")
+    return buf.getvalue()
+
+
+def _kitty_apc(ctrl: str, data: str) -> str:
+    """Emit a kitty APC escape for *data*, chunked into ≤4096-byte ``m`` pieces."""
+    chunk = 4096
+    if len(data) <= chunk:
+        return f"\x1b_G{ctrl},m=0;{data}\x1b\\"
+    out = [f"\x1b_G{ctrl},m=1;{data[:chunk]}\x1b\\"]
+    rest = data[chunk:]
+    while rest:
+        piece, rest = rest[:chunk], rest[chunk:]
+        out.append(f"\x1b_Gm={1 if rest else 0};{piece}\x1b\\")
+    return "".join(out)
+
+
+def _encode_kitty(frame, *, cell_cols: int | None = None, cell_rows: int | None = None) -> str:
+    """Encode one frame via the kitty graphics protocol (transmit + display).
+
+    ``a=T`` transmits & displays at the cursor; ``c``/``r`` request a display
+    box in terminal cells so successive frames overwrite the same area.
+    """
+    ctrl = "f=100,a=T,q=2"
+    if cell_cols:
+        ctrl += f",c={cell_cols}"
+    if cell_rows:
+        ctrl += f",r={cell_rows}"
+    return _kitty_apc(ctrl, base64.standard_b64encode(_png_bytes(frame)).decode("ascii"))
+
+
+# ─────────────────────────────────────────────────────────────────────────
+# kitty Unicode placeholders
+#
+# Ink (the TUI's React-for-terminal layer) owns the screen and measures every
+# cell's width, so it can't host raw kitty image escapes (no width to count,
+# clobbered on the next repaint). kitty's *Unicode placeholder* protocol is the
+# grid-safe path: transmit the image once (q=2, virtual placement U=1), then the
+# host app prints ordinary-width placeholder cells (U+10EEEE + diacritics) whose
+# foreground color encodes the image id. Ink counts those as width-1 text, so
+# layout stays correct and the terminal paints the image underneath.
+#   https://sw.kovidgoyal.net/kitty/graphics-protocol/#unicode-placeholders
+# ─────────────────────────────────────────────────────────────────────────
+
+_KITTY_PLACEHOLDER = "\U0010eeee"
+
+# Row/column diacritics, in order (index → diacritic). Verbatim from kitty's
+# gen/rowcolumn-diacritics.txt (Unicode 6.0.0, combining class 230). Index i is
+# the diacritic that encodes the number i; we only ever need the row index.
+_ROWCOL_DIACRITICS: tuple[int, ...] = (
+    0x0305, 0x030D, 0x030E, 0x0310, 0x0312, 0x033D, 0x033E, 0x033F, 0x0346, 0x034A,
+    0x034B, 0x034C, 0x0350, 0x0351, 0x0352, 0x0357, 0x035B, 0x0363, 0x0364, 0x0365,
+    0x0366, 0x0367, 0x0368, 0x0369, 0x036A, 0x036B, 0x036C, 0x036D, 0x036E, 0x036F,
+    0x0483, 0x0484, 0x0485, 0x0486, 0x0487, 0x0592, 0x0593, 0x0594, 0x0595, 0x0597,
+    0x0598, 0x0599, 0x059C, 0x059D, 0x059E, 0x059F, 0x05A0, 0x05A1, 0x05A8, 0x05A9,
+    0x05AB, 0x05AC, 0x05AF, 0x05C4, 0x0610, 0x0611, 0x0612, 0x0613, 0x0614, 0x0615,
+    0x0616, 0x0617, 0x0657, 0x0658, 0x0659, 0x065A, 0x065B, 0x065D, 0x065E, 0x06D6,
+    0x06D7, 0x06D8, 0x06D9, 0x06DA, 0x06DB, 0x06DC, 0x06DF, 0x06E0, 0x06E1, 0x06E2,
+    0x06E4, 0x06E7, 0x06E8, 0x06EB, 0x06EC, 0x0730, 0x0732, 0x0733, 0x0735, 0x0736,
+    0x073A, 0x073D, 0x073F, 0x0740, 0x0741, 0x0743, 0x0745, 0x0747, 0x0749, 0x074A,
+    0x07EB, 0x07EC, 0x07ED, 0x07EE, 0x07EF, 0x07F0, 0x07F1, 0x07F3, 0x0816, 0x0817,
+    0x0818, 0x0819, 0x081B, 0x081C, 0x081D, 0x081E, 0x081F, 0x0820, 0x0821, 0x0822,
+    0x0823, 0x0825, 0x0826, 0x0827, 0x0829, 0x082A, 0x082B, 0x082C, 0x082D, 0x0951,
+    0x0953, 0x0954, 0x0F82, 0x0F83, 0x0F86, 0x0F87, 0x135D, 0x135E, 0x135F, 0x17DD,
+    0x193A, 0x1A17, 0x1A75, 0x1A76, 0x1A77, 0x1A78, 0x1A79, 0x1A7A, 0x1A7B, 0x1A7C,
+    0x1B6B, 0x1B6D, 0x1B6E, 0x1B6F, 0x1B70, 0x1B71, 0x1B72, 0x1B73, 0x1CD0, 0x1CD1,
+    0x1CD2, 0x1CDA, 0x1CDB, 0x1CE0, 0x1DC0, 0x1DC1, 0x1DC3, 0x1DC4, 0x1DC5, 0x1DC6,
+    0x1DC7, 0x1DC8, 0x1DC9, 0x1DCB, 0x1DCC, 0x1DD1, 0x1DD2, 0x1DD3, 0x1DD4, 0x1DD5,
+    0x1DD6, 0x1DD7, 0x1DD8, 0x1DD9, 0x1DDA, 0x1DDB, 0x1DDC, 0x1DDD, 0x1DDE, 0x1DDF,
+    0x1DE0, 0x1DE1, 0x1DE2, 0x1DE3, 0x1DE4, 0x1DE5, 0x1DE6, 0x1DFE, 0x20D0, 0x20D1,
+    0x20D4, 0x20D5, 0x20D6, 0x20D7, 0x20DB, 0x20DC, 0x20E1, 0x20E7, 0x20E9, 0x20F0,
+    0x2CEF, 0x2CF0, 0x2CF1, 0x2DE0, 0x2DE1, 0x2DE2, 0x2DE3, 0x2DE4, 0x2DE5, 0x2DE6,
+    0x2DE7, 0x2DE8, 0x2DE9, 0x2DEA, 0x2DEB, 0x2DEC, 0x2DED, 0x2DEE, 0x2DEF, 0x2DF0,
+    0x2DF1, 0x2DF2, 0x2DF3, 0x2DF4, 0x2DF5, 0x2DF6, 0x2DF7, 0x2DF8, 0x2DF9, 0x2DFA,
+    0x2DFB, 0x2DFC, 0x2DFD, 0x2DFE, 0x2DFF, 0xA66F, 0xA67C, 0xA67D, 0xA6F0, 0xA6F1,
+    0xA8E0, 0xA8E1, 0xA8E2, 0xA8E3, 0xA8E4, 0xA8E5, 0xA8E6, 0xA8E7, 0xA8E8, 0xA8E9,
+    0xA8EA, 0xA8EB, 0xA8EC, 0xA8ED, 0xA8EE, 0xA8EF, 0xA8F0, 0xA8F1, 0xAAB0, 0xAAB2,
+    0xAAB3, 0xAAB7, 0xAAB8, 0xAABE, 0xAABF, 0xAAC1, 0xFE20, 0xFE21, 0xFE22, 0xFE23,
+    0xFE24, 0xFE25, 0xFE26, 0x10A0F, 0x10A38, 0x1D185, 0x1D186, 0x1D187, 0x1D188,
+    0x1D189, 0x1D1AA, 0x1D1AB, 0x1D1AC, 0x1D1AD, 0x1D242, 0x1D243, 0x1D244,
+)
+
+
+def kitty_image_id(slug: str) -> int:
+    """Stable per-pet image id in ``[1, 0x7FFF]``.
+
+    The id is encoded in the placeholder's 24-bit foreground color, so it must
+    be non-zero and fit comfortably under ``0xFFFFFF``. A small CRC keeps it
+    deterministic per slug (so re-renders reuse the same terminal-side image)
+    while making collisions between two different pets unlikely.
+    """
+    import zlib
+
+    return (zlib.crc32(slug.encode("utf-8")) % 0x7FFE) + 1
+
+
+def kitty_color_hex(image_id: int) -> str:
+    """Hex foreground color (``#rrggbb``) that encodes *image_id* for kitty."""
+    return "#%06x" % (image_id & 0xFFFFFF)
+
+
+def kitty_placeholder_rows(cols: int, rows: int) -> list[str]:
+    """Build the placeholder text grid for an *rows*×*cols* image.
+
+    Each line is one row of the grid: the first cell carries the row diacritic
+    (column defaults to 0), and the remaining ``cols-1`` bare placeholders let
+    the terminal auto-increment the column. The foreground color (the image id)
+    is applied by the caller / Ink, not embedded here.
+    """
+    cols = max(1, cols)
+    out: list[str] = []
+    for r in range(max(1, rows)):
+        idx = min(r, len(_ROWCOL_DIACRITICS) - 1)
+        first = _KITTY_PLACEHOLDER + chr(_ROWCOL_DIACRITICS[idx])
+        out.append(first + _KITTY_PLACEHOLDER * (cols - 1))
+    return out
+
+
+def _encode_kitty_virtual(frame, *, image_id: int, cols: int, rows: int) -> str:
+    """Transmit a frame as a kitty *virtual* placement for Unicode placeholders.
+
+    ``a=T`` transmits and creates the placement in one shot; ``U=1`` marks it
+    virtual (no on-screen output, cursor untouched); ``q=2`` suppresses the
+    terminal's OK/error replies that would otherwise corrupt the host app's
+    output. Re-sending with the same ``i`` replaces the image, so the static
+    placeholder cells animate underneath.
+    """
+    ctrl = f"a=T,U=1,i={image_id},c={cols},r={rows},f=100,q=2"
+    return _kitty_apc(ctrl, base64.standard_b64encode(_png_bytes(frame)).decode("ascii"))
+
+
+def _encode_iterm(frame, *, cell_cols: int | None = None, cell_rows: int | None = None) -> str:
+    """Encode one frame as an iTerm2 inline image (OSC 1337 File)."""
+    payload = base64.standard_b64encode(_png_bytes(frame)).decode("ascii")
+    size = len(payload)
+    args = [f"inline=1", f"size={size}", "preserveAspectRatio=1"]
+    if cell_cols:
+        args.append(f"width={cell_cols}")
+    if cell_rows:
+        args.append(f"height={cell_rows}")
+    return f"\x1b]1337;File={';'.join(args)}:{payload}\x07"
+
+
+def _encode_sixel(frame) -> str:
+    """Encode one frame as DEC sixel.
+
+    Quantizes to an adaptive palette (≤255 colors) and emits the sixel band
+    stream.  Pillow has no sixel writer, so this is a compact hand-rolled
+    encoder.  Transparent pixels render as background (color register skipped).
+    """
+    from PIL import Image
+
+    rgba = frame
+    # Composite onto transparent-as-skip: track alpha to decide background.
+    pal = rgba.convert("RGB").quantize(colors=255, method=Image.MEDIANCUT)
+    palette = pal.getpalette() or []
+    px = pal.load()
+    alpha = rgba.getchannel("A").load()
+    w, h = pal.size
+
+    out = ["\x1bP0;1;0q", '"1;1;%d;%d' % (w, h)]
+    # Color register definitions (sixel uses 0..100 scale).
+    used = sorted({px[x, y] for y in range(h) for x in range(w)})
+    for idx in used:
+        r = palette[idx * 3] if idx * 3 < len(palette) else 0
+        g = palette[idx * 3 + 1] if idx * 3 + 1 < len(palette) else 0
+        b = palette[idx * 3 + 2] if idx * 3 + 2 < len(palette) else 0
+        out.append("#%d;2;%d;%d;%d" % (idx, r * 100 // 255, g * 100 // 255, b * 100 // 255))
+
+    # Emit in 6-row bands.
+    for band in range(0, h, 6):
+        for color_idx in used:
+            line = ["#%d" % color_idx]
+            run_char = None
+            run_len = 0
+
+            def flush():
+                nonlocal run_char, run_len
+                if run_char is None:
+                    return
+                if run_len > 3:
+                    line.append("!%d%s" % (run_len, run_char))
+                else:
+                    line.append(run_char * run_len)
+                run_char, run_len = None, 0
+
+            for x in range(w):
+                bits = 0
+                for bit in range(6):
+                    y = band + bit
+                    if y < h and alpha[x, y] > 32 and px[x, y] == color_idx:
+                        bits |= 1 << bit
+                ch = chr(63 + bits)
+                if ch == run_char:
+                    run_len += 1
+                else:
+                    flush()
+                    run_char, run_len = ch, 1
+            flush()
+            out.append("".join(line) + "$")  # carriage return within band
+        out.append("-")  # next band
+    out.append("\x1b\\")
+    return "".join(out)
+
+
+_HALF_BLOCK = "▀"
+
+# A single half-block cell: top pixel + bottom pixel as (r, g, b, a) tuples.
+Cell = tuple[tuple[int, int, int, int], tuple[int, int, int, int]]
+
+
+def _downscale_cells(frame, *, target_cols: int) -> list[list[Cell]]:
+    """Downscale a frame to a grid of half-block cells.
+
+    Each cell pairs a top and bottom pixel so one terminal row encodes two
+    pixel rows.  Returns rows of ``((tr,tg,tb,ta),(br,bg,bb,ba))`` — the
+    framework-neutral representation shared by the ANSI encoder (CLI) and the
+    structured ``cells`` API (Ink).
+    """
+    from PIL import Image
+
+    target_cols = max(4, target_cols)
+    aspect = frame.height / max(1, frame.width)
+    target_rows = max(2, int(round(target_cols * aspect * 0.5)) * 2)
+    small = frame.resize((target_cols, target_rows), Image.LANCZOS).convert("RGBA")
+    px = small.load()
+
+    grid: list[list[Cell]] = []
+    for y in range(0, target_rows, 2):
+        row: list[Cell] = []
+        for x in range(target_cols):
+            top = px[x, y]
+            bottom = px[x, y + 1] if y + 1 < target_rows else (0, 0, 0, 0)
+            row.append((top, bottom))
+        grid.append(row)
+    return grid
+
+
+def _encode_unicode(frame, *, target_cols: int) -> str:
+    """Downscale to truecolor ANSI half-blocks (one char = 2 vertical pixels)."""
+    lines: list[str] = []
+    for row in _downscale_cells(frame, target_cols=target_cols):
+        cells: list[str] = []
+        for (tr, tg, tb, ta), (br, bg, bb, ba) in row:
+            if ta < 32 and ba < 32:
+                cells.append("\x1b[0m ")  # fully transparent → blank
+                continue
+            cells.append(f"\x1b[38;2;{tr};{tg};{tb}m\x1b[48;2;{br};{bg};{bb}m{_HALF_BLOCK}")
+        lines.append("".join(cells) + "\x1b[0m")
+    return "\n".join(lines)
+
+
+# ─────────────────────────────────────────────────────────────────────────
+# Public renderer
+# ─────────────────────────────────────────────────────────────────────────
+
+class PetRenderer:
+    """Holds a pet's spritesheet and yields encoded frames per (state, index).
+
+    Construct once per pet, then call :meth:`frame` on an animation timer.
+    Cheap to call repeatedly — decoded frames are cached.
+    """
+
+    def __init__(
+        self,
+        spritesheet: str | Path,
+        *,
+        mode: str = "unicode",
+        scale: float = DEFAULT_SCALE,
+        unicode_cols: int = 20,
+        frame_w: int = FRAME_W,
+        frame_h: int = FRAME_H,
+        frames_per_state: int = FRAMES_PER_STATE,
+    ) -> None:
+        self.spritesheet = str(spritesheet)
+        self.mode = mode if mode in RENDER_MODES else "unicode"
+        self.scale = scale
+        self.unicode_cols = unicode_cols
+        self.frame_w = frame_w
+        self.frame_h = frame_h
+        self.frames_per_state = frames_per_state
+
+    @property
+    def available(self) -> bool:
+        return self.mode != "off" and Path(self.spritesheet).is_file()
+
+    def frame_count(self, state: PetState | str) -> int:
+        return len(self._frames(state))
+
+    def _frames(self, state: PetState | str):
+        value = state.value if isinstance(state, PetState) else str(state)
+        scale_w = max(1, int(self.frame_w * self.scale))
+        scale_h = max(1, int(self.frame_h * self.scale))
+        return _frames_for(
+            self.spritesheet,
+            value,
+            self.frame_w,
+            self.frame_h,
+            self.frames_per_state,
+            scale_w,
+            scale_h,
+        )
+
+    def cells(self, state: PetState | str, index: int, *, cols: int | None = None) -> list[list[Cell]]:
+        """Return one frame as a half-block cell grid (framework-neutral).
+
+        Used by the TUI, which renders the grid with native Ink color props
+        instead of raw ANSI.  Returns ``[]`` when no frame is available.
+        """
+        frames = self._frames(state)
+        if not frames:
+            return []
+        frame = frames[index % len(frames)]
+        return _downscale_cells(frame, target_cols=cols or self.unicode_cols)
+
+    def _cell_box(self, frame) -> tuple[int, int]:
+        """Terminal cell box for a scaled frame (~8×16 px per cell).
+
+        Must match :meth:`frame` graphics sizing — kitty stretches the image to
+        fill ``c``×``r`` cells, so these must reflect the scaled pixel
+        dimensions, not a native-aspect column count (that upscales small pets).
+        """
+        return max(1, frame.width // 8), max(1, frame.height // 16)
+
+    def kitty_payload(self, state: PetState | str, *, image_id: int) -> dict | None:
+        """Build the kitty Unicode-placeholder payload for one state.
+
+        Returns ``{cols, rows, placeholder, frames}`` where ``frames`` is a
+        list of transmit escapes (one per animation frame, all reusing
+        ``image_id``) and ``placeholder`` is the static text grid Ink paints.
+        Placement geometry is derived from the scaled frame pixels (via
+        :meth:`_cell_box`), not ``unicode_cols`` — kitty upscales to fill
+        ``c``×``r`` cells. ``None`` when no frame is available.
+        """
+        frames = self._frames(state)
+        if not frames:
+            return None
+        cols, rows = self._cell_box(frames[0])
+        return {
+            "cols": cols,
+            "rows": rows,
+            "placeholder": kitty_placeholder_rows(cols, rows),
+            "frames": [
+                _encode_kitty_virtual(f, image_id=image_id, cols=cols, rows=rows) for f in frames
+            ],
+        }
+
+    def frame(self, state: PetState | str, index: int) -> str:
+        """Return the encoded escape string for one frame, or ``""``.
+
+        ``index`` is taken modulo the available frame count so callers can pass
+        a free-running counter.
+        """
+        if self.mode == "off":
+            return ""
+        frames = self._frames(state)
+        if not frames:
+            return ""
+        frame = frames[index % len(frames)]
+        cell_cols, cell_rows = self._cell_box(frame)
+
+        try:
+            if self.mode == "kitty":
+                return _encode_kitty(frame, cell_cols=cell_cols, cell_rows=cell_rows)
+            if self.mode == "iterm":
+                return _encode_iterm(frame, cell_cols=cell_cols, cell_rows=cell_rows)
+            if self.mode == "sixel":
+                return _encode_sixel(frame)
+            return _encode_unicode(frame, target_cols=self.unicode_cols)
+        except Exception as exc:  # noqa: BLE001 - degrade silently
+            logger.debug("pet frame encode failed (mode=%s): %s", self.mode, exc)
+            return ""
+
+
+def build_renderer(
+    spritesheet: str | Path,
+    *,
+    configured_mode: str | None = None,
+    scale: float = DEFAULT_SCALE,
+    unicode_cols: int = 20,
+    stream=None,
+) -> PetRenderer:
+    """Convenience factory: resolve the mode from config+env, then construct."""
+    mode = resolve_mode(configured_mode, stream=stream)
+    return PetRenderer(
+        spritesheet,
+        mode=mode,
+        scale=scale,
+        unicode_cols=unicode_cols,
+    )

agent/pet/state.py

@@ -0,0 +1,81 @@
+"""Map agent activity → a :class:`PetState`.
+
+This is the one place the "what is the agent doing right now?" → "which
+animation row?" decision lives.  Each surface feeds it the signals it already
+tracks:
+
+- CLI    — ``KawaiiSpinner`` waiting/thinking state + tool outcomes.
+- TUI    — gateway ``tool.start/complete`` + ``message.delta/complete`` events.
+- Desktop — the ``$busy``/``$awaitingResponse``/tool-event nanostores
+            (re-implemented in TS, but mirroring this priority order).
+
+Keeping the priority order here (and documenting it) lets the TypeScript
+mirror stay faithful without a second design.
+"""
+
+from __future__ import annotations
+
+from collections.abc import Iterable
+from typing import Any
+
+from agent.pet.constants import PetState
+
+
+def todos_all_done(todos: Iterable[Any] | None) -> bool:
+    """True iff there's ≥1 todo and every one is completed/cancelled.
+
+    The "celebrate" beat (``JUMP``) fires when a plan finishes; this mirrors
+    the TUI's ``isTodoDone`` so the trigger is defined once across surfaces.
+    Accepts dicts (``{"status": ...}``) or objects with a ``status`` attr.
+    """
+    items = list(todos or [])
+    if not items:
+        return False
+
+    def _status(t: Any) -> Any:
+        return t.get("status") if isinstance(t, dict) else getattr(t, "status", None)
+
+    return all(_status(t) in ("completed", "cancelled") for t in items)
+
+
+def derive_pet_state(
+    *,
+    busy: bool = False,
+    awaiting_input: bool = False,
+    error: bool = False,
+    celebrate: bool = False,
+    just_completed: bool = False,
+    tool_running: bool = False,
+    reasoning: bool = False,
+) -> PetState:
+    """Resolve the animation state from coarse activity signals.
+
+    Priority (highest first) — only one row can show at a time, so the most
+    salient signal wins:
+
+    1. ``error``          → ``FAILED``  (a tool/turn just failed)
+    2. ``celebrate``      → ``JUMP``    (explicit success beat, e.g. todos done)
+    3. ``just_completed`` → ``WAVE``    (turn finished cleanly / greeting)
+    4. ``awaiting_input`` → ``WAITING`` (blocked on the user — a clarify/approval
+       prompt is open; this outranks the in-flight signals below because the turn
+       is paused on *you*, even though a tool is technically mid-call)
+    5. ``tool_running``   → ``RUN``     (a tool is executing)
+    6. ``reasoning``      → ``REVIEW``  (model is thinking / reading)
+    7. ``busy``           → ``RUN``     (turn in flight, unspecified work)
+    8. otherwise          → ``IDLE``
+    """
+    if error:
+        return PetState.FAILED
+    if celebrate:
+        return PetState.JUMP
+    if just_completed:
+        return PetState.WAVE
+    if awaiting_input:
+        return PetState.WAITING
+    if tool_running:
+        return PetState.RUN
+    if reasoning:
+        return PetState.REVIEW
+    if busy:
+        return PetState.RUN
+    return PetState.IDLE

