fix: proactive Codex CLI sync before refresh + retry on failure

OpenAI OAuth refresh tokens are single-use and rotate on every refresh.
When the Codex CLI (or another Hermes profile) refreshes its token,
the pool entry's refresh_token becomes stale. Previously, the sync from
~/.codex/auth.json only ran for EXHAUSTED entries in _available_entries().

Now:
1. _refresh_entry() proactively syncs from ~/.codex/auth.json BEFORE
   attempting the OAuth refresh, picking up tokens refreshed by the
   Codex CLI or VS Code extension.
2. On refresh failure, re-syncs and retries once (mirrors the existing
   Anthropic retry pattern), handling the race where the CLI refreshes
   between the proactive sync and the actual refresh call.
3. If the synced entry has a valid (non-expired) token, uses it
   directly without an unnecessary refresh round-trip.

.env.example

@@ -81,6 +81,14 @@
 # HF_TOKEN=
 # OPENCODE_GO_BASE_URL=https://opencode.ai/zen/go/v1  # Override default base URL
 
+# =============================================================================
+# LLM PROVIDER (Qwen OAuth)
+# =============================================================================
+# Qwen OAuth reuses your local Qwen CLI login (qwen auth qwen-oauth).
+# No API key needed — credentials come from ~/.qwen/oauth_creds.json.
+# Optional base URL override:
+# HERMES_QWEN_BASE_URL=https://portal.qwen.ai/v1
+
 # =============================================================================
 # TOOL API KEYS
 # =============================================================================

.github/workflows/docker-publish.yml

@@ -8,6 +8,9 @@ on:
   release:
     types: [published]
 
+permissions:
+  contents: read
+
 concurrency:
   group: docker-${{ github.ref }}
   cancel-in-progress: true
@@ -17,22 +20,29 @@ jobs:
     # Only run on the upstream repository, not on forks
     if: github.repository == 'NousResearch/hermes-agent'
     runs-on: ubuntu-latest
-    timeout-minutes: 30
+    timeout-minutes: 60
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
         with:
           submodules: recursive
 
+      - name: Set up QEMU
+        uses: docker/setup-qemu-action@v3
+
       - name: Set up Docker Buildx
         uses: docker/setup-buildx-action@v3
 
-      - name: Build image
+      # Build amd64 only so we can `load` the image for smoke testing.
+      # `load: true` cannot export a multi-arch manifest to the local daemon.
+      # The multi-arch build follows on push to main / release.
+      - name: Build image (amd64, smoke test)
         uses: docker/build-push-action@v6
         with:
           context: .
           file: Dockerfile
           load: true
+          platforms: linux/amd64
           tags: nousresearch/hermes-agent:test
           cache-from: type=gha
           cache-to: type=gha,mode=max
@@ -51,26 +61,28 @@ jobs:
           username: ${{ secrets.DOCKERHUB_USERNAME }}
           password: ${{ secrets.DOCKERHUB_TOKEN }}
 
-      - name: Push image (main branch)
+      - name: Push multi-arch image (main branch)
         if: github.event_name == 'push' && github.ref == 'refs/heads/main'
         uses: docker/build-push-action@v6
         with:
           context: .
           file: Dockerfile
           push: true
+          platforms: linux/amd64,linux/arm64
           tags: |
             nousresearch/hermes-agent:latest
             nousresearch/hermes-agent:${{ github.sha }}
           cache-from: type=gha
           cache-to: type=gha,mode=max
 
-      - name: Push image (release)
+      - name: Push multi-arch image (release)
         if: github.event_name == 'release'
         uses: docker/build-push-action@v6
         with:
           context: .
           file: Dockerfile
           push: true
+          platforms: linux/amd64,linux/arm64
           tags: |
             nousresearch/hermes-agent:latest
             nousresearch/hermes-agent:${{ github.event.release.tag_name }}

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

@@ -27,8 +27,8 @@ jobs:
         with:
           python-version: '3.11'
 
-      - name: Install Python dependencies
-        run: python -m pip install ascii-guard pyyaml
+      - name: Install ascii-guard
+        run: python -m pip install ascii-guard==2.3.0 pyyaml==6.0.3
 
       - name: Extract skill metadata for dashboard
         run: python3 website/scripts/extract-skills.py

.github/workflows/nix.yml

@@ -27,8 +27,8 @@ jobs:
     timeout-minutes: 30
     steps:
       - uses: actions/checkout@v4
-      - uses: DeterminateSystems/nix-installer-action@main
-      - uses: DeterminateSystems/magic-nix-cache-action@main
+      - uses: DeterminateSystems/nix-installer-action@ef8a148080ab6020fd15196c2084a2eea5ff2d25  # v22
+      - uses: DeterminateSystems/magic-nix-cache-action@565684385bcd71bad329742eefe8d12f2e765b39  # v13
       - name: Check flake
         if: runner.os == 'Linux'
         run: nix flake check --print-build-logs

Dockerfile

@@ -1,5 +1,8 @@
 FROM debian:13.4
 
+# Disable Python stdout buffering to ensure logs are printed immediately
+ENV PYTHONUNBUFFERED=1
+
 # Install system dependencies in one layer, clear APT cache
 RUN apt-get update && \
     apt-get install -y --no-install-recommends \

RELEASE_v0.8.0.md

@@ -2,18 +2,22 @@
 
 **Release Date:** April 8, 2026
 
-> The intelligence release — native Google AI Studio provider, live model switching across all platforms, self-optimized GPT/Codex guidance, smart inactivity timeouts, approval buttons, interactive model pickers, MCP OAuth 2.1, and 209 merged PRs with 82 resolved issues.
+> The intelligence release — background task auto-notifications, free MiMo v2 Pro on Nous Portal, live model switching across all platforms, self-optimized GPT/Codex guidance, native Google AI Studio, smart inactivity timeouts, approval buttons, MCP OAuth 2.1, and 209 merged PRs with 82 resolved issues.
 
 ---
 
 ## ✨ Highlights
 
-- **Google AI Studio (Gemini) Native Provider** — Direct access to Gemini models through Google's AI Studio API. Includes automatic models.dev registry integration for real-time context length detection across any provider. ([#5577](https://github.com/NousResearch/hermes-agent/pull/5577))
+- **Background Process Auto-Notifications (`notify_on_complete`)** — Background tasks can now automatically notify the agent when they finish. Start a long-running process (AI model training, test suites, deployments, builds) and the agent gets notified on completion — no polling needed. The agent can keep working on other things and pick up results when they land. ([#5779](https://github.com/NousResearch/hermes-agent/pull/5779))
+
+- **Free Xiaomi MiMo v2 Pro on Nous Portal** — Nous Portal now supports the free-tier Xiaomi MiMo v2 Pro model for auxiliary tasks (compression, vision, summarization), with free-tier model gating and pricing display in model selection. ([#6018](https://github.com/NousResearch/hermes-agent/pull/6018), [#5880](https://github.com/NousResearch/hermes-agent/pull/5880))
 
 - **Live Model Switching (`/model` Command)** — Switch models and providers mid-session from CLI, Telegram, Discord, Slack, or any gateway platform. Aggregator-aware resolution keeps you on OpenRouter/Nous when possible, with automatic cross-provider fallback when needed. Interactive model pickers on Telegram and Discord with inline buttons. ([#5181](https://github.com/NousResearch/hermes-agent/pull/5181), [#5742](https://github.com/NousResearch/hermes-agent/pull/5742))
 
 - **Self-Optimized GPT/Codex Tool-Use Guidance** — The agent diagnosed and patched 5 failure modes in GPT and Codex tool calling through automated behavioral benchmarking, dramatically improving reliability on OpenAI models. Includes execution discipline guidance and thinking-only prefill continuation for structured reasoning. ([#6120](https://github.com/NousResearch/hermes-agent/pull/6120), [#5414](https://github.com/NousResearch/hermes-agent/pull/5414), [#5931](https://github.com/NousResearch/hermes-agent/pull/5931))
 
+- **Google AI Studio (Gemini) Native Provider** — Direct access to Gemini models through Google's AI Studio API. Includes automatic models.dev registry integration for real-time context length detection across any provider. ([#5577](https://github.com/NousResearch/hermes-agent/pull/5577))
+
 - **Inactivity-Based Agent Timeouts** — Gateway and cron timeouts now track actual tool activity instead of wall-clock time. Long-running tasks that are actively working will never be killed — only truly idle agents time out. ([#5389](https://github.com/NousResearch/hermes-agent/pull/5389), [#5440](https://github.com/NousResearch/hermes-agent/pull/5440))
 
 - **Approval Buttons on Slack & Telegram** — Dangerous command approval via native platform buttons instead of typing `/approve`. Slack gets thread context preservation; Telegram gets emoji reactions for approval status. ([#5890](https://github.com/NousResearch/hermes-agent/pull/5890), [#5975](https://github.com/NousResearch/hermes-agent/pull/5975))

agent/anthropic_adapter.py

@@ -163,6 +163,17 @@ def _is_oauth_token(key: str) -> bool:
     return True
 
 
+def _normalize_base_url_text(base_url) -> str:
+    """Normalize SDK/base transport URL values to a plain string for inspection.
+
+    Some client objects expose ``base_url`` as an ``httpx.URL`` instead of a raw
+    string.  Provider/auth detection should accept either shape.
+    """
+    if not base_url:
+        return ""
+    return str(base_url).strip()
+
+
 def _is_third_party_anthropic_endpoint(base_url: str | None) -> bool:
     """Return True for non-Anthropic endpoints using the Anthropic Messages API.
 
