fix: persist non-NULL system prompt on fresh turn setup (#45499)

build_turn_context() created the DB session row via _ensure_db_session()
before the system prompt was restored/built, so a fresh API/gateway agent
carrying client-managed history inserted a row with system_prompt=NULL. That
tripped the misleading 'stored system prompt is null; rebuilding from scratch
... investigate the previous turn's write path' warning and a guaranteed
first-turn prefix cache miss. Move row creation to after _cached_system_prompt
is populated.

Verified live (OpenRouter + claude-sonnet-4.5): persistent-agent turns show
cache_read jumping to the full prefix on turn 2+ (write 24411 -> read 24411),
and the persisted system_prompt is non-NULL so fresh-agent restore keeps the
prefix cache warm.

Tests: turn-context ordering regression asserting _ensure_db_session runs
after _cached_system_prompt is populated.

.github/workflows/ci.yml

@@ -12,7 +12,6 @@ name: CI
 
 on:
   pull_request:
-    branches: [main]
   push:
     branches: [main]
 

.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]

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

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")
@@ -1506,6 +1518,7 @@ 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 +1536,35 @@ 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 +1654,27 @@ 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/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/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_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
@@ -1396,10 +1397,12 @@ 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,
@@ -2663,10 +2666,12 @@ 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,
@@ -3537,16 +3542,38 @@ 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 +3583,12 @@ 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 +4079,19 @@ 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 +4521,10 @@ 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 +4536,44 @@ 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/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
@@ -177,6 +178,167 @@ 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 _delegate_task_goal_parts(tasks: Any, *, per_goal_len: int) -> tuple[int, list[str]]:
     if not isinstance(tasks, list):
         return 0, []
@@ -253,6 +415,23 @@ 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 ''}\""
@@ -943,7 +1122,7 @@ 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 +1130,7 @@ 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":

agent/learn_prompt.py

@@ -0,0 +1,133 @@
+#!/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: the human you are authoring this for, first; "Hermes Agent" second.
+  Never credit only the tool.
+- 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/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 +1,618 @@
+"""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
+
+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 detect_terminal_graphics() -> str:
+    """Best-effort detection of the richest graphics protocol available.
+
+    Env-based (non-blocking — we never issue a DA1/terminal query that could
+    hang a pipe).  Returns one of ``kitty`` / ``iterm`` / ``sixel`` /
+    ``unicode``.  Conservative: unknown terminals get ``unicode``, which works
+    anywhere with truecolor.
+    """
+    term = os.environ.get("TERM", "").lower()
+    term_program = os.environ.get("TERM_PROGRAM", "").lower()
+
+    # The VS Code / Cursor integrated terminal sets TERM_PROGRAM=vscode
+    # authoritatively but does NOT scrub the terminal env vars it inherits when
+    # launched from another emulator (ITERM_SESSION_ID, KITTY_WINDOW_ID, …).
+    # Trusting those leaks emits an image protocol the embedded xterm.js can't
+    # display — you get a blank frame. Inline images there are opt-in
+    # (terminal.integrated.enableImages), so default to half-blocks, which
+    # always render in its truecolor grid. Users who enabled images can pin
+    # display.pet.render_mode explicitly.
+    if term_program == "vscode":
+        return "unicode"
+
+    # kitty graphics protocol
+    if os.environ.get("KITTY_WINDOW_ID") or "kitty" in term or "ghostty" in term:
+        return "kitty"
+    if term_program in {"ghostty"}:
+        return "kitty"
+
+    # WezTerm speaks both kitty and iterm; prefer kitty (richer placement).
+    if term_program == "wezterm" or os.environ.get("WEZTERM_PANE"):
+        return "kitty"
+
+    # iTerm2 inline images
+    if term_program == "iterm.app" or os.environ.get("ITERM_SESSION_ID"):
+        return "iterm"
+
+    # sixel-capable terminals (env heuristics only)
+    if term_program in {"mintty"} or "foot" in term or "mlterm" in term:
+        return "sixel"
+    if "sixel" in term:
+        return "sixel"
+
+    return "unicode"
+
+
+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/prompt_builder.py

@@ -709,7 +709,24 @@ 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/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/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

@@ -69,12 +69,35 @@ 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 +302,11 @@ 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 ───────────────────────
@@ -581,13 +609,40 @@ 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
@@ -768,6 +823,11 @@ 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 +863,16 @@ 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
@@ -1402,6 +1465,11 @@ 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 +1496,11 @@ 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

@@ -166,6 +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 +1,164 @@
+"""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
+
+
+def verify_on_stop_enabled(config: dict[str, Any] | None = None) -> bool:
+    """Return whether edit -> verify-before-finish behavior is enabled."""
+    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
+    if isinstance(agent_cfg, dict) and "verify_on_stop" in agent_cfg:
+        return bool(agent_cfg.get("verify_on_stop"))
+    return True
+
+
+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/electron/main.cjs

@@ -12,6 +12,7 @@ const {
   powerMonitor,
   protocol,
   safeStorage,
+  screen,
   session,
   shell,
   systemPreferences
@@ -56,6 +57,7 @@ const {
 const { gitRootForIpc } = require('./git-root.cjs')
 const { worktreesForIpc } = require('./git-worktrees.cjs')
 const { OFFICIAL_REPO_HTTPS_URL, isOfficialSshRemote } = require('./update-remote.cjs')
+const { resolveBehindCount, shouldCountCommits } = require('./update-count.cjs')
 const { runRebuildWithRetry } = require('./update-rebuild.cjs')
 const {
   buildPosixCleanupScript,
@@ -67,6 +69,13 @@ const {
   uninstallArgsForMode
 } = require('./desktop-uninstall.cjs')
 const { isPackagedInstallPath: isPackagedInstallPathUnderRoots } = require('./workspace-cwd.cjs')
+const {
+  MIN_WIDTH: WINDOW_MIN_WIDTH,
+  MIN_HEIGHT: WINDOW_MIN_HEIGHT,
+  sanitizeWindowState,
+  computeWindowOptions,
+  debounce
+} = require('./window-state.cjs')
 const {
   authModeFromStatus,
   buildGatewayWsUrl,
@@ -320,6 +329,7 @@ const BOOTSTRAP_MARKER_SCHEMA_VERSION = 1
 
 const DESKTOP_CONNECTION_CONFIG_PATH = path.join(app.getPath('userData'), 'connection.json')
 const DESKTOP_UPDATE_CONFIG_PATH = path.join(app.getPath('userData'), 'updates.json')
+const DESKTOP_WINDOW_STATE_PATH = path.join(app.getPath('userData'), 'window-state.json')
 // active-profile.json records which Hermes profile the desktop launches its
 // local backend as. When set, startHermes() passes `hermes --profile <name>
 // dashboard …`, which deterministically pins HERMES_HOME (see
@@ -1522,6 +1532,36 @@ function writeDesktopUpdateConfig(config) {
   writeFileAtomic(DESKTOP_UPDATE_CONFIG_PATH, JSON.stringify(config, null, 2))
 }
 
+// ─── Main-window geometry persistence (window-state.json) ──────────────────
+
+function readWindowState() {
+  try {
+    return sanitizeWindowState(JSON.parse(fs.readFileSync(DESKTOP_WINDOW_STATE_PATH, 'utf8')))
+  } catch {
+    return null
+  }
+}
+
+// Persist the window's restored (non-maximized) bounds plus its maximized flag.
+// getNormalBounds() keeps the pre-maximize size, so un-maximizing next session
+// lands back where the user actually sized the window.
+function persistWindowState() {
+  if (!mainWindow || mainWindow.isDestroyed() || mainWindow.isMinimized()) return
+  try {
+    const { x, y, width, height } = mainWindow.getNormalBounds()
+    fs.mkdirSync(path.dirname(DESKTOP_WINDOW_STATE_PATH), { recursive: true })
+    writeFileAtomic(
+      DESKTOP_WINDOW_STATE_PATH,
+      JSON.stringify({ x, y, width, height, isMaximized: mainWindow.isMaximized() }, null, 2)
+    )
+  } catch (err) {
+    rememberLog(`[window-state] persist failed: ${err?.message || err}`)
+  }
+}
+
+// resized/moved fire many times mid-drag on Linux; debounce to one write.
+const schedulePersistWindowState = debounce(persistWindowState, 250)
+
 // Match the backend's source resolution but bias toward a real git checkout.
 // Dev → SOURCE_REPO_ROOT. Packaged/CLI install → ACTIVE_HERMES_ROOT.
 // HERMES_DESKTOP_HERMES_ROOT always wins so devs can pin a worktree.
@@ -1667,15 +1707,34 @@ async function checkUpdates() {
   }
 
   const git = args => runGit(args, { cwd: updateRoot }).then(r => r.stdout.trim())
-  const [currentSha, targetSha, countStr, dirtyStr, currentBranch] = await Promise.all([
+  const [currentSha, targetSha, dirtyStr, currentBranch, shallowStr, mergeBaseStr] = await Promise.all([
     git(['rev-parse', 'HEAD']),
     git(['rev-parse', `origin/${branch}`]),
-    git(['rev-list', `HEAD..origin/${branch}`, '--count']),
     git(['status', '--porcelain']),
-    git(['rev-parse', '--abbrev-ref', 'HEAD'])
+    git(['rev-parse', '--abbrev-ref', 'HEAD']),
+    git(['rev-parse', '--is-shallow-repository']),
+    // merge-base exits non-zero with empty stdout when HEAD shares no common
+    // ancestor with the freshly fetched tip — exactly the shallow-clone case.
+    git(['merge-base', 'HEAD', `origin/${branch}`])
   ])
 
-  const behind = Number.parseInt(countStr, 10) || 0
+  const isShallow = shallowStr === 'true'
+  const hasMergeBase = Boolean(mergeBaseStr)
+  // Only enumerate the commit count when it is meaningful. On a shallow checkout
+  // with no merge-base, `rev-list --count` walks the entire remote ancestry
+  // (thousands of commits, see #51922) and resolveBehindCount discards the
+  // result anyway in favour of a SHA compare — so skip the expensive query.
+  const countStr = shouldCountCommits({ isShallow, hasMergeBase })
+    ? await git(['rev-list', `HEAD..origin/${branch}`, '--count'])
+    : ''
+
+  const behind = resolveBehindCount({
+    countStr,
+    currentSha,
+    targetSha,
+    isShallow,
+    hasMergeBase
+  })
   const commits = behind > 0 ? await readCommitLog(updateRoot, branch) : []
 
   return {
@@ -5385,13 +5444,149 @@ function createNewSessionWindow() {
   return spawnSecondaryWindow({ newSession: true })
 }
 
+// The pet overlay: a single transparent, frameless, always-on-top window that
+// hosts ONLY the floating mascot. Shift-clicking the in-window pet "pops it out"
+// here so it can leave the app's bounds and stay visible while Hermes is
+// minimized (Codex-style task-completion glance). It carries no gateway
+// connection of its own — the main renderer is the single source of truth and
+// pushes pet state over IPC (hermes:pet-overlay:state); the overlay just renders
+// it. Control flows back (pop-in, composer submit) via hermes:pet-overlay:control.
+let petOverlayWindow = null
+
+function petOverlayUrl() {
+  if (DEV_SERVER) {
+    return `${DEV_SERVER.endsWith('/') ? DEV_SERVER.slice(0, -1) : DEV_SERVER}/?win=overlay#/`
+  }
+
+  return `${pathToFileURL(resolveRendererIndex()).toString()}?win=overlay#/`
+}
+
+function spawnPetOverlayWindow(bounds) {
+  const win = new BrowserWindow({
+    width: Math.max(80, Math.round(bounds?.width || 220)),
+    height: Math.max(80, Math.round(bounds?.height || 220)),
+    x: Number.isFinite(bounds?.x) ? Math.round(bounds.x) : undefined,
+    y: Number.isFinite(bounds?.y) ? Math.round(bounds.y) : undefined,
+    frame: false,
+    transparent: true,
+    resizable: false,
+    movable: true,
+    minimizable: false,
+    maximizable: false,
+    fullscreenable: false,
+    // Windows/Linux need this so the helper window does not get its own
+    // taskbar/alt-tab entry. On macOS, cmd-tab is app-level and this can make
+    // the whole app look like it vanished when the only newly-created visible
+    // window is a frameless overlay. Use NSPanel + Mission Control hiding below
+    // instead, leaving the main Hermes app as the Dock/cmd-tab anchor.
+    skipTaskbar: !IS_MAC,
+    hasShadow: false,
+    alwaysOnTop: true,
+    // macOS panels are non-activating helper windows and can float over full
+    // screen spaces without becoming the app's main switcher window.
+    type: IS_MAC ? 'panel' : undefined,
+    hiddenInMissionControl: IS_MAC,
+    // Non-activating: the overlay must never become the app's key/main window,
+    // or it (a frameless, taskbar-skipping panel) becomes the app's switcher
+    // anchor and the Hermes icon drops out of cmd/alt-tab — especially when the
+    // main window is minimized. We flip this on only while the composer needs
+    // the keyboard (see hermes:pet-overlay:set-focusable).
+    focusable: false,
+    show: false,
+    // Fully transparent — the renderer paints only the sprite + bubble.
+    backgroundColor: '#00000000',
+    webPreferences: {
+      preload: path.join(__dirname, 'preload.cjs'),
+      contextIsolation: true,
+      sandbox: true,
+      nodeIntegration: false,
+      devTools: true,
+      // Keep the sprite animating + bubble updating while the main window is
+      // minimized/blurred — the whole point of the overlay.
+      backgroundThrottling: false
+    }
+  })
+
+  // Float above other apps and follow the user across desktops so the pet is
+  // always reachable. `floating` + `type: panel` is the macOS NSPanel path; the
+  // more aggressive `screen-saver` level can interfere with normal app/window
+  // switching semantics.
+  win.setAlwaysOnTop(true, IS_MAC ? 'floating' : 'screen-saver')
+  win.setHiddenInMissionControl?.(true)
+  try {
+    // Electron docs: macOS may transform process type on each
+    // setVisibleOnAllWorkspaces() call unless skipTransformProcessType=true,
+    // which briefly hides the Dock/cmd-tab presence. Keep Hermes in the normal
+    // ForegroundApplication class so shift-clicking the pet never drops the app
+    // out of app switchers.
+    win.setVisibleOnAllWorkspaces(
+      true,
+      IS_MAC ? { visibleOnFullScreen: true, skipTransformProcessType: true } : undefined
+    )
+  } catch {
+    // Not supported everywhere — best effort.
+  }
+
+  wireCommonWindowHandlers(win)
+
+  win.once('ready-to-show', () => {
+    if (!win.isDestroyed()) win.showInactive()
+  })
+
+  win.on('closed', () => {
+    if (petOverlayWindow === win) {
+      petOverlayWindow = null
+    }
+
+    // If the overlay went away on its own (e.g. ⌘W), tell the main renderer to
+    // pop the pet back in so it doesn't stay hidden. Harmless echo when we're
+    // the ones who closed it (popInPet already cleared the active flag).
+    if (mainWindow && !mainWindow.isDestroyed()) {
+      mainWindow.webContents.send('hermes:pet-overlay:control', { type: 'pop-in' })
+    }
+  })
+
+  win.loadURL(petOverlayUrl())
+
+  return win
+}
+
+function openPetOverlay(bounds) {
+  if (petOverlayWindow && !petOverlayWindow.isDestroyed()) {
+    if (bounds) {
+      petOverlayWindow.setBounds({
+        x: Math.round(bounds.x),
+        y: Math.round(bounds.y),
+        width: Math.max(80, Math.round(bounds.width)),
+        height: Math.max(80, Math.round(bounds.height))
+      })
+    }
+
+    petOverlayWindow.showInactive()
+
+    return petOverlayWindow
+  }
+
+  petOverlayWindow = spawnPetOverlayWindow(bounds)
+
+  return petOverlayWindow
+}
+
+function closePetOverlay() {
+  if (petOverlayWindow && !petOverlayWindow.isDestroyed()) {
+    petOverlayWindow.close()
+  }
+
+  petOverlayWindow = null
+}
+
 function createWindow() {
   const icon = getAppIconPath()
+  const savedWindowState = readWindowState()
   mainWindow = new BrowserWindow({
-    width: 1220,
-    height: 800,
-    minWidth: 400,
-    minHeight: 620,
+    ...computeWindowOptions(savedWindowState, screen.getAllDisplays()),
+    minWidth: WINDOW_MIN_WIDTH,
+    minHeight: WINDOW_MIN_HEIGHT,
     title: 'Hermes',
     // Frameless title bar on every platform so the renderer can paint the
     // "hide sidebar" button (and other left-side titlebar tools) flush with
@@ -5433,6 +5628,8 @@ function createWindow() {
     }
   }
 
+  if (savedWindowState?.isMaximized) mainWindow.maximize()
+
   mainWindow.once('ready-to-show', () => {
     if (mainWindow && !mainWindow.isDestroyed()) mainWindow.show()
   })
@@ -5442,6 +5639,19 @@ function createWindow() {
   mainWindow.on('will-leave-full-screen', () => sendWindowStateChanged(false))
   mainWindow.on('leave-full-screen', () => sendWindowStateChanged(false))
 
+  // Reopen where the user left off. resized/moved settle once per drag; close is
+  // the cross-platform backstop, flushed synchronously before the window is gone.
+  mainWindow.on('resized', schedulePersistWindowState)
+  mainWindow.on('moved', schedulePersistWindowState)
+  mainWindow.on('maximize', schedulePersistWindowState)
+  mainWindow.on('unmaximize', schedulePersistWindowState)
+  mainWindow.on('close', () => schedulePersistWindowState.flush())
+
+  // The overlay rides the main window — closing the app's primary window must
+  // tear it down too (otherwise it strands as an orphan that blocks
+  // window-all-closed from quitting on Windows/Linux).
+  mainWindow.on('closed', () => closePetOverlay())
+
   wireCommonWindowHandlers(mainWindow)
 
   mainWindow.webContents.on('render-process-gone', (_event, details) => {
@@ -5562,6 +5772,116 @@ ipcMain.handle('hermes:window:openNewSession', async () => {
 
   return { ok: true }
 })
+
+// --- Pet overlay (pop-out mascot) -----------------------------------------
+// `request` is `{ bounds, screen }`. A fresh pop-out passes viewport-space
+// bounds (screen=false): convert to screen space by adding the main window's
+// content origin so the pet lands where it sat in-window. A remembered/dragged
+// spot passes screen-space bounds (screen=true) and is used as-is. We return the
+// resolved screen bounds so the renderer can persist exactly where it opened.
+ipcMain.handle('hermes:pet-overlay:open', async (_event, request) => {
+  const bounds = request && request.bounds ? request.bounds : request
+  const isScreen = Boolean(request && request.screen)
+  let screenBounds = bounds
+
+  try {
+    if (bounds && !isScreen && mainWindow && !mainWindow.isDestroyed()) {
+      const content = mainWindow.getContentBounds()
+      screenBounds = {
+        x: content.x + (bounds.x || 0),
+        y: content.y + (bounds.y || 0),
+        width: bounds.width,
+        height: bounds.height
+      }
+    }
+  } catch {
+    // Fall back to raw bounds if the window geometry is unavailable.
+  }
+
+  openPetOverlay(screenBounds)
+
+  return { ok: true, bounds: screenBounds }
+})
+ipcMain.handle('hermes:pet-overlay:close', async () => {
+  closePetOverlay()
+
+  return { ok: true }
+})
+// Drag: the overlay reports a new absolute screen position (it already knows the
+// pointer's screen coords), we just move the window.
+ipcMain.on('hermes:pet-overlay:set-bounds', (_event, bounds) => {
+  if (!petOverlayWindow || petOverlayWindow.isDestroyed() || !bounds) {
+    return
+  }
+
+  petOverlayWindow.setBounds({
+    x: Math.round(bounds.x),
+    y: Math.round(bounds.y),
+    width: Math.max(80, Math.round(bounds.width)),
+    height: Math.max(80, Math.round(bounds.height))
+  })
+})
+// Click-through: the overlay window is a full rectangle but only the pet pixels
+// should be interactive. The renderer toggles this as the cursor enters/leaves
+// the sprite so transparent margins pass clicks to whatever is behind.
+ipcMain.on('hermes:pet-overlay:ignore-mouse', (_event, ignore) => {
+  if (petOverlayWindow && !petOverlayWindow.isDestroyed()) {
+    petOverlayWindow.setIgnoreMouseEvents(Boolean(ignore), { forward: true })
+  }
+})
+// The overlay is a non-activating panel (focusable:false) so it never steals
+// the app's cmd/alt-tab anchor from the main window. But the pop-up composer
+// needs the keyboard, so the renderer asks us to flip it focusable + focus it
+// while the composer is open, then back to non-activating when it closes.
+ipcMain.on('hermes:pet-overlay:set-focusable', (_event, focusable) => {
+  if (!petOverlayWindow || petOverlayWindow.isDestroyed()) {
+    return
+  }
+
+  petOverlayWindow.setFocusable(Boolean(focusable))
+  if (focusable) {
+    petOverlayWindow.focus()
+  }
+})
+// Main renderer → overlay: forward the latest pet state for the overlay to render.
+ipcMain.on('hermes:pet-overlay:state', (_event, payload) => {
+  if (petOverlayWindow && !petOverlayWindow.isDestroyed()) {
+    petOverlayWindow.webContents.send('hermes:pet-overlay:state', payload)
+  }
+})
+// Overlay → main renderer: control messages (pop back in, composer submit).
+ipcMain.on('hermes:pet-overlay:control', (_event, payload) => {
+  if (!mainWindow || mainWindow.isDestroyed()) {
+    return
+  }
+
+  // Double-click toggles the app window: hide it away if it's up front, bring it
+  // back if it's minimized/buried. Pure window control — nothing for the
+  // renderer to do, so don't forward it.
+  if (payload && payload.type === 'toggle-app') {
+    if (mainWindow.isMinimized() || !mainWindow.isVisible()) {
+      mainWindow.show()
+      mainWindow.focus()
+    } else {
+      mainWindow.minimize()
+    }
+
+    return
+  }
+
+  // The mail icon means "take me to the app": raise the main window (it may be
+  // minimized or buried) before the renderer navigates to the latest thread.
+  if (payload && payload.type === 'open-app') {
+    if (mainWindow.isMinimized()) {
+      mainWindow.restore()
+    }
+
+    mainWindow.show()
+    mainWindow.focus()
+  }
+
+  mainWindow.webContents.send('hermes:pet-overlay:control', payload)
+})
 ipcMain.handle('hermes:bootstrap:reset', async () => {
   // Renderer's "Reload and retry" path. Clear the latched failure and
   // reset connection state so the next startHermes() call restarts the
@@ -6772,6 +7092,10 @@ function configureSpellChecker() {
 }
 
 app.on('before-quit', () => {
+  // The always-on-top overlay isn't a "real" app window; close it so a stray
+  // pet can't keep the process alive or float over a quit app.
+  closePetOverlay()
+
   // Quitting mid-install should stop the installer, not orphan it.
   if (bootstrapAbortController) {
     try {

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),

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 +1,79 @@
+'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/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 +1,135 @@
+/**
+ * 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/package.json

@@ -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/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/window-state.test.cjs",
     "typecheck": "tsc -p . --noEmit",
     "lint": "eslint src/ electron/",
     "lint:fix": "eslint src/ electron/ --fix",

apps/desktop/src/app/agents/index.tsx

@@ -9,9 +9,9 @@ import { type Translations, useI18n } from '@/i18n'
 import { AlertCircle, CheckCircle2, Sparkles } from '@/lib/icons'
 import { useEnterAnimation } from '@/lib/use-enter-animation'
 import { cn } from '@/lib/utils'
-import { $activeSessionId } from '@/store/session'
 import {
   $subagentsBySession,
+  allSubagents,
   buildSubagentTree,
   type SubagentNode,
   type SubagentStatus,
@@ -77,15 +77,12 @@ interface AgentsViewProps {
 
 export function AgentsView({ onClose }: AgentsViewProps) {
   const { t } = useI18n()
-  const activeSessionId = useStore($activeSessionId)
   const subagentsBySession = useStore($subagentsBySession)
 
-  const activeSubagents = useMemo(
-    () => (activeSessionId ? (subagentsBySession[activeSessionId] ?? []) : []),
-    [activeSessionId, subagentsBySession]
-  )
-
-  const tree = useMemo(() => buildSubagentTree(activeSubagents), [activeSubagents])
+  // Aggregate every session, matching the status-bar indicator — a subagent
+  // running in a background session must still be visible here, or the two
+  // desync ("Agents N running" vs an empty tree).
+  const tree = useMemo(() => buildSubagentTree(allSubagents(subagentsBySession)), [subagentsBySession])
 
   return (
     <OverlayView

apps/desktop/src/app/chat/composer/composer-text-guard.test.tsx

@@ -0,0 +1,106 @@
+// @vitest-environment jsdom
+import { act, cleanup, render } from '@testing-library/react'
+import { useCallback, useRef } from 'react'
+import { afterEach, describe, expect, it, vi } from 'vitest'
+
+afterEach(cleanup)
+
+// Regression repro for #49903: on desktop v0.17.0 the composer threw an
+// uncaught `Error: Composer is not available` at startup and the input went
+// unresponsive. The throw comes from @assistant-ui/core's composer-runtime —
+// every *mutator* (setText/send/…) does `if (!core) throw new Error("Composer
+// is not available")` when the thread's composer core isn't bound yet. Unlike
+// the read path (`s.composer.text`, which is null-safe: `runtime?.text ?? ""`),
+// the mutators have no graceful fallback. ChatBar's mount-time effects (draft
+// restore, clearDraft, external inserts) push text via `aui.composer().setText`
+// before the core binds, and the popout refactor (#49488) widened that window,
+// so the throw surfaced as an uncaught error that wedged the input.
+//
+// The fix wraps every `aui.composer().setText` call in a `setComposerText`
+// helper that swallows the unbound-core throw — the contentEditable DOM +
+// draftRef already hold the text and the draft⇄editor sync re-applies it once
+// the core attaches, so nothing is lost. This Harness mirrors that helper
+// faithfully (same try/catch shape) over a fake `aui` whose composer can be
+// toggled bound/unbound, the way the assistant-ui runtime behaves across mount.
+
+interface FakeComposer {
+  setText: (value: string) => void
+}
+
+// Mirror of index.tsx's `useAui()` composer surface: composer() returns a
+// runtime whose setText throws exactly like @assistant-ui/core when unbound.
+function makeFakeAui(bound: { current: boolean }, applied: string[]) {
+  const composer: FakeComposer = {
+    setText(value: string) {
+      if (!bound.current) {
+        throw new Error('Composer is not available')
+      }
+
+      applied.push(value)
+    }
+  }
+
+  return { composer: () => composer }
+}
+
+function Harness({
+  bound,
+  applied,
+  onError
+}: {
+  applied: string[]
+  bound: { current: boolean }
+  onError: (err: unknown) => void
+}) {
+  const aui = useRef(makeFakeAui(bound, applied)).current
+
+  // Verbatim mirror of the production `setComposerText` helper in index.tsx.
+  const setComposerText = useCallback(
+    (value: string) => {
+      try {
+        aui.composer().setText(value)
+      } catch {
+        // Composer core not bound yet — swallow so the input stays usable.
+      }
+    },
+    [aui]
+  )
+
+  // A draft-restore-on-mount that fires while the core may still be unbound,
+  // exactly like loadIntoComposer/clearDraft do on startup.
+  try {
+    setComposerText('restored draft')
+  } catch (err) {
+    onError(err)
+  }
+
+  return null
+}
+
+describe('setComposerText guard (#49903)', () => {
+  it('swallows the unbound-core throw at startup instead of crashing the renderer', () => {
+    const applied: string[] = []
+    const bound = { current: false }
+    const onError = vi.fn()
+
+    expect(() => render(<Harness applied={applied} bound={bound} onError={onError} />)).not.toThrow()
+
+    // The guard absorbed the throw — nothing escaped to the renderer, and no
+    // assistant-ui write landed (core was unbound).
+    expect(onError).not.toHaveBeenCalled()
+    expect(applied).toEqual([])
+  })
+
+  it('writes through to the composer once the core is bound', () => {
+    const applied: string[] = []
+    const bound = { current: true }
+    const onError = vi.fn()
+
+    act(() => {
+      render(<Harness applied={applied} bound={bound} onError={onError} />)
+    })
+
+    expect(onError).not.toHaveBeenCalled()
+    expect(applied).toEqual(['restored draft'])
+  })
+})

apps/desktop/src/app/chat/composer/focus.ts

@@ -34,6 +34,7 @@ interface InsertRefsDetail {
 const FOCUS_EVENT = 'hermes:composer-focus'
 const INSERT_EVENT = 'hermes:composer-insert'
 const INSERT_REFS_EVENT = 'hermes:composer-insert-refs'
+const VOICE_TOGGLE_EVENT = 'hermes:composer-voice-toggle'
 
 let activeTarget: ComposerTarget = 'main'
 
@@ -105,6 +106,13 @@ export const requestComposerInsertRefs = (
 export const onComposerInsertRefsRequest = (handler: (detail: InsertRefsDetail) => void) =>
   subscribe<InsertRefsDetail>(INSERT_REFS_EVENT, handler)
 
+/** Toggle the active composer's voice conversation — the `composer.voice`
+ *  hotkey (Ctrl+B) reaching into the composer that owns the voice state. */
+export const requestVoiceToggle = () => dispatch<{ at: number }>(VOICE_TOGGLE_EVENT, { at: Date.now() })
+
+export const onComposerVoiceToggleRequest = (handler: () => void) =>
+  subscribe<{ at: number }>(VOICE_TOGGLE_EVENT, () => handler())
+
 /**
  * Focus a composer input across React commit + browser focus restore.
  *

apps/desktop/src/app/chat/composer/index.tsx

@@ -79,7 +79,8 @@ import {
   markActiveComposer,
   onComposerFocusRequest,
   onComposerInsertRefsRequest,
-  onComposerInsertRequest
+  onComposerInsertRequest,
+  onComposerVoiceToggleRequest
 } from './focus'
 import { HelpHint } from './help-hint'
 import { useAtCompletions } from './hooks/use-at-completions'
@@ -193,6 +194,32 @@ export function ChatBar({
 }: ChatBarProps) {
   const aui = useAui()
   const draft = useAuiState(s => s.composer.text)
+
+  // assistant-ui's composer *mutators* (setText/send/…) throw "Composer is not
+  // available" when the thread's composer core isn't bound yet — and unlike the
+  // read path (`s.composer.text`, which is null-safe), there's no graceful
+  // fallback. There's a startup/thread-swap window where this ChatBar's mount
+  // effects (draft restore, clearDraft, external inserts) run before the core
+  // binds; the popout refactor (#49488) widened it by moving the composer out
+  // of the contain wrapper into a sibling of the thread, so the throw began
+  // surfacing as an uncaught error that wedged the desktop input (#49903).
+  //
+  // Guard every mutation: if the core isn't ready, no-op the assistant-ui write.
+  // The contentEditable DOM + draftRef already hold the text, and the
+  // draft⇄editor sync reconciles composer state once the core attaches, so the
+  // draft is never lost — only the (premature) state push is skipped.
+  const setComposerText = useCallback(
+    (value: string) => {
+      try {
+        aui.composer().setText(value)
+      } catch {
+        // Composer core not bound yet — DOM/draftRef carry the text; the sync
+        // effect re-applies it after bind. Swallow so the input stays usable.
+      }
+    },
+    [aui]
+  )
+
   const attachments = useStore($composerAttachments)
   const queuedPromptsBySession = useStore($queuedPromptsBySession)
   const statusItemsBySession = useStore($statusItemsBySession)
@@ -370,7 +397,7 @@ export function ChatBar({
       const next = `${base}${sep}${value}`
 
       draftRef.current = next
-      aui.composer().setText(next)
+      setComposerText(next)
 
       const editor = editorRef.current
 
@@ -381,7 +408,7 @@ export function ChatBar({
 
       setFocusRequestId(id => id + 1)
     },
-    [aui]
+    [setComposerText]
   )
 
   useEffect(() => {
@@ -591,7 +618,7 @@ export function ChatBar({
     const nextDraft = `${currentDraft}${sep}${text}`
 
     draftRef.current = nextDraft
-    aui.composer().setText(nextDraft)
+    setComposerText(nextDraft)
 
     // Push the new text into the contentEditable editor directly. Setting the
     // assistant-ui composer state alone is not enough: the draft→editor sync
@@ -624,7 +651,7 @@ export function ChatBar({
     }
 
     draftRef.current = nextDraft
-    aui.composer().setText(nextDraft)
+    setComposerText(nextDraft)
     requestMainFocus()
 
     return true
@@ -710,7 +737,7 @@ export function ChatBar({
 
     if (nextDraft !== draftRef.current) {
       draftRef.current = nextDraft
-      aui.composer().setText(nextDraft)
+      setComposerText(nextDraft)
     }
 
     window.setTimeout(refreshTrigger, 0)
@@ -836,7 +863,7 @@ export function ChatBar({
       renderComposerContents(editor, prefix)
       placeCaretEnd(editor)
       draftRef.current = composerPlainText(editor)
-      aui.composer().setText(draftRef.current)
+      setComposerText(draftRef.current)
       closeTrigger()
       runAction()
       requestMainFocus()
@@ -864,7 +891,7 @@ export function ChatBar({
 
     const finish = () => {
       draftRef.current = composerPlainText(editor)
-      aui.composer().setText(draftRef.current)
+      setComposerText(draftRef.current)
       requestMainFocus()
       keepTriggerOpen ? window.setTimeout(refreshTrigger, 0) : closeTrigger()
     }
@@ -1316,17 +1343,17 @@ export function ChatBar({
   }
 
   const clearDraft = useCallback(() => {
-    aui.composer().setText('')
+    setComposerText('')
     draftRef.current = ''
 
     if (editorRef.current) {
       editorRef.current.replaceChildren()
     }
-  }, [aui])
+  }, [setComposerText])
 
   const loadIntoComposer = (text: string, attachments: ComposerAttachment[]) => {
     draftRef.current = text
-    aui.composer().setText(text)
+    setComposerText(text)
     $composerAttachments.set(cloneAttachments(attachments))
 
     const editor = editorRef.current
@@ -1699,7 +1726,7 @@ export function ChatBar({
 
       if (domText !== draftRef.current) {
         draftRef.current = domText
-        aui.composer().setText(domText)
+        setComposerText(domText)
       }
     }
 
@@ -1818,6 +1845,24 @@ export function ChatBar({
     pendingResponse
   })
 
+  // The `composer.voice` hotkey (Ctrl+B) toggles the conversation. Starting
+  // with STT unconfigured lets the conversation surface its own "configure
+  // speech-to-text" notice rather than silently no-opping.
+  const toggleVoiceConversation = useCallback(() => {
+    if (disabled) {
+      return
+    }
+
+    if (voiceConversationActive) {
+      setVoiceConversationActive(false)
+      void conversation.end()
+    } else {
+      setVoiceConversationActive(true)
+    }
+  }, [conversation, disabled, voiceConversationActive])
+
+  useEffect(() => onComposerVoiceToggleRequest(toggleVoiceConversation), [toggleVoiceConversation])
+
   const contextMenu = (
     <ContextMenu
       onInsertText={insertText}

apps/desktop/src/app/command-palette/index.tsx

@@ -20,6 +20,7 @@ import {
   Clock,
   Cpu,
   Download,
+  Egg,
   Globe,
   type IconComponent,
   Info,
@@ -29,6 +30,7 @@ import {
   Moon,
   Package,
   Palette,
+  PawPrint,
   Plus,
   RefreshCw,
   Settings,
@@ -40,8 +42,9 @@ import {
   Zap
 } from '@/lib/icons'
 import { cn } from '@/lib/utils'
-import { $commandPaletteOpen, closeCommandPalette, setCommandPaletteOpen } from '@/store/command-palette'
+import { $commandPaletteOpen, $commandPalettePage, closeCommandPalette, setCommandPaletteOpen } from '@/store/command-palette'
 import { $bindings } from '@/store/keybinds'
+import { openPetGenerate } from '@/store/pet-generate'
 import { runGatewayRestart } from '@/store/system-actions'
 import { luminance } from '@/themes/color'
 import { type ThemeMode, useTheme } from '@/themes/context'
@@ -64,6 +67,7 @@ import { fieldCopyForSchemaKey } from '../settings/field-copy'
 import { prettyName } from '../settings/helpers'
 
 import { MarketplaceThemePage } from './marketplace-theme-page'
+import { PetInlineToggle, PetPalettePage } from './pet-palette-page'
 
 interface PaletteItem {
   /** Keybind action id — its live combo renders as a hotkey hint. */
@@ -207,6 +211,7 @@ function themeSupportsMode(name: string, target: 'light' | 'dark'): boolean {
 export function CommandPalette() {
   const { t } = useI18n()
   const open = useStore($commandPaletteOpen)
+  const pendingPage = useStore($commandPalettePage)
   const bindings = useStore($bindings)
   const navigate = useNavigate()
   const { availableThemes, resolvedMode, setMode, setTheme, themeName } = useTheme()
@@ -252,6 +257,14 @@ export function CommandPalette() {
     }
   }, [open])
 
+  // Deep-link into a nested page (e.g. `/pet list` → pets picker).
+  useEffect(() => {
+    if (open && pendingPage) {
+      setPage(pendingPage)
+      $commandPalettePage.set(null)
+    }
+  }, [open, pendingPage])
+
   const go = useCallback((path: string) => () => navigate(path), [navigate])
 
   // Step up one nested page (or back to the root list), clearing the filter so
@@ -391,6 +404,20 @@ export function CommandPalette() {
             keywords: ['appearance', 'color mode', 'brightness', 'dark', 'light', 'system'],
             label: cc.changeColorMode,
             to: 'color-mode'
+          },
+          {
+            icon: PawPrint,
+            id: 'appearance-pets',
+            keywords: ['pet', 'petdex', 'mascot', 'pets', '/pet', 'paw'],
+            label: cc.pets.title,
+            to: 'pets'
+          },
+          {
+            icon: Egg,
+            id: 'appearance-generate-pet',
+            keywords: ['pet', 'generate', 'create', 'make', 'new pet', 'mascot', 'hatch', 'ai'],
+            label: cc.generatePet.title,
+            run: () => openPetGenerate()
           }
         ]
       },
@@ -559,6 +586,12 @@ export function CommandPalette() {
           }
         ]
       },
+      // Server-driven page: browse petdex gallery, adopt/switch, toggle off.
+      pets: {
+        title: t.commandCenter.pets.title,
+        placeholder: t.commandCenter.pets.placeholder,
+        groups: []
+      },
       // Server-driven page: items come from the Marketplace, rendered by
       // <MarketplaceThemePage> (loader + live search + per-row install).
       'install-theme': {
@@ -629,49 +662,57 @@ export function CommandPalette() {
                   event.preventDefault()
                   event.stopPropagation()
                   goBack()
+
+                  return
                 }
               }}
               onValueChange={setSearch}
               placeholder={placeholder}
+              right={page === 'pets' ? <PetInlineToggle /> : undefined}
               value={search}
             />
             <CommandList className="dt-portal-scrollbar max-h-[min(20rem,56vh)]">
-              {page === 'install-theme' ? (
+              {/* Server-driven pages render their own list; the rest show groups. */}
+              {page === 'pets' ? (
+                <PetPalettePage onGenerate={() => { closeCommandPalette(); openPetGenerate() }} search={search} />
+              ) : page === 'install-theme' ? (
                 <MarketplaceThemePage onPickTheme={setTheme} search={search} />
               ) : (
-                <CommandEmpty>{t.commandCenter.noResults}</CommandEmpty>
-              )}
-              {visibleGroups.map((group, index) => (
-                <CommandGroup
-                  className={HUD_HEADING}
-                  heading={group.heading}
-                  key={group.heading ?? `palette-group-${index}`}
-                >
-                  {group.items.map(item => {
-                    const Icon = item.icon
-                    const combo = item.action ? bindings[item.action]?.[0] : undefined
+                <>
+                  <CommandEmpty>{t.commandCenter.noResults}</CommandEmpty>
+                  {visibleGroups.map((group, index) => (
+                    <CommandGroup
+                      className={HUD_HEADING}
+                      heading={group.heading}
+                      key={group.heading ?? `palette-group-${index}`}
+                    >
+                      {group.items.map(item => {
+                        const Icon = item.icon
+                        const combo = item.action ? bindings[item.action]?.[0] : undefined
 