agent/pet/store.py

@@ -0,0 +1,503 @@
+"""On-disk pet store — install / list / resolve pets.
+
+Pets live under ``get_hermes_home()/pets/<slug>/`` so every profile gets its
+own set (we deliberately do **not** reuse petdex's ``~/.codex/pets`` default —
+that's owned by the petdex npm CLI and isn't profile-aware).  Each installed
+pet directory holds:
+
+    pets/<slug>/
+        pet.json            # {id, displayName, description, spritesheetPath}
+        spritesheet.webp    # (or .png)
+
+The active pet is resolved from the caller-supplied ``display.pet.slug`` config
+value (falling back to the first installed pet), so this module stays free of
+the config loader.
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+import re
+from dataclasses import dataclass
+from pathlib import Path
+
+from hermes_constants import get_hermes_home
+
+logger = logging.getLogger(__name__)
+
+_DOWNLOAD_TIMEOUT = 60.0
+
+
+class PetStoreError(RuntimeError):
+    """Raised on install/IO failures."""
+
+
+@dataclass(frozen=True)
+class InstalledPet:
+    """A pet present on disk."""
+
+    slug: str
+    display_name: str
+    description: str
+    directory: Path
+    spritesheet: Path
+    created_by: str = ""  # "generator" for pets hatched locally; "" for petdex installs
+
+    @property
+    def exists(self) -> bool:
+        return self.spritesheet.is_file()
+
+    @property
+    def generated(self) -> bool:
+        return self.created_by == "generator"
+
+
+def pets_dir() -> Path:
+    """Return the profile-scoped pets directory (created on demand)."""
+    path = get_hermes_home() / "pets"
+    path.mkdir(parents=True, exist_ok=True)
+    return path
+
+
+def _read_pet_json(directory: Path) -> dict:
+    pet_json = directory / "pet.json"
+    if not pet_json.is_file():
+        return {}
+    try:
+        return json.loads(pet_json.read_text(encoding="utf-8"))
+    except (OSError, ValueError) as exc:
+        logger.debug("unreadable pet.json in %s: %s", directory, exc)
+        return {}
+
+
+def _resolve_spritesheet(directory: Path, meta: dict) -> Path:
+    """Find the spritesheet for a pet dir.
+
+    Honors ``spritesheetPath`` from pet.json, else probes the conventional
+    filenames (``spritesheet.{webp,png}`` and petdex R2's ``sprite.webp``).
+    """
+    declared = str(meta.get("spritesheetPath", "") or "").strip()
+    if declared:
+        candidate = directory / declared
+        if candidate.is_file():
+            return candidate
+    for name in ("spritesheet.webp", "spritesheet.png", "sprite.webp", "sprite.png"):
+        candidate = directory / name
+        if candidate.is_file():
+            return candidate
+    # Default expectation even if missing, so callers get a stable path.
+    return directory / "spritesheet.webp"
+
+
+def _safe_slug(slug: str) -> str:
+    """Normalize a slug to a single bare path segment.
+
+    Pet slugs index into ``pets_dir()/<slug>/`` for load/remove, so a value
+    carrying path separators (``../``, absolute paths) could escape the pets
+    directory. Strip every separator and reject ``.``/``..`` so callers can
+    only ever name a direct child of the pets directory.
+    """
+    segment = Path(str(slug).strip()).name
+    if segment in ("", ".", ".."):
+        return ""
+    return segment
+
+
+def load_pet(slug: str) -> InstalledPet | None:
+    """Return the :class:`InstalledPet` for *slug*, or ``None`` if absent."""
+    slug = _safe_slug(slug)
+    if not slug:
+        return None
+    directory = pets_dir() / slug
+    if not directory.is_dir():
+        return None
+    meta = _read_pet_json(directory)
+    return InstalledPet(
+        slug=slug,
+        display_name=str(meta.get("displayName", "") or slug),
+        description=str(meta.get("description", "") or ""),
+        directory=directory,
+        spritesheet=_resolve_spritesheet(directory, meta),
+        created_by=str(meta.get("createdBy", "") or ""),
+    )
+
+
+def installed_pets() -> list[InstalledPet]:
+    """Return every installed pet (dirs containing a usable spritesheet)."""
+    out: list[InstalledPet] = []
+    for child in sorted(pets_dir().iterdir()):
+        if not child.is_dir():
+            continue
+        pet = load_pet(child.name)
+        if pet and pet.exists:
+            out.append(pet)
+    return out
+
+
+def resolve_active_pet(configured_slug: str | None = None) -> InstalledPet | None:
+    """Resolve which pet to display.
+
+    Precedence: the configured slug (``display.pet.slug``) if it's installed,
+    otherwise the first installed pet alphabetically, otherwise ``None``.
+    """
+    if configured_slug:
+        pet = load_pet(configured_slug.strip())
+        if pet and pet.exists:
+            return pet
+    pets = installed_pets()
+    return pets[0] if pets else None
+
+
+def install_pet(slug: str, *, force: bool = False, timeout: float = _DOWNLOAD_TIMEOUT) -> InstalledPet:
+    """Download *slug* from the manifest into the pets directory.
+
+    Idempotent: a fully-installed pet is returned as-is unless *force*.  Raises
+    :class:`PetStoreError` / :class:`~agent.pet.manifest.ManifestError` on
+    failure.
+    """
+    from agent.pet.manifest import find_entry
+
+    slug = _safe_slug(slug)
+    if not slug:
+        raise PetStoreError("invalid pet slug")
+    existing = load_pet(slug)
+    if existing and existing.exists and not force:
+        return existing
+
+    entry = find_entry(slug, timeout=timeout)
+    if entry is None:
+        raise PetStoreError(f"pet '{slug}' is not in the petdex manifest")
+
+    # Host-pin every asset URL to petdex. The manifest is trusted (HTTPS from
+    # petdex.dev), but pin the asset hosts too so a compromised/spoofed manifest
+    # can't redirect the download at an arbitrary host. Matches thumbnail_png.
+    if not _is_petdex_host(entry.spritesheet_url):
+        raise PetStoreError(f"refusing non-petdex spritesheet host for '{slug}'")
+
+    directory = pets_dir() / slug
+    directory.mkdir(parents=True, exist_ok=True)
+
+    sprite_ext = ".png" if entry.spritesheet_url.lower().split("?")[0].endswith(".png") else ".webp"
+    sprite_path = directory / f"spritesheet{sprite_ext}"
+
+    _download(entry.spritesheet_url, sprite_path, timeout=timeout)
+
+    # Fetch the upstream pet.json if present; otherwise synthesize a minimal
+    # one so the local layout is self-describing.
+    meta: dict = {}
+    if entry.pet_json_url and _is_petdex_host(entry.pet_json_url):
+        try:
+            meta = _download_json(entry.pet_json_url, timeout=timeout)
+        except Exception as exc:  # noqa: BLE001 - non-fatal, fall back below
+            logger.debug("pet.json fetch failed for %s: %s", slug, exc)
+    if not isinstance(meta, dict) or not meta:
+        meta = {"id": slug, "displayName": entry.display_name, "description": ""}
+    meta["spritesheetPath"] = sprite_path.name
+    meta.setdefault("id", slug)
+    meta.setdefault("displayName", entry.display_name)
+    (directory / "pet.json").write_text(json.dumps(meta, indent=2), encoding="utf-8")
+
+    pet = load_pet(slug)
+    if pet is None or not pet.exists:
+        raise PetStoreError(f"install of '{slug}' did not produce a spritesheet")
+    return pet
+
+
+def slugify(name: str) -> str:
+    """Lowercase, hyphenate, and strip a display name into a filesystem slug."""
+    slug = re.sub(r"[^a-z0-9]+", "-", (name or "").strip().lower()).strip("-")
+    return slug or "pet"
+
+
+def unique_slug(name: str) -> str:
+    """A :func:`slugify` result that doesn't collide with an existing pet dir."""
+    base = slugify(name)
+    slug = base
+    counter = 2
+    while (pets_dir() / slug).exists():
+        slug = f"{base}-{counter}"
+        counter += 1
+    return slug
+
+
+def _write_spritesheet(source, dest: Path) -> None:
+    """Write *source* (PIL image, bytes, or path) as a lossless WebP at *dest*."""
+    if isinstance(source, (bytes, bytearray)):
+        dest.write_bytes(bytes(source))
+        return
+
+    from PIL import Image
+
+    if isinstance(source, (str, Path)):
+        with Image.open(source) as opened:
+            image = opened.convert("RGBA")
+    else:
+        image = source.convert("RGBA")
+    image.save(dest, format="WEBP", lossless=True, quality=100, method=6, exact=True)
+
+
+def register_local_pet(
+    spritesheet,
+    *,
+    slug: str,
+    display_name: str = "",
+    description: str = "",
+) -> InstalledPet:
+    """Write a locally-generated pet into the store and return it.
+
+    *spritesheet* may be a PIL image, raw WebP/PNG bytes, or a path. The pet
+    appears in :func:`installed_pets` immediately, and because :func:`install_pet`
+    returns an already-on-disk pet before consulting the manifest, it can be
+    adopted (``pet.select`` / ``/pet <slug>``) without a manifest entry.
+    """
+    slug = slugify(slug)
+    directory = pets_dir() / slug
+    directory.mkdir(parents=True, exist_ok=True)
+    sprite_path = directory / "spritesheet.webp"
+    try:
+        _write_spritesheet(spritesheet, sprite_path)
+    except Exception as exc:  # noqa: BLE001 - normalize to one error type
+        raise PetStoreError(f"could not write spritesheet for '{slug}': {exc}") from exc
+
+    meta = {
+        "id": slug,
+        "displayName": display_name or slug,
+        "description": description or "",
+        "spritesheetPath": sprite_path.name,
+        "createdBy": "generator",
+    }
+    (directory / "pet.json").write_text(json.dumps(meta, indent=2), encoding="utf-8")
+
+    pet = load_pet(slug)
+    if pet is None or not pet.exists:
+        raise PetStoreError(f"register of generated pet '{slug}' did not produce a spritesheet")
+    return pet
+
+
+def export_pet(slug: str) -> tuple[str, bytes]:
+    """Zip an installed pet's folder (pet.json + spritesheet) → (filename, bytes).
+
+    Dotfiles (cached thumbs, backups) are skipped so the archive is a clean,
+    re-importable pet package. Raises :class:`PetStoreError` if not installed.
+    """
+    import io
+    import zipfile
+
+    root = pets_dir()
+    directory = root / slug.strip()
+    # Guard against traversal: the target must be a direct child of pets_dir.
+    if directory.resolve().parent != root.resolve() or not directory.is_dir():
+        raise PetStoreError(f"pet '{slug}' is not installed")
+
+    name = directory.name
+    buf = io.BytesIO()
+    with zipfile.ZipFile(buf, "w", zipfile.ZIP_DEFLATED) as archive:
+        for path in sorted(directory.iterdir()):
+            if path.is_file() and not path.name.startswith("."):
+                archive.write(path, f"{name}/{path.name}")
+    return f"{name}.zip", buf.getvalue()
+
+
+_THUMB_FRAME_W = 192
+_THUMB_FRAME_H = 208
+_THUMB_W = 96  # rendered ~40px; 2x+ keeps it crisp on HiDPI
+
+
+def _thumbs_dir() -> Path:
+    path = pets_dir() / ".thumbs"
+    path.mkdir(parents=True, exist_ok=True)
+    return path
+
+
+def _is_petdex_host(url: str) -> bool:
+    """True only for petdex.dev hosts — bounds server-side fetch (anti-SSRF)."""
+    from urllib.parse import urlparse
+
+    try:
+        host = (urlparse(url).hostname or "").lower()
+    except ValueError:
+        return False
+    return host == "petdex.dev" or host.endswith(".petdex.dev")
+
+
+def thumbnail_png(slug: str, *, source_url: str = "", timeout: float = 30.0) -> bytes | None:
+    """Return a small idle-frame PNG for *slug*, cached on disk.
+
+    Crops the top-left (idle, frame 0) cell of the spritesheet and downsamples
+    it to a thumbnail. Source preference: an installed spritesheet on disk, else
+    *source_url* — but only when it points at petdex (so the gateway never
+    fetches an arbitrary client-supplied URL). Returns ``None`` when there's no
+    usable source or Pillow/network fails; callers render a placeholder.
+
+    Doing this server-side sidesteps the renderer's CSP / R2 hotlink limits that
+    break a direct ``<img src=cdn>`` and lets the result ride the authenticated
+    gateway as a same-origin data URL.
+    """
+    slug = slug.strip()
+    if not slug:
+        return None
+
+    cache = _thumbs_dir() / f"{slug}.png"
+    if cache.is_file():
+        try:
+            return cache.read_bytes()
+        except OSError:
+            pass
+
+    sheet_bytes: bytes | None = None
+    pet = load_pet(slug)
+    if pet and pet.exists:
+        try:
+            sheet_bytes = pet.spritesheet.read_bytes()
+        except OSError:
+            sheet_bytes = None
+
+    if sheet_bytes is None and source_url and _is_petdex_host(source_url):
+        try:
+            import httpx
+
+            resp = httpx.get(
+                source_url,
+                timeout=timeout,
+                follow_redirects=True,
+                headers={"User-Agent": "hermes-agent-petdex"},
+            )
+            resp.raise_for_status()
+            sheet_bytes = resp.content
+        except Exception as exc:  # noqa: BLE001 - cosmetic, degrade to placeholder
+            logger.debug("thumb fetch failed for %s: %s", slug, exc)
+
+    if not sheet_bytes:
+        return None
+
+    try:
+        import io
+
+        from PIL import Image
+
+        with Image.open(io.BytesIO(sheet_bytes)) as im:
+            frame = im.convert("RGBA").crop(
+                (0, 0, min(_THUMB_FRAME_W, im.width), min(_THUMB_FRAME_H, im.height))
+            )
+            height = round(_THUMB_W * _THUMB_FRAME_H / _THUMB_FRAME_W)
+            frame = frame.resize((_THUMB_W, height), Image.NEAREST)
+            buf = io.BytesIO()
+            frame.save(buf, format="PNG")
+            data = buf.getvalue()
+    except Exception as exc:  # noqa: BLE001
+        logger.debug("thumb crop failed for %s: %s", slug, exc)
+        return None
+
+    try:
+        cache.write_bytes(data)
+    except OSError:
+        pass
+    return data
+
+
+def remove_pet(slug: str) -> bool:
+    """Delete an installed pet directory.  Returns True if anything was removed."""
+    import shutil
+
+    slug = _safe_slug(slug)
+    if not slug:
+        return False
+
+    # The cached thumbnail lives in pets/.thumbs/<slug>.png — OUTSIDE the pet
+    # dir, so rmtree won't catch it. Drop it too, or a later pet that reuses this
+    # slug renders this one's stale thumbnail.
+    try:
+        (_thumbs_dir() / f"{slug}.png").unlink(missing_ok=True)
+    except OSError:
+        pass
+
+    directory = pets_dir() / slug
+    if not directory.is_dir():
+        return False
+    shutil.rmtree(directory, ignore_errors=True)
+    return not directory.exists()
+
+
+def rename_pet(slug: str, display_name: str) -> str | None:
+    """Rename a pet's ``displayName`` AND realign its slug/dir to match.
+
+    Generated pets are hatched under a provisional, prompt-derived slug; when
+    the user names the pet on the reveal screen we make that name the real
+    identity so lists/subtitles show what they typed, not the prompt. The dir is
+    renamed to ``slugify(name)`` (and the cached thumbnail moved alongside it)
+    whenever that yields a free, different slug — otherwise the slug is left as
+    is. Returns the resulting slug on success, or ``None`` on failure.
+    """
+    slug = _safe_slug(slug)
+    display_name = (display_name or "").strip()
+    if not slug or not display_name:
+        return None
+    directory = pets_dir() / slug
+    pet_json = directory / "pet.json"
+    if not pet_json.is_file():
+        return None
+    try:
+        meta = json.loads(pet_json.read_text(encoding="utf-8"))
+    except (OSError, ValueError):
+        meta = {}
+    if not isinstance(meta, dict):
+        meta = {}
+    meta["displayName"] = display_name
+
+    new_slug = slug
+    desired = slugify(display_name)
+    if desired and desired != slug and not (pets_dir() / desired).exists():
+        try:
+            directory.rename(pets_dir() / desired)
+            try:
+                (_thumbs_dir() / f"{slug}.png").rename(_thumbs_dir() / f"{desired}.png")
+            except OSError:
+                pass
+            directory = pets_dir() / desired
+            pet_json = directory / "pet.json"
+            new_slug = desired
+            meta["id"] = new_slug
+        except OSError:
+            new_slug = slug  # keep the provisional slug if the move fails
+
+    try:
+        pet_json.write_text(json.dumps(meta, indent=2), encoding="utf-8")
+    except OSError:
+        return None
+    return new_slug
+
+
+def _download(url: str, dest: Path, *, timeout: float) -> None:
+    import httpx
+
+    try:
+        with httpx.stream(
+            "GET",
+            url,
+            timeout=timeout,
+            follow_redirects=True,
+            headers={"User-Agent": "hermes-agent-petdex"},
+        ) as resp:
+            resp.raise_for_status()
+            tmp = dest.with_suffix(dest.suffix + ".part")
+            with tmp.open("wb") as fh:
+                for chunk in resp.iter_bytes():
+                    fh.write(chunk)
+            tmp.replace(dest)
+    except Exception as exc:  # noqa: BLE001
+        raise PetStoreError(f"download failed for {url}: {exc}") from exc
+
+
+def _download_json(url: str, *, timeout: float) -> dict:
+    import httpx
+
+    resp = httpx.get(
+        url,
+        timeout=timeout,
+        follow_redirects=True,
+        headers={"User-Agent": "hermes-agent-petdex"},
+    )
+    resp.raise_for_status()
+    data = resp.json()
+    return data if isinstance(data, dict) else {}

agent/pet/terminal.py

@@ -0,0 +1,74 @@
+from __future__ import annotations
+import os
+from dataclasses import dataclass, field
+
+
+RENDER_MODES = ("auto", "kitty", "iterm", "sixel", "unicode", "off")
+
+@dataclass(frozen=True)
+class TerminalProfile:
+    """Declarative description of a terminal's detection fingerprint."""
+    mode: str  # one of RENDER_MODES
+    # Any of these env vars being set (non-empty) is a match
+    env_vars: tuple[str, ...] = ()
+    # $TERM_PROGRAM must equal one of these (case-insensitive)
+    term_program_vals: tuple[str, ...] = ()
+    # Any of these strings appearing in $TERM is a match
+    term_substrings: tuple[str, ...] = ()
+
+    def matches(self) -> bool:
+        if any(os.environ.get(v) for v in self.env_vars):
+            return True
+        tp = os.environ.get("TERM_PROGRAM", "").lower()
+        if tp and tp in self.term_program_vals:
+            return True
+        term = os.environ.get("TERM", "").lower()
+        if any(s in term for s in self.term_substrings):
+            return True
+        return False
+
+
+# ── Registry (priority order — first match wins) ───────────────────────────
+#
+# VS Code / Cursor override comes first: their embedded xterm.js leaks
+# KITTY_WINDOW_ID / ITERM_SESSION_ID from the parent shell, so we must
+# short-circuit before any of the real-graphics entries.
+#
+_TERMINAL_PROFILES: tuple[TerminalProfile, ...] = (
+    TerminalProfile(
+        mode="unicode",
+        term_program_vals=("vscode",),
+    ),
+
+    TerminalProfile(
+        mode="kitty",
+        env_vars=("KITTY_WINDOW_ID","GHOSTTY_RESOURCES_DIR","WEZTERM_PANE",),
+        term_program_vals=("kitty", "ghostty",),
+        term_substrings=("kitty","ghostty","wezterm", ),
+    ),
+    TerminalProfile(
+        mode="iterm",
+        env_vars=("ITERM_SESSION_ID",),
+        term_program_vals=("iterm.app",),
+    ),
+    TerminalProfile(
+        mode="sixel",
+        env_vars=("WT_SESSION",),
+        term_substrings=("mlterm","contour","sixel",),
+        term_program_vals=("mintty","foot","contour",),
+    ),
+)
+
+
+def detect_terminal_graphics() -> str:
+    """Return the richest graphics mode available in the current terminal.
+
+    Non-blocking — no DA1/terminal query is issued, as those can break pipes
+    in strange ways.
+    Walks ``_TERMINAL_PROFILES` in priority order; returns the first match's mode, or ``"unicode"`` as the
+    safe universal fallback.
+    """
+    for profile in _TERMINAL_PROFILES:
+        if profile.matches():
+            return profile.mode
+    return "unicode"

agent/prompt_builder.py

@@ -243,7 +243,10 @@ KANBAN_GUIDANCE = (
     "- **Workspace.** `cd $HERMES_KANBAN_WORKSPACE` first. For a `worktree` kind "
     "with no `.git`, `git worktree add <path> "
     "${HERMES_KANBAN_BRANCH:-wt/$HERMES_KANBAN_TASK}` from the main repo, then "
-    "cd there.\n"
+    "cd there. For a project-linked task the workspace is a fresh "
+    "`<repo>/.worktrees/<task-id>` and `$HERMES_KANBAN_BRANCH` a deterministic "
+    "`<project-slug>/<task-id>` — the main repo is two levels up, so run "
+    "`git worktree add` from there.\n"
     "- **Deliverables.** Files a human wants go in "
     "`kanban_complete(artifacts=[<absolute paths>])` (top-level param; paths in "
     "`metadata` are NOT uploaded). Files must exist at completion.\n"
@@ -709,7 +712,24 @@ PLATFORM_HINTS = {
         "(those are only intercepted on messaging platforms like Telegram, "
         "Discord, Slack, etc.; on the CLI they render as literal text). "
         "When referring to a file you created or changed, just state its "
-        "absolute path in plain text; the user can open it from there."
+        "absolute path in plain text; the user can open it from there. "
+        "Cron jobs scheduled from this session are LOCAL-ONLY: their output is "
+        "saved (viewable via cronjob action='list') but is NOT delivered back "
+        "into this terminal — there is no live-delivery channel here. If the "
+        "user wants to be notified when a job runs, the job's `deliver` must "
+        "target a gateway-connected messaging platform (e.g. deliver='telegram' "
+        "or 'all'). Do not promise the user that a deliver='origin' or "
+        "default-deliver cron job will message them in this session."
+    ),
+    "tui": (
+        "You are running in the Hermes terminal UI (TUI). "
+        "Cron jobs scheduled from this session are LOCAL-ONLY: their output is "
+        "saved (viewable via cronjob action='list') but is NOT delivered back "
+        "into this TUI session — there is no live-delivery channel here. If the "
+        "user wants to be notified when a job runs, the job's `deliver` must "
+        "target a gateway-connected messaging platform (e.g. deliver='telegram' "
+        "or 'all'). Do not promise the user that a deliver='origin' or "
+        "default-deliver cron job will message them in this session."
     ),
     "sms": (
         "You are communicating via SMS. Keep responses concise and use plain text "

agent/reasoning_timeouts.py

@@ -0,0 +1,216 @@
+"""Per-reasoning-model stale-timeout floor for known reasoning models.
+
+Reasoning models (those that emit extended thinking blocks before their
+first content token) routinely exceed Hermes's default chat-model
+stale detectors:
+
+* Stream stale detector:   ``HERMES_STREAM_STALE_TIMEOUT``     default 180s
+                           ``agent/chat_completion_helpers.py:2544``
+* Non-stream stale detector: ``HERMES_API_CALL_STALE_TIMEOUT``  default 90s
+                           ``run_agent.py:1140``
+
+For NVIDIA Nemotron 3 Ultra on the hosted NIM gateway the empirical
+upstream idle kill is ~120s (first-party reproduction at
+NVIDIA/NemoClaw#4846 — TTFB ~31s, stream dies at 120s). The same
+failure mode exists on OpenAI o1/o3, Anthropic Opus 4.x thinking,
+DeepSeek R1, Qwen QwQ, xAI Grok reasoning — every cloud reasoning
+model hits upstream-proxies / load-balancers with idle timeouts
+shorter than the model's thinking phase. Result: the stale detector
+kills the connection mid-think, surfacing as
+``BrokenPipeError``/``RemoteProtocolError`` on the next read.
+
+This module provides a floor that the existing stale-detector scaling
+blocks consult via :func:`get_reasoning_stale_timeout_floor` and
+apply as ``max(default, floor)``. It is a FLOOR:
+
+* Never overrides explicit user config (``providers.<id>.models.<model>.stale_timeout_seconds``
+  or ``request_timeout_seconds`` already wins — this code never runs
+  in that branch).
+* Never lowers an existing threshold.
+* Has zero effect on non-reasoning models — they are not in the
+  allowlist and the resolver returns ``None``.
+
+Matching uses start-anchored regex on the slug-only component of
+the model name (after stripping any aggregator prefix like
+``openai/``, ``x-ai/``, ``anthropic/``).  The right-anchor matches
+end-of-string or a ``-``/``.``/``_`` slug separator, so ``qwen3-235b``
+matches the ``qwen3`` family entry (a future model slug would be
+``qwen3-235b-instruct`` and would also match) but ``some-other-qwen3``
+does NOT match ``qwen3`` (the ``-qwen3`` is not at start of slug).
+
+The ``o1`` case is the most delicate: a model named
+``llama-4-70b-o1-preview`` is a hypothetical community derivative that
+should NOT trigger the reasoning-model floor for the user (the user
+chose a non-OpenAI model, not a reasoning model).  The start-of-slug
+anchor naturally excludes this — the matched ``o1-preview`` is at
+position 11 of the slug, not at position 0.  The previous substring-
+with-trailing-hyphen design would have over-matched here, which is
+why start-of-slug anchoring is the right shape.
+
+Fixes #52217.
+"""
+
+from __future__ import annotations
+
+import re
+from typing import Optional
+
+
+# (slug, floor_seconds).  Each slug is matched as a discrete
+# word-boundary component via the wrapper regex in ``_match_any``
+# below.  Order is irrelevant — the first regex match wins.
+_REASONING_STALE_TIMEOUT_FLOORS: tuple[tuple[str, int], ...] = (
+    # NVIDIA Nemotron — reasoning models behind hosted NIM with
+    # documented 60-180s upstream idle kill (NVIDIA/NemoClaw#4846:
+    # 120s measured).
+    ("nemotron-3-ultra", 600),
+    ("nemotron-3-super", 600),
+    ("nemotron-3-nano",  300),
+    # DeepSeek — R1 reasoning model on hosted NIM / DeepSeek direct.
+    ("deepseek-r1", 600),
+    ("deepseek-reasoner", 600),
+    # Qwen — QwQ reasoning + Qwen3 thinking variants.  QwQ-32B
+    # preview is the stable slug; ``qwen3`` covers the family of
+    # thinking-mode Qwen3 models (qwen3-235b-a22b, qwen3-32b, etc.)
+    # without over-matching every Qwen3 instruct variant — the
+    # right-anchor requires the slug to be at the start of the
+    # remaining model name, so ``qwen3-235b-instruct`` (instruct is
+    # NOT a thinking variant) would still match.  Acceptable
+    # trade-off: instruct variants of qwen3 get the 180s floor
+    # even though they don't reason.  The cost is a slightly longer
+    # wait on a hung provider; the alternative (matching only
+    # ``qwen3-.*-thinking``) breaks the moment NVIDIA or Alibaba
+    # ships a slightly different naming shape.
+    ("qwq-32b", 300),
+    ("qwen3", 180),
+    # OpenAI o-series — known multi-minute TTFB.  Each variant
+    # enumerated explicitly so bare ``o1`` doesn't over-match
+    # ``olmo-1`` or hypothetical future community derivatives.
+    ("o1", 600),
+    ("o1-mini", 600),
+    ("o1-pro", 600),
+    ("o1-preview", 600),
+    ("o3", 600),
+    ("o3-pro", 600),
+    ("o3-mini", 300),
+    ("o4-mini", 300),
+    # Anthropic Claude 4.x thinking variants.  Anchored at
+    # ``claude-opus-4`` so non-thinking Claude 3.x or future
+    # non-reasoning Claude variants don't match.
+    ("claude-opus-4", 240),
+    ("claude-sonnet-4.5", 180),
+    ("claude-sonnet-4.6", 180),
+    # xAI Grok reasoning variants.  Explicit reasoning-only keys
+    # plus one for the ``non-reasoning`` variant so users picking
+    # the fast variant don't get the 300s floor.  Bare ``grok-3``,
+    # ``grok-4`` etc. don't match — only the explicit reasoning /
+    # non-reasoning pairs.
+    ("grok-4-fast-reasoning", 300),
+    ("grok-4.20-reasoning", 300),
+    ("grok-4-fast-non-reasoning", 180),
+)
+
+
+# Pre-compile each pattern.  Wrapper = start-of-slug + slug + end-or-
+# separator, where ``start-of-slug`` means start-of-string OR
+# immediately after the last ``/`` (aggregator separator) and
+# ``end-or-separator`` means end-of-string OR a ``-``/``.``/``_``.
+#
+# Why start-of-slug and not start-of-string: aggregator prefixes
+# like ``openai/`` should not affect matching — the slug identity is
+# the part after the last ``/``.  Stripping the aggregator prefix in
+# :func:`get_reasoning_stale_timeout_floor` before regex matching
+# gives the wrapper a clean start-of-string anchor.
+#
+# Why end-or-separator on the right: ``openai/o3-mini`` must match
+# the ``o3-mini`` slug (the right anchor is end-of-string).  And
+# ``openai/o3-mini-2025-01-31`` must also match ``o3-mini`` (the right
+# anchor is the ``-`` separator).  But ``openai/o3-mini-fork`` should
+# NOT match ``o3-mini`` if we wanted to exclude forks — though the
+# pattern ``o3-mini-fork`` would be matched as a derivative anyway,
+# so we accept that community forks inheriting the same prefix are
+# treated as reasoning models (a reasonable default — the upstream
+# gateway timing is the same).
+_PATTERN_CACHE: dict[str, re.Pattern[str]] = {}
+
+
+def _get_pattern(slug: str) -> re.Pattern[str]:
+    compiled = _PATTERN_CACHE.get(slug)
+    if compiled is None:
+        compiled = re.compile(
+            r"^"
+            + re.escape(slug)
+            + r"(?:$|[\-._])"
+        )
+        _PATTERN_CACHE[slug] = compiled
+    return compiled
+
+
+def _match_any(model_lower: str) -> Optional[float]:
+    """Return the floor for the first matching slug, else None.
+
+    Each table entry is matched as a start-of-slug prefix with the
+    slug-separator-or-end-of-string right-anchor.  Table iteration
+    order is irrelevant: longest slug wins (so ``o3-mini`` beats
+    ``o3`` on a model like ``openai/o3-mini``).
+    """
+    # Sort by slug length descending so longer / more-specific slugs
+    # win on shared prefixes (o3-mini beats o3).
+    sorted_floors = sorted(
+        _REASONING_STALE_TIMEOUT_FLOORS, key=lambda kv: -len(kv[0])
+    )
+    for slug, floor in sorted_floors:
+        if _get_pattern(slug).search(model_lower):
+            return float(floor)
+    return None
+
+
+def get_reasoning_stale_timeout_floor(model: object) -> Optional[float]:
+    """Return the stale-timeout floor (seconds) for a known reasoning model.
+
+    Returns ``None`` when the model is not in the allowlist or the
+    argument is empty / not a string.  Matching uses
+    word-boundary-anchored regex on the lowercased model name, so
+    ``openai/o3-mini`` matches the ``o3-mini`` slug but
+    ``olmo-1`` does NOT match ``o1`` (the ``o1`` substring is not
+    at a word boundary inside ``olmo-1``).
+
+    Aggregator prefixes (``openai/``, ``x-ai/``, ``anthropic/`` etc.)
+    are preserved through matching — the ``/`` is itself a word
+    boundary, so ``openai/o3-mini`` matches ``o3-mini`` because the
+    ``/`` before ``o3-mini`` satisfies the left-anchor alternation.
+
+    This is a FLOOR — callers must apply it as ``max(default, floor)``
+    and only when no explicit user-configured per-model
+    ``stale_timeout_seconds`` exists.
+
+    >>> get_reasoning_stale_timeout_floor("nvidia/nemotron-3-ultra-550b-a55b")
+    600.0
+    >>> get_reasoning_stale_timeout_floor("openai/o3-mini")
+    300.0
+    >>> get_reasoning_stale_timeout_floor("deepseek/deepseek-r1")
+    600.0
+    >>> get_reasoning_stale_timeout_floor("qwen/qwen3-235b-a22b-thinking")
+    180.0
+    >>> get_reasoning_stale_timeout_floor("x-ai/grok-4-fast-reasoning")
+    300.0
+    >>> get_reasoning_stale_timeout_floor("anthropic/claude-opus-4-6")
+    240.0
+    >>> get_reasoning_stale_timeout_floor("gpt-4o") is None
+    True
+    >>> get_reasoning_stale_timeout_floor("olmo-1") is None
+    True
+    >>> get_reasoning_stale_timeout_floor(None) is None
+    True
+    """
+    if not model or not isinstance(model, str):
+        return None
+    name = model.strip().lower()
+    if not name:
+        return None
+    # Strip aggregator prefix (everything before and including the
+    # last ``/``).  The wrapper regex anchors at start-of-string, so
+    # the slug identity is the bare model name.
+    if "/" in name:
+        name = name.rsplit("/", 1)[1]
+    return _match_any(name)

agent/retry_utils.py

@@ -8,6 +8,7 @@ rate-limited provider concurrently.
 import random
 import threading
 import time
+from typing import Any
 
 # Monotonic counter for jitter seed uniqueness within the same process.
 # Protected by a lock to avoid race conditions in concurrent retry paths
@@ -15,6 +16,14 @@ import time
 _jitter_counter = 0
 _jitter_lock = threading.Lock()
 
+# Z.AI Coding Plan's GLM-5.2 endpoint often returns HTTP 429 code 1305
+# ("The service may be temporarily overloaded...") for otherwise valid
+# Hermes requests. Short retries tend to hammer the same overloaded window;
+# after a few normal retries, progressively widen the wait window. Keep the
+# cap interactive-friendly: a simple TUI message should fail visibly in minutes,
+# not sit silent for 20+ minutes.
+_ZAI_CODING_OVERLOAD_LONG_BACKOFF = (30.0, 60.0, 90.0, 120.0)
+
 
 def jittered_backoff(
     attempt: int,
@@ -55,3 +64,66 @@ def jittered_backoff(
     jitter = rng.uniform(0, jitter_ratio * delay)
 
     return delay + jitter
+
+
+def _error_text(error: Any) -> str:
+    """Best-effort flattened provider error text for retry classification."""
+    parts = [
+        error,
+        getattr(error, "message", None),
+        getattr(error, "body", None),
+        getattr(error, "response", None),
+    ]
+    return " ".join(str(part) for part in parts if part is not None).lower()
+
+
+def is_zai_coding_overload_error(*, base_url: str | None, model: str | None, error: Any) -> bool:
+    """Return True for Z.AI Coding Plan transient overload 429s.
+
+    The coding-plan endpoint reports overload as HTTP 429 with body code 1305
+    and message "The service may be temporarily overloaded...". Treat only
+    that narrow shape specially so ordinary quota/billing 429s still fail fast
+    through the existing classifier.
+    """
+    base = (base_url or "").lower()
+    model_name = (model or "").lower()
+    status = getattr(error, "status_code", None)
+    text = _error_text(error)
+    return (
+        status == 429
+        and "api.z.ai/api/coding/paas/v4" in base
+        and "glm-5.2" in model_name
+        and ("1305" in text or "temporarily overloaded" in text)
+    )
+
+
+def adaptive_rate_limit_backoff(
+    attempt: int,
+    *,
+    base_url: str | None,
+    model: str | None,
+    error: Any,
+    default_wait: float,
+    short_attempts: int = 3,
+) -> tuple[float, str | None]:
+    """Provider-aware rate-limit backoff.
+
+    For most providers this returns ``default_wait`` unchanged. For Z.AI
+    Coding Plan GLM-5.2 overloads, keep the first ``short_attempts`` retries on
+    the normal short exponential schedule, then switch to progressively longer
+    waits (30s → 60s → 90s → 120s, capped) plus light jitter.
+
+    ``attempt`` is 1-based, matching the retry loop's logged attempt number.
+    Returns ``(wait_seconds, reason_label)`` where ``reason_label`` is suitable
+    for status/log decoration when a provider-specific policy fired.
+    """
+    if not is_zai_coding_overload_error(base_url=base_url, model=model, error=error):
+        return default_wait, None
+    if attempt <= short_attempts:
+        return default_wait, "zai_coding_overload_short"
+
+    idx = min(attempt - short_attempts - 1, len(_ZAI_CODING_OVERLOAD_LONG_BACKOFF) - 1)
+    base_delay = _ZAI_CODING_OVERLOAD_LONG_BACKOFF[idx]
+    # A smaller jitter ratio keeps long waits readable while still avoiding
+    # synchronized retry storms across concurrent Hermes sessions.
+    return jittered_backoff(1, base_delay=base_delay, max_delay=base_delay, jitter_ratio=0.2), "zai_coding_overload_long"

agent/skill_utils.py

@@ -507,6 +507,34 @@ def get_all_skills_dirs() -> List[Path]:
     return dirs
 
 
+def _resolve_for_skill_ownership(path) -> Path:
+    path_obj = path if isinstance(path, Path) else Path(str(path))
+    try:
+        return path_obj.expanduser().resolve()
+    except (OSError, RuntimeError):
+        return path_obj.expanduser().absolute()
+
+
+def is_external_skill_path(path) -> bool:
+    """Return True when ``path`` lives under a configured external skills dir.
+
+    ``skills.external_dirs`` are externally owned: Hermes can discover and view
+    their skills, and foreground user-directed tool calls may still edit them,
+    but autonomous lifecycle maintenance must treat them as read-only. This
+    helper centralizes the ownership boundary so curator/reporting/tool paths do
+    not each need to re-interpret the config.
+    """
+    candidate = _resolve_for_skill_ownership(path)
+    for root in get_external_skills_dirs():
+        resolved_root = _resolve_for_skill_ownership(root)
+        try:
+            candidate.relative_to(resolved_root)
+            return True
+        except ValueError:
+            continue
+    return False
+
+
 # ── Condition extraction ──────────────────────────────────────────────────
 