@@ -170,9 +181,10 @@ def _is_third_party_anthropic_endpoint(base_url: str | None) -> bool:
     with their own API keys via x-api-key, not Anthropic OAuth tokens. OAuth
     detection should be skipped for these endpoints.
     """
-    if not base_url:
+    normalized = _normalize_base_url_text(base_url)
+    if not normalized:
         return False  # No base_url = direct Anthropic API
-    normalized = base_url.rstrip("/").lower()
+    normalized = normalized.rstrip("/").lower()
     if "anthropic.com" in normalized:
         return False  # Direct Anthropic API — OAuth applies
     return True  # Any other endpoint is a third-party proxy
@@ -182,12 +194,13 @@ def _requires_bearer_auth(base_url: str | None) -> bool:
     """Return True for Anthropic-compatible providers that require Bearer auth.
 
     Some third-party /anthropic endpoints implement Anthropic's Messages API but
-    require Authorization: Bearer instead of Anthropic's native x-api-key header.
+    require Authorization: Bearer *** of Anthropic's native x-api-key header.
     MiniMax's global and China Anthropic-compatible endpoints follow this pattern.
     """
-    if not base_url:
+    normalized = _normalize_base_url_text(base_url)
+    if not normalized:
         return False
-    normalized = base_url.rstrip("/").lower()
+    normalized = normalized.rstrip("/").lower()
     return normalized.startswith(("https://api.minimax.io/anthropic", "https://api.minimaxi.com/anthropic"))
 
 
@@ -203,13 +216,14 @@ def build_anthropic_client(api_key: str, base_url: str = None):
         )
     from httpx import Timeout
 
+    normalized_base_url = _normalize_base_url_text(base_url)
     kwargs = {
         "timeout": Timeout(timeout=900.0, connect=10.0),
     }
-    if base_url:
-        kwargs["base_url"] = base_url
+    if normalized_base_url:
+        kwargs["base_url"] = normalized_base_url
 
-    if _requires_bearer_auth(base_url):
+    if _requires_bearer_auth(normalized_base_url):
         # Some Anthropic-compatible providers (e.g. MiniMax) expect the API key in
         # Authorization: Bearer even for regular API keys. Route those endpoints
         # through auth_token so the SDK sends Bearer auth instead of x-api-key.
@@ -942,12 +956,18 @@ def _convert_content_to_anthropic(content: Any) -> Any:
 
 def convert_messages_to_anthropic(
     messages: List[Dict],
+    base_url: str | None = None,
 ) -> Tuple[Optional[Any], List[Dict]]:
     """Convert OpenAI-format messages to Anthropic format.
 
     Returns (system_prompt, anthropic_messages).
     System messages are extracted since Anthropic takes them as a separate param.
     system_prompt is a string or list of content blocks (when cache_control present).
+
+    When *base_url* is provided and points to a third-party Anthropic-compatible
+    endpoint, all thinking block signatures are stripped.  Signatures are
+    Anthropic-proprietary — third-party endpoints cannot validate them and will
+    reject them with HTTP 400 "Invalid signature in thinking block".
     """
     system = None
     result = []
@@ -1134,7 +1154,14 @@ def convert_messages_to_anthropic(
     # orphan stripping, message merging) invalidates the signature,
     # causing HTTP 400 "Invalid signature in thinking block".
     #
-    # Strategy (following clawdbot/OpenClaw pattern):
+    # Signatures are Anthropic-proprietary.  Third-party endpoints
+    # (MiniMax, Azure AI Foundry, self-hosted proxies) cannot validate
+    # them and will reject them outright.  When targeting a third-party
+    # endpoint, strip ALL thinking/redacted_thinking blocks from every
+    # assistant message — the third-party will generate its own
+    # thinking blocks if it supports extended thinking.
+    #
+    # For direct Anthropic (strategy following clawdbot/OpenClaw):
     # 1. Strip thinking/redacted_thinking from all assistant messages
     #    EXCEPT the last one — preserves reasoning continuity on the
     #    current tool-use chain while avoiding stale signature errors.
@@ -1143,6 +1170,7 @@ def convert_messages_to_anthropic(
     # 3. Strip cache_control from thinking/redacted_thinking blocks —
     #    cache markers can interfere with signature validation.
     _THINKING_TYPES = frozenset(("thinking", "redacted_thinking"))
+    _is_third_party = _is_third_party_anthropic_endpoint(base_url)
 
     last_assistant_idx = None
     for i in range(len(result) - 1, -1, -1):
@@ -1154,16 +1182,19 @@ def convert_messages_to_anthropic(
         if m.get("role") != "assistant" or not isinstance(m.get("content"), list):
             continue
 
-        if idx != last_assistant_idx:
-            # Strip ALL thinking blocks from non-latest assistant messages
+        if _is_third_party or idx != last_assistant_idx:
+            # Third-party endpoint: strip ALL thinking blocks from every
+            # assistant message — signatures are Anthropic-proprietary.
+            # Direct Anthropic: strip from non-latest assistant messages only.
             stripped = [
                 b for b in m["content"]
                 if not (isinstance(b, dict) and b.get("type") in _THINKING_TYPES)
             ]
             m["content"] = stripped or [{"type": "text", "text": "(thinking elided)"}]
         else:
-            # Latest assistant: keep signed thinking blocks for reasoning
-            # continuity; downgrade unsigned ones to plain text.
+            # Latest assistant on direct Anthropic: keep signed thinking
+            # blocks for reasoning continuity; downgrade unsigned ones to
+            # plain text.
             new_content = []
             for b in m["content"]:
                 if not isinstance(b, dict) or b.get("type") not in _THINKING_TYPES:
@@ -1203,28 +1234,53 @@ def build_anthropic_kwargs(
     is_oauth: bool = False,
     preserve_dots: bool = False,
     context_length: Optional[int] = None,
+    base_url: str | None = None,
 ) -> Dict[str, Any]:
     """Build kwargs for anthropic.messages.create().
 
-    When *max_tokens* is None, the model's native output limit is used
-    (e.g. 128K for Opus 4.6, 64K for Sonnet 4.6).  If *context_length*
-    is provided, the effective limit is clamped so it doesn't exceed
-    the context window.
+    Naming note — two distinct concepts, easily confused:
+      max_tokens     = OUTPUT token cap for a single response.
+                       Anthropic's API calls this "max_tokens" but it only
+                       limits the *output*.  Anthropic's own native SDK
+                       renamed it "max_output_tokens" for clarity.
+      context_length = TOTAL context window (input tokens + output tokens).
+                       The API enforces: input_tokens + max_tokens ≤ context_length.
+                       Stored on the ContextCompressor; reduced on overflow errors.
+
+    When *max_tokens* is None the model's native output ceiling is used
+    (e.g. 128K for Opus 4.6, 64K for Sonnet 4.6).
+
+    When *context_length* is provided and the model's native output ceiling
+    exceeds it (e.g. a local endpoint with an 8K window), the output cap is
+    clamped to context_length − 1.  This only kicks in for unusually small
+    context windows; for full-size models the native output cap is always
+    smaller than the context window so no clamping happens.
+    NOTE: this clamping does not account for prompt size — if the prompt is
+    large, Anthropic may still reject the request.  The caller must detect
+    "max_tokens too large given prompt" errors and retry with a smaller cap
+    (see parse_available_output_tokens_from_error + _ephemeral_max_output_tokens).
 
     When *is_oauth* is True, applies Claude Code compatibility transforms:
     system prompt prefix, tool name prefixing, and prompt sanitization.
 
     When *preserve_dots* is True, model name dots are not converted to hyphens
     (for Alibaba/DashScope anthropic-compatible endpoints: qwen3.5-plus).
+
+    When *base_url* points to a third-party Anthropic-compatible endpoint,
+    thinking block signatures are stripped (they are Anthropic-proprietary).
     """
-    system, anthropic_messages = convert_messages_to_anthropic(messages)
+    system, anthropic_messages = convert_messages_to_anthropic(messages, base_url=base_url)
     anthropic_tools = convert_tools_to_anthropic(tools) if tools else []
 
     model = normalize_model_name(model, preserve_dots=preserve_dots)
+    # effective_max_tokens = output cap for this call (≠ total context window)
     effective_max_tokens = max_tokens or _get_anthropic_max_output(model)
 