-                    return (
-                      <CommandItem
-                        className={cn(HUD_ITEM, HUD_TEXT)}
-                        key={item.id}
-                        keywords={item.keywords}
-                        onSelect={() => handleSelect(item)}
-                        value={`${item.label} ${item.keywords?.join(' ') ?? ''} ${item.id}`}
-                      >
-                        <Icon className="size-3.5 shrink-0 text-muted-foreground" />
-                        <span className="truncate">{item.label}</span>
-                        {combo && <KbdCombo className="ml-auto opacity-55" combo={combo} size="sm" />}
-                        {item.to && (
-                          <ChevronRight
-                            className={cn('size-3.5 shrink-0 text-muted-foreground/70', !combo && 'ml-auto')}
-                          />
-                        )}
-                      </CommandItem>
-                    )
-                  })}
-                </CommandGroup>
-              ))}
+                        return (
+                          <CommandItem
+                            className={cn(HUD_ITEM, HUD_TEXT)}
+                            key={item.id}
+                            keywords={item.keywords}
+                            onSelect={() => handleSelect(item)}
+                            value={`${item.label} ${item.keywords?.join(' ') ?? ''} ${item.id}`}
+                          >
+                            <Icon className="size-3.5 shrink-0 text-muted-foreground" />
+                            <span className="truncate">{item.label}</span>
+                            {combo && <KbdCombo className="ml-auto opacity-55" combo={combo} size="sm" />}
+                            {item.to && (
+                              <ChevronRight
+                                className={cn('size-3.5 shrink-0 text-muted-foreground/70', !combo && 'ml-auto')}
+                              />
+                            )}
+                          </CommandItem>
+                        )
+                      })}
+                    </CommandGroup>
+                  ))}
+                </>
+              )}
             </CommandList>
           </Command>
         </DialogPrimitive.Content>

apps/desktop/src/app/command-palette/pet-palette-page.tsx

@@ -0,0 +1,212 @@
+/**
+ * Cmd-K "Pets…" page — browse the petdex gallery, adopt/switch, toggle off.
+ *
+ * A thin view over the `pet-gallery` store: it subscribes to the shared atoms
+ * and calls the store's actions. The store owns fetching, caching, the thumb
+ * cache, and optimistic mutations, so reopening this page is instant and a
+ * toggle never re-pulls the network gallery.
+ */
+
+import { useStore } from '@nanostores/react'
+import { useEffect, useMemo } from 'react'
+
+import { HUD_ITEM, HUD_TEXT } from '@/app/floating-hud'
+import { useGatewayRequest } from '@/app/gateway/hooks/use-gateway-request'
+import { PetThumb } from '@/components/pet/pet-thumb'
+import { useI18n } from '@/i18n'
+import { triggerHaptic } from '@/lib/haptics'
+import { Check, Egg, Loader2, PawPrint } from '@/lib/icons'
+import { cn } from '@/lib/utils'
+import {
+  $petBusy,
+  $petGallery,
+  $petGalleryError,
+  $petGalleryStatus,
+  adoptPet,
+  loadPetGallery,
+  loadPetThumb,
+  rankedGalleryPets,
+  setPetEnabled
+} from '@/store/pet-gallery'
+
+interface PetPalettePageProps {
+  search: string
+  /** Navigate to the "generate a pet" page (rendered as a header action). */
+  onGenerate?: () => void
+}
+
+export function PetPalettePage({ search, onGenerate }: PetPalettePageProps) {
+  const { t } = useI18n()
+  const copy = t.commandCenter.pets
+  const { requestGateway } = useGatewayRequest()
+
+  const gallery = useStore($petGallery)
+  const status = useStore($petGalleryStatus)
+  const error = useStore($petGalleryError)
+  const busy = useStore($petBusy)
+
+  useEffect(() => {
+    void loadPetGallery(requestGateway)
+  }, [requestGateway])
+
+  const enabled = gallery?.enabled ?? false
+  const active = gallery?.active ?? ''
+
+  const shown = useMemo(() => rankedGalleryPets(gallery, search).slice(0, 50), [gallery, search])
+
+  const adopt = (slug: string) => {
+    void adoptPet(requestGateway, slug, copy.adoptFailed).then(ok => ok && triggerHaptic('crisp'))
+  }
+
+  if (status === 'loading' && !gallery) {
+    return <Status icon={<Loader2 className="size-3.5 animate-spin" />} text={copy.loading} />
+  }
+
+  if (status === 'stale') {
+    return <Status text={copy.staleBackend} tone="error" />
+  }
+
+  if (!gallery?.pets.length && error) {
+    return <Status text={error} tone="error" />
+  }
+
+  const mutating = Boolean(busy)
+
+  return (
+    <div role="listbox">
+      {onGenerate && (
+        <button
+          className={cn(
+            'flex w-full items-center gap-2 rounded-md text-left text-foreground transition-colors hover:bg-(--chrome-action-hover)',
+            HUD_ITEM,
+            HUD_TEXT
+          )}
+          onClick={onGenerate}
+          onMouseDown={event => event.preventDefault()}
+          type="button"
+        >
+          <span className="flex size-8 shrink-0 items-center justify-center rounded-md bg-(--chrome-action-hover)">
+            <Egg className="size-4" />
+          </span>
+          <span className="font-medium">{t.commandCenter.generatePet.title}</span>
+        </button>
+      )}
+
+      {error && <p className="px-2 pb-1 pt-1.5 text-[0.6875rem] text-(--ui-red)">{error}</p>}
+
+      {shown.length === 0 ? (
+        <Status text={copy.empty} />
+      ) : (
+        shown.map(pet => {
+          const isActive = enabled && pet.slug === active
+          const isBusy = busy === pet.slug
+
+          return (
+            <button
+              className={cn(
+                'flex w-full items-center gap-2 rounded-md text-left transition-colors hover:bg-(--chrome-action-hover) disabled:opacity-60',
+                HUD_ITEM,
+                HUD_TEXT,
+                isActive && 'bg-(--chrome-action-hover)/70'
+              )}
+              disabled={mutating && !isBusy}
+              key={pet.slug}
+              onClick={() => adopt(pet.slug)}
+              onMouseDown={event => event.preventDefault()}
+              role="option"
+              type="button"
+            >
+              <PetThumb
+                alt={pet.displayName}
+                load={(slug, url) => loadPetThumb(requestGateway, slug, url)}
+                size={32}
+                slug={pet.slug}
+                url={pet.spritesheetUrl}
+              />
+              <span className="flex min-w-0 flex-col">
+                <span className="flex items-center gap-1.5">
+                  <span className="truncate font-medium">{pet.displayName}</span>
+                  {pet.generated && (
+                    <span className="shrink-0 rounded-full bg-primary/15 px-1.5 py-px text-[0.625rem] font-medium text-primary">
+                      {copy.generatedTag}
+                    </span>
+                  )}
+                </span>
+                <span className="truncate text-[0.6875rem] text-muted-foreground/80">
+                  {pet.slug}
+                  {pet.installed ? ` · ${copy.installed}` : ''}
+                </span>
+              </span>
+              <span className="ml-auto flex shrink-0 items-center text-[0.6875rem] text-muted-foreground">
+                {isBusy ? (
+                  <Loader2 className="size-3 animate-spin" />
+                ) : isActive ? (
+                  <Check className="size-3.5 text-foreground" />
+                ) : null}
+              </span>
+            </button>
+          )
+        })
+      )}
+    </div>
+  )
+}
+
+/**
+ * Single on/off toggle, rendered inline on the palette's search row (see
+ * `CommandInput`'s `right` slot). The paw lights up when pets are on. Reads the
+ * same shared gallery atoms, so it stays in sync with the list below.
+ */
+export function PetInlineToggle() {
+  const { t } = useI18n()
+  const copy = t.commandCenter.pets
+  const { requestGateway } = useGatewayRequest()
+  const gallery = useStore($petGallery)
+  const busy = useStore($petBusy)
+
+  if (!gallery) {
+    return null
+  }
+
+  const enabled = gallery.enabled
+
+  const toggle = () => {
+    void setPetEnabled(requestGateway, !enabled, {
+      noneAvailable: copy.noneAvailable,
+      fallback: copy.toggleFailed
+    }).then(ok => ok && triggerHaptic('crisp'))
+  }
+
+  return (
+    <button
+      aria-label={enabled ? copy.turnOff : copy.turnOn}
+      aria-pressed={enabled}
+      className={cn(
+        'flex shrink-0 items-center justify-center rounded-md p-1.5 transition-colors disabled:opacity-50',
+        enabled ? 'bg-(--chrome-action-hover) text-foreground' : 'text-muted-foreground hover:bg-(--chrome-action-hover)/60'
+      )}
+      disabled={Boolean(busy)}
+      onClick={toggle}
+      // Don't steal focus from the search input on click.
+      onMouseDown={event => event.preventDefault()}
+      title={enabled ? copy.turnOff : copy.turnOn}
+      type="button"
+    >
+      {busy ? <Loader2 className="size-4 animate-spin" /> : <PawPrint className="size-4" />}
+    </button>
+  )
+}
+
+function Status({ icon, text, tone }: { icon?: React.ReactNode; text: string; tone?: 'error' }) {
+  return (
+    <div
+      className={cn(
+        'flex items-center justify-center gap-2 px-2 py-6 text-xs',
+        tone === 'error' ? 'text-(--ui-red)' : 'text-muted-foreground'
+      )}
+    >
+      {icon}
+      {text}
+    </div>
+  )
+}

apps/desktop/src/app/desktop-controller.tsx

@@ -41,6 +41,8 @@ import {
   unpinSession
 } from '../store/layout'
 import { respondToApprovalAction } from '../store/native-notifications'
+import { setPetActivity } from '../store/pet'
+import { setPetOverlayOpenAppHandler, setPetOverlaySubmitHandler } from '../store/pet-overlay'
 import { $filePreviewTarget, $previewTarget, closeActiveRightRailTab } from '../store/preview'
 import {
   $activeGatewayProfile,
@@ -52,6 +54,7 @@ import {
 } from '../store/profile'
 import {
   $activeSessionId,
+  $attentionSessionIds,
   $currentCwd,
   $freshDraftReady,
   $gatewayState,
@@ -105,6 +108,7 @@ import { useKeybinds } from './hooks/use-keybinds'
 import { SIDEBAR_COLLAPSE_MEDIA_QUERY } from './layout-constants'
 import { ModelPickerOverlay } from './model-picker-overlay'
 import { ModelVisibilityOverlay } from './model-visibility-overlay'
+import { PetGenerateOverlay } from './pet-generate/pet-generate-overlay'
 import { RightSidebarPane } from './right-sidebar'
 import { $terminalTakeover } from './right-sidebar/store'
 import { PersistentTerminal, TerminalSlot } from './right-sidebar/terminal/persistent'
@@ -841,6 +845,53 @@ export function DesktopController() {
     updateSessionState
   })
 
+  // The popped-out pet drives two actions back into the app: send a prompt, and
+  // open the most recent thread. Both are registered ONCE through refs that track
+  // the latest callbacks — re-registering on every `submitText`/`resumeSession`
+  // identity change left a brief window where the handler was nulled (cleanup
+  // before re-register), which could drop a submit fired from the overlay (e.g.
+  // creating a session from the new-session screen). The ref form keeps a stable,
+  // always-current handler. Primary window only — it owns the overlay.
+  const submitTextRef = useRef(submitText)
+  submitTextRef.current = submitText
+  const resumeSessionRef = useRef(resumeSession)
+  resumeSessionRef.current = resumeSession
+
+  useEffect(() => {
+    if (isSecondaryWindow()) {
+      return
+    }
+
+    setPetOverlaySubmitHandler(text => void submitTextRef.current(text))
+    // Mail icon: $sessions is ordered most-recent-first; the pet is global (not
+    // per session) so "most recent" is the right target. main.cjs already raised
+    // the window before forwarding this.
+    setPetOverlayOpenAppHandler(() => {
+      const recent = $sessions.get()[0]
+
+      if (recent?.id) {
+        void resumeSessionRef.current(recent.id)
+      }
+    })
+
+    return () => {
+      setPetOverlaySubmitHandler(null)
+      setPetOverlayOpenAppHandler(null)
+    }
+  }, [])
+
+  // Mirror "a session is blocked on the user" (clarify/approval) into the pet's
+  // awaitingInput flag so it shows the `waiting` pose. Lives on $petActivity so
+  // it rides the same atom the pop-out overlay mirrors — no session list needed
+  // there. Every window keeps its own in-window pet in sync.
+  useEffect(() => {
+    const sync = () => setPetActivity({ awaitingInput: $attentionSessionIds.get().length > 0 })
+
+    sync()
+
+    return $attentionSessionIds.listen(sync)
+  }, [])
+
   useGatewayBoot({
     handleGatewayEvent: handleDesktopGatewayEvent,
     onConnectionReady: c => {
@@ -978,6 +1029,7 @@ export function DesktopController() {
       <GatewayConnectingOverlay />
       <BootFailureOverlay />
       <CommandPalette />
+      <PetGenerateOverlay />
       <SessionSwitcher />
 
       {settingsOpen && (

apps/desktop/src/app/gateway/hooks/use-gateway-boot.ts

@@ -40,6 +40,13 @@ import {
 } from '@/store/session'
 import type { RpcEvent } from '@/types/hermes'
 
+// After this many consecutive failed reconnects (≈45s with the 1→15s backoff)
+// raise a recoverable boot error. Otherwise a dropped remote gateway loops the
+// backoff forever behind the fullscreen CONNECTING overlay with no way to reach
+// Settings / sign in / switch to local — the "lost connection breaks the app"
+// dead end. The next successful reconnect clears it.
+const RECONNECT_ESCALATE_AFTER = 6
+
 interface GatewayBootOptions {
   handleGatewayEvent: (event: RpcEvent) => void
   onConnectionReady: (
@@ -105,6 +112,10 @@ export function useGatewayBoot({
     // tick — a stale OAuth ticket fails every attempt and would otherwise stack
     // identical error toasts (and their haptics). Reset on the next clean open.
     let reauthNotified = false
+    // Raised once the reconnect loop crosses RECONNECT_ESCALATE_AFTER so the
+    // recovery overlay replaces the dead-end CONNECTING screen. Reset on a clean
+    // open or a manual/wake-driven reconnect.
+    let escalated = false
 
     // Wrap the live getter in a call so TS control-flow analysis doesn't narrow
     // `connectionState` to a constant across the early-return guards (the state
@@ -171,6 +182,11 @@ export function useGatewayBoot({
         reconnecting = false
 
         if (!cancelled && !gatewayOpen()) {
+          if (reconnectAttempt >= RECONNECT_ESCALATE_AFTER && !escalated) {
+            escalated = true
+            failDesktopBoot(translateNow('boot.errors.gatewayConnectionLost'))
+          }
+
           scheduleReconnect()
         }
       }
@@ -197,6 +213,7 @@ export function useGatewayBoot({
 
       clearReconnectTimer()
       reconnectAttempt = 0
+      escalated = false
       reconnectSecondaryGateways()
 
       if (!gatewayOpen()) {
@@ -230,6 +247,7 @@ export function useGatewayBoot({
       if (st === 'open') {
         reconnectAttempt = 0
         reauthNotified = false
+        escalated = false
         clearReconnectTimer()
 
         // A revalidate-driven reconnect can rebuild the backend in place when the

apps/desktop/src/app/gateway/hooks/use-gateway-request.ts

@@ -94,7 +94,7 @@ export function useGatewayRequest() {
   }, [])
 
   const requestGateway = useCallback(
-    async <T>(method: string, params: Record<string, unknown> = {}) => {
+    async <T>(method: string, params: Record<string, unknown> = {}, timeoutMs?: number, signal?: AbortSignal) => {
       const gateway = gatewayRef.current
 
       if (!gateway) {
@@ -102,7 +102,7 @@ export function useGatewayRequest() {
       }
 
       try {
-        return await gateway.request<T>(method, params)
+        return await gateway.request<T>(method, params, timeoutMs, signal)
       } catch (error) {
         const message = error instanceof Error ? error.message : String(error)
 
@@ -128,7 +128,7 @@ export function useGatewayRequest() {
           throw error
         }
 
-        return recovered.request<T>(method, params)
+        return recovered.request<T>(method, params, timeoutMs, signal)
       }
     },
     [ensureGatewayOpen]

apps/desktop/src/app/hooks/use-keybinds.ts

@@ -40,7 +40,7 @@ import {
 import { openNewSessionInNewWindow } from '@/store/windows'
 import { useTheme } from '@/themes/context'
 
-import { requestComposerFocus } from '../chat/composer/focus'
+import { requestComposerFocus, requestVoiceToggle } from '../chat/composer/focus'
 import { SIDEBAR_COLLAPSE_MEDIA_QUERY } from '../layout-constants'
 import {
   AGENTS_ROUTE,
@@ -114,6 +114,7 @@ export function useKeybinds(deps: KeybindRuntimeDeps): void {
 
     'composer.focus': () => requestComposerFocus('main'),
     'composer.modelPicker': () => setModelPickerOpen(true),
+    'composer.voice': requestVoiceToggle,
 
     'nav.commandPalette': toggleCommandPalette,
     'nav.commandCenter': deps.toggleCommandCenter,

apps/desktop/src/app/hooks/use-route-overlay-active.ts

@@ -0,0 +1,19 @@
+import { useLocation } from 'react-router-dom'
+
+import { appViewForPath, isOverlayView } from '@/app/routes'
+
+/**
+ * True while a full-screen route overlay (settings, agents, command-center, …)
+ * is showing.
+ *
+ * A portaled Radix modal sits above the app shell, so it would cover such a
+ * route. Any modal that sends the user to one (e.g. "set up image generation" →
+ * `/settings`) can `if (useRouteOverlayActive()) return null` to *yield* the
+ * screen — its open state lives in a store, so it stays open — and reappear,
+ * re-running its mount effects (a free refresh), when the route overlay closes.
+ */
+export function useRouteOverlayActive(): boolean {
+  const { pathname } = useLocation()
+
+  return isOverlayView(appViewForPath(pathname))
+}

apps/desktop/src/app/pet-generate/components/draft-grid.tsx

@@ -0,0 +1,125 @@
+import { PixelEggSprite } from '@/components/pet/pixel-egg-sprite'
+import { Button } from '@/components/ui/button'
+import { Codicon } from '@/components/ui/codicon'
+import { Tip } from '@/components/ui/tooltip'
+import { useI18n } from '@/i18n'
+import { PawPrint } from '@/lib/icons'
+import { selectableCardClass } from '@/lib/selectable-card'
+import { cn } from '@/lib/utils'
+
+const VARIANT_COUNT = 4
+
+interface DraftGridProps {
+  drafts: { index: number; dataUri: string }[]
+  generating: boolean
+  hasDrafts: boolean
+  onCancel: () => void
+  onHatch: () => void
+  onRemix: (draft: { index: number; dataUri: string }) => void
+  onSelect: (index: number) => void
+  selected: number | null
+}
+
+export function DraftGrid({
+  drafts,
+  generating,
+  hasDrafts,
+  onCancel,
+  onHatch,
+  onRemix,
+  onSelect,
+  selected
+}: DraftGridProps) {
+  const { t } = useI18n()
+  const copy = t.commandCenter.generatePet
+
+  const slots = generating
+    ? Array.from({ length: VARIANT_COUNT }, (_, i) => drafts.find(draft => draft.index === i) ?? null)
+    : drafts
+
+  return (
+    <div className="flex flex-col gap-2">
+      <div className="flex items-center justify-between text-[length:var(--conversation-caption-font-size)] text-(--ui-text-tertiary)">
+        <span className={cn(generating && 'shimmer shimmer-color-primary opacity-40', !generating && 'invisible')}>
+          {copy.generating}
+        </span>
+        <span className="tabular-nums">
+          {Math.min(drafts.length, VARIANT_COUNT)}/{VARIANT_COUNT}
+        </span>
+      </div>
+
+      <div className="grid grid-cols-2 gap-2">
+        {slots.map((draft, i) => {
+          // A streamed draft is selectable immediately — even mid-generation —
+          // so the user can commit to one without waiting for the rest.
+          const isSelected = draft != null && selected === draft.index
+
+          return (
+            <div className="group relative aspect-[192/208]" key={draft ? `draft-${draft.index}` : `slot-${i}`}>
+              <button
+                className={cn(
+                  'absolute inset-0 flex items-center justify-center overflow-hidden',
+                  selectableCardClass({ active: isSelected, prominent: true })
+                )}
+                disabled={draft == null}
+                onClick={() => draft != null && onSelect(draft.index)}
+                type="button"
+              >
+                {draft != null ? (
+                  // Hatches into place as each draft streams back.
+                  <img
+                    alt=""
+                    className="pet-reveal size-full object-contain p-1.5"
+                    draggable={false}
+                    src={draft.dataUri}
+                  />
+                ) : (
+                  // Incubating: a creme egg bouncing on its contact shadow.
+                  <div className="relative z-10 flex flex-col items-center">
+                    <PixelEggSprite index={i} mode="bounce" size={48} />
+                    <span className="pet-egg-shadow pet-egg-shadow--sm" style={{ marginTop: '-0.3rem' }} />
+                  </div>
+                )}
+              </button>
+
+              {/* Remix: branch a new round off this look. Revealed on hover/focus. */}
+              {draft != null && !generating && (
+                <Tip label={copy.remix}>
+                  <Button
+                    aria-label={copy.remix}
+                    className={cn(
+                      'absolute right-1 top-1 z-20',
+                      'text-(--ui-text-tertiary) opacity-10 transition',
+                      'hover:bg-transparent hover:text-foreground focus-visible:opacity-100 group-hover:opacity-100'
+                    )}
+                    onClick={event => {
+                      event.stopPropagation()
+                      onRemix(draft)
+                    }}
+                    size="icon-xs"
+                    type="button"
+                    variant="ghost"
+                  >
+                    <Codicon name="git-branch" size={12} />
+                  </Button>
+                </Tip>
+              )}
+            </div>
+          )
+        })}
+      </div>
+
+      {/* Same abort/go-back text link in both states (sits right under the grid);
+          once drafts land, the full-width Hatch drops in below it. */}
+      <Button className="self-center" onClick={onCancel} size="xs" variant="text">
+        {t.common.cancel}
+      </Button>
+      {hasDrafts && (
+        <Button className="w-full" disabled={selected === null} onClick={onHatch}>
+          <PawPrint />
+          {copy.hatch}
+        </Button>
+      )}
+    </div>
+  )
+}

apps/desktop/src/app/pet-generate/components/empty-hint.tsx

@@ -0,0 +1,27 @@
+import { Button } from '@/components/ui/button'
+
+interface EmptyHintProps {
+  onExample: (prompt: string) => void
+}
+
+// Creative seed prompts — specifics make better pets (petdex's own advice).
+// Short chips that wrap into a tight, centered cluster (capped width → 2 rows).
+const EXAMPLE_PROMPTS = ['bubble-tea otter', 'sock elf', 'pixel dragon', 'office cat', 'neon axolotl', 'moss golem']
+
+export function EmptyHint({ onExample }: EmptyHintProps) {
+  return (
+    <div className="flex max-w-[300px] flex-wrap place-content-center place-items-center gap-2">
+      {EXAMPLE_PROMPTS.map(example => (
+        <Button
+          className="h-auto w-fit rounded-full font-normal"
+          key={example}
+          onClick={() => onExample(`a ${example}`)}
+          size="xs"
+          variant="outline"
+        >
+          {example}
+        </Button>
+      ))}
+    </div>
+  )
+}

apps/desktop/src/app/pet-generate/components/generate-unavailable.tsx

@@ -0,0 +1,52 @@
+import { Button } from '@/components/ui/button'
+import { ExternalLink } from '@/lib/external-link'
+import { PawPrint, Settings2 } from '@/lib/icons'
+
+interface GenerateUnavailableProps {
+  onSetup: () => void
+}
+
+// Shown when no reference-capable image backend is configured: generation is
+// impossible, so we replace the prompt entirely with a friendly path to set one
+// up (in-app) plus where to grab a key.
+export function GenerateUnavailable({ onSetup }: GenerateUnavailableProps) {
+  return (
+    <div className="flex flex-col items-center gap-4 text-center">
+      <span className="grid size-11 place-items-center rounded-full bg-primary/10 text-primary">
+        <PawPrint className="size-5" />
+      </span>
+      <div className="space-y-1.5">
+        <p className="text-[length:var(--conversation-text-font-size)] font-semibold">Add an image backend to generate</p>
+        <p className="mx-auto max-w-[19rem] text-[length:var(--conversation-caption-font-size)] leading-relaxed text-(--ui-text-tertiary)">
+          Hatching a custom pet needs a provider that can ground on a reference image.
+        </p>
+      </div>
+      <Button onClick={onSetup} size="sm">
+        <Settings2 className="size-4" />
+        Set up image generation
+      </Button>
+      <p className="flex flex-wrap items-center justify-center gap-x-1.5 text-[0.6875rem] text-(--ui-text-tertiary)">
+        <span>Grab a key from</span>
+        <ExternalLink href="https://portal.nousresearch.com" showExternalIcon={false}>
+          Nous Portal
+        </ExternalLink>
+        <span>·</span>
+        <ExternalLink
+          className="opacity-40 transition-opacity hover:opacity-100"
+          href="https://openrouter.ai/keys"
+          showExternalIcon={false}
+        >
+          OpenRouter
+        </ExternalLink>
+        <span>·</span>
+        <ExternalLink
+          className="opacity-40 transition-opacity hover:opacity-100"
+          href="https://platform.openai.com/api-keys"
+          showExternalIcon={false}
+        >
+          OpenAI
+        </ExternalLink>
+      </p>
+    </div>
+  )
+}

apps/desktop/src/app/pet-generate/components/hatch-preview.tsx

@@ -0,0 +1,137 @@
+import { useEffect, useState } from 'react'
+
+import { PetSprite } from '@/components/pet/pet-sprite'
+import { PetStarShower } from '@/components/pet/pet-star-shower'
+import { PixelEggSprite } from '@/components/pet/pixel-egg-sprite'
+import { Alert, AlertDescription } from '@/components/ui/alert'
+import { Button } from '@/components/ui/button'
+import { Input } from '@/components/ui/input'
+import { useI18n } from '@/i18n'
+import { triggerHaptic } from '@/lib/haptics'
+import { Loader2, PawPrint, RefreshCw } from '@/lib/icons'
+import { type PetInfo } from '@/store/pet'
+
+import { frameCountForRow } from '../lib/frame-count'
+
+const PREVIEW_SCALE = 0.7
+const PREVIEW_STATE_MS = 1400
+
+const PREVIEW_ROWS = ['idle', 'waving', 'running-right', 'running-left', 'running', 'review', 'jumping', 'failed', 'waiting']
+
+interface HatchPreviewProps {
+  pet: PetInfo
+  adopting: boolean
+  error: string | null
+  onAdopt: (name: string) => void
+  onDiscard: () => void
+}
+
+export function HatchPreview({ pet, adopting, error, onAdopt, onDiscard }: HatchPreviewProps) {
+  const { t } = useI18n()
+  const copy = t.commandCenter.generatePet
+  // Empty so the "Name your pet" placeholder shows; blank adopt keeps the
+  // provisional name from the prompt.
+  const [name, setName] = useState('')
+  // Play the egg's crack/hatch frames once before swapping in the live pet.
+  const [revealed, setRevealed] = useState(false)
+  // Right after the egg cracks the pet plays its "yay" jump a couple times, then
+  // hands off to the normal state-cycling preview.
+  const [celebrating, setCelebrating] = useState(false)
+  const [stateIndex, setStateIndex] = useState(0)
+  const previewRows = (pet.stateRows?.length ? pet.stateRows : PREVIEW_ROWS).filter(row => frameCountForRow(pet, row) > 0)
+  const rows = previewRows.length > 0 ? previewRows : ['idle']
+  const activeRow = rows[stateIndex % rows.length] ?? 'idle'
+  const canJump = frameCountForRow(pet, 'jumping') > 0
+  const rowOverride = celebrating && canJump ? 'jumping' : activeRow
+
+  useEffect(() => {
+    const id = setInterval(() => setStateIndex(i => (i + 1) % rows.length), PREVIEW_STATE_MS)
+
+    return () => clearInterval(id)
+  }, [rows.length])
+
+  // On reveal: celebrate (jump) ~2 loops, then drop into the cycling preview.
+  useEffect(() => {
+    if (!revealed) {
+      return
+    }
+
+    setCelebrating(true)
+
+    const id = setTimeout(() => {
+      setCelebrating(false)
+      setStateIndex(0)
+    }, 2 * (pet.loopMs ?? 1100))
+
+    return () => clearTimeout(id)
+  }, [revealed, pet.loopMs])
+
+  useEffect(() => {
+    setStateIndex(0)
+    setName('')
+    setRevealed(false)
+    setCelebrating(false)
+  }, [pet.slug])
+
+  const previewInfo: PetInfo = { ...pet, scale: PREVIEW_SCALE }
+
+  return (
+    <div className="flex flex-col items-center gap-2">
+      {/* Fills the (now narrow) dialog so the pet frame is the screen width. */}
+      <div className="relative flex aspect-[192/208] w-full items-center justify-center overflow-hidden rounded-lg border border-(--ui-stroke-tertiary) bg-(--ui-bg-quinary)">
+        {revealed ? (
+          <>
+            <div className="relative inline-block">
+              <span aria-hidden className="pet-contact-shadow" />
+              <div className="pet-reveal relative z-10">
+                <PetSprite info={previewInfo} rowOverride={rowOverride} />
+              </div>
+            </div>
+            <PetStarShower />
+          </>
+        ) : (
+          // The egg cracks open, then we swap in the live pet.
+          <PixelEggSprite
+            mode="hatch"
+            onDone={() => {
+              setRevealed(true)
+              triggerHaptic('crisp')
+            }}
+            size={150}
+          />
+        )}
+      </div>
+
+      <Input
+        autoFocus
+        className="w-full"
+        onChange={event => setName(event.target.value)}
+        onKeyDown={event => {
+          if (event.key === 'Enter') {
+            event.preventDefault()
+            onAdopt(name)
+          }
+        }}
+        placeholder={copy.namePlaceholder}
+        value={name}
+      />
+
+      {error && (
+        <Alert variant="destructive">
+          <AlertDescription>{error}</AlertDescription>
+        </Alert>
+      )}
+
+      <div className="flex w-full items-center gap-1.5">
+        <Button disabled={adopting} onClick={onDiscard} variant="ghost">
+          <RefreshCw />
+          {copy.startOver}
+        </Button>
+        <Button className="flex-1" disabled={adopting} onClick={() => onAdopt(name)}>
+          {adopting ? <Loader2 className="animate-spin" /> : <PawPrint />}
+          {copy.adopt}
+        </Button>
+      </div>
+    </div>
+  )
+}

apps/desktop/src/app/pet-generate/components/hatching-view.tsx

@@ -0,0 +1,24 @@
+import { PetEggHatch } from '@/components/pet/pet-egg-hatch'
+import { useI18n } from '@/i18n'
+import { cancelHatch, type PetHatchStage } from '@/store/pet-generate'
+
+interface HatchingViewProps {
+  stage: PetHatchStage | null
+}
+
+// The hatch progress screen — a beating egg with a phase-tracking subtitle
+// (per-row → composing → saving).
+export function HatchingView({ stage }: HatchingViewProps) {
+  const { t } = useI18n()
+  const copy = t.commandCenter.generatePet
+
+  const subtitle = stage
+    ? stage.phase === 'row'
+      ? copy.hatchRow(stage.state ?? '', stage.done ?? 0, stage.total ?? 0)
+      : stage.phase === 'compose'
+        ? copy.hatchComposing
+        : copy.hatchSaving
+    : copy.hatchingSub
+
+  return <PetEggHatch cancelLabel={t.common.cancel} onCancel={cancelHatch} subtitle={subtitle} />
+}

apps/desktop/src/app/pet-generate/components/provider-picker.tsx

@@ -0,0 +1,51 @@
+import { useStore } from '@nanostores/react'
+
+import { DropdownMenu, DropdownMenuContent, DropdownMenuItem, DropdownMenuTrigger } from '@/components/ui/dropdown-menu'
+import { Check, ChevronDown } from '@/lib/icons'
+import { $petGenProvider, $petGenProviders, setPetGenProvider } from '@/store/pet-generate'
+
+// Image-backend picker for pet generation — the composer's model-pill pattern:
+// a quiet trigger + a dropdown of options. No per-option notes: every backend
+// resolves to the same faithful OpenAI image model, so there's no tradeoff to
+// describe. Hidden unless there are 2+ reference-capable backends (nothing to pick).
+export function ProviderPicker() {
+  const providers = useStore($petGenProviders)
+  const picked = useStore($petGenProvider)
+
+  if (providers.length < 2) {
+    return null
+  }
+
+  const fallback = providers.find(p => p.default) ?? providers[0]
+  const current = providers.find(p => p.name === picked) ?? fallback
+
+  return (
+    <DropdownMenu>
+      <DropdownMenuTrigger asChild>
+        {/* Plain text affordance (matches "Add a reference"), not a padded pill. */}
+        <button
+          className="flex h-6 items-center gap-1 text-[0.6875rem] text-(--ui-text-tertiary) transition hover:text-foreground"
+          type="button"
+        >
+          {current?.label}
+          <ChevronDown className="size-3" />
+        </button>
+      </DropdownMenuTrigger>
+      {/* The picker lives inside the pet-gen Dialog (z-130) and portals to body,
+          so lift its menu above the dialog or it opens behind it. */}
+      <DropdownMenuContent align="start" className="z-[140]">
+        {providers.map(provider => (
+          <DropdownMenuItem
+            className="flex items-center gap-1.5"
+            key={provider.name}
+            // Picking the default clears the override (no need to pin it).
+            onSelect={() => setPetGenProvider(provider.default ? '' : provider.name)}
+          >
+            <span className="min-w-0 flex-1 truncate font-medium text-foreground">{provider.label}</span>
+            {provider.name === current?.name && <Check className="size-3.5 text-primary" />}
+          </DropdownMenuItem>
+        ))}
+      </DropdownMenuContent>
+    </DropdownMenu>
+  )
+}

apps/desktop/src/app/pet-generate/components/reference-chip.tsx

@@ -0,0 +1,48 @@
+import { useState } from 'react'
+
+import { ImageLightbox } from '@/components/chat/zoomable-image'
+import { useImageDownload } from '@/hooks/use-image-download'
+import { useI18n } from '@/i18n'
+import { X } from '@/lib/icons'
+
+interface ReferenceChipProps {
+  name: string
+  onRemove: () => void
+  src: string
+}
+
+// The reference photo as an attachment chip: filename + thumbnail that opens
+// the shared image viewer (lightbox), with a remove affordance.
+export function ReferenceChip({ name, onRemove, src }: ReferenceChipProps) {
+  const { t } = useI18n()
+  const { download, saving } = useImageDownload(src)
+  const [viewing, setViewing] = useState(false)
+
+  return (
+    <div className="ml-auto flex h-6 items-center gap-2 self-start rounded-lg border border-border/60 bg-background/50 pl-1 pr-2">
+      <button className="shrink-0" onClick={() => setViewing(true)} title={t.desktop.openImage} type="button">
+        <img alt={name} className="size-4 rounded-md object-cover" src={src} />
+      </button>
+
+      <span className="max-w-40 truncate text-[0.64rem] font-medium text-foreground/50">{name || 'Reference'}</span>
+      <button
+        aria-label="Remove reference"
+        className="text-(--ui-text-tertiary) transition not-hover:opacity-50"
+        onClick={onRemove}
+        type="button"
+      >
+        <X className="size-3" />
+      </button>
+
+      <ImageLightbox
+        alt={name}
+        copy={t.desktop}
+        onClick={download}
+        onOpenChange={setViewing}
+        open={viewing}
+        saving={saving}
+        src={src}
+      />
+    </div>
+  )
+}

apps/desktop/src/app/pet-generate/lib/frame-count.ts

@@ -0,0 +1,26 @@
+import { type PetInfo } from '@/store/pet'
+
+// Sprite row → the PetInfo frame-count key it resolves to (directional walks and
+// aliases collapse onto their base state).
+const ROW_TO_FRAME_KEY: Record<string, string> = {
+  idle: 'idle',
+  wave: 'wave',
+  waving: 'wave',
+  jump: 'jump',
+  jumping: 'jump',
+  run: 'run',
+  running: 'run',
+  'running-right': 'run',
+  'running-left': 'run',
+  failed: 'failed',
+  review: 'review',
+  waiting: 'waiting'
+}
+
+// Real frame count for a row, preferring the concrete per-row count, then the
+// per-state count, then the mapped base state, then the sheet-wide default.
+export function frameCountForRow(pet: PetInfo, row: string): number {
+  const mapped = ROW_TO_FRAME_KEY[row]
+
+  return pet.framesByRow?.[row] ?? pet.framesByState?.[row] ?? (mapped ? pet.framesByState?.[mapped] : undefined) ?? pet.framesPerState ?? 0
+}

apps/desktop/src/app/pet-generate/lib/read-reference-image.ts

@@ -0,0 +1,49 @@
+const DEFAULT_MAX_INPUT_BYTES = 16 * 1024 * 1024
+
+function loadImage(url: string): Promise<HTMLImageElement> {
+  const img = new Image()
+
+  return new Promise((resolve, reject) => {
+    img.onload = () => resolve(img)
+    img.onerror = () => reject(new Error('unreadable image'))
+    img.src = url
+  })
+}
+
+// Read an image file as a downscaled PNG data URL. We decode from an object URL
+// (not readAsDataURL) so large files don't inflate into giant base64 strings
+// before we scale them down for generation.
+export async function readReferenceImage(
+  file: File,
+  max = 1024,
+  maxInputBytes = DEFAULT_MAX_INPUT_BYTES
+): Promise<string> {
+  if (file.size > maxInputBytes) {
+    throw new Error('reference image too large')
+  }
+
+  const objectUrl = URL.createObjectURL(file)
+
+  try {
+    const img = await loadImage(objectUrl)
+    const scale = Math.min(1, max / Math.max(img.width, img.height))
+    const width = Math.max(1, Math.round(img.width * scale))
+    const height = Math.max(1, Math.round(img.height * scale))
+
+    const canvas = document.createElement('canvas')
+    canvas.width = width
+    canvas.height = height
+
+    const ctx = canvas.getContext('2d')
+
+    if (!ctx) {
+      throw new Error('could not create canvas context')
+    }
+
+    ctx.drawImage(img, 0, 0, width, height)
+
+    return canvas.toDataURL('image/png')
+  } finally {
+    URL.revokeObjectURL(objectUrl)
+  }
+}