 

agent/thinking_timeout_guidance.py

@@ -0,0 +1,136 @@
+"""Thinking-timeout detection and user-facing guidance for reasoning models.
+
+When a known reasoning model (NVIDIA Nemotron 3 Ultra, OpenAI o1/o3,
+Anthropic Opus 4.x thinking, DeepSeek R1, Qwen QwQ, xAI Grok reasoning)
+hits a transport-layer error before the first content token arrives, the
+upstream proxy has almost certainly idle-killed a long thinking stream —
+not a true context overflow or a configuration error.  The user needs
+distinct guidance for this case:
+
+    "The model's thinking phase exceeded the upstream proxy's idle
+     timeout before the first content token arrived.  This is a known
+     issue with reasoning models behind cloud gateways (NVIDIA NIM,
+     OpenAI, Anthropic, DeepSeek).  Workarounds in priority order:
+     1. Set `providers.<provider>.models.<model>.stale_timeout_seconds: 900`
+        in `~/.hermes/config.yaml` to extend the per-call timeout...
+     2. Lower `reasoning_budget` or set `reasoning_effort: medium`...
+     3. Use a smaller / faster reasoning model..."
+
+The existing `_is_stream_drop` guidance at
+``agent/conversation_loop.py:3464-3486`` fires for large-file-write
+stream drops ("try execute_code with Python's open() for large files")
+which is the WRONG advice for the thinking-timeout case.  This module
+provides the detection and the message as standalone helpers so the
+detection logic is unit-testable without driving the full retry loop,
+and the message text can be regression-tested for spelling and accuracy.
+
+Part 2 of Fixes #52310.
+"""
+
+from __future__ import annotations
+
+from typing import Optional
+
+
+# Substring set that identifies a transport-layer failure on the
+# response stream.  Same shape as the existing
+# ``_SERVER_DISCONNECT_PATTERNS`` in ``agent/error_classifier.py:394``
+# but extended to also catch the OSS-level error signature
+# (``broken pipe`` / ``errno 32``) that the upstream kill surfaces
+# to the OpenAI SDK wrapper.
+_THINKING_TIMEOUT_SUBSTRINGS: tuple[str, ...] = (
+    "broken pipe",
+    "errno 32",
+    "remote protocol",
+    "connection reset",
+    "connection lost",
+    "peer closed",
+    "server disconnected",
+)
+
+
+def is_thinking_timeout(classified: object, model: str, error_msg: str) -> bool:
+    """Return True when a reasoning model's thinking phase hit a transport kill.
+
+    Args:
+        classified: a :class:`agent.error_classifier.ClassifiedError` instance
+            (duck-typed here to avoid an import cycle in unit tests).
+        model: the model slug at failure time (e.g.
+            ``"nvidia/nemotron-3-ultra-550b-a55b"``).
+        error_msg: lowercased string representation of the underlying
+            exception (typically ``str(api_error).lower()``).
+
+    Returns True when ALL conditions hold:
+        1. ``classified.reason == FailoverReason.timeout`` (the classifier
+           override at ``agent/error_classifier.py:720-738`` ensures this
+           is the case for reasoning models even on large sessions).
+        2. ``api_error`` has no ``.status_code`` attribute set (transport
+           disconnect, not an HTTP error).
+        3. ``model`` is in the reasoning-model allowlist (reuses
+           ``agent.reasoning_timeouts.get_reasoning_stale_timeout_floor``).
+        4. ``error_msg`` contains one of the transport-kill substrings.
+
+    Non-reasoning models always return False.  Non-transport errors
+    (billing / rate_limit / auth / context_overflow / format_error)
+    always return False.  HTTP-status errors always return False.
+    """
+    # Import here (not at module top) to keep this helper cheap to
+    # import even from callers that don't need it.  ``agent.reasoning_timeouts``
+    # is small and dependency-free.
+    from agent.reasoning_timeouts import get_reasoning_stale_timeout_floor
+
+    # Condition 1: classifier says timeout.  Use a string/value check
+    # rather than importing FailoverReason so this module has zero
+    # import cycles from the error_classifier package.
+    reason = getattr(classified, "reason", None)
+    reason_value = getattr(reason, "value", None)
+    if reason_value != "timeout":
+        return False
+
+    # Condition 2: no HTTP status code (transport, not API error).
+    # Caller is expected to gate on ``getattr(api_error, "status_code", None) is None``
+    # before calling this helper; the surface here is just the post-gate
+    # boolean so the caller can pass an already-prepped error_msg.
+
+    # Condition 3: reasoning model allowlist.
+    if get_reasoning_stale_timeout_floor(model) is None:
+        return False
+
+    # Condition 4: transport-kill substring in the error message.
+    error_msg_lower = (error_msg or "").lower()
+    return any(p in error_msg_lower for p in _THINKING_TIMEOUT_SUBSTRINGS)
+
+
+def build_thinking_timeout_guidance(
+    provider: str, model: str, model_label: Optional[str] = None,
+) -> str:
+    """Return the user-facing guidance string appended to ``_final_response``.
+
+    Args:
+        provider: provider slug (e.g. ``"nvidia"``, ``"openai"``).
+        model: bare model slug the user would put in their config
+            (e.g. ``"nemotron-3-ultra-550b-a55b"`` if the user uses
+            NVIDIA direct, or the full ``"nvidia/nemotron-3-ultra-550b-a55b"``
+            if they go through an aggregator).  Used verbatim in the
+            config snippet so the user can copy-paste.
+        model_label: optional short label for the model name in the
+            prose (e.g. ``"Nemotron 3 Ultra"``).  Falls back to the
+            slug if not provided.
+    """
+    label = model_label or model
+    return (
+        "\n\nThe model's thinking phase exceeded the upstream proxy's "
+        "idle timeout before the first content token arrived. This is a "
+        f"known issue with reasoning models (like {label}) behind cloud "
+        "gateways (NVIDIA NIM, OpenAI, Anthropic, DeepSeek). Workarounds "
+        "in priority order:\n"
+        f"1. Set `providers.{provider}.models.{model}.stale_timeout_seconds: 900` "
+        "in `~/.hermes/config.yaml` to extend the per-call timeout. "
+        "(Hermes's built-in floor is 600s for known reasoning models — "
+        "if you still see this after raising, the upstream cap is even "
+        "shorter.)\n"
+        "2. Lower `reasoning_budget` or set `reasoning_effort: medium` on this "
+        "model if the provider supports it.\n"
+        "3. Use a smaller / faster reasoning model if the task doesn't "
+        "require deep thinking."
+    )

agent/tool_dispatch_helpers.py

@@ -11,7 +11,8 @@ Pure module-level utilities extracted from ``run_agent.py``:
   ``_append_subdir_hint_to_multimodal`` — envelope helpers for the
   ``{"_multimodal": True, "content": [...], "text_summary": ...}`` dict
   shape returned by tools like ``computer_use``.
-* ``_extract_file_mutation_targets`` / ``_extract_error_preview`` —
+* ``_extract_file_mutation_targets`` / ``_extract_landed_file_mutation_paths`` /
+  ``_extract_error_preview`` —
   per-turn file-mutation verifier inputs.
 * ``_trajectory_normalize_msg`` — strip image blobs from a message for
   trajectory saving.
@@ -269,6 +270,35 @@ def _extract_file_mutation_targets(tool_name: str, args: Dict[str, Any]) -> List
     return []
 
 
+def _extract_landed_file_mutation_paths(
+    tool_name: str,
+    args: Dict[str, Any],
+    result: Any,
+) -> List[str]:
+    """Return the concrete file paths a successful mutation reports."""
+    targets = _extract_file_mutation_targets(tool_name, args)
+    if tool_name not in _FILE_MUTATING_TOOLS or not isinstance(result, str):
+        return targets
+    try:
+        data = json.loads(result.strip())
+    except Exception:
+        return targets
+    if not isinstance(data, dict):
+        return targets
+
+    files = data.get("files_modified")
+    if isinstance(files, list):
+        landed = [str(p) for p in files if p]
+        if landed:
+            return landed
+
+    resolved = data.get("resolved_path")
+    if resolved:
+        return [str(resolved)]
+
+    return targets
+
+
 def _extract_error_preview(result: Any, max_len: int = 180) -> str:
     """Pull a one-line error summary out of a tool result for footer display."""
     text = _multimodal_text_summary(result) if result is not None else ""
@@ -411,6 +441,7 @@ __all__ = [
     "_multimodal_text_summary",
     "_append_subdir_hint_to_multimodal",
     "_extract_file_mutation_targets",
+    "_extract_landed_file_mutation_paths",
     "_extract_error_preview",
     "_trajectory_normalize_msg",
     "make_tool_result_message",

agent/tool_executor.py

@@ -26,6 +26,7 @@ from agent.display import (
     build_tool_preview as _build_tool_preview,
     get_cute_tool_message as _get_cute_tool_message_impl,
     get_tool_emoji as _get_tool_emoji,
+    redact_tool_args_for_display as _redact_tool_args_for_display,
     _detect_tool_failure,
 )
 from agent.tool_guardrails import ToolGuardrailDecision
@@ -69,12 +70,35 @@ def _budget_for_agent(agent) -> BudgetConfig:
 _MAX_TOOL_WORKERS = 8
 
 
+def _flush_session_db_after_tool_progress(
+    agent,
+    messages: list,
+    *,
+    stage: str,
+) -> None:
+    """Best-effort incremental SessionDB flush for tool-call progress.
+
+    Tool execution can perform side effects that terminate or restart the
+    current Hermes process before the normal turn-end persistence path runs.
+    Flush the already-appended assistant/tool messages immediately so the
+    transcript survives destructive-but-valid tool calls.
+    """
+    try:
+        agent._flush_messages_to_session_db(messages)
+    except Exception as exc:
+        logger.warning("Incremental tool-call persistence failed after %s: %s", stage, exc)
+
+
 def _ra():
     """Lazy reference to ``run_agent`` so patches like ``run_agent._set_interrupt`` work."""
     import run_agent
     return run_agent
 
 
+def _is_interpreter_shutdown_submit_error(exc: RuntimeError) -> bool:
+    return "cannot schedule new futures after interpreter shutdown" in str(exc)
+
+
 def _emit_terminal_post_tool_call(
     agent,
     *,
@@ -279,6 +303,11 @@ def execute_tool_calls_concurrent(agent, assistant_message, messages: list, effe
                 f"[Tool execution cancelled — {tc.function.name} was skipped due to user interrupt]",
                 tc.id,
             ))
+            _flush_session_db_after_tool_progress(
+                agent,
+                messages,
+                stage=f"cancelled tool result {tc.function.name}",
+            )
         return
 
     # ── Parse args + pre-execution bookkeeping ───────────────────────
@@ -441,10 +470,11 @@ def execute_tool_calls_concurrent(agent, assistant_message, messages: list, effe
     if not agent.quiet_mode and getattr(agent, "tool_progress_mode", "all") != "off":
         print(f"  ⚡ Concurrent: {num_tools} tool calls — {tool_names_str}")
         for i, (tc, name, args, middleware_trace, block_result, blocked_by_guardrail) in enumerate(parsed_calls, 1):
-            args_str = json.dumps(args, ensure_ascii=False)
+            display_args = _redact_tool_args_for_display(name, args) or args
+            args_str = json.dumps(display_args, ensure_ascii=False)
             if agent.verbose_logging:
-                print(f"  📞 Tool {i}: {name}({list(args.keys())})")
-                print(agent._wrap_verbose("Args: ", json.dumps(args, indent=2, ensure_ascii=False)))
+                print(f"  📞 Tool {i}: {name}({list(display_args.keys())})")
+                print(agent._wrap_verbose("Args: ", json.dumps(display_args, indent=2, ensure_ascii=False)))
             else:
                 args_preview = args_str[:agent.log_prefix_chars] + "..." if len(args_str) > agent.log_prefix_chars else args_str
                 print(f"  📞 Tool {i}: {name}({list(args.keys())}) - {args_preview}")
@@ -454,8 +484,9 @@ def execute_tool_calls_concurrent(agent, assistant_message, messages: list, effe
             continue
         if agent.tool_progress_callback:
             try:
-                preview = _build_tool_preview(name, args)
-                agent.tool_progress_callback("tool.started", name, preview, args)
+                display_args = _redact_tool_args_for_display(name, args) or args
+                preview = _build_tool_preview(name, display_args)
+                agent.tool_progress_callback("tool.started", name, preview, display_args)
             except Exception as cb_err:
                 logging.debug(f"Tool progress callback error: {cb_err}")
 
@@ -464,7 +495,8 @@ def execute_tool_calls_concurrent(agent, assistant_message, messages: list, effe
             continue
         if agent.tool_start_callback:
             try:
-                agent.tool_start_callback(tc.id, name, args)
+                display_args = _redact_tool_args_for_display(name, args) or args
+                agent.tool_start_callback(tc.id, name, display_args)
             except Exception as cb_err:
                 logging.debug(f"Tool start callback error: {cb_err}")
 
@@ -581,13 +613,40 @@ def execute_tool_calls_concurrent(agent, assistant_message, messages: list, effe
         if runnable_calls:
             max_workers = min(len(runnable_calls), _MAX_TOOL_WORKERS)
             with concurrent.futures.ThreadPoolExecutor(max_workers=max_workers) as executor:
-                for i, tc, name, args in runnable_calls:
+                for submit_index, (i, tc, name, args) in enumerate(runnable_calls):
                     # Propagate the agent turn's ContextVars (e.g.
                     # _approval_session_key) AND thread-local approval/sudo
                     # callbacks into the worker thread; clears callbacks on exit.
-                    f = executor.submit(
-                        propagate_context_to_thread(_run_tool), i, tc, name, args, parsed_calls[i][3]
-                    )
+                    try:
+                        f = executor.submit(
+                            propagate_context_to_thread(_run_tool), i, tc, name, args, parsed_calls[i][3]
+                        )
+                    except RuntimeError as submit_error:
+                        if not _is_interpreter_shutdown_submit_error(submit_error):
+                            raise
+                        skipped_calls = runnable_calls[submit_index:]
+                        logger.warning(
+                            "interpreter shutdown while scheduling concurrent tools; "
+                            "skipping %d unsubmitted tool(s)",
+                            len(skipped_calls),
+                        )
+                        for skipped_i, _tc, skipped_name, skipped_args in skipped_calls:
+                            if results[skipped_i] is None:
+                                middleware_trace = parsed_calls[skipped_i][3]
+                                result = (
+                                    f"Error executing tool '{skipped_name}': "
+                                    "Python interpreter is shutting down; tool was not started"
+                                )
+                                results[skipped_i] = (
+                                    skipped_name,
+                                    skipped_args,
+                                    result,
+                                    0.0,
+                                    True,
+                                    False,
+                                    middleware_trace,
+                                )
+                        break
                     futures.append(f)
 
                 # Wait for all to complete with periodic heartbeats so the
@@ -737,7 +796,8 @@ def execute_tool_calls_concurrent(agent, assistant_message, messages: list, effe
 
         if not blocked and agent.tool_complete_callback:
             try:
-                agent.tool_complete_callback(tc.id, name, args, function_result)
+                display_args = _redact_tool_args_for_display(name, args) or args
+                agent.tool_complete_callback(tc.id, name, display_args, function_result)
             except Exception as cb_err:
                 logging.debug(f"Tool complete callback error: {cb_err}")
 
@@ -768,6 +828,11 @@ def execute_tool_calls_concurrent(agent, assistant_message, messages: list, effe
         # String results pass through unchanged.
         _tool_content = agent._tool_result_content_for_active_model(name, function_result)
         messages.append(make_tool_result_message(name, _tool_content, tc.id))
+        _flush_session_db_after_tool_progress(
+            agent,
+            messages,
+            stage=f"tool result {name}",
+        )
 
         # ── Per-tool /steer drain ───────────────────────────────────
         # Same as the sequential path: drain between each collected
@@ -803,13 +868,16 @@ def execute_tool_calls_sequential(agent, assistant_message, messages: list, effe
                 agent._vprint(f"{agent.log_prefix}⚡ Interrupt: skipping {len(remaining_calls)} tool call(s)", force=True)
             for skipped_tc in remaining_calls:
                 skipped_name = skipped_tc.function.name
-                skip_msg = {
-                    "role": "tool",
-                    "name": skipped_name,
-                    "content": f"[Tool execution cancelled — {skipped_name} was skipped due to user interrupt]",
-                    "tool_call_id": skipped_tc.id,
-                }
-                messages.append(skip_msg)
+                messages.append(make_tool_result_message(
+                    skipped_name,
+                    f"[Tool execution cancelled — {skipped_name} was skipped due to user interrupt]",
+                    skipped_tc.id,
+                ))
+                _flush_session_db_after_tool_progress(
+                    agent,
+                    messages,
+                    stage=f"cancelled tool result {skipped_name}",
+                )
             break
 
         function_name = tool_call.function.name
@@ -891,10 +959,11 @@ def execute_tool_calls_sequential(agent, assistant_message, messages: list, effe
             agent._iters_since_skill = 0
 
         if not agent.quiet_mode and getattr(agent, "tool_progress_mode", "all") != "off":
-            args_str = json.dumps(function_args, ensure_ascii=False)
+            display_args = _redact_tool_args_for_display(function_name, function_args) or function_args
+            args_str = json.dumps(display_args, ensure_ascii=False)
             if agent.verbose_logging:
-                print(f"  📞 Tool {i}: {function_name}({list(function_args.keys())})")
-                print(agent._wrap_verbose("Args: ", json.dumps(function_args, indent=2, ensure_ascii=False)))
+                print(f"  📞 Tool {i}: {function_name}({list(display_args.keys())})")
+                print(agent._wrap_verbose("Args: ", json.dumps(display_args, indent=2, ensure_ascii=False)))
             else:
                 args_preview = args_str[:agent.log_prefix_chars] + "..." if len(args_str) > agent.log_prefix_chars else args_str
                 print(f"  📞 Tool {i}: {function_name}({list(function_args.keys())}) - {args_preview}")
@@ -915,14 +984,16 @@ def execute_tool_calls_sequential(agent, assistant_message, messages: list, effe
 
         if not _execution_blocked and agent.tool_progress_callback:
             try:
-                preview = _build_tool_preview(function_name, function_args)
-                agent.tool_progress_callback("tool.started", function_name, preview, function_args)
+                display_args = _redact_tool_args_for_display(function_name, function_args) or function_args
+                preview = _build_tool_preview(function_name, display_args)
+                agent.tool_progress_callback("tool.started", function_name, preview, display_args)
             except Exception as cb_err:
                 logging.debug(f"Tool progress callback error: {cb_err}")
 
         if not _execution_blocked and agent.tool_start_callback:
             try:
-                agent.tool_start_callback(tool_call.id, function_name, function_args)
+                display_args = _redact_tool_args_for_display(function_name, function_args) or function_args
+                agent.tool_start_callback(tool_call.id, function_name, display_args)
             except Exception as cb_err:
                 logging.debug(f"Tool start callback error: {cb_err}")
 
@@ -1152,7 +1223,8 @@ def execute_tool_calls_sequential(agent, assistant_message, messages: list, effe
             if agent._should_emit_quiet_tool_messages():
                 face = random.choice(KawaiiSpinner.get_waiting_faces())
                 emoji = _get_tool_emoji(function_name)
-                preview = _build_tool_preview(function_name, function_args) or function_name
+                display_args = _redact_tool_args_for_display(function_name, function_args) or function_args
+                preview = _build_tool_preview(function_name, display_args) or function_name
                 spinner = KawaiiSpinner(f"{face} {emoji} {preview}", spinner_type='dots', print_fn=agent._print_fn)
                 spinner.start()
             _ce_result = None
@@ -1185,7 +1257,8 @@ def execute_tool_calls_sequential(agent, assistant_message, messages: list, effe
             if agent._should_emit_quiet_tool_messages() and agent._should_start_quiet_spinner():
                 face = random.choice(KawaiiSpinner.get_waiting_faces())
                 emoji = _get_tool_emoji(function_name)
-                preview = _build_tool_preview(function_name, function_args) or function_name
+                display_args = _redact_tool_args_for_display(function_name, function_args) or function_args
+                preview = _build_tool_preview(function_name, display_args) or function_name
                 spinner = KawaiiSpinner(f"{face} {emoji} {preview}", spinner_type='dots', print_fn=agent._print_fn)
                 spinner.start()
             _mem_result = None
@@ -1216,7 +1289,8 @@ def execute_tool_calls_sequential(agent, assistant_message, messages: list, effe
             if agent._should_emit_quiet_tool_messages() and agent._should_start_quiet_spinner():
                 face = random.choice(KawaiiSpinner.get_waiting_faces())
                 emoji = _get_tool_emoji(function_name)
-                preview = _build_tool_preview(function_name, function_args) or function_name
+                display_args = _redact_tool_args_for_display(function_name, function_args) or function_args
+                preview = _build_tool_preview(function_name, display_args) or function_name
                 spinner = KawaiiSpinner(f"{face} {emoji} {preview}", spinner_type='dots', print_fn=agent._print_fn)
                 spinner.start()
             _spinner_result = None
@@ -1378,7 +1452,8 @@ def execute_tool_calls_sequential(agent, assistant_message, messages: list, effe
 
         if not _execution_blocked and agent.tool_complete_callback:
             try:
-                agent.tool_complete_callback(tool_call.id, function_name, function_args, function_result)
+                display_args = _redact_tool_args_for_display(function_name, function_args) or function_args
+                agent.tool_complete_callback(tool_call.id, function_name, display_args, function_result)
             except Exception as cb_err:
                 logging.debug(f"Tool complete callback error: {cb_err}")
 
@@ -1402,6 +1477,11 @@ def execute_tool_calls_sequential(agent, assistant_message, messages: list, effe
         # (see parallel path for rationale). String results pass through.
         _tool_content = agent._tool_result_content_for_active_model(function_name, function_result)
         messages.append(make_tool_result_message(function_name, _tool_content, tool_call.id))
+        _flush_session_db_after_tool_progress(
+            agent,
+            messages,
+            stage=f"tool result {function_name}",
+        )
 
         # ── Per-tool /steer drain ───────────────────────────────────
         # Drain pending steer BETWEEN individual tool calls so the
@@ -1428,6 +1508,11 @@ def execute_tool_calls_sequential(agent, assistant_message, messages: list, effe
                     f"[Tool execution skipped — {skipped_name} was not started. User sent a new message]",
                     skipped_tc.id,
                 ))
+                _flush_session_db_after_tool_progress(
+                    agent,
+                    messages,
+                    stage=f"skipped tool result {skipped_name}",
+                )
             break
 
         if agent.tool_delay > 0 and i < len(assistant_message.tool_calls):

agent/transports/codex.py

@@ -5,12 +5,47 @@ This transport owns format conversion and normalization — NOT client lifecycle
 streaming, or the _run_codex_stream() call path.
 """
 
+import hashlib
+import json
 from typing import Any, Dict, List, Optional
 
 from agent.transports.base import ProviderTransport
 from agent.transports.types import NormalizedResponse, ToolCall
 
 
+def _content_cache_key(instructions: str, tools: Optional[List[Dict[str, Any]]]) -> Optional[str]:
+    """Content-address the prompt cache key from the static request prefix.
+
+    Returns ``pck_<sha256[:24]>`` of (instructions + sorted tool schemas), or
+    None when there is nothing static to key on. The cache key is a routing
+    hint only — never a correctness boundary — so two requests sharing a system
+    prompt and tool set intentionally resolve to the same warm prefix bucket.
+
+    The fix this exists for: recurring cron jobs build session_id as
+    ``cron_<id>_<timestamp>``, so using session_id as the cache key made every
+    fire cache-cold. The static prefix (identity + tools) is identical across
+    fires, so hashing it gives a stable key that stays warm within the
+    provider's cache TTL. Sorting tools by name keeps the hash insertion-order
+    independent.
+    """
+    if not instructions and not tools:
+        return None
+    tools_part = ""
+    if tools:
+        sorted_tools = sorted(
+            (t for t in tools if isinstance(t, dict)),
+            key=lambda t: str(t.get("name") or t.get("type") or ""),
+        )
+        tools_part = json.dumps(
+            sorted_tools, sort_keys=True, ensure_ascii=False, separators=(",", ":")
+        )
+    # \x00 separator so instructions ending in the tool JSON can't collide with
+    # a request whose instructions contain that JSON and whose tools are empty.
+    content = f"{instructions or ''}\x00{tools_part}"
+    digest = hashlib.sha256(content.encode("utf-8", errors="replace")).hexdigest()[:24]
+    return f"pck_{digest}"
+
+
 class ResponsesApiTransport(ProviderTransport):
     """Transport for api_mode='codex_responses'.
 
@@ -71,7 +106,10 @@ class ResponsesApiTransport(ProviderTransport):
         params:
             instructions: str — system prompt (extracted from messages[0] if not given)
             reasoning_config: dict | None — {effort, enabled}
-            session_id: str | None — used for prompt_cache_key + xAI conv header
+            session_id: str | None — transcript/session id; drives the xAI
+                x-grok-conv-id header and the Codex cache-scope headers, and is
+                the fallback prompt_cache_key when there is no static prefix to
+                content-address
             max_tokens: int | None — max_output_tokens
             timeout: float | None — per-request timeout forwarded to the SDK
             request_overrides: dict | None — extra kwargs merged in
@@ -212,10 +250,17 @@ class ResponsesApiTransport(ProviderTransport):
             kwargs["parallel_tool_calls"] = True
 
         session_id = params.get("session_id")
+        # prompt_cache_key is content-addressed from the static prefix
+        # (instructions + tools), NOT session_id — recurring cron jobs carry a
+        # per-fire timestamp in session_id (cron_<id>_<ts>) that made every run
+        # cache-cold. session_id is left untouched for transcript isolation and
+        # the cache-scope routing headers below. Falls back to session_id when
+        # there is no static content to hash.
+        cache_key = _content_cache_key(instructions, response_tools) or session_id
         # xAI Responses takes prompt_cache_key in extra_body (set further
         # down); GitHub Models opts out of cache-key routing entirely.
-        if not is_github_responses and not is_xai_responses and session_id:
-            kwargs["prompt_cache_key"] = session_id
+        if not is_github_responses and not is_xai_responses and cache_key:
+            kwargs["prompt_cache_key"] = cache_key
 
         if reasoning_enabled and is_xai_responses:
             from agent.model_metadata import grok_supports_reasoning_effort
@@ -326,7 +371,7 @@ class ResponsesApiTransport(ProviderTransport):
             merged_extra_body: Dict[str, Any] = {}
             if isinstance(existing_extra_body, dict):
                 merged_extra_body.update(existing_extra_body)
-            merged_extra_body.setdefault("prompt_cache_key", session_id)
+            merged_extra_body.setdefault("prompt_cache_key", cache_key)
             kwargs["extra_body"] = merged_extra_body
 
         return kwargs

agent/turn_context.py

@@ -29,7 +29,10 @@ from dataclasses import dataclass
 from typing import Any, Dict, List, Optional
 
 from agent.iteration_budget import IterationBudget
-from agent.model_metadata import estimate_request_tokens_rough
+from agent.model_metadata import (
+    estimate_messages_tokens_rough,
+    estimate_request_tokens_rough,
+)
 
 logger = logging.getLogger(__name__)
 
@@ -57,6 +60,34 @@ def _compression_made_progress(
     return orig_tokens > 0 and new_tokens < orig_tokens * 0.95
 
 
+def _should_run_preflight_estimate(
+    messages: List[Dict[str, Any]],
+    protect_first_n: int,
+    protect_last_n: int,
+    threshold_tokens: int,
+) -> bool:
+    """Cheap gate for the (expensive) full preflight token estimate.
+
+    Returns ``True`` when either:
+      (a) message count exceeds the protected ranges (the historical gate), or
+      (b) a cheap char-based estimate already crosses the configured threshold
+          — the few-but-huge case from issue #27405 that the count-only gate
+          would silently skip (a handful of very large messages never trips
+          the count condition, so compression was never attempted and the
+          turn hit a hard context-overflow error).
+
+    Branch (b) uses ``estimate_messages_tokens_rough`` (the shared char-based
+    estimator) so a single large base64 image isn't mistaken for ~250K tokens.
+    It intentionally undercounts vs. the full request estimate — it omits the
+    system prompt and tool schemas — because it is only a *hint* deciding
+    whether to pay for the authoritative ``estimate_request_tokens_rough``,
+    which (together with ``should_compress``) makes the real decision.
+    """
+    if len(messages) > protect_first_n + protect_last_n + 1:
+        return True
+    return estimate_messages_tokens_rough(messages) >= threshold_tokens
+
+
 @dataclass
 class TurnContext:
     """Values produced by the turn prologue and consumed by the turn loop."""
@@ -111,7 +142,13 @@ def build_turn_context(
     # Guard stdio against OSError from broken pipes (systemd/headless/daemon).
     install_safe_stdio()
 
-    agent._ensure_db_session()
+    # NOTE: the DB session row is created later, AFTER the system prompt is
+    # restored/built (see _ensure_db_session() below the system-prompt block).
+    # Creating it here — before _cached_system_prompt is populated — inserts a
+    # row with system_prompt=NULL on a fresh API/gateway agent that carries
+    # client-managed history, which then trips the "stored system prompt is
+    # null; rebuilding from scratch" warning and a needless first-turn prefix
+    # cache miss. (Issue #45499.)
 