-    # Clamp to context window if the user set a lower context_length
-    # (e.g. custom endpoint with limited capacity).
+    # Clamp output cap to fit inside the total context window.
+    # Only matters for small custom endpoints where context_length < native
+    # output ceiling.  For standard Anthropic models context_length (e.g.
+    # 200K) is always larger than the output ceiling (e.g. 128K), so this
+    # branch is not taken.
     if context_length and effective_max_tokens > context_length:
         effective_max_tokens = max(context_length - 1, 1)
 

agent/auxiliary_client.py

@@ -629,11 +629,19 @@ def _nous_base_url() -> str:
 
 
 def _read_codex_access_token() -> Optional[str]:
-    """Read a valid, non-expired Codex OAuth access token from Hermes auth store."""
+    """Read a valid, non-expired Codex OAuth access token from Hermes auth store.
+
+    If a credential pool exists but currently has no selectable runtime entry
+    (for example all pool slots are marked exhausted), fall back to the
+    profile's auth.json token instead of hard-failing. This keeps explicit
+    fallback-to-Codex working when the pool state is stale but the stored OAuth
+    token is still valid.
+    """
     pool_present, entry = _select_pool_entry("openai-codex")
     if pool_present:
         token = _pool_runtime_api_key(entry)
-        return token or None
+        if token:
+            return token
 
     try:
         from hermes_cli.auth import _read_codex_tokens
@@ -694,7 +702,7 @@ def _resolve_api_key_provider() -> Tuple[Optional[OpenAI], Optional[str]]:
             logger.debug("Auxiliary text client: %s (%s) via pool", pconfig.name, model)
             extra = {}
             if "api.kimi.com" in base_url.lower():
-                extra["default_headers"] = {"User-Agent": "KimiCLI/1.0"}
+                extra["default_headers"] = {"User-Agent": "KimiCLI/1.3"}
             elif "api.githubcopilot.com" in base_url.lower():
                 from hermes_cli.models import copilot_default_headers
 
@@ -713,7 +721,7 @@ def _resolve_api_key_provider() -> Tuple[Optional[OpenAI], Optional[str]]:
         logger.debug("Auxiliary text client: %s (%s)", pconfig.name, model)
         extra = {}
         if "api.kimi.com" in base_url.lower():
-            extra["default_headers"] = {"User-Agent": "KimiCLI/1.0"}
+            extra["default_headers"] = {"User-Agent": "KimiCLI/1.3"}
         elif "api.githubcopilot.com" in base_url.lower():
             from hermes_cli.models import copilot_default_headers
 
@@ -834,7 +842,7 @@ def _read_main_provider() -> str:
         if isinstance(model_cfg, dict):
             provider = model_cfg.get("provider", "")
             if isinstance(provider, str) and provider.strip():
-                return _normalize_aux_provider(provider)
+                return provider.strip().lower()
     except Exception:
         pass
     return ""
@@ -894,9 +902,13 @@ def _try_codex() -> Tuple[Optional[Any], Optional[str]]:
     pool_present, entry = _select_pool_entry("openai-codex")
     if pool_present:
         codex_token = _pool_runtime_api_key(entry)
-        if not codex_token:
-            return None, None
-        base_url = _pool_runtime_base_url(entry, _CODEX_AUX_BASE_URL) or _CODEX_AUX_BASE_URL
+        if codex_token:
+            base_url = _pool_runtime_base_url(entry, _CODEX_AUX_BASE_URL) or _CODEX_AUX_BASE_URL
+        else:
+            codex_token = _read_codex_access_token()
+            if not codex_token:
+                return None, None
+            base_url = _CODEX_AUX_BASE_URL
     else:
         codex_token = _read_codex_access_token()
         if not codex_token:
@@ -1035,6 +1047,32 @@ def _is_payment_error(exc: Exception) -> bool:
     return False
 
 
+def _is_connection_error(exc: Exception) -> bool:
+    """Detect connection/network errors that warrant provider fallback.
+
+    Returns True for errors indicating the provider endpoint is unreachable
+    (DNS failure, connection refused, TLS errors, timeouts).  These are
+    distinct from API errors (4xx/5xx) which indicate the provider IS
+    reachable but returned an error.
+    """
+    from openai import APIConnectionError, APITimeoutError
+
+    if isinstance(exc, (APIConnectionError, APITimeoutError)):
+        return True
+    # urllib3 / httpx / httpcore connection errors
+    err_type = type(exc).__name__
+    if any(kw in err_type for kw in ("Connection", "Timeout", "DNS", "SSL")):
+        return True
+    err_lower = str(exc).lower()
+    if any(kw in err_lower for kw in (
+        "connection refused", "name or service not known",
+        "no route to host", "network is unreachable",
+        "timed out", "connection reset",
+    )):
+        return True
+    return False
+
+
 def _try_payment_fallback(
     failed_provider: str,
     task: str = None,
@@ -1099,7 +1137,7 @@ def _resolve_auto() -> Tuple[Optional[OpenAI], Optional[str]]:
     main_model = _read_main_model()
     if (main_provider and main_model
             and main_provider not in _AGGREGATOR_PROVIDERS
-            and main_provider not in ("auto", "custom", "")):
+            and main_provider not in ("auto", "")):
         client, resolved = resolve_provider_client(main_provider, main_model)
         if client is not None:
             logger.info("Auxiliary auto-detect: using main provider %s (%s)",
@@ -1157,7 +1195,7 @@ def _to_async_client(sync_client, model: str):
 
         async_kwargs["default_headers"] = copilot_default_headers()
     elif "api.kimi.com" in base_lower:
-        async_kwargs["default_headers"] = {"User-Agent": "KimiCLI/1.0"}
+        async_kwargs["default_headers"] = {"User-Agent": "KimiCLI/1.3"}
     return AsyncOpenAI(**async_kwargs), model
 
 
@@ -1277,7 +1315,13 @@ def resolve_provider_client(
                 )
                 return None, None
             final_model = model or _read_main_model() or "gpt-4o-mini"
-            client = OpenAI(api_key=custom_key, base_url=custom_base)
+            extra = {}
+            if "api.kimi.com" in custom_base.lower():
+                extra["default_headers"] = {"User-Agent": "KimiCLI/1.3"}
+            elif "api.githubcopilot.com" in custom_base.lower():
+                from hermes_cli.models import copilot_default_headers
+                extra["default_headers"] = copilot_default_headers()
+            client = OpenAI(api_key=custom_key, base_url=custom_base, **extra)
             return (_to_async_client(client, final_model) if async_mode
                     else (client, final_model))
         # Try custom first, then codex, then API-key providers
@@ -1356,7 +1400,7 @@ def resolve_provider_client(
         # Provider-specific headers
         headers = {}
         if "api.kimi.com" in base_url.lower():
-            headers["User-Agent"] = "KimiCLI/1.0"
+            headers["User-Agent"] = "KimiCLI/1.3"
         elif "api.githubcopilot.com" in base_url.lower():
             from hermes_cli.models import copilot_default_headers
 
@@ -1470,19 +1514,25 @@ def _preferred_main_vision_provider() -> Optional[str]:
 def get_available_vision_backends() -> List[str]:
     """Return the currently available vision backends in auto-selection order.
 
-    Order: OpenRouter → Nous → active provider.  This is the single source
-    of truth for setup, tool gating, and runtime auto-routing of vision tasks.
+    Order: active provider → OpenRouter → Nous → stop.  This is the single
+    source of truth for setup, tool gating, and runtime auto-routing of
+    vision tasks.
     """
-    available = [p for p in _VISION_AUTO_PROVIDER_ORDER
-                 if _strict_vision_backend_available(p)]
-    # Also check the user's active provider (may be DeepSeek, Alibaba, named
-    # custom, etc.) — resolve_provider_client handles all provider types.
+    available: List[str] = []
+    # 1. Active provider — if the user configured a provider, try it first.
     main_provider = _read_main_provider()
-    if (main_provider and main_provider not in ("auto", "")
-            and main_provider not in available):
-        client, _ = resolve_provider_client(main_provider, _read_main_model())
-        if client is not None:
-            available.append(main_provider)
+    if main_provider and main_provider not in ("auto", ""):
+        if main_provider in _VISION_AUTO_PROVIDER_ORDER:
+            if _strict_vision_backend_available(main_provider):
+                available.append(main_provider)
+        else:
+            client, _ = resolve_provider_client(main_provider, _read_main_model())
+            if client is not None:
+                available.append(main_provider)
+    # 2. OpenRouter, 3. Nous — skip if already covered by main provider.
+    for p in _VISION_AUTO_PROVIDER_ORDER:
+        if p not in available and _strict_vision_backend_available(p):
+            available.append(p)
     return available
 