apps/desktop/src/app/pet-generate/pet-generate-content.tsx

@@ -0,0 +1,336 @@
+import { useStore } from '@nanostores/react'
+import { useEffect, useRef, useState } from 'react'
+import { useNavigate } from 'react-router-dom'
+
+import { useGatewayRequest } from '@/app/gateway/hooks/use-gateway-request'
+import { SETTINGS_ROUTE } from '@/app/routes'
+import { Alert, AlertDescription } from '@/components/ui/alert'
+import { ConfirmDialog } from '@/components/ui/confirm-dialog'
+import { DialogHeader, DialogTitle } from '@/components/ui/dialog'
+import { GenerateButton } from '@/components/ui/generate-button'
+import { Input } from '@/components/ui/input'
+import { useI18n } from '@/i18n'
+import { triggerHaptic } from '@/lib/haptics'
+import { Egg, ImageIcon } from '@/lib/icons'
+import { cn } from '@/lib/utils'
+import {
+  $petGenAvailable,
+  $petGenDrafts,
+  $petGenError,
+  $petGenInput,
+  $petGenPreview,
+  $petGenRefImage,
+  $petGenRefName,
+  $petGenRemixConfirmed,
+  $petGenSelected,
+  $petGenStage,
+  $petGenStatus,
+  adoptHatched,
+  cancelGenerate,
+  checkPetGenAvailable,
+  cleanPetName,
+  closePetGenerate,
+  discardDrafts,
+  discardHatched,
+  generateDrafts,
+  hatchSelected,
+  markRemixConfirmed
+} from '@/store/pet-generate'
+
+import { DraftGrid } from './components/draft-grid'
+import { EmptyHint } from './components/empty-hint'
+import { GenerateUnavailable } from './components/generate-unavailable'
+import { HatchPreview } from './components/hatch-preview'
+import { HatchingView } from './components/hatching-view'
+import { ProviderPicker } from './components/provider-picker'
+import { ReferenceChip } from './components/reference-chip'
+import { readReferenceImage } from './lib/read-reference-image'
+
+// The generate → hatch → adopt controller. A thin view over the `pet-generate`
+// store; the store owns the steps and persists inputs across close/reopen.
+export function PetGenerateContent() {
+  const { t } = useI18n()
+  const copy = t.commandCenter.generatePet
+  const { requestGateway } = useGatewayRequest()
+  const navigate = useNavigate()
+
+  const status = useStore($petGenStatus)
+  const error = useStore($petGenError)
+  const available = useStore($petGenAvailable)
+  // `null` = not yet probed → stay optimistic (show the prompt); only the
+  // confirmed-no-backend case swaps in the setup card.
+  const unavailable = available === false
+  const drafts = useStore($petGenDrafts)
+  const selected = useStore($petGenSelected)
+  const preview = useStore($petGenPreview)
+  const stage = useStore($petGenStage)
+
+  // Inputs live in atoms so they survive a close/reopen (and background runs).
+  const prompt = useStore($petGenInput)
+  const refImage = useStore($petGenRefImage)
+  const refName = useStore($petGenRefName)
+  const fileRef = useRef<HTMLInputElement>(null)
+
+  // The draft awaiting the one-time "remix regenerates" confirmation.
+  const [remixPending, setRemixPending] = useState<{ dataUri: string } | null>(null)
+
+  // Probe backend availability on open — and again whenever the content
+  // remounts (e.g. after returning from the providers settings), so adding a
+  // key flips the setup card to the prompt with no manual refresh.
+  useEffect(() => {
+    void checkPetGenAvailable(requestGateway)
+  }, [requestGateway])
+
+  const busy = status === 'generating' || status === 'hatching'
+  const hasDrafts = drafts.length > 0
+  const generating = status === 'generating'
+
+  // The idle "describe a pet" state — egg + suggestions get generous, equidistant
+  // breathing room (gap-4) from the prompt; the working states stay compact.
+  const isEmptyState =
+    !hasDrafts &&
+    !generating &&
+    status !== 'hatching' &&
+    status !== 'preview' &&
+    status !== 'adopting' &&
+    status !== 'stale'
+
+  const generate = () => {
+    if ((prompt.trim() || refImage) && !busy) {
+      void generateDrafts(requestGateway, { prompt: prompt.trim(), referenceImage: refImage ?? undefined })
+    }
+  }
+
+  const clearReference = () => {
+    $petGenRefImage.set(null)
+    $petGenRefName.set('')
+  }
+
+  const pickReference = (file: File | undefined) => {
+    if (!file) {
+      return
+    }
+
+    const mapReferenceError = (reason: unknown): string => {
+      const message = reason instanceof Error ? reason.message.toLowerCase() : ''
+
+      return message.includes('too large') ? copy.referenceImageTooLarge : copy.referenceImageInvalid
+    }
+
+    void readReferenceImage(file)
+      .then(dataUrl => {
+        $petGenRefImage.set(dataUrl)
+        $petGenRefName.set(file.name)
+        // Clear picker-only errors once the reference is valid again.
+
+        if ($petGenStatus.get() === 'error' && $petGenDrafts.get().length === 0) {
+          $petGenStatus.set('idle')
+          $petGenError.set(null)
+        }
+      })
+      .catch(reason => {
+        $petGenRefImage.set(null)
+        $petGenRefName.set('')
+        $petGenError.set(mapReferenceError(reason))
+
+        if (!busy) {
+          $petGenStatus.set('error')
+        }
+      })
+  }
+
+  // One-click an example prompt straight into a draft round.
+  const runExample = (example: string) => {
+    $petGenInput.set(example)
+    void generateDrafts(requestGateway, { prompt: example })
+  }
+
+  // A remix re-runs generation grounded on an existing draft — same prompt, stay
+  // on step 2 — so the user explores variations without starting over.
+  const runRemix = (draft: { dataUri: string }) => {
+    void generateDrafts(requestGateway, { prompt: prompt.trim(), referenceImage: draft.dataUri })
+  }
+
+  // Slow, and it replaces the current drafts — so confirm once, then remember it.
+  const remixDraft = (draft: { dataUri: string }) => {
+    if (busy) {
+      return
+    }
+
+    if ($petGenRemixConfirmed.get()) {
+      runRemix(draft)
+
+      return
+    }
+
+    setRemixPending(draft)
+  }
+
+  // Hatch the selected draft. The user can pick one before the rest stream in —
+  // if so, abort the remaining generations first (keeping the drafts we have).
+  // The prompt is grounding text, not a label; the user names it on reveal.
+  const hatch = () => {
+    if (selected === null) {
+      return
+    }
+
+    if (generating) {
+      cancelGenerate()
+    }
+
+    void hatchSelected(requestGateway, { name: cleanPetName(prompt), prompt: prompt.trim() })
+  }
+
+  const adopt = (finalName: string) => {
+    void adoptHatched(requestGateway, finalName).then(out => {
+      if (out.ok) {
+        triggerHaptic('crisp')
+        closePetGenerate()
+      }
+    })
+  }
+
+  // The header title tracks the phase instead of sticking on "Generate a pet".
+  const headerTitle =
+    status === 'hatching' ? copy.spawning : status === 'preview' || status === 'adopting' ? copy.hatched : copy.title
+
+  // Send the user to set up a key without closing — the overlay yields to the
+  // settings route (useRouteOverlayActive) and reappears + re-checks on return.
+  const setupImageGen = () => navigate(`${SETTINGS_ROUTE}?tab=providers`)
+
+  // Prompt input only belongs on the describe/draft screens (and never when
+  // there's no backend to generate with).
+  const showPrompt = !unavailable && status !== 'hatching' && status !== 'preview' && status !== 'adopting'
+
+  return (
+    <>
+      {unavailable ? (
+        <DialogTitle className="sr-only">{copy.title}</DialogTitle>
+      ) : (
+        <DialogHeader>
+          <DialogTitle icon={Egg}>{headerTitle}</DialogTitle>
+        </DialogHeader>
+      )}
+
+      <div className={cn('flex min-h-0 flex-1 flex-col', isEmptyState ? 'gap-4' : 'gap-2.5')}>
+        {/* Concept prompt with the inline sparkle generate/stop affordance (the
+            same primitive as the commit-message + project-idea fields). */}
+        {showPrompt && (
+          <div className="flex flex-col gap-1.5">
+            <div className="relative">
+              <Input
+                autoFocus
+                className="pr-9"
+                onChange={event => $petGenInput.set(event.target.value)}
+                onKeyDown={event => {
+                  if (event.key === 'Enter') {
+                    event.preventDefault()
+                    generate()
+                  }
+                }}
+                placeholder={copy.placeholder}
+                value={prompt}
+              />
+              <GenerateButton
+                className="absolute right-1 top-1/2 -translate-y-1/2"
+                disabled={!prompt.trim() && !refImage}
+                generating={generating}
+                generatingLabel={t.common.cancel}
+                label={copy.generate}
+                // Inline cancel should match step-2 cancel semantics: abort and
+                // return to step 1 (prompt retained for quick tweaks).
+                onCancel={discardDrafts}
+                onGenerate={generate}
+              />
+            </div>
+
+            <div className="flex items-center gap-2">
+              <ProviderPicker />
+              {refImage ? (
+                <ReferenceChip name={refName} onRemove={clearReference} src={refImage} />
+              ) : (
+                <button
+                  className="ml-auto flex h-6 items-center gap-1.5 text-[0.6875rem] text-(--ui-text-tertiary) transition hover:text-foreground"
+                  onClick={() => fileRef.current?.click()}
+                  type="button"
+                >
+                  <ImageIcon className="size-3" />
+                  Add a reference
+                </button>
+              )}
+            </div>
+
+            {/* Optional reference photo — make a pet from the user's own image.
+                Styled like the chat composer's attachment pill. */}
+            <Input
+              accept="image/*"
+              className="hidden"
+              onChange={event => {
+                pickReference(event.target.files?.[0])
+                event.target.value = ''
+              }}
+              ref={fileRef}
+              type="file"
+            />
+          </div>
+        )}
+
+        {/* Hatch failed but the drafts are still here — show why above the grid so
+            the user can re-pick and retry without losing their options. */}
+        {status === 'error' && hasDrafts && (
+          <Alert variant="destructive">
+            <AlertDescription>{error || copy.genericError}</AlertDescription>
+          </Alert>
+        )}
+
+        {unavailable ? (
+          <GenerateUnavailable onSetup={setupImageGen} />
+        ) : status === 'stale' ? (
+          <Alert variant="destructive">
+            <AlertDescription>{copy.staleBackend}</AlertDescription>
+          </Alert>
+        ) : status === 'hatching' ? (
+          <HatchingView stage={stage} />
+        ) : (status === 'preview' || status === 'adopting') && preview ? (
+          <HatchPreview
+            adopting={status === 'adopting'}
+            error={error}
+            onAdopt={adopt}
+            onDiscard={() => void discardHatched(requestGateway)}
+            pet={preview}
+          />
+        ) : !hasDrafts && !generating ? (
+          // Doubles as the error-empty state — the failure reason rides the
+          // dialog's footer banner, so here we just offer the retry sparks.
+          <EmptyHint onExample={runExample} />
+        ) : (
+          <DraftGrid
+            drafts={drafts}
+            generating={generating}
+            hasDrafts={hasDrafts}
+            onCancel={discardDrafts}
+            onHatch={hatch}
+            onRemix={remixDraft}
+            onSelect={index => $petGenSelected.set(index)}
+            selected={selected}
+          />
+        )}
+      </div>
+
+      <ConfirmDialog
+        confirmLabel={copy.remix}
+        description={copy.remixConfirmBody}
+        onClose={() => setRemixPending(null)}
+        onConfirm={() => {
+          markRemixConfirmed()
+
+          if (remixPending) {
+            runRemix(remixPending)
+          }
+        }}
+        open={remixPending !== null}
+        title={copy.remixConfirmTitle}
+      />
+    </>
+  )
+}

apps/desktop/src/app/pet-generate/pet-generate-overlay.tsx

@@ -0,0 +1,86 @@
+/**
+ * "Hatch a Pet" — a dedicated, Pokédex-style overlay for pet generation.
+ *
+ * Previously generation lived as a cramped nested page inside the Cmd-K command
+ * palette (~34rem popover). This is its own full Radix dialog with room to
+ * breathe: a device-framed header, its own concept prompt, a roomy draft grid
+ * that streams in live, and the egg-hatch + reveal flow. It's a thin view over
+ * the `pet-generate` store; the store owns the generate → hatch → adopt steps.
+ *
+ * This file is just the dialog shell + sizing; the flow lives in
+ * `PetGenerateContent`, and each screen is its own atomic component under
+ * `./components`.
+ */
+
+import { useStore } from '@nanostores/react'
+
+import { useGatewayRequest } from '@/app/gateway/hooks/use-gateway-request'
+import { useRouteOverlayActive } from '@/app/hooks/use-route-overlay-active'
+import { Dialog, DialogContent } from '@/components/ui/dialog'
+import { useI18n } from '@/i18n'
+import { cn } from '@/lib/utils'
+import {
+  $petGenDrafts,
+  $petGenerateOpen,
+  $petGenError,
+  $petGenStatus,
+  cleanupPetGenOnClose,
+  closePetGenerate
+} from '@/store/pet-generate'
+
+import { PetGenerateContent } from './pet-generate-content'
+
+export function PetGenerateOverlay() {
+  const { t } = useI18n()
+  const { requestGateway } = useGatewayRequest()
+  const open = useStore($petGenerateOpen)
+  const status = useStore($petGenStatus)
+  const error = useStore($petGenError)
+  const drafts = useStore($petGenDrafts)
+
+  // Yield the screen to a full-screen route overlay (e.g. /settings while the
+  // user adds an image-gen key) without tearing down — the store keeps us open,
+  // and we reappear + re-check on return.
+  if (useRouteOverlayActive()) {
+    return null
+  }
+
+  const handleOpenChange = (next: boolean) => {
+    if (!next) {
+      cleanupPetGenOnClose(requestGateway)
+      // Never interrupt in-flight work. Generating/hatching continues in the
+      // background; only an unadopted finished preview is discarded on close.
+      closePetGenerate()
+    }
+  }
+
+  // The draft screen needs room for the 2×2 grid; the single-pet screens
+  // (hatch egg, reveal) shrink to the pet's frame so it isn't lost in a wide box.
+  // `fitContent` lets the dialog size to content; the `min-w` floors each phase.
+  const single = status === 'hatching' || status === 'preview' || status === 'adopting'
+  const copy = t.commandCenter.generatePet
+
+  // The footer banner narrates the dialog's async state: the failure reason on a
+  // dead-end error, else the "you can close this, we'll notify you" reassurance
+  // while a generate/hatch runs in the background. On step 1, show a neutral ETA.
+  const working = status === 'generating' || status === 'hatching'
+  const errored = status === 'error' && drafts.length === 0
+  const stepOne = status === 'idle' || status === 'ready'
+  const banner = errored ? error || copy.genericError : working ? copy.backgroundHint : stepOne ? copy.slowProviderHint : undefined
+
+  return (
+    <Dialog onOpenChange={handleOpenChange} open={open}>
+      <DialogContent
+        aria-describedby={undefined}
+        banner={banner}
+        bannerTone={errored ? 'error' : 'info'}
+        // Cap the width so a long banner (e.g. a provider refusal) wraps instead
+        // of stretching the dialog out; the min-w floors each phase.
+        className={cn('gap-4 text-center', single ? 'min-w-[17rem] max-w-[20rem]' : 'min-w-[19rem] max-w-[22rem]')}
+        fitContent
+      >
+        {open && <PetGenerateContent />}
+      </DialogContent>
+    </Dialog>
+  )
+}

apps/desktop/src/app/pet-overlay/overlay-root.tsx

@@ -0,0 +1,38 @@
+import { StrictMode } from 'react'
+import { createRoot } from 'react-dom/client'
+
+import { ErrorBoundary } from '@/components/error-boundary'
+import { ThemeProvider } from '@/themes/context'
+
+import { PetOverlayApp } from './pet-overlay-app'
+
+/**
+ * Boot the pet-overlay window. Loaded by the same bundle as the main app but
+ * via `?win=overlay`, so it shares CSS/atoms while mounting a minimal, transparent
+ * surface (no app shell, no gateway, no I18n — the bubble strings are inline).
+ *
+ * The index.html boot script paints an OPAQUE themed background to avoid a flash
+ * in normal windows; the overlay must be see-through, so we force every host
+ * layer transparent with a late, high-specificity style tag.
+ */
+export function mountPetOverlay(): void {
+  const style = document.createElement('style')
+  style.textContent = 'html,body,#root{background:transparent !important;}'
+  document.head.appendChild(style)
+
+  const root = document.getElementById('root')
+
+  if (!root) {
+    return
+  }
+
+  createRoot(root).render(
+    <StrictMode>
+      <ErrorBoundary label="pet-overlay">
+        <ThemeProvider>
+          <PetOverlayApp />
+        </ThemeProvider>
+      </ErrorBoundary>
+    </StrictMode>
+  )
+}

apps/desktop/src/app/pet-overlay/pet-overlay-app.tsx

@@ -0,0 +1,345 @@
+import { useStore } from '@nanostores/react'
+import { useEffect, useRef, useState } from 'react'
+
+import { PetBubble } from '@/components/pet/pet-bubble'
+import { PetSprite } from '@/components/pet/pet-sprite'
+import { Mail } from '@/lib/icons'
+import { $petActivity, $petInfo, setPetInfo } from '@/store/pet'
+import { setAwaitingResponse, setBusy } from '@/store/session'
+
+/**
+ * The pop-out overlay's only view: a transparent, draggable mascot with a mini
+ * composer.
+ *
+ * This runs in a separate, gateway-less BrowserWindow (`?win=overlay`). It is a
+ * pure puppet — the main renderer pushes the live pet state over IPC and we
+ * mirror it into the same atoms the in-window pet reads, so `PetSprite` /
+ * `PetBubble` render identically with zero extra logic.
+ *
+ * The window is a full rectangle but mostly transparent; we toggle OS-level
+ * mouse click-through so only the sprite (or the open composer) is interactive
+ * and the empty margins pass clicks through to whatever is behind.
+ *
+ * Gestures on the pet: drag to move it anywhere on screen (even outside the
+ * app), shift-click to pop it back into the window, single-click to open a small
+ * composer, double-click to toggle the app window (minimize ↔ restore). A mail
+ * icon (shown only when a turn finished while you were away) raises the app on
+ * the most recent thread.
+ */
+
+// Below this much pointer travel, a press counts as a click, not a drag.
+const CLICK_SLOP_PX = 3
+// A second click within this window is a double-click (raise app) and cancels
+// the deferred single-click (open composer), so a double never flashes it open.
+const DOUBLE_CLICK_MS = 250
+
+interface DragState {
+  startX: number
+  startY: number
+  offX: number
+  offY: number
+  width: number
+  height: number
+  moved: boolean
+}
+
+export function PetOverlayApp() {
+  const info = useStore($petInfo)
+  const [composerOpen, setComposerOpen] = useState(false)
+  const [draft, setDraft] = useState('')
+  // Mirrored from the main renderer: a finish landed while you were away.
+  const [unread, setUnread] = useState(false)
+
+  const dragRef = useRef<DragState | null>(null)
+  const petRef = useRef<HTMLDivElement | null>(null)
+  const inputRef = useRef<HTMLInputElement | null>(null)
+  const ignoreRef = useRef(true)
+  const composerOpenRef = useRef(false)
+  const clickTimerRef = useRef<ReturnType<typeof setTimeout> | undefined>(undefined)
+
+  const setIgnore = (ignore: boolean) => {
+    if (ignoreRef.current !== ignore) {
+      ignoreRef.current = ignore
+      window.hermesDesktop?.petOverlay?.setIgnoreMouse(ignore)
+    }
+  }
+
+  // Mirror pushed state into the shared atoms so PetSprite/PetBubble just work.
+  useEffect(() => {
+    const off = window.hermesDesktop?.petOverlay?.onState(payload => {
+      setPetInfo(payload.info)
+      $petActivity.set(payload.activity ?? {})
+      setBusy(Boolean(payload.busy))
+      setAwaitingResponse(Boolean(payload.awaiting))
+      setUnread(Boolean(payload.unread))
+    })
+
+    // Tell the main renderer we're mounted so it pushes the current frame (the
+    // subscribe-time pushes during open() can land before this view exists).
+    window.hermesDesktop?.petOverlay?.control({ type: 'ready' })
+
+    return off
+  }, [])
+
+  // Click-through: make only the sprite (or an open composer) interactive. With
+  // ignore+forward, the renderer still receives mousemove so we can re-enable
+  // hit-testing the moment the cursor returns to the pet.
+  useEffect(() => {
+    setIgnore(true)
+
+    const onMove = (ev: MouseEvent) => {
+      if (dragRef.current || composerOpenRef.current) {
+        setIgnore(false)
+
+        return
+      }
+
+      const el = petRef.current
+
+      if (!el) {
+        return
+      }
+
+      const r = el.getBoundingClientRect()
+      const over = ev.clientX >= r.left && ev.clientX <= r.right && ev.clientY >= r.top && ev.clientY <= r.bottom
+      setIgnore(!over)
+    }
+
+    window.addEventListener('mousemove', onMove)
+
+    return () => {
+      window.removeEventListener('mousemove', onMove)
+      clearTimeout(clickTimerRef.current)
+    }
+  }, [])
+
+  // The whole window must stay interactive while the composer is open (so the
+  // input keeps focus); focus it on open. The overlay is a non-activating panel
+  // (so it never steals the app's cmd/alt-tab anchor) — flip it focusable while
+  // the composer needs the keyboard, then back to non-activating when it closes.
+  useEffect(() => {
+    composerOpenRef.current = composerOpen
+
+    window.hermesDesktop?.petOverlay?.setFocusable(composerOpen)
+
+    if (composerOpen) {
+      setIgnore(false)
+      // The OS window has to become key first (setFocusable + focus happen in
+      // the main process), so focus the input on the next frame.
+      requestAnimationFrame(() => inputRef.current?.focus())
+    }
+  }, [composerOpen])
+
+  const onPetPointerDown = (e: React.PointerEvent) => {
+    if (e.button !== 0) {
+      return
+    }
+
+    ;(e.target as Element).setPointerCapture?.(e.pointerId)
+    dragRef.current = {
+      height: window.outerHeight,
+      moved: false,
+      offX: e.screenX - window.screenX,
+      offY: e.screenY - window.screenY,
+      startX: e.screenX,
+      startY: e.screenY,
+      width: window.outerWidth
+    }
+  }
+
+  const onPetPointerMove = (e: React.PointerEvent) => {
+    const drag = dragRef.current
+
+    if (!drag) {
+      return
+    }
+
+    if (Math.hypot(e.screenX - drag.startX, e.screenY - drag.startY) > CLICK_SLOP_PX) {
+      drag.moved = true
+    }
+
+    window.hermesDesktop?.petOverlay?.setBounds({
+      height: drag.height,
+      width: drag.width,
+      x: e.screenX - drag.offX,
+      y: e.screenY - drag.offY
+    })
+  }
+
+  const onPetPointerUp = (e: React.PointerEvent) => {
+    const drag = dragRef.current
+    dragRef.current = null
+    ;(e.target as Element).releasePointerCapture?.(e.pointerId)
+
+    if (!drag) {
+      return
+    }
+
+    if (drag.moved) {
+      // A drag cancels any deferred single-click so the composer can't pop open
+      // after you reposition the pet.
+      clearTimeout(clickTimerRef.current)
+      clickTimerRef.current = undefined
+
+      // Remember the spot on the desktop (screen coords) so the pet reopens here
+      // next time / after a restart.
+      window.hermesDesktop?.petOverlay?.control({
+        bounds: { height: drag.height, width: drag.width, x: e.screenX - drag.offX, y: e.screenY - drag.offY },
+        type: 'bounds'
+      })
+
+      return
+    }
+
+    // Shift-click always pops the pet back in (no double-click ambiguity).
+    if (e.shiftKey) {
+      window.hermesDesktop?.petOverlay?.control({ type: 'pop-in' })
+
+      return
+    }
+
+    // Double-click toggles the app window (minimize ↔ restore); defer the
+    // single-click composer toggle so a double never flashes the composer open.
+    if (clickTimerRef.current) {
+      clearTimeout(clickTimerRef.current)
+      clickTimerRef.current = undefined
+      window.hermesDesktop?.petOverlay?.control({ type: 'toggle-app' })
+
+      return
+    }
+
+    clickTimerRef.current = setTimeout(() => {
+      clickTimerRef.current = undefined
+      setComposerOpen(open => !open)
+    }, DOUBLE_CLICK_MS)
+  }
+
+  const send = () => {
+    const text = draft.trim()
+
+    if (text) {
+      window.hermesDesktop?.petOverlay?.control({ text, type: 'submit' })
+    }
+
+    setDraft('')
+    setComposerOpen(false)
+  }
+
+  const openApp = () => {
+    // Hide the icon immediately; the main renderer also clears the source flag.
+    setUnread(false)
+    window.hermesDesktop?.petOverlay?.control({ type: 'open-app' })
+  }
+
+  if (!info.enabled || !info.spritesheetBase64) {
+    return null
+  }
+
+  return (
+    <div
+      onPointerDown={e => {
+        // Click on the transparent backdrop (not the pet/composer) dismisses
+        // the composer.
+        if (composerOpen && e.target === e.currentTarget) {
+          setComposerOpen(false)
+        }
+      }}
+      style={{
+        alignItems: 'center',
+        background: 'transparent',
+        display: 'flex',
+        flexDirection: 'column',
+        height: '100vh',
+        justifyContent: 'flex-end',
+        paddingBottom: 24,
+        userSelect: 'none',
+        width: '100vw'
+      }}
+    >
+      {composerOpen && (
+        <input
+          onChange={e => setDraft(e.target.value)}
+          onKeyDown={e => {
+            if (e.key === 'Enter' && !e.shiftKey) {
+              e.preventDefault()
+              send()
+            } else if (e.key === 'Escape') {
+              setComposerOpen(false)
+            }
+          }}
+          placeholder="Message…"
+          ref={inputRef}
+          style={{
+            background: 'var(--ui-bg-elevated)',
+            border: '1px solid var(--ui-stroke-secondary)',
+            borderRadius: 2,
+            boxShadow: '0 6px 18px rgba(0,0,0,0.28)',
+            color: 'var(--foreground)',
+            fontSize: 12,
+            marginBottom: 8,
+            outline: 'none',
+            padding: '4px 8px',
+            width: 184
+          }}
+          value={draft}
+        />
+      )}
+
+      <div
+        onPointerDown={onPetPointerDown}
+        onPointerMove={onPetPointerMove}
+        onPointerUp={onPetPointerUp}
+        ref={petRef}
+        style={{
+          alignItems: 'center',
+          cursor: 'grab',
+          display: 'flex',
+          flexDirection: 'column',
+          position: 'relative',
+          touchAction: 'none'
+        }}
+      >
+        <div style={{ marginBottom: 4 }}>
+          <PetBubble />
+        </div>
+        <div style={{ lineHeight: 0, position: 'relative' }}>
+          <PetSprite info={info} />
+
+          {/* Mail icon: only when a finish landed while you were away. Jumps to
+              the app's most recent thread. Anchored to the sprite (kept inside
+              its box so the overlay's click-through hit-test still catches it);
+              stopPropagation keeps a click from starting a window drag. */}
+          {unread && (
+            <button
+              aria-label="Open in Hermes"
+              onClick={openApp}
+              onPointerDown={e => e.stopPropagation()}
+              onPointerUp={e => e.stopPropagation()}
+              style={{
+                alignItems: 'center',
+                background: 'var(--ui-bg-elevated)',
+                border: '1px solid var(--ui-stroke-secondary)',
+                borderRadius: 999,
+                boxShadow: '0 4px 14px rgba(0,0,0,0.22)',
+                color: 'var(--foreground)',
+                cursor: 'pointer',
+                display: 'inline-flex',
+                height: 24,
+                justifyContent: 'center',
+                padding: 0,
+                position: 'absolute',
+                right: 0,
+                top: 0,
+                width: 24
+              }}
+              title="Open in Hermes"
+              type="button"
+            >
+              <Mail style={{ height: 13, width: 13 }} />
+            </button>
+          )}
+        </div>
+      </div>
+    </div>
+  )
+}

apps/desktop/src/app/session/hooks/use-message-stream.ts

@@ -34,6 +34,7 @@ import { $gateway } from '@/store/gateway'
 import { dispatchNativeNotification } from '@/store/native-notifications'
 import { notify } from '@/store/notifications'
 import { requestDesktopOnboarding } from '@/store/onboarding'
+import { flashPetActivity, markPetUnread, setPetActivity } from '@/store/pet'
 import { clearAllPrompts, setApprovalRequest, setSecretRequest, setSudoRequest } from '@/store/prompts'
 import {
   setCurrentBranch,
@@ -870,10 +871,18 @@ export function useMessageStream({
         if (sessionId) {
           appendReasoningDelta(sessionId, coerceThinkingText(payload?.text))
         }
+
+        if (isActiveEvent) {
+          setPetActivity({ reasoning: true })
+        }
       } else if (event.type === 'reasoning.available') {
         if (sessionId) {
           appendReasoningDelta(sessionId, coerceThinkingText(payload?.text), true)
         }
+
+        if (isActiveEvent) {
+          setPetActivity({ reasoning: true })
+        }
       } else if (event.type === 'message.complete') {
         if (!sessionId) {
           return
@@ -895,6 +904,20 @@ export function useMessageStream({
 
         if (isActiveEvent) {
           setTurnStartedAt(null)
+
+          // Pet beat: a finished turn always celebrates — go straight to the
+          // jump, never linger on the run/reason pose. One atom update (clears
+          // toolRunning/reasoning AND sets celebrate together) so no stray "run"
+          // frame leaks to the sprite — including the popped-out overlay, which
+          // mirrors each activity change. The jump runs ~2 loops, then settles.
+          flashPetActivity({ celebrate: true, reasoning: false, toolRunning: false }, 2200)
+
+          // Light up the pet's mail icon if the user wasn't looking when the turn
+          // finished — a glanceable "new message" hint on the popped-out overlay.
+          // Cleared when they open the app via the mail icon or refocus the window.
+          if (typeof document !== 'undefined' && !document.hasFocus()) {
+            markPetUnread()
+          }
         }
 
         if (payload?.usage) {
@@ -907,10 +930,19 @@ export function useMessageStream({
 
         flushQueuedDeltas(sessionId)
         upsertToolCall(sessionId, toTodoPayload(payload) ?? payload, 'running', event.type)
+
+        if (isActiveEvent) {
+          setPetActivity({ reasoning: false, toolRunning: true })
+        }
       } else if (event.type === 'tool.complete') {
         if (sessionId) {
           flushQueuedDeltas(sessionId)
           upsertToolCall(sessionId, toTodoPayload(payload) ?? payload, 'complete', event.type)
+
+          if (isActiveEvent) {
+            setPetActivity({ toolRunning: false })
+          }
+
           // A pending clarify blocks the turn, so the first tool.complete after
           // one is the clarify resolving — drop the "needs input" flag here so
           // the sidebar indicator clears as soon as it's answered, not only at
@@ -1120,6 +1152,11 @@ export function useMessageStream({
           compactedTurnRef.current.delete(sessionId)
         }
 
+        if (isActiveEvent) {
+          setPetActivity({ reasoning: false, toolRunning: false })
+          flashPetActivity({ error: true })
+        }
+
         dispatchNativeNotification({
           body: errorMessage,
           kind: 'turnError',

apps/desktop/src/app/session/hooks/use-prompt-actions.ts

@@ -27,6 +27,7 @@ import { triggerHaptic } from '@/lib/haptics'
 import { setMutableRef } from '@/lib/mutable-ref'
 import { isProviderSetupErrorMessage } from '@/lib/provider-setup-errors'
 import { setSessionYolo } from '@/lib/yolo-session'
+import { openCommandPalettePage } from '@/store/command-palette'
 import {
   $composerAttachments,
   clearComposerAttachments,
@@ -40,6 +41,8 @@ import { resetSessionBackground } from '@/store/composer-status'
 import { clearPreviewArtifacts } from '@/store/preview-status'
 import { clearNotifications, notify, notifyError } from '@/store/notifications'
 import { requestDesktopOnboarding } from '@/store/onboarding'
+import { setPetScale } from '@/store/pet-gallery'
+import { $petGenInput, openPetGenerate } from '@/store/pet-generate'
 import { $activeGatewayProfile, $newChatProfile, ensureGatewayProfile, normalizeProfileKey } from '@/store/profile'
 import {
   $busy,
@@ -59,8 +62,8 @@ import { clearSessionSubagents } from '@/store/subagents'
 import { clearSessionTodos } from '@/store/todos'
 
 import type {
-  ClientSessionState,
   BrowserManageResponse,
+  ClientSessionState,
   FileAttachResponse,
   HandoffFailResponse,
   HandoffRequestResponse,
@@ -552,7 +555,14 @@ export function usePromptActions({
     async (rawText: string, options?: SubmitTextOptions) => {
       const visibleText = rawText.trim()
       const usingComposerAttachments = !options?.attachments
-      const attachments = options?.attachments ?? $composerAttachments.get()
+      // Drop undefined/null holes a session switch or draft restore can leave in
+      // the attachments array (same bug class as AttachmentList #49624). Without
+      // this, the sibling iterations below (a.kind / a.label / a.refText, and the
+      // sync step) throw "Cannot read properties of undefined (reading 'refText')"
+      // and break the chat surface.
+      const attachments = (options?.attachments ?? $composerAttachments.get()).filter(
+        (a): a is ComposerAttachment => Boolean(a)
+      )
 
       const terminalContextBlocks = terminalContextBlocksFromDraft(rawText).join('\n\n')
       const hasImage = attachments.some(a => a.kind === 'image')
@@ -565,14 +575,17 @@ export function usePromptActions({
       let attachmentRefs = attachments.map(optimisticAttachmentRef).filter((r): r is string => Boolean(r))
 
       const buildContextText = (atts: ComposerAttachment[]): string => {
-        const contextRefs = atts
+        // atts may be the post-sync array, which can reintroduce holes; filter
+        // before touching a.refText / a.kind.
+        const present = atts.filter((a): a is ComposerAttachment => Boolean(a))
+        const contextRefs = present
           .map(a => a.refText)
           .filter(Boolean)
           .join('\n')
 
         return (
           [contextRefs, terminalContextBlocks, visibleText].filter(Boolean).join('\n\n') ||
-          (atts.some(a => a.kind === 'image') ? 'What do you see in this image?' : '')
+          (present.some(a => a.kind === 'image') ? 'What do you see in this image?' : '')
         )
       }
 
@@ -1176,6 +1189,47 @@ export function usePromptActions({
             renderSlashOutput(`error: ${err instanceof Error ? err.message : String(err)}`)
           }
         },
+        // /hatch opens the pet generator overlay (the desktop's rich, multi-step
+        // generate→pick→hatch→adopt flow). A typed description seeds the prompt
+        // so `/hatch a cyber fox` lands on the composer step prefilled.
+        hatch: async ({ arg }) => {
+          const concept = arg.trim()
+
+          if (concept) {
+            $petGenInput.set(concept)
+          }
+
+          openPetGenerate()
+        },
+        pet: async ctx => {
+          const [sub = '', rawValue = ''] = ctx.arg.trim().split(/\s+/)
+          const lower = sub.toLowerCase()
+
+          if (lower === 'list' || lower === 'gallery' || lower === 'browse' || lower === 'all') {
+            openCommandPalettePage('pets')
+
+            return
+          }
+
+          // `/pet scale <n>` resizes the floating pet locally (instant) and
+          // persists via the store — no round-trip to the slash worker.
+          if (lower === 'scale') {
+            const value = Number(rawValue)
+
+            if (!rawValue || Number.isNaN(value)) {
+              const resolved = await withSlashOutput(ctx)
+              resolved?.render('usage: /pet scale <factor>  (e.g. /pet scale 0.5)')
+
+              return
+            }
+
+            setPetScale(requestGateway, value)
+
+            return
+          }
+
+          await runExec(ctx)
+        },
         // /browser connect|disconnect|status manages the live CDP connection on
         // the gateway host, mirroring the TUI's browser.manage RPC. It mutates
         // BROWSER_CDP_URL (and may launch Chrome) in the gateway process — only
@@ -1392,6 +1446,7 @@ export function usePromptActions({
 
   const cancelRun = useCallback(async () => {
     const sessionId = activeSessionId || activeSessionIdRef.current
+
     const releaseBusy = () => {
       setMutableRef(busyRef, false)
       setBusy(false)

apps/desktop/src/app/session/hooks/use-session-state-cache.ts

@@ -9,6 +9,7 @@ import {
   $busy,
   $messages,
   noteSessionActivity,
+  onSessionWatchdogClear,
   setCurrentFastMode,
   setCurrentModel,
   setCurrentPersonality,
@@ -276,6 +277,31 @@ export function useSessionStateCache({
     [ensureSessionState, syncSessionStateToView]
   )
 
+  // When the store watchdog force-clears a stuck session (8 min of stream
+  // silence — a hung or looping turn that never delivered its terminal event),
+  // also drop that session's busy/awaiting flags here. Clearing the sidebar dot
+  // alone leaves the composer wedged on "Thinking"/Stop; updateSessionState
+  // re-syncs `$busy` when the healed session is the one on screen.
+  useEffect(
+    () =>
+      onSessionWatchdogClear(storedSessionId => {
+        const runtimeId = runtimeIdByStoredSessionIdRef.current.get(storedSessionId)
+        const state = runtimeId ? sessionStateByRuntimeIdRef.current.get(runtimeId) : undefined
+
+        if (!runtimeId || !state?.busy) {
+          return
+        }
+
+        updateSessionState(runtimeId, current => ({
+          ...current,
+          awaitingResponse: false,
+          busy: false,
+          needsInput: false
+        }))
+      }),
+    [updateSessionState]
+  )
+
   return {
     activeSessionIdRef,
     ensureSessionState,

apps/desktop/src/app/settings/appearance-settings.tsx

@@ -1,30 +1,31 @@
 import { useStore } from '@nanostores/react'
-import { useState } from 'react'
+import { useQuery } from '@tanstack/react-query'
+import { useEffect, useState } from 'react'
 
 import { LanguageSwitcher } from '@/components/language-switcher'
 import { SegmentedControl } from '@/components/ui/segmented-control'
+import type { DesktopMarketplaceSearchItem } from '@/global'
 import { useI18n } from '@/i18n'
 import { triggerHaptic } from '@/lib/haptics'
 import { Check, Download, Loader2, Palette, Trash2 } from '@/lib/icons'
+import { selectableCardClass } from '@/lib/selectable-card'
 import { cn } from '@/lib/utils'
 import { $activeGatewayProfile, $profiles, normalizeProfileKey } from '@/store/profile'
 import { $toolViewMode, setToolViewMode } from '@/store/tool-view'
 import { $translucency, setTranslucency } from '@/store/translucency'
-import { useTheme } from '@/themes/context'
+import { getBaseColors, useTheme } from '@/themes/context'
 import { installVscodeThemeFromMarketplace } from '@/themes/install'
-import { isUserTheme, removeUserTheme, resolveTheme } from '@/themes/user-themes'
+import { isUserTheme, removeUserTheme } from '@/themes/user-themes'
 
 import { MODE_OPTIONS } from './constants'
+import { PetSettings } from './pet-settings'
 import { ListRow, SectionHeading, SettingsContent } from './primitives'
 
-function ThemePreview({ name }: { name: string }) {
-  const t = resolveTheme(name)
-
-  if (!t) {
-    return null
-  }
-
-  const c = t.colors
+function ThemePreview({ name, mode }: { name: string; mode: 'light' | 'dark' }) {
+  // Preview in the *current* mode: the dark palette in Dark, and the light
+  // palette in Light — synthesizing one for dark-only themes — so every card
+  // tracks the Light/Dark toggle, exactly like the app itself does.
+  const c = getBaseColors(name, mode)
 
   return (
     <div
@@ -57,90 +58,200 @@ function ThemePreview({ name }: { name: string }) {
   )
 }
 