     # Tell auxiliary_client what the live main provider/model are for this turn.
     try:
@@ -278,6 +315,11 @@ def build_turn_context(
 
     active_system_prompt = agent._cached_system_prompt
 
+    # Create the DB session row now that _cached_system_prompt is populated, so
+    # the persisted snapshot is written non-NULL on the first turn (Issue
+    # #45499). Idempotent: _ensure_db_session() no-ops once the row exists.
+    agent._ensure_db_session()
+
     # Crash-resilience: persist the inbound user turn as soon as the session row exists.
     try:
         agent._persist_session(messages, conversation_history)
@@ -289,10 +331,14 @@ def build_turn_context(
         )
 
     # ── Preflight context compression ──
-    if (
-        agent.compression_enabled
-        and len(messages) > agent.context_compressor.protect_first_n
-                            + agent.context_compressor.protect_last_n + 1
+    # Gate the (expensive) full token estimate behind a cheap pre-check.
+    # See ``_should_run_preflight_estimate`` for the OR semantics that fix
+    # issue #27405 (a few very large messages slipping past the count gate).
+    if agent.compression_enabled and _should_run_preflight_estimate(
+        messages,
+        agent.context_compressor.protect_first_n,
+        agent.context_compressor.protect_last_n,
+        agent.context_compressor.threshold_tokens,
     ):
         _preflight_tokens = estimate_request_tokens_rough(
             messages,
@@ -392,6 +438,8 @@ def build_turn_context(
 
     # Per-turn file-mutation verifier state.
     agent._turn_failed_file_mutations = {}
+    agent._turn_file_mutation_paths = set()
+    agent._verification_stop_nudges = 0
 
     # Record the execution thread so interrupt()/clear_interrupt() can scope
     # the tool-level interrupt signal to THIS agent's thread only.

agent/turn_finalizer.py

@@ -122,10 +122,14 @@ def finalize_turn(
                 )
 
     # Determine if conversation completed successfully
+    normal_text_response = str(_turn_exit_reason).startswith("text_response(")
     completed = (
         final_response is not None
-        and api_call_count < agent.max_iterations
         and not failed
+        and (
+            api_call_count < agent.max_iterations
+            or normal_text_response
+        )
     )
 
     # Post-loop cleanup must never lose the response.  Trajectory save,
@@ -162,6 +166,25 @@ def finalize_turn(
     # same empty-response loop again.
     try:
         agent._drop_trailing_empty_response_scaffolding(messages)
+
+        # When the turn was interrupted and the last message is a tool
+        # result, append a synthetic assistant message to close the
+        # tool-call sequence. Without this, the session persists a
+        # ``tool → user`` alternation that strict providers (Gemini,
+        # Claude) reject, causing them to hallucinate a continuation of
+        # the user's message on the next turn (#48879).
+        #
+        # ``_drop_trailing_empty_response_scaffolding`` only rewinds the
+        # tool tail when an empty-response scaffolding flag is present; a
+        # clean ``/stop`` interrupt after a successful tool sets no such
+        # flag, so the tool result survives as the tail and we close it
+        # here instead. On an interrupt ``final_response`` is typically
+        # empty, so fall back to an explicit placeholder rather than
+        # persisting an empty-content assistant turn.
+        if interrupted:
+            from agent.message_sanitization import close_interrupted_tool_sequence
+            close_interrupted_tool_sequence(messages, final_response)
+
         agent._persist_session(messages, conversation_history)
     except Exception as _persist_err:
         _cleanup_errors.append(f"persist_session: {_persist_err}")

agent/verification_evidence.py

@@ -0,0 +1,618 @@
+"""Coding verification evidence ledger.
+
+This module records what the agent actually proved while working in a code
+workspace. It is deliberately passive: it never decides to run a suite, never
+blocks completion, and never upgrades targeted checks into "repo green".
+"""
+
+from __future__ import annotations
+
+import json
+import re
+import shlex
+import sqlite3
+import tempfile
+import threading
+from dataclasses import dataclass
+from datetime import datetime, timedelta, timezone
+from pathlib import Path
+from typing import Any, Optional
+
+from hermes_constants import get_hermes_home
+
+
+_DB_LOCK = threading.Lock()
+_MAX_OUTPUT_SUMMARY_CHARS = 2000
+_MAX_EVIDENCE_AGE_DAYS = 30
+_MAX_EVENTS_PER_SESSION_ROOT = 100
+_MAX_TOTAL_UNREFERENCED_EVENTS = 10_000
+_AD_HOC_SCRIPT_NAME_PREFIXES = ("hermes-verify-", "hermes-ad-hoc-")
+_VERIFY_SCHEMA_VERSION = 1
+_SHELL_SPLIT_RE = re.compile(r"\s*(?:&&|\|\||;)\s*")
+
+
+@dataclass(frozen=True)
+class VerificationEvidence:
+    """A classified command result worth recording."""
+
+    command: str
+    canonical_command: str
+    kind: str
+    scope: str
+    status: str
+    exit_code: int
+    cwd: str
+    root: str
+    session_id: str
+    output_summary: str = ""
+
+
+def _utc_now() -> str:
+    return datetime.now(timezone.utc).isoformat()
+
+
+def _retention_cutoff() -> str:
+    return (datetime.now(timezone.utc) - timedelta(days=_MAX_EVIDENCE_AGE_DAYS)).isoformat()
+
+
+def _db_path() -> Path:
+    return get_hermes_home() / "verification_evidence.db"
+
+
+def _connect() -> sqlite3.Connection:
+    path = _db_path()
+    path.parent.mkdir(parents=True, exist_ok=True)
+    conn = sqlite3.connect(path)
+    conn.execute("PRAGMA journal_mode=WAL")
+    conn.execute("PRAGMA busy_timeout=5000")
+    conn.row_factory = sqlite3.Row
+    _ensure_schema(conn)
+    return conn
+
+
+def _ensure_schema(conn: sqlite3.Connection) -> None:
+    conn.execute(
+        """
+        CREATE TABLE IF NOT EXISTS meta (
+            key TEXT PRIMARY KEY,
+            value TEXT NOT NULL
+        )
+        """
+    )
+    conn.execute(
+        """
+        CREATE TABLE IF NOT EXISTS verification_events (
+            id INTEGER PRIMARY KEY AUTOINCREMENT,
+            created_at TEXT NOT NULL,
+            session_id TEXT NOT NULL,
+            cwd TEXT NOT NULL,
+            root TEXT NOT NULL,
+            command TEXT NOT NULL,
+            canonical_command TEXT NOT NULL,
+            kind TEXT NOT NULL,
+            scope TEXT NOT NULL,
+            status TEXT NOT NULL,
+            exit_code INTEGER NOT NULL,
+            output_summary TEXT NOT NULL
+        )
+        """
+    )
+    conn.execute(
+        """
+        CREATE TABLE IF NOT EXISTS verification_state (
+            session_id TEXT NOT NULL,
+            root TEXT NOT NULL,
+            last_event_id INTEGER,
+            last_edit_at TEXT,
+            changed_paths_json TEXT NOT NULL DEFAULT '[]',
+            PRIMARY KEY (session_id, root)
+        )
+        """
+    )
+    conn.execute(
+        """
+        CREATE INDEX IF NOT EXISTS idx_verification_events_session_root
+        ON verification_events(session_id, root, id DESC)
+        """
+    )
+    conn.execute(
+        "INSERT OR REPLACE INTO meta(key, value) VALUES ('schema_version', ?)",
+        (str(_VERIFY_SCHEMA_VERSION),),
+    )
+    conn.commit()
+
+
+def _split_segment_tokens(command: str) -> list[list[str]]:
+    segments: list[list[str]] = []
+    for segment in _SHELL_SPLIT_RE.split(command.strip()):
+        if not segment:
+            continue
+        try:
+            tokens = shlex.split(segment)
+        except ValueError:
+            continue
+        if tokens:
+            segments.append(tokens)
+    return segments
+
+
+def _clean_token(token: str) -> str:
+    token = token.strip()
+    while token.startswith("./"):
+        token = token[2:]
+    return token
+
+
+def _canonical_tokens(canonical: str) -> list[str]:
+    try:
+        return [_clean_token(t) for t in shlex.split(canonical) if t]
+    except ValueError:
+        return []
+
+
+def _find_subsequence(tokens: list[str], needle: list[str]) -> Optional[int]:
+    if not tokens or not needle or len(needle) > len(tokens):
+        return None
+    cleaned = [_clean_token(t) for t in tokens]
+    for idx in range(0, len(cleaned) - len(needle) + 1):
+        if cleaned[idx:idx + len(needle)] == needle:
+            return idx
+    return None
+
+
+def _strip_command_prefix(tokens: list[str]) -> list[str]:
+    """Remove harmless command prefixes before matching canonical commands."""
+    remaining = list(tokens)
+    if remaining and remaining[0] == "env":
+        remaining = remaining[1:]
+    while remaining and "=" in remaining[0] and not remaining[0].startswith("-"):
+        remaining = remaining[1:]
+    while remaining and remaining[0] in {"command", "time", "noglob"}:
+        remaining = remaining[1:]
+    return remaining
+
+
+def _equivalent_needles(needle: list[str]) -> list[list[str]]:
+    """Return command spellings equivalent to the detected canonical command."""
+    candidates = [needle]
+    if len(needle) >= 3 and needle[1] == "run":
+        package_manager = needle[0]
+        script_name = needle[2]
+        if package_manager in {"npm", "pnpm", "yarn", "bun"}:
+            candidates.append([package_manager, script_name])
+    if len(needle) == 1 and "/" in needle[0]:
+        candidates.extend([["bash", needle[0]], ["sh", needle[0]]])
+    if needle == ["pytest"]:
+        candidates.extend(
+            [
+                ["python", "-m", "pytest"],
+                ["python3", "-m", "pytest"],
+                ["uv", "run", "pytest"],
+                ["poetry", "run", "pytest"],
+                ["pipenv", "run", "pytest"],
+            ]
+        )
+    return candidates
+
+
+def _find_canonical_match(command: str, canonical_commands: list[str]) -> Optional[tuple[str, list[str]]]:
+    """Return ``(canonical, trailing_args)`` for the first detected command."""
+
+    segments = _split_segment_tokens(command)
+    for canonical in canonical_commands:
+        needle = _canonical_tokens(canonical)
+        if not needle:
+            continue
+        for tokens in segments:
+            candidate_tokens = _strip_command_prefix(tokens)
+            for candidate in _equivalent_needles(needle):
+                if candidate_tokens[:len(candidate)] == candidate:
+                    return canonical, candidate_tokens[len(candidate):]
+    return None
+
+
+def _kind_for_command(canonical: str) -> str:
+    lowered = canonical.lower()
+    if any(word in lowered for word in ("lint", "eslint", "ruff")):
+        return "lint"
+    if any(word in lowered for word in ("typecheck", "tsc", "mypy", "pyright", "ty")):
+        return "typecheck"
+    if "build" in lowered:
+        return "build"
+    if "fmt" in lowered or "format" in lowered:
+        return "format"
+    if "check" in lowered and "test" not in lowered:
+        return "check"
+    return "test"
+
+
+def _looks_like_target(arg: str) -> bool:
+    if not arg or arg.startswith("-") or "=" in arg:
+        return False
+    return (
+        "/" in arg
+        or "\\" in arg
+        or "::" in arg
+        or arg.endswith((".py", ".js", ".jsx", ".ts", ".tsx", ".rs", ".go", ".java"))
+        or arg.startswith(("test_", "tests", "spec", "__tests__"))
+    )
+
+
+def _scope_for_args(args: list[str]) -> str:
+    return "targeted" if any(_looks_like_target(arg) for arg in args) else "full"
+
+
+def _is_under_temp_dir(token: str) -> bool:
+    if not token or token.startswith("-"):
+        return False
+    try:
+        path = Path(token).expanduser()
+        if not path.is_absolute():
+            return False
+        resolved = path.resolve()
+        temp_root = Path(tempfile.gettempdir()).resolve()
+        return resolved == temp_root or temp_root in resolved.parents
+    except Exception:
+        return False
+
+
+def _is_under_root(token: str, root: str | Path | None) -> bool:
+    if not root:
+        return False
+    try:
+        path = Path(token).expanduser().resolve()
+        root_path = Path(root).expanduser().resolve()
+        return path == root_path or root_path in path.parents
+    except Exception:
+        return False
+
+
+def _is_temp_script_path(token: str, root: str | Path | None) -> bool:
+    try:
+        name = Path(token).expanduser().name
+    except Exception:
+        return False
+    return (
+        name.startswith(_AD_HOC_SCRIPT_NAME_PREFIXES)
+        and _is_under_temp_dir(token)
+        and not _is_under_root(token, root)
+    )
+
+
+def _ad_hoc_script_args(tokens: list[str], root: str | Path | None) -> Optional[list[str]]:
+    candidate_tokens = _strip_command_prefix(tokens)
+    if not candidate_tokens:
+        return None
+    command = candidate_tokens[0]
+    if _is_temp_script_path(command, root):
+        return candidate_tokens[1:]
+    if command in {"python", "python3", "node", "bash", "sh", "ruby", "perl"}:
+        for idx, token in enumerate(candidate_tokens[1:], start=1):
+            if token == "--":
+                continue
+            if _is_temp_script_path(token, root):
+                return candidate_tokens[idx + 1:]
+            if not token.startswith("-"):
+                return None
+    return None
+
+
+def _find_ad_hoc_match(command: str, root: str | Path | None) -> Optional[list[str]]:
+    for tokens in _split_segment_tokens(command):
+        trailing_args = _ad_hoc_script_args(tokens, root)
+        if trailing_args is not None:
+            return trailing_args
+    return None
+
+
+def _summarize_output(output: str) -> str:
+    text = (output or "").strip()
+    if len(text) <= _MAX_OUTPUT_SUMMARY_CHARS:
+        return text
+    head = _MAX_OUTPUT_SUMMARY_CHARS // 3
+    tail = _MAX_OUTPUT_SUMMARY_CHARS - head
+    return (
+        text[:head]
+        + f"\n... [{len(text) - _MAX_OUTPUT_SUMMARY_CHARS} chars omitted] ...\n"
+        + text[-tail:]
+    )
+
+
+def _prune_old_events(conn: sqlite3.Connection, *, session_id: str, root: str) -> None:
+    """Bound ledger growth without deleting the current state pointer."""
+    cutoff = _retention_cutoff()
+    conn.execute(
+        """
+        DELETE FROM verification_events
+        WHERE session_id = ?
+          AND root = ?
+          AND id NOT IN (
+              SELECT id FROM verification_events
+              WHERE session_id = ? AND root = ?
+              ORDER BY id DESC
+              LIMIT ?
+          )
+        """,
+        (session_id, root, session_id, root, _MAX_EVENTS_PER_SESSION_ROOT),
+    )
+    conn.execute(
+        """
+        DELETE FROM verification_state
+        WHERE (
+            last_edit_at IS NOT NULL
+            AND last_edit_at < ?
+        )
+        OR (
+            last_edit_at IS NULL
+            AND last_event_id IN (
+                SELECT id FROM verification_events
+                WHERE created_at < ?
+            )
+        )
+        """,
+        (cutoff, cutoff),
+    )
+    conn.execute(
+        """
+        DELETE FROM verification_events
+        WHERE created_at < ?
+          AND id NOT IN (
+              SELECT last_event_id FROM verification_state
+              WHERE last_event_id IS NOT NULL
+          )
+        """,
+        (cutoff,),
+    )
+    conn.execute(
+        """
+        DELETE FROM verification_events
+        WHERE id NOT IN (
+            SELECT id FROM verification_events
+            ORDER BY id DESC
+            LIMIT ?
+        )
+          AND id NOT IN (
+              SELECT last_event_id FROM verification_state
+              WHERE last_event_id IS NOT NULL
+          )
+        """,
+        (_MAX_TOTAL_UNREFERENCED_EVENTS,),
+    )
+
+
+def classify_verification_command(
+    command: str,
+    *,
+    cwd: str | Path | None = None,
+    session_id: str | None = None,
+    exit_code: int = 0,
+    output: str = "",
+) -> Optional[VerificationEvidence]:
+    """Classify a terminal command as verification evidence, if applicable."""
+
+    if not command or not isinstance(command, str):
+        return None
+    try:
+        from agent.coding_context import project_facts_for
+
+        facts = project_facts_for(cwd)
+    except Exception:
+        facts = None
+    if not facts:
+        return None
+
+    verify_commands = list(facts.get("verifyCommands") or [])
+    match = _find_canonical_match(command, verify_commands)
+    is_ad_hoc = False
+    if match is None and not verify_commands:
+        ad_hoc_args = _find_ad_hoc_match(command, facts.get("root"))
+        if ad_hoc_args is not None:
+            match = ("ad-hoc verification script", ad_hoc_args)
+            is_ad_hoc = True
+    if match is None:
+        return None
+
+    canonical, trailing_args = match
+    return VerificationEvidence(
+        command=command,
+        canonical_command=canonical,
+        kind="ad_hoc" if is_ad_hoc else _kind_for_command(canonical),
+        scope="targeted" if is_ad_hoc else _scope_for_args(trailing_args),
+        status="passed" if int(exit_code) == 0 else "failed",
+        exit_code=int(exit_code),
+        cwd=str(Path(cwd or ".").resolve()),
+        root=str(facts.get("root") or Path(cwd or ".").resolve()),
+        session_id=str(session_id or "default"),
+        output_summary=_summarize_output(output),
+    )
+
+
+def record_terminal_result(
+    *,
+    command: str,
+    cwd: str | Path | None,
+    session_id: str | None,
+    exit_code: int,
+    output: str = "",
+) -> Optional[dict[str, Any]]:
+    """Record a foreground terminal result when it is verification evidence."""
+
+    evidence = classify_verification_command(
+        command,
+        cwd=cwd,
+        session_id=session_id,
+        exit_code=exit_code,
+        output=output,
+    )
+    if evidence is None:
+        return None
+
+    created_at = _utc_now()
+    with _DB_LOCK:
+        with _connect() as conn:
+            cur = conn.execute(
+                """
+                INSERT INTO verification_events(
+                    created_at, session_id, cwd, root, command, canonical_command,
+                    kind, scope, status, exit_code, output_summary
+                ) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
+                """,
+                (
+                    created_at,
+                    evidence.session_id,
+                    evidence.cwd,
+                    evidence.root,
+                    evidence.command,
+                    evidence.canonical_command,
+                    evidence.kind,
+                    evidence.scope,
+                    evidence.status,
+                    evidence.exit_code,
+                    evidence.output_summary,
+                ),
+            )
+            if cur.lastrowid is None:
+                raise RuntimeError("verification event insert did not return an id")
+            event_id = int(cur.lastrowid)
+            conn.execute(
+                """
+                INSERT INTO verification_state(
+                    session_id, root, last_event_id, last_edit_at, changed_paths_json
+                ) VALUES (?, ?, ?, NULL, '[]')
+                ON CONFLICT(session_id, root) DO UPDATE SET
+                    last_event_id = excluded.last_event_id,
+                    last_edit_at = NULL,
+                    changed_paths_json = '[]'
+                """,
+                (evidence.session_id, evidence.root, event_id),
+            )
+            _prune_old_events(conn, session_id=evidence.session_id, root=evidence.root)
+            conn.commit()
+
+    return {"id": event_id, **evidence.__dict__, "created_at": created_at}
+
+
+def mark_workspace_edited(
+    *,
+    session_id: str | None,
+    cwd: str | Path | None,
+    paths: list[str] | tuple[str, ...] | None = None,
+) -> Optional[dict[str, Any]]:
+    """Mark verification evidence stale after a successful file edit."""
+
+    try:
+        from agent.coding_context import project_facts_for
+
+        facts = project_facts_for(cwd)
+    except Exception:
+        facts = None
+    if not facts:
+        return None
+
+    sid = str(session_id or "default")
+    root = str(facts.get("root") or Path(cwd or ".").resolve())
+    changed_paths = sorted({str(p) for p in (paths or []) if p})
+    edited_at = _utc_now()
+
+    with _DB_LOCK:
+        with _connect() as conn:
+            row = conn.execute(
+                """
+                SELECT changed_paths_json FROM verification_state
+                WHERE session_id = ? AND root = ?
+                """,
+                (sid, root),
+            ).fetchone()
+            existing: set[str] = set()
+            if row is not None:
+                try:
+                    existing = set(json.loads(row["changed_paths_json"] or "[]"))
+                except (TypeError, ValueError):
+                    existing = set()
+            merged = sorted((existing | set(changed_paths)))[-200:]
+            conn.execute(
+                """
+                INSERT INTO verification_state(
+                    session_id, root, last_event_id, last_edit_at, changed_paths_json
+                ) VALUES (?, ?, NULL, ?, ?)
+                ON CONFLICT(session_id, root) DO UPDATE SET
+                    last_edit_at = excluded.last_edit_at,
+                    changed_paths_json = excluded.changed_paths_json
+                """,
+                (sid, root, edited_at, json.dumps(merged)),
+            )
+            conn.commit()
+
+    return {"session_id": sid, "root": root, "last_edit_at": edited_at, "changed_paths": changed_paths}
+
+
+def verification_status(
+    *,
+    session_id: str | None,
+    cwd: str | Path | None,
+) -> dict[str, Any]:
+    """Return the best known verification state for a session/workspace."""
+
+    try:
+        from agent.coding_context import project_facts_for
+
+        facts = project_facts_for(cwd)
+    except Exception:
+        facts = None
+    if not facts:
+        return {"status": "not_applicable", "evidence": None}
+
+    sid = str(session_id or "default")
+    root = str(facts.get("root") or Path(cwd or ".").resolve())
+    with _DB_LOCK:
+        with _connect() as conn:
+            state = conn.execute(
+                """
+                SELECT last_event_id, last_edit_at, changed_paths_json
+                FROM verification_state
+                WHERE session_id = ? AND root = ?
+                """,
+                (sid, root),
+            ).fetchone()
+            if state is None:
+                return {
+                    "status": "unverified",
+                    "evidence": None,
+                    "root": root,
+                    "session_id": sid,
+                    "changed_paths": [],
+                }
+            event = None
+            if state["last_event_id"] is not None:
+                event = conn.execute(
+                    "SELECT * FROM verification_events WHERE id = ?",
+                    (state["last_event_id"],),
+                ).fetchone()
+
+    changed_paths: list[str] = []
+    try:
+        changed_paths = json.loads(state["changed_paths_json"] or "[]")
+    except (TypeError, ValueError):
+        changed_paths = []
+
+    if event is None:
+        return {
+            "status": "unverified",
+            "evidence": None,
+            "root": root,
+            "session_id": sid,
+            "changed_paths": changed_paths,
+        }
+
+    evidence = dict(event)
+    if state["last_edit_at"] and state["last_edit_at"] > evidence["created_at"]:
+        status = "stale"
+    else:
+        status = evidence["status"]
+    return {
+        "status": status,
+        "evidence": evidence,
+        "root": root,
+        "session_id": sid,
+        "changed_paths": changed_paths,
+    }

agent/verification_stop.py

@@ -0,0 +1,240 @@
+"""Turn-end verification guard for coding edits.
+
+This module is intentionally policy-only. It never runs checks itself; it turns
+the passive verification ledger into a bounded follow-up when the model tries to
+finish immediately after editing code without fresh evidence.
+"""
+
+from __future__ import annotations
+
+import os
+import tempfile
+from pathlib import Path
+from typing import Any, Iterable
+
+
+_MAX_CHANGED_PATHS_IN_NUDGE = 8
+
+# Session identities (platform or source) that are NOT human conversational
+# messaging surfaces: interactive coding surfaces (CLI, TUI, desktop, codex,
+# local, gateway) and programmatic callers (API server, webhooks, tools).
+# Verify-on-stop stays ON by default for these. Any other resolved gateway
+# platform is a conversational messaging surface (Telegram, Discord, WhatsApp,
+# Signal, Slack, etc.) where the verification narrative would reach a human as
+# chat noise, so it defaults OFF. Mirrors LOCAL_SESSION_SOURCE_IDS in
+# apps/desktop/src/lib/session-source.ts; keep roughly in sync when adding a
+# local or programmatic surface. Default-deny by design: an unrecognized
+# identity is treated as messaging (OFF) so a new chat platform never leaks the
+# verification receipt before this set is updated.
+_NON_MESSAGING_SESSION_SURFACES = frozenset(
+    {
+        "",
+        "cli",
+        "codex",
+        "desktop",
+        "gateway",
+        "local",
+        "tui",
+        "tool",
+        "api_server",
+        "webhook",
+        "msgraph_webhook",
+    }
+)
+
+
+def _session_is_messaging_surface() -> bool:
+    """Return whether this turn is delivered over a human messaging channel.
+
+    The gateway binds the platform value (e.g. ``telegram``) to
+    ``HERMES_SESSION_PLATFORM``; the CLI and TUI set ``HERMES_SESSION_SOURCE``
+    (e.g. ``cli``, ``tui``) instead. Both are consulted via the session-context
+    helper (with an ``os.environ`` fallback), alongside the ``HERMES_PLATFORM``
+    override, matching the sibling platform resolution in
+    ``agent/skill_commands.py`` and ``agent/prompt_builder.py``. A turn is a
+    messaging surface when a resolved identity is present and is not a known
+    non-messaging surface.
+    """
+    try:
+        from gateway.session_context import get_session_env
+
+        platform = (
+            os.getenv("HERMES_PLATFORM")
+            or get_session_env("HERMES_SESSION_PLATFORM", "")
+        )
+        source = get_session_env("HERMES_SESSION_SOURCE", "")
+    except Exception:
+        platform = os.getenv("HERMES_PLATFORM", "") or os.environ.get(
+            "HERMES_SESSION_PLATFORM", ""
+        )
+        source = os.environ.get("HERMES_SESSION_SOURCE", "")
+    for identity in (platform, source):
+        identity = str(identity or "").strip().lower()
+        if identity and identity not in _NON_MESSAGING_SESSION_SURFACES:
+            return True
+    return False
+
+
+def verify_on_stop_enabled(config: dict[str, Any] | None = None) -> bool:
+    """Return whether edit -> verify-before-finish behavior is enabled.
+
+    Precedence: an explicit ``HERMES_VERIFY_ON_STOP`` env var wins, then an
+    explicit boolean ``agent.verify_on_stop`` config value, then a surface-aware
+    default. The config default is the sentinel ``"auto"`` (see
+    ``DEFAULT_CONFIG``), which resolves to ON for interactive coding surfaces
+    (CLI, TUI, desktop) and programmatic callers, and OFF for conversational
+    messaging surfaces (Telegram, Discord, etc.) where the verification
+    narrative would otherwise reach a human as chat noise.
+    """
+    env = os.environ.get("HERMES_VERIFY_ON_STOP")
+    if env is not None:
+        return env.strip().lower() not in {"0", "false", "no", "off"}
+    if config is None:
+        try:
+            from hermes_cli.config import load_config
+
+            config = load_config()
+        except Exception:
+            config = {}
+    agent_cfg = (config or {}).get("agent") if isinstance(config, dict) else None
+    cfg_val = agent_cfg.get("verify_on_stop") if isinstance(agent_cfg, dict) else None
+    if isinstance(cfg_val, bool):
+        return cfg_val
+    if isinstance(cfg_val, str):
+        token = cfg_val.strip().lower()
+        if token in {"1", "true", "yes", "on"}:
+            return True
+        if token in {"0", "false", "no", "off"}:
+            return False
+    # "auto", missing, or any other value -> surface-aware default.
+    return not _session_is_messaging_surface()
+
+
+def _candidate_cwds(paths: Iterable[str]) -> list[Path]:
+    candidates: list[Path] = []
+    seen: set[str] = set()
+    for raw in paths:
+        if not raw:
+            continue
+        try:
+            path = Path(raw).expanduser()
+            candidate = path if path.is_dir() else path.parent
+            resolved = str(candidate.resolve())
+        except Exception:
+            continue
+        if resolved not in seen:
+            seen.add(resolved)
+            candidates.append(Path(resolved))
+    return candidates
+
+
+def _verification_snapshot(
+    *,
+    session_id: str | None,
+    changed_paths: list[str],
+) -> tuple[dict[str, Any], dict[str, Any]] | None:
+    """Return ``(status, facts)`` for the first edited workspace needing proof."""
+    try:
+        from agent.coding_context import project_facts_for
+        from agent.verification_evidence import verification_status
+    except Exception:
+        return None
+
+    first_snapshot: tuple[dict[str, Any], dict[str, Any]] | None = None
+    for cwd in _candidate_cwds(changed_paths):
+        facts = project_facts_for(cwd)
+        if not facts:
+            continue
+        status = verification_status(session_id=session_id, cwd=cwd)
+        snapshot = (status, facts)
+        if first_snapshot is None:
+            first_snapshot = snapshot
+        if str(status.get("status") or "unverified") != "passed":
+            return snapshot
+    return first_snapshot
+
+
+def _format_changed_paths(paths: list[str]) -> str:
+    shown = paths[:_MAX_CHANGED_PATHS_IN_NUDGE]
+    lines = [f"- `{path}`" for path in shown]
+    remaining = len(paths) - len(shown)
+    if remaining > 0:
+        lines.append(f"- ... and {remaining} more")
+    return "\n".join(lines)
+
+
+def _status_detail(status: dict[str, Any]) -> str:
+    state = str(status.get("status") or "unverified")
+    evidence = status.get("evidence") if isinstance(status.get("evidence"), dict) else None
+    if not evidence:
+        return state
+
+    command = evidence.get("canonical_command") or evidence.get("command")
+    summary = str(evidence.get("output_summary") or "").strip()
+    parts = [state]
+    if command:
+        parts.append(f"last command `{command}`")
+    if summary:
+        max_summary = 1200
+        if len(summary) > max_summary:
+            summary = summary[:max_summary].rstrip() + "\n... [truncated]"
+        parts.append(f"last output:\n{summary}")
+    return "\n".join(parts)
+
+
+def build_verify_on_stop_nudge(
+    *,
+    session_id: str | None,
+    changed_paths: Iterable[str],
+    attempts: int = 0,
+    max_attempts: int = 2,
+) -> str | None:
+    """Return a synthetic follow-up when edited code lacks fresh verification."""
+    paths = sorted({str(p) for p in changed_paths if p})
+    if not paths or attempts >= max_attempts:
+        return None
+
+    snapshot = _verification_snapshot(session_id=session_id, changed_paths=paths)
+    if snapshot is None:
+        return None
+    status, facts = snapshot
+
+    verify_commands = [
+        str(cmd).strip()
+        for cmd in (facts.get("verifyCommands") or [])
+        if str(cmd).strip()
+    ]
+
+    state = str(status.get("status") or "unverified")
+    if state == "passed":
+        return None
+
+    if verify_commands:
+        command_instruction = (
+            "Run the relevant verification command now ("
+            + ", ".join(f"`{cmd}`" for cmd in verify_commands[:3])
+            + (", ..." if len(verify_commands) > 3 else "")
+            + "), read any failure, repair the code, and summarize what passed."
+        )
+    else:
+        temp_dir = tempfile.gettempdir()
+        command_instruction = (
+            "No canonical test/lint/build command was detected. Create a focused "
+            f"temporary verification script under `{temp_dir}` using an OS-safe "
+            "`tempfile` path with a `hermes-verify-` filename prefix, run it "
+            "against the changed behavior, clean it up when possible, and "
+            "summarize it explicitly as ad-hoc verification rather than suite "
+            "green."
+        )
+
+    return (
+        "[System: You edited code in this turn, but the workspace does not have "
+        "fresh passing verification evidence yet.\n\n"
+        f"Verification status: {_status_detail(status)}\n\n"
+        f"Changed paths:\n{_format_changed_paths(paths)}\n\n"
+        f"{command_instruction} If verification is not possible, explain the "
+        "concrete blocker instead of claiming the work is fully verified.]"
+    )
+
+
+__all__ = ["build_verify_on_stop_nudge", "verify_on_stop_enabled"]

apps/desktop/components.json

@@ -17,5 +17,5 @@
     "lib": "@/lib",
     "hooks": "@/hooks"
   },
-  "iconLibrary": "lucide"
+  "iconLibrary": "tabler"
 }

apps/desktop/electron/backend-env.cjs

@@ -61,10 +61,7 @@ function buildDesktopBackendPath({
   const venvBin = venvRoot ? pathModule.join(venvRoot, platform === 'win32' ? 'Scripts' : 'bin') : null
   const saneEntries = platform === 'win32' ? [] : POSIX_SANE_PATH_ENTRIES
 