 
@@ -1529,28 +1579,37 @@ def resolve_vision_provider_client(
 
     if requested == "auto":
         # Vision auto-detection order:
-        #   1. OpenRouter  (known vision-capable default model)
-        #   2. Nous Portal (known vision-capable default model)
-        #   3. Active provider + model (user's main chat config)
+        #   1. Active provider + model (user's main chat config)
+        #   2. OpenRouter  (known vision-capable default model)
+        #   3. Nous Portal (known vision-capable default model)
         #   4. Stop
-        for candidate in _VISION_AUTO_PROVIDER_ORDER:
-            sync_client, default_model = _resolve_strict_vision_backend(candidate)
-            if sync_client is not None:
-                return _finalize(candidate, sync_client, default_model)
-
-        # Fall back to the user's active provider + model.
         main_provider = _read_main_provider()
         main_model = _read_main_model()
         if main_provider and main_provider not in ("auto", ""):
-            sync_client, resolved_model = resolve_provider_client(
-                main_provider, main_model)
+            if main_provider in _VISION_AUTO_PROVIDER_ORDER:
+                # Known strict backend — use its defaults.
+                sync_client, default_model = _resolve_strict_vision_backend(main_provider)
+                if sync_client is not None:
+                    return _finalize(main_provider, sync_client, default_model)
+            else:
+                # Exotic provider (DeepSeek, Alibaba, named custom, etc.)
+                rpc_client, rpc_model = resolve_provider_client(
+                    main_provider, main_model)
+                if rpc_client is not None:
+                    logger.info(
+                        "Vision auto-detect: using active provider %s (%s)",
+                        main_provider, rpc_model or main_model,
+                    )
+                    return _finalize(
+                        main_provider, rpc_client, rpc_model or main_model)
+
+        # Fall back through aggregators.
+        for candidate in _VISION_AUTO_PROVIDER_ORDER:
+            if candidate == main_provider:
+                continue  # already tried above
+            sync_client, default_model = _resolve_strict_vision_backend(candidate)
             if sync_client is not None:
-                logger.info(
-                    "Vision auto-detect: using active provider %s (%s)",
-                    main_provider, resolved_model or main_model,
-                )
-                return _finalize(
-                    main_provider, sync_client, resolved_model or main_model)
+                return _finalize(candidate, sync_client, default_model)
 
         logger.debug("Auxiliary vision client: none available")
         return None, None, None
@@ -2066,7 +2125,18 @@ def call_llm(
         # try alternative providers instead of giving up.  This handles the
         # common case where a user runs out of OpenRouter credits but has
         # Codex OAuth or another provider available.
-        if _is_payment_error(first_err):
+        #
+        # ── Connection error fallback ────────────────────────────────
+        # When a provider endpoint is unreachable (DNS failure, connection
+        # refused, timeout), try alternative providers.  This handles stale
+        # Codex/OAuth tokens that authenticate but whose endpoint is down,
+        # and providers the user never configured that got picked up by
+        # the auto-detection chain.
+        should_fallback = _is_payment_error(first_err) or _is_connection_error(first_err)
+        if should_fallback:
+            reason = "payment error" if _is_payment_error(first_err) else "connection error"
+            logger.info("Auxiliary %s: %s on %s (%s), trying fallback",
+                        task or "call", reason, resolved_provider, first_err)
             fb_client, fb_model, fb_label = _try_payment_fallback(
                 resolved_provider, task)
             if fb_client is not None:

agent/context_compressor.py

@@ -154,12 +154,15 @@ class ContextCompressor:
 
     def _prune_old_tool_results(
         self, messages: List[Dict[str, Any]], protect_tail_count: int,
+        protect_tail_tokens: int | None = None,
     ) -> tuple[List[Dict[str, Any]], int]:
         """Replace old tool result contents with a short placeholder.
 
-        Walks backward from the end, protecting the most recent
-        ``protect_tail_count`` messages. Older tool results get their
-        content replaced with a placeholder string.
+        Walks backward from the end, protecting the most recent messages that
+        fall within ``protect_tail_tokens`` (when provided) OR the last
+        ``protect_tail_count`` messages (backward-compatible default).
+        When both are given, the token budget takes priority and the message
+        count acts as a hard minimum floor.
 
         Returns (pruned_messages, pruned_count).
         """
@@ -168,7 +171,29 @@ class ContextCompressor:
 
         result = [m.copy() for m in messages]
         pruned = 0
-        prune_boundary = len(result) - protect_tail_count
+
+        # Determine the prune boundary
+        if protect_tail_tokens is not None and protect_tail_tokens > 0:
+            # Token-budget approach: walk backward accumulating tokens
+            accumulated = 0
+            boundary = len(result)
+            min_protect = min(protect_tail_count, len(result) - 1)
+            for i in range(len(result) - 1, -1, -1):
+                msg = result[i]
+                content_len = len(msg.get("content") or "")
+                msg_tokens = content_len // _CHARS_PER_TOKEN + 10
+                for tc in msg.get("tool_calls") or []:
+                    if isinstance(tc, dict):
+                        args = tc.get("function", {}).get("arguments", "")
+                        msg_tokens += len(args) // _CHARS_PER_TOKEN
+                if accumulated + msg_tokens > protect_tail_tokens and (len(result) - i) >= min_protect:
+                    boundary = i
+                    break
+                accumulated += msg_tokens
+                boundary = i
+            prune_boundary = max(boundary, len(result) - min_protect)
+        else:
+            prune_boundary = len(result) - protect_tail_count
 
         for i in range(prune_boundary):
             msg = result[i]
@@ -199,30 +224,39 @@ class ContextCompressor:
         budget = int(content_tokens * _SUMMARY_RATIO)
         return max(_MIN_SUMMARY_TOKENS, min(budget, self.max_summary_tokens))
 
+    # Truncation limits for the summarizer input.  These bound how much of
+    # each message the summary model sees — the budget is the *summary*
+    # model's context window, not the main model's.
+    _CONTENT_MAX = 6000       # total chars per message body
+    _CONTENT_HEAD = 4000      # chars kept from the start
+    _CONTENT_TAIL = 1500      # chars kept from the end
+    _TOOL_ARGS_MAX = 1500     # tool call argument chars
+    _TOOL_ARGS_HEAD = 1200    # kept from the start of tool args
+
     def _serialize_for_summary(self, turns: List[Dict[str, Any]]) -> str:
         """Serialize conversation turns into labeled text for the summarizer.
 
-        Includes tool call arguments and result content (up to 3000 chars
-        per message) so the summarizer can preserve specific details like
-        file paths, commands, and outputs.
+        Includes tool call arguments and result content (up to
+        ``_CONTENT_MAX`` chars per message) so the summarizer can preserve
+        specific details like file paths, commands, and outputs.
         """
         parts = []
         for msg in turns:
             role = msg.get("role", "unknown")
             content = msg.get("content") or ""
 
-            # Tool results: keep more content than before (3000 chars)
+            # Tool results: keep enough content for the summarizer
             if role == "tool":
                 tool_id = msg.get("tool_call_id", "")
-                if len(content) > 3000:
-                    content = content[:2000] + "\n...[truncated]...\n" + content[-800:]
+                if len(content) > self._CONTENT_MAX:
+                    content = content[:self._CONTENT_HEAD] + "\n...[truncated]...\n" + content[-self._CONTENT_TAIL:]
                 parts.append(f"[TOOL RESULT {tool_id}]: {content}")
                 continue
 
             # Assistant messages: include tool call names AND arguments
             if role == "assistant":
-                if len(content) > 3000:
-                    content = content[:2000] + "\n...[truncated]...\n" + content[-800:]
+                if len(content) > self._CONTENT_MAX:
+                    content = content[:self._CONTENT_HEAD] + "\n...[truncated]...\n" + content[-self._CONTENT_TAIL:]
                 tool_calls = msg.get("tool_calls", [])
                 if tool_calls:
                     tc_parts = []
@@ -232,8 +266,8 @@ class ContextCompressor:
                             name = fn.get("name", "?")
                             args = fn.get("arguments", "")
                             # Truncate long arguments but keep enough for context
-                            if len(args) > 500:
-                                args = args[:400] + "..."
+                            if len(args) > self._TOOL_ARGS_MAX:
+                                args = args[:self._TOOL_ARGS_HEAD] + "..."
                             tc_parts.append(f"  {name}({args})")
                         else:
                             fn = getattr(tc, "function", None)
@@ -244,8 +278,8 @@ class ContextCompressor:
                 continue
 
             # User and other roles
-            if len(content) > 3000:
-                content = content[:2000] + "\n...[truncated]...\n" + content[-800:]
+            if len(content) > self._CONTENT_MAX:
+                content = content[:self._CONTENT_HEAD] + "\n...[truncated]...\n" + content[-self._CONTENT_TAIL:]
             parts.append(f"[{role.upper()}]: {content}")
 
         return "\n\n".join(parts)
@@ -310,6 +344,9 @@ Update the summary using this exact structure. PRESERVE all existing information
 ## Critical Context
 [Any specific values, error messages, configuration details, or data that would be lost without explicit preservation]
 
+## Tools & Patterns
+[Which tools were used, how they were used effectively, and any tool-specific discoveries. Accumulate across compactions.]
+
 Target ~{summary_budget} tokens. Be specific — include file paths, command outputs, error messages, and concrete values rather than vague descriptions.
 
 Write only the summary body. Do not include any preamble or prefix."""
@@ -348,6 +385,9 @@ Use this exact structure:
 ## Critical Context
 [Any specific values, error messages, configuration details, or data that would be lost without explicit preservation]
 
+## Tools & Patterns
+[Which tools were used, how they were used effectively, and any tool-specific discoveries (e.g., preferred flags, working invocations, successful command patterns)]
+
 Target ~{summary_budget} tokens. Be specific — include file paths, command outputs, error messages, and concrete values rather than vague descriptions. The goal is to prevent the next assistant from repeating work or losing important details.
 
 Write only the summary body. Do not include any preamble or prefix."""
@@ -518,13 +558,20 @@ Write only the summary body. Do not include any preamble or prefix."""
         derived from ``summary_target_ratio * context_length``, so it
         scales automatically with the model's context window.
 