-function VscodeThemeInstaller() {
+function useDebounced<T>(value: T, delayMs: number): T {
+  const [debounced, setDebounced] = useState(value)
+
+  useEffect(() => {
+    const handle = setTimeout(() => setDebounced(value), delayMs)
+
+    return () => clearTimeout(handle)
+  }, [value, delayMs])
+
+  return debounced
+}
+
+const compactNumber = new Intl.NumberFormat(undefined, { notation: 'compact', maximumFractionDigits: 1 })
+
+/**
+ * Live VS Code Marketplace theme search (the same backend as the Cmd-K "Install
+ * theme…" page). Renders below the local grid when there's a query: each row
+ * downloads + converts + installs via `installVscodeThemeFromMarketplace` and
+ * activates it. Extensions already imported locally are marked installed.
+ */
+function MarketplaceThemeResults({
+  query,
+  installedExtIds,
+  onInstalled
+}: {
+  query: string
+  installedExtIds: Set<string>
+  onInstalled: (name: string) => void
+}) {
   const { t } = useI18n()
-  const { setTheme } = useTheme()
-  const a = t.settings.appearance
-  const [id, setId] = useState('')
-  const [busy, setBusy] = useState(false)
-  const [status, setStatus] = useState<{ kind: 'error' | 'success'; text: string } | null>(null)
+  const copy = t.commandCenter.installTheme
+  const debounced = useDebounced(query.trim(), 300)
+  const [installingId, setInstallingId] = useState<string | null>(null)
+  const [installedHere, setInstalledHere] = useState<Record<string, true>>({})
+  const [error, setError] = useState<string | null>(null)
 
-  const install = async () => {
-    const trimmed = id.trim()
+  const search = useQuery({
+    enabled: debounced.length > 0,
+    queryFn: () => window.hermesDesktop?.themes?.searchMarketplace(debounced) ?? Promise.resolve([]),
+    queryKey: ['marketplace-themes-settings', debounced],
+    staleTime: 5 * 60 * 1000
+  })
 
-    if (!trimmed || busy) {
+  const install = async (item: DesktopMarketplaceSearchItem) => {
+    if (installingId) {
       return
     }
 
-    setBusy(true)
-    setStatus(null)
+    setInstallingId(item.extensionId)
+    setError(null)
 
     try {
-      const theme = await installVscodeThemeFromMarketplace(trimmed)
+      const theme = await installVscodeThemeFromMarketplace(item.extensionId)
 
       triggerHaptic('crisp')
-      setTheme(theme.name)
-      setStatus({ kind: 'success', text: a.installed(theme.label) })
-      setId('')
-    } catch (error) {
-      setStatus({ kind: 'error', text: error instanceof Error ? error.message : a.installError })
+      setInstalledHere(prev => ({ ...prev, [item.extensionId]: true }))
+      onInstalled(theme.name)
+    } catch (e) {
+      setError(e instanceof Error ? e.message : copy.error)
     } finally {
-      setBusy(false)
+      setInstallingId(null)
     }
   }
 
-  return (
-    <div className="mt-3">
-      <div className="flex flex-wrap items-center gap-2">
-        <input
-          className="min-w-0 flex-1 rounded-lg border border-(--ui-stroke-tertiary) bg-(--ui-bg-quinary) px-3 py-1.5 font-mono text-[length:var(--conversation-caption-font-size)] outline-none placeholder:text-(--ui-text-tertiary) focus:border-(--ui-stroke-secondary)"
-          disabled={busy}
-          onChange={event => {
-            setId(event.target.value)
-            setStatus(null)
-          }}
-          onKeyDown={event => {
-            if (event.key === 'Enter') {
-              void install()
-            }
-          }}
-          placeholder={a.installPlaceholder}
-          spellCheck={false}
-          value={id}
-        />
-        <button
-          className="inline-flex items-center gap-1.5 rounded-lg border border-(--ui-stroke-secondary) bg-(--ui-bg-tertiary) px-3 py-1.5 text-[length:var(--conversation-caption-font-size)] font-medium transition hover:bg-(--chrome-action-hover) disabled:opacity-50"
-          disabled={busy || !id.trim()}
-          onClick={() => void install()}
-          type="button"
-        >
-          {busy ? <Loader2 className="size-3.5 animate-spin" /> : <Download className="size-3.5" />}
-          {busy ? a.installing : a.installButton}
-        </button>
-      </div>
-      {status && (
-        <p
-          className={cn(
-            'mt-2 text-[length:var(--conversation-caption-font-size)] leading-(--conversation-caption-line-height)',
-            status.kind === 'error' ? 'text-(--ui-red)' : 'text-(--ui-text-tertiary)'
-          )}
-        >
-          {status.text}
+  if (!debounced) {
+    return null
+  }
+
+  const header = (
+    <p className="mb-2 mt-4 text-[length:var(--conversation-caption-font-size)] font-medium text-(--ui-text-tertiary)">
+      From the VS Code Marketplace
+    </p>
+  )
+
+  if (search.isLoading) {
+    return (
+      <>
+        {header}
+        <p className="flex items-center gap-2 text-[length:var(--conversation-caption-font-size)] text-(--ui-text-tertiary)">
+          <Loader2 className="size-3.5 animate-spin" />
+          {copy.loading}
         </p>
-      )}
-    </div>
+      </>
+    )
+  }
+
+  if (search.isError) {
+    return (
+      <>
+        {header}
+        <p className="text-[length:var(--conversation-caption-font-size)] text-(--ui-red)">{copy.error}</p>
+      </>
+    )
+  }
+
+  const results = search.data ?? []
+
+  if (results.length === 0) {
+    return (
+      <>
+        {header}
+        <p className="text-[length:var(--conversation-caption-font-size)] text-(--ui-text-tertiary)">{copy.empty}</p>
+      </>
+    )
+  }
+
+  return (
+    <>
+      {header}
+      {error && <p className="mb-2 text-[length:var(--conversation-caption-font-size)] text-(--ui-red)">{error}</p>}
+      <div className="grid gap-2 sm:grid-cols-2">
+        {results.map(item => {
+          const busy = installingId === item.extensionId
+          const done = installedHere[item.extensionId] || installedExtIds.has(item.extensionId)
+
+          return (
+            <button
+              className={cn(
+                'flex items-center gap-2.5 px-2.5 py-2 text-left disabled:opacity-60',
+                selectableCardClass({ prominent: done })
+              )}
+              disabled={Boolean(installingId) && !busy}
+              key={item.extensionId}
+              onClick={() => void install(item)}
+              type="button"
+            >
+              <Palette className="size-4 shrink-0 text-(--ui-text-tertiary)" />
+              <span className="min-w-0 flex-1">
+                <span className="block truncate text-[length:var(--conversation-text-font-size)] font-medium">
+                  {item.displayName}
+                </span>
+                <span className="block truncate text-[length:var(--conversation-caption-font-size)] text-(--ui-text-tertiary)">
+                  {item.publisher}
+                  {item.installs > 0 ? ` · ${copy.installs(compactNumber.format(item.installs))}` : ''}
+                </span>
+              </span>
+              <span className="shrink-0 text-(--ui-text-tertiary)">
+                {busy ? (
+                  <Loader2 className="size-4 animate-spin" />
+                ) : done ? (
+                  <Check className="size-4 text-(--ui-green)" />
+                ) : (
+                  <Download className="size-4" />
+                )}
+              </span>
+            </button>
+          )
+        })}
+      </div>
+    </>
   )
 }
 
 export function AppearanceSettings() {
   const { t, isSavingLocale } = useI18n()
-  const { themeName, mode, availableThemes, setTheme, setMode } = useTheme()
+  const { themeName, mode, resolvedMode, availableThemes, setTheme, setMode } = useTheme()
   const toolViewMode = useStore($toolViewMode)
   const translucency = useStore($translucency)
   const profiles = useStore($profiles)
   const activeProfileKey = normalizeProfileKey(useStore($activeGatewayProfile))
   const a = t.settings.appearance
 
+  const [query, setQuery] = useState('')
+
+  // One box does double duty: filter installed themes live (below), and run a
+  // name search against the VS Code Marketplace (the Cmd-K "Install theme…"
+  // backend) for anything not already installed.
+  const needle = query.trim().toLowerCase()
+
+  const filteredThemes = availableThemes
+    .filter(
+      theme =>
+        !needle ||
+        theme.label.toLowerCase().includes(needle) ||
+        theme.name.toLowerCase().includes(needle) ||
+        theme.description.toLowerCase().includes(needle)
+    )
+    // Active theme first; stable sort keeps the rest in their original order.
+    .sort((a, b) => Number(b.name === themeName) - Number(a.name === themeName))
+
+  // Marketplace imports describe themselves as "VS Code · <publisher.extension>";
+  // pull those ids back out so search results already imported show as installed.
+  const MARKETPLACE_DESC_PREFIX = 'VS Code · '
+
+  const installedExtIds = new Set(
+    availableThemes
+      .map(theme =>
+        theme.description.startsWith(MARKETPLACE_DESC_PREFIX)
+          ? theme.description.slice(MARKETPLACE_DESC_PREFIX.length)
+          : ''
+      )
+      .filter(Boolean)
+  )
+
   // Themes save per profile. Surface that only when the user actually has more
   // than one profile (single-profile installs never see the distinction).
   const showProfileNote = profiles.length > 1
@@ -163,7 +274,7 @@ export function AppearanceSettings() {
           {a.intro}
         </p>
 
-        <div className="mt-2 divide-y divide-(--ui-stroke-tertiary)">
+        <div className="mt-2">
           <ListRow
             action={<LanguageSwitcher />}
             description={isSavingLocale ? t.language.saving : t.language.description}
@@ -171,18 +282,107 @@ export function AppearanceSettings() {
           />
 
           <ListRow
-            action={
-              <SegmentedControl
-                onChange={id => {
-                  triggerHaptic('crisp')
-                  setMode(id)
-                }}
-                options={modeOptions}
-                value={mode}
-              />
+            below={
+              <>
+                {/* One search box: filters your installed themes (the grid)
+                    and live-searches the VS Code Marketplace below. */}
+                <div className="mt-3">
+                  <input
+                    className="w-full rounded-lg border border-(--ui-stroke-tertiary) bg-(--ui-bg-quinary) px-3 py-1.5 text-[length:var(--conversation-caption-font-size)] outline-none placeholder:text-(--ui-text-tertiary) focus:border-(--ui-stroke-secondary)"
+                    onChange={event => setQuery(event.target.value)}
+                    placeholder="Search your themes or the VS Code Marketplace…"
+                    spellCheck={false}
+                    value={query}
+                  />
+                </div>
+
+                {/* Fixed-height scroll area so the (growing) theme list never
+                    runs the page long; the grid scrolls inside it. */}
+                <div className="mt-3 max-h-96 overflow-y-auto pr-1">
+                  {filteredThemes.length === 0 ? (
+                    needle ? (
+                      <p className="text-[length:var(--conversation-caption-font-size)] text-(--ui-text-tertiary)">
+                        No installed themes match "{query.trim()}".
+                      </p>
+                    ) : null
+                  ) : (
+                    <div className="grid gap-3 sm:grid-cols-2 xl:grid-cols-3">
+                      {filteredThemes.map(theme => {
+                        const active = themeName === theme.name
+                        const removable = isUserTheme(theme.name)
+
+                        return (
+                          <div className="group relative" key={theme.name}>
+                            <button
+                              className={cn('w-full p-2 text-left', selectableCardClass({ active, prominent: true }))}
+                              onClick={() => {
+                                triggerHaptic('crisp')
+                                setTheme(theme.name)
+                              }}
+                              type="button"
+                            >
+                              <ThemePreview mode={resolvedMode} name={theme.name} />
+                              <div className="mt-3 px-1">
+                                <div className="truncate text-[length:var(--conversation-text-font-size)] font-medium">
+                                  {theme.label}
+                                </div>
+                                <div className="mt-0.5 line-clamp-2 text-[length:var(--conversation-caption-font-size)] leading-(--conversation-caption-line-height) text-(--ui-text-tertiary)">
+                                  {theme.description}
+                                </div>
+                              </div>
+                            </button>
+                            {removable && (
+                              <button
+                                aria-label={a.removeTheme}
+                                className="absolute right-1.5 top-1.5 grid size-6 place-items-center rounded-md bg-(--ui-bg-elevated)/80 text-(--ui-text-tertiary) opacity-0 backdrop-blur-sm transition hover:text-(--ui-red) focus-visible:opacity-100 group-hover:opacity-100"
+                                onClick={() => {
+                                  triggerHaptic('crisp')
+                                  removeUserTheme(theme.name)
+
+                                  // Re-normalize off the now-missing skin → default.
+                                  if (active) {
+                                    setTheme(theme.name)
+                                  }
+                                }}
+                                title={a.removeTheme}
+                                type="button"
+                              >
+                                <Trash2 className="size-3.5" />
+                              </button>
+                            )}
+                          </div>
+                        )
+                      })}
+                    </div>
+                  )}
+                  <MarketplaceThemeResults
+                    installedExtIds={installedExtIds}
+                    onInstalled={name => setTheme(name)}
+                    query={query}
+                  />
+                </div>
+                {showProfileNote && (
+                  <p className="mt-3 text-[length:var(--conversation-caption-font-size)] leading-(--conversation-caption-line-height) text-(--ui-text-tertiary)">
+                    {a.themeProfileNote(activeProfileName)}
+                  </p>
+                )}
+              </>
             }
-            description={a.colorModeDesc}
-            title={a.colorMode}
+            description={a.themeDesc}
+            title={
+              <div className="flex items-center justify-between gap-3">
+                <span>{a.themeTitle}</span>
+                <SegmentedControl
+                  onChange={id => {
+                    triggerHaptic('crisp')
+                    setMode(id)
+                  }}
+                  options={modeOptions}
+                  value={mode}
+                />
+              </div>
+            }
+            wide
           />
 
           <ListRow
@@ -211,80 +411,6 @@ export function AppearanceSettings() {
             title={a.translucencyTitle}
           />
 
-          <ListRow
-            below={
-              <>
-                <div className="mt-3 grid gap-3 sm:grid-cols-2 xl:grid-cols-3">
-                  {availableThemes.map(theme => {
-                    const active = themeName === theme.name
-                    const removable = isUserTheme(theme.name)
-
-                    return (
-                      <div className="group relative" key={theme.name}>
-                        <button
-                          className={cn(
-                            'w-full rounded-lg border border-(--ui-stroke-tertiary) bg-(--ui-bg-quinary) p-2 text-left transition hover:bg-(--chrome-action-hover)',
-                            active && 'border-(--ui-stroke-secondary) bg-(--ui-bg-tertiary)'
-                          )}
-                          onClick={() => {
-                            triggerHaptic('crisp')
-                            setTheme(theme.name)
-                          }}
-                          type="button"
-                        >
-                          <ThemePreview name={theme.name} />
-                          <div className="mt-3 flex items-start justify-between gap-3 px-1">
-                            <div className="min-w-0">
-                              <div className="truncate text-[length:var(--conversation-text-font-size)] font-medium">
-                                {theme.label}
-                              </div>
-                              <div className="mt-0.5 line-clamp-2 text-[length:var(--conversation-caption-font-size)] leading-(--conversation-caption-line-height) text-(--ui-text-tertiary)">
-                                {theme.description}
-                              </div>
-                            </div>
-                            {active && (
-                              <span className="mt-0.5 grid size-5 shrink-0 place-items-center rounded-full bg-primary text-primary-foreground">
-                                <Check className="size-3.5" />
-                              </span>
-                            )}
-                          </div>
-                        </button>
-                        {removable && (
-                          <button
-                            aria-label={a.removeTheme}
-                            className="absolute right-1.5 top-1.5 grid size-6 place-items-center rounded-md bg-(--ui-bg-elevated)/80 text-(--ui-text-tertiary) opacity-0 backdrop-blur-sm transition hover:text-(--ui-red) focus-visible:opacity-100 group-hover:opacity-100"
-                            onClick={() => {
-                              triggerHaptic('crisp')
-                              removeUserTheme(theme.name)
-
-                              // Re-normalize off the now-missing skin → default.
-                              if (active) {
-                                setTheme(theme.name)
-                              }
-                            }}
-                            title={a.removeTheme}
-                            type="button"
-                          >
-                            <Trash2 className="size-3.5" />
-                          </button>
-                        )}
-                      </div>
-                    )
-                  })}
-                </div>
-                <VscodeThemeInstaller />
-                {showProfileNote && (
-                  <p className="mt-3 text-[length:var(--conversation-caption-font-size)] leading-(--conversation-caption-line-height) text-(--ui-text-tertiary)">
-                    {a.themeProfileNote(activeProfileName)}
-                  </p>
-                )}
-              </>
-            }
-            description={a.themeDesc}
-            title={a.themeTitle}
-            wide
-          />
-
           <ListRow
             action={
               <SegmentedControl
@@ -301,6 +427,10 @@ export function AppearanceSettings() {
           />
         </div>
       </div>
+
+      <div className="mt-6">
+        <PetSettings />
+      </div>
     </SettingsContent>
   )
 }

apps/desktop/src/app/settings/config-settings.tsx

@@ -26,6 +26,26 @@ import { ModelSettings } from './model-settings'
 import { EmptyState, ListRow, LoadingState, SettingsContent } from './primitives'
 import { ProviderConfigPanel } from './provider-config-panel'
 
+// On the Voice page, only surface the sub-fields of the *selected* TTS/STT
+// provider — otherwise every provider's options render at once (the "totally
+// crazy" wall of ~30 fields). Top-level keys (tts.provider, stt.enabled,
+// voice.*) always show; STT provider fields hide entirely when STT is off.
+export function voiceFieldVisible(key: string, config: HermesConfigRecord): boolean {
+  const match = /^(tts|stt)\.([^.]+)\./.exec(key)
+
+  if (!match) {
+    return true
+  }
+
+  const [, domain, provider] = match
+
+  if (domain === 'stt' && !getNested(config, 'stt.enabled')) {
+    return false
+  }
+
+  return provider === String(getNested(config, `${domain}.provider`) ?? '')
+}
+
 function ConfigField({
   schemaKey,
   schema,
@@ -356,6 +376,9 @@ export function ConfigSettings({
     return <LoadingState label={c.loading} />
   }
 
+  const visibleFields =
+    activeSectionId === 'voice' ? fields.filter(([key]) => voiceFieldVisible(key, config)) : fields
+
   return (
     <SettingsContent>
       {activeSectionId === 'model' && (
@@ -363,11 +386,11 @@ export function ConfigSettings({
           <ModelSettings onMainModelChanged={onMainModelChanged} />
         </div>
       )}
-      {fields.length === 0 ? (
+      {visibleFields.length === 0 ? (
         <EmptyState description={c.emptyDesc} title={c.emptyTitle} />
       ) : (
         <div className="grid gap-1">
-          {fields.map(([key, field]) => (
+          {visibleFields.map(([key, field]) => (
             <div className="scroll-mt-6 rounded-lg" id={`setting-field-${key}`} key={key}>
               <ConfigField
                 descriptionExtra={

apps/desktop/src/app/settings/pet-settings.tsx

@@ -0,0 +1,359 @@
+import { useStore } from '@nanostores/react'
+import { type ReactNode, useEffect, useState } from 'react'
+
+import { useGatewayRequest } from '@/app/gateway/hooks/use-gateway-request'
+import { PetThumb } from '@/components/pet/pet-thumb'
+import { Button } from '@/components/ui/button'
+import { ConfirmDialog } from '@/components/ui/confirm-dialog'
+import { Dialog, DialogContent, DialogFooter, DialogHeader, DialogTitle } from '@/components/ui/dialog'
+import { Input } from '@/components/ui/input'
+import { SegmentedControl } from '@/components/ui/segmented-control'
+import { useI18n } from '@/i18n'
+import { triggerHaptic } from '@/lib/haptics'
+import { Download, Loader2, PawPrint, Pencil, Trash2 } from '@/lib/icons'
+import { selectableCardClass } from '@/lib/selectable-card'
+import { cn } from '@/lib/utils'
+import { $petInfo } from '@/store/pet'
+import {
+  $petBusy,
+  $petGallery,
+  $petGalleryError,
+  $petGalleryStatus,
+  adoptPet,
+  exportPet as exportPetAction,
+  loadPetGallery,
+  loadPetThumb,
+  PET_SCALE_DEFAULT,
+  PET_SCALE_MAX,
+  PET_SCALE_MIN,
+  type GalleryPet,
+  rankedGalleryPets,
+  removePet as removePetAction,
+  renamePet as renamePetAction,
+  setPetEnabled,
+  setPetScale
+} from '@/store/pet-gallery'
+import { $gatewayState } from '@/store/session'
+
+import { ListRow, SectionHeading } from './primitives'
+
+/**
+ * Appearance opt-in for the floating petdex mascot. A thin view over the shared
+ * `pet-gallery` store — it subscribes to the atoms and calls the store actions,
+ * so the gallery is fetched once + cached and adopt/toggle/remove patch local
+ * state instead of re-pulling the network gallery. The floating mascot polls
+ * `pet.info`, so picking a pet here lights it up within a couple seconds.
+ */
+export function PetSettings() {
+  const { t } = useI18n()
+  const copy = t.settings.appearance.pet
+  const { requestGateway } = useGatewayRequest()
+  const gatewayState = useStore($gatewayState)
+  const gallery = useStore($petGallery)
+  const status = useStore($petGalleryStatus)
+  const error = useStore($petGalleryError)
+  const busySlug = useStore($petBusy)
+  const petInfo = useStore($petInfo)
+  const [query, setQuery] = useState('')
+  const [confirmDelete, setConfirmDelete] = useState<GalleryPet | null>(null)
+  const [renameTarget, setRenameTarget] = useState<GalleryPet | null>(null)
+  const [renameValue, setRenameValue] = useState('')
+  const scale = petInfo.scale ?? PET_SCALE_DEFAULT
+
+  useEffect(() => {
+    if (gatewayState !== 'open') {
+      return
+    }
+
+    void loadPetGallery(requestGateway)
+  }, [gatewayState, requestGateway])
+
+  const enabled = gallery?.enabled ?? false
+  const active = gallery?.active ?? ''
+  const pets = gallery?.pets ?? []
+  const staleBackend = status === 'stale'
+
+  const selectPet = (slug: string) => {
+    void adoptPet(requestGateway, slug, copy.adoptFailed(slug)).then(ok => ok && triggerHaptic('crisp'))
+  }
+
+  const removePet = (slug: string) => {
+    void removePetAction(requestGateway, slug, copy.uninstallFailed(slug)).then(ok => ok && triggerHaptic('crisp'))
+  }
+
+  const exportPet = (slug: string) => {
+    void exportPetAction(requestGateway, slug, copy.exportFailed(slug)).then(ok => ok && triggerHaptic('crisp'))
+  }
+
+  const saveRename = () => {
+    if (!renameTarget || !renameValue.trim()) {
+      return
+    }
+
+    // Optimistic: the rename paints instantly, so close now and let the RPC
+    // settle in the background (it rolls back + surfaces an error on failure).
+    const { slug } = renameTarget
+    setRenameTarget(null)
+    triggerHaptic('crisp')
+    void renamePetAction(requestGateway, slug, renameValue, copy.renameFailed(slug))
+  }
+
+  const toggle = (on: boolean) => {
+    void setPetEnabled(requestGateway, on, {
+      noneAvailable: copy.noneAvailable,
+      fallback: on ? copy.turnOnFailed : copy.turnOffFailed
+    }).then(ok => ok && triggerHaptic('crisp'))
+  }
+
+  // The petdex catalog is thousands of entries, so rank + cap how many render.
+  const RENDER_CAP = 60
+  const sorted = rankedGalleryPets(gallery, query)
+  const shown = sorted.slice(0, RENDER_CAP)
+
+  return (
+    <div>
+      <SectionHeading icon={PawPrint} title={copy.title} />
+      <p className="max-w-2xl text-[length:var(--conversation-caption-font-size)] leading-(--conversation-caption-line-height) text-(--ui-text-tertiary)">
+        {copy.intro}
+      </p>
+
+      {staleBackend && (
+        <p className="mt-2 rounded-lg border border-(--ui-stroke-tertiary) bg-(--ui-bg-quinary) px-3 py-2 text-[length:var(--conversation-caption-font-size)] leading-(--conversation-caption-line-height) text-(--ui-text-tertiary)">
+          {copy.restartHint}
+        </p>
+      )}
+
+      <div className="mt-2">
+        <ListRow
+          below={
+            <>
+              <input
+                className="mt-3 w-full rounded-lg border border-(--ui-stroke-tertiary) bg-(--ui-bg-quinary) px-3 py-1.5 text-[length:var(--conversation-caption-font-size)] outline-none placeholder:text-(--ui-text-tertiary) focus:border-(--ui-stroke-secondary)"
+                onChange={event => setQuery(event.target.value)}
+                placeholder={copy.searchPlaceholder}
+                spellCheck={false}
+                value={query}
+              />
+              {/* Fixed-height scroll area so filtering never grows/shrinks the
+                  page (no layout thrash); the grid scrolls inside it. */}
+              <div className="mt-3 h-72 overflow-y-auto pr-1">
+                {pets.length === 0 ? (
+                  <p className="text-[length:var(--conversation-caption-font-size)] text-(--ui-text-tertiary)">
+                    {copy.unreachable}
+                  </p>
+                ) : shown.length === 0 ? (
+                  <p className="text-[length:var(--conversation-caption-font-size)] text-(--ui-text-tertiary)">
+                    {copy.noMatch(query)}
+                  </p>
+                ) : (
+                  <div className="grid gap-2 sm:grid-cols-2 xl:grid-cols-3">
+                    {shown.map(pet => {
+                      const isActive = enabled && active === pet.slug
+                      const isBusy = busySlug === pet.slug
+
+                      return (
+                        <div className="group relative" key={pet.slug}>
+                          <button
+                            className={cn(
+                              'flex w-full items-center gap-2.5 px-2.5 py-2 text-left disabled:opacity-50',
+                              selectableCardClass({ active: isActive, prominent: pet.installed })
+                            )}
+                            disabled={isBusy}
+                            onClick={() => void selectPet(pet.slug)}
+                            type="button"
+                          >
+                            <PetThumb
+                              alt={pet.displayName}
+                              load={(slug, url) => loadPetThumb(requestGateway, slug, url)}
+                              slug={pet.slug}
+                              url={pet.spritesheetUrl}
+                            />
+                            <span className="min-w-0 flex-1">
+                              <span className="flex items-center gap-1.5">
+                                <span className="truncate text-[length:var(--conversation-text-font-size)] font-medium">
+                                  {pet.displayName}
+                                </span>
+                                {pet.generated && (
+                                  <span className="shrink-0 rounded-full bg-primary/15 px-1.5 py-px text-[0.625rem] font-medium text-primary">
+                                    {copy.generatedTag}
+                                  </span>
+                                )}
+                              </span>
+                              <span className="block truncate text-[length:var(--conversation-caption-font-size)] text-(--ui-text-tertiary)">
+                                {pet.slug}
+                                {pet.installed ? ` · ${copy.installedTag}` : ''}
+                              </span>
+                            </span>
+                            {isBusy && <Loader2 className="size-4 shrink-0 animate-spin text-(--ui-text-tertiary)" />}
+                          </button>
+                          {!isBusy && (pet.installed || pet.generated) && (
+                            <div className="absolute right-1.5 top-1.5 flex gap-1 opacity-0 transition focus-within:opacity-100 group-hover:opacity-100">
+                              {pet.generated && (
+                                <PetAction
+                                  icon={<Pencil className="size-3.5" />}
+                                  label={copy.rename(pet.displayName)}
+                                  onClick={() => {
+                                    setRenameValue(pet.displayName)
+                                    setRenameTarget(pet)
+                                  }}
+                                />
+                              )}
+                              {pet.generated && (
+                                <PetAction
+                                  icon={<Download className="size-3.5" />}
+                                  label={copy.exportPet(pet.displayName)}
+                                  onClick={() => exportPet(pet.slug)}
+                                />
+                              )}
+                              {pet.installed && (
+                                // Generated pets have no remote source — deletion is
+                                // permanent, so confirm; petdex pets just uninstall.
+                                <PetAction
+                                  danger
+                                  icon={<Trash2 className="size-3.5" />}
+                                  label={pet.generated ? copy.delete(pet.displayName) : copy.uninstall(pet.displayName)}
+                                  onClick={() => (pet.generated ? setConfirmDelete(pet) : removePet(pet.slug))}
+                                />
+                              )}
+                            </div>
+                          )}
+                        </div>
+                      )
+                    })}
+                  </div>
+                )}
+              </div>
+              {/* Always-present status line so its appearance never shifts layout. */}
+              <p className="mt-2 min-h-4 text-[length:var(--conversation-caption-font-size)] text-(--ui-text-tertiary)">
+                {error ? (
+                  <span className="text-(--ui-red)">{error}</span>
+                ) : sorted.length > RENDER_CAP ? (
+                  copy.countCapped(RENDER_CAP, sorted.length)
+                ) : (
+                  copy.count(sorted.length)
+                )}
+              </p>
+            </>
+          }
+          description={copy.chooseDesc}
+          title={
+            <div className="flex items-center justify-between gap-3">
+              <span>{copy.chooseTitle}</span>
+              <SegmentedControl
+                onChange={id => void toggle(id === 'on')}
+                options={[
+                  { id: 'off', label: copy.off },
+                  { id: 'on', label: copy.on }
+                ]}
+                value={enabled ? 'on' : 'off'}
+              />
+            </div>
+          }
+          wide
+        />
+
+        {enabled && (
+          <ListRow
+            action={
+              <div className="flex items-center gap-3">
+                <input
+                  aria-label={copy.scaleTitle}
+                  className="h-1 w-40 cursor-pointer appearance-none rounded-full bg-(--ui-stroke-tertiary)"
+                  max={PET_SCALE_MAX}
+                  min={PET_SCALE_MIN}
+                  onChange={event => {
+                    triggerHaptic('selection')
+                    setPetScale(requestGateway, Number(event.target.value))
+                  }}
+                  step={0.05}
+                  style={{ accentColor: 'var(--dt-primary)' }}
+                  type="range"
+                  value={scale}
+                />
+                <span className="w-9 text-right text-[length:var(--conversation-caption-font-size)] tabular-nums text-(--ui-text-tertiary)">
+                  {`${Math.round(scale * 100)}%`}
+                </span>
+              </div>
+            }
+            description={copy.scaleDesc}
+            title={copy.scaleTitle}
+          />
+        )}
+      </div>
+
+      <ConfirmDialog
+        confirmLabel={copy.deleteConfirm}
+        description={copy.deleteBody}
+        destructive
+        onClose={() => setConfirmDelete(null)}
+        onConfirm={async () => {
+          if (confirmDelete) {
+            const ok = await removePetAction(requestGateway, confirmDelete.slug, copy.uninstallFailed(confirmDelete.slug))
+            if (!ok) {
+              throw new Error(copy.uninstallFailed(confirmDelete.slug))
+            }
+            triggerHaptic('crisp')
+          }
+        }}
+        open={confirmDelete !== null}
+        title={confirmDelete ? copy.deleteTitle(confirmDelete.displayName) : ''}
+      />
+
+      <Dialog onOpenChange={open => !open && setRenameTarget(null)} open={renameTarget !== null}>
+        <DialogContent className="max-w-sm">
+          <DialogHeader>
+            <DialogTitle>{copy.renameTitle}</DialogTitle>
+          </DialogHeader>
+          <Input
+            autoFocus
+            onChange={event => setRenameValue(event.target.value)}
+            onKeyDown={event => {
+              if (event.key === 'Enter') {
+                event.preventDefault()
+                saveRename()
+              }
+            }}
+            placeholder={copy.renamePlaceholder}
+            value={renameValue}
+          />
+          <DialogFooter>
+            <Button onClick={() => setRenameTarget(null)} type="button" variant="ghost">
+              {t.common.cancel}
+            </Button>
+            <Button disabled={!renameValue.trim()} onClick={saveRename}>
+              {copy.renameSave}
+            </Button>
+          </DialogFooter>
+        </DialogContent>
+      </Dialog>
+    </div>
+  )
+}
+
+/** A single hover-revealed icon action on a pet card (rename / export / delete). */
+function PetAction({
+  danger,
+  icon,
+  label,
+  onClick
+}: {
+  danger?: boolean
+  icon: ReactNode
+  label: string
+  onClick: () => void
+}) {
+  return (
+    <button
+      aria-label={label}
+      className={cn(
+        'grid size-6 place-items-center rounded-md bg-(--ui-bg-elevated)/80 text-(--ui-text-tertiary) backdrop-blur-sm transition',
+        danger ? 'hover:text-(--ui-red)' : 'hover:text-foreground'
+      )}
+      onClick={onClick}
+      title={label}
+      type="button"
+    >
+      {icon}
+    </button>
+  )
+}

apps/desktop/src/app/settings/voice-field-visible.test.ts

@@ -0,0 +1,48 @@
+import { describe, expect, it } from 'vitest'
+
+import type { HermesConfigRecord } from '@/types/hermes'
+
+import { voiceFieldVisible } from './config-settings'
+
+const cfg = (over: Record<string, unknown> = {}): HermesConfigRecord =>
+  ({
+    tts: { provider: 'edge', edge: {}, openai: {} },
+    stt: { enabled: true, provider: 'local', local: {}, groq: {} },
+    ...over
+  }) as unknown as HermesConfigRecord
+
+describe('voiceFieldVisible', () => {
+  it('always shows top-level + non-provider keys', () => {
+    const config = cfg()
+
+    for (const key of ['tts.provider', 'stt.enabled', 'stt.provider', 'voice.auto_tts', 'voice.record_key']) {
+      expect(voiceFieldVisible(key, config)).toBe(true)
+    }
+  })
+
+  it('shows only the selected TTS provider sub-fields', () => {
+    const config = cfg()
+    expect(voiceFieldVisible('tts.edge.voice', config)).toBe(true)
+    expect(voiceFieldVisible('tts.openai.voice', config)).toBe(false)
+    expect(voiceFieldVisible('tts.elevenlabs.voice_id', config)).toBe(false)
+  })
+
+  it('shows only the selected STT provider sub-fields', () => {
+    const config = cfg()
+    expect(voiceFieldVisible('stt.local.model', config)).toBe(true)
+    expect(voiceFieldVisible('stt.groq.model', config)).toBe(false)
+  })
+
+  it('hides every STT provider sub-field when STT is disabled', () => {
+    const config = cfg({ stt: { enabled: false, provider: 'local', local: {} } })
+    expect(voiceFieldVisible('stt.local.model', config)).toBe(false)
+    // ...but the enable/provider toggles themselves stay visible.
+    expect(voiceFieldVisible('stt.enabled', config)).toBe(true)
+    expect(voiceFieldVisible('stt.provider', config)).toBe(true)
+  })
+
+  it('tracks a provider switch', () => {
+    expect(voiceFieldVisible('tts.openai.voice', cfg({ tts: { provider: 'openai', openai: {} } }))).toBe(true)
+    expect(voiceFieldVisible('tts.edge.voice', cfg({ tts: { provider: 'openai', openai: {} } }))).toBe(false)
+  })
+})

apps/desktop/src/app/shell/app-shell.tsx

@@ -4,6 +4,7 @@ import { useSyncExternalStore } from 'react'
 
 import { NotificationStack } from '@/components/notifications'
 import { PaneShell } from '@/components/pane-shell'
+import { FloatingPet } from '@/components/pet/floating-pet'
 import { SidebarProvider } from '@/components/ui/sidebar'
 import { useMediaQuery } from '@/hooks/use-media-query'
 import {
@@ -202,6 +203,10 @@ export function AppShell({
       {/* Mounted at the shell root (after overlays) so success/error toasts
           surface above every route and overlay — not just the chat view. */}
       <NotificationStack />
+
+      {/* Petdex floating mascot — in-window, always-on-top, reactive to agent
+          activity. Renders nothing unless a pet is installed + enabled. */}
+      <FloatingPet />
     </SidebarProvider>
   )
 }

apps/desktop/src/app/shell/hooks/use-statusbar-items.tsx

@@ -22,8 +22,6 @@ import type { RuntimeReadinessResult } from '@/lib/runtime-readiness'
 import { contextBarLabel, LiveDuration, usageContextLabel } from '@/lib/statusbar'
 import { cn } from '@/lib/utils'
 import { setGlobalYolo, setSessionYolo } from '@/lib/yolo-session'
-import { $desktopActionTasks } from '@/store/activity'
-import { $previewServerRestartStatus } from '@/store/preview'
 import {
   $activeSessionId,
   $busy,
@@ -31,11 +29,10 @@ import {
   $currentUsage,
   $sessionStartedAt,
   $turnStartedAt,
-  $workingSessionIds,
   $yoloActive,
   setYoloActive
 } from '@/store/session'
-import { $subagentsBySession, activeSubagentCount } from '@/store/subagents'
+import { $subagentsBySession, activeSubagentCount, failedSubagentCount } from '@/store/subagents'
 import { $gatewayRestarting } from '@/store/system-actions'
 import {
   $backendUpdateApply,
@@ -90,12 +87,9 @@ export function useStatusbarItems({
   const yoloActive = useStore($yoloActive)
   const busy = useStore($busy)
   const currentUsage = useStore($currentUsage)
-  const desktopActionTasks = useStore($desktopActionTasks)
   const gatewayRestarting = useStore($gatewayRestarting)
-  const previewServerRestartStatus = useStore($previewServerRestartStatus)
   const sessionStartedAt = useStore($sessionStartedAt)
   const turnStartedAt = useStore($turnStartedAt)
-  const workingSessionIds = useStore($workingSessionIds)
   const subagentsBySession = useStore($subagentsBySession)
   const updateStatus = useStore($updateStatus)
   const updateApply = useStore($updateApply)
@@ -159,24 +153,17 @@ export function useStatusbarItems({
     [gatewayLogLines, gatewayState, inferenceStatus, openCommandCenterSection, statusSnapshot]
   )
 
-  const { bgFailed, bgRunning, subagentsRunning } = useMemo(() => {
-    const actions = Object.values(desktopActionTasks)
-    const running = actions.filter(t => t.status.running).length
-    const failed = actions.filter(t => !t.status.running && (t.status.exit_code ?? 0) !== 0).length
-    const previewRunning = previewServerRestartStatus === 'running' ? 1 : 0
-    const previewFailed = previewServerRestartStatus === 'error' ? 1 : 0
-
-    const subagentsRunning = Object.values(subagentsBySession).reduce(
-      (sum, items) => sum + activeSubagentCount(items),
-      0
-    )
+  // The indicator must speak the same scope as the Spawn-tree panel it opens:
+  // every session's subagents, never background system actions (gateway
+  // restarts, toolset installs) which surface in their own panels.
+  const { subagentsFailed, subagentsRunning } = useMemo(() => {
+    const lists = Object.values(subagentsBySession)
 
     return {
-      bgFailed: failed + previewFailed,
-      bgRunning: workingSessionIds.length + running + previewRunning,
-      subagentsRunning
+      subagentsFailed: lists.reduce((sum, items) => sum + failedSubagentCount(items), 0),
+      subagentsRunning: lists.reduce((sum, items) => sum + activeSubagentCount(items), 0)
     }
-  }, [desktopActionTasks, previewServerRestartStatus, subagentsBySession, workingSessionIds])
+  }, [subagentsBySession])
 
   const gatewayOpen = gatewayState === 'open'
   const gatewayConnecting = gatewayState === 'connecting'
@@ -321,20 +308,18 @@ export function useStatusbarItems({
       {
         className: cn(
           agentsOpen && 'bg-accent/55 text-foreground',
-          bgFailed > 0 && 'text-destructive hover:text-destructive'
+          subagentsFailed > 0 && 'text-destructive hover:text-destructive'
         ),
         detail:
           subagentsRunning > 0
             ? copy.subagents(subagentsRunning)
-            : bgFailed > 0
-              ? copy.failed(bgFailed)
-              : bgRunning > 0
-                ? copy.running(bgRunning)
-                : undefined,
+            : subagentsFailed > 0
+              ? copy.failed(subagentsFailed)
+              : undefined,
         icon:
-          bgFailed > 0 ? (
+          subagentsFailed > 0 ? (
             <AlertCircle className="size-3" />
-          ) : bgRunning > 0 || subagentsRunning > 0 ? (
+          ) : subagentsRunning > 0 ? (
             <Loader2 className="size-3 animate-spin" />
           ) : (
             <Sparkles className="size-3" />
@@ -356,8 +341,6 @@ export function useStatusbarItems({
     ],
     [
       agentsOpen,
-      bgFailed,
-      bgRunning,
       commandCenterOpen,
       copy,
       gatewayMenuContent,
@@ -367,6 +350,7 @@ export function useStatusbarItems({
       inferenceReady,
       inferenceStatus?.reason,
       openAgents,
+      subagentsFailed,
       subagentsRunning,
       toggleCommandCenter
     ]

apps/desktop/src/app/updates-overlay.tsx

@@ -382,6 +382,8 @@ function ApplyingView({ apply, isBackend }: { apply: UpdateApplyState; isBackend
   const u = t.updates
   const label = u.stages[apply.stage as DesktopUpdateStage] ?? u.stages.idle
   const body = isBackend ? u.applyingBodyBackend : u.applyingBody
+  const currentMessage = apply.message.trim()
+  const recentLog = apply.log.slice(-4)
 
   const percent =
     typeof apply.percent === 'number' && Number.isFinite(apply.percent)
@@ -397,6 +399,12 @@ function ApplyingView({ apply, isBackend }: { apply: UpdateApplyState; isBackend
         <DialogDescription className="text-center text-sm">
           {body}
         </DialogDescription>
+
+        {currentMessage ? (
+          <p className="max-w-lg break-words text-center text-xs leading-5 text-muted-foreground">
+            {currentMessage}
+          </p>
+        ) : null}
       </div>
 