-  return appendUniquePathEntries(
-    [hermesNodeBin, venvBin, currentPath, saneEntries],
-    { delimiter }
-  )
+  return appendUniquePathEntries([hermesNodeBin, venvBin, currentPath, saneEntries], { delimiter })
 }
 
 function normalizeHermesHomeRoot(hermesHome, { pathModule = pathModuleForPlatform(process.platform) } = {}) {

apps/desktop/electron/backend-env.test.cjs

@@ -76,10 +76,7 @@ test('normalizeHermesHomeRoot maps profile homes back to the global Hermes root'
     normalizeHermesHomeRoot('C:\\Users\\test\\AppData\\Local\\hermes\\profiles\\oracle', { pathModule: path.win32 }),
     'C:\\Users\\test\\AppData\\Local\\hermes'
   )
-  assert.equal(
-    normalizeHermesHomeRoot('/Users/test/.hermes', { pathModule: path.posix }),
-    '/Users/test/.hermes'
-  )
+  assert.equal(normalizeHermesHomeRoot('/Users/test/.hermes', { pathModule: path.posix }), '/Users/test/.hermes')
 })
 
 test('Windows PATH casing and delimiter are preserved without POSIX sane entries', () => {
@@ -104,8 +101,5 @@ test('Windows PATH casing and delimiter are preserved without POSIX sane entries
 })
 
 test('appendUniquePathEntries drops empty entries and keeps first occurrence', () => {
-  assert.equal(
-    appendUniquePathEntries([':/a::/b', ['/a', '/c']], { delimiter: ':' }),
-    '/a:/b:/c'
-  )
+  assert.equal(appendUniquePathEntries([':/a::/b', ['/a', '/c']], { delimiter: ':' }), '/a:/b:/c')
 })

apps/desktop/electron/backend-ready.cjs

@@ -1,3 +1,5 @@
+const fs = require('node:fs')
+
 const _READY_RE = /^HERMES_DASHBOARD_READY port=(\d+)/m
 
 // The announcement clock starts the instant the backend process is spawned —
@@ -94,9 +96,76 @@ function waitForDashboardPort(child, timeoutMs = resolvePortAnnounceTimeoutMs())
   })
 }
 
+function readDashboardReadyFile(readyFile) {
+  if (!readyFile) return null
+  try {
+    const parsed = JSON.parse(fs.readFileSync(readyFile, 'utf8'))
+    const port = Number(parsed?.port)
+    return Number.isInteger(port) && port > 0 ? port : null
+  } catch {
+    return null
+  }
+}
+
+function waitForDashboardReadyFile(readyFile, child, timeoutMs = resolvePortAnnounceTimeoutMs()) {
+  return new Promise((resolve, reject) => {
+    let done = false
+    let interval = null
+
+    function cleanup() {
+      if (done) return
+      done = true
+      clearTimeout(timer)
+      if (interval) clearInterval(interval)
+      child.off('exit', onExit)
+      child.off('error', onError)
+    }
+
+    function check() {
+      const port = readDashboardReadyFile(readyFile)
+      if (port) {
+        cleanup()
+        resolve(port)
+      }
+    }
+
+    function onExit(code, signal) {
+      cleanup()
+      reject(new Error(`Hermes backend: exited before port announcement (${signal || code})`))
+    }
+
+    function onError(err) {
+      cleanup()
+      reject(err)
+    }
+
+    const timer = setTimeout(() => {
+      cleanup()
+      reject(new Error(`Timed out waiting for Hermes backend port announcement (${timeoutMs}ms)`))
+    }, timeoutMs)
+
+    child.on('exit', onExit)
+    child.on('error', onError)
+    interval = setInterval(check, 50)
+    if (typeof interval.unref === 'function') interval.unref()
+    check()
+  })
+}
+
+function waitForDashboardPortAnnouncement(child, options = {}) {
+  const timeoutMs = options.timeoutMs ?? resolvePortAnnounceTimeoutMs()
+  if (options.readyFile) {
+    return waitForDashboardReadyFile(options.readyFile, child, timeoutMs)
+  }
+  return waitForDashboardPort(child, timeoutMs)
+}
+
 module.exports = {
   waitForDashboardPort,
+  waitForDashboardPortAnnouncement,
+  waitForDashboardReadyFile,
+  readDashboardReadyFile,
   resolvePortAnnounceTimeoutMs,
   DEFAULT_PORT_ANNOUNCE_TIMEOUT_MS,
-  MIN_PORT_ANNOUNCE_TIMEOUT_MS,
+  MIN_PORT_ANNOUNCE_TIMEOUT_MS
 }

apps/desktop/electron/backend-ready.test.cjs

@@ -14,12 +14,18 @@
 const test = require('node:test')
 const assert = require('node:assert/strict')
 const { EventEmitter } = require('node:events')
+const fs = require('node:fs')
+const os = require('node:os')
+const path = require('node:path')
 
 const {
+  readDashboardReadyFile,
   waitForDashboardPort,
+  waitForDashboardPortAnnouncement,
+  waitForDashboardReadyFile,
   resolvePortAnnounceTimeoutMs,
   DEFAULT_PORT_ANNOUNCE_TIMEOUT_MS,
-  MIN_PORT_ANNOUNCE_TIMEOUT_MS,
+  MIN_PORT_ANNOUNCE_TIMEOUT_MS
 } = require('./backend-ready.cjs')
 
 // A minimal stand-in for a spawned child process: an EventEmitter with a
@@ -119,3 +125,75 @@ test('a late announcement after timeout does not throw (listeners torn down)', a
     child.stdout.emit('data', 'HERMES_DASHBOARD_READY port=9999\n')
   })
 })
+
+// ---------------------------------------------------------------------------
+// ready-file port announcement
+// ---------------------------------------------------------------------------
+
+function mkTmpReadyFile() {
+  const dir = fs.mkdtempSync(path.join(os.tmpdir(), 'hermes-ready-test-'))
+  return {
+    dir,
+    file: path.join(dir, 'ready.json'),
+    cleanup: () => fs.rmSync(dir, { recursive: true, force: true })
+  }
+}
+
+test('readDashboardReadyFile returns a valid port from JSON', () => {
+  const tmp = mkTmpReadyFile()
+  try {
+    fs.writeFileSync(tmp.file, JSON.stringify({ port: 4567 }))
+    assert.equal(readDashboardReadyFile(tmp.file), 4567)
+  } finally {
+    tmp.cleanup()
+  }
+})
+
+test('readDashboardReadyFile ignores missing, malformed, or invalid files', () => {
+  const tmp = mkTmpReadyFile()
+  try {
+    assert.equal(readDashboardReadyFile(tmp.file), null)
+    fs.writeFileSync(tmp.file, '{')
+    assert.equal(readDashboardReadyFile(tmp.file), null)
+    fs.writeFileSync(tmp.file, JSON.stringify({ port: 0 }))
+    assert.equal(readDashboardReadyFile(tmp.file), null)
+  } finally {
+    tmp.cleanup()
+  }
+})
+
+test('waitForDashboardReadyFile resolves when the ready file appears', async () => {
+  const tmp = mkTmpReadyFile()
+  const child = makeFakeChild()
+  try {
+    const p = waitForDashboardReadyFile(tmp.file, child, 1000)
+    setTimeout(() => fs.writeFileSync(tmp.file, JSON.stringify({ port: 8765 })), 20)
+    assert.equal(await p, 8765)
+  } finally {
+    tmp.cleanup()
+  }
+})
+
+test('waitForDashboardPortAnnouncement uses ready file when provided', async () => {
+  const tmp = mkTmpReadyFile()
+  const child = makeFakeChild()
+  try {
+    const p = waitForDashboardPortAnnouncement(child, { readyFile: tmp.file, timeoutMs: 1000 })
+    setTimeout(() => fs.writeFileSync(tmp.file, JSON.stringify({ port: 9876 })), 20)
+    assert.equal(await p, 9876)
+  } finally {
+    tmp.cleanup()
+  }
+})
+
+test('waitForDashboardReadyFile rejects when the child exits before file readiness', async () => {
+  const tmp = mkTmpReadyFile()
+  const child = makeFakeChild()
+  try {
+    const p = waitForDashboardReadyFile(tmp.file, child, 1000)
+    child.emit('exit', 1, null)
+    await assert.rejects(p, /exited before port announcement/)
+  } finally {
+    tmp.cleanup()
+  }
+})

apps/desktop/electron/bootstrap-runner.cjs

@@ -179,7 +179,13 @@ function downloadInstallScript(commit, destPath) {
   })
 }
 
-async function resolveInstallScript({ installStamp, sourceRepoRoot, hermesHome, emit, _download = downloadInstallScript }) {
+async function resolveInstallScript({
+  installStamp,
+  sourceRepoRoot,
+  hermesHome,
+  emit,
+  _download = downloadInstallScript
+}) {
   // 1. Dev shortcut: prefer a local checkout's installer so we can iterate
   //    without pushing. SOURCE_REPO_ROOT comes from main.cjs (path.resolve
   //    of APP_ROOT/../..).
@@ -293,15 +299,19 @@ function spawnPowerShell(scriptPath, args, { emit, stageName, abortSignal, herme
     const ps = process.platform === 'win32' ? resolveWindowsPowerShell() : 'pwsh'
     const fullArgs = ['-NoProfile', '-ExecutionPolicy', 'Bypass', '-File', scriptPath, ...args]
 
-    const child = spawn(ps, fullArgs, hiddenWindowsChildOptions({
-      stdio: ['ignore', 'pipe', 'pipe'],
-      env: {
-        ...process.env,
-        // Pass HERMES_HOME through so install.ps1 respects the caller's
-        // choice rather than re-computing the default.
-        HERMES_HOME: hermesHome || process.env.HERMES_HOME || ''
-      }
-    }))
+    const child = spawn(
+      ps,
+      fullArgs,
+      hiddenWindowsChildOptions({
+        stdio: ['ignore', 'pipe', 'pipe'],
+        env: {
+          ...process.env,
+          // Pass HERMES_HOME through so install.ps1 respects the caller's
+          // choice rather than re-computing the default.
+          HERMES_HOME: hermesHome || process.env.HERMES_HOME || ''
+        }
+      })
+    )
 
     let stdout = ''
     let stderr = ''

apps/desktop/electron/connection-config.cjs

@@ -261,12 +261,7 @@ function cookiesHaveSession(cookies) {
  */
 function cookiesHaveLiveSession(cookies) {
   if (!Array.isArray(cookies)) return false
-  return cookies.some(
-    c =>
-      c &&
-      c.value &&
-      (AT_COOKIE_VARIANTS.includes(c.name) || RT_COOKIE_VARIANTS.includes(c.name))
-  )
+  return cookies.some(c => c && c.value && (AT_COOKIE_VARIANTS.includes(c.name) || RT_COOKIE_VARIANTS.includes(c.name)))
 }
 
 module.exports = {

apps/desktop/electron/desktop-uninstall.cjs

@@ -138,10 +138,7 @@ function buildPosixCleanupScript({ desktopPid, pythonExe, pythonPath, agentRoot,
   if (pythonPath) {
     lines.push(`export PYTHONPATH=${q(pythonPath)}\${PYTHONPATH:+:$PYTHONPATH}`)
   }
-  lines.push(
-    `cd ${q(agentRoot)} 2>/dev/null || true`,
-    `${q(pythonExe)} ${uninstallArgs.map(q).join(' ')} || true`
-  )
+  lines.push(`cd ${q(agentRoot)} 2>/dev/null || true`, `${q(pythonExe)} ${uninstallArgs.map(q).join(' ')} || true`)
   if (appPath) {
     lines.push(`rm -rf ${q(appPath)} || true`)
   }
@@ -169,7 +166,15 @@ function buildPosixCleanupScript({ desktopPid, pythonExe, pythonPath, agentRoot,
  * Removal: even after the desktop PID is gone, Windows releases directory
  * handles lazily, so a single `rmdir /s /q` can half-fail — retry up to 10x.
  */
-function buildWindowsCleanupScript({ desktopPid, pythonExe, pythonPath, agentRoot, uninstallArgs, appPath, hermesHome }) {
+function buildWindowsCleanupScript({
+  desktopPid,
+  pythonExe,
+  pythonPath,
+  agentRoot,
+  uninstallArgs,
+  appPath,
+  hermesHome
+}) {
   const pid = Number(desktopPid) || 0
   // cmd.exe has no string escaping inside quotes; strip embedded quotes (paths
   // under %LOCALAPPDATA% never contain them). `&`/`^` in a path would still be

apps/desktop/electron/desktop-uninstall.test.cjs

@@ -101,10 +101,7 @@ test('resolveRemovableAppPath uses APPIMAGE on Linux when set', () => {
 })
 
 test('resolveRemovableAppPath finds the unpacked dir on Linux', () => {
-  assert.equal(
-    resolveRemovableAppPath('/opt/hermes/linux-unpacked/hermes', 'linux', {}),
-    '/opt/hermes/linux-unpacked'
-  )
+  assert.equal(resolveRemovableAppPath('/opt/hermes/linux-unpacked/hermes', 'linux', {}), '/opt/hermes/linux-unpacked')
   // A system-package install (/usr/bin) → null, left to apt/dnf.
   assert.equal(resolveRemovableAppPath('/usr/bin/hermes', 'linux', {}), null)
 })

apps/desktop/electron/embed-referer.cjs

@@ -0,0 +1,48 @@
+'use strict'
+
+const { session } = require('electron')
+
+const EMBED_SESSION_PARTITION = 'persist:hermes-embed'
+const EMBED_REFERER = 'https://www.youtube.com/'
+const YOUTUBE_REFERER_HOST_RE =
+  /(^|\.)(youtube\.com|youtube-nocookie\.com|googlevideo\.com|ytimg\.com|youtubei\.googleapis\.com)$/i
+
+function installEmbedRefererForSession(embedSession) {
+  if (!embedSession) {
+    return
+  }
+
+  embedSession.webRequest.onBeforeSendHeaders((details, callback) => {
+    let host = ''
+
+    try {
+      host = new URL(details.url).hostname
+    } catch {
+      host = ''
+    }
+
+    if (!YOUTUBE_REFERER_HOST_RE.test(host)) {
+      callback({ requestHeaders: details.requestHeaders })
+      return
+    }
+
+    const headers = { ...details.requestHeaders }
+
+    if (!headers.Referer && !headers.referer) {
+      headers.Referer = EMBED_REFERER
+    }
+
+    callback({ requestHeaders: headers })
+  })
+}
+
+/** Stamp Referer on YouTube requests in the embed webview partition only. */
+function installEmbedReferer() {
+  try {
+    installEmbedRefererForSession(session.fromPartition(EMBED_SESSION_PARTITION))
+  } catch {
+    // Non-fatal: embeds still render; YouTube may show referer errors.
+  }
+}
+
+module.exports = { installEmbedReferer }

apps/desktop/electron/fs-read-dir.cjs

@@ -92,9 +92,7 @@ async function readDirForIpc(dirPath, options = {}) {
   try {
     const dirents = await fsImpl.promises.readdir(resolved, { withFileTypes: true })
     const visibleDirents = dirents.filter(dirent => !FS_READDIR_HIDDEN.has(dirent.name))
-    const entries = await mapWithStatConcurrency(visibleDirents, dirent =>
-      entryForDirent(dirent, resolved, fsImpl)
-    )
+    const entries = await mapWithStatConcurrency(visibleDirents, dirent => entryForDirent(dirent, resolved, fsImpl))
 
     entries.sort((a, b) => Number(b.isDirectory) - Number(a.isDirectory) || a.name.localeCompare(b.name))
 

apps/desktop/electron/fs-read-dir.test.cjs

@@ -349,7 +349,10 @@ test('readDirForIpc bounds concurrent stats while preserving complete sorted out
   assert.equal(result.error, undefined)
   assert.equal(result.entries.length, names.length)
   assert.equal(statCalls.length, names.length)
-  assert.equal(statCalls.some(fullPath => fullPath.endsWith(`${path.sep}node_modules`)), false)
+  assert.equal(
+    statCalls.some(fullPath => fullPath.endsWith(`${path.sep}node_modules`)),
+    false
+  )
   assert.ok(peak > 1, `expected concurrent stats, observed peak ${peak}`)
   assert.ok(peak <= 16, `expected at most 16 concurrent stats, observed peak ${peak}`)
   assert.deepEqual(
@@ -357,8 +360,5 @@ test('readDirForIpc bounds concurrent stats while preserving complete sorted out
     expectedNames
   )
   assert.equal(result.entries.find(entry => entry.name === failedName)?.isDirectory, false)
-  assert.equal(
-    result.entries.filter(entry => entry.isDirectory).length,
-    successfulDirectoryNames.size
-  )
+  assert.equal(result.entries.filter(entry => entry.isDirectory).length, successfulDirectoryNames.size)
 })

apps/desktop/electron/git-repo-scan.cjs

@@ -0,0 +1,96 @@
+'use strict'
+
+// Repo-first discovery: walk bounded roots for git repos using only Node's `fs`
+// — no native addon, so it just works for anyone who pulls main (no
+// electron-rebuild). Mirrors how GitHub Desktop scans: stop at the first `.git`
+// (don't descend into a repo), cap depth, and skip heavy non-repo trees so the
+// first scan stays fast. Results are cached by the backend after the first run.
+
+const fs = require('node:fs')
+const os = require('node:os')
+const path = require('node:path')
+
+const fsp = fs.promises
+
+// Shallow on purpose: real projects live a few levels under home
+// (`~/www/repo`, `~/code/org/repo`); deeper `.git` dirs are almost always
+// fixtures/vendored/eval checkouts (e.g. `~/www/ha-evals/tasks/*/repo`). Repos
+// you actually use but keep deeper still surface via session-derived discovery,
+// so this only prunes noise, never repos with history.
+const DEFAULT_MAX_DEPTH = 3
+const MAX_CONCURRENCY = 32
+
+// Big trees that are never themselves repos and would waste the walk. Anything
+// hidden (dotdirs like .cache/.Trash/.npm) is skipped wholesale below, so this
+// only needs the non-hidden heavyweights.
+const JUNK_DIRS = new Set(['Applications', 'Library', 'node_modules', 'site-packages', 'vendor', 'venv'])
+
+async function mapLimit(items, limit, fn) {
+  let cursor = 0
+
+  async function worker() {
+    while (cursor < items.length) {
+      const index = cursor
+      cursor += 1
+      await fn(items[index])
+    }
+  }
+
+  await Promise.all(Array.from({ length: Math.min(limit, items.length) }, worker))
+}
+
+/**
+ * Scan `roots` (default: the home dir) for git repositories. Returns deduped
+ * `{ root, label }` entries. `options.maxDepth` caps recursion (default 3).
+ */
+async function scanGitRepos(roots, options = {}) {
+  const maxDepth = Number(options.maxDepth) || DEFAULT_MAX_DEPTH
+  const searchRoots = Array.isArray(roots) && roots.length > 0 ? roots : [os.homedir()]
+  const found = new Map()
+
+  async function walk(dir, depth) {
+    if (depth > maxDepth) {
+      return
+    }
+
+    let entries
+    try {
+      entries = await fsp.readdir(dir, { withFileTypes: true })
+    } catch {
+      return // unreadable / permission denied
+    }
+
+    // A `.git` DIRECTORY marks a real repo root (a main checkout). A `.git`
+    // FILE is a linked worktree or submodule — those belong to their parent
+    // repo as lanes, not as separate projects, so we don't list them (and we
+    // keep descending in case a real repo sits deeper). This is what kills the
+    // worktree/eval-repo duplicate explosion.
+    if (entries.some(entry => entry.name === '.git' && entry.isDirectory())) {
+      const root = dir.replace(/[/\\]+$/, '')
+      found.set(root, path.basename(root) || root)
+
+      return
+    }
+
+    const subdirs = []
+    for (const entry of entries) {
+      // Real directories only (skip symlinks to avoid loops), no hidden dirs, no
+      // known heavy trees.
+      if (!entry.isDirectory() || entry.name.startsWith('.') || JUNK_DIRS.has(entry.name)) {
+        continue
+      }
+
+      subdirs.push(path.join(dir, entry.name))
+    }
+
+    await mapLimit(subdirs, MAX_CONCURRENCY, sub => walk(sub, depth + 1))
+  }
+
+  await mapLimit(searchRoots.map(root => String(root || '').trim()).filter(Boolean), MAX_CONCURRENCY, root =>
+    walk(root, 0)
+  )
+
+  return [...found.entries()].map(([root, label]) => ({ label, root }))
+}
+
+module.exports = { scanGitRepos }