-        Never cuts inside a tool_call/result group. Falls back to the old
-        ``protect_last_n`` if the budget would protect fewer messages.
+        Token budget is the primary criterion.  A hard minimum of 3 messages
+        is always protected, but the budget is allowed to exceed by up to
+        1.5x to avoid cutting inside an oversized message (tool output, file
+        read, etc.).  If even the minimum 3 messages exceed 1.5x the budget
+        the cut is placed right after the head so compression still runs.
+
+        Never cuts inside a tool_call/result group.
         """
         if token_budget is None:
             token_budget = self.tail_token_budget
         n = len(messages)
-        min_tail = self.protect_last_n
+        # Hard minimum: always keep at least 3 messages in the tail
+        min_tail = min(3, n - head_end - 1) if n - head_end > 1 else 0
+        soft_ceiling = int(token_budget * 1.5)
         accumulated = 0
         cut_idx = n  # start from beyond the end
 
@@ -537,21 +584,21 @@ Write only the summary body. Do not include any preamble or prefix."""
                 if isinstance(tc, dict):
                     args = tc.get("function", {}).get("arguments", "")
                     msg_tokens += len(args) // _CHARS_PER_TOKEN
-            if accumulated + msg_tokens > token_budget and (n - i) >= min_tail:
+            # Stop once we exceed the soft ceiling (unless we haven't hit min_tail yet)
+            if accumulated + msg_tokens > soft_ceiling and (n - i) >= min_tail:
                 break
             accumulated += msg_tokens
             cut_idx = i
 
-        # Ensure we protect at least protect_last_n messages
+        # Ensure we protect at least min_tail messages
         fallback_cut = n - min_tail
         if cut_idx > fallback_cut:
             cut_idx = fallback_cut
 
         # If the token budget would protect everything (small conversations),
-        # fall back to the fixed protect_last_n approach so compression can
-        # still remove middle turns.
+        # force a cut after the head so compression can still remove middle turns.
         if cut_idx <= head_end:
-            cut_idx = fallback_cut
+            cut_idx = max(fallback_cut, head_end + 1)
 
         # Align to avoid splitting tool groups
         cut_idx = self._align_boundary_backward(messages, cut_idx)
@@ -576,12 +623,13 @@ Write only the summary body. Do not include any preamble or prefix."""
         up so the API never receives mismatched IDs.
         """
         n_messages = len(messages)
-        if n_messages <= self.protect_first_n + self.protect_last_n + 1:
+        # Only need head + 3 tail messages minimum (token budget decides the real tail size)
+        _min_for_compress = self.protect_first_n + 3 + 1
+        if n_messages <= _min_for_compress:
             if not self.quiet_mode:
                 logger.warning(
                     "Cannot compress: only %d messages (need > %d)",
-                    n_messages,
-                    self.protect_first_n + self.protect_last_n + 1,
+                    n_messages, _min_for_compress,
                 )
             return messages
 
@@ -589,7 +637,8 @@ Write only the summary body. Do not include any preamble or prefix."""
 
         # Phase 1: Prune old tool results (cheap, no LLM call)
         messages, pruned_count = self._prune_old_tool_results(
-            messages, protect_tail_count=self.protect_last_n * 3,
+            messages, protect_tail_count=self.protect_last_n,
+            protect_tail_tokens=self.tail_token_budget,
         )
         if pruned_count and not self.quiet_mode:
             logger.info("Pre-compression: pruned %d old tool result(s)", pruned_count)

agent/credential_pool.py

@@ -18,12 +18,14 @@ import hermes_cli.auth as auth_mod
 from hermes_cli.auth import (
     CODEX_ACCESS_TOKEN_REFRESH_SKEW_SECONDS,
     DEFAULT_AGENT_KEY_MIN_TTL_SECONDS,
+    KIMI_CODE_BASE_URL,
     PROVIDER_REGISTRY,
     _codex_access_token_is_expiring,
     _decode_jwt_claims,
     _import_codex_cli_tokens,
     _load_auth_store,
     _load_provider_state,
+    _resolve_kimi_base_url,
     _resolve_zai_base_url,
     read_credential_pool,
     write_credential_pool,
@@ -64,10 +66,10 @@ SUPPORTED_POOL_STRATEGIES = {
 }
 
 # Cooldown before retrying an exhausted credential.
-# 429 (rate-limited) cools down faster since quotas reset frequently.
-# 402 (billing/quota) and other codes use a longer default.
+# 429 (rate-limited) and 402 (billing/quota) both cool down after 1 hour.
+# Provider-supplied reset_at timestamps override these defaults.
 EXHAUSTED_TTL_429_SECONDS = 60 * 60          # 1 hour
-EXHAUSTED_TTL_DEFAULT_SECONDS = 24 * 60 * 60 # 24 hours
+EXHAUSTED_TTL_DEFAULT_SECONDS = 60 * 60      # 1 hour
 
 # Pool key prefix for custom OpenAI-compatible endpoints.
 # Custom endpoints all share provider='custom' but are keyed by their
@@ -511,6 +513,13 @@ class CredentialPool:
                     except Exception as wexc:
                         logger.debug("Failed to write refreshed token to credentials file: %s", wexc)
             elif self.provider == "openai-codex":
+                # Proactively sync from ~/.codex/auth.json before refresh.
+                # The Codex CLI (or another Hermes profile) may have already
+                # consumed our refresh_token.  Syncing first avoids a
+                # "refresh_token_reused" error when the CLI has a newer pair.
+                synced = self._sync_codex_entry_from_cli(entry)
+                if synced is not entry:
+                    entry = synced
                 refreshed = auth_mod.refresh_codex_oauth_pure(
                     entry.access_token,
                     entry.refresh_token,
@@ -596,6 +605,35 @@ class CredentialPool:
                     # Credentials file had a valid (non-expired) token — use it directly
                     logger.debug("Credentials file has valid token, using without refresh")
                     return synced
+            # For openai-codex: the refresh_token may have been consumed by
+            # the Codex CLI between our proactive sync and the refresh call.
+            # Re-sync and retry once.
+            if self.provider == "openai-codex":
+                synced = self._sync_codex_entry_from_cli(entry)
+                if synced.refresh_token != entry.refresh_token:
+                    logger.debug("Retrying Codex refresh with synced token from ~/.codex/auth.json")
+                    try:
+                        refreshed = auth_mod.refresh_codex_oauth_pure(
+                            synced.access_token,
+                            synced.refresh_token,
+                        )
+                        updated = replace(
+                            synced,
+                            access_token=refreshed["access_token"],
+                            refresh_token=refreshed["refresh_token"],
+                            last_refresh=refreshed.get("last_refresh"),
+                            last_status=STATUS_OK,
+                            last_status_at=None,
+                            last_error_code=None,
+                        )
+                        self._replace_entry(synced, updated)
+                        self._persist()
+                        return updated
+                    except Exception as retry_exc:
+                        logger.debug("Codex retry refresh also failed: %s", retry_exc)
+                elif not self._entry_needs_refresh(synced):
+                    logger.debug("Codex CLI has valid token, using without refresh")
+                    return synced
             self._mark_exhausted(entry, None)
             return None
 
@@ -1084,7 +1122,9 @@ def _seed_from_env(provider: str, entries: List[PooledCredential]) -> Tuple[bool
         active_sources.add(source)
         auth_type = AUTH_TYPE_OAUTH if provider == "anthropic" and not token.startswith("sk-ant-api") else AUTH_TYPE_API_KEY
         base_url = env_url or pconfig.inference_base_url
-        if provider == "zai":
+        if provider == "kimi-coding":
+            base_url = _resolve_kimi_base_url(token, pconfig.inference_base_url, env_url)
+        elif provider == "zai":
             base_url = _resolve_zai_base_url(token, pconfig.inference_base_url, env_url)
         changed |= _upsert_entry(
             entries,