       <div className="h-2 overflow-hidden rounded-full bg-muted">
@@ -409,6 +417,16 @@ function ApplyingView({ apply, isBackend }: { apply: UpdateApplyState; isBackend
         />
       </div>
 
+      {recentLog.length > 1 ? (
+        <div className="max-h-24 overflow-hidden rounded-md border border-border/70 bg-muted/35 px-3 py-2 text-left font-mono text-[11px] leading-4 text-muted-foreground">
+          {recentLog.map((entry, index) => (
+            <div className="truncate" key={`${entry.at}-${index}`}>
+              {entry.message}
+            </div>
+          ))}
+        </div>
+      ) : null}
+
       <p className="text-center text-xs text-muted-foreground">{u.applyingClose}</p>
     </div>
   )

apps/desktop/src/components/assistant-ui/tool-fallback-model.test.ts

@@ -1,9 +1,13 @@
-import { describe, expect, it } from 'vitest'
+import { afterEach, describe, expect, it } from 'vitest'
+
+import { setRuntimeI18nLocale } from '@/i18n'
 
 import {
   buildToolView,
+  clampForDisplay,
   countDiffLineStats,
   inlineDiffFromResult,
+  MAX_TOOL_RENDER_CHARS,
   type ToolPart
 } from './tool-fallback-model'
 
@@ -17,6 +21,10 @@ const part = (overrides: Partial<ToolPart>): ToolPart => ({
   ...overrides
 })
 
+afterEach(() => {
+  setRuntimeI18nLocale('en')
+})
+
 describe('buildToolView image handling', () => {
   // vision_analyze reports the input image as a local path; an <img> pointed at
   // a bare path resolves against the renderer origin and 404s, so we render the
@@ -40,8 +48,7 @@ describe('buildToolView image handling', () => {
 })
 
 describe('buildToolView terminal exit-code status', () => {
-  const terminal = (result: Record<string, unknown>) =>
-    buildToolView(part({ result, toolName: 'terminal' }), '')
+  const terminal = (result: Record<string, unknown>) => buildToolView(part({ result, toolName: 'terminal' }), '')
 
   // A non-zero exit code with real output is not a failure (grep no-match,
   // diff differences, piped commands surfacing the last stage's code, etc.) —
@@ -110,6 +117,207 @@ describe('buildToolView file edit diffs', () => {
   })
 })
 
+describe('buildToolView title actions', () => {
+  it('marks the pending action separately from the rest of the title', () => {
+    const read = buildToolView(part({ args: { path: '/tmp/demo.txt' }, result: undefined, toolName: 'read_file' }), '')
+
+    const web = buildToolView(
+      part({ args: { url: 'https://example.com/docs' }, result: undefined, toolName: 'web_extract' }),
+      ''
+    )
+
+    const terminal = buildToolView(
+      part({ args: { command: 'npm test -- --runInBand' }, result: undefined, toolName: 'terminal' }),
+      ''
+    )
+
+    const code = buildToolView(
+      part({ args: { code: 'print("hello")' }, result: undefined, toolName: 'execute_code' }),
+      ''
+    )
+
+    expect(read.title).toBe('Reading demo.txt')
+    expect(read.titleAction).toEqual({ prefix: '', text: 'Reading', suffix: ' demo.txt' })
+    expect(web.title).toBe('Reading example.com/docs')
+    expect(web.titleAction).toEqual({ prefix: '', text: 'Reading', suffix: ' example.com/docs' })
+    expect(terminal.title).toBe('Running npm test -- --runInBand')
+    expect(terminal.titleAction).toEqual({ prefix: '', text: 'Running', suffix: ' npm test -- --runInBand' })
+    expect(code.title).toBe('Scripting print("hello")')
+    expect(code.titleAction).toEqual({ prefix: '', text: 'Scripting', suffix: ' print("hello")' })
+  })
+
+  it('does not mark completed tool titles as pending actions', () => {
+    const view = buildToolView(part({ args: { url: 'https://example.com/docs' }, toolName: 'web_extract' }), '')
+
+    expect(view.title).toBe('Read example.com/docs')
+    expect(view.titleAction).toBeUndefined()
+  })
+
+  it('uses the filename for completed read_file rows', () => {
+    const view = buildToolView(
+      part({ args: { path: './package.json' }, result: { content: '1|{"name":"demo"}' }, toolName: 'read_file' }),
+      ''
+    )
+
+    expect(view.title).toBe('Read package.json')
+    expect(view.subtitle).toBe('')
+    expect(view.titleAction).toBeUndefined()
+  })
+
+  it('adds a compact line range to line-scoped read_file rows', () => {
+    const view = buildToolView(
+      part({
+        args: { limit: 10, offset: 25, path: './src/main.ts' },
+        result: { content: '25|function toggleDock() {\n26|  dock.classList.toggle("hidden");\n34|}' },
+        toolName: 'read_file'
+      }),
+      ''
+    )
+
+    expect(view.title).toBe('Read main.ts L25-34')
+    expect(view.subtitle).toBe('')
+  })
+
+  it('uses the requested positive offset/limit for read_file row line ranges', () => {
+    const view = buildToolView(
+      part({
+        args: { limit: 5, offset: 1, path: './package.json' },
+        result: { content: '1|{\n2|  "name": "bb-rainbows",\n3|  "private": true,\n4|  "version": "0.0.1",\n5|  "type": "module",\n6|  "description": "extra"' },
+        toolName: 'read_file'
+      }),
+      ''
+    )
+
+    expect(view.title).toBe('Read package.json L1-5')
+  })
+
+  it('uses inherited backend context for live read_file rows', () => {
+    const view = buildToolView(
+      part({
+        args: { context: 'package.json L1-5', path: './package.json' },
+        result: undefined,
+        toolName: 'read_file'
+      }),
+      ''
+    )
+
+    expect(view.title).toBe('Reading package.json L1-5')
+    expect(view.titleAction).toEqual({ prefix: '', text: 'Reading', suffix: ' package.json L1-5' })
+  })
+
+  it('uses returned line numbers for negative-offset read_file rows', () => {
+    const view = buildToolView(
+      part({
+        args: { limit: 2, offset: -2, path: './src/main.ts' },
+        result: { content: '99|lastLine();\n100|done();' },
+        toolName: 'read_file'
+      }),
+      ''
+    )
+
+    expect(view.title).toBe('Read main.ts L99-100')
+  })
+
+  it('renders compact terminal titles for session 20260624_231846_bdbd1e commands', () => {
+    const rows = [
+      [
+        'cd /Users/brooklyn/www/bb-rainbows && pnpm run lint 2>&1 | tail -20; echo "lint_exit=${PIPESTATUS[0]}"',
+        'Ran pnpm run lint'
+      ],
+      [
+        'cd /Users/brooklyn/www/bb-rainbows && pnpm run build 2>&1 | tail -20; echo "build_exit=${PIPESTATUS[0]}"',
+        'Ran pnpm run build'
+      ],
+      [
+        'which node pnpm corepack; node -v; echo "---"; corepack --version 2>&1; echo "---pnpm via corepack---"; pnpm --version 2>&1 | tail -5',
+        'Ran which node pnpm corepack + 3 commands'
+      ],
+      [
+        'echo "--- proto pnpm direct ---"; ~/.proto/tools/node/24.11.0/bin/pnpm --version 2>&1 | tail -3; echo "--- proto node ---"; ls ~/.proto/tools/node/ 2>&1; echo "--- corepack cache ---"; ls ~/.cache/node/corepack/v1/pnpm/ 2>&1',
+        'Ran ~/.proto/tools/node/24.11.0/bin/pnpm --version + 2 commands'
+      ],
+      [
+        'cd /Users/brooklyn/www/bb-rainbows && COREPACK_ENABLE_DOWNLOAD_PROMPT=0 corepack pnpm@10.20.0 --version 2>&1 | tail -3',
+        'Ran COREPACK_ENABLE_DOWNLOAD_PROMPT=0 corepack pnpm@10.20.0 --version'
+      ],
+      [
+        'cd /Users/brooklyn/www/bb-rainbows && COREPACK_ENABLE_DOWNLOAD_PROMPT=0 corepack use pnpm@10.20.0 2>&1 | tail -10; echo "exit=$?"',
+        'Ran COREPACK_ENABLE_DOWNLOAD_PROMPT=0 corepack use pnpm@10.20.0'
+      ]
+    ] as const
+
+    for (const [command, expectedTitle] of rows) {
+      const view = buildToolView(part({ args: { command }, result: { output: 'ok', exit_code: 0 }, toolName: 'terminal' }), '')
+
+      expect(view.title).toBe(expectedTitle)
+    }
+  })
+
+  it('uses inherited backend context for live terminal rows', () => {
+    const view = buildToolView(
+      part({
+        args: {
+          command: 'cd /Users/brooklyn/www/bb-rainbows && pnpm run lint 2>&1 | tail -20',
+          context: 'pnpm run lint'
+        },
+        result: undefined,
+        toolName: 'terminal'
+      }),
+      ''
+    )
+
+    expect(view.title).toBe('Running pnpm run lint')
+    expect(view.subtitle).toBe('')
+    expect(view.titleAction).toEqual({ prefix: '', text: 'Running', suffix: ' pnpm run lint' })
+  })
+
+  it('uses the runtime locale for title text and action placement', () => {
+    setRuntimeI18nLocale('ja')
+
+    const read = buildToolView(part({ args: { path: '/tmp/demo.txt' }, result: undefined, toolName: 'read_file' }), '')
+
+    const web = buildToolView(
+      part({ args: { url: 'https://example.com/docs' }, result: undefined, toolName: 'web_extract' }),
+      ''
+    )
+
+    expect(read.title).toBe('demo.txt を読み取り中')
+    expect(read.titleAction).toEqual({ prefix: 'demo.txt を', text: '読み取り中', suffix: '' })
+    expect(web.title).toBe('example.com/docs を読み取り中')
+    expect(web.titleAction).toEqual({ prefix: 'example.com/docs を', text: '読み取り中', suffix: '' })
+  })
+})
+
+describe('clampForDisplay', () => {
+  it('passes short payloads through untouched', () => {
+    expect(clampForDisplay('hello')).toBe('hello')
+    expect(clampForDisplay('x'.repeat(MAX_TOOL_RENDER_CHARS))).toHaveLength(MAX_TOOL_RENDER_CHARS)
+  })
+
+  it('truncates oversized payloads and reports the omitted count', () => {
+    const oversized = 'x'.repeat(MAX_TOOL_RENDER_CHARS + 5_000)
+    const clamped = clampForDisplay(oversized)
+
+    expect(clamped.length).toBeLessThan(oversized.length)
+    expect(clamped.startsWith('x'.repeat(MAX_TOOL_RENDER_CHARS))).toBe(true)
+    expect(clamped).toContain('5,000 more characters truncated')
+    expect(clamped).toContain('Copy')
+  })
+})
+
+// A large tool result (e.g. a 100KB read_file during a `/learn` run) must not
+// be serialized into the rendered rawResult at full size — that JSON.stringify
+// payload is what floods the renderer when many rows stack up.
+describe('buildToolView caps serialized result size', () => {
+  it('clamps rawResult for an oversized result', () => {
+    const huge = 'y'.repeat(MAX_TOOL_RENDER_CHARS * 3)
+    const view = buildToolView(part({ result: { content: huge }, toolName: 'read_file' }), '')
+
+    expect(view.rawResult.length).toBeLessThanOrEqual(MAX_TOOL_RENDER_CHARS + 200)
+    expect(view.rawResult).toContain('truncated')
+  })
+})
+
 describe('countDiffLineStats', () => {
   it('counts added and removed lines', () => {
     expect(

apps/desktop/src/components/assistant-ui/tool-fallback-model.ts

@@ -1,6 +1,7 @@
+import { type ToolTitleKey, translateNow } from '@/i18n'
 import { normalizeExternalUrl } from '@/lib/external-link'
+import { summarizeShellCommand } from '@/lib/summarize-command'
 import { extractToolErrorMessage, formatToolResultSummary } from '@/lib/tool-result-summary'
-import { translateNow } from '@/i18n'
 
 export type ToolTone = 'agent' | 'browser' | 'default' | 'file' | 'image' | 'terminal' | 'web'
 export type ToolStatus = 'error' | 'running' | 'success' | 'warning'
@@ -20,6 +21,12 @@ export interface SearchResultRow {
   url: string
 }
 
+export interface ToolTitleAction {
+  prefix: string
+  suffix: string
+  text: string
+}
+
 interface CountMetric {
   count: number
   noun: string
@@ -51,6 +58,7 @@ export interface ToolView {
   status: ToolStatus
   subtitle: string
   title: string
+  titleAction?: ToolTitleAction
   tone: ToolTone
 }
 
@@ -58,6 +66,12 @@ interface ToolMeta {
   done: string
   icon?: string
   pending: string
+  pendingAction: string
+  tone: ToolTone
+}
+
+interface ToolMetaSpec {
+  icon?: string
   tone: ToolTone
 }
 
@@ -112,44 +126,135 @@ function fileEditBasename(path: string): string {
   return normalized.split('/').filter(Boolean).pop() || normalized
 }
 
-const TOOL_META: Record<string, ToolMeta> = {
-  browser_click: { done: 'Clicked page element', pending: 'Clicking page element', icon: 'globe', tone: 'browser' },
-  browser_fill: { done: 'Filled form field', pending: 'Filling form field', icon: 'globe', tone: 'browser' },
-  browser_navigate: { done: 'Opened page', pending: 'Opening page', icon: 'globe', tone: 'browser' },
+function numericField(record: Record<string, unknown>, key: string): number | undefined {
+  const value = record[key]
+
+  return typeof value === 'number' && Number.isFinite(value) ? value : undefined
+}
+
+function readFileLineLabel(args: Record<string, unknown>, result: Record<string, unknown>): string {
+  if (numericField(args, 'offset') === undefined && numericField(args, 'limit') === undefined) {
+    return ''
+  }
+
+  const content = firstStringField(result, ['content'])
+  const offset = numericField(args, 'offset')
+  const limit = numericField(args, 'limit')
+
+  if (offset !== undefined && offset > 0) {
+    if (limit === undefined || limit <= 1) {
+      return `L${offset}`
+    }
+
+    return `L${offset}-${offset + limit - 1}`
+  }
+
+  const lines = content
+    .split('\n')
+    .map(line => /^(\d+)\|/.exec(line)?.[1])
+    .filter((line): line is string => !!line)
+    .map(Number)
+
+  if (lines.length === 0) {
+    return ''
+  }
+
+  const start = lines[0]!
+  const end = lines[lines.length - 1]!
+
+  return start === end ? `L${start}` : `L${start}-${end}`
+}
+
+function readFileDisplayTarget(args: Record<string, unknown>, result: Record<string, unknown>): string {
+  const inherited = firstStringField(args, ['context', 'preview'])
+
+  if (inherited) {
+    return inherited
+  }
+
+  const path = firstStringField(args, ['path', 'file', 'filepath'])
+
+  if (!path) {
+    return ''
+  }
+
+  const lineLabel = readFileLineLabel(args, result)
+
+  return [fileEditBasename(path), lineLabel].filter(Boolean).join(' ')
+}
+
+const TOOL_META: Record<ToolTitleKey, ToolMetaSpec> = {
+  browser_click: {
+    icon: 'globe',
+    tone: 'browser'
+  },
+  browser_fill: {
+    icon: 'globe',
+    tone: 'browser'
+  },
+  browser_navigate: {
+    icon: 'globe',
+    tone: 'browser'
+  },
   browser_snapshot: {
-    done: 'Captured page snapshot',
-    pending: 'Capturing page snapshot',
     icon: 'globe',
     tone: 'browser'
   },
   browser_take_screenshot: {
-    done: 'Captured screenshot',
-    pending: 'Capturing screenshot',
     icon: 'file-media',
     tone: 'browser'
   },
-  browser_type: { done: 'Typed on page', pending: 'Typing on page', icon: 'globe', tone: 'browser' },
-  clarify: { done: 'Asked a question', pending: 'Asking a question', icon: 'question', tone: 'agent' },
-  cronjob: { done: 'Cron job', pending: 'Scheduling cron job', icon: 'watch', tone: 'agent' },
-  edit_file: { done: 'Edited file', pending: 'Editing file', icon: 'edit', tone: 'file' },
-  execute_code: { done: 'Ran code', pending: 'Running code', icon: 'terminal', tone: 'terminal' },
-  image_generate: { done: 'Generated image', pending: 'Generating image', icon: 'file-media', tone: 'image' },
-  list_files: { done: 'Listed files', pending: 'Listing files', icon: 'files', tone: 'file' },
-  patch: { done: 'Patched file', pending: 'Patching file', icon: 'edit', tone: 'file' },
-  read_file: { done: 'Read file', pending: 'Reading file', icon: 'file', tone: 'file' },
-  search_files: { done: 'Searched files', pending: 'Searching files', icon: 'search', tone: 'file' },
+  browser_type: {
+    icon: 'globe',
+    tone: 'browser'
+  },
+  clarify: {
+    icon: 'question',
+    tone: 'agent'
+  },
+  cronjob: {
+    icon: 'watch',
+    tone: 'agent'
+  },
+  edit_file: { icon: 'edit', tone: 'file' },
+  execute_code: {
+    icon: 'terminal',
+    tone: 'terminal'
+  },
+  image_generate: {
+    icon: 'file-media',
+    tone: 'image'
+  },
+  list_files: {
+    icon: 'files',
+    tone: 'file'
+  },
+  patch: { icon: 'edit', tone: 'file' },
+  read_file: { icon: 'file', tone: 'file' },
+  search_files: {
+    icon: 'search',
+    tone: 'file'
+  },
   session_search_recall: {
-    done: 'Searched session history',
-    pending: 'Searching session history',
     icon: 'search',
     tone: 'agent'
   },
-  terminal: { done: 'Ran command', pending: 'Running command', icon: 'terminal', tone: 'terminal' },
-  todo: { done: 'Updated todos', pending: 'Updating todos', icon: 'tools', tone: 'agent' },
-  vision_analyze: { done: 'Analyzed image', pending: 'Analyzing image', icon: 'eye', tone: 'image' },
-  web_extract: { done: 'Read webpage', pending: 'Reading webpage', icon: 'globe', tone: 'web' },
-  web_search: { done: 'Searched web', pending: 'Searching web', icon: 'search', tone: 'web' },
-  write_file: { done: 'Edited file', pending: 'Editing file', icon: 'edit', tone: 'file' }
+  terminal: {
+    icon: 'terminal',
+    tone: 'terminal'
+  },
+  todo: { icon: 'tools', tone: 'agent' },
+  vision_analyze: {
+    icon: 'eye',
+    tone: 'image'
+  },
+  web_extract: { icon: 'globe', tone: 'web' },
+  web_search: { icon: 'search', tone: 'web' },
+  write_file: { icon: 'edit', tone: 'file' }
+}
+
+function isToolTitleKey(name: string): name is ToolTitleKey {
+  return name in TOOL_META
 }
 
 const INLINE_CODE_SPLIT_RE = /(`[^`\n]+`)/g
@@ -171,27 +276,45 @@ function titleForTool(name: string): string {
   )
 }
 
-const PREFIX_META: { icon?: string; prefix: string; tone: ToolTone; verb: string }[] = [
-  { prefix: 'browser_', verb: 'Browser', icon: 'globe', tone: 'browser' },
-  { prefix: 'web_', verb: 'Web', icon: 'globe', tone: 'web' }
+const PREFIX_META: { icon?: string; labelKey: string; prefix: string; tone: ToolTone }[] = [
+  { prefix: 'browser_', labelKey: 'browser', icon: 'globe', tone: 'browser' },
+  { prefix: 'web_', labelKey: 'web', icon: 'globe', tone: 'web' }
 ]
 
 function toolMeta(name: string): ToolMeta {
-  if (TOOL_META[name]) {
-    return TOOL_META[name]
+  if (isToolTitleKey(name)) {
+    const meta = TOOL_META[name]
+
+    return {
+      done: translateNow(`assistant.tool.titles.${name}.done`),
+      pending: translateNow(`assistant.tool.titles.${name}.pending`),
+      pendingAction: translateNow(`assistant.tool.titles.${name}.pendingAction`),
+      icon: meta.icon,
+      tone: meta.tone
+    }
   }
 
   const action = titleForTool(name)
   const prefix = PREFIX_META.find(p => name.startsWith(p.prefix))
 
-  return prefix
-    ? {
-        done: `${prefix.verb} ${action}`,
-        pending: `Running ${prefix.verb.toLowerCase()} ${action.toLowerCase()}`,
-        icon: prefix.icon,
-        tone: prefix.tone
-      }
-    : { done: action, pending: `Running ${action.toLowerCase()}`, tone: 'default' }
+  if (prefix) {
+    const prefixLabel = translateNow(`assistant.tool.prefixes.${prefix.labelKey}`)
+
+    return {
+      done: translateNow('assistant.tool.titleTemplates.prefixedDone', prefixLabel, action),
+      pending: translateNow('assistant.tool.titleTemplates.runningPrefixedTool', prefixLabel, action),
+      pendingAction: translateNow('assistant.tool.actions.running'),
+      icon: prefix.icon,
+      tone: prefix.tone
+    }
+  }
+
+  return {
+    done: action,
+    pending: translateNow('assistant.tool.titleTemplates.runningTool', action),
+    pendingAction: translateNow('assistant.tool.actions.running'),
+    tone: 'default'
+  }
 }
 
 function isRecord(value: unknown): value is Record<string, unknown> {
@@ -238,8 +361,26 @@ function contextValue(value: unknown): string {
   return typeof value === 'string' ? value : ''
 }
 
+// Each tool result is server-capped (~100KB), but a turn over a big directory
+// stacks many rows; painting/serializing them all floods the renderer (freeze,
+// then OOM). Clamp every inline-painted payload to a bounded slice — the row's
+// Copy button still reads the uncapped `view.detail` for the full output.
+export const MAX_TOOL_RENDER_CHARS = 20_000
+
+export function clampForDisplay(value: string, max = MAX_TOOL_RENDER_CHARS): string {
+  if (value.length <= max) {
+    return value
+  }
+
+  const omitted = value.length - max
+
+  return `${value.slice(0, max)}\n\n… ${omitted.toLocaleString()} more characters truncated — use Copy for the full output.`
+}
+
 function prettyJson(value: unknown): string {
-  return typeof value === 'string' ? value : JSON.stringify(value, null, 2)
+  const raw = typeof value === 'string' ? value : JSON.stringify(value, null, 2)
+
+  return clampForDisplay(raw ?? '')
 }
 
 function parseMaybeObject(value: unknown): Record<string, unknown> {
@@ -949,8 +1090,13 @@ function fallbackDetailText(args: unknown, result: unknown): string {
 }
 
 function cronScalar(value: unknown): string {
-  if (typeof value === 'string') return value.trim()
-  if (typeof value === 'number' && Number.isFinite(value)) return String(value)
+  if (typeof value === 'string') {
+    return value.trim()
+  }
+
+  if (typeof value === 'number' && Number.isFinite(value)) {
+    return String(value)
+  }
 
   return ''
 }
@@ -958,7 +1104,9 @@ function cronScalar(value: unknown): string {
 function formatCronTime(iso: string): string {
   const ts = Date.parse(iso)
 
-  if (Number.isNaN(ts)) return iso
+  if (Number.isNaN(ts)) {
+    return iso
+  }
 
   return new Date(ts).toLocaleString(undefined, {
     month: 'short',
@@ -968,10 +1116,7 @@ function formatCronTime(iso: string): string {
   })
 }
 
-function cronjobSubtitle(
-  argsRecord: Record<string, unknown>,
-  resultRecord: Record<string, unknown>
-): string {
+function cronjobSubtitle(argsRecord: Record<string, unknown>, resultRecord: Record<string, unknown>): string {
   const jobs = Array.isArray(resultRecord.jobs) ? resultRecord.jobs : null
 
   if (jobs) {
@@ -980,7 +1125,9 @@ function cronjobSubtitle(
 
   const message = firstStringField(resultRecord, ['message'])
 
-  if (message) return message
+  if (message) {
+    return message
+  }
 
   const action = firstStringField(argsRecord, ['action']) || 'manage'
   const name = firstStringField(resultRecord, ['name']) || firstStringField(argsRecord, ['name', 'job_id'])
@@ -989,14 +1136,13 @@ function cronjobSubtitle(
   return name ? `${label} ${name}` : `Cron ${action}`
 }
 
-function cronjobDetail(
-  argsRecord: Record<string, unknown>,
-  resultRecord: Record<string, unknown>
-): string {
+function cronjobDetail(argsRecord: Record<string, unknown>, resultRecord: Record<string, unknown>): string {
   const jobs = Array.isArray(resultRecord.jobs) ? resultRecord.jobs : null
 
   if (jobs) {
-    if (!jobs.length) return 'No cron jobs scheduled'
+    if (!jobs.length) {
+      return 'No cron jobs scheduled'
+    }
 
     return jobs
       .slice(0, 20)
@@ -1011,12 +1157,14 @@ function cronjobDetail(
   }
 
   const nextRun = cronScalar(resultRecord.next_run_at)
+
   const rows: [string, string][] = [
     ['Schedule', cronScalar(resultRecord.schedule)],
     ['Repeat', cronScalar(resultRecord.repeat)],
     ['Delivery', cronScalar(resultRecord.deliver)],
     ['Next run', nextRun ? formatCronTime(nextRun) : '']
   ]
+
   const lines = rows.filter(([, value]) => value).map(([key, value]) => `${key}: ${value}`)
 
   return lines.length ? lines.join('\n') : fallbackDetailText(argsRecord, resultRecord)
@@ -1090,9 +1238,9 @@ function toolSubtitle(
       }
     }
 
-    const command = firstStringField(argsRecord, ['command', 'code']) || contextValue(argsRecord)
+    const command = firstStringField(argsRecord, ['context', 'preview', 'command', 'code']) || contextValue(argsRecord)
 
-    return command ? compactPreview(command, 120) : 'Executed command'
+    return command ? '' : 'Executed command'
   }
 
   if (toolName === 'read_file' || isFileEditTool(toolName)) {
@@ -1200,7 +1348,7 @@ function toolDetailText(
     }
   }
 
-  if (part.toolName === 'read_file') {
+  if (part.toolName === 'read_file' && part.result !== undefined) {
     const content = firstStringField(resultRecord, ['content', 'text', 'data', 'body'])
 
     if (content) {
@@ -1259,6 +1407,7 @@ export function toolCopyPayload(part: ToolPart, view: ToolView): { label: string
     url: translateNow('assistant.tool.copyUrl'),
     generic: translateNow('common.copy')
   }
+
   const args = parseMaybeObject(part.args)
   const result = parseMaybeObject(part.result)
   const detail = view.detail.trim()
@@ -1310,7 +1459,7 @@ export function toolCopyPayload(part: ToolPart, view: ToolView): { label: string
     }
   }
 
-  if (part.toolName === 'read_file') {
+  if (part.toolName === 'read_file' && part.result !== undefined) {
     if (hasSubstantialOutput) {
       return { label: copy.file, text: detail }
     }
@@ -1341,39 +1490,104 @@ export function toolCopyPayload(part: ToolPart, view: ToolView): { label: string
   return { label: copy.generic, text: view.title }
 }
 
+interface ToolTitleParts {
+  action?: ToolTitleAction
+  title: string
+}
+
+function titlePartsFromAction(title: string, action?: string): ToolTitleParts {
+  if (!action) {
+    return { title }
+  }
+
+  const actionStart = title.indexOf(action)
+
+  if (actionStart < 0) {
+    return { title }
+  }
+
+  return {
+    action: {
+      prefix: title.slice(0, actionStart),
+      suffix: title.slice(actionStart + action.length),
+      text: action
+    },
+    title
+  }
+}
+
 function dynamicTitle(
   part: ToolPart,
   args: Record<string, unknown>,
   result: Record<string, unknown>,
-  fallback: string
-): string {
+  fallback: ToolTitleParts
+): ToolTitleParts {
   const verb = (gerund: string, past: string) => (part.result === undefined ? gerund : past)
 
+  const titledAction = (action: string, title: string): ToolTitleParts =>
+    titlePartsFromAction(title, part.result === undefined ? action : undefined)
+
   if (part.toolName === 'web_extract') {
     const url = findFirstUrl(args, result)
+    const action = verb(translateNow('assistant.tool.actions.reading'), translateNow('assistant.tool.actions.read'))
 
-    return url ? `${verb('Reading', 'Read')} ${hostnameOf(url)}` : fallback
+    return url
+      ? titledAction(action, translateNow('assistant.tool.titleTemplates.actionTarget', action, hostnameOf(url)))
+      : fallback
   }
 
   if (part.toolName === 'browser_navigate') {
     const url = findFirstUrl(args, result)
+    const action = verb(translateNow('assistant.tool.actions.opening'), translateNow('assistant.tool.actions.opened'))
 
-    return url ? `${verb('Opening', 'Opened')} ${hostnameOf(url)}` : fallback
+    return url
+      ? titledAction(action, translateNow('assistant.tool.titleTemplates.actionTarget', action, hostnameOf(url)))
+      : fallback
   }
 
   if (part.toolName === 'web_search') {
     const query = firstStringField(args, ['search_term', 'query']) || contextValue(args)
 
-    return query ? `${verb('Searching', 'Searched')} “${compactPreview(query, 48)}”` : fallback
+    const action = verb(
+      translateNow('assistant.tool.actions.searching'),
+      translateNow('assistant.tool.actions.searched')
+    )
+
+    return query
+      ? titledAction(
+          action,
+          translateNow('assistant.tool.titleTemplates.actionQuoted', action, compactPreview(query, 48))
+        )
+      : fallback
+  }
+
+  if (part.toolName === 'read_file') {
+    const target = readFileDisplayTarget(args, result)
+    const action = verb(translateNow('assistant.tool.actions.reading'), translateNow('assistant.tool.actions.read'))
+
+    return target
+      ? titledAction(action, translateNow('assistant.tool.titleTemplates.actionTarget', action, target))
+      : fallback
   }
 
   if (part.toolName === 'terminal' || part.toolName === 'execute_code') {
-    const command = firstStringField(args, ['command', 'code']) || contextValue(args)
+    const command =
+      firstStringField(args, ['context', 'preview']) || firstStringField(args, ['command', 'code']) || contextValue(args)
 
     if (command) {
-      const verbText = part.toolName === 'execute_code' ? verb('Running code', 'Ran code') : verb('Running', 'Ran')
+      const action =
+        part.toolName === 'execute_code'
+          ? verb(translateNow('assistant.tool.actions.runningCode'), translateNow('assistant.tool.actions.ranCode'))
+          : verb(translateNow('assistant.tool.actions.running'), translateNow('assistant.tool.actions.ran'))
 
-      return `${verbText} · ${compactPreview(command, 160)}`
+      return titledAction(
+        action,
+        translateNow(
+          'assistant.tool.titleTemplates.actionCommand',
+          action,
+          compactPreview(summarizeShellCommand(command), 160)
+        )
+      )
     }
   }
 
@@ -1381,7 +1595,7 @@ function dynamicTitle(
     const path = fileEditPath(args, result)
 
     if (path) {
-      return fileEditBasename(path)
+      return { title: fileEditBasename(path) }
     }
   }
 
@@ -1395,7 +1609,15 @@ export function buildToolView(part: ToolPart, inlineDiff: string): ToolView {
   const status = toolStatus(part, resultRecord)
   const error = toolErrorText(part, resultRecord)
   const baseTitle = part.result === undefined ? meta.pending : meta.done
-  const title = dynamicTitle(part, argsRecord, resultRecord, baseTitle)
+
+  const titleParts = dynamicTitle(
+    part,
+    argsRecord,
+    resultRecord,
+    titlePartsFromAction(baseTitle, part.result === undefined ? meta.pendingAction : undefined)
+  )
+
+  const title = titleParts.title
   const titleEnriched = title !== baseTitle
   const baseSubtitle = error || toolSubtitle(part, argsRecord, resultRecord)
 
@@ -1449,6 +1671,7 @@ export function buildToolView(part: ToolPart, inlineDiff: string): ToolView {
     status,
     subtitle,
     title,
+    titleAction: titleParts.action,
     tone: meta.tone
   }
 }

apps/desktop/src/components/assistant-ui/tool-fallback.tsx

@@ -33,6 +33,7 @@ import { $toolDisclosureOpen, $toolViewMode, setToolDisclosureOpen } from '@/sto
 import { PendingToolApproval } from './tool-approval'
 import {
   buildToolView,
+  clampForDisplay,
   cleanVisibleText,
   countDiffLineStats,
   inlineDiffFromResult,
@@ -45,7 +46,8 @@ import {
   toolCopyPayload,
   type ToolPart,
   toolPartDisclosureId,
-  type ToolStatus
+  type ToolStatus,
+  type ToolTitleAction
 } from './tool-fallback-model'
 
 // `true` when a ToolEntry is rendered inside an embedding wrapper that owns
@@ -104,7 +106,7 @@ function rawTechnicalTrace(args: unknown, result: unknown): string {
     })
     .filter(Boolean)
 
-  return parts.join('\n')
+  return clampForDisplay(parts.join('\n'))
 }
 
 function statusGlyph(status: ToolStatus, copy: ToolStatusCopy): ReactNode {
@@ -202,6 +204,39 @@ function LinkifiedText({ className, text }: { className?: string; text: string }
   return <SharedLinkifiedText className={className} pretty text={cleanVisibleText(text)} />
 }
 
+function ToolTitle({
+  isPending,
+  status,
+  title,
+  titleAction
+}: {
+  isPending: boolean
+  status: ToolStatus
+  title: string
+  titleAction?: ToolTitleAction
+}) {
+  return (
+    <FadeText
+      className={cn(
+        TOOL_HEADER_TITLE_CLASS,
+        isPending && 'text-(--ui-text-tertiary)',
+        status === 'error' && 'text-destructive',
+        status === 'warning' && 'text-amber-700 dark:text-amber-300'
+      )}
+    >
+      {isPending && titleAction ? (
+        <>
+          {titleAction.prefix}
+          <span className="shimmer">{titleAction.text}</span>
+          {titleAction.suffix}
+        </>
+      ) : (
+        title
+      )}
+    </FadeText>
+  )
+}
+
 interface ToolEntryProps {
   part: ToolPart
 }
@@ -220,13 +255,25 @@ function ToolEntry({ part }: ToolEntryProps) {
   const messageRunning = useAuiState(selectMessageRunning)
   const embedded = useContext(ToolEmbedContext)
   const toolViewMode = useStore($toolViewMode)
-  const disclosureId = `tool-entry:${messageId}:${toolPartDisclosureId(part)}`
+
+  // `ToolFallback` rebuilds the `part` wrapper each render, defeating the memos
+  // below and re-running buildToolView (full JSON.stringify of result) on every
+  // stream delta — the freeze on big `/learn` runs. Re-derive a stable part from
+  // the referentially-stable args/result so the memos hold across deltas.
+  const { args, isError, result, toolCallId, toolName } = part
+
+  const stablePart = useMemo<ToolPart>(
+    () => ({ args, isError, result, toolCallId, toolName, type: 'tool-call' }),
+    [args, isError, result, toolCallId, toolName]
+  )
+
+  const disclosureId = `tool-entry:${messageId}:${toolPartDisclosureId(stablePart)}`
   const dismissed = useStore($toolRowDismissed(disclosureId))
-  const isPending = messageRunning && part.result === undefined
+  const isPending = messageRunning && result === undefined
   const liveDiffs = useStore($toolInlineDiffs)
-  const sideDiff = part.toolCallId ? liveDiffs[part.toolCallId] || '' : ''
-  const inlineDiff = stripInlineDiffChrome(sideDiff) || inlineDiffFromResult(part.result)
-  const isFileEdit = isFileEditTool(part.toolName)
+  const sideDiff = toolCallId ? liveDiffs[toolCallId] || '' : ''
+  const inlineDiff = stripInlineDiffChrome(sideDiff) || inlineDiffFromResult(result)
+  const isFileEdit = isFileEditTool(toolName)
   const defaultOpen = Boolean(inlineDiff)
   const open = useDisclosureOpen(disclosureId, defaultOpen)
   const canDismiss = !isPending && !embedded
@@ -237,13 +284,14 @@ function ToolEntry({ part }: ToolEntryProps) {
   const enterRef = useEnterAnimation(messageRunning && !embedded, `tool-entry:${disclosureId}`)
   const elapsed = useElapsedSeconds(isPending, `tool:${disclosureId}`)
 
-  // Stale parts (no result, but message stopped running) get a synthetic
-  // empty result so buildToolView treats them as completed-no-output.
+  // Stale parts (no result, but message stopped running) get a synthetic empty
+  // result so buildToolView treats them as completed-no-output. Keyed on
+  // stablePart so it recomputes only when this tool's data changes.
   const view = useMemo(() => {
-    const p = !isPending && part.result === undefined ? { ...part, result: {} } : part
+    const p = !isPending && result === undefined ? { ...stablePart, result: {} } : stablePart
 
     return buildToolView(p, inlineDiff)
-  }, [inlineDiff, isPending, part])
+  }, [inlineDiff, isPending, result, stablePart])
 
   // Surface a previewable artifact (HTML file / localhost URL) as a compact link
   // in the composer status stack rather than a bulky inline card. Uses the same
@@ -313,7 +361,9 @@ function ToolEntry({ part }: ToolEntryProps) {
     view.imageUrl || view.inlineDiff || showDetail || hasSearchHits || toolViewMode === 'technical'
   )
 