apps/desktop/electron/git-review-ops.cjs

@@ -0,0 +1,684 @@
+'use strict'
+
+// Git ops backing the coding rail + Codex-style review pane. Built on `simple-git`
+// (a maintained wrapper around the system git binary — same git the rest of the
+// app shells to, no native build) so we read structured status()/diffSummary()
+// results instead of hand-parsing porcelain. Reads degrade to null/empty on a
+// non-repo / remote backend; mutations reject so the renderer can toast.
+
+const { execFile } = require('node:child_process')
+const fs = require('node:fs/promises')
+const path = require('node:path')
+
+const simpleGit = require('simple-git')
+
+const { resolveRequestedPathForIpc } = require('./hardening.cjs')
+
+const COMMIT_CONTEXT_DIFF_MAX_CHARS = 120_000
+const COMMIT_CONTEXT_UNTRACKED_MAX = 80
+const UNTRACKED_LINE_COUNT_CONCURRENCY = 16
+const UNTRACKED_LINE_COUNT_MAX_BYTES = 1024 * 1024
+
+// GUI-launched Electron apps on macOS inherit only a minimal PATH (no
+// /opt/homebrew/bin or /usr/local/bin), so `gh` — and the `git` gh shells out
+// to — aren't found. Augment PATH with the resolved gh dir + the common
+// package-manager bins so gh runs the same way it does in a terminal.
+function ghEnv(ghBin) {
+  const extra = [ghBin ? path.dirname(ghBin) : '', '/opt/homebrew/bin', '/usr/local/bin', '/usr/bin'].filter(
+    dir => dir && dir !== '.'
+  )
+
+  return { ...process.env, PATH: [...extra, process.env.PATH].filter(Boolean).join(path.delimiter) }
+}
+
+// Run the `gh` CLI in a repo. Resolves { ok, stdout } so callers branch on
+// availability/auth without a throw. gh missing/unauthed → ok:false.
+function runGh(args, cwd, ghBin) {
+  return new Promise(resolve => {
+    execFile(
+      ghBin || 'gh',
+      args,
+      { cwd, env: ghEnv(ghBin), windowsHide: true, timeout: 30_000, maxBuffer: 8 * 1024 * 1024 },
+      (err, stdout) => resolve({ ok: !err, stdout: String(stdout || '') })
+    )
+  })
+}
+
+function gitFor(cwd, gitBin) {
+  return simpleGit({ baseDir: cwd, binary: gitBin || 'git', maxConcurrentProcesses: 4, trimmed: false })
+}
+
+// simple-git reports renames as `old => new` (and `dir/{old => new}/f`); resolve
+// to the NEW path so the row addresses the real file for diff/stage.
+function resolveRenamePath(raw) {
+  const path = String(raw || '').trim()
+
+  if (!path.includes(' => ')) {
+    return path
+  }
+
+  const brace = path.match(/^(.*)\{(.*) => (.*)\}(.*)$/)
+
+  if (brace) {
+    const [, prefix, , to, suffix] = brace
+
+    return `${prefix}${to}${suffix}`.replace(/\/{2,}/g, '/')
+  }
+
+  return path.split(' => ').pop().trim()
+}
+
+// DiffResult.files → Map<path, {added, removed}> (binary files carry no line
+// delta).
+function countsByPath(summary) {
+  const map = new Map()
+
+  for (const file of summary.files) {
+    map.set(resolveRenamePath(file.file), {
+      added: file.binary ? 0 : file.insertions,
+      removed: file.binary ? 0 : file.deletions
+    })
+  }
+
+  return map
+}
+
+// Untracked files don't appear in diffSummary(); count insertions from disk so
+// the review tree can show +N for new files (matches an all-add diff view).
+// Insertions = line count: newline bytes, plus one for a final unterminated
+// line. Binary (NUL byte) → 0, mirroring git numstat's "-".
+async function untrackedInsertions(cwd, relPath) {
+  try {
+    const fullPath = path.join(cwd, relPath)
+    const stat = await fs.stat(fullPath)
+
+    if (!stat.isFile() || stat.size > UNTRACKED_LINE_COUNT_MAX_BYTES) {
+      return 0
+    }
+
+    const buf = await fs.readFile(fullPath)
+
+    if (buf.includes(0)) {
+      return 0
+    }
+
+    let lines = 0
+
+    for (const byte of buf) {
+      if (byte === 10) {
+        lines++
+      }
+    }
+
+    return buf.length > 0 && buf[buf.length - 1] !== 10 ? lines + 1 : lines
+  } catch {
+    return 0
+  }
+}
+
+function capText(text, maxChars, label = 'truncated') {
+  const value = String(text || '')
+
+  if (value.length <= maxChars) {
+    return value
+  }
+
+  return `${value.slice(0, maxChars)}\n# ${label}: ${value.length - maxChars} chars omitted\n`
+}
+
+async function fillUntrackedCounts(cwd, files) {
+  const pending = files.filter(file => file.status === '?' && file.added === 0 && file.removed === 0)
+
+  for (let i = 0; i < pending.length; i += UNTRACKED_LINE_COUNT_CONCURRENCY) {
+    await Promise.all(
+      pending.slice(i, i + UNTRACKED_LINE_COUNT_CONCURRENCY).map(async file => {
+        file.added = await untrackedInsertions(cwd, file.path)
+      })
+    )
+  }
+}
+
+// Resolve the base ref for "all branch changes": merge-base with the remote
+// default branch (origin/HEAD), falling back to common trunk names.
+async function branchBase(git) {
+  const candidates = []
+
+  try {
+    const head = (await git.revparse(['--abbrev-ref', 'origin/HEAD'])).trim()
+
+    if (head) {
+      candidates.push(head)
+    }
+  } catch {
+    // No origin/HEAD configured.
+  }
+
+  candidates.push('origin/main', 'origin/master', 'main', 'master')
+
+  for (const ref of candidates) {
+    try {
+      const base = (await git.raw(['merge-base', 'HEAD', ref])).trim()
+
+      if (base) {
+        return base
+      }
+    } catch {
+      // Ref doesn't exist; try the next candidate.
+    }
+  }
+
+  return null
+}
+
+// Resolve the repo's default branch NAME ("main" / "master" / …), preferring
+// the remote's HEAD, then common local trunk names. Null when none is found
+// (e.g. a fresh repo with only a feature branch). Used to offer "branch off the
+// trunk" regardless of which branch you're currently on.
+async function defaultBranchName(git) {
+  try {
+    const head = (await git.revparse(['--abbrev-ref', 'origin/HEAD'])).trim()
+
+    // "origin/main" → "main"; skip the bare "origin/HEAD" placeholder.
+    if (head && head !== 'origin/HEAD') {
+      return head.replace(/^origin\//, '')
+    }
+  } catch {
+    // No origin/HEAD configured.
+  }
+
+  // Prefer a local trunk, then a remote-only one (returns the clean name either
+  // way) so "branch off main" works even before main is checked out locally.
+  for (const ref of [
+    'refs/heads/main',
+    'refs/heads/master',
+    'refs/remotes/origin/main',
+    'refs/remotes/origin/master'
+  ]) {
+    try {
+      await git.raw(['rev-parse', '--verify', '--quiet', ref])
+
+      return ref.replace(/^refs\/(?:heads|remotes\/origin)\//, '')
+    } catch {
+      // Ref doesn't exist; try the next candidate.
+    }
+  }
+
+  return null
+}
+
+// A status file's single-letter classification, preferring the staged (index)
+// code over the worktree code; untracked wins (simple-git marks both '?').
+function statusLetter(file) {
+  if (file.index === '?' || file.working_dir === '?') {
+    return '?'
+  }
+
+  const code = file.index && file.index !== ' ' ? file.index : file.working_dir
+
+  return (code || 'M').toUpperCase()
+}
+
+const isStaged = file => Boolean(file.index && file.index !== ' ' && file.index !== '?')
+
+async function reviewList(repoPath, scope, baseRef, gitBin) {
+  let cwd
+
+  try {
+    cwd = resolveRequestedPathForIpc(repoPath, { purpose: 'Review list' })
+  } catch {
+    return { files: [], base: null }
+  }
+
+  const git = gitFor(cwd, gitBin)
+
+  try {
+    if (scope === 'branch' || scope === 'lastTurn') {
+      const base = scope === 'branch' ? await branchBase(git) : baseRef
+
+      if (!base) {
+        return { files: [], base: null }
+      }
+
+      const range = scope === 'branch' ? `${base}...HEAD` : base
+      const summary = await git.diffSummary([range])
+      const files = summary.files.map(file => ({
+        path: resolveRenamePath(file.file),
+        added: file.binary ? 0 : file.insertions,
+        removed: file.binary ? 0 : file.deletions,
+        status: 'M',
+        staged: false
+      }))
+
+      // "Last turn" also surfaces files created since the baseline (untracked).
+      if (scope === 'lastTurn') {
+        const status = await git.status()
+
+        for (const path of status.not_added) {
+          if (!files.some(f => f.path === path)) {
+            files.push({ path, added: 0, removed: 0, status: '?', staged: false })
+          }
+        }
+      }
+
+      files.sort((a, b) => a.path.localeCompare(b.path))
+      await fillUntrackedCounts(cwd, files)
+
+      return { files, base }
+    }
+
+    // Default: uncommitted (staged + unstaged + untracked), one row per path.
+    const [status, staged, unstaged] = await Promise.all([
+      git.status(),
+      git.diffSummary(['--cached']),
+      git.diffSummary([])
+    ])
+    const stagedCounts = countsByPath(staged)
+    const unstagedCounts = countsByPath(unstaged)
+
+    const files = status.files.map(file => {
+      const filePath = resolveRenamePath(file.path)
+      const sc = stagedCounts.get(filePath) || { added: 0, removed: 0 }
+      const uc = unstagedCounts.get(filePath) || { added: 0, removed: 0 }
+
+      return {
+        path: filePath,
+        added: sc.added + uc.added,
+        removed: sc.removed + uc.removed,
+        status: statusLetter(file),
+        staged: isStaged(file)
+      }
+    })
+
+    files.sort((a, b) => a.path.localeCompare(b.path))
+    await fillUntrackedCounts(cwd, files)
+
+    return { files, base: null }
+  } catch {
+    return { files: [], base: null }
+  }
+}
+
+async function reviewDiff(repoPath, filePath, scope, baseRef, staged, gitBin) {
+  let cwd
+
+  try {
+    cwd = resolveRequestedPathForIpc(repoPath, { purpose: 'Review diff' })
+  } catch {
+    return ''
+  }
+
+  const git = gitFor(cwd, gitBin)
+  const safe = args => git.diff(args).catch(() => '')
+
+  if (scope === 'branch') {
+    const base = await branchBase(git)
+
+    return base ? safe([`${base}...HEAD`, '--', filePath]) : ''
+  }
+
+  if (scope === 'lastTurn') {
+    return baseRef ? safe([baseRef, '--', filePath]) : ''
+  }
+
+  if (staged) {
+    return safe(['--cached', '--', filePath])
+  }
+
+  const worktree = await safe(['--', filePath])
+
+  if (worktree.trim()) {
+    return worktree
+  }
+
+  // Untracked file: no worktree diff exists, so synthesize an all-add diff via
+  // --no-index (exits non-zero by design when files differ, so go around
+  // simple-git's reject-on-nonzero with a raw execFile).
+  return new Promise(resolve => {
+    execFile(
+      gitBin || 'git',
+      ['diff', '--no-index', '--', '/dev/null', filePath],
+      { cwd, windowsHide: true, timeout: 30_000, maxBuffer: 32 * 1024 * 1024 },
+      (_err, stdout) => resolve(String(stdout || ''))
+    )
+  })
+}
+
+// Working-tree-vs-HEAD diff for ONE file — the "what changed since the last
+// commit" view used by the file preview. Unlike reviewDiff this never synthesizes
+// a full-add for a clean tracked file (so a pristine file shows no diff); it only
+// all-adds a genuinely untracked file.
+async function fileDiffVsHead(repoPath, filePath, gitBin) {
+  let cwd
+
+  try {
+    cwd = resolveRequestedPathForIpc(repoPath, { purpose: 'File diff' })
+  } catch {
+    return ''
+  }
+
+  const git = gitFor(cwd, gitBin)
+  const head = await git.diff(['HEAD', '--', filePath]).catch(() => '')
+
+  if (head.trim()) {
+    return head
+  }
+
+  // No tracked changes vs HEAD. Only synthesize an all-add diff for a file git
+  // doesn't know yet; a clean tracked file must return empty.
+  const status = await git.raw(['status', '--porcelain', '--', filePath]).catch(() => '')
+
+  if (!status.trim().startsWith('??')) {
+    return ''
+  }
+
+  return new Promise(resolve => {
+    execFile(
+      gitBin || 'git',
+      ['diff', '--no-index', '--', '/dev/null', filePath],
+      { cwd, windowsHide: true, timeout: 30_000, maxBuffer: 32 * 1024 * 1024 },
+      (_err, stdout) => resolve(String(stdout || ''))
+    )
+  })
+}
+
+async function reviewStage(repoPath, filePath, gitBin) {
+  const cwd = resolveRequestedPathForIpc(repoPath, { purpose: 'Review stage' })
+
+  await gitFor(cwd, gitBin).raw(filePath ? ['add', '--', filePath] : ['add', '-A'])
+
+  return { ok: true }
+}
+
+async function reviewUnstage(repoPath, filePath, gitBin) {
+  const cwd = resolveRequestedPathForIpc(repoPath, { purpose: 'Review unstage' })
+
+  await gitFor(cwd, gitBin).raw(filePath ? ['reset', '-q', 'HEAD', '--', filePath] : ['reset', '-q', 'HEAD'])
+
+  return { ok: true }
+}
+
+// Discard changes back to the committed state. Destructive — the renderer
+// confirms first. Restores tracked files and removes untracked ones.
+async function reviewRevert(repoPath, filePath, gitBin) {
+  const cwd = resolveRequestedPathForIpc(repoPath, { purpose: 'Review revert' })
+  const git = gitFor(cwd, gitBin)
+
+  if (filePath) {
+    await git.raw(['checkout', 'HEAD', '--', filePath]).catch(() => undefined)
+    await git.raw(['clean', '-fd', '--', filePath]).catch(() => undefined)
+  } else {
+    await git.raw(['checkout', 'HEAD', '--', '.']).catch(() => undefined)
+    await git.raw(['clean', '-fd']).catch(() => undefined)
+  }
+
+  return { ok: true }
+}
+
+// Resolve a ref to a commit sha (captures the turn baseline for "Last turn").
+async function reviewRevParse(repoPath, ref, gitBin) {
+  let cwd
+
+  try {
+    cwd = resolveRequestedPathForIpc(repoPath, { purpose: 'Review rev-parse' })
+  } catch {
+    return null
+  }
+
+  try {
+    return (await gitFor(cwd, gitBin).revparse([ref || 'HEAD'])).trim() || null
+  } catch {
+    return null
+  }
+}
+
+// Commit the working tree. Mirrors VS Code: if nothing is staged, stage
+// everything first ("commit all"), then commit. Optionally push afterward,
+// setting upstream on the first push.
+async function reviewCommit(repoPath, message, push, gitBin) {
+  const cwd = resolveRequestedPathForIpc(repoPath, { purpose: 'Review commit' })
+  const git = gitFor(cwd, gitBin)
+  const status = await git.status()
+
+  if (status.staged.length === 0) {
+    await git.raw(['add', '-A'])
+  }
+
+  await git.commit(message)
+
+  if (push) {
+    const fresh = await git.status()
+
+    if (fresh.tracking) {
+      await git.push()
+    } else if (fresh.current) {
+      await git.raw(['push', '-u', 'origin', fresh.current])
+    }
+  }
+
+  return { ok: true }
+}
+
+// Gather the context the model needs to draft a commit message: the diff of
+// what *will* be committed (staged when anything is staged, else everything
+// vs HEAD — mirroring reviewCommit's "stage all when nothing staged" rule),
+// the names of untracked files (which carry no diff), and recent commit
+// subjects for style. Diff is capped so the payload stays bounded. Reads only.
+async function reviewCommitContext(repoPath, gitBin) {
+  let cwd
+
+  try {
+    cwd = resolveRequestedPathForIpc(repoPath, { purpose: 'Review commit context' })
+  } catch {
+    return { diff: '', recent: '' }
+  }
+
+  const git = gitFor(cwd, gitBin)
+  const safe = args => git.diff(args).catch(() => '')
+
+  let status
+  try {
+    status = await git.status()
+  } catch {
+    return { diff: '', recent: '' }
+  }
+
+  // What will land: staged changes if any, otherwise all tracked changes vs HEAD.
+  let diff = capText(
+    status.staged.length > 0 ? await safe(['--cached']) : await safe(['HEAD']),
+    COMMIT_CONTEXT_DIFF_MAX_CHARS,
+    'diff truncated for commit-message generation'
+  )
+
+  // Untracked files have no diff — list them so new files aren't invisible.
+  const untracked = status.not_added || []
+  if (untracked.length > 0) {
+    const visible = untracked.slice(0, COMMIT_CONTEXT_UNTRACKED_MAX)
+    const omitted = untracked.length - visible.length
+    const note =
+      `\n# New (untracked) files:\n${visible.map(p => `#   ${p}`).join('\n')}\n` +
+      (omitted > 0 ? `#   ... ${omitted} more omitted\n` : '')
+
+    diff = diff ? `${diff}${note}` : note
+  }
+
+  const recent = await git.raw(['log', '-n', '10', '--pretty=format:%s']).catch(() => '')
+
+  return { diff: diff || '', recent: String(recent || '').trim() }
+}
+
+async function reviewPush(repoPath, gitBin) {
+  const cwd = resolveRequestedPathForIpc(repoPath, { purpose: 'Review push' })
+  const git = gitFor(cwd, gitBin)
+  const status = await git.status()
+
+  if (status.tracking) {
+    await git.push()
+  } else if (status.current) {
+    await git.raw(['push', '-u', 'origin', status.current])
+  }
+
+  return { ok: true }
+}
+
+// gh availability + auth + whether this branch already has a PR. Reads only;
+// drives the PR button's enabled/label state. `ghReady` is false when gh is
+// missing OR not authenticated — either way the PR action can't run.
+async function reviewShipInfo(repoPath, ghBin) {
+  let cwd
+
+  try {
+    cwd = resolveRequestedPathForIpc(repoPath, { purpose: 'Review ship info' })
+  } catch {
+    return { ghReady: false, pr: null }
+  }
+
+  const auth = await runGh(['auth', 'status'], cwd, ghBin)
+
+  if (!auth.ok) {
+    return { ghReady: false, pr: null }
+  }
+
+  const view = await runGh(['pr', 'view', '--json', 'url,state,number'], cwd, ghBin)
+
+  if (!view.ok) {
+    // gh exits non-zero when no PR exists for the branch — that's not an error.
+    return { ghReady: true, pr: null }
+  }
+
+  try {
+    const pr = JSON.parse(view.stdout)
+
+    return { ghReady: true, pr: pr && pr.url ? { url: pr.url, state: pr.state, number: pr.number } : null }
+  } catch {
+    return { ghReady: true, pr: null }
+  }
+}
+
+// Create a PR for the current branch (pushing first so gh has a remote ref),
+// letting gh fill title/body from the commits. Returns the new PR url.
+async function reviewCreatePr(repoPath, gitBin, ghBin) {
+  const cwd = resolveRequestedPathForIpc(repoPath, { purpose: 'Review create PR' })
+
+  await reviewPush(repoPath, gitBin).catch(() => undefined)
+
+  const created = await runGh(['pr', 'create', '--fill'], cwd, ghBin)
+
+  if (!created.ok) {
+    throw new Error('gh pr create failed (is gh installed and authenticated?)')
+  }
+
+  const url = created.stdout.trim().split('\n').filter(Boolean).pop() || ''
+
+  return { url }
+}
+
+// Compact working-tree status for the composer coding rail: branch, ahead/behind,
+// per-state change counts, +/- vs HEAD, and a capped changed-file list.
+async function repoStatus(repoPath, gitBin) {
+  let cwd
+
+  try {
+    cwd = resolveRequestedPathForIpc(repoPath, { purpose: 'Repo status' })
+  } catch {
+    return null
+  }
+
+  // Session cwds can point at a deleted worktree for a moment (or forever in a
+  // stale row). simple-git throws at construction time on a missing baseDir, so
+  // fail soft and hide the coding rail instead of spamming IPC handler errors.
+  try {
+    const stat = await fs.stat(cwd)
+    if (!stat.isDirectory()) {
+      return null
+    }
+  } catch {
+    return null
+  }
+
+  let git
+  try {
+    git = gitFor(cwd, gitBin)
+  } catch {
+    return null
+  }
+  let status
+
+  try {
+    status = await git.status()
+  } catch {
+    // Not a repo / git unavailable / remote backend.
+    return null
+  }
+
+  const detached = typeof status.detached === 'boolean' ? status.detached : !status.current
+  const files = status.files.map(file => ({
+    path: file.path,
+    staged: isStaged(file),
+    unstaged: Boolean(file.working_dir && file.working_dir !== ' ' && file.working_dir !== '?'),
+    untracked: file.index === '?' || file.working_dir === '?',
+    conflicted: file.index === 'U' || file.working_dir === 'U'
+  }))
+
+  const result = {
+    branch: detached ? null : status.current || null,
+    defaultBranch: await defaultBranchName(git),
+    detached,
+    ahead: status.ahead || 0,
+    behind: status.behind || 0,
+    staged: files.filter(f => f.staged).length,
+    unstaged: files.filter(f => f.unstaged).length,
+    untracked: status.not_added.length,
+    conflicted: status.conflicted.length,
+    changed: files.length,
+    added: 0,
+    removed: 0,
+    files: files.slice(0, 200)
+  }
+
+  // +/- vs HEAD (staged + unstaged tracked changes). No HEAD yet → leave 0.
+  try {
+    const summary = await git.diffSummary(['HEAD'])
+    result.added = summary.insertions
+    result.removed = summary.deletions
+  } catch {
+    // No commits yet.
+  }
+
+  // `git diff HEAD` ignores untracked files, so a turn that only creates new
+  // files (the common case — a fresh module, a demo dir) showed +0 in the rail
+  // while the review pane counted them. Fold untracked insertions into `added`
+  // so the rail matches reality. Bounded (size cap + concurrency) like the
+  // review tree; only the capped file slice is counted so a huge untracked tree
+  // can't stall the probe.
+  try {
+    const untracked = status.not_added.slice(0, 500)
+    for (let i = 0; i < untracked.length; i += UNTRACKED_LINE_COUNT_CONCURRENCY) {
+      const batch = await Promise.all(
+        untracked.slice(i, i + UNTRACKED_LINE_COUNT_CONCURRENCY).map(path => untrackedInsertions(cwd, path))
+      )
+      result.added += batch.reduce((sum, n) => sum + n, 0)
+    }
+  } catch {
+    // Best-effort: a probe failure just leaves untracked lines uncounted.
+  }
+
+  return result
+}
+
+module.exports = {
+  branchBase,
+  fileDiffVsHead,
+  repoStatus,
+  resolveRenamePath,
+  reviewCommit,
+  reviewCommitContext,
+  reviewCreatePr,
+  reviewDiff,
+  reviewList,
+  reviewPush,
+  reviewRevParse,
+  reviewRevert,
+  reviewShipInfo,
+  reviewStage,
+  reviewUnstage
+}

apps/desktop/electron/git-review-ops.test.cjs

@@ -0,0 +1,22 @@
+'use strict'
+
+const assert = require('node:assert/strict')
+const test = require('node:test')
+
+const { resolveRenamePath } = require('./git-review-ops.cjs')
+
+test('resolveRenamePath: plain path is unchanged', () => {
+  assert.equal(resolveRenamePath('src/a.ts'), 'src/a.ts')
+})
+
+test('resolveRenamePath: simple rename resolves to the new path', () => {
+  assert.equal(resolveRenamePath('old.ts => new.ts'), 'new.ts')
+})
+
+test('resolveRenamePath: brace rename resolves to the new path', () => {
+  assert.equal(resolveRenamePath('src/{old => new}/file.ts'), 'src/new/file.ts')
+})
+
+test('resolveRenamePath: brace rename collapsing a segment', () => {
+  assert.equal(resolveRenamePath('src/{lib => }/file.ts'), 'src/file.ts')
+})

apps/desktop/electron/git-worktree-ops.cjs

@@ -0,0 +1,350 @@
+'use strict'
+
+// Git-driven worktree operations for the desktop "Start work" flow: spin up a
+// fresh worktree the lightest way (`git worktree add -b`), list real worktrees,
+// and remove them. Git is the source of truth; the renderer just drives these.
+
+const path = require('node:path')
+const fs = require('node:fs')
+const { execFile } = require('node:child_process')
+
+const { resolveRequestedPathForIpc } = require('./hardening.cjs')
+
+function runGit(gitBin, args, cwd) {
+  return new Promise((resolve, reject) => {
+    execFile(
+      gitBin,
+      args,
+      { cwd, windowsHide: true, timeout: 30_000, maxBuffer: 8 * 1024 * 1024 },
+      (err, stdout, stderr) => {
+        if (err) {
+          err.stderr = String(stderr || '')
+          reject(err)
+
+          return
+        }
+
+        resolve(String(stdout || ''))
+      }
+    )
+  })
+}
+
+// Parse `git worktree list --porcelain`. The first record is the main worktree.
+function parseWorktrees(out) {
+  const trees = []
+  let cur = null
+
+  for (const line of out.split('\n')) {
+    if (line.startsWith('worktree ')) {
+      if (cur) {
+        trees.push(cur)
+      }
+
+      cur = { path: line.slice(9).trim(), branch: null, detached: false, bare: false, locked: false }
+    } else if (!cur) {
+      continue
+    } else if (line.startsWith('branch ')) {
+      cur.branch = line
+        .slice(7)
+        .trim()
+        .replace(/^refs\/heads\//, '')
+    } else if (line === 'detached') {
+      cur.detached = true
+    } else if (line === 'bare') {
+      cur.bare = true
+    } else if (line.startsWith('locked')) {
+      cur.locked = true
+    }
+  }
+
+  if (cur) {
+    trees.push(cur)
+  }
+
+  return trees
+}
+
+async function listWorktrees(repoPath, gitBin) {
+  let resolved
+
+  try {
+    resolved = resolveRequestedPathForIpc(repoPath, { purpose: 'Worktree list' })
+  } catch {
+    return []
+  }
+
+  try {
+    const out = await runGit(gitBin, ['worktree', 'list', '--porcelain'], resolved)
+
+    return parseWorktrees(out).map((tree, index) => ({
+      path: tree.path,
+      branch: tree.branch,
+      isMain: index === 0,
+      detached: tree.detached,
+      locked: tree.locked
+    }))
+  } catch {
+    return []
+  }
+}
+
+// A git-ref-safe branch name (spaces → "-", drop forbidden chars, trim edges),
+// or "" when nothing usable remains. Mirrors the renderer's `gitRef`, so a bad
+// value can't reach `git` no matter the caller (the GUI also enforces live).
+function sanitizeBranch(name) {
+  return String(name || '')
+    .replace(/\s+/g, '-')
+    .replace(/[^\w./-]/g, '')
+    .replace(/-{2,}/g, '-')
+    .replace(/\/{2,}/g, '/')
+    .replace(/\.{2,}/g, '.')
+    .replace(/^[-./]+|[-./]+$/g, '')
+}
+
+function slugify(name) {
+  const slug = String(name || '')
+    .trim()
+    .toLowerCase()
+    .replace(/[^a-z0-9]+/g, '-')
+    .replace(/^-+|-+$/g, '')
+    .slice(0, 40)
+    .replace(/-+$/g, '')
+
+  return slug || 'work'
+}
+
+const TRUNK_BRANCHES = ['main', 'master']
+
+async function gitLine(gitBin, args, cwd) {
+  try {
+    return (await runGit(gitBin, args, cwd)).trim()
+  } catch {
+    return ''
+  }
+}
+
+async function defaultBranch(gitBin, cwd) {
+  const remote = (
+    await gitLine(gitBin, ['symbolic-ref', '--quiet', '--short', 'refs/remotes/origin/HEAD'], cwd)
+  ).replace(/^origin\//, '')
+
+  if (remote) {
+    return remote
+  }
+
+  const configured = await gitLine(gitBin, ['config', '--get', 'init.defaultBranch'], cwd)
+
+  if (configured) {
+    return configured
+  }
+
+  for (const branch of TRUNK_BRANCHES) {
+    if (await gitLine(gitBin, ['show-ref', '--verify', `refs/heads/${branch}`], cwd)) {
+      return branch
+    }
+  }
+
+  return ''
+}
+
+// A brand-new project folder isn't a git repo — and a freshly-init'd one has no
+// commit to branch from — so `git worktree add` would fail. Make the dir a repo
+// with a root commit on the user's behalf so worktrees "just work". No-op for a
+// repo that already has commits; never touches the user's files (the seed commit
+// is `--allow-empty`), and never inits a dir that already lives inside a repo.
+async function ensureGitRepo(gitBin, dir) {
+  let needsRoot = false
+
+  try {
+    const inside = (await runGit(gitBin, ['rev-parse', '--is-inside-work-tree'], dir)).trim()
+
+    if (inside !== 'true') {
+      await runGit(gitBin, ['init'], dir)
+      needsRoot = true
+    } else {
+      // Repo exists; a worktree still needs a HEAD to branch from.
+      try {
+        await runGit(gitBin, ['rev-parse', '--verify', 'HEAD'], dir)
+      } catch {
+        needsRoot = true
+      }
+    }
+  } catch {
+    await runGit(gitBin, ['init'], dir)
+    needsRoot = true
+  }
+
+  if (needsRoot) {
+    // Inline identity so the seed commit lands even with no global git config.
+    await runGit(
+      gitBin,
+      [
+        '-c',
+        'user.email=hermes@localhost',
+        '-c',
+        'user.name=Hermes',
+        'commit',
+        '--allow-empty',
+        '-m',
+        'Initial commit'
+      ],
+      dir
+    )
+  }
+}
+
+// Resolve the repo's MAIN worktree root, so `.worktrees/` always nests under the
+// primary checkout even when called from a linked worktree.
+async function mainRoot(gitBin, cwd) {
+  const list = await listWorktrees(cwd, gitBin)
+  const main = list.find(tree => tree.isMain)
+
+  return main ? main.path : cwd
+}
+
+function uniqueDir(base) {
+  let dir = base
+  let n = 1
+
+  while (fs.existsSync(dir)) {
+    n += 1
+    dir = `${base}-${n}`
+  }
+
+  return dir
+}
+
+async function addExistingBranchWorktree(gitBin, root, name) {
+  const branch = sanitizeBranch(name)
+
+  if (!branch) {
+    throw new Error('Branch name is required.')
+  }
+
+  if (branch === (await defaultBranch(gitBin, root))) {
+    await runGit(gitBin, ['switch', branch], root)
+
+    return { path: root, branch, repoRoot: root }
+  }
+
+  const dir = uniqueDir(path.join(root, '.worktrees', slugify(branch)))
+  await runGit(gitBin, ['worktree', 'add', dir, branch], root)
+
+  return { path: dir, branch, repoRoot: root }
+}
+
+async function addWorktree(repoPath, options, gitBin) {
+  const resolved = resolveRequestedPathForIpc(repoPath, { purpose: 'Worktree add' })
+  // A new project's folder may not be a git repo yet — init it (with a root
+  // commit) so the worktree has something to branch from.
+  await ensureGitRepo(gitBin, resolved)
+  const root = await mainRoot(gitBin, resolved)
+  const opts = options || {}
+
+  if (opts.existingBranch) {
+    return addExistingBranchWorktree(gitBin, root, opts.existingBranch)
+  }
+
+  const slug = slugify(opts.name || `work-${Date.now().toString(36)}`)
+  const branch = sanitizeBranch(opts.branch) || `hermes/${slug}`
+  const dir = uniqueDir(path.join(root, '.worktrees', slug))
+
+  const args = ['worktree', 'add', '-b', branch, dir]
+
+  if (opts.base) {
+    args.push(String(opts.base))
+  }
+
+  try {
+    await runGit(gitBin, args, root)
+  } catch (err) {
+    // Branch name may already exist — retry checking out the existing branch
+    // into a fresh worktree dir instead of failing the whole flow.
+    if (/already exists/i.test(err.stderr || '')) {
+      await runGit(gitBin, ['worktree', 'add', dir, branch], root)
+    } else {
+      throw err
+    }
+  }
+
+  return { path: dir, branch, repoRoot: root }
+}
+
+async function removeWorktree(repoPath, worktreePath, options, gitBin) {
+  const resolvedRepo = resolveRequestedPathForIpc(repoPath, { purpose: 'Worktree remove (repo)' })
+  const resolvedTree = resolveRequestedPathForIpc(worktreePath, { purpose: 'Worktree remove (tree)' })
+  const root = await mainRoot(gitBin, resolvedRepo)
+  const args = ['worktree', 'remove']
+
+  if (options && options.force) {
+    args.push('--force')
+  }
+
+  args.push(resolvedTree)
+  await runGit(gitBin, args, root)
+
+  return { removed: resolvedTree }
+}
+
+// List local branches for the "convert a branch into a worktree" picker, most
+// recently committed first. Each carries whether it's already checked out in a
+// worktree and, when checked out, that worktree's path. Empty on a non-repo /
+// remote backend where the probe can't run.
+async function listBranches(repoPath, gitBin) {
+  let resolved
+
+  try {
+    resolved = resolveRequestedPathForIpc(repoPath, { purpose: 'Branch list' })
+  } catch {
+    return []
+  }
+
+  try {
+    const out = await runGit(
+      gitBin,
+      ['for-each-ref', '--format=%(refname:short)', '--sort=-committerdate', 'refs/heads'],
+      resolved
+    )
+    const trees = await listWorktrees(resolved, gitBin)
+    const pathByBranch = new Map(trees.filter(tree => tree.branch).map(tree => [tree.branch, tree.path]))
+    const trunk = await defaultBranch(gitBin, resolved)
+
+    return out
+      .split('\n')
+      .map(line => line.trim())
+      .filter(Boolean)
+      .map(name => ({
+        name,
+        checkedOut: pathByBranch.has(name),
+        isDefault: Boolean(trunk && name === trunk),
+        worktreePath: pathByBranch.get(name) || null
+      }))
+  } catch {
+    return []
+  }
+}
+
+async function switchBranch(repoPath, branch, gitBin) {
+  const resolved = resolveRequestedPathForIpc(repoPath, { purpose: 'Branch switch' })
+  const target = sanitizeBranch(branch)
+
+  if (!target) {
+    throw new Error('Branch name is required.')
+  }
+
+  await runGit(gitBin, ['switch', target], resolved)
+
+  return { branch: target }
+}
+
+module.exports = {
+  addWorktree,
+  ensureGitRepo,
+  listBranches,
+  listWorktrees,
+  parseWorktrees,
+  removeWorktree,
+  sanitizeBranch,
+  switchBranch
+}

apps/desktop/electron/git-worktree-ops.test.cjs

@@ -0,0 +1,214 @@
+'use strict'
+
+const assert = require('node:assert/strict')
+const { execFileSync } = require('node:child_process')
+const fs = require('node:fs')
+const os = require('node:os')
+const path = require('node:path')
+const test = require('node:test')
+
+const {
+  addWorktree,
+  ensureGitRepo,
+  listBranches,
+  parseWorktrees,
+  sanitizeBranch,
+  switchBranch
+} = require('./git-worktree-ops.cjs')
+
+test('sanitizeBranch: spaces → hyphens, forbidden chars dropped, edges trimmed', () => {
+  assert.equal(sanitizeBranch('beach vibes'), 'beach-vibes')
+  assert.equal(sanitizeBranch('feat/cool thing'), 'feat/cool-thing')
+  assert.equal(sanitizeBranch('  wip~^:? '), 'wip')
+  assert.equal(sanitizeBranch('///'), '')
+})
+
+test('parseWorktrees: main checkout + linked worktree', () => {
+  const out = [
+    'worktree /repo',
+    'HEAD abc123',
+    'branch refs/heads/main',
+    '',
+    'worktree /repo/.worktrees/feat',
+    'HEAD def456',
+    'branch refs/heads/hermes/feat',
+    ''
+  ].join('\n')
+
+  const trees = parseWorktrees(out)
+
+  assert.equal(trees.length, 2)
+  assert.equal(trees[0].path, '/repo')
+  assert.equal(trees[0].branch, 'main')
+  assert.equal(trees[1].path, '/repo/.worktrees/feat')
+  assert.equal(trees[1].branch, 'hermes/feat')
+})
+
+test('parseWorktrees: detached + locked flags', () => {
+  const out = ['worktree /repo/wt', 'HEAD abc', 'detached', 'locked reason', ''].join('\n')
+  const trees = parseWorktrees(out)
+
+  assert.equal(trees.length, 1)
+  assert.equal(trees[0].detached, true)
+  assert.equal(trees[0].locked, true)
+  assert.equal(trees[0].branch, null)
+})
+
+test('parseWorktrees: empty input', () => {
+  assert.deepEqual(parseWorktrees(''), [])
+})
+
+test('ensureGitRepo: inits a plain dir with a root commit so worktrees branch', async () => {
+  const dir = fs.mkdtempSync(path.join(os.tmpdir(), 'hermes-wt-'))
+  const git = (...args) => execFileSync('git', args, { cwd: dir }).toString().trim()
+
+  try {
+    await ensureGitRepo('git', dir)
+    assert.match(git('rev-parse', '--verify', 'HEAD'), /^[0-9a-f]{7,}$/)
+
+    // The whole point: a worktree can now branch off the seeded root commit.
+    execFileSync('git', ['worktree', 'add', '-b', 'wt', path.join(dir, '.worktrees', 'wt')], { cwd: dir })
+    assert.ok(fs.existsSync(path.join(dir, '.worktrees', 'wt')))
+
+    // Idempotent: an already-committed repo gets no extra commit.
+    await ensureGitRepo('git', dir)
+    assert.equal(git('rev-list', '--count', 'HEAD'), '1')
+  } finally {
+    fs.rmSync(dir, { recursive: true, force: true })
+  }
+})
+
+test('switchBranch: switches a normal checkout branch', async () => {
+  const dir = fs.mkdtempSync(path.join(os.tmpdir(), 'hermes-switch-'))
+  const git = (...args) => execFileSync('git', args, { cwd: dir }).toString().trim()
+
+  try {
+    await ensureGitRepo('git', dir)
+    execFileSync('git', ['branch', 'feature'], { cwd: dir })
+
+    await switchBranch(dir, 'feature', 'git')
+
+    assert.equal(git('branch', '--show-current'), 'feature')
+  } finally {
+    fs.rmSync(dir, { recursive: true, force: true })
+  }
+})
+
+test('listBranches: lists locals and flags the checked-out branch', async () => {
+  const dir = fs.mkdtempSync(path.join(os.tmpdir(), 'hermes-branches-'))
+
+  try {
+    await ensureGitRepo('git', dir)
+    const current = execFileSync('git', ['branch', '--show-current'], { cwd: dir }).toString().trim()
+    execFileSync('git', ['branch', 'feature'], { cwd: dir })
+
+    const branches = await listBranches(dir, 'git')
+    const names = branches.map(b => b.name).sort()
+
+    assert.deepEqual(names, [current, 'feature'].sort())
+    // The repo's own checkout is flagged; the unused branch is convertible.
+    assert.equal(branches.find(b => b.name === current).checkedOut, true)
+    assert.equal(branches.find(b => b.name === current).isDefault, true)
+    assert.equal(fs.realpathSync(branches.find(b => b.name === current).worktreePath), fs.realpathSync(dir))
+    assert.equal(branches.find(b => b.name === 'feature').checkedOut, false)
+    assert.equal(branches.find(b => b.name === 'feature').isDefault, false)
+    assert.equal(branches.find(b => b.name === 'feature').worktreePath, null)
+  } finally {
+    fs.rmSync(dir, { recursive: true, force: true })
+  }
+})
+
+test('listBranches: flags a free default branch as default, not checked out', async () => {
+  const dir = fs.mkdtempSync(path.join(os.tmpdir(), 'hermes-branches-default-'))
+  const git = (...args) => execFileSync('git', args, { cwd: dir }).toString().trim()
+
+  try {
+    await ensureGitRepo('git', dir)
+    const trunk = git('branch', '--show-current')
+    execFileSync('git', ['switch', '-c', 'rawr'], { cwd: dir })
+
+    const branches = await listBranches(dir, 'git')
+    const defaultBranch = branches.find(b => b.name === trunk)
+
+    assert.equal(defaultBranch.checkedOut, false)
+    assert.equal(defaultBranch.isDefault, true)
+    assert.equal(defaultBranch.worktreePath, null)
+  } finally {
+    fs.rmSync(dir, { recursive: true, force: true })
+  }
+})
+
+test('listBranches: a branch claimed by a worktree is flagged checked out', async () => {
+  const dir = fs.mkdtempSync(path.join(os.tmpdir(), 'hermes-branches-wt-'))
+
+  try {
+    await ensureGitRepo('git', dir)
+    execFileSync('git', ['branch', 'feature'], { cwd: dir })
+    // addWorktree converts the existing "feature" branch into a worktree.
+    const result = await addWorktree(dir, { existingBranch: 'feature' }, 'git')
+
+    assert.equal(result.branch, 'feature')
+    assert.ok(fs.existsSync(result.path))
+
+    const branches = await listBranches(dir, 'git')
+
+    assert.equal(branches.find(b => b.name === 'feature').checkedOut, true)
+  } finally {
+    fs.rmSync(dir, { recursive: true, force: true })
+  }
+})
+
+test('listBranches: empty on a non-repo path', async () => {
+  const dir = fs.mkdtempSync(path.join(os.tmpdir(), 'hermes-nonrepo-'))
+
+  try {
+    assert.deepEqual(await listBranches(dir, 'git'), [])
+  } finally {
+    fs.rmSync(dir, { recursive: true, force: true })
+  }
+})
+
+test('addWorktree: existingBranch checks the branch out without a new branch', async () => {
+  const dir = fs.mkdtempSync(path.join(os.tmpdir(), 'hermes-convert-'))
+  const git = (...args) => execFileSync('git', args, { cwd: dir }).toString().trim()
+
+  try {
+    await ensureGitRepo('git', dir)
+    execFileSync('git', ['branch', 'cool/feature'], { cwd: dir })
+
+    const before = git('branch', '--list').split('\n').length
+    const result = await addWorktree(dir, { existingBranch: 'cool/feature' }, 'git')
+
+    // No new branch was created — only the existing one is checked out.
+    assert.equal(git('branch', '--list').split('\n').length, before)
+    assert.equal(result.branch, 'cool/feature')
+    // Dir is named off the branch slug, nested under the main repo's .worktrees.
+    assert.match(result.path, /[/\\]\.worktrees[/\\]cool-feature/)
+    assert.equal(
+      execFileSync('git', ['branch', '--show-current'], { cwd: result.path }).toString().trim(),
+      'cool/feature'
+    )
+  } finally {
+    fs.rmSync(dir, { recursive: true, force: true })
+  }
+})
+
+test('addWorktree: existing default branch switches the main checkout, not .worktrees/main', async () => {
+  const dir = fs.mkdtempSync(path.join(os.tmpdir(), 'hermes-convert-default-'))
+  const git = (...args) => execFileSync('git', args, { cwd: dir }).toString().trim()
+
+  try {
+    await ensureGitRepo('git', dir)
+    const trunk = git('branch', '--show-current')
+    execFileSync('git', ['switch', '-c', 'rawr'], { cwd: dir })
+
+    const result = await addWorktree(dir, { existingBranch: trunk }, 'git')
+
+    assert.equal(result.branch, trunk)
+    assert.equal(fs.realpathSync(result.path), fs.realpathSync(dir))
+    assert.equal(git('branch', '--show-current'), trunk)
+    assert.equal(fs.existsSync(path.join(dir, '.worktrees', trunk)), false)
+  } finally {
+    fs.rmSync(dir, { recursive: true, force: true })
+  }
+})