agent/error_classifier.py

@@ -0,0 +1,792 @@
+"""API error classification for smart failover and recovery.
+
+Provides a structured taxonomy of API errors and a priority-ordered
+classification pipeline that determines the correct recovery action
+(retry, rotate credential, fallback to another provider, compress
+context, or abort).
+
+Replaces scattered inline string-matching with a centralized classifier
+that the main retry loop in run_agent.py consults for every API failure.
+"""
+
+from __future__ import annotations
+
+import enum
+import logging
+import re
+from dataclasses import dataclass, field
+from typing import Any, Dict, Optional
+
+logger = logging.getLogger(__name__)
+
+
+# ── Error taxonomy ──────────────────────────────────────────────────────
+
+class FailoverReason(enum.Enum):
+    """Why an API call failed — determines recovery strategy."""
+
+    # Authentication / authorization
+    auth = "auth"                        # Transient auth (401/403) — refresh/rotate
+    auth_permanent = "auth_permanent"    # Auth failed after refresh — abort
+
+    # Billing / quota
+    billing = "billing"                  # 402 or confirmed credit exhaustion — rotate immediately
+    rate_limit = "rate_limit"            # 429 or quota-based throttling — backoff then rotate
+
+    # Server-side
+    overloaded = "overloaded"            # 503/529 — provider overloaded, backoff
+    server_error = "server_error"        # 500/502 — internal server error, retry
+
+    # Transport
+    timeout = "timeout"                  # Connection/read timeout — rebuild client + retry
+
+    # Context / payload
+    context_overflow = "context_overflow"  # Context too large — compress, not failover
+    payload_too_large = "payload_too_large"  # 413 — compress payload
+
+    # Model
+    model_not_found = "model_not_found"  # 404 or invalid model — fallback to different model
+
+    # Request format
+    format_error = "format_error"        # 400 bad request — abort or strip + retry
+
+    # Provider-specific
+    thinking_signature = "thinking_signature"  # Anthropic thinking block sig invalid
+    long_context_tier = "long_context_tier"    # Anthropic "extra usage" tier gate
+
+    # Catch-all
+    unknown = "unknown"                  # Unclassifiable — retry with backoff
+
+
+# ── Classification result ───────────────────────────────────────────────
+
+@dataclass
+class ClassifiedError:
+    """Structured classification of an API error with recovery hints."""
+
+    reason: FailoverReason
+    status_code: Optional[int] = None
+    provider: Optional[str] = None
+    model: Optional[str] = None
+    message: str = ""
+    error_context: Dict[str, Any] = field(default_factory=dict)
+
+    # Recovery action hints — the retry loop checks these instead of
+    # re-classifying the error itself.
+    retryable: bool = True
+    should_compress: bool = False
+    should_rotate_credential: bool = False
+    should_fallback: bool = False
+
+    @property
+    def is_auth(self) -> bool:
+        return self.reason in (FailoverReason.auth, FailoverReason.auth_permanent)
+
+    @property
+    def is_transient(self) -> bool:
+        """Error is expected to resolve on retry (with or without backoff)."""
+        return self.reason in (
+            FailoverReason.rate_limit,
+            FailoverReason.overloaded,
+            FailoverReason.server_error,
+            FailoverReason.timeout,
+            FailoverReason.unknown,
+        )
+
+
+# ── Provider-specific patterns ──────────────────────────────────────────
+
+# Patterns that indicate billing exhaustion (not transient rate limit)
+_BILLING_PATTERNS = [
+    "insufficient credits",
+    "insufficient_quota",
+    "credit balance",
+    "credits have been exhausted",
+    "top up your credits",
+    "payment required",
+    "billing hard limit",
+    "exceeded your current quota",
+    "account is deactivated",
+    "plan does not include",
+]
+
+# Patterns that indicate rate limiting (transient, will resolve)
+_RATE_LIMIT_PATTERNS = [
+    "rate limit",
+    "rate_limit",
+    "too many requests",
+    "throttled",
+    "requests per minute",
+    "tokens per minute",
+    "requests per day",
+    "try again in",
+    "please retry after",
+    "resource_exhausted",
+]
+
+# Usage-limit patterns that need disambiguation (could be billing OR rate_limit)
+_USAGE_LIMIT_PATTERNS = [
+    "usage limit",
+    "quota",
+    "limit exceeded",
+    "key limit exceeded",
+]
+
+# Patterns confirming usage limit is transient (not billing)
+_USAGE_LIMIT_TRANSIENT_SIGNALS = [
+    "try again",
+    "retry",
+    "resets at",
+    "reset in",
+    "wait",
+    "requests remaining",
+    "periodic",
+    "window",
+]
+
+# Payload-too-large patterns detected from message text (no status_code attr).
+# Proxies and some backends embed the HTTP status in the error message.
+_PAYLOAD_TOO_LARGE_PATTERNS = [
+    "request entity too large",
+    "payload too large",
+    "error code: 413",
+]
+
+# Context overflow patterns
+_CONTEXT_OVERFLOW_PATTERNS = [
+    "context length",
+    "context size",
+    "maximum context",
+    "token limit",
+    "too many tokens",
+    "reduce the length",
+    "exceeds the limit",
+    "context window",
+    "prompt is too long",
+    "prompt exceeds max length",
+    "max_tokens",
+    "maximum number of tokens",
+    # Chinese error messages (some providers return these)
+    "超过最大长度",
+    "上下文长度",
+]
+
+# Model not found patterns
+_MODEL_NOT_FOUND_PATTERNS = [
+    "is not a valid model",
+    "invalid model",
+    "model not found",
+    "model_not_found",
+    "does not exist",
+    "no such model",
+    "unknown model",
+    "unsupported model",
+]
+
+# Auth patterns (non-status-code signals)
+_AUTH_PATTERNS = [
+    "invalid api key",
+    "invalid_api_key",
+    "authentication",
+    "unauthorized",
+    "forbidden",
+    "invalid token",
+    "token expired",
+    "token revoked",
+    "access denied",
+]
+
+# Anthropic thinking block signature patterns
+_THINKING_SIG_PATTERNS = [
+    "signature",  # Combined with "thinking" check
+]
+
+# Transport error type names
+_TRANSPORT_ERROR_TYPES = frozenset({
+    "ReadTimeout", "ConnectTimeout", "PoolTimeout",
+    "ConnectError", "RemoteProtocolError",
+    "ConnectionError", "ConnectionResetError",
+    "ConnectionAbortedError", "BrokenPipeError",
+    "TimeoutError", "ReadError",
+    "ServerDisconnectedError",
+    # OpenAI SDK errors (not subclasses of Python builtins)
+    "APIConnectionError",
+    "APITimeoutError",
+})
+
+# Server disconnect patterns (no status code, but transport-level)
+_SERVER_DISCONNECT_PATTERNS = [
+    "server disconnected",
+    "peer closed connection",
+    "connection reset by peer",
+    "connection was closed",
+    "network connection lost",
+    "unexpected eof",
+    "incomplete chunked read",
+]
+
+
+# ── Classification pipeline ─────────────────────────────────────────────
+
+def classify_api_error(
+    error: Exception,
+    *,
+    provider: str = "",
+    model: str = "",
+    approx_tokens: int = 0,
+    context_length: int = 200000,
+    num_messages: int = 0,
+) -> ClassifiedError:
+    """Classify an API error into a structured recovery recommendation.
+
+    Priority-ordered pipeline:
+      1. Special-case provider-specific patterns (thinking sigs, tier gates)
+      2. HTTP status code + message-aware refinement
+      3. Error code classification (from body)
+      4. Message pattern matching (billing vs rate_limit vs context vs auth)
+      5. Transport error heuristics
+      6. Server disconnect + large session → context overflow
+      7. Fallback: unknown (retryable with backoff)
+
+    Args:
+        error: The exception from the API call.
+        provider: Current provider name (e.g. "openrouter", "anthropic").
+        model: Current model slug.
+        approx_tokens: Approximate token count of the current context.
+        context_length: Maximum context length for the current model.
+
+    Returns:
+        ClassifiedError with reason and recovery action hints.
+    """
+    status_code = _extract_status_code(error)
+    error_type = type(error).__name__
+    body = _extract_error_body(error)
+    error_code = _extract_error_code(body)
+
+    # Build a comprehensive error message string for pattern matching.
+    # str(error) alone may not include the body message (e.g. OpenAI SDK's
+    # APIStatusError.__str__ returns the first arg, not the body).  Append
+    # the body message so patterns like "try again" in 402 disambiguation
+    # are detected even when only present in the structured body.
+    #
+    # Also extract metadata.raw — OpenRouter wraps upstream provider errors
+    # inside {"error": {"message": "Provider returned error", "metadata":
+    # {"raw": "<actual error JSON>"}}} and the real error message (e.g.
+    # "context length exceeded") is only in the inner JSON.
+    _raw_msg = str(error).lower()
+    _body_msg = ""
+    _metadata_msg = ""
+    if isinstance(body, dict):
+        _err_obj = body.get("error", {})
+        if isinstance(_err_obj, dict):
+            _body_msg = (_err_obj.get("message") or "").lower()
+            # Parse metadata.raw for wrapped provider errors
+            _metadata = _err_obj.get("metadata", {})
+            if isinstance(_metadata, dict):
+                _raw_json = _metadata.get("raw") or ""
+                if isinstance(_raw_json, str) and _raw_json.strip():
+                    try:
+                        import json
+                        _inner = json.loads(_raw_json)
+                        if isinstance(_inner, dict):
+                            _inner_err = _inner.get("error", {})
+                            if isinstance(_inner_err, dict):
+                                _metadata_msg = (_inner_err.get("message") or "").lower()
+                    except (json.JSONDecodeError, TypeError):
+                        pass
+        if not _body_msg:
+            _body_msg = (body.get("message") or "").lower()
+    # Combine all message sources for pattern matching
+    parts = [_raw_msg]
+    if _body_msg and _body_msg not in _raw_msg:
+        parts.append(_body_msg)
+    if _metadata_msg and _metadata_msg not in _raw_msg and _metadata_msg not in _body_msg:
+        parts.append(_metadata_msg)
+    error_msg = " ".join(parts)
+    provider_lower = (provider or "").strip().lower()
+    model_lower = (model or "").strip().lower()
+
+    def _result(reason: FailoverReason, **overrides) -> ClassifiedError:
+        defaults = {
+            "reason": reason,
+            "status_code": status_code,
+            "provider": provider,
+            "model": model,
+            "message": _extract_message(error, body),
+        }
+        defaults.update(overrides)
+        return ClassifiedError(**defaults)
+
+    # ── 1. Provider-specific patterns (highest priority) ────────────
+
+    # Anthropic thinking block signature invalid (400).
+    # Don't gate on provider — OpenRouter proxies Anthropic errors, so the
+    # provider may be "openrouter" even though the error is Anthropic-specific.
+    # The message pattern ("signature" + "thinking") is unique enough.
+    if (
+        status_code == 400
+        and "signature" in error_msg
+        and "thinking" in error_msg
+    ):
+        return _result(
+            FailoverReason.thinking_signature,
+            retryable=True,
+            should_compress=False,
+        )
+
+    # Anthropic long-context tier gate (429 "extra usage" + "long context")
+    if (
+        status_code == 429
+        and "extra usage" in error_msg
+        and "long context" in error_msg
+    ):
+        return _result(
+            FailoverReason.long_context_tier,
+            retryable=True,
+            should_compress=True,
+        )
+
+    # ── 2. HTTP status code classification ──────────────────────────
+
+    if status_code is not None:
+        classified = _classify_by_status(
+            status_code, error_msg, error_code, body,
+            provider=provider_lower, model=model_lower,
+            approx_tokens=approx_tokens, context_length=context_length,
+            num_messages=num_messages,
+            result_fn=_result,
+        )
+        if classified is not None:
+            return classified
+
+    # ── 3. Error code classification ────────────────────────────────
+
+    if error_code:
+        classified = _classify_by_error_code(error_code, error_msg, _result)
+        if classified is not None:
+            return classified
+
+    # ── 4. Message pattern matching (no status code) ────────────────
+
+    classified = _classify_by_message(
+        error_msg, error_type,
+        approx_tokens=approx_tokens,
+        context_length=context_length,
+        result_fn=_result,
+    )
+    if classified is not None:
+        return classified
+
+    # ── 5. Server disconnect + large session → context overflow ─────
+    # Must come BEFORE generic transport error catch — a disconnect on
+    # a large session is more likely context overflow than a transient
+    # transport hiccup.  Without this ordering, RemoteProtocolError
+    # always maps to timeout regardless of session size.
+
+    is_disconnect = any(p in error_msg for p in _SERVER_DISCONNECT_PATTERNS)
+    if is_disconnect and not status_code:
+        is_large = approx_tokens > context_length * 0.6 or approx_tokens > 120000 or num_messages > 200