-  const copyAction = useMemo(() => toolCopyPayload(part, view), [part, view])
+  // copyAction reads the uncapped view.detail; clampForDisplay below only bounds
+  // what's painted, so the row's Copy button still yields the full output.
+  const copyAction = useMemo(() => toolCopyPayload(stablePart, view), [stablePart, view])
 
   const diffStats = useMemo(
     () => (isFileEdit && view.inlineDiff ? countDiffLineStats(view.inlineDiff) : null),
@@ -398,16 +448,7 @@ function ToolEntry({ part }: ToolEntryProps) {
               icon={view.icon}
               status={leadingStatus(isPending, view.status)}
             />
-            <FadeText
-              className={cn(
-                TOOL_HEADER_TITLE_CLASS,
-                isPending && 'shimmer text-(--ui-text-tertiary)',
-                view.status === 'error' && 'text-destructive',
-                view.status === 'warning' && 'text-amber-700 dark:text-amber-300'
-              )}
-            >
-              {view.title}
-            </FadeText>
+            <ToolTitle isPending={isPending} status={view.status} title={view.title} titleAction={view.titleAction} />
             {!isPending && view.countLabel && <span className={TOOL_HEADER_DURATION_CLASS}>{view.countLabel}</span>}
             {showDiffStats && diffStats && (
               <span className="flex shrink-0 items-center gap-1 font-mono text-[0.625rem] tabular-nums">
@@ -466,7 +507,7 @@ function ToolEntry({ part }: ToolEntryProps) {
                         detailSections.summary && 'mt-1.5'
                       )}
                     >
-                      {detailSections.body}
+                      {clampForDisplay(detailSections.body)}
                     </pre>
                   )}
                 </div>
@@ -481,7 +522,7 @@ function ToolEntry({ part }: ToolEntryProps) {
                   <div className="space-y-0.5">
                     {view.stderr && <p className={TOOL_SECTION_LABEL_CLASS}>stdout</p>}
                     <pre className={cn(TOOL_SECTION_PRE_CLASS, 'whitespace-pre-wrap wrap-anywhere')}>
-                      {view.rendersAnsi ? <AnsiText text={view.stdout} /> : view.stdout}
+                      {view.rendersAnsi ? <AnsiText text={clampForDisplay(view.stdout)} /> : clampForDisplay(view.stdout)}
                     </pre>
                   </div>
                 )}
@@ -494,7 +535,7 @@ function ToolEntry({ part }: ToolEntryProps) {
                         'whitespace-pre-wrap wrap-anywhere text-(--ui-text-tertiary)'
                       )}
                     >
-                      {view.rendersAnsi ? <AnsiText text={view.stderr} /> : view.stderr}
+                      {view.rendersAnsi ? <AnsiText text={clampForDisplay(view.stderr)} /> : clampForDisplay(view.stderr)}
                     </pre>
                   </div>
                 )}
@@ -504,10 +545,10 @@ function ToolEntry({ part }: ToolEntryProps) {
                 {view.detailLabel && <p className={TOOL_SECTION_LABEL_CLASS}>{view.detailLabel}</p>}
                 {renderDetailAsCode ? (
                   <pre className={cn(TOOL_SECTION_PRE_CLASS, 'whitespace-pre-wrap wrap-anywhere')}>
-                    {view.rendersAnsi ? <AnsiText text={view.detail} /> : view.detail}
+                    {view.rendersAnsi ? <AnsiText text={clampForDisplay(view.detail)} /> : clampForDisplay(view.detail)}
                   </pre>
                 ) : (
-                  <CompactMarkdown className={cn(TOOL_SECTION_SURFACE_CLASS, 'wrap-anywhere')} text={view.detail} />
+                  <CompactMarkdown className={cn(TOOL_SECTION_SURFACE_CLASS, 'wrap-anywhere')} text={clampForDisplay(view.detail)} />
                 )}
               </div>
             ))}

apps/desktop/src/components/desktop-install-overlay.tsx

@@ -407,7 +407,11 @@ export function DesktopInstallOverlay({ enabled = true }: DesktopInstallOverlayP
 
   const totalCount = stages.length
   const failed = Boolean(state.error)
-  const progressPct = totalCount > 0 ? Math.round((completedCount / totalCount) * 100) : 0
+  // Count the running stage as half-done so the bar advances *during* a long
+  // stage instead of sitting frozen at the last completed step while its logs
+  // stream (e.g. "0 of 2" pinned at 0% for the whole first stage).
+  const progressUnits = completedCount + (!failed && currentStage ? 0.5 : 0)
+  const progressPct = totalCount > 0 ? Math.round((progressUnits / totalCount) * 100) : 0
   const currentStartedAt = currentStage ? state.stages[currentStage]?.startedAt : null
   const currentElapsed = typeof currentStartedAt === 'number' ? formatElapsed(now - currentStartedAt) : ''
 

apps/desktop/src/components/pet/floating-pet.tsx

@@ -0,0 +1,368 @@
+import { useStore } from '@nanostores/react'
+import { useCallback, useEffect, useRef, useState } from 'react'
+
+import { useGatewayRequest } from '@/app/gateway/hooks/use-gateway-request'
+import { persistString, storedString } from '@/lib/storage'
+import { $petInfo, clearPetUnread, type PetInfo, petProfile, setPetInfo } from '@/store/pet'
+import { resetPetGallery } from '@/store/pet-gallery'
+import { $petOverlayActive, initPetOverlayBridge, popOutPet, restorePetOverlay } from '@/store/pet-overlay'
+import { $activeGatewayProfile, normalizeProfileKey } from '@/store/profile'
+import { $gatewayState } from '@/store/session'
+import { isSecondaryWindow } from '@/store/windows'
+import { useTheme } from '@/themes/context'
+
+import { PetSprite } from './pet-sprite'
+
+// v2: positions are now top/left anchored (v1 stored bottom-anchored values,
+// which dragged inverted). Bumping the key discards stale v1 coordinates.
+const POSITION_KEY = 'hermes.desktop.pet-position.v2'
+
+interface Point {
+  x: number
+  y: number
+}
+
+interface PetInfoMeta {
+  enabled: boolean
+  slug?: string
+  displayName?: string
+  scale?: number
+  spritesheetRevision?: string
+}
+
+function samePetRevision(info: PetInfo, meta: PetInfoMeta): boolean {
+  return (
+    info.enabled &&
+    Boolean(info.spritesheetBase64) &&
+    info.slug === meta.slug &&
+    info.displayName === meta.displayName &&
+    info.scale === meta.scale &&
+    info.spritesheetRevision === meta.spritesheetRevision
+  )
+}
+
+function clampToViewport({ x, y }: Point): Point {
+  const maxX = Math.max(0, (window.innerWidth || 800) - 80)
+  const maxY = Math.max(0, (window.innerHeight || 600) - 80)
+
+  return { x: Math.min(Math.max(0, x), maxX), y: Math.min(Math.max(0, y), maxY) }
+}
+
+// The sprite art faces left by default, so mirror it when the pet's center sits
+// on the left half of the window — it always faces inward, toward the content.
+function facing(leftX: number, petW: number): string {
+  return leftX + petW / 2 < (window.innerWidth || 800) / 2 ? 'scaleX(-1)' : 'none'
+}
+
+function loadPosition(): Point {
+  try {
+    const raw = storedString(POSITION_KEY)
+
+    if (raw) {
+      const parsed = JSON.parse(raw) as Point
+
+      if (typeof parsed.x === 'number' && typeof parsed.y === 'number') {
+        return clampToViewport(parsed)
+      }
+    }
+  } catch {
+    // fall through to default
+  }
+
+  // Default: lower-left corner (top/left anchored).
+  return clampToViewport({ x: 24, y: (window.innerHeight || 600) - 220 })
+}
+
+/**
+ * In-window floating petdex mascot. Always-on-top within the app, draggable,
+ * and reactive to agent activity via `$petState`. Fetches the active pet via
+ * the shared `pet.info` RPC; renders nothing until a pet is installed +
+ * enabled.
+ *
+ * Adopting a pet is fully in-app: type `/pet boba` in the composer. That
+ * writes `display.pet.*` from the slash worker, so we keep polling `pet.info`
+ * while no pet is active and the mascot pops in within a few seconds — no
+ * reload, no CLI. Once a pet is live we still refresh more slowly so generated
+ * pets rewritten on disk (or renamed/rebuilt by the hatch flow) repaint without
+ * restarting the app.
+ *
+ * Promotion to a separate frameless OS-level window is a follow-up — the
+ * sprite + state logic here is reused as-is, only the host changes.
+ */
+const PET_POLL_MS = 3000
+const PET_ACTIVE_REFRESH_MS = 15000
+
+export function FloatingPet() {
+  const { requestGateway } = useGatewayRequest()
+  const { resolvedMode } = useTheme()
+  const gatewayState = useStore($gatewayState)
+  const info = useStore($petInfo)
+  const overlayActive = useStore($petOverlayActive)
+
+  const [position, setPosition] = useState<Point>(loadPosition)
+  const containerRef = useRef<HTMLDivElement | null>(null)
+  // The facing mirror lives on the sprite wrapper, not the container, so the
+  // speech bubble (a container child) never renders flipped/backwards.
+  const spriteWrapRef = useRef<HTMLDivElement | null>(null)
+  const petW = (info.frameW ?? 192) * (info.scale ?? 0.33)
+  // Soft contact shadow, sized off the pet so every scale/species grounds the
+  // same way (cf. lairp's per-actor feet ellipse). Lighter on light backgrounds.
+  const shadowW = Math.round(petW * 0.55)
+  const shadowH = Math.max(3, Math.round(shadowW * 0.28))
+  const shadowAlpha = resolvedMode === 'light' ? 0.2 : 0.55
+  // Live drag offset (pointer → element top-left). Drag updates the DOM
+  // directly to avoid a React re-render (and canvas reflow) per pointermove —
+  // state is only committed on release.
+  const dragRef = useRef<{ dx: number; dy: number; x: number; y: number } | null>(null)
+
+  // Fetch pet.info on connect. Poll quickly while inactive so an in-app
+  // `/pet <slug>` appears, then slowly while active so regenerated spritesheets
+  // and row-count metadata replace the cached base64 payload.
+  const active = info.enabled && Boolean(info.spritesheetBase64)
+  useEffect(() => {
+    if (gatewayState !== 'open') {
+      return
+    }
+
+    let cancelled = false
+
+    const pull = async () => {
+      try {
+        if (active) {
+          try {
+            const meta = await requestGateway<PetInfoMeta>('pet.info.meta', { profile: petProfile() })
+            if (cancelled || !meta) {
+              return
+            }
+            if (!meta.enabled) {
+              setPetInfo({ enabled: false })
+              return
+            }
+            if (samePetRevision($petInfo.get(), meta)) {
+              return
+            }
+          } catch {
+            // Older gateways may not have pet.info.meta yet; fall back to pet.info.
+          }
+        }
+
+        const next = await requestGateway<PetInfo>('pet.info', { profile: petProfile() })
+
+        if (!cancelled && next) {
+          const current = $petInfo.get()
+          if (
+            next.enabled &&
+            current.enabled &&
+            current.slug === next.slug &&
+            current.displayName === next.displayName &&
+            current.scale === next.scale &&
+            current.spritesheetRevision &&
+            current.spritesheetRevision === next.spritesheetRevision
+          ) {
+            return
+          }
+          setPetInfo(next)
+        }
+      } catch {
+        // cosmetic feature — never surface gateway errors
+      }
+    }
+
+    void pull()
+    const timer = window.setInterval(() => void pull(), active ? PET_ACTIVE_REFRESH_MS : PET_POLL_MS)
+    window.addEventListener('focus', pull)
+
+    return () => {
+      cancelled = true
+      window.removeEventListener('focus', pull)
+      window.clearInterval(timer)
+    }
+  }, [gatewayState, active, requestGateway])
+
+  // Pets are per-profile. When the active profile changes, drop the previous
+  // profile's mascot + gallery cache so the poll above refetches the new
+  // profile's pet (its config + pets dir resolve per-profile on the backend).
+  const profileRef = useRef(normalizeProfileKey($activeGatewayProfile.get()))
+  useEffect(
+    () =>
+      $activeGatewayProfile.subscribe(next => {
+        const key = normalizeProfileKey(next)
+
+        if (key === profileRef.current) {
+          return
+        }
+
+        profileRef.current = key
+        setPetInfo({ enabled: false })
+        resetPetGallery()
+      }),
+    []
+  )
+
+  // Wire the overlay control channel once, only in the primary window — the
+  // pop-out overlay belongs to it (main.cjs positions it against the main
+  // window and routes control messages back to it).
+  useEffect(() => {
+    if (isSecondaryWindow()) {
+      return
+    }
+
+    return initPetOverlayBridge()
+  }, [])
+
+  // Returning to the app (by any route, not just the mail icon) clears the pet's
+  // "new message" hint — you've seen it now.
+  useEffect(() => {
+    if (isSecondaryWindow()) {
+      return
+    }
+
+    const onFocus = () => clearPetUnread()
+    window.addEventListener('focus', onFocus)
+
+    return () => window.removeEventListener('focus', onFocus)
+  }, [])
+
+  // Restore a popped-out pet on boot, once the pet has loaded (so we never spawn
+  // an empty overlay window). Primary window only; runs at most once.
+  const restoredRef = useRef(false)
+  useEffect(() => {
+    if (isSecondaryWindow() || restoredRef.current || !active) {
+      return
+    }
+
+    restoredRef.current = true
+    restorePetOverlay()
+  }, [active])
+
+  // A window resize must never strand the pet off-screen — re-clamp the
+  // committed position (and persist it) whenever the viewport shrinks.
+  useEffect(() => {
+    const onResize = () =>
+      setPosition(prev => {
+        const next = clampToViewport(prev)
+
+        if (next.x === prev.x && next.y === prev.y) {
+          return prev
+        }
+
+        persistString(POSITION_KEY, JSON.stringify(next))
+
+        return next
+      })
+
+    window.addEventListener('resize', onResize)
+
+    return () => window.removeEventListener('resize', onResize)
+  }, [])
+
+  const onPointerDown = useCallback((e: React.PointerEvent) => {
+    const el = containerRef.current
+
+    if (!el) {
+      return
+    }
+
+    const rect = el.getBoundingClientRect()
+
+    // Shift-click pops the pet out into a free-floating desktop overlay (it can
+    // leave the window and stays visible while Hermes is minimized) instead of
+    // starting an in-window drag. Primary window only — the overlay is anchored
+    // to it.
+    if (e.shiftKey && !isSecondaryWindow()) {
+      popOutPet({ height: rect.height, width: rect.width, x: rect.left, y: rect.top })
+
+      return
+    }
+
+    dragRef.current = { dx: e.clientX - rect.left, dy: e.clientY - rect.top, x: rect.left, y: rect.top }
+    el.setPointerCapture(e.pointerId)
+    el.style.cursor = 'grabbing'
+  }, [])
+
+  const onPointerMove = useCallback(
+    (e: React.PointerEvent) => {
+      const drag = dragRef.current
+      const el = containerRef.current
+
+      if (!drag || !el) {
+        return
+      }
+
+      const next = clampToViewport({ x: e.clientX - drag.dx, y: e.clientY - drag.dy })
+      drag.x = next.x
+      drag.y = next.y
+      // Mutate the DOM directly — no setState, so no re-render while dragging. The
+      // mirror follows the pointer across the midline for the same reason; it
+      // rides the sprite wrapper so the bubble stays upright.
+      el.style.left = `${next.x}px`
+      el.style.top = `${next.y}px`
+
+      if (spriteWrapRef.current) {
+        spriteWrapRef.current.style.transform = facing(next.x, petW)
+      }
+    },
+    [petW]
+  )
+
+  const onPointerUp = useCallback((e: React.PointerEvent) => {
+    const drag = dragRef.current
+
+    if (drag) {
+      dragRef.current = null
+      const committed = { x: drag.x, y: drag.y }
+      setPosition(committed)
+      persistString(POSITION_KEY, JSON.stringify(committed))
+    }
+
+    const el = containerRef.current
+
+    if (el) {
+      el.style.cursor = 'grab'
+      el.releasePointerCapture?.(e.pointerId)
+    }
+  }, [])
+
+  // While popped out, the desktop overlay window owns the mascot — hide the
+  // in-window one so there aren't two.
+  if (!info.enabled || !info.spritesheetBase64 || overlayActive) {
+    return null
+  }
+
+  return (
+    <div
+      onPointerDown={onPointerDown}
+      onPointerMove={onPointerMove}
+      onPointerUp={onPointerUp}
+      ref={containerRef}
+      style={{
+        cursor: 'grab',
+        left: position.x,
+        pointerEvents: 'auto',
+        position: 'fixed',
+        top: position.y,
+        touchAction: 'none',
+        userSelect: 'none',
+        zIndex: 60
+      }}
+    >
+      <div
+        aria-hidden
+        style={{
+          background: `radial-gradient(ellipse at center, rgba(0,0,0,${shadowAlpha}) 0%, rgba(0,0,0,0) 70%)`,
+          bottom: -shadowH * 0.4,
+          height: shadowH,
+          left: '50%',
+          pointerEvents: 'none',
+          position: 'absolute',
+          transform: 'translateX(-50%)',
+          width: shadowW,
+          zIndex: 0
+        }}
+      />
+      <div ref={spriteWrapRef} style={{ lineHeight: 0, position: 'relative', transform: facing(position.x, petW), zIndex: 1 }}>
+        <PetSprite info={info} />
+      </div>
+    </div>
+  )
+}

apps/desktop/src/components/pet/pet-bubble.tsx

@@ -0,0 +1,142 @@
+import { useStore } from '@nanostores/react'
+import { useEffect, useState } from 'react'
+
+import { AlertCircle, Clock, type IconComponent } from '@/lib/icons'
+import { $petActivity, $petState, type PetState } from '@/store/pet'
+
+/**
+ * Speech bubble + status glyph for the popped-out pet overlay — the
+ * "notification" half of the mascot. It externalizes what the agent is doing
+ * (Codex-style) so a glance at the desktop pet replaces switching back to the
+ * window. The in-window pet doesn't show it (the app itself is the surface);
+ * only the overlay renders it.
+ *
+ * Text is derived purely from the same `$petState` / `$petActivity` the sprite
+ * already reacts to, so it never drifts from the animation. The bubble is shown
+ * only when there's something worth saying (working / reviewing / a transient
+ * done/error beat / waiting on the user) and is hidden at plain idle.
+ */
+
+type Tone = 'error' | 'wait'
+
+interface Spec {
+  lines: string[]
+  glyph?: IconComponent
+  tone?: Tone
+}
+
+// Phrasings per mood, picked at random (no immediate repeat) for a bit of life.
+// Keep them short — the bubble is tiny and never wraps.
+const SPECS: Partial<Record<PetState, Spec>> = {
+  run: {
+    lines: ['working…', 'on it…', 'crunching…', 'tinkering…', 'cooking…', 'in the weeds…', 'wiring it up…', 'making moves…', 'heads down…', 'hammering away…']
+  },
+  review: {
+    lines: ['thinking…', 'reading…', 'reviewing…', 'pondering…', 'connecting dots…', 'sizing it up…', 'tracing it…', 'mulling…', 'scheming…', 'hmm…']
+  },
+  failed: {
+    glyph: AlertCircle,
+    lines: ['hit a snag', 'welp', 'that broke', 'oof', 'snagged'],
+    tone: 'error'
+  },
+  waiting: {
+    glyph: Clock,
+    lines: ['your turn', 'all yours', 'over to you', 'ball’s in your court', 'awaiting orders'],
+    tone: 'wait'
+  }
+}
+
+const TONE_COLOR: Record<Tone, string> = {
+  error: 'var(--ui-red)',
+  wait: 'var(--ui-yellow)'
+}
+
+// Random pick that avoids repeating the line we're already showing.
+function pick(lines: string[], prev: string): string {
+  if (lines.length <= 1) {
+    return lines[0] ?? ''
+  }
+
+  let next = prev
+
+  while (next === prev) {
+    next = lines[Math.floor(Math.random() * lines.length)]
+  }
+
+  return next
+}
+
+export function PetBubble() {
+  const state = useStore($petState)
+  const activity = useStore($petActivity)
+  const [line, setLine] = useState('')
+
+  // Finish beats are carried by the sprite/mail icon; idle only speaks up when
+  // it's actually the user's turn. Everything else maps to a mood spec.
+  const specKey: null | PetState =
+    state in SPECS ? state : state === 'idle' && activity.awaitingInput ? 'waiting' : null
+  const rotating = specKey === 'run' || specKey === 'review'
+
+  // Pick a fresh line on every mood change, then keep rotating (random, no
+  // repeat) only while the agent is actively working/thinking.
+  useEffect(() => {
+    const spec = specKey ? SPECS[specKey] : null
+
+    if (!spec) {
+      setLine('')
+
+      return
+    }
+
+    setLine(prev => pick(spec.lines, prev))
+
+    if (!rotating || spec.lines.length <= 1) {
+      return
+    }
+
+    const id = window.setInterval(() => setLine(prev => pick(spec.lines, prev)), 2600)
+
+    return () => window.clearInterval(id)
+  }, [specKey, rotating])
+
+  const spec = specKey ? SPECS[specKey] : null
+
+  if (!spec) {
+    return null
+  }
+
+  const Glyph = spec.glyph
+  const text = line || spec.lines[0]
+  const hasText = Boolean(text)
+
+  return (
+    <div
+      style={{
+        alignItems: 'center',
+        // Solid, theme-driven surface (the prior --ui-bg-card mixes in
+        // `transparent`, so the bubble was see-through).
+        background: 'var(--ui-bg-elevated)',
+        border: '1px solid var(--ui-stroke-secondary)',
+        borderRadius: hasText ? 10 : 999,
+        boxShadow: '0 4px 14px rgba(0,0,0,0.22)',
+        color: 'var(--foreground)',
+        display: 'inline-flex',
+        fontSize: 11,
+        fontWeight: 500,
+        gap: hasText ? 5 : 0,
+        lineHeight: 1,
+        // Glyph-only bubbles collapse to a tight, symmetric badge.
+        padding: hasText ? '5px 8px' : 5,
+        pointerEvents: 'none',
+        whiteSpace: 'nowrap'
+      }}
+    >
+      {Glyph && (
+        <span style={{ display: 'inline-flex' }}>
+          <Glyph style={{ color: spec.tone ? TONE_COLOR[spec.tone] : 'currentColor', height: 13, width: 13 }} />
+        </span>
+      )}
+      {text}
+    </div>
+  )
+}

apps/desktop/src/components/pet/pet-egg-hatch.tsx

@@ -0,0 +1,68 @@
+/**
+ * Egg-hatch visuals for the pet generation flow (Cmd-K → Pets → Generate).
+ *
+ * `PetEggHatch` is the incubation beat shown while `pet.hatch` runs: a wobbling
+ * egg that reads as "something is about to hatch" instead of a bare spinner. The
+ * reveal celebration is the canvas `PetStarShower`. Motion is disabled under
+ * `prefers-reduced-motion`.
+ */
+
+import { PixelEggSprite } from '@/components/pet/pixel-egg-sprite'
+import { Button } from '@/components/ui/button'
+
+interface PetEggHatchProps {
+  subtitle?: string
+  onCancel?: () => void
+  cancelLabel?: string
+}
+
+/**
+ * Thin progress bar. Determinate when given done/total (hatch rows stream one by
+ * one, so a real percentage is meaningful); indeterminate otherwise (drafts
+ * return together, so a count would just snap 0→100).
+ */
+export function PetProgress({ done, total }: { done?: number; total?: number }) {
+  const determinate = typeof done === 'number' && typeof total === 'number' && total > 0
+  const pct = determinate ? Math.min(100, Math.round((done / total) * 100)) : 0
+
+  return (
+    <div
+      aria-valuemax={100}
+      aria-valuemin={0}
+      aria-valuenow={determinate ? pct : undefined}
+      className="pet-progress"
+      role="progressbar"
+    >
+      {determinate ? (
+        <div className="pet-progress__fill" style={{ width: `${pct}%` }} />
+      ) : (
+        <div className="pet-progress__indeterminate" />
+      )}
+    </div>
+  )
+}
+
+export function PetEggHatch({ subtitle, onCancel, cancelLabel }: PetEggHatchProps) {
+  return (
+    <div className="flex flex-col items-center justify-center gap-3">
+      <div className="flex flex-col items-center">
+        <PixelEggSprite mode="bounce" size={88} />
+        {/* The egg sprite has transparent canvas below the art, so pull the
+            shadow up ~a fifth of its size to sit at the egg's base. */}
+        <span className="pet-egg-shadow" style={{ marginTop: '-0.55rem' }} />
+      </div>
+
+      {subtitle && (
+        <p className="shimmer shimmer-color-primary whitespace-nowrap text-center text-[length:var(--conversation-caption-font-size)] leading-snug text-(--ui-text-tertiary)">
+          {subtitle}
+        </p>
+      )}
+
+      {onCancel && (
+        <Button onClick={onCancel} size="xs" variant="text">
+          {cancelLabel ?? 'Cancel'}
+        </Button>
+      )}
+    </div>
+  )
+}

apps/desktop/src/components/pet/pet-sprite.tsx

@@ -0,0 +1,236 @@
+import { memo, useEffect, useMemo, useRef } from 'react'
+
+import { $petState, type PetInfo, type PetState } from '@/store/pet'
+
+const DEFAULT_FRAME_W = 192
+const DEFAULT_FRAME_H = 208
+const DEFAULT_FRAMES = 6
+const DEFAULT_LOOP_MS = 1100
+// Mirrors agent.pet.constants.DEFAULT_SCALE — fallback only; the gateway sends
+// the configured scale.
+const DEFAULT_SCALE = 0.33
+// Mirrors agent.pet.constants.CODEX_STATE_ROWS (Petdex current taxonomy).
+const DEFAULT_STATE_ROWS = [
+  'idle',
+  'running-right',
+  'running-left',
+  'waving',
+  'jumping',
+  'failed',
+  'waiting',
+  'running',
+  'review'
+]
+
+const STATE_ALIASES: Record<PetState, string[]> = {
+  idle: ['idle'],
+  wave: ['wave', 'waving'],
+  jump: ['jump', 'jumping'],
+  run: ['run', 'running'],
+  failed: ['failed'],
+  review: ['review'],
+  waiting: ['waiting']
+}
+
+const ROW_TO_STATE: Record<string, PetState> = {
+  idle: 'idle',
+  wave: 'wave',
+  waving: 'wave',
+  jump: 'jump',
+  jumping: 'jump',
+  run: 'run',
+  running: 'run',
+  'running-right': 'run',
+  'running-left': 'run',
+  failed: 'failed',
+  review: 'review',
+  waiting: 'waiting'
+}
+
+interface PetSpriteProps {
+  info: PetInfo
+  /** On-screen scale multiplier applied on top of the pet's native scale. */
+  zoom?: number
+  /**
+   * Force a specific animation state instead of reading the live `$petState`.
+   * Used by the generate-flow preview to showcase every row without driving (or
+   * being driven by) the real agent activity that moves the floating mascot.
+   */
+  stateOverride?: PetState
+  /** Force a concrete row name from `info.stateRows` (e.g. `running-right`). */
+  rowOverride?: string
+}
+
+/**
+ * Canvas renderer for a petdex spritesheet — the one piece that must be
+ * TypeScript (the engine's decode/encode is Python). Draws the row matching the
+ * live `$petState`, stepping `framesPerState` frames across a `loopMs` loop.
+ *
+ * State is read from `$petState` via a ref + subscription rather than a prop,
+ * so the frequent activity-driven state changes during an agent turn update the
+ * canvas (inside its RAF loop) WITHOUT triggering a React re-render. Combined
+ * with `memo`, this component effectively never re-renders after mount until
+ * the pet itself changes.
+ */
+function PetSpriteImpl({ info, zoom = 1, stateOverride, rowOverride }: PetSpriteProps) {
+  const canvasRef = useRef<HTMLCanvasElement | null>(null)
+  const stateRef = useRef<PetState>($petState.get())
+  const overrideRef = useRef<PetState | undefined>(stateOverride)
+  const rowOverrideRef = useRef<string | undefined>(rowOverride)
+
+  // Keep the override current without re-running the RAF setup effect.
+  useEffect(() => {
+    overrideRef.current = stateOverride
+  }, [stateOverride])
+
+  useEffect(() => {
+    rowOverrideRef.current = rowOverride
+  }, [rowOverride])
+
+  const frameW = info.frameW ?? DEFAULT_FRAME_W
+  const frameH = info.frameH ?? DEFAULT_FRAME_H
+  const frames = info.framesPerState ?? DEFAULT_FRAMES
+  const framesByState = info.framesByState
+  const framesByRow = info.framesByRow
+  const loopMs = info.loopMs ?? DEFAULT_LOOP_MS
+  const scale = (info.scale ?? DEFAULT_SCALE) * zoom
+  const rows = info.stateRows ?? DEFAULT_STATE_ROWS
+
+  const drawW = Math.round(frameW * scale)
+  const drawH = Math.round(frameH * scale)
+
+  const image = useMemo(() => {
+    if (!info.spritesheetBase64) {
+      return null
+    }
+
+    const img = new Image()
+    img.src = `data:${info.mime ?? 'image/webp'};base64,${info.spritesheetBase64}`
+
+    return img
+  }, [info.spritesheetBase64, info.mime])
+
+  useEffect(() => {
+    const canvas = canvasRef.current
+
+    if (!canvas || !image) {
+      return
+    }
+
+    const ctx = canvas.getContext('2d')
+
+    if (!ctx) {
+      return
+    }
+
+    // Track state via subscription, not a prop — no re-render on activity ticks.
+    stateRef.current = $petState.get()
+
+    const unsubState = $petState.listen(next => {
+      stateRef.current = next
+    })
+
+    let raf = 0
+    let frame = 0
+    let lastStep = performance.now()
+    let drawnFrame = -1
+    let drawnRow = -1
+    let activeRow = -1
+    let activeCount = -1
+
+    const rowIndexForState = (s: PetState): number => {
+      for (const key of STATE_ALIASES[s] ?? [s]) {
+        const idx = rows.indexOf(key)
+        if (idx >= 0) {
+          return idx
+        }
+      }
+      return 0
+    }
+
+    // Resolve a state to the row it draws and its real frame count. A state
+    // with no real frames (ragged sheet, empty row) falls back to idle rather
+    // than flashing blank padding.
+    const resolve = (s: PetState): { row: number; count: number } => {
+      const real = framesByState?.[s] ?? frames
+
+      if (real > 0) {
+        return { row: rowIndexForState(s), count: real }
+      }
+
+      return { row: rowIndexForState('idle'), count: Math.max(1, framesByState?.idle ?? frames) }
+    }
+
+    const resolveRow = (rowName: string): { row: number; count: number } => {
+      const row = rows.indexOf(rowName)
+      const state = ROW_TO_STATE[rowName]
+      const count = Math.max(
+        1,
+        framesByRow?.[rowName] ?? framesByState?.[rowName] ?? (state ? framesByState?.[state] : 0) ?? frames
+      )
+      return { row: row >= 0 ? row : rowIndexForState(state ?? 'idle'), count }
+    }
+
+    const render = (now: number) => {
+      const forcedRow = rowOverrideRef.current
+      const { row, count } = forcedRow ? resolveRow(forcedRow) : resolve(overrideRef.current ?? stateRef.current)
+
+      if (row !== activeRow || count !== activeCount) {
+        activeRow = row
+        activeCount = count
+        frame = 0
+        lastStep = now
+        drawnFrame = -1
+      }
+
+      // Per-state step keeps every state's loop ~loopMs even when frame counts
+      // differ; counts vary per row so derive the cadence here, not once.
+      const stepMs = loopMs / count
+
+      if (now - lastStep >= stepMs) {
+        frame += 1
+        lastStep = now
+      }
+
+      frame %= count
+
+      // Only touch the canvas when the visible cell actually changes. The RAF
+      // ticks at ~60Hz but the sprite only steps ~5Hz, so this skips ~90% of
+      // the clear+draw work and keeps the main thread free.
+      if ((frame !== drawnFrame || row !== drawnRow) && image.complete && image.naturalWidth > 0) {
+        const sx = frame * frameW
+        const sy = row * frameH
+        ctx.clearRect(0, 0, canvas.width, canvas.height)
+        ctx.imageSmoothingEnabled = false
+        ctx.drawImage(image, sx, sy, frameW, frameH, 0, 0, drawW, drawH)
+        drawnFrame = frame
+        drawnRow = row
+      }
+
+      raf = requestAnimationFrame(render)
+    }
+
+    raf = requestAnimationFrame(render)
+
+    return () => {
+      cancelAnimationFrame(raf)
+      unsubState()
+    }
+  }, [image, frameW, frameH, frames, framesByState, framesByRow, loopMs, drawW, drawH, rows])
+
+  return (
+    <canvas
+      aria-label={info.displayName ? `${info.displayName} pet` : 'pet'}
+      height={drawH}
+      ref={canvasRef}
+      style={{ height: drawH, width: drawW }}
+      width={drawW}
+    />
+  )
+}
+
+/**
+ * Memoized so a parent re-render (e.g. a position commit on drag-end) doesn't
+ * re-run the canvas setup. Props change only when the pet itself changes.
+ */
+export const PetSprite = memo(PetSpriteImpl)

apps/desktop/src/components/pet/pet-star-shower.tsx

@@ -0,0 +1,204 @@
+import { useEffect, useRef } from 'react'
+
+/**
+ * Canvas hatch celebration layered over a freshly revealed pet: a one-shot
+ * sunburst of rotating god-rays, a fast radial star burst (confetti physics —
+ * velocity + decay + gravity + spin), and a light trickle of rising twinkle
+ * motes. Additive (`lighter`) so the sparkles bloom. No glow-halo flash.
+ *
+ * Sized to its container (absolute inset-0, pointer-events: none) and disabled
+ * under `prefers-reduced-motion`.
+ */
+
+const GOLD = '#ffd76a'
+const BURST = 15
+const VELOCITY = 500
+const DECAY = 0.9
+const GRAVITY = 90
+const RAY_COUNT = 24
+const GOLD_MIX = 0.6
+const MOTE_MS = 333 // ~3 / sec
+
+interface Star {
+  x: number
+  y: number
+  vx: number
+  vy: number
+  size: number
+  rot: number
+  vrot: number
+  phase: number
+  twinkle: number
+  life: number
+  ttl: number
+  color: string
+  rise: boolean
+}
+
+function readAccent(el: HTMLElement): string {
+  return getComputedStyle(el).getPropertyValue('--ui-accent').trim() || '#9aa0ff'
+}
+
+function sparkle(ctx: CanvasRenderingContext2D, size: number, rot: number, color: string): void {
+  ctx.rotate(rot)
+  ctx.fillStyle = color
+  for (const [rx, ry] of [
+    [size, size * 0.26],
+    [size * 0.26, size]
+  ]) {
+    ctx.beginPath()
+    ctx.moveTo(0, -ry)
+    ctx.lineTo(rx, 0)
+    ctx.lineTo(0, ry)
+    ctx.lineTo(-rx, 0)
+    ctx.closePath()
+    ctx.fill()
+  }
+  const core = Math.max(1, Math.round(size * 0.4))
+  ctx.fillStyle = '#fff'
+  ctx.fillRect(-core / 2, -core / 2, core, core)
+}
+
+export function PetStarShower() {
+  const canvasRef = useRef<HTMLCanvasElement | null>(null)
+
+  useEffect(() => {
+    const canvas = canvasRef.current
+    const ctx = canvas?.getContext('2d')
+    const parent = canvas?.parentElement
+    if (!canvas || !ctx || !parent) {
+      return
+    }
+    if (window.matchMedia?.('(prefers-reduced-motion: reduce)').matches) {
+      return
+    }
+
+    const accent = readAccent(canvas)
+    const dpr = Math.min(window.devicePixelRatio || 1, 3)
+    let w = 0
+    let h = 0
+    let cx = 0
+    let cy = 0
+    const resize = () => {
+      const r = parent.getBoundingClientRect()
+      w = r.width
+      h = r.height
+      cx = w / 2
+      cy = h * 0.54
+      canvas.width = Math.round(w * dpr)
+      canvas.height = Math.round(h * dpr)
+      ctx.setTransform(dpr, 0, 0, dpr, 0, 0)
+    }
+    resize()
+    const ro = new ResizeObserver(resize)
+    ro.observe(parent)
+
+    const pick = () => (Math.random() < GOLD_MIX ? GOLD : Math.random() < 0.5 ? accent : '#ffffff')
+    const stars: Star[] = []
+    for (let i = 0; i < BURST; i++) {
+      const a = Math.random() * Math.PI * 2
+      const sp = VELOCITY * (0.4 + Math.random() * 0.7)
+      stars.push({
+        x: cx, y: cy, vx: Math.cos(a) * sp, vy: Math.sin(a) * sp,
+        size: 3.5 + Math.random() * 5.5, rot: Math.random() * 6.28, vrot: (Math.random() - 0.5) * 8,
+        phase: 0, twinkle: 0, life: 0, ttl: 0.8 + Math.random() * 0.7, color: pick(), rise: false
+      })
+    }
+    const rays = { life: 0, ttl: 0.9, rot: Math.random() * 6.28 }
+
+    let raf = 0
+    let last = performance.now()
+    let acc = 0
+    let raysAlive = true
+
+    const tick = (now: number) => {
+      raf = requestAnimationFrame(tick)
+      const ms = now - last
+      last = now
+      const dt = Math.min(0.05, ms / 1000)
+      const decay = Math.pow(DECAY, dt * 60)
+      acc += ms
+      if (acc >= MOTE_MS && stars.length < 40) {
+        acc = 0
+        stars.push({
+          x: cx + (Math.random() - 0.5) * w * 0.85, y: cy + Math.random() * h * 0.25,
+          vx: (Math.random() - 0.5) * 14, vy: -(14 + Math.random() * 26),
+          size: 2.5 + Math.random() * 3.5, rot: Math.random() * 6.28, vrot: (Math.random() - 0.5) * 2,
+          phase: Math.random() * 6.28, twinkle: 5 + Math.random() * 4, life: 0, ttl: 1.2 + Math.random(),
+          color: pick(), rise: true
+        })
+      }
+
+      ctx.clearRect(0, 0, w, h)
+      ctx.globalCompositeOperation = 'lighter'
+
+      // Sunburst god-rays — one-shot bloom + slow spin.
+      if (raysAlive) {
+        rays.life += dt
+        rays.rot += dt * 0.6
+        const t = rays.life / rays.ttl
+        if (t >= 1) {
+          raysAlive = false
+        } else {
+          const len = Math.max(w, h) * 0.62 * (1 - (1 - t) ** 2)
+          ctx.save()
+          ctx.translate(cx, cy)
+          ctx.rotate(rays.rot)
+          for (let i = 0; i < RAY_COUNT; i++) {
+            ctx.rotate((Math.PI * 2) / RAY_COUNT)
+            const a = (1 - t) * 0.3 * (i % 2 ? 0.65 : 1)
+            const wd = len * 0.05
+            const g = ctx.createLinearGradient(0, 0, 0, -len)
+            g.addColorStop(0, `rgba(255,255,255,${a})`)
+            g.addColorStop(1, 'rgba(255,255,255,0)')
+            ctx.fillStyle = g
+            ctx.beginPath()
+            ctx.moveTo(-wd, 0)
+            ctx.lineTo(wd, 0)
+            ctx.lineTo(0, -len)
+            ctx.closePath()
+            ctx.fill()
+          }
+          ctx.restore()
+        }
+      }
+
+      for (let i = stars.length - 1; i >= 0; i--) {
+        const s = stars[i]
+        s.life += dt
+        if (s.rise) {
+          s.vy += 7 * dt
+          s.phase += s.twinkle * dt
+        } else {
+          s.vx *= decay
+          s.vy = s.vy * decay + GRAVITY * dt
+        }
+        s.x += s.vx * dt
+        s.y += s.vy * dt
+        s.rot += s.vrot * dt
+        if (s.life >= s.ttl || s.y < -12) {
+          stars.splice(i, 1)
+          continue
+        }
+        const fade = s.rise
+          ? Math.min(1, s.life * 5, (s.ttl - s.life) * 3) * (0.45 + 0.55 * Math.abs(Math.sin(s.phase)))
+          : Math.min(1, (s.ttl - s.life) * 3)
+        ctx.save()
+        ctx.globalAlpha = fade
+        ctx.translate(Math.round(s.x), Math.round(s.y))
+        sparkle(ctx, s.size, s.rot, s.color)
+        ctx.restore()
+      }
+
+      ctx.globalCompositeOperation = 'source-over'
+    }
+    raf = requestAnimationFrame(tick)
+
+    return () => {
+      cancelAnimationFrame(raf)
+      ro.disconnect()
+    }
+  }, [])
+
+  return <canvas className="pointer-events-none absolute inset-0 z-10 h-full w-full" ref={canvasRef} />
+}