apps/desktop/electron/git-worktrees.cjs

@@ -1,174 +0,0 @@
-'use strict'
-
-// Resolve git-worktree relationships for a set of session cwds, reading git's
-// on-disk metadata directly (no `git` spawn per path):
-//
-//   - A normal checkout has a `.git` DIRECTORY at its root → it's the main
-//     worktree; its repo root IS that directory's parent.
-//   - A linked worktree has a `.git` FILE: `gitdir: <repo>/.git/worktrees/<name>`.
-//     That admin dir's `commondir` points back at the shared `<repo>/.git`, whose
-//     parent is the main repo root.
-//
-// Grouping by repoRoot therefore clusters a repo's main checkout with all of its
-// linked worktrees, regardless of how the worktree directories are named. The
-// branch (read from the worktree's own HEAD) gives each worktree a meaningful
-// label.
-
-const fs = require('node:fs')
-const path = require('node:path')
-const { resolveRequestedPathForIpc } = require('./hardening.cjs')
-
-// Walk up from `start` to the nearest ancestor that carries a `.git` entry
-// (file for a linked worktree, dir for the main checkout). Capped so a stray
-// path can't loop forever.
-function findGitHost(start, fsImpl) {
-  let dir = start
-
-  for (let i = 0; i < 64; i += 1) {
-    const dotgit = path.join(dir, '.git')
-
-    try {
-      if (fsImpl.existsSync(dotgit)) {
-        return dir
-      }
-    } catch {
-      return null
-    }
-
-    const parent = path.dirname(dir)
-
-    if (parent === dir) {
-      return null
-    }
-
-    dir = parent
-  }
-
-  return null
-}
-
-function readBranch(gitDir, fsImpl) {
-  try {
-    const head = fsImpl.readFileSync(path.join(gitDir, 'HEAD'), 'utf8').trim()
-    const ref = head.match(/^ref:\s*refs\/heads\/(.+)$/)
-
-    if (ref) {
-      return ref[1]
-    }
-
-    // Detached HEAD: surface a short sha so the worktree still gets a label.
-    return /^[0-9a-f]{7,40}$/i.test(head) ? head.slice(0, 8) : null
-  } catch {
-    return null
-  }
-}
-
-// Given the directory that owns the `.git` entry, resolve its worktree identity.
-function resolveFromHost(host, fsImpl) {
-  const dotgit = path.join(host, '.git')
-  let stat
-
-  try {
-    stat = fsImpl.statSync(dotgit)
-  } catch {
-    return null
-  }
-
-  if (stat.isDirectory()) {
-    return {
-      repoRoot: host,
-      worktreeRoot: host,
-      isMainWorktree: true,
-      branch: readBranch(dotgit, fsImpl)
-    }
-  }
-
-  // Linked worktree: `.git` is a file pointing at the admin dir.
-  let contents
-
-  try {
-    contents = fsImpl.readFileSync(dotgit, 'utf8').trim()
-  } catch {
-    return null
-  }
-
-  const match = contents.match(/^gitdir:\s*(.+)$/m)
-
-  if (!match) {
-    return null
-  }
-
-  const adminDir = path.resolve(host, match[1].trim())
-
-  // `commondir` resolves to the shared `<repo>/.git`; fall back to walking two
-  // levels up from `<repo>/.git/worktrees/<name>` if it's missing.
-  let commonDir
-
-  try {
-    const rel = fsImpl.readFileSync(path.join(adminDir, 'commondir'), 'utf8').trim()
-    commonDir = path.resolve(adminDir, rel)
-  } catch {
-    commonDir = path.dirname(path.dirname(adminDir))
-  }
-
-  return {
-    repoRoot: path.dirname(commonDir),
-    worktreeRoot: host,
-    isMainWorktree: false,
-    branch: readBranch(adminDir, fsImpl)
-  }
-}
-
-function resolveWorktree(startPath, fsImpl = fs) {
-  let resolved
-
-  try {
-    resolved = resolveRequestedPathForIpc(startPath, { purpose: 'Worktree lookup' })
-  } catch {
-    return null
-  }
-
-  let start = resolved
-
-  try {
-    const stat = fsImpl.statSync(resolved)
-
-    if (!stat.isDirectory()) {
-      start = path.dirname(resolved)
-    }
-  } catch {
-    return null
-  }
-
-  const host = findGitHost(start, fsImpl)
-
-  if (!host) {
-    return null
-  }
-
-  return resolveFromHost(host, fsImpl)
-}
-
-// Batch entry point for the renderer: maps each requested cwd to its worktree
-// info (or null when it isn't inside a git checkout / can't be read). Dedupes so
-// many sessions sharing a cwd cost one lookup.
-async function worktreesForIpc(cwds, options = {}) {
-  const fsImpl = options.fs || fs
-  const list = Array.isArray(cwds) ? cwds : []
-  const out = {}
-
-  for (const cwd of list) {
-    if (typeof cwd !== 'string' || !cwd.trim() || cwd in out) {
-      continue
-    }
-
-    out[cwd] = resolveWorktree(cwd, fsImpl)
-  }
-
-  return out
-}
-
-module.exports = {
-  resolveWorktree,
-  worktreesForIpc
-}

apps/desktop/electron/hardening.cjs

@@ -186,7 +186,10 @@ async function statForIpc(fsImpl, resolvedPath, purpose, typeLabel) {
     if (code === 'ENOENT' || code === 'ENOTDIR') {
       throw ipcPathError(code || 'ENOENT', `${purpose} failed: ${typeLabel} does not exist.`)
     }
-    throw ipcPathError(code || 'read-error', `${purpose} failed: ${error instanceof Error ? error.message : String(error)}`)
+    throw ipcPathError(
+      code || 'read-error',
+      `${purpose} failed: ${error instanceof Error ? error.message : String(error)}`
+    )
   }
 }
 
@@ -201,7 +204,10 @@ async function realpathForIpc(fsImpl, resolvedPath, purpose) {
     return realPath
   } catch (error) {
     const code = error && typeof error === 'object' ? error.code : ''
-    throw ipcPathError(code || 'read-error', `${purpose} failed: ${error instanceof Error ? error.message : String(error)}`)
+    throw ipcPathError(
+      code || 'read-error',
+      `${purpose} failed: ${error instanceof Error ? error.message : String(error)}`
+    )
   }
 }
 

apps/desktop/electron/oauth-net-request.test.cjs

@@ -30,5 +30,8 @@ test('setJsonRequestHeaders does not set Electron-restricted Content-Length', ()
   setJsonRequestHeaders(request)
 
   assert.deepEqual(headers, [['Content-Type', 'application/json']])
-  assert.equal(headers.some(([name]) => name.toLowerCase() === 'content-length'), false)
+  assert.equal(
+    headers.some(([name]) => name.toLowerCase() === 'content-length'),
+    false
+  )
 })

apps/desktop/electron/preload.cjs

@@ -7,6 +7,32 @@ contextBridge.exposeInMainWorld('hermesDesktop', {
   getGatewayWsUrl: profile => ipcRenderer.invoke('hermes:gateway:ws-url', profile),
   openSessionWindow: (sessionId, opts) => ipcRenderer.invoke('hermes:window:openSession', sessionId, opts),
   openNewSessionWindow: () => ipcRenderer.invoke('hermes:window:openNewSession'),
+  petOverlay: {
+    // Main renderer → main process: window lifecycle + drag. `request` is
+    // `{ bounds, screen }`; resolves with the screen bounds it actually used.
+    open: request => ipcRenderer.invoke('hermes:pet-overlay:open', request),
+    close: () => ipcRenderer.invoke('hermes:pet-overlay:close'),
+    setBounds: bounds => ipcRenderer.send('hermes:pet-overlay:set-bounds', bounds),
+    setIgnoreMouse: ignore => ipcRenderer.send('hermes:pet-overlay:ignore-mouse', ignore),
+    // Flip the overlay focusable (and focus it) while the composer needs keys.
+    setFocusable: focusable => ipcRenderer.send('hermes:pet-overlay:set-focusable', focusable),
+    // Main renderer → overlay (forwarded by main): push the latest pet state.
+    pushState: payload => ipcRenderer.send('hermes:pet-overlay:state', payload),
+    // Overlay → main renderer (forwarded by main): pop back in / composer submit.
+    control: payload => ipcRenderer.send('hermes:pet-overlay:control', payload),
+    // Overlay subscribes to state pushes.
+    onState: callback => {
+      const listener = (_event, payload) => callback(payload)
+      ipcRenderer.on('hermes:pet-overlay:state', listener)
+      return () => ipcRenderer.removeListener('hermes:pet-overlay:state', listener)
+    },
+    // Main renderer subscribes to overlay control messages.
+    onControl: callback => {
+      const listener = (_event, payload) => callback(payload)
+      ipcRenderer.on('hermes:pet-overlay:control', listener)
+      return () => ipcRenderer.removeListener('hermes:pet-overlay:control', listener)
+    }
+  },
   getBootProgress: () => ipcRenderer.invoke('hermes:boot-progress:get'),
   getConnectionConfig: profile => ipcRenderer.invoke('hermes:connection-config:get', profile),
   saveConnectionConfig: payload => ipcRenderer.invoke('hermes:connection-config:save', payload),
@@ -44,6 +70,7 @@ contextBridge.exposeInMainWorld('hermesDesktop', {
   setTranslucency: payload => ipcRenderer.send('hermes:translucency', payload),
   setPreviewShortcutActive: active => ipcRenderer.send('hermes:previewShortcutActive', Boolean(active)),
   openExternal: url => ipcRenderer.invoke('hermes:openExternal', url),
+  openPreviewInBrowser: url => ipcRenderer.invoke('hermes:openPreviewInBrowser', url),
   fetchLinkTitle: url => ipcRenderer.invoke('hermes:fetchLinkTitle', url),
   sanitizeWorkspaceCwd: cwd => ipcRenderer.invoke('hermes:workspace:sanitize', cwd),
   settings: {
@@ -55,7 +82,35 @@ contextBridge.exposeInMainWorld('hermesDesktop', {
   getRecentLogs: () => ipcRenderer.invoke('hermes:logs:recent'),
   readDir: dirPath => ipcRenderer.invoke('hermes:fs:readDir', dirPath),
   gitRoot: startPath => ipcRenderer.invoke('hermes:fs:gitRoot', startPath),
-  worktrees: cwds => ipcRenderer.invoke('hermes:fs:worktrees', cwds),
+  revealPath: targetPath => ipcRenderer.invoke('hermes:fs:reveal', targetPath),
+  renamePath: (targetPath, newName) => ipcRenderer.invoke('hermes:fs:rename', targetPath, newName),
+  writeTextFile: (filePath, content) => ipcRenderer.invoke('hermes:fs:writeText', filePath, content),
+  trashPath: targetPath => ipcRenderer.invoke('hermes:fs:trash', targetPath),
+  git: {
+    worktreeList: repoPath => ipcRenderer.invoke('hermes:git:worktreeList', repoPath),
+    worktreeAdd: (repoPath, options) => ipcRenderer.invoke('hermes:git:worktreeAdd', repoPath, options),
+    worktreeRemove: (repoPath, worktreePath, options) =>
+      ipcRenderer.invoke('hermes:git:worktreeRemove', repoPath, worktreePath, options),
+    branchSwitch: (repoPath, branch) => ipcRenderer.invoke('hermes:git:branchSwitch', repoPath, branch),
+    branchList: repoPath => ipcRenderer.invoke('hermes:git:branchList', repoPath),
+    repoStatus: repoPath => ipcRenderer.invoke('hermes:git:repoStatus', repoPath),
+    fileDiff: (repoPath, filePath) => ipcRenderer.invoke('hermes:git:fileDiff', repoPath, filePath),
+    scanRepos: (roots, options) => ipcRenderer.invoke('hermes:git:scanRepos', roots, options),
+    review: {
+      list: (repoPath, scope, baseRef) => ipcRenderer.invoke('hermes:git:review:list', repoPath, scope, baseRef),
+      diff: (repoPath, filePath, scope, baseRef, staged) =>
+        ipcRenderer.invoke('hermes:git:review:diff', repoPath, filePath, scope, baseRef, staged),
+      stage: (repoPath, filePath) => ipcRenderer.invoke('hermes:git:review:stage', repoPath, filePath),
+      unstage: (repoPath, filePath) => ipcRenderer.invoke('hermes:git:review:unstage', repoPath, filePath),
+      revert: (repoPath, filePath) => ipcRenderer.invoke('hermes:git:review:revert', repoPath, filePath),
+      revParse: (repoPath, ref) => ipcRenderer.invoke('hermes:git:review:revParse', repoPath, ref),
+      commit: (repoPath, message, push) => ipcRenderer.invoke('hermes:git:review:commit', repoPath, message, push),
+      commitContext: repoPath => ipcRenderer.invoke('hermes:git:review:commitContext', repoPath),
+      push: repoPath => ipcRenderer.invoke('hermes:git:review:push', repoPath),
+      shipInfo: repoPath => ipcRenderer.invoke('hermes:git:review:shipInfo', repoPath),
+      createPr: repoPath => ipcRenderer.invoke('hermes:git:review:createPr', repoPath)
+    }
+  },
   terminal: {
     dispose: id => ipcRenderer.invoke('hermes:terminal:dispose', id),
     resize: (id, size) => ipcRenderer.invoke('hermes:terminal:resize', id, size),

apps/desktop/electron/titlebar-overlay-width.cjs

@@ -0,0 +1,11 @@
+// Pre-layout fallback for WCO right-edge reservation (--titlebar-tools-right).
+// Live width comes from navigator.windowControlsOverlay in the renderer.
+
+const OVERLAY_FALLBACK_WIDTH = 144
+
+/** @param {{ isWindows?: boolean, isWsl?: boolean }} opts */
+function nativeOverlayWidth({ isWindows = false, isWsl = false } = {}) {
+  return isWindows || isWsl ? OVERLAY_FALLBACK_WIDTH : 0
+}
+
+module.exports = { OVERLAY_FALLBACK_WIDTH, nativeOverlayWidth }

apps/desktop/electron/titlebar-overlay-width.test.cjs

@@ -0,0 +1,29 @@
+const assert = require('node:assert/strict')
+const test = require('node:test')
+
+const { OVERLAY_FALLBACK_WIDTH, nativeOverlayWidth } = require('./titlebar-overlay-width.cjs')
+
+// This static reservation is only the pre-layout FALLBACK. Once laid out the
+// renderer reads the exact width from navigator.windowControlsOverlay
+// (use-window-controls-overlay-width.ts) and uses these values only when the WCO
+// API is unavailable.
+
+test('Windows reserves the overlay fallback width', () => {
+  assert.equal(nativeOverlayWidth({ isWindows: true }), OVERLAY_FALLBACK_WIDTH)
+})
+
+test('WSLg paints the same WCO, so it reserves the same fallback width', () => {
+  // The original bug: WSL fell through to 0, so the right tools sat under the
+  // controls and the title overran into them.
+  assert.equal(nativeOverlayWidth({ isWsl: true }), OVERLAY_FALLBACK_WIDTH)
+})
+
+test('plain Linux and macOS reserve nothing', () => {
+  assert.equal(nativeOverlayWidth({ isWindows: false, isWsl: false }), 0)
+  assert.equal(nativeOverlayWidth(), 0)
+  assert.equal(nativeOverlayWidth({}), 0)
+})
+
+test('the fallback width is a sane positive pixel value', () => {
+  assert.ok(Number.isInteger(OVERLAY_FALLBACK_WIDTH) && OVERLAY_FALLBACK_WIDTH > 0)
+})

apps/desktop/electron/update-count.cjs

@@ -0,0 +1,28 @@
+'use strict'
+
+// Whether `git rev-list HEAD..origin/<branch> --count` produces a meaningful
+// number worth computing. On a SHALLOW checkout (installer clones with
+// --depth 1) the local history often shares no merge-base with the freshly
+// fetched origin tip, so the count enumerates the entire remote ancestry and
+// returns a bogus huge number (e.g. 12104) — see #51922. resolveBehindCount
+// discards that bogus count in favour of a SHA compare, so the caller should
+// SKIP the expensive rev-list entirely in that case rather than run it and
+// throw the result away.
+function shouldCountCommits({ isShallow, hasMergeBase }) {
+  return !(isShallow && !hasMergeBase)
+}
+
+// Resolve how many commits the local checkout is behind origin for the desktop
+// update indicator. When the count isn't meaningful (shallow + no merge-base)
+// fall back to a binary up-to-date check by SHA, exactly like the official-SSH
+// path in checkUpdates() and the CLI guard in hermes_cli/banner.py. Full clones
+// (developers / Docker dev images) keep the exact count path unchanged.
+function resolveBehindCount({ countStr, currentSha, targetSha, isShallow, hasMergeBase }) {
+  if (!shouldCountCommits({ isShallow, hasMergeBase })) {
+    if (currentSha && targetSha && currentSha === targetSha) return 0
+    return 1 // behind by an unknown amount — show a generic "update available"
+  }
+  return Number.parseInt(countStr, 10) || 0
+}
+
+module.exports = { resolveBehindCount, shouldCountCommits }

apps/desktop/electron/update-count.test.cjs

@@ -0,0 +1,127 @@
+'use strict'
+const test = require('node:test')
+const assert = require('node:assert/strict')
+const { resolveBehindCount, shouldCountCommits } = require('./update-count.cjs')
+
+// FAIL-BEFORE: pre-fix the function did `Number.parseInt(countStr) || 0`
+// unconditionally, so a shallow checkout with no merge-base surfaced the bogus
+// rev-list count (e.g. 12104). This asserts the new shallow/no-merge-base branch.
+test('shallow checkout with no merge-base does NOT trust the bogus rev-list count', () => {
+  assert.equal(
+    resolveBehindCount({
+      countStr: '12104',
+      currentSha: 'aaa',
+      targetSha: 'bbb',
+      isShallow: true,
+      hasMergeBase: false
+    }),
+    1
+  )
+})
+
+test('shallow checkout with no merge-base but identical SHA reports up-to-date', () => {
+  assert.equal(
+    resolveBehindCount({
+      countStr: '12104',
+      currentSha: 'abc',
+      targetSha: 'abc',
+      isShallow: true,
+      hasMergeBase: false
+    }),
+    0
+  )
+})
+
+test('shallow checkout WITH a merge-base keeps the exact count (reliable)', () => {
+  assert.equal(
+    resolveBehindCount({
+      countStr: '3',
+      currentSha: 'aaa',
+      targetSha: 'bbb',
+      isShallow: true,
+      hasMergeBase: true
+    }),
+    3
+  )
+})
+
+test('full (non-shallow) clone keeps the exact count path unchanged', () => {
+  assert.equal(
+    resolveBehindCount({
+      countStr: '7',
+      currentSha: 'aaa',
+      targetSha: 'bbb',
+      isShallow: false,
+      hasMergeBase: true
+    }),
+    7
+  )
+})
+
+test('up-to-date full clone reports 0', () => {
+  assert.equal(
+    resolveBehindCount({
+      countStr: '0',
+      currentSha: 'x',
+      targetSha: 'x',
+      isShallow: false,
+      hasMergeBase: true
+    }),
+    0
+  )
+})
+
+test('non-numeric count falls back to 0 (defensive, unchanged behaviour)', () => {
+  assert.equal(
+    resolveBehindCount({
+      countStr: '',
+      currentSha: 'aaa',
+      targetSha: 'bbb',
+      isShallow: false,
+      hasMergeBase: true
+    }),
+    0
+  )
+})
+
+// shouldCountCommits gates the expensive `rev-list --count` in checkUpdates().
+// FAIL-BEFORE: in the shallow + no-merge-base case the caller ran rev-list
+// unconditionally and discarded the bogus result; this predicate lets the
+// caller SKIP the whole-ancestry enumeration in exactly that case (#51922).
+test('shallow checkout with no merge-base SKIPS the rev-list count', () => {
+  assert.equal(shouldCountCommits({ isShallow: true, hasMergeBase: false }), false)
+})
+
+test('shallow checkout WITH a merge-base still runs the count', () => {
+  assert.equal(shouldCountCommits({ isShallow: true, hasMergeBase: true }), true)
+})
+
+test('full (non-shallow) clone always runs the count', () => {
+  assert.equal(shouldCountCommits({ isShallow: false, hasMergeBase: true }), true)
+  assert.equal(shouldCountCommits({ isShallow: false, hasMergeBase: false }), true)
+})
+
+// The skip path produces an empty countStr; resolveBehindCount must NOT trust
+// it and must fall through to the SHA compare (mirrors the live call site).
+test('skipped-count path resolves via SHA compare, never via empty countStr', () => {
+  assert.equal(
+    resolveBehindCount({
+      countStr: '',
+      currentSha: 'aaa',
+      targetSha: 'bbb',
+      isShallow: true,
+      hasMergeBase: false
+    }),
+    1
+  )
+  assert.equal(
+    resolveBehindCount({
+      countStr: '',
+      currentSha: 'same',
+      targetSha: 'same',
+      isShallow: true,
+      hasMergeBase: false
+    }),
+    0
+  )
+})

apps/desktop/electron/update-relaunch.test.cjs

@@ -62,7 +62,10 @@ test('resolveUnpackedRelease is null for AppImage / .deb / .rpm / dev / unresolv
   assert.equal(resolveUnpackedRelease('/usr/lib/hermes/hermes', ROOT, 'linux'), null)
   assert.equal(resolveUnpackedRelease('/opt/Hermes/hermes', ROOT, 'linux'), null)
   // dev electron
-  assert.equal(resolveUnpackedRelease('/home/u/.hermes/hermes-agent/node_modules/electron/dist/electron', ROOT, 'linux'), null)
+  assert.equal(
+    resolveUnpackedRelease('/home/u/.hermes/hermes-agent/node_modules/electron/dist/electron', ROOT, 'linux'),
+    null
+  )
   // empty / missing
   assert.equal(resolveUnpackedRelease('', ROOT, 'linux'), null)
   assert.equal(resolveUnpackedRelease(path.join(UNPACKED, 'hermes'), '', 'linux'), null)

apps/desktop/electron/update-remote.cjs

@@ -39,7 +39,9 @@ function canonicalGitHubRemote(url) {
 }
 
 function isSshRemote(url) {
-  const value = String(url || '').trim().toLowerCase()
+  const value = String(url || '')
+    .trim()
+    .toLowerCase()
   return value.startsWith('git@') || value.startsWith('ssh://')
 }
 

apps/desktop/electron/vscode-marketplace.cjs

@@ -26,7 +26,11 @@ const REQUEST_TIMEOUT_MS = 20_000
 const ID_RE = /^[\w-]+\.[\w-]+$/
 
 /** Minimal HTTPS helper with redirect-following, timeout, and a size cap. */
-function request(url, { method = 'GET', headers = {}, body = null, maxBytes = MAX_VSIX_BYTES } = {}, redirectsLeft = MAX_REDIRECTS) {
+function request(
+  url,
+  { method = 'GET', headers = {}, body = null, maxBytes = MAX_VSIX_BYTES } = {},
+  redirectsLeft = MAX_REDIRECTS
+) {
   return new Promise((resolve, reject) => {
     const req = https.request(url, { method, headers }, res => {
       const status = res.statusCode ?? 0
@@ -42,7 +46,13 @@ function request(url, { method = 'GET', headers = {}, body = null, maxBytes = MA
         const next = new URL(res.headers.location, url).toString()
         res.resume()
         // Redirects to the CDN are plain GETs (drop the POST body).
-        resolve(request(next, { method: 'GET', headers: { 'User-Agent': headers['User-Agent'] }, maxBytes }, redirectsLeft - 1))
+        resolve(
+          request(
+            next,
+            { method: 'GET', headers: { 'User-Agent': headers['User-Agent'] }, maxBytes },
+            redirectsLeft - 1
+          )
+        )
 
         return
       }

apps/desktop/electron/window-state.cjs

@@ -0,0 +1,117 @@
+/**
+ * Pure geometry helpers for window-state.json — restoring the main window's
+ * size, position, and maximized flag across launches. Side-effect-free so the
+ * part that actually matters (rejecting garbage + off-screen bounds) is
+ * unit-testable without booting Electron; main.cjs owns the file I/O and the
+ * live `screen` displays.
+ */
+
+// Defaults mirror the historical hardcoded BrowserWindow size; MIN_* mirror its
+// minWidth/minHeight so a restored size never undershoots what the live window
+// allows. A fresh install (no saved state) is byte-identical to before.
+const DEFAULT_WIDTH = 1220
+const DEFAULT_HEIGHT = 800
+const MIN_WIDTH = 400
+const MIN_HEIGHT = 620
+
+// Keep at least this much of the window over a display work area before we trust
+// a saved position, so the title bar stays grabbable after a monitor unplugs.
+const MIN_VISIBLE = 48
+
+const finite = v => typeof v === 'number' && Number.isFinite(v)
+const clamp = (v, lo, hi) => Math.max(lo, Math.min(v, hi))
+
+// Parse raw JSON → clean state, or null if garbage. width/height are required
+// and floored; x/y survive only as a finite pair; isMaximized is strict.
+function sanitizeWindowState(raw) {
+  if (!raw || typeof raw !== 'object' || !finite(raw.width) || !finite(raw.height)) return null
+
+  const state = {
+    width: Math.max(MIN_WIDTH, Math.round(raw.width)),
+    height: Math.max(MIN_HEIGHT, Math.round(raw.height)),
+    isMaximized: raw.isMaximized === true
+  }
+  if (finite(raw.x) && finite(raw.y)) {
+    state.x = Math.round(raw.x)
+    state.y = Math.round(raw.y)
+  }
+  return state
+}
+
+// True when `bounds` overlaps some display's work area by ≥ MIN_VISIBLE on both
+// axes. `displays` is Electron's screen.getAllDisplays() shape.
+function onScreen(bounds, displays) {
+  if (!Array.isArray(displays)) return false
+  return displays.some(({ workArea: a } = {}) => {
+    if (!a) return false
+    const x = Math.min(bounds.x + bounds.width, a.x + a.width) - Math.max(bounds.x, a.x)
+    const y = Math.min(bounds.y + bounds.height, a.y + a.height) - Math.max(bounds.y, a.y)
+    return x >= MIN_VISIBLE && y >= MIN_VISIBLE
+  })
+}
+
+// Sanitized state (or null) → BrowserWindow size/position options. Always sets
+// width/height, capped to the largest current display so a size saved on a
+// since-disconnected bigger monitor can't exceed any screen the user now has.
+// Sets x/y only when still on-screen; otherwise Electron centers the window.
+function computeWindowOptions(state, displays) {
+  const opts = {
+    width: finite(state?.width) ? state.width : DEFAULT_WIDTH,
+    height: finite(state?.height) ? state.height : DEFAULT_HEIGHT
+  }
+
+  const cap = (Array.isArray(displays) ? displays : []).reduce(
+    (m, { workArea: a } = {}) =>
+      a && finite(a.width) && finite(a.height)
+        ? { width: Math.max(m.width, a.width), height: Math.max(m.height, a.height) }
+        : m,
+    { width: 0, height: 0 }
+  )
+  if (cap.width && cap.height) {
+    opts.width = clamp(opts.width, MIN_WIDTH, cap.width)
+    opts.height = clamp(opts.height, MIN_HEIGHT, cap.height)
+  }
+
+  if (
+    state &&
+    finite(state.x) &&
+    finite(state.y) &&
+    onScreen({ x: state.x, y: state.y, width: opts.width, height: opts.height }, displays)
+  ) {
+    opts.x = state.x
+    opts.y = state.y
+  }
+  return opts
+}
+
+// Trailing debounce: collapse a burst of resize/move events (Linux fires many
+// mid-drag) into a single run `delayMs` after the last. `.flush()` runs now and
+// cancels the pending timer — used on close, before the window is gone.
+function debounce(fn, delayMs) {
+  let timer = null
+  const debounced = () => {
+    clearTimeout(timer)
+    timer = setTimeout(() => {
+      timer = null
+      fn()
+    }, delayMs)
+  }
+  debounced.flush = () => {
+    clearTimeout(timer)
+    timer = null
+    fn()
+  }
+  return debounced
+}
+
+module.exports = {
+  DEFAULT_WIDTH,
+  DEFAULT_HEIGHT,
+  MIN_WIDTH,
+  MIN_HEIGHT,
+  MIN_VISIBLE,
+  sanitizeWindowState,
+  onScreen,
+  computeWindowOptions,
+  debounce
+}