+        if is_large:
+            return _result(
+                FailoverReason.context_overflow,
+                retryable=True,
+                should_compress=True,
+            )
+        return _result(FailoverReason.timeout, retryable=True)
+
+    # ── 6. Transport / timeout heuristics ───────────────────────────
+
+    if error_type in _TRANSPORT_ERROR_TYPES or isinstance(error, (TimeoutError, ConnectionError, OSError)):
+        return _result(FailoverReason.timeout, retryable=True)
+
+    # ── 7. Fallback: unknown ────────────────────────────────────────
+
+    return _result(FailoverReason.unknown, retryable=True)
+
+
+# ── Status code classification ──────────────────────────────────────────
+
+def _classify_by_status(
+    status_code: int,
+    error_msg: str,
+    error_code: str,
+    body: dict,
+    *,
+    provider: str,
+    model: str,
+    approx_tokens: int,
+    context_length: int,
+    num_messages: int = 0,
+    result_fn,
+) -> Optional[ClassifiedError]:
+    """Classify based on HTTP status code with message-aware refinement."""
+
+    if status_code == 401:
+        # Not retryable on its own — credential pool rotation and
+        # provider-specific refresh (Codex, Anthropic, Nous) run before
+        # the retryability check in run_agent.py.  If those succeed, the
+        # loop `continue`s.  If they fail, retryable=False ensures we
+        # hit the client-error abort path (which tries fallback first).
+        return result_fn(
+            FailoverReason.auth,
+            retryable=False,
+            should_rotate_credential=True,
+            should_fallback=True,
+        )
+
+    if status_code == 403:
+        # OpenRouter 403 "key limit exceeded" is actually billing
+        if "key limit exceeded" in error_msg or "spending limit" in error_msg:
+            return result_fn(
+                FailoverReason.billing,
+                retryable=False,
+                should_rotate_credential=True,
+                should_fallback=True,
+            )
+        return result_fn(
+            FailoverReason.auth,
+            retryable=False,
+            should_fallback=True,
+        )
+
+    if status_code == 402:
+        return _classify_402(error_msg, result_fn)
+
+    if status_code == 404:
+        if any(p in error_msg for p in _MODEL_NOT_FOUND_PATTERNS):
+            return result_fn(
+                FailoverReason.model_not_found,
+                retryable=False,
+                should_fallback=True,
+            )
+        # Generic 404 — could be model or endpoint
+        return result_fn(
+            FailoverReason.model_not_found,
+            retryable=False,
+            should_fallback=True,
+        )
+
+    if status_code == 413:
+        return result_fn(
+            FailoverReason.payload_too_large,
+            retryable=True,
+            should_compress=True,
+        )
+
+    if status_code == 429:
+        # Already checked long_context_tier above; this is a normal rate limit
+        return result_fn(
+            FailoverReason.rate_limit,
+            retryable=True,
+            should_rotate_credential=True,
+            should_fallback=True,
+        )
+
+    if status_code == 400:
+        return _classify_400(
+            error_msg, error_code, body,
+            provider=provider, model=model,
+            approx_tokens=approx_tokens,
+            context_length=context_length,
+            num_messages=num_messages,
+            result_fn=result_fn,
+        )
+
+    if status_code in (500, 502):
+        return result_fn(FailoverReason.server_error, retryable=True)
+
+    if status_code in (503, 529):
+        return result_fn(FailoverReason.overloaded, retryable=True)
+
+    # Other 4xx — non-retryable
+    if 400 <= status_code < 500:
+        return result_fn(
+            FailoverReason.format_error,
+            retryable=False,
+            should_fallback=True,
+        )
+
+    # Other 5xx — retryable
+    if 500 <= status_code < 600:
+        return result_fn(FailoverReason.server_error, retryable=True)
+
+    return None
+
+
+def _classify_402(error_msg: str, result_fn) -> ClassifiedError:
+    """Disambiguate 402: billing exhaustion vs transient usage limit.
+
+    The key insight from OpenClaw: some 402s are transient rate limits
+    disguised as payment errors.  "Usage limit, try again in 5 minutes"
+    is NOT a billing problem — it's a periodic quota that resets.
+    """
+    # Check for transient usage-limit signals first
+    has_usage_limit = any(p in error_msg for p in _USAGE_LIMIT_PATTERNS)
+    has_transient_signal = any(p in error_msg for p in _USAGE_LIMIT_TRANSIENT_SIGNALS)
+
+    if has_usage_limit and has_transient_signal:
+        # Transient quota — treat as rate limit, not billing
+        return result_fn(
+            FailoverReason.rate_limit,
+            retryable=True,
+            should_rotate_credential=True,
+            should_fallback=True,
+        )
+
+    # Confirmed billing exhaustion
+    return result_fn(
+        FailoverReason.billing,
+        retryable=False,
+        should_rotate_credential=True,
+        should_fallback=True,
+    )
+
+
+def _classify_400(
+    error_msg: str,
+    error_code: str,
+    body: dict,
+    *,
+    provider: str,
+    model: str,
+    approx_tokens: int,
+    context_length: int,
+    num_messages: int = 0,
+    result_fn,
+) -> ClassifiedError:
+    """Classify 400 Bad Request — context overflow, format error, or generic."""
+
+    # Context overflow from 400
+    if any(p in error_msg for p in _CONTEXT_OVERFLOW_PATTERNS):
+        return result_fn(
+            FailoverReason.context_overflow,
+            retryable=True,
+            should_compress=True,
+        )
+
+    # Some providers return model-not-found as 400 instead of 404 (e.g. OpenRouter).
+    if any(p in error_msg for p in _MODEL_NOT_FOUND_PATTERNS):
+        return result_fn(
+            FailoverReason.model_not_found,
+            retryable=False,
+            should_fallback=True,
+        )
+
+    # Some providers return rate limit / billing errors as 400 instead of 429/402.
+    # Check these patterns before falling through to format_error.
+    if any(p in error_msg for p in _RATE_LIMIT_PATTERNS):
+        return result_fn(
+            FailoverReason.rate_limit,
+            retryable=True,
+            should_rotate_credential=True,
+            should_fallback=True,
+        )
+    if any(p in error_msg for p in _BILLING_PATTERNS):
+        return result_fn(
+            FailoverReason.billing,
+            retryable=False,
+            should_rotate_credential=True,
+            should_fallback=True,
+        )
+
+    # Generic 400 + large session → probable context overflow
+    # Anthropic sometimes returns a bare "Error" message when context is too large
+    err_body_msg = ""
+    if isinstance(body, dict):
+        err_obj = body.get("error", {})
+        if isinstance(err_obj, dict):
+            err_body_msg = (err_obj.get("message") or "").strip().lower()
+        # Responses API (and some providers) use flat body: {"message": "..."}
+        if not err_body_msg:
+            err_body_msg = (body.get("message") or "").strip().lower()
+    is_generic = len(err_body_msg) < 30 or err_body_msg in ("error", "")
+    is_large = approx_tokens > context_length * 0.4 or approx_tokens > 80000 or num_messages > 80
+
+    if is_generic and is_large:
+        return result_fn(
+            FailoverReason.context_overflow,
+            retryable=True,
+            should_compress=True,
+        )
+
+    # Non-retryable format error
+    return result_fn(
+        FailoverReason.format_error,
+        retryable=False,
+        should_fallback=True,
+    )
+
+
+# ── Error code classification ───────────────────────────────────────────
+
+def _classify_by_error_code(
+    error_code: str, error_msg: str, result_fn,
+) -> Optional[ClassifiedError]:
+    """Classify by structured error codes from the response body."""
+    code_lower = error_code.lower()
+
+    if code_lower in ("resource_exhausted", "throttled", "rate_limit_exceeded"):
+        return result_fn(
+            FailoverReason.rate_limit,
+            retryable=True,
+            should_rotate_credential=True,
+        )
+
+    if code_lower in ("insufficient_quota", "billing_not_active", "payment_required"):
+        return result_fn(
+            FailoverReason.billing,
+            retryable=False,
+            should_rotate_credential=True,
+            should_fallback=True,
+        )
+
+    if code_lower in ("model_not_found", "model_not_available", "invalid_model"):
+        return result_fn(
+            FailoverReason.model_not_found,
+            retryable=False,
+            should_fallback=True,
+        )
+
+    if code_lower in ("context_length_exceeded", "max_tokens_exceeded"):
+        return result_fn(
+            FailoverReason.context_overflow,
+            retryable=True,
+            should_compress=True,
+        )
+
+    return None
+
+
+# ── Message pattern classification ──────────────────────────────────────
+
+def _classify_by_message(
+    error_msg: str,
+    error_type: str,
+    *,
+    approx_tokens: int,
+    context_length: int,
+    result_fn,
+) -> Optional[ClassifiedError]:
+    """Classify based on error message patterns when no status code is available."""
+
+    # Payload-too-large patterns (from message text when no status_code)
+    if any(p in error_msg for p in _PAYLOAD_TOO_LARGE_PATTERNS):
+        return result_fn(
+            FailoverReason.payload_too_large,
+            retryable=True,
+            should_compress=True,
+        )
+
+    # Billing patterns
+    if any(p in error_msg for p in _BILLING_PATTERNS):
+        return result_fn(
+            FailoverReason.billing,
+            retryable=False,
+            should_rotate_credential=True,
+            should_fallback=True,
+        )
+
+    # Rate limit patterns
+    if any(p in error_msg for p in _RATE_LIMIT_PATTERNS):
+        return result_fn(
+            FailoverReason.rate_limit,
+            retryable=True,
+            should_rotate_credential=True,
+            should_fallback=True,
+        )
+
+    # Context overflow patterns
+    if any(p in error_msg for p in _CONTEXT_OVERFLOW_PATTERNS):
+        return result_fn(
+            FailoverReason.context_overflow,
+            retryable=True,
+            should_compress=True,
+        )
+
+    # Auth patterns
+    if any(p in error_msg for p in _AUTH_PATTERNS):
+        return result_fn(
+            FailoverReason.auth,
+            retryable=True,
+            should_rotate_credential=True,
+        )
+
+    # Model not found patterns
+    if any(p in error_msg for p in _MODEL_NOT_FOUND_PATTERNS):
+        return result_fn(
+            FailoverReason.model_not_found,
+            retryable=False,
+            should_fallback=True,
+        )
+
+    return None
+
+
+# ── Helpers ─────────────────────────────────────────────────────────────
+
+def _extract_status_code(error: Exception) -> Optional[int]:
+    """Walk the error and its cause chain to find an HTTP status code."""
+    current = error
+    for _ in range(5):  # Max depth to prevent infinite loops
+        code = getattr(current, "status_code", None)
+        if isinstance(code, int):
+            return code
+        # Some SDKs use .status instead of .status_code
+        code = getattr(current, "status", None)
+        if isinstance(code, int) and 100 <= code < 600:
+            return code
+        # Walk cause chain
+        cause = getattr(current, "__cause__", None) or getattr(current, "__context__", None)
+        if cause is None or cause is current:
+            break
+        current = cause
+    return None
+
+
+def _extract_error_body(error: Exception) -> dict:
+    """Extract the structured error body from an SDK exception."""
+    body = getattr(error, "body", None)
+    if isinstance(body, dict):
+        return body
+    # Some errors have .response.json()
+    response = getattr(error, "response", None)
+    if response is not None:
+        try:
+            json_body = response.json()
+            if isinstance(json_body, dict):
+                return json_body
+        except Exception:
+            pass
+    return {}
+
+
+def _extract_error_code(body: dict) -> str:
+    """Extract an error code string from the response body."""
+    if not body:
+        return ""
+    error_obj = body.get("error", {})
+    if isinstance(error_obj, dict):
+        code = error_obj.get("code") or error_obj.get("type") or ""
+        if isinstance(code, str) and code.strip():
+            return code.strip()
+    # Top-level code
+    code = body.get("code") or body.get("error_code") or ""
+    if isinstance(code, (str, int)):
+        return str(code).strip()
+    return ""
+
+
+def _extract_message(error: Exception, body: dict) -> str:
+    """Extract the most informative error message."""
+    # Try structured body first
+    if body:
+        error_obj = body.get("error", {})
+        if isinstance(error_obj, dict):
+            msg = error_obj.get("message", "")
+            if isinstance(msg, str) and msg.strip():
+                return msg.strip()[:500]
+        msg = body.get("message", "")
+        if isinstance(msg, str) and msg.strip():
+            return msg.strip()[:500]
+    # Fallback to str(error)
+    return str(error)[:500]