apps/desktop/src/components/pet/pet-thumb.tsx

@@ -0,0 +1,79 @@
+import { useEffect, useRef, useState } from 'react'
+
+import { PawPrint } from '@/lib/icons'
+
+// petdex frames are a fixed 192×208 grid; the box matches that aspect.
+const THUMB_W = 40
+const THUMB_H = Math.round((THUMB_W * 208) / 192)
+
+export type PetThumbLoader = (slug: string, url?: string) => Promise<string | null>
+
+/**
+ * Idle-frame preview for one pet. The backend crops + caches the frame and
+ * returns it as a same-origin data URI (`pet.thumb`), which dodges the renderer
+ * CSP / R2 hotlink rules that break a direct `<img src=cdn>`.
+ */
+export function PetThumb({
+  slug,
+  url,
+  alt,
+  load,
+  size = THUMB_W
+}: {
+  slug: string
+  url?: string
+  alt: string
+  load: PetThumbLoader
+  /** Width in px; height follows the petdex frame aspect. */
+  size?: number
+}) {
+  const [src, setSrc] = useState<string | null>(null)
+  const boxRef = useRef<HTMLSpanElement | null>(null)
+  const height = Math.round((size * 208) / 192)
+
+  useEffect(() => {
+    const el = boxRef.current
+
+    if (!el || src) {
+      return
+    }
+
+    const observer = new IntersectionObserver(
+      entries => {
+        if (entries.some(entry => entry.isIntersecting)) {
+          observer.disconnect()
+          void load(slug, url).then(uri => {
+            if (uri) {
+              setSrc(uri)
+            }
+          })
+        }
+      },
+      { rootMargin: '120px' }
+    )
+
+    observer.observe(el)
+
+    return () => observer.disconnect()
+  }, [slug, url, src, load])
+
+  return (
+    <span
+      className="grid shrink-0 place-items-center overflow-hidden rounded-md bg-(--ui-bg-tertiary) text-(--ui-text-tertiary)"
+      ref={boxRef}
+      style={{ height, width: size }}
+    >
+      {src ? (
+        <img
+          alt={alt}
+          aria-hidden
+          className="pointer-events-none size-full object-contain"
+          src={src}
+          style={{ imageRendering: 'pixelated' }}
+        />
+      ) : (
+        <PawPrint className="size-4" />
+      )}
+    </span>
+  )
+}

apps/desktop/src/components/pet/pixel-egg-sprite.tsx

@@ -0,0 +1,234 @@
+import { type CSSProperties, useEffect, useRef } from 'react'
+
+import eggSheetUrl from './pet-egg-sheet.png'
+
+/**
+ * Animated pixel egg — the iamcrog "bouncing hatching egg" 12-frame sheet
+ * (32×32 cells, stacked vertically), drawn to a canvas and recolored to a warm
+ * white/creme shell.
+ *
+ * The sheet's shell is mid-gray, so a plain multiply only darkens it (still
+ * gray). Instead we remap each pixel's luminance through a creme ramp via a 256-
+ * entry LUT: near-black stays a warm dark outline, midtones become creme shadow,
+ * highlights go near-white. Done on a 32×32 offscreen then nearest-neighbor
+ * scaled up so it stays crisp.
+ *
+ * Frames 0–5 are the intact squash/stretch bounce; 6–11 are the crack/hatch.
+ * `mode="bounce"` loops 0–5 (never shows a crack); `mode="hatch"` plays 6–11
+ * once then calls onDone.
+ */
+
+const FRAME = 32
+const TOTAL_FRAMES = 12
+const BOUNCE_FRAMES = 6 // 0..5 — intact egg only; cracks start at frame 6
+const HATCH_START = 6 // first crack frame
+// Per-frame speed *while* a bounce is playing.
+const BOUNCE_MS = 250
+const HATCH_MS = 190
+// Harvest-Moon idle: the egg rests on frame 0 for a long, randomized gap between
+// bounces so it reads as "occasionally stirs", not "constantly animating".
+const REST_MIN_MS = 2600
+const REST_MAX_MS = 6200
+
+// Creme ramp endpoints: warm dark outline → creme shadow → near-white highlight.
+const OUTLINE: [number, number, number] = [78, 66, 58]
+const SHADOW: [number, number, number] = [214, 198, 168]
+const HIGHLIGHT: [number, number, number] = [253, 249, 238]
+const OUTLINE_CUTOFF = 46
+
+const lerp = (a: number, b: number, t: number) => a + (b - a) * t
+
+// Precompute the luminance→creme mapping once (shared across every egg). Below
+// the cutoff it's the flat outline; above, a SHADOW→HIGHLIGHT ramp.
+const CREME_LUT = (() => {
+  const lut = new Uint8ClampedArray(256 * 3)
+  for (let g = 0; g < 256; g++) {
+    const dark = g < OUTLINE_CUTOFF
+    const t = dark ? 0 : (g - OUTLINE_CUTOFF) / (255 - OUTLINE_CUTOFF)
+    const from = dark ? OUTLINE : SHADOW
+    const to = dark ? OUTLINE : HIGHLIGHT
+    lut.set([lerp(from[0], to[0], t), lerp(from[1], to[1], t), lerp(from[2], to[2], t)], g * 3)
+  }
+  return lut
+})()
+
+let _sheet: HTMLImageElement | null = null
+let _sheetLoading: Promise<HTMLImageElement> | null = null
+
+function loadSheet(): Promise<HTMLImageElement> {
+  if (_sheet?.complete) {
+    return Promise.resolve(_sheet)
+  }
+  if (!_sheetLoading) {
+    _sheetLoading = new Promise((resolve, reject) => {
+      const img = new Image()
+      img.onload = () => {
+        _sheet = img
+        resolve(img)
+      }
+      img.onerror = reject
+      img.src = eggSheetUrl
+    })
+  }
+  return _sheetLoading
+}
+
+interface PixelEggSpriteProps {
+  mode: 'bounce' | 'hatch'
+  /** On-screen size (px, square). */
+  size: number
+  /**
+   * Slot position in a grid of eggs. Used to deterministically spread each egg's
+   * first bounce across the rest window so neighbours never stir together (random
+   * jitter alone can collide with only a handful of eggs).
+   */
+  index?: number
+  className?: string
+  style?: CSSProperties
+  /** Fired once when a `hatch` run reaches the final frame. */
+  onDone?: () => void
+}
+
+export function PixelEggSprite({ mode, size, index = 0, className, style, onDone }: PixelEggSpriteProps) {
+  const canvasRef = useRef<HTMLCanvasElement | null>(null)
+  const onDoneRef = useRef(onDone)
+  onDoneRef.current = onDone
+
+  useEffect(() => {
+    const canvas = canvasRef.current
+    const ctx = canvas?.getContext('2d')
+    if (!canvas || !ctx) {
+      return
+    }
+
+    const dpr = Math.min(window.devicePixelRatio || 1, 3)
+    const dim = Math.round(size * dpr)
+    canvas.width = dim
+    canvas.height = dim
+
+    const lastFrame = TOTAL_FRAMES - 1
+    // Mild per-egg speed jitter so bounces don't feel mechanical.
+    const frameMs = (mode === 'bounce' ? BOUNCE_MS : HATCH_MS) * (0.85 + Math.random() * 0.3)
+    const restMs = () => REST_MIN_MS + Math.random() * (REST_MAX_MS - REST_MIN_MS)
+    // First bounce: a deterministic per-slot slice of the rest window (so two
+    // eggs never start together) plus a little random jitter on top.
+    const firstDelay = ((index % 4) + 1) * (REST_MIN_MS / 4) + Math.random() * REST_MIN_MS
+
+    // 32×32 offscreen we recolor per frame, then scale up nearest-neighbor.
+    const off = document.createElement('canvas')
+    off.width = FRAME
+    off.height = FRAME
+    const offCtx = off.getContext('2d', { willReadFrequently: true })
+
+    let sheet: HTMLImageElement | null = null
+    void loadSheet().then(img => {
+      sheet = img
+    })
+
+    const render = (frame: number) => {
+      if (!sheet || !offCtx) {
+        return
+      }
+      offCtx.clearRect(0, 0, FRAME, FRAME)
+      offCtx.imageSmoothingEnabled = false
+      offCtx.drawImage(sheet, 0, frame * FRAME, FRAME, FRAME, 0, 0, FRAME, FRAME)
+      const img = offCtx.getImageData(0, 0, FRAME, FRAME)
+      const d = img.data
+      for (let i = 0; i < d.length; i += 4) {
+        if (d[i + 3] === 0) {
+          continue
+        }
+        const g = d[i] * 3
+        d[i] = CREME_LUT[g]
+        d[i + 1] = CREME_LUT[g + 1]
+        d[i + 2] = CREME_LUT[g + 2]
+      }
+      offCtx.putImageData(img, 0, 0)
+
+      ctx.clearRect(0, 0, dim, dim)
+      ctx.imageSmoothingEnabled = false
+      ctx.drawImage(off, 0, 0, FRAME, FRAME, 0, 0, dim, dim)
+    }
+
+    let raf = 0
+    let step = 0
+    let finished = false
+    // bounce: `nextAt` is when the next thing happens — the next bounce frame, or
+    // the start of a new bounce after a rest. hatch: `lastHatch` time-gates frames.
+    let resting = mode === 'bounce'
+    let nextAt = 0
+    let lastHatch = 0
+
+    const tick = (now: number) => {
+      raf = requestAnimationFrame(tick)
+      if (!sheet) {
+        return
+      }
+
+      if (mode === 'hatch') {
+        if (!lastHatch) {
+          lastHatch = now
+          render(HATCH_START)
+          return
+        }
+        if (now - lastHatch < frameMs) {
+          return
+        }
+        lastHatch = now
+        const frame = Math.min(HATCH_START + step, lastFrame)
+        render(frame)
+        if (frame >= lastFrame) {
+          if (!finished) {
+            finished = true
+            onDoneRef.current?.()
+          }
+          return // hold the cracked-open last frame
+        }
+        step += 1
+        return
+      }
+
+      // bounce: rest on frame 0, play 0..5, then rest again.
+      if (!nextAt) {
+        render(0)
+        nextAt = now + firstDelay // staggered first bounce, per slot
+        return
+      }
+      if (now < nextAt) {
+        return
+      }
+
+      if (resting) {
+        resting = false
+        step = 0
+        render(0)
+        nextAt = now + frameMs
+        return
+      }
+
+      step += 1
+      if (step >= BOUNCE_FRAMES) {
+        resting = true
+        render(0)
+        nextAt = now + restMs()
+        return
+      }
+      render(step)
+      nextAt = now + frameMs
+    }
+
+    raf = requestAnimationFrame(tick)
+
+    return () => {
+      cancelAnimationFrame(raf)
+    }
+  }, [mode, size, index])
+
+  return (
+    <canvas
+      className={className}
+      ref={canvasRef}
+      style={{ width: size, height: size, imageRendering: 'pixelated', ...style }}
+    />
+  )
+}

apps/desktop/src/components/ui/command.tsx

@@ -17,7 +17,12 @@ function Command({ className, ...props }: React.ComponentProps<typeof CommandPri
   )
 }
 
-function CommandInput({ className, ...props }: React.ComponentProps<typeof CommandPrimitive.Input>) {
+interface CommandInputProps extends React.ComponentProps<typeof CommandPrimitive.Input> {
+  /** Inline trailing slot, rendered on the right of the search row. */
+  right?: React.ReactNode
+}
+
+function CommandInput({ className, right, ...props }: CommandInputProps) {
   return (
     <div className="flex h-11 items-center gap-2 border-b border-border px-3" data-slot="command-input-wrapper">
       <SearchIcon className="size-4 shrink-0 text-muted-foreground" />
@@ -29,6 +34,7 @@ function CommandInput({ className, ...props }: React.ComponentProps<typeof Comma
         data-slot="command-input"
         {...props}
       />
+      {right}
     </div>
   )
 }

apps/desktop/src/components/ui/dialog.tsx

@@ -35,16 +35,98 @@ function DialogOverlay({ className, ...props }: React.ComponentProps<typeof Dial
   )
 }
 
+type DialogBannerTone = 'error' | 'warn' | 'info'
+
+// Tinted, edge-to-edge bottom banner per tone. Error/warn keep their semantic
+// destructive/primary tokens; info derives from the dialog's own bubble
+// background so it reads as part of the themed dialog — lifted 30% toward white
+// in light mode, deepened 20% toward black in dark mode.
+const DIALOG_BANNER_TONES: Record<DialogBannerTone, string> = {
+  error: 'bg-destructive/12 text-destructive',
+  warn: 'bg-primary/12 text-primary',
+  info: 'bg-[color-mix(in_srgb,var(--ui-chat-bubble-background),white_30%)] text-[color-mix(in_srgb,var(--ui-chat-bubble-background),black_60%)] dark:bg-[color-mix(in_srgb,var(--ui-chat-bubble-background),black_20%)] dark:text-[color-mix(in_srgb,var(--ui-chat-bubble-background),white_60%)]'
+}
+
 function DialogContent({
   className,
   children,
   showCloseButton = true,
+  fitContent = false,
+  banner,
+  bannerTone = 'error',
   ...props
 }: React.ComponentProps<typeof DialogPrimitive.Content> & {
   showCloseButton?: boolean
+  // Size the dialog to its content (capped at the viewport) instead of the
+  // default fixed `max-w-lg`. For content that has no intrinsic width (grids,
+  // full-width inputs) pair it with a `min-w-*` in `className`.
+  fitContent?: boolean
+  // A dialog-level notice rendered as a banner flush to the bottom edge (tinted,
+  // inherited bottom radius) so it reads as part of the dialog, not a floating
+  // alert. Falsy → no banner. Tone picks the colour.
+  banner?: React.ReactNode
+  bannerTone?: DialogBannerTone
 }) {
   const { t } = useI18n()
 
+  const widthClass = fitContent ? 'w-auto max-w-[92vw]' : 'w-full max-w-lg'
+
+  const closeButton = showCloseButton ? (
+    <DialogPrimitive.Close asChild data-slot="dialog-close-button">
+      <Button
+        aria-label={t.common.close}
+        className="absolute right-2.5 top-2.5 z-20 text-(--ui-text-tertiary) hover:bg-(--chrome-action-hover) hover:text-foreground"
+        size="icon-xs"
+        variant="ghost"
+      >
+        <X className="size-4" />
+        <span className="sr-only">{t.common.close}</span>
+      </Button>
+    </DialogPrimitive.Close>
+  ) : null
+
+  // With a banner, the border can't live on the scroll/clip box (it would draw a
+  // line around the banner too). The white body keeps its own bottom radius and
+  // sits over the tinted footer; the outer shell only clips the banner to the
+  // dialog's rounded bottom edge.
+  if (banner) {
+    return (
+      <DialogPortal>
+        <DialogOverlay />
+        <DialogPrimitive.Content
+          className={cn(
+            'fixed left-1/2 top-1/2 z-[130] pointer-events-auto flex max-h-[85vh] -translate-x-1/2 -translate-y-1/2 flex-col overflow-hidden rounded-xl bg-(--ui-chat-bubble-background) text-[length:var(--conversation-text-font-size)] text-foreground shadow-nous duration-200 data-[state=closed]:animate-out data-[state=closed]:fade-out-0 data-[state=closed]:zoom-out-95 data-[state=open]:animate-in data-[state=open]:fade-in-0 data-[state=open]:zoom-in-95',
+            widthClass,
+            className,
+            // Callers often pass `gap-*` for the no-banner grid layout — suppress
+            // it here so the banner can tuck under the body's rounded bottom edge.
+            'gap-0'
+          )}
+          data-slot="dialog-content"
+          {...props}
+        >
+          {/* Scroll lives on an inner box so this shell keeps a painted bottom radius. */}
+          <div className="relative z-10 overflow-hidden rounded-xl border border-b-0 border-(--stroke-nous) bg-(--ui-chat-bubble-background)">
+            <div className="grid max-h-[calc(85vh-5rem)] min-h-0 gap-3 overflow-y-auto p-4">{children}</div>
+          </div>
+          <div
+            className={cn(
+              // Overlap by one corner radius so the white bottom lobes read clearly
+              // over the tint instead of meeting it on a straight seam.
+              'relative z-0 -mt-[var(--radius-xl)] px-4 pb-2.5 pt-[calc(var(--radius-xl)+0.625rem)] text-center text-[length:var(--conversation-tool-font-size)] leading-relaxed shadow-[inset_0_7px_7px_-4px_rgb(0_0_0/0.28)]',
+              DIALOG_BANNER_TONES[bannerTone]
+            )}
+            data-slot="dialog-banner"
+            role={bannerTone === 'error' ? 'alert' : 'status'}
+          >
+            {banner}
+          </div>
+          {closeButton}
+        </DialogPrimitive.Content>
+      </DialogPortal>
+    )
+  }
+
   return (
     <DialogPortal>
       <DialogOverlay />
@@ -53,26 +135,15 @@ function DialogContent({
           // Cap height at 85vh and let long content scroll inside the dialog
           // instead of overflowing off-screen (long cron titles, tool detail
           // dumps, etc.). Individual dialogs can still override via className.
-          'fixed left-1/2 top-1/2 z-[130] pointer-events-auto grid max-h-[85vh] w-full max-w-lg -translate-x-1/2 -translate-y-1/2 gap-3 overflow-y-auto rounded-xl border border-(--stroke-nous) bg-(--ui-chat-bubble-background) p-4 text-[length:var(--conversation-text-font-size)] text-foreground shadow-nous duration-200 data-[state=closed]:animate-out data-[state=closed]:fade-out-0 data-[state=closed]:zoom-out-95 data-[state=open]:animate-in data-[state=open]:fade-in-0 data-[state=open]:zoom-in-95',
+          'fixed left-1/2 top-1/2 z-[130] pointer-events-auto grid max-h-[85vh] -translate-x-1/2 -translate-y-1/2 gap-3 overflow-y-auto rounded-xl border border-(--stroke-nous) bg-(--ui-chat-bubble-background) p-4 text-[length:var(--conversation-text-font-size)] text-foreground shadow-nous duration-200 data-[state=closed]:animate-out data-[state=closed]:fade-out-0 data-[state=closed]:zoom-out-95 data-[state=open]:animate-in data-[state=open]:fade-in-0 data-[state=open]:zoom-in-95',
+          widthClass,
           className
         )}
         data-slot="dialog-content"
         {...props}
       >
         {children}
-        {showCloseButton && (
-          <DialogPrimitive.Close asChild data-slot="dialog-close-button">
-            <Button
-              aria-label={t.common.close}
-              className="absolute right-2.5 top-2.5 text-(--ui-text-tertiary) hover:bg-(--chrome-action-hover) hover:text-foreground"
-              size="icon-xs"
-              variant="ghost"
-            >
-              <X className="size-4" />
-              <span className="sr-only">{t.common.close}</span>
-            </Button>
-          </DialogPrimitive.Close>
-        )}
+        {closeButton}
       </DialogPrimitive.Content>
     </DialogPortal>
   )

apps/desktop/src/components/ui/generate-button.tsx

@@ -0,0 +1,62 @@
+import type * as React from 'react'
+
+import { Button } from '@/components/ui/button'
+import { Codicon } from '@/components/ui/codicon'
+import { Tip } from '@/components/ui/tooltip'
+import { Square } from '@/lib/icons'
+import { cn } from '@/lib/utils'
+
+interface GenerateButtonProps extends Omit<React.ComponentProps<typeof Button>, 'children' | 'onClick'> {
+  /** True while a generation is in flight. */
+  generating: boolean
+  /** Start a generation. */
+  onGenerate: () => void
+  /** Cancel an in-flight generation. When omitted, the button just spins while
+   *  generating (for one-shots that can't be cancelled). */
+  onCancel?: () => void
+  /** Tooltip + aria label at rest (and while generating if no `generatingLabel`). */
+  label: string
+  /** Tooltip while generating (e.g. "Stop" with cancel, "Generating…" without). */
+  generatingLabel?: string
+  iconSize?: number | string
+}
+
+/** The sparkle "generate with AI" affordance — icon + tooltip, shared by the
+ *  commit-message box and the new-project idea field so they stay one pattern.
+ *  Sparkle → click generates; with `onCancel`, a Stop square appears mid-run;
+ *  without it, the sparkle spins until the one-shot resolves. */
+export function GenerateButton({
+  generating,
+  onGenerate,
+  onCancel,
+  label,
+  generatingLabel,
+  disabled,
+  iconSize = 12,
+  className,
+  ...rest
+}: GenerateButtonProps) {
+  const tip = generating ? (generatingLabel ?? label) : label
+  const cancellable = generating && !!onCancel
+
+  return (
+    <Tip label={tip}>
+      <Button
+        aria-label={tip}
+        className={cn('text-muted-foreground/80 hover:text-foreground', className)}
+        disabled={generating ? !onCancel : disabled}
+        onClick={cancellable ? onCancel : onGenerate}
+        size="icon-xs"
+        type="button"
+        variant="ghost"
+        {...rest}
+      >
+        {cancellable ? (
+          <Square className="fill-current" size={11} />
+        ) : (
+          <Codicon name="sparkle" size={iconSize} spinning={generating} />
+        )}
+      </Button>
+    </Tip>
+  )
+}

apps/desktop/src/global.d.ts

@@ -1,3 +1,10 @@
+import type {
+  PetOverlayBounds,
+  PetOverlayControl,
+  PetOverlayOpenRequest,
+  PetOverlayStatePayload
+} from './store/pet-overlay'
+
 export {}
 
 declare global {
@@ -26,6 +33,20 @@ declare global {
       openSessionWindow: (sessionId: string, opts?: { watch?: boolean }) => Promise<{ ok: boolean; error?: string }>
       // Open (or focus) a compact secondary window on the new-session draft.
       openNewSessionWindow: () => Promise<{ ok: boolean; error?: string }>
+      // The pop-out pet overlay: a transparent always-on-top window hosting only
+      // the mascot. The main renderer drives it (open/close/drag + state push);
+      // the overlay sends control messages back (pop-in, composer submit).
+      petOverlay: {
+        open: (request: PetOverlayOpenRequest) => Promise<{ ok: boolean; bounds?: PetOverlayBounds }>
+        close: () => Promise<{ ok: boolean }>
+        setBounds: (bounds: PetOverlayBounds) => void
+        setIgnoreMouse: (ignore: boolean) => void
+        setFocusable: (focusable: boolean) => void
+        pushState: (payload: PetOverlayStatePayload) => void
+        control: (payload: PetOverlayControl) => void
+        onState: (callback: (payload: PetOverlayStatePayload) => void) => () => void
+        onControl: (callback: (payload: PetOverlayControl) => void) => () => void
+      }
       getBootProgress: () => Promise<DesktopBootProgress>
       getConnectionConfig: (profile?: null | string) => Promise<DesktopConnectionConfig>
       saveConnectionConfig: (payload: DesktopConnectionConfigInput) => Promise<DesktopConnectionConfig>

apps/desktop/src/hermes-profile-scope.test.ts

@@ -0,0 +1,49 @@
+import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest'
+
+import {
+  checkHermesUpdate,
+  getActionStatus,
+  getStatus,
+  restartGateway,
+  setApiRequestProfile,
+  updateHermes
+} from './hermes'
+
+// Contract: every backend-targeted action helper must carry the active gateway
+// profile, so a multi-profile / global-remote user's restart, status poll, and
+// update hit the backend they're actually on — not the primary/default. The
+// System-panel "restart does nothing" bug was these helpers dropping it.
+describe('backend action helpers are profile-scoped', () => {
+  const api = vi.fn(async (_req: { path: string; profile?: string }) => ({}) as never)
+
+  beforeEach(() => {
+    ;(window as { hermesDesktop?: unknown }).hermesDesktop = { api }
+    api.mockClear()
+  })
+
+  afterEach(() => {
+    setApiRequestProfile(null)
+    delete (window as { hermesDesktop?: unknown }).hermesDesktop
+  })
+
+  const lastProfile = () => api.mock.calls.at(-1)?.[0].profile
+
+  it('omits profile when none is active (single-profile users unaffected)', () => {
+    void getStatus()
+    expect(lastProfile()).toBeUndefined()
+  })
+
+  it('forwards the active profile to every backend action', () => {
+    setApiRequestProfile('coder')
+
+    void getStatus()
+    void restartGateway()
+    void updateHermes()
+    void checkHermesUpdate()
+    void getActionStatus('gateway-restart')
+
+    for (const call of api.mock.calls) {
+      expect(call[0].profile).toBe('coder')
+    }
+  })
+})

apps/desktop/src/hermes.ts

@@ -274,6 +274,7 @@ export function getGlobalModelInfo(): Promise<ModelInfoResponse> {
 
 export function getStatus(): Promise<StatusResponse> {
   return window.hermesDesktop.api<StatusResponse>({
+    ...profileScoped(),
     path: '/api/status'
   })
 }
@@ -756,6 +757,7 @@ export function setModelAssignment(body: ModelAssignmentRequest): Promise<ModelA
 
 export function restartGateway(): Promise<ActionResponse> {
   return window.hermesDesktop.api<ActionResponse>({
+    ...profileScoped(),
     path: '/api/gateway/restart',
     method: 'POST'
   })
@@ -763,6 +765,7 @@ export function restartGateway(): Promise<ActionResponse> {
 
 export function updateHermes(): Promise<ActionResponse> {
   return window.hermesDesktop.api<ActionResponse>({
+    ...profileScoped(),
     path: '/api/hermes/update',
     method: 'POST'
   })
@@ -773,12 +776,14 @@ export function updateHermes(): Promise<ActionResponse> {
  *  distinct from the Electron client clone's git state. */
 export function checkHermesUpdate(force = false): Promise<BackendUpdateCheckResponse> {
   return window.hermesDesktop.api<BackendUpdateCheckResponse>({
+    ...profileScoped(),
     path: `/api/hermes/update/check${force ? '?force=true' : ''}`
   })
 }
 
 export function getActionStatus(name: string, lines = 200): Promise<ActionStatusResponse> {
   return window.hermesDesktop.api<ActionStatusResponse>({
+    ...profileScoped(),
     path: `/api/actions/${encodeURIComponent(name)}/status?lines=${Math.max(1, lines)}`
   })
 }

apps/desktop/src/i18n/en.ts

@@ -57,6 +57,7 @@ export const en: Translations = {
       backgroundExitedDuringStartup: 'Hermes background process exited during startup.',
       backendStopped: 'Backend stopped',
       desktopBootFailed: 'Desktop boot failed',
+      gatewayConnectionLost: 'Lost connection to the gateway',
       gatewaySignInRequired: 'Gateway sign-in required',
       ipcBridgeUnavailable: 'Desktop IPC bridge is unavailable.'
     },
@@ -211,6 +212,7 @@ export const en: Translations = {
       'session.togglePin': 'Pin / unpin current session',
       'composer.focus': 'Focus composer',
       'composer.modelPicker': 'Open model picker',
+      'composer.voice': 'Start / stop voice conversation',
       'view.toggleSidebar': 'Toggle sessions sidebar',
       'view.toggleRightSidebar': 'Toggle file browser',
       'view.showFiles': 'Show file browser',
@@ -372,7 +374,44 @@ export const en: Translations = {
       installError: 'Could not install that theme.',
       installed: name => `Installed “${name}”.`,
       removeTheme: 'Remove theme',
-      importedBadge: 'Imported'
+      importedBadge: 'Imported',
+      pet: {
+        title: 'Pet',
+        intro:
+          'Adopt an animated petdex mascot that floats over the app and reacts to what Hermes is doing — running while tools execute, celebrating on success, sulking on errors.',
+        restartHint:
+          'Pets need a quick restart — the running app started before this feature was added. Quit and reopen Hermes, then come back here.',
+        on: 'On',
+        off: 'Off',
+        scaleTitle: 'Size',
+        scaleDesc: 'Resize the floating mascot. Applies everywhere instantly.',
+        chooseTitle: 'Choose a pet',
+        chooseDesc: 'Picking one installs it (if needed) and makes it active.',
+        searchPlaceholder: 'Search pets…',
+        unreachable: "Couldn't reach the petdex gallery. Check your connection and reopen this page.",
+        noMatch: query => `No pets match "${query}".`,
+        installedTag: 'installed',
+        generatedTag: 'Generated',
+        countCapped: (cap, total) => `Showing ${cap} of ${total} — type to narrow it down.`,
+        count: n => `${n} pet${n === 1 ? '' : 's'}.`,
+        uninstall: name => `Uninstall ${name}`,
+        delete: name => `Delete ${name}`,
+        deleteTitle: name => `Delete ${name}?`,
+        deleteBody: "This permanently deletes the pet — it can't be reinstalled.",
+        deleteConfirm: 'Delete',
+        rename: name => `Rename ${name}`,
+        renameTitle: 'Rename pet',
+        renamePlaceholder: 'Name your pet',
+        renameSave: 'Save',
+        exportPet: name => `Export ${name}`,
+        adoptFailed: slug => `Could not adopt ${slug}`,
+        uninstallFailed: slug => `Could not uninstall ${slug}`,
+        renameFailed: slug => `Could not rename ${slug}`,
+        exportFailed: slug => `Could not export ${slug}`,
+        noneAvailable: 'No pets available to turn on right now.',
+        turnOnFailed: 'Could not turn the pet on.',
+        turnOffFailed: 'Could not turn the pet off.'
+      }
     },
     fieldLabels: FIELD_LABELS,
     fieldDescriptions: FIELD_DESCRIPTIONS,
@@ -723,8 +762,53 @@ export const en: Translations = {
     commandCenter: 'Command Center',
     appearance: 'Appearance',
     settings: 'Settings',
-    changeTheme: 'Change theme...',
+    changeTheme: 'Change theme',
     changeColorMode: 'Change color mode...',
+    pets: {
+      title: 'Pets',
+      placeholder: 'Search pets…',
+      loading: 'Loading petdex gallery…',
+      error: 'Could not reach the petdex gallery.',
+      staleBackend: 'Restart Hermes to use pets — the backend predates this feature.',
+      empty: 'No matching pets.',
+      turnOff: 'Turn off',
+      turnOn: 'Turn on',
+      installed: 'Installed',
+      generatedTag: 'Generated',
+      adoptFailed: 'Could not adopt that pet.',
+      toggleFailed: 'Could not toggle the pet.',
+      noneAvailable: 'No pets available — pick one below to install.'
+    },
+    generatePet: {
+      title: 'Generate a pet',
+      placeholder: 'Describe a pet to generate…',
+      promptHint: 'Type a description, then press Enter to draft four looks.',
+      readyHint: 'Press Enter to draft four looks from your description.',
+      generate: 'Generate',
+      generating: 'Generating…',
+      retry: 'Retry',
+      hatch: 'Hatch',
+      spawning: 'Spawning…',
+      hatching: 'Hatching your pet…',
+      hatchingSub: 'Bringing it to life…',
+      hatched: 'It hatched!',
+      hatchRow: (_state, done, total) => `Sketching frame ${done} of ${total}…`,
+      hatchComposing: 'Piecing it together…',
+      hatchSaving: 'Almost there…',
+      namePlaceholder: 'Name your pet',
+      staleBackend: 'Update Hermes to generate pets.',
+      backgroundHint: 'You can close this — Hermes will notify you when it’s done.',
+      slowProviderHint: 'This can take several minutes',
+      remix: 'Remix',
+      remixConfirmTitle: 'Remix this look?',
+      remixConfirmBody:
+        'This generates a fresh set of drafts using this one as the starting point. It can take several minutes.',
+      genericError: 'Generation failed — try again or pick a suggestion.',
+      referenceImageTooLarge: 'Reference image is too large. Use one under 16 MB.',
+      referenceImageInvalid: 'Could not read that reference image. Try a PNG, JPG, WebP, or GIF.',
+      adopt: 'Adopt',
+      startOver: 'Start over'
+    },
     installTheme: {
       title: 'Install theme...',
       placeholder: 'Search the VS Code Marketplace...',
@@ -1766,7 +1850,8 @@ export const en: Translations = {
       restoreCheckpoint: 'Restore checkpoint',
       restoreFromHere: 'Restore checkpoint — rerun from this prompt',
       restoreTitle: 'Restore to this checkpoint?',
-      restoreBody: 'Everything after this prompt is removed from the conversation, and the prompt runs again from here.',
+      restoreBody:
+        'Everything after this prompt is removed from the conversation, and the prompt runs again from here.',
       restoreConfirm: 'Restore & rerun',
       restoreNext: 'Restore next checkpoint',
       goForward: 'Go forward',
@@ -1822,7 +1907,67 @@ export const en: Translations = {
       statusRunning: 'Running',
       statusError: 'Error',
       statusRecovered: 'Recovered',
-      statusDone: 'Done'
+      statusDone: 'Done',
+      actions: {
+        read: 'Read',
+        reading: 'Reading',
+        opened: 'Opened',
+        opening: 'Opening',
+        searched: 'Searched',
+        searching: 'Searching',
+        ran: 'Ran',
+        running: 'Running',
+        ranCode: 'Ran code',
+        runningCode: 'Scripting'
+      },
+      prefixes: {
+        browser: 'Browser',
+        web: 'Web'
+      },
+      titleTemplates: {
+        actionCommand: (action, command) => `${action} ${command}`,
+        actionQuoted: (action, value) => `${action} “${value}”`,
+        actionTarget: (action, target) => `${action} ${target}`,
+        prefixedDone: (prefix, action) => `${prefix} ${action}`,
+        runningPrefixedTool: (prefix, action) => `Running ${prefix.toLowerCase()} ${action.toLowerCase()}`,
+        runningTool: action => `Running ${action.toLowerCase()}`
+      },
+      titles: {
+        browser_click: { done: 'Clicked page element', pending: 'Clicking page element', pendingAction: 'Clicking' },
+        browser_fill: { done: 'Filled form field', pending: 'Filling form field', pendingAction: 'Filling' },
+        browser_navigate: { done: 'Opened page', pending: 'Opening page', pendingAction: 'Opening' },
+        browser_snapshot: {
+          done: 'Captured page snapshot',
+          pending: 'Capturing page snapshot',
+          pendingAction: 'Capturing'
+        },
+        browser_take_screenshot: {
+          done: 'Captured screenshot',
+          pending: 'Capturing screenshot',
+          pendingAction: 'Capturing'
+        },
+        browser_type: { done: 'Typed on page', pending: 'Typing on page', pendingAction: 'Typing' },
+        clarify: { done: 'Asked a question', pending: 'Asking a question', pendingAction: 'Asking' },
+        cronjob: { done: 'Cron job', pending: 'Scheduling cron job', pendingAction: 'Scheduling' },
+        edit_file: { done: 'Edited file', pending: 'Editing file', pendingAction: 'Editing' },
+        execute_code: { done: 'Ran code', pending: 'Scripting', pendingAction: 'Scripting' },
+        image_generate: { done: 'Generated image', pending: 'Generating image', pendingAction: 'Generating' },
+        list_files: { done: 'Listed files', pending: 'Listing files', pendingAction: 'Listing' },
+        patch: { done: 'Patched file', pending: 'Patching file', pendingAction: 'Patching' },
+        read_file: { done: 'Read file', pending: 'Reading file', pendingAction: 'Reading' },
+        search_files: { done: 'Searched files', pending: 'Searching files', pendingAction: 'Searching' },
+        session_search_recall: {
+          done: 'Searched session history',
+          pending: 'Searching session history',
+          pendingAction: 'Searching'
+        },
+        terminal: { done: 'Ran command', pending: 'Running command', pendingAction: 'Running' },
+        todo: { done: 'Updated todos', pending: 'Updating todos', pendingAction: 'Updating' },
+        vision_analyze: { done: 'Analyzed image', pending: 'Analyzing image', pendingAction: 'Analyzing' },
+        web_extract: { done: 'Read webpage', pending: 'Reading webpage', pendingAction: 'Reading' },
+        web_search: { done: 'Searched web', pending: 'Searching web', pendingAction: 'Searching' },
+        write_file: { done: 'Edited file', pending: 'Editing file', pendingAction: 'Editing' }
+      }
     }
   },
 