apps/desktop/electron/window-state.test.cjs

@@ -0,0 +1,150 @@
+/**
+ * Unit tests for the pure window-state geometry helpers. These cover the logic
+ * that protects the user: garbage rejection, off-screen fallback, oversized
+ * clamping, and the debounce that collapses mid-drag write storms.
+ */
+
+const test = require('node:test')
+const assert = require('node:assert/strict')
+
+const {
+  DEFAULT_WIDTH,
+  DEFAULT_HEIGHT,
+  MIN_WIDTH,
+  MIN_HEIGHT,
+  sanitizeWindowState,
+  onScreen,
+  computeWindowOptions,
+  debounce
+} = require('./window-state.cjs')
+
+// A single 1920×1080 monitor (work area trimmed for the taskbar).
+const PRIMARY = [{ workArea: { x: 0, y: 0, width: 1920, height: 1040 } }]
+// A laptop panel left behind after a bigger external monitor is unplugged.
+const LAPTOP = [{ workArea: { x: 0, y: 0, width: 1366, height: 728 } }]
+
+// ─── sanitizeWindowState ───────────────────────────────────────────────────
+
+test('sanitizeWindowState rejects missing/garbage input', () => {
+  for (const bad of [
+    null,
+    undefined,
+    'nope',
+    42,
+    {},
+    { width: 'x', height: 800 },
+    { width: NaN, height: 800 },
+    { width: 1000 }
+  ]) {
+    assert.equal(sanitizeWindowState(bad), null)
+  }
+})
+
+test('sanitizeWindowState keeps a valid full state and rounds HiDPI fractions', () => {
+  assert.deepEqual(sanitizeWindowState({ x: 100.6, y: 50.2, width: 1400.4, height: 900.7, isMaximized: true }), {
+    x: 101,
+    y: 50,
+    width: 1400,
+    height: 901,
+    isMaximized: true
+  })
+})
+
+test('sanitizeWindowState floors size to the minimums', () => {
+  const state = sanitizeWindowState({ width: 10, height: 10 })
+  assert.equal(state.width, MIN_WIDTH)
+  assert.equal(state.height, MIN_HEIGHT)
+})
+
+test('sanitizeWindowState drops a partial position but keeps the size', () => {
+  assert.deepEqual(sanitizeWindowState({ x: 100, width: 1400, height: 900 }), {
+    width: 1400,
+    height: 900,
+    isMaximized: false
+  })
+})
+
+test('sanitizeWindowState treats isMaximized strictly', () => {
+  assert.equal(sanitizeWindowState({ width: 1400, height: 900, isMaximized: 'yes' }).isMaximized, false)
+})
+
+// ─── onScreen ──────────────────────────────────────────────────────────────
+
+test('onScreen accepts a window on the primary or a secondary display', () => {
+  const dual = [...PRIMARY, { workArea: { x: 1920, y: 0, width: 2560, height: 1400 } }]
+  assert.equal(onScreen({ x: 100, y: 100, width: 1220, height: 800 }, PRIMARY), true)
+  assert.equal(onScreen({ x: 2200, y: 200, width: 1220, height: 800 }, dual), true)
+})
+
+test('onScreen rejects off-screen, slivers, and bad input', () => {
+  assert.equal(onScreen({ x: 3000, y: 100, width: 1220, height: 800 }, PRIMARY), false) // past right edge
+  assert.equal(onScreen({ x: 100, y: -900, width: 1220, height: 800 }, PRIMARY), false) // above top
+  assert.equal(onScreen({ x: 1910, y: 100, width: 1220, height: 800 }, PRIMARY), false) // ~10px sliver
+  assert.equal(onScreen({ x: 0, y: 0, width: 1220, height: 800 }, []), false)
+  assert.equal(onScreen({ x: 0, y: 0, width: 1220, height: 800 }, null), false)
+})
+
+// ─── computeWindowOptions ──────────────────────────────────────────────────
+
+test('computeWindowOptions falls back to defaults with no saved state', () => {
+  assert.deepEqual(computeWindowOptions(null, PRIMARY), { width: DEFAULT_WIDTH, height: DEFAULT_HEIGHT })
+})
+
+test('computeWindowOptions restores an on-screen position', () => {
+  const saved = sanitizeWindowState({ x: 200, y: 150, width: 1400, height: 900 })
+  assert.deepEqual(computeWindowOptions(saved, PRIMARY), { width: 1400, height: 900, x: 200, y: 150 })
+})
+
+test('computeWindowOptions keeps the size but drops an off-screen position', () => {
+  const saved = sanitizeWindowState({ x: 5000, y: 150, width: 1400, height: 900 })
+  assert.deepEqual(computeWindowOptions(saved, PRIMARY), { width: 1400, height: 900 })
+})
+
+test('computeWindowOptions clamps a size larger than the only display', () => {
+  const saved = sanitizeWindowState({ width: 2560, height: 1440 })
+  assert.deepEqual(computeWindowOptions(saved, LAPTOP), { width: 1366, height: 728 })
+})
+
+test('computeWindowOptions keeps the MIN floor on a sub-minimum display', () => {
+  const tiny = [{ workArea: { x: 0, y: 0, width: 360, height: 480 } }]
+  const saved = sanitizeWindowState({ width: 2000, height: 1500 })
+  assert.deepEqual(computeWindowOptions(saved, tiny), { width: MIN_WIDTH, height: MIN_HEIGHT })
+})
+
+test('computeWindowOptions does not clamp when displays are unknown', () => {
+  const saved = sanitizeWindowState({ width: 2560, height: 1440 })
+  assert.deepEqual(computeWindowOptions(saved, []), { width: 2560, height: 1440 })
+})
+
+// ─── debounce ──────────────────────────────────────────────────────────────
+
+test('debounce coalesces a burst into one trailing run', t => {
+  t.mock.timers.enable({ apis: ['setTimeout'] })
+  let calls = 0
+  const d = debounce(() => {
+    calls += 1
+  }, 250)
+
+  d()
+  d()
+  d()
+  assert.equal(calls, 0)
+  t.mock.timers.tick(249)
+  assert.equal(calls, 0)
+  t.mock.timers.tick(1)
+  assert.equal(calls, 1)
+})
+
+test('debounce.flush runs now and cancels the pending timer', t => {
+  t.mock.timers.enable({ apis: ['setTimeout'] })
+  let calls = 0
+  const d = debounce(() => {
+    calls += 1
+  }, 250)
+
+  d()
+  d.flush()
+  assert.equal(calls, 1)
+  t.mock.timers.tick(1000)
+  assert.equal(calls, 1)
+})

apps/desktop/electron/windows-child-process.test.cjs

@@ -12,7 +12,8 @@ function readElectronFile(name) {
 }
 
 function requireHiddenChildOptions(source, needle) {
-  const index = source.indexOf(needle)
+  const match = needle instanceof RegExp ? needle.exec(source) : null
+  const index = needle instanceof RegExp ? (match?.index ?? -1) : source.indexOf(needle)
   assert.notEqual(index, -1, `missing call site: ${needle}`)
   const snippet = source.slice(index, index + 700)
   assert.match(
@@ -28,14 +29,28 @@ test('desktop background child processes opt into hidden Windows consoles', () =
   assert.match(source, /function hiddenWindowsChildOptions\(options = \{\}\)/)
 
   requireHiddenChildOptions(source, "execFileSync(\n          'reg'")
-  requireHiddenChildOptions(source, 'execFileSync(pyExe')
-  requireHiddenChildOptions(source, 'spawn(resolveGitBinary()')
+  requireHiddenChildOptions(source, /execFileSync\(\s*pyExe/)
+  requireHiddenChildOptions(source, /spawn\(\s*resolveGitBinary\(\)/)
   requireHiddenChildOptions(source, "execFileSync('taskkill'")
-  requireHiddenChildOptions(source, 'spawn(command, args')
+  requireHiddenChildOptions(source, /spawn\(\s*command,\s*args/)
   requireHiddenChildOptions(source, "spawn('curl'")
-  requireHiddenChildOptions(source, 'spawn(backend.command, backend.args')
-  requireHiddenChildOptions(source, 'hermesProcess = spawn(backend.command, backend.args')
-  requireHiddenChildOptions(source, "spawn(py, ['-m', 'hermes_cli.main', 'uninstall', '--gui-summary']")
+  requireHiddenChildOptions(source, /spawn\(\s*backend\.command,\s*backend\.args/)
+  requireHiddenChildOptions(source, /hermesProcess = spawn\(\s*backend\.command,\s*backend\.args/)
+  requireHiddenChildOptions(source, /spawn\(\s*py,\s*\['-m', 'hermes_cli\.main', 'uninstall', '--gui-summary'\]/)
+
+  assert.match(source, /function unwrapWindowsVenvHermesCommand\(command, dashboardArgs\)/)
+  assert.match(source, /existing Hermes no-console Python at/)
+  assert.match(source, /function getNoConsoleVenvPython\(venvRoot\)/)
+  assert.match(source, /function toNoConsolePython\(pythonPath\)/)
+  assert.match(source, /function applyWindowsNoConsoleSpawnHints\(backend\)/)
+  assert.match(source, /function readVenvHome\(venvRoot\)/)
+  assert.match(source, /path\.join\(venvRoot, 'Scripts', 'pythonw\.exe'\)/)
+  assert.match(source, /backendStartFailure/)
+  assert.match(source, /HERMES_DESKTOP_READY_FILE/)
+  assert.match(source, /readyFile: true/)
+  assert.match(source, /function getVenvSitePackagesEntries\(venvRoot\)/)
+  assert.match(source, /path\.join\(venvRoot, 'Lib', 'site-packages'\)/)
+  assert.match(source, /args: \['-m', 'hermes_cli\.main', \.\.\.dashboardArgs\]/)
 })
 
 test('intentional or interactive desktop child processes stay documented', () => {

apps/desktop/electron/windows-user-env.cjs

@@ -21,8 +21,7 @@ const { execFileSync } = require('node:child_process')
 // the requested value line isn't present.
 function parseRegQueryValue(stdout, name) {
   if (!stdout || !name) return null
-  const typePattern =
-    /^(\S+)\s+(?:REG_SZ|REG_EXPAND_SZ|REG_MULTI_SZ|REG_DWORD|REG_QWORD|REG_BINARY|REG_NONE)\s+(.*)$/
+  const typePattern = /^(\S+)\s+(?:REG_SZ|REG_EXPAND_SZ|REG_MULTI_SZ|REG_DWORD|REG_QWORD|REG_BINARY|REG_NONE)\s+(.*)$/
   for (const rawLine of String(stdout).split(/\r?\n/)) {
     const line = rawLine.trim()
     const match = line.match(typePattern)
@@ -47,10 +46,7 @@ function expandWindowsEnvRefs(value, env = process.env) {
 // Read a User-scoped env var from HKCU\Environment. Windows-only: returns null
 // off-Windows (without spawning), on any spawn error, when `reg` exits non-zero
 // (the value doesn't exist), or when the value is empty.
-function readWindowsUserEnvVar(
-  name,
-  { platform = process.platform, env = process.env, exec = execFileSync } = {}
-) {
+function readWindowsUserEnvVar(name, { platform = process.platform, env = process.env, exec = execFileSync } = {}) {
   if (platform !== 'win32' || !name) return null
   let stdout
   try {

apps/desktop/electron/windows-user-env.test.cjs

@@ -1,21 +1,12 @@
 const assert = require('node:assert/strict')
 const { test } = require('node:test')
 
-const {
-  expandWindowsEnvRefs,
-  parseRegQueryValue,
-  readWindowsUserEnvVar
-} = require('./windows-user-env.cjs')
+const { expandWindowsEnvRefs, parseRegQueryValue, readWindowsUserEnvVar } = require('./windows-user-env.cjs')
 
 // ── parseRegQueryValue ─────────────────────────────────────────────────────
 
 test('parseRegQueryValue extracts a REG_SZ value', () => {
-  const out = [
-    '',
-    'HKEY_CURRENT_USER\\Environment',
-    '    HERMES_HOME    REG_SZ    F:\\Hermes\\data',
-    ''
-  ].join('\r\n')
+  const out = ['', 'HKEY_CURRENT_USER\\Environment', '    HERMES_HOME    REG_SZ    F:\\Hermes\\data', ''].join('\r\n')
   assert.equal(parseRegQueryValue(out, 'HERMES_HOME'), 'F:\\Hermes\\data')
 })
 
@@ -39,10 +30,7 @@ test('parseRegQueryValue returns null when the value line is absent', () => {
 // ── expandWindowsEnvRefs ───────────────────────────────────────────────────
 
 test('expandWindowsEnvRefs expands %VAR% case-insensitively', () => {
-  assert.equal(
-    expandWindowsEnvRefs('%UserProfile%\\h', { USERPROFILE: 'C:\\Users\\jeff' }),
-    'C:\\Users\\jeff\\h'
-  )
+  assert.equal(expandWindowsEnvRefs('%UserProfile%\\h', { USERPROFILE: 'C:\\Users\\jeff' }), 'C:\\Users\\jeff\\h')
 })
 
 test('expandWindowsEnvRefs leaves literal paths and unknown refs intact', () => {

apps/desktop/electron/workspace-cwd.cjs

@@ -14,11 +14,7 @@ function isPackagedInstallPath(dir, { installRoots, isPackaged }) {
     return false
   }
 
-  const roots = new Set(
-    (installRoots ?? [])
-      .filter(Boolean)
-      .map(candidate => path.resolve(String(candidate)))
-  )
+  const roots = new Set((installRoots ?? []).filter(Boolean).map(candidate => path.resolve(String(candidate))))
 
   for (const root of roots) {
     if (resolved === root) {

apps/desktop/electron/workspace-cwd.test.cjs

@@ -13,33 +13,21 @@ const { isPackagedInstallPath } = require('./workspace-cwd.cjs')
 const installRoot = path.resolve('/opt/Hermes')
 
 test('isPackagedInstallPath returns false when not packaged', () => {
-  assert.equal(
-    isPackagedInstallPath(installRoot, { isPackaged: false, installRoots: [installRoot] }),
-    false
-  )
+  assert.equal(isPackagedInstallPath(installRoot, { isPackaged: false, installRoots: [installRoot] }), false)
 })
 
 test('isPackagedInstallPath flags the install root itself', () => {
-  assert.equal(
-    isPackagedInstallPath(installRoot, { isPackaged: true, installRoots: [installRoot] }),
-    true
-  )
+  assert.equal(isPackagedInstallPath(installRoot, { isPackaged: true, installRoots: [installRoot] }), true)
 })
 
 test('isPackagedInstallPath flags paths nested under the install root', () => {
   const nested = path.join(installRoot, 'resources', 'app.asar')
 
-  assert.equal(
-    isPackagedInstallPath(nested, { isPackaged: true, installRoots: [installRoot] }),
-    true
-  )
+  assert.equal(isPackagedInstallPath(nested, { isPackaged: true, installRoots: [installRoot] }), true)
 })
 
 test('isPackagedInstallPath ignores paths outside the install root', () => {
   const homeProject = path.resolve('/home/user/projects/demo')
 
-  assert.equal(
-    isPackagedInstallPath(homeProject, { isPackaged: true, installRoots: [installRoot] }),
-    false
-  )
+  assert.equal(isPackagedInstallPath(homeProject, { isPackaged: true, installRoots: [installRoot] }), false)
 })

apps/desktop/electron/wsl-clipboard-image.cjs

@@ -0,0 +1,92 @@
+// Pull a Windows-host clipboard image from inside WSL2 via PowerShell (WSLg
+// bridges text but not images). Returns PNG bytes or null; exec injectable.
+
+const { execFileSync } = require('node:child_process')
+
+// STA is mandatory: System.Windows.Forms.Clipboard throws ThreadStateException
+// off a single-threaded apartment. We emit base64 (not raw bytes) so the PNG
+// survives stdout's text decoding intact, and write with [Console]::Out.Write
+// to avoid a trailing newline.
+const PS_SCRIPT = [
+  'Add-Type -AssemblyName System.Windows.Forms,System.Drawing',
+  '$img = [System.Windows.Forms.Clipboard]::GetImage()',
+  'if ($null -eq $img) { exit 0 }',
+  '$ms = New-Object System.IO.MemoryStream',
+  '$img.Save($ms, [System.Drawing.Imaging.ImageFormat]::Png)',
+  '[Console]::Out.Write([System.Convert]::ToBase64String($ms.ToArray()))'
+].join('\n')
+
+// PowerShell's -EncodedCommand takes UTF-16LE base64. Encoding the whole script
+// this way sidesteps every layer of WSL→Windows quoting (spaces, quotes,
+// brackets, newlines) that plain -Command arguments would mangle.
+function encodePowerShellCommand(script) {
+  return Buffer.from(String(script), 'utf16le').toString('base64')
+}
+
+// Locate powershell.exe. The bare name resolves through WSL's Windows-interop
+// PATH on every standard WSL2 setup; the absolute fallback covers a stripped
+// PATH. Returns the first candidate — execFile surfaces ENOENT if it's wrong
+// and we fall back to null.
+function powershellCandidates() {
+  return ['powershell.exe', '/mnt/c/Windows/System32/WindowsPowerShell/v1.0/powershell.exe']
+}
+
+function decodeClipboardImageBase64(stdout) {
+  const b64 = String(stdout || '').trim()
+  if (!b64) return null
+
+  let buffer
+  try {
+    buffer = Buffer.from(b64, 'base64')
+  } catch {
+    return null
+  }
+
+  // Guard against partial / garbage output: require a real PNG signature.
+  const PNG_SIGNATURE = Buffer.from([0x89, 0x50, 0x4e, 0x47, 0x0d, 0x0a, 0x1a, 0x0a])
+  if (buffer.length < PNG_SIGNATURE.length || !buffer.subarray(0, PNG_SIGNATURE.length).equals(PNG_SIGNATURE)) {
+    return null
+  }
+
+  return buffer
+}
+
+// Read the Windows clipboard image from inside WSL. Returns a PNG Buffer, or
+// null when there's no image, PowerShell is unreachable, or output is invalid.
+// Linux-only by contract (caller gates on IS_WSL); never throws.
+function readWslWindowsClipboardImage({ exec = execFileSync, candidates = powershellCandidates() } = {}) {
+  const encoded = encodePowerShellCommand(PS_SCRIPT)
+
+  for (const ps of candidates) {
+    try {
+      const stdout = exec(
+        ps,
+        ['-NoProfile', '-NonInteractive', '-STA', '-ExecutionPolicy', 'Bypass', '-EncodedCommand', encoded],
+        {
+          encoding: 'utf8',
+          windowsHide: true,
+          timeout: 8000,
+          // A 4K screenshot base64s to a few MB; give stdout generous headroom.
+          maxBuffer: 64 * 1024 * 1024,
+          // PowerShell writes progress/CLIXML noise to stderr — ignore it.
+          stdio: ['ignore', 'pipe', 'ignore']
+        }
+      )
+      const decoded = decodeClipboardImageBase64(stdout)
+      if (decoded) return decoded
+      // Empty stdout = no image on the clipboard; stop, don't try fallbacks.
+      if (String(stdout || '').trim() === '') return null
+    } catch {
+      // This powershell.exe candidate is missing/failed — try the next one.
+    }
+  }
+
+  return null
+}
+
+module.exports = {
+  decodeClipboardImageBase64,
+  encodePowerShellCommand,
+  powershellCandidates,
+  readWslWindowsClipboardImage
+}

apps/desktop/electron/wsl-clipboard-image.test.cjs

@@ -0,0 +1,114 @@
+const assert = require('node:assert/strict')
+const test = require('node:test')
+
+const {
+  decodeClipboardImageBase64,
+  encodePowerShellCommand,
+  powershellCandidates,
+  readWslWindowsClipboardImage
+} = require('./wsl-clipboard-image.cjs')
+
+const PNG_SIGNATURE = Buffer.from([0x89, 0x50, 0x4e, 0x47, 0x0d, 0x0a, 0x1a, 0x0a])
+
+function fakePngBuffer(extraBytes = 16) {
+  return Buffer.concat([PNG_SIGNATURE, Buffer.alloc(extraBytes, 0x42)])
+}
+
+test('encodePowerShellCommand produces UTF-16LE base64 PowerShell can decode', () => {
+  const encoded = encodePowerShellCommand('Write-Output "hi"')
+  const roundTripped = Buffer.from(encoded, 'base64').toString('utf16le')
+  assert.equal(roundTripped, 'Write-Output "hi"')
+})
+
+test('decodeClipboardImageBase64 returns a Buffer for valid PNG base64', () => {
+  const png = fakePngBuffer()
+  const decoded = decodeClipboardImageBase64(png.toString('base64'))
+  assert.ok(Buffer.isBuffer(decoded))
+  assert.ok(decoded.equals(png))
+})
+
+test('decodeClipboardImageBase64 trims surrounding whitespace before decoding', () => {
+  const png = fakePngBuffer()
+  const decoded = decodeClipboardImageBase64(`\n  ${png.toString('base64')}  \r\n`)
+  assert.ok(decoded && decoded.equals(png))
+})
+
+test('decodeClipboardImageBase64 returns null for empty / whitespace input', () => {
+  assert.equal(decodeClipboardImageBase64(''), null)
+  assert.equal(decodeClipboardImageBase64('   \n  '), null)
+  assert.equal(decodeClipboardImageBase64(null), null)
+  assert.equal(decodeClipboardImageBase64(undefined), null)
+})
+
+test('decodeClipboardImageBase64 rejects base64 without a PNG signature', () => {
+  // Valid base64, but the decoded bytes are not a PNG.
+  const notPng = Buffer.from('this is not a png at all').toString('base64')
+  assert.equal(decodeClipboardImageBase64(notPng), null)
+})
+
+test('readWslWindowsClipboardImage decodes the first candidate that returns a PNG', () => {
+  const png = fakePngBuffer()
+  const calls = []
+  const exec = (cmd, args) => {
+    calls.push({ cmd, args })
+    return png.toString('base64')
+  }
+
+  const result = readWslWindowsClipboardImage({ exec, candidates: ['powershell.exe'] })
+  assert.ok(result && result.equals(png))
+  assert.equal(calls.length, 1)
+  assert.equal(calls[0].cmd, 'powershell.exe')
+  // -STA is mandatory for System.Windows.Forms.Clipboard.
+  assert.ok(calls[0].args.includes('-STA'))
+  assert.ok(calls[0].args.includes('-EncodedCommand'))
+})
+
+test('readWslWindowsClipboardImage returns null and stops when stdout is empty (no image)', () => {
+  let count = 0
+  const exec = () => {
+    count += 1
+    return ''
+  }
+
+  const result = readWslWindowsClipboardImage({
+    exec,
+    candidates: ['powershell.exe', '/mnt/c/Windows/System32/WindowsPowerShell/v1.0/powershell.exe']
+  })
+  assert.equal(result, null)
+  // Empty stdout means "no image on the clipboard" — don't probe further candidates.
+  assert.equal(count, 1)
+})
+
+test('readWslWindowsClipboardImage falls through to the next candidate when one throws', () => {
+  const png = fakePngBuffer()
+  const seen = []
+  const exec = cmd => {
+    seen.push(cmd)
+    if (cmd === 'powershell.exe') {
+      throw Object.assign(new Error('not found'), { code: 'ENOENT' })
+    }
+    return png.toString('base64')
+  }
+
+  const result = readWslWindowsClipboardImage({
+    exec,
+    candidates: ['powershell.exe', '/mnt/c/Windows/System32/WindowsPowerShell/v1.0/powershell.exe']
+  })
+  assert.ok(result && result.equals(png))
+  assert.deepEqual(seen, ['powershell.exe', '/mnt/c/Windows/System32/WindowsPowerShell/v1.0/powershell.exe'])
+})
+
+test('readWslWindowsClipboardImage returns null when every candidate throws', () => {
+  const exec = () => {
+    throw new Error('boom')
+  }
+
+  const result = readWslWindowsClipboardImage({ exec, candidates: ['a', 'b'] })
+  assert.equal(result, null)
+})
+
+test('powershellCandidates lists the bare name first, then the absolute fallback', () => {
+  const candidates = powershellCandidates()
+  assert.equal(candidates[0], 'powershell.exe')
+  assert.ok(candidates.some(c => c.endsWith('WindowsPowerShell/v1.0/powershell.exe')))
+})

apps/desktop/package.json

@@ -18,7 +18,7 @@
     "profile:main": "wait-on http://127.0.0.1:5174 && cross-env XCURSOR_SIZE=24 HERMES_DESKTOP_DEV_SERVER=http://127.0.0.1:5174 electron --inspect=9229 .",
     "profile:main:cpu": "wait-on http://127.0.0.1:5174 && cross-env XCURSOR_SIZE=24 NODE_OPTIONS=--cpu-prof HERMES_DESKTOP_DEV_SERVER=http://127.0.0.1:5174 electron .",
     "start": "npm run build && electron .",
-    "build": "node scripts/assert-root-install.cjs && node scripts/write-build-stamp.cjs && node scripts/stage-native-deps.cjs && tsc -b && vite build && npm run postbuild",
+    "build": "node scripts/assert-root-install.cjs && node scripts/write-build-stamp.cjs && node scripts/stage-native-deps.cjs && tsc -b && vite build &&  node scripts/bundle-electron-main.mjs && npm run postbuild",
     "postbuild": "node scripts/assert-dist-built.cjs",
     "prebuilder": "node scripts/patch-electron-builder-mac-binary.cjs",
     "builder": "cross-env NODE_OPTIONS=--max-old-space-size=16384 node scripts/run-electron-builder.cjs",
@@ -37,7 +37,7 @@
     "test:desktop:nsis": "node scripts/test-desktop.mjs nsis",
     "test:desktop:existing": "node scripts/test-desktop.mjs existing",
     "test:desktop:fresh": "node scripts/test-desktop.mjs fresh",
-    "test:desktop:platforms": "node --test electron/bootstrap-platform.test.cjs electron/hardening.test.cjs electron/backend-env.test.cjs electron/backend-probes.test.cjs electron/backend-ready.test.cjs electron/bootstrap-runner.test.cjs electron/connection-config.test.cjs electron/dashboard-token.test.cjs electron/gateway-ws-probe.test.cjs electron/oauth-net-request.test.cjs electron/desktop-uninstall.test.cjs electron/session-windows.test.cjs electron/link-title-window.test.cjs electron/workspace-cwd.test.cjs electron/fs-read-dir.test.cjs electron/git-root.test.cjs electron/windows-child-process.test.cjs electron/update-remote.test.cjs electron/update-rebuild.test.cjs electron/update-marker.test.cjs electron/update-relaunch.test.cjs electron/windows-user-env.test.cjs",
+    "test:desktop:platforms": "node --test electron/bootstrap-platform.test.cjs electron/hardening.test.cjs electron/backend-env.test.cjs electron/backend-probes.test.cjs electron/backend-ready.test.cjs electron/bootstrap-runner.test.cjs electron/connection-config.test.cjs electron/dashboard-token.test.cjs electron/gateway-ws-probe.test.cjs electron/oauth-net-request.test.cjs electron/desktop-uninstall.test.cjs electron/session-windows.test.cjs electron/link-title-window.test.cjs electron/workspace-cwd.test.cjs electron/fs-read-dir.test.cjs electron/git-root.test.cjs electron/git-worktree-ops.test.cjs electron/windows-child-process.test.cjs electron/update-remote.test.cjs electron/update-count.test.cjs electron/update-rebuild.test.cjs electron/update-marker.test.cjs electron/update-relaunch.test.cjs electron/windows-user-env.test.cjs electron/wsl-clipboard-image.test.cjs electron/titlebar-overlay-width.test.cjs electron/window-state.test.cjs",
     "typecheck": "tsc -p . --noEmit",
     "lint": "eslint src/ electron/",
     "lint:fix": "eslint src/ electron/ --fix",
@@ -51,11 +51,17 @@
     "@assistant-ui/react-streamdown": "^0.1.11",
     "@audiowave/react": "^0.6.2",
     "@chenglou/pretext": "^0.0.6",
+    "@codemirror/commands": "^6.10.4",
+    "@codemirror/language": "^6.12.4",
+    "@codemirror/language-data": "^6.5.2",
+    "@codemirror/state": "^6.7.0",
+    "@codemirror/view": "^6.43.3",
     "@dnd-kit/core": "^6.3.1",
     "@dnd-kit/sortable": "^10.0.0",
     "@dnd-kit/utilities": "^3.2.2",
     "@hermes/shared": "file:../shared",
     "@icons-pack/react-simple-icons": "=13.11.1",
+    "@lezer/highlight": "^1.2.3",
     "@nanostores/react": "^1.1.0",
     "@nous-research/ui": "^0.13.0",
     "@radix-ui/react-slot": "^1.2.4",
@@ -75,11 +81,13 @@
     "clsx": "^2.1.1",
     "cmdk": "^1.1.1",
     "dnd-core": "^14.0.1",
+    "dompurify": "^3.4.11",
     "hast-util-from-html-isomorphic": "^2.0.0",
     "hast-util-to-text": "^4.0.2",
     "ignore": "^7.0.5",
     "katex": "^0.16.45",
     "leva": "^0.10.1",
+    "mermaid": "^11.15.0",
     "motion": "^12.38.0",
     "nanostores": "^1.3.0",
     "node-pty": "1.1.0",
@@ -93,6 +101,7 @@
     "remark-math": "^6.0.0",
     "remend": "^1.3.0",
     "shiki": "^4.0.2",
+    "simple-git": "^3.36.0",
     "streamdown": "^2.5.0",
     "tailwind-merge": "^3.5.0",
     "tailwindcss": "^4.2.4",