@@ -1865,7 +2010,8 @@ export const en: Translations = {
     editFailed: 'Edit failed',
     resumeFailed: 'Resume failed',
     resumeStrandedTitle: "Couldn't load this session",
-    resumeStrandedBody: 'The connection to this session failed and automatic retries gave up. Check that the gateway is running, then try again.',
+    resumeStrandedBody:
+      'The connection to this session failed and automatic retries gave up. Check that the gateway is running, then try again.',
     resumeRetry: 'Retry',
     nothingToBranch: 'Nothing to branch',
     branchNeedsChat: 'Start or resume a chat before branching.',

apps/desktop/src/i18n/index.ts

@@ -17,4 +17,4 @@ export {
   normalizeLocale
 } from './languages'
 export { setRuntimeI18nLocale, translateNow } from './runtime'
-export type { Locale, Translations } from './types'
+export type { Locale, ToolTitleKey, Translations } from './types'

apps/desktop/src/i18n/ja.ts

@@ -57,6 +57,7 @@ export const ja = defineLocale({
       backgroundExitedDuringStartup: '起動中に Hermes バックグラウンドプロセスが終了しました。',
       backendStopped: 'バックエンドが停止しました',
       desktopBootFailed: 'デスクトップの起動に失敗しました',
+      gatewayConnectionLost: 'ゲートウェイへの接続が切断されました',
       gatewaySignInRequired: 'ゲートウェイへのサインインが必要です',
       ipcBridgeUnavailable: 'デスクトップ IPC ブリッジが利用できません。'
     },
@@ -200,8 +201,7 @@ export const ja = defineLocale({
     },
     notifications: {
       title: '通知',
-      intro:
-        'アプリ内トーストとは別の、ネイティブのデスクトップ通知です。設定は端末ごとに保存されます。',
+      intro: 'アプリ内トーストとは別の、ネイティブのデスクトップ通知です。設定は端末ごとに保存されます。',
       enableAll: '通知を有効にする',
       enableAllDesc: 'マスタースイッチ。オフにすると以下のすべての通知を無効にします。',
       focusedHint: '完了通知は Hermes がバックグラウンドにあるときのみ表示されます。',
@@ -287,7 +287,44 @@ export const ja = defineLocale({
       installError: 'そのテーマをインストールできませんでした。',
       installed: name => `「${name}」をインストールしました。`,
       removeTheme: 'テーマを削除',
-      importedBadge: 'インポート済み'
+      importedBadge: 'インポート済み',
+      pet: {
+        title: 'ペット',
+        intro:
+          'アプリ上に浮かぶ petdex のアニメーションマスコットを採用しましょう。ツール実行中は走り、成功すると喜び、エラーでしょんぼりと、Hermes の状態に反応します。',
+        restartHint:
+          'ペット機能には再起動が必要です。この機能が追加される前に起動したアプリが動作中です。Hermes を終了して再度開き、このページに戻ってください。',
+        scaleTitle: 'サイズ',
+        scaleDesc: '浮遊マスコットの大きさを変更します。すべての画面に即時反映されます。',
+        on: 'オン',
+        off: 'オフ',
+        chooseTitle: 'ペットを選ぶ',
+        chooseDesc: '選ぶと（必要に応じて）インストールされ、アクティブになります。',
+        searchPlaceholder: 'ペットを検索…',
+        unreachable: 'petdex ギャラリーに接続できませんでした。接続を確認してこのページを開き直してください。',
+        noMatch: query => `「${query}」に一致するペットがありません。`,
+        installedTag: 'インストール済み',
+        generatedTag: '生成',
+        countCapped: (cap, total) => `${total} 件中 ${cap} 件を表示中——入力して絞り込めます。`,
+        count: n => `${n} 件のペット。`,
+        uninstall: name => `${name} をアンインストール`,
+        delete: name => `${name} を削除`,
+        deleteTitle: name => `${name} を削除しますか？`,
+        deleteBody: 'ペットを完全に削除します。再インストールはできません。',
+        deleteConfirm: '削除',
+        rename: name => `${name} の名前を変更`,
+        renameTitle: 'ペットの名前を変更',
+        renamePlaceholder: 'ペットに名前を付ける',
+        renameSave: '保存',
+        exportPet: name => `${name} をエクスポート`,
+        adoptFailed: slug => `${slug} を採用できませんでした`,
+        uninstallFailed: slug => `${slug} をアンインストールできませんでした`,
+        renameFailed: slug => `${slug} の名前を変更できませんでした`,
+        exportFailed: slug => `${slug} をエクスポートできませんでした`,
+        noneAvailable: 'オンにできるペットがありません。',
+        turnOnFailed: 'ペットをオンにできませんでした。',
+        turnOffFailed: 'ペットをオフにできませんでした。'
+      }
     },
     fieldLabels: defineFieldCopy({
       model: 'デフォルトモデル',
@@ -843,8 +880,52 @@ export const ja = defineLocale({
     commandCenter: 'コマンドセンター',
     appearance: '外観',
     settings: '設定',
-    changeTheme: 'テーマを変更...',
+    changeTheme: 'テーマを変更',
     changeColorMode: 'カラーモードを変更...',
+    pets: {
+      title: 'ペット',
+      placeholder: 'ペットを検索…',
+      loading: 'petdex ギャラリーを読み込み中…',
+      error: 'petdex ギャラリーに接続できません。',
+      staleBackend: 'ペット機能を使うには Hermes を再起動してください。',
+      empty: '一致するペットがありません。',
+      turnOff: 'オフ',
+      turnOn: 'オン',
+      installed: 'インストール済み',
+      generatedTag: '生成',
+      adoptFailed: 'ペットを採用できませんでした。',
+      toggleFailed: 'ペットを切り替えできませんでした。',
+      noneAvailable: '利用可能なペットがありません。'
+    },
+    generatePet: {
+      title: 'ペットを生成',
+      placeholder: '生成するペットを説明…',
+      promptHint: '説明を入力して Enter を押すと、4 つの見た目を生成します。',
+      readyHint: 'Enter を押すと、説明から 4 つの見た目を生成します。',
+      generate: '生成',
+      generating: '生成中…',
+      retry: '再試行',
+      hatch: '孵化',
+      spawning: 'スポーン中…',
+      hatching: 'ペットを孵化しています…',
+      hatchingSub: '命を吹き込んでいます…',
+      hatched: '孵化しました！',
+      hatchRow: (_state, done, total) => `フレームを描画中… ${done}/${total}`,
+      hatchComposing: 'まとめています…',
+      hatchSaving: 'もうすぐです…',
+      namePlaceholder: 'ペットに名前を付ける',
+      staleBackend: 'ペットを生成するには Hermes を更新してください。',
+      backgroundHint: 'このウィンドウは閉じても大丈夫です。完了したら Hermes が通知します。',
+      slowProviderHint: '数分かかることがあります',
+      remix: 'リミックス',
+      remixConfirmTitle: 'この見た目でリミックスしますか？',
+      remixConfirmBody: 'これを起点に新しい候補を生成します。数分かかることがあります。',
+      genericError: '生成に失敗しました。もう一度試すか、候補を選んでください。',
+      referenceImageTooLarge: '参照画像が大きすぎます。16 MB 未満の画像を使ってください。',
+      referenceImageInvalid: '参照画像を読み込めませんでした。PNG/JPG/WebP/GIF を試してください。',
+      adopt: '迎え入れる',
+      startOver: 'やり直す'
+    },
     installTheme: {
       title: 'テーマをインストール...',
       placeholder: 'VS Code Marketplace を検索...',
@@ -1420,7 +1501,8 @@ export const ja = defineLocale({
     queueSend: '送信',
     queueDelete: '削除',
     queueStuckTitle: 'キュー内のメッセージを送信できません',
-    queueStuckBody: 'キューに入れたターンの送信が繰り返し失敗しました。まだキューに残っています。もう一度送信してください。',
+    queueStuckBody:
+      'キューに入れたターンの送信が繰り返し失敗しました。まだキューに残っています。もう一度送信してください。',
     previewUnavailable: 'プレビューは利用できません',
     previewLabel: label => `${label} のプレビュー`,
     couldNotPreview: label => `${label} をプレビューできませんでした`,
@@ -1519,7 +1601,8 @@ export const ja = defineLocale({
     copy: 'コピー',
     copied: 'コピーしました',
     done: '完了',
-    applyingBody: 'Hermes アップデーターが独自のウィンドウで引き継ぎ、完了後に自動的に Hermes を再度開きます。更新中はご自分で Hermes を開き直さないでください。',
+    applyingBody:
+      'Hermes アップデーターが独自のウィンドウで引き継ぎ、完了後に自動的に Hermes を再度開きます。更新中はご自分で Hermes を開き直さないでください。',
     applyingBodyBackend: 'リモートバックエンドが更新を適用して再起動します。復帰すると Hermes が自動的に再接続します。',
     applyingClose: 'このウィンドウは更新中に閉じ、その後 Hermes が自動的に再度開きます。',
     errorTitle: '更新が完了しませんでした',
@@ -1951,7 +2034,83 @@ export const ja = defineLocale({
       statusRunning: '実行中',
       statusError: 'エラー',
       statusRecovered: '回復しました',
-      statusDone: '完了'
+      statusDone: '完了',
+      actions: {
+        read: '読み取り完了',
+        reading: '読み取り中',
+        opened: 'オープン済み',
+        opening: 'オープン中',
+        searched: '検索完了',
+        searching: '検索中',
+        ran: '実行完了',
+        running: '実行中',
+        ranCode: 'コード実行完了',
+        runningCode: 'スクリプト作成中'
+      },
+      prefixes: {
+        browser: 'ブラウザー',
+        web: 'Web'
+      },
+      titleTemplates: {
+        actionCommand: (action, command) => `${action} ${command}`,
+        actionQuoted: (action, value) => `「${value}」を${action}`,
+        actionTarget: (action, target) => `${target} を${action}`,
+        prefixedDone: (prefix, action) => `${prefix} ${action}`,
+        runningPrefixedTool: (prefix, action) => `${prefix} ${action}を実行中`,
+        runningTool: action => `${action}を実行中`
+      },
+      titles: {
+        browser_click: {
+          done: 'ページ要素をクリックしました',
+          pending: 'ページ要素をクリック中',
+          pendingAction: 'クリック中'
+        },
+        browser_fill: { done: 'フォーム欄に入力しました', pending: 'フォーム欄に入力中', pendingAction: '入力中' },
+        browser_navigate: { done: 'ページを開きました', pending: 'ページをオープン中', pendingAction: 'オープン中' },
+        browser_snapshot: {
+          done: 'ページスナップショットを取得しました',
+          pending: 'ページスナップショットを取得中',
+          pendingAction: '取得中'
+        },
+        browser_take_screenshot: {
+          done: 'スクリーンショットを取得しました',
+          pending: 'スクリーンショットを取得中',
+          pendingAction: '取得中'
+        },
+        browser_type: { done: 'ページに入力しました', pending: 'ページに入力中', pendingAction: '入力中' },
+        clarify: { done: '質問しました', pending: '質問中', pendingAction: '質問中' },
+        cronjob: { done: 'Cron ジョブ', pending: 'Cron ジョブをスケジュール中', pendingAction: 'スケジュール中' },
+        edit_file: { done: 'ファイルを編集しました', pending: 'ファイルを編集中', pendingAction: '編集中' },
+        execute_code: { done: 'コードを実行しました', pending: 'スクリプト作成中', pendingAction: 'スクリプト作成中' },
+        image_generate: { done: '画像を生成しました', pending: '画像を生成中', pendingAction: '生成中' },
+        list_files: {
+          done: 'ファイルを一覧表示しました',
+          pending: 'ファイルを一覧表示中',
+          pendingAction: '一覧表示中'
+        },
+        patch: {
+          done: 'ファイルにパッチを適用しました',
+          pending: 'ファイルにパッチ適用中',
+          pendingAction: 'パッチ適用中'
+        },
+        read_file: { done: 'ファイルを読み取りました', pending: 'ファイルを読み取り中', pendingAction: '読み取り中' },
+        search_files: { done: 'ファイルを検索しました', pending: 'ファイルを検索中', pendingAction: '検索中' },
+        session_search_recall: {
+          done: 'セッション履歴を検索しました',
+          pending: 'セッション履歴を検索中',
+          pendingAction: '検索中'
+        },
+        terminal: { done: 'コマンドを実行しました', pending: 'コマンドを実行中', pendingAction: '実行中' },
+        todo: { done: 'Todo を更新しました', pending: 'Todo を更新中', pendingAction: '更新中' },
+        vision_analyze: { done: '画像を分析しました', pending: '画像を分析中', pendingAction: '分析中' },
+        web_extract: {
+          done: 'Web ページを読み取りました',
+          pending: 'Web ページを読み取り中',
+          pendingAction: '読み取り中'
+        },
+        web_search: { done: 'Web を検索しました', pending: 'Web を検索中', pendingAction: '検索中' },
+        write_file: { done: 'ファイルを編集しました', pending: 'ファイルを編集中', pendingAction: '編集中' }
+      }
     }
   },
 
@@ -1995,7 +2154,8 @@ export const ja = defineLocale({
     editFailed: '編集に失敗しました',
     resumeFailed: '再開に失敗しました',
     resumeStrandedTitle: 'このセッションを読み込めませんでした',
-    resumeStrandedBody: 'このセッションへの接続に失敗し、自動再試行も停止しました。ゲートウェイが実行中か確認してから、もう一度お試しください。',
+    resumeStrandedBody:
+      'このセッションへの接続に失敗し、自動再試行も停止しました。ゲートウェイが実行中か確認してから、もう一度お試しください。',
     resumeRetry: '再試行',
     nothingToBranch: 'ブランチするものがありません',
     branchNeedsChat: 'ブランチする前にチャットを開始または再開してください。',

apps/desktop/src/i18n/types.ts

@@ -7,6 +7,36 @@
 
 export type Locale = 'en' | 'zh' | 'zh-hant' | 'ja'
 
+export type ToolTitleKey =
+  | 'browser_click'
+  | 'browser_fill'
+  | 'browser_navigate'
+  | 'browser_snapshot'
+  | 'browser_take_screenshot'
+  | 'browser_type'
+  | 'clarify'
+  | 'cronjob'
+  | 'edit_file'
+  | 'execute_code'
+  | 'image_generate'
+  | 'list_files'
+  | 'patch'
+  | 'read_file'
+  | 'search_files'
+  | 'session_search_recall'
+  | 'terminal'
+  | 'todo'
+  | 'vision_analyze'
+  | 'web_extract'
+  | 'web_search'
+  | 'write_file'
+
+interface ToolTitleCopy {
+  done: string
+  pending: string
+  pendingAction: string
+}
+
 interface ModeOptionCopy {
   label: string
   description: string
@@ -72,6 +102,7 @@ export interface Translations {
       backgroundExitedDuringStartup: string
       backendStopped: string
       desktopBootFailed: string
+      gatewayConnectionLost: string
       gatewaySignInRequired: string
       ipcBridgeUnavailable: string
     }
@@ -270,6 +301,41 @@ export interface Translations {
       installed: (name: string) => string
       removeTheme: string
       importedBadge: string
+      pet: {
+        title: string
+        intro: string
+        restartHint: string
+        on: string
+        off: string
+        scaleTitle: string
+        scaleDesc: string
+        chooseTitle: string
+        chooseDesc: string
+        searchPlaceholder: string
+        unreachable: string
+        noMatch: (query: string) => string
+        installedTag: string
+        generatedTag: string
+        countCapped: (cap: number, total: number) => string
+        count: (n: number) => string
+        uninstall: (name: string) => string
+        delete: (name: string) => string
+        deleteTitle: (name: string) => string
+        deleteBody: string
+        deleteConfirm: string
+        rename: (name: string) => string
+        renameTitle: string
+        renamePlaceholder: string
+        renameSave: string
+        exportPet: (name: string) => string
+        adoptFailed: (slug: string) => string
+        uninstallFailed: (slug: string) => string
+        renameFailed: (slug: string) => string
+        exportFailed: (slug: string) => string
+        noneAvailable: string
+        turnOnFailed: string
+        turnOffFailed: string
+      }
     }
     fieldLabels: Record<string, string>
     fieldDescriptions: Record<string, string>
@@ -602,6 +668,50 @@ export interface Translations {
     settings: string
     changeTheme: string
     changeColorMode: string
+    pets: {
+      title: string
+      placeholder: string
+      loading: string
+      error: string
+      staleBackend: string
+      empty: string
+      turnOff: string
+      turnOn: string
+      installed: string
+      generatedTag: string
+      adoptFailed: string
+      toggleFailed: string
+      noneAvailable: string
+    }
+    generatePet: {
+      title: string
+      placeholder: string
+      promptHint: string
+      readyHint: string
+      generate: string
+      generating: string
+      retry: string
+      hatch: string
+      spawning: string
+      hatching: string
+      hatchingSub: string
+      hatched: string
+      hatchRow: (state: string, done: number, total: number) => string
+      hatchComposing: string
+      hatchSaving: string
+      namePlaceholder: string
+      staleBackend: string
+      backgroundHint: string
+      slowProviderHint: string
+      remix: string
+      remixConfirmTitle: string
+      remixConfirmBody: string
+      genericError: string
+      referenceImageTooLarge: string
+      referenceImageInvalid: string
+      adopt: string
+      startOver: string
+    }
     installTheme: {
       title: string
       placeholder: string
@@ -1457,6 +1567,31 @@ export interface Translations {
       statusError: string
       statusRecovered: string
       statusDone: string
+      actions: {
+        read: string
+        reading: string
+        opened: string
+        opening: string
+        searched: string
+        searching: string
+        ran: string
+        running: string
+        ranCode: string
+        runningCode: string
+      }
+      prefixes: {
+        browser: string
+        web: string
+      }
+      titleTemplates: {
+        actionCommand: (action: string, command: string) => string
+        actionQuoted: (action: string, value: string) => string
+        actionTarget: (action: string, target: string) => string
+        prefixedDone: (prefix: string, action: string) => string
+        runningPrefixedTool: (prefix: string, action: string) => string
+        runningTool: (action: string) => string
+      }
+      titles: Record<ToolTitleKey, ToolTitleCopy>
     }
   }
 

apps/desktop/src/i18n/zh-hant.ts

@@ -57,6 +57,7 @@ export const zhHant = defineLocale({
       backgroundExitedDuringStartup: 'Hermes 背景程序在啟動期間結束。',
       backendStopped: '後端已停止',
       desktopBootFailed: '桌面啟動失敗',
+      gatewayConnectionLost: '與閘道的連線已中斷',
       gatewaySignInRequired: '需要閘道登入',
       ipcBridgeUnavailable: '桌面 IPC 橋接器不可用。'
     },
@@ -276,7 +277,43 @@ export const zhHant = defineLocale({
       installError: '無法安裝該主題。',
       installed: name => `已安裝「${name}」。`,
       removeTheme: '移除主題',
-      importedBadge: '已匯入'
+      importedBadge: '已匯入',
+      pet: {
+        title: '寵物',
+        intro:
+          '領養一隻懸浮在應用上的 petdex 動畫寵物，它會根據 Hermes 的狀態做出反應——工具執行時奔跑、成功時歡呼、出錯時沮喪。',
+        restartHint: '寵物功能需要重新啟動——目前執行的應用在此功能加入前啟動。請結束並重新開啟 Hermes，然後回到此處。',
+        scaleTitle: '大小',
+        scaleDesc: '調整懸浮寵物的大小，所有介面即時生效。',
+        on: '開啟',
+        off: '關閉',
+        chooseTitle: '選擇寵物',
+        chooseDesc: '選擇後會自動安裝（如需）並設為目前寵物。',
+        searchPlaceholder: '搜尋寵物…',
+        unreachable: '無法連線至 petdex 畫廊。請檢查網路連線並重新開啟此頁面。',
+        noMatch: query => `沒有符合「${query}」的寵物。`,
+        installedTag: '已安裝',
+        generatedTag: '生成',
+        countCapped: (cap, total) => `顯示 ${total} 個中的 ${cap} 個——輸入關鍵字以縮小範圍。`,
+        count: n => `${n} 個寵物。`,
+        uninstall: name => `解除安裝 ${name}`,
+        delete: name => `刪除 ${name}`,
+        deleteTitle: name => `刪除 ${name}？`,
+        deleteBody: '此操作會永久刪除寵物，且無法重新安裝。',
+        deleteConfirm: '刪除',
+        rename: name => `重新命名 ${name}`,
+        renameTitle: '重新命名寵物',
+        renamePlaceholder: '為寵物取個名字',
+        renameSave: '儲存',
+        exportPet: name => `匯出 ${name}`,
+        adoptFailed: slug => `無法領養 ${slug}`,
+        uninstallFailed: slug => `無法解除安裝 ${slug}`,
+        renameFailed: slug => `無法重新命名 ${slug}`,
+        exportFailed: slug => `無法匯出 ${slug}`,
+        noneAvailable: '目前沒有可開啟的寵物。',
+        turnOnFailed: '無法開啟寵物。',
+        turnOffFailed: '無法關閉寵物。'
+      }
     },
     fieldLabels: defineFieldCopy({
       model: '預設模型',
@@ -815,8 +852,52 @@ export const zhHant = defineLocale({
     commandCenter: '命令中心',
     appearance: '外觀',
     settings: '設定',
-    changeTheme: '變更主題...',
+    changeTheme: '變更主題',
     changeColorMode: '變更色彩模式...',
+    pets: {
+      title: '寵物',
+      placeholder: '搜尋寵物…',
+      loading: '正在載入 petdex 畫廊…',
+      error: '無法連線至 petdex 畫廊。',
+      staleBackend: '請重新啟動 Hermes 以使用寵物功能。',
+      empty: '沒有符合的寵物。',
+      turnOff: '關閉',
+      turnOn: '開啟',
+      installed: '已安裝',
+      generatedTag: '生成',
+      adoptFailed: '無法領養該寵物。',
+      toggleFailed: '無法切換寵物顯示。',
+      noneAvailable: '尚無可用寵物——請在下方選擇一個安裝。'
+    },
+    generatePet: {
+      title: '生成寵物',
+      placeholder: '描述要生成的寵物……',
+      promptHint: '輸入描述，然後按 Enter 生成四種造型。',
+      readyHint: '按 Enter 依描述生成四種造型。',
+      generate: '生成',
+      generating: '生成中……',
+      retry: '重試',
+      hatch: '孵化',
+      spawning: '召喚中……',
+      hatching: '正在孵化你的寵物……',
+      hatchingSub: '正在注入生命……',
+      hatched: '孵化成功！',
+      hatchRow: (_state, done, total) => `正在繪製畫面…… ${done}/${total}`,
+      hatchComposing: '正在拼合……',
+      hatchSaving: '快好了……',
+      namePlaceholder: '為寵物命名',
+      staleBackend: '請更新 Hermes 以生成寵物。',
+      backgroundHint: '你可以關閉此視窗——完成後 Hermes 會通知你。',
+      slowProviderHint: '這可能需要幾分鐘',
+      remix: '混合生成',
+      remixConfirmTitle: '以此造型混合生成？',
+      remixConfirmBody: '將以此造型為起點生成一組新草圖，可能需要幾分鐘。',
+      genericError: '生成失敗——請重試或選一個建議。',
+      referenceImageTooLarge: '參考圖片過大。請使用小於 16 MB 的圖片。',
+      referenceImageInvalid: '無法讀取該參考圖片。請嘗試 PNG、JPG、WebP 或 GIF。',
+      adopt: '領養',
+      startOver: '重新開始'
+    },
     installTheme: {
       title: '安裝主題...',
       placeholder: '搜尋 VS Code Marketplace...',
@@ -1470,7 +1551,8 @@ export const zhHant = defineLocale({
     copy: '複製',
     copied: '已複製',
     done: '完成',
-    applyingBody: 'Hermes 更新程式會在自己的視窗中接管，並在完成後自動重新開啟 Hermes。更新期間請勿自行重新開啟 Hermes。',
+    applyingBody:
+      'Hermes 更新程式會在自己的視窗中接管，並在完成後自動重新開啟 Hermes。更新期間請勿自行重新開啟 Hermes。',
     applyingBodyBackend: '遠端後端正在套用更新並將重新啟動。恢復後 Hermes 會自動重新連線。',
     applyingClose: '此視窗會在更新期間關閉，隨後 Hermes 會自動重新開啟。',
     errorTitle: '更新未完成',
@@ -1892,7 +1974,59 @@ export const zhHant = defineLocale({
       statusRunning: '執行中',
       statusError: '錯誤',
       statusRecovered: '已復原',
-      statusDone: '完成'
+      statusDone: '完成',
+      actions: {
+        read: '已讀取',
+        reading: '正在讀取',
+        opened: '已開啟',
+        opening: '正在開啟',
+        searched: '已搜尋',
+        searching: '正在搜尋',
+        ran: '已執行',
+        running: '正在執行',
+        ranCode: '已執行程式碼',
+        runningCode: '正在撰寫腳本'
+      },
+      prefixes: {
+        browser: '瀏覽器',
+        web: '網頁'
+      },
+      titleTemplates: {
+        actionCommand: (action, command) => `${action} ${command}`,
+        actionQuoted: (action, value) => `${action}「${value}」`,
+        actionTarget: (action, target) => `${action} ${target}`,
+        prefixedDone: (prefix, action) => `${prefix}${action}`,
+        runningPrefixedTool: (prefix, action) => `正在執行${prefix}${action}`,
+        runningTool: action => `正在執行 ${action}`
+      },
+      titles: {
+        browser_click: { done: '已點擊頁面元素', pending: '正在點擊頁面元素', pendingAction: '正在點擊' },
+        browser_fill: { done: '已填寫表單欄位', pending: '正在填寫表單欄位', pendingAction: '正在填寫' },
+        browser_navigate: { done: '已開啟頁面', pending: '正在開啟頁面', pendingAction: '正在開啟' },
+        browser_snapshot: { done: '已擷取頁面快照', pending: '正在擷取頁面快照', pendingAction: '正在擷取' },
+        browser_take_screenshot: { done: '已擷取截圖', pending: '正在擷取截圖', pendingAction: '正在擷取' },
+        browser_type: { done: '已在頁面輸入', pending: '正在頁面輸入', pendingAction: '正在輸入' },
+        clarify: { done: '已提問', pending: '正在提問', pendingAction: '正在提問' },
+        cronjob: { done: 'Cron 工作', pending: '正在安排 Cron 工作', pendingAction: '正在安排' },
+        edit_file: { done: '已編輯檔案', pending: '正在編輯檔案', pendingAction: '正在編輯' },
+        execute_code: { done: '已執行程式碼', pending: '正在撰寫腳本', pendingAction: '正在撰寫腳本' },
+        image_generate: { done: '已生成圖片', pending: '正在生成圖片', pendingAction: '正在生成' },
+        list_files: { done: '已列出檔案', pending: '正在列出檔案', pendingAction: '正在列出' },
+        patch: { done: '已修補檔案', pending: '正在修補檔案', pendingAction: '正在修補' },
+        read_file: { done: '已讀取檔案', pending: '正在讀取檔案', pendingAction: '正在讀取' },
+        search_files: { done: '已搜尋檔案', pending: '正在搜尋檔案', pendingAction: '正在搜尋' },
+        session_search_recall: {
+          done: '已搜尋工作階段歷史',
+          pending: '正在搜尋工作階段歷史',
+          pendingAction: '正在搜尋'
+        },
+        terminal: { done: '已執行指令', pending: '正在執行指令', pendingAction: '正在執行' },
+        todo: { done: '已更新待辦', pending: '正在更新待辦', pendingAction: '正在更新' },
+        vision_analyze: { done: '已分析圖片', pending: '正在分析圖片', pendingAction: '正在分析' },
+        web_extract: { done: '已讀取網頁', pending: '正在讀取網頁', pendingAction: '正在讀取' },
+        web_search: { done: '已搜尋網頁', pending: '正在搜尋網頁', pendingAction: '正在搜尋' },
+        write_file: { done: '已編輯檔案', pending: '正在編輯檔案', pendingAction: '正在編輯' }
+      }
     }
   },
 

apps/desktop/src/i18n/zh.ts

@@ -57,6 +57,7 @@ export const zh: Translations = {
       backgroundExitedDuringStartup: 'Hermes 后台进程在启动期间退出。',
       backendStopped: '后端已停止',
       desktopBootFailed: '桌面启动失败',
+      gatewayConnectionLost: '与网关的连接已断开',
       gatewaySignInRequired: '需要登录网关',
       ipcBridgeUnavailable: '桌面 IPC 桥不可用。'
     },
@@ -206,6 +207,7 @@ export const zh: Translations = {
       'session.togglePin': '固定/取消固定当前会话',
       'composer.focus': '聚焦输入框',
       'composer.modelPicker': '打开模型选择器',
+      'composer.voice': '开始 / 停止语音对话',
       'view.toggleSidebar': '切换会话侧边栏',
       'view.toggleRightSidebar': '切换文件浏览器',
       'view.showFiles': '显示文件浏览器',
@@ -364,7 +366,43 @@ export const zh: Translations = {
       installError: '无法安装该主题。',
       installed: name => `已安装「${name}」。`,
       removeTheme: '移除主题',
-      importedBadge: '已导入'
+      importedBadge: '已导入',
+      pet: {
+        title: '宠物',
+        intro:
+          '领养一只悬浮在应用上的 petdex 动画宠物，它会根据 Hermes 的状态做出反应——工具执行时奔跑、成功时欢呼、出错时沮丧。',
+        restartHint: '宠物功能需要重启——当前运行的应用在此功能加入前启动。请退出并重新打开 Hermes，然后回到此处。',
+        scaleTitle: '大小',
+        scaleDesc: '调整悬浮宠物的大小，所有界面即时生效。',
+        on: '开启',
+        off: '关闭',
+        chooseTitle: '选择宠物',
+        chooseDesc: '选择后会自动安装（如需）并设为当前宠物。',
+        searchPlaceholder: '搜索宠物…',
+        unreachable: '无法连接到 petdex 画廊。请检查网络连接并重新打开此页面。',
+        noMatch: query => `没有匹配「${query}」的宠物。`,
+        installedTag: '已安装',
+        generatedTag: '生成',
+        countCapped: (cap, total) => `显示 ${total} 个中的 ${cap} 个——输入关键词以缩小范围。`,
+        count: n => `${n} 个宠物。`,
+        uninstall: name => `卸载 ${name}`,
+        delete: name => `删除 ${name}`,
+        deleteTitle: name => `删除 ${name}？`,
+        deleteBody: '此操作会永久删除宠物，且无法重新安装。',
+        deleteConfirm: '删除',
+        rename: name => `重命名 ${name}`,
+        renameTitle: '重命名宠物',
+        renamePlaceholder: '给宠物起个名字',
+        renameSave: '保存',
+        exportPet: name => `导出 ${name}`,
+        adoptFailed: slug => `无法领养 ${slug}`,
+        uninstallFailed: slug => `无法卸载 ${slug}`,
+        renameFailed: slug => `无法重命名 ${slug}`,
+        exportFailed: slug => `无法导出 ${slug}`,
+        noneAvailable: '当前没有可开启的宠物。',
+        turnOnFailed: '无法开启宠物。',
+        turnOffFailed: '无法关闭宠物。'
+      }
     },
     fieldLabels: defineFieldCopy({
       model: '默认模型',
@@ -912,8 +950,52 @@ export const zh: Translations = {
     commandCenter: '命令中心',
     appearance: '外观',
     settings: '设置',
-    changeTheme: '更改主题...',
+    changeTheme: '更改主题',
     changeColorMode: '更改颜色模式...',
+    pets: {
+      title: '宠物',
+      placeholder: '搜索宠物…',
+      loading: '正在加载 petdex 画廊…',
+      error: '无法连接到 petdex 画廊。',
+      staleBackend: '请重启 Hermes 以使用宠物功能——当前后端版本过旧。',
+      empty: '没有匹配的宠物。',
+      turnOff: '关闭',
+      turnOn: '开启',
+      installed: '已安装',
+      generatedTag: '生成',
+      adoptFailed: '无法领养该宠物。',
+      toggleFailed: '无法切换宠物显示。',
+      noneAvailable: '暂无可用宠物——请在下方选择一个安装。'
+    },
+    generatePet: {
+      title: '生成宠物',
+      placeholder: '描述要生成的宠物……',
+      promptHint: '输入描述，然后按 Enter 生成四种造型。',
+      readyHint: '按 Enter 根据描述生成四种造型。',
+      generate: '生成',
+      generating: '生成中……',
+      retry: '重试',
+      hatch: '孵化',
+      spawning: '召唤中……',
+      hatching: '正在孵化你的宠物……',
+      hatchingSub: '正在注入生命……',
+      hatched: '孵化成功！',
+      hatchRow: (_state, done, total) => `正在绘制画面…… ${done}/${total}`,
+      hatchComposing: '正在拼合……',
+      hatchSaving: '马上就好……',
+      namePlaceholder: '给宠物起个名字',
+      staleBackend: '请更新 Hermes 以生成宠物。',
+      backgroundHint: '你可以关闭此窗口——完成后 Hermes 会通知你。',
+      slowProviderHint: '这可能需要几分钟',
+      remix: '混合生成',
+      remixConfirmTitle: '以此造型混合生成？',
+      remixConfirmBody: '将以此造型为起点生成一组新草图，可能需要几分钟。',
+      genericError: '生成失败——请重试或选择一个建议。',
+      referenceImageTooLarge: '参考图过大。请使用小于 16 MB 的图片。',
+      referenceImageInvalid: '无法读取该参考图。请尝试 PNG、JPG、WebP 或 GIF。',
+      adopt: '领养',
+      startOver: '重新开始'
+    },
     installTheme: {
       title: '安装主题...',
       placeholder: '搜索 VS Code Marketplace...',
@@ -1570,11 +1652,13 @@ export const zh: Translations = {
     manualBody: '你是从命令行安装的 Hermes，因此更新也需要在那里运行。请将此命令粘贴到终端：',
     manualPickedUp: '下次启动 Hermes 时会使用新版本。',
     guiSkewTitle: '请更新桌面应用',
-    guiSkewBody: '后端已更新，但此桌面应用包未更改。请更新或重新安装 Hermes 桌面应用（你的 AppImage / .deb / .rpm）以保持一致。',
+    guiSkewBody:
+      '后端已更新，但此桌面应用包未更改。请更新或重新安装 Hermes 桌面应用（你的 AppImage / .deb / .rpm）以保持一致。',
     copy: '复制',
     copied: '已复制',
     done: '完成',
-    applyingBody: 'Hermes 更新器会在自己的窗口中接管，并在完成后自动重新打开 Hermes。更新期间请不要自行重新打开 Hermes。',
+    applyingBody:
+      'Hermes 更新器会在自己的窗口中接管，并在完成后自动重新打开 Hermes。更新期间请不要自行重新打开 Hermes。',
     applyingBodyBackend: '远程后端正在应用更新并将重启。恢复后 Hermes 会自动重新连接。',
     applyingClose: '此窗口会在更新期间关闭，随后 Hermes 会自动重新打开。',
     errorTitle: '更新未完成',
@@ -1998,7 +2082,55 @@ export const zh: Translations = {
       statusRunning: '运行中',
       statusError: '错误',
       statusRecovered: '已恢复',
-      statusDone: '完成'
+      statusDone: '完成',
+      actions: {
+        read: '已读取',
+        reading: '正在读取',
+        opened: '已打开',
+        opening: '正在打开',
+        searched: '已搜索',
+        searching: '正在搜索',
+        ran: '已运行',
+        running: '正在运行',
+        ranCode: '已运行代码',
+        runningCode: '正在编写脚本'
+      },
+      prefixes: {
+        browser: '浏览器',
+        web: '网页'
+      },
+      titleTemplates: {
+        actionCommand: (action, command) => `${action} ${command}`,
+        actionQuoted: (action, value) => `${action}“${value}”`,
+        actionTarget: (action, target) => `${action} ${target}`,
+        prefixedDone: (prefix, action) => `${prefix}${action}`,
+        runningPrefixedTool: (prefix, action) => `正在运行${prefix}${action}`,
+        runningTool: action => `正在运行 ${action}`
+      },
+      titles: {
+        browser_click: { done: '已点击页面元素', pending: '正在点击页面元素', pendingAction: '正在点击' },
+        browser_fill: { done: '已填写表单字段', pending: '正在填写表单字段', pendingAction: '正在填写' },
+        browser_navigate: { done: '已打开页面', pending: '正在打开页面', pendingAction: '正在打开' },
+        browser_snapshot: { done: '已捕获页面快照', pending: '正在捕获页面快照', pendingAction: '正在捕获' },
+        browser_take_screenshot: { done: '已捕获截图', pending: '正在捕获截图', pendingAction: '正在捕获' },
+        browser_type: { done: '已在页面输入', pending: '正在页面输入', pendingAction: '正在输入' },
+        clarify: { done: '已提问', pending: '正在提问', pendingAction: '正在提问' },
+        cronjob: { done: 'Cron 任务', pending: '正在安排 Cron 任务', pendingAction: '正在安排' },
+        edit_file: { done: '已编辑文件', pending: '正在编辑文件', pendingAction: '正在编辑' },
+        execute_code: { done: '已运行代码', pending: '正在编写脚本', pendingAction: '正在编写脚本' },
+        image_generate: { done: '已生成图片', pending: '正在生成图片', pendingAction: '正在生成' },
+        list_files: { done: '已列出文件', pending: '正在列出文件', pendingAction: '正在列出' },
+        patch: { done: '已修补文件', pending: '正在修补文件', pendingAction: '正在修补' },
+        read_file: { done: '已读取文件', pending: '正在读取文件', pendingAction: '正在读取' },
+        search_files: { done: '已搜索文件', pending: '正在搜索文件', pendingAction: '正在搜索' },
+        session_search_recall: { done: '已搜索会话历史', pending: '正在搜索会话历史', pendingAction: '正在搜索' },
+        terminal: { done: '已运行命令', pending: '正在运行命令', pendingAction: '正在运行' },
+        todo: { done: '已更新待办', pending: '正在更新待办', pendingAction: '正在更新' },
+        vision_analyze: { done: '已分析图片', pending: '正在分析图片', pendingAction: '正在分析' },
+        web_extract: { done: '已读取网页', pending: '正在读取网页', pendingAction: '正在读取' },
+        web_search: { done: '已搜索网页', pending: '正在搜索网页', pendingAction: '正在搜索' },
+        write_file: { done: '已编辑文件', pending: '正在编辑文件', pendingAction: '正在编辑' }
+      }
     }
   },
 

apps/desktop/src/lib/chat-runtime.test.ts

@@ -2,7 +2,7 @@ import { describe, expect, it } from 'vitest'
 
 import type { ComposerAttachment } from '@/store/composer'
 
-import { coerceThinkingText, optimisticAttachmentRef, parseCommandDispatch } from './chat-runtime'
+import { attachmentDisplayText, coerceThinkingText, optimisticAttachmentRef, parseCommandDispatch } from './chat-runtime'
 
 const DATA_URL = 'data:image/png;base64,iVBORw0KGgoAAAANS'
 
@@ -36,6 +36,32 @@ describe('optimisticAttachmentRef', () => {
       '@file:src/a.ts'
     )
   })
+
+  // Session switches / draft restores can leave undefined|null holes in the
+  // composer attachments array. AttachmentList already filters them (#49624),
+  // but the submit path maps the same array through these helpers — an unguarded
+  // hole threw "Cannot read properties of undefined (reading 'refText')",
+  // crashing the chat surface (blank pane). The helpers must no-op on holes.
+  it('returns null for an undefined attachment instead of throwing', () => {
+    expect(() => optimisticAttachmentRef(undefined as unknown as ComposerAttachment)).not.toThrow()
+    expect(optimisticAttachmentRef(undefined as unknown as ComposerAttachment)).toBeNull()
+  })
+
+  it('returns null for a null attachment instead of throwing', () => {
+    expect(optimisticAttachmentRef(null as unknown as ComposerAttachment)).toBeNull()
+  })
+})
+
+describe('attachmentDisplayText', () => {
+  it('returns null for undefined|null instead of reading .kind/.refText on a hole', () => {
+    expect(() => attachmentDisplayText(undefined as unknown as ComposerAttachment)).not.toThrow()
+    expect(attachmentDisplayText(undefined as unknown as ComposerAttachment)).toBeNull()
+    expect(attachmentDisplayText(null as unknown as ComposerAttachment)).toBeNull()
+  })
+
+  it('still resolves a normal file ref', () => {
+    expect(attachmentDisplayText(attachment({ kind: 'file', refText: '@file:src/a.ts' }))).toBe('@file:src/a.ts')
+  })
 })
 
 describe('coerceThinkingText', () => {

apps/desktop/src/lib/chat-runtime.ts

@@ -155,6 +155,13 @@ export function pathLabel(path: string): string {
 }
 
 export function attachmentDisplayText(attachment: ComposerAttachment): string | null {
+  // Session switches / draft restores can leave undefined holes in the
+  // composer attachments array (see AttachmentList's filter(Boolean) + #49624).
+  // Every consumer funnels through here, so guard the chokepoint too.
+  if (!attachment) {
+    return null
+  }
+
   if (attachment.kind === 'terminal' && attachment.detail) {
     return `\`\`\`terminal\n${attachment.detail.trim()}\n\`\`\``
   }
@@ -188,6 +195,10 @@ export function attachmentDisplayText(attachment: ComposerAttachment): string |
  * through to `attachmentDisplayText`.
  */
 export function optimisticAttachmentRef(attachment: ComposerAttachment): string | null {
+  if (!attachment) {
+    return null
+  }
+
   if (attachment.kind === 'image' && attachment.previewUrl?.startsWith('data:')) {
     return attachment.previewUrl
   }

apps/desktop/src/lib/desktop-slash-commands.test.ts

@@ -52,6 +52,16 @@ describe('desktop slash command curation', () => {
     expect(desktopSlashUnavailableMessage('/personality')).toBeNull()
   })
 
+  it('routes /pet through the desktop action handler and drops /pets', () => {
+    expect(resolveDesktopCommand('/pet')?.surface).toEqual({ kind: 'action', action: 'pet' })
+    expect(resolveDesktopCommand('/pet')?.args).toBe(true)
+    expect(isDesktopSlashSuggestion('/pet')).toBe(true)
+    expect(isDesktopSlashCommand('/pet')).toBe(true)
+    expect(resolveDesktopCommand('/pets')?.surface).toEqual({ kind: 'unavailable', reason: 'settings' })
+    expect(isDesktopSlashSuggestion('/pets')).toBe(false)
+    expect(isDesktopSlashCommand('/pets')).toBe(false)
+  })
+
   it('treats /browser as an executable action command (local-gateway connect)', () => {
     // /browser used to be terminal-only; it now resolves to a desktop action
     // handler that routes browser.manage RPC when the gateway is local.