agent/model_metadata.py

@@ -26,12 +26,14 @@ _PROVIDER_PREFIXES: frozenset[str] = frozenset({
     "openrouter", "nous", "openai-codex", "copilot", "copilot-acp",
     "gemini", "zai", "kimi-coding", "minimax", "minimax-cn", "anthropic", "deepseek",
     "opencode-zen", "opencode-go", "ai-gateway", "kilocode", "alibaba",
+    "qwen-oauth",
     "custom", "local",
     # Common aliases
     "google", "google-gemini", "google-ai-studio",
     "glm", "z-ai", "z.ai", "zhipu", "github", "github-copilot",
     "github-models", "kimi", "moonshot", "claude", "deep-seek",
     "opencode", "zen", "go", "vercel", "kilo", "dashscope", "aliyun", "qwen",
+    "qwen-portal",
 })
 
 
@@ -134,7 +136,7 @@ DEFAULT_CONTEXT_LENGTHS = {
     "deepseek-ai/DeepSeek-V3.2": 65536,
     "moonshotai/Kimi-K2.5": 262144,
     "moonshotai/Kimi-K2-Thinking": 262144,
-    "minimaxai/minimax-m2.5": 1048576,
+    "MiniMaxAI/MiniMax-M2.5": 1048576,
     "XiaomiMiMo/MiMo-V2-Flash": 32768,
     "mimo-v2-pro": 1048576,
     "mimo-v2-omni": 1048576,
@@ -187,6 +189,7 @@ _URL_TO_PROVIDER: Dict[str, str] = {
     "api.minimax": "minimax",
     "dashscope.aliyuncs.com": "alibaba",
     "dashscope-intl.aliyuncs.com": "alibaba",
+    "portal.qwen.ai": "qwen-oauth",
     "openrouter.ai": "openrouter",
     "generativelanguage.googleapis.com": "gemini",
     "inference-api.nousresearch.com": "nous",
@@ -194,6 +197,7 @@ _URL_TO_PROVIDER: Dict[str, str] = {
     "api.githubcopilot.com": "copilot",
     "models.github.ai": "copilot",
     "api.fireworks.ai": "fireworks",
+    "opencode.ai": "opencode-go",
 }
 
 
@@ -599,6 +603,49 @@ def parse_context_limit_from_error(error_msg: str) -> Optional[int]:
     return None
 
 
+def parse_available_output_tokens_from_error(error_msg: str) -> Optional[int]:
+    """Detect an "output cap too large" error and return how many output tokens are available.
+
+    Background — two distinct context errors exist:
+      1. "Prompt too long"  — the INPUT itself exceeds the context window.
+           Fix: compress history and/or halve context_length.
+      2. "max_tokens too large" — input is fine, but input + requested_output > window.
+           Fix: reduce max_tokens (the output cap) for this call.
+           Do NOT touch context_length — the window hasn't shrunk.
+
+    Anthropic's API returns errors like:
+      "max_tokens: 32768 > context_window: 200000 - input_tokens: 190000 = available_tokens: 10000"
+
+    Returns the number of output tokens that would fit (e.g. 10000 above), or None if
+    the error does not look like a max_tokens-too-large error.
+    """
+    error_lower = error_msg.lower()
+
+    # Must look like an output-cap error, not a prompt-length error.
+    is_output_cap_error = (
+        "max_tokens" in error_lower
+        and ("available_tokens" in error_lower or "available tokens" in error_lower)
+    )
+    if not is_output_cap_error:
+        return None
+
+    # Extract the available_tokens figure.
+    # Anthropic format: "… = available_tokens: 10000"
+    patterns = [
+        r'available_tokens[:\s]+(\d+)',
+        r'available\s+tokens[:\s]+(\d+)',
+        # fallback: last number after "=" in expressions like "200000 - 190000 = 10000"
+        r'=\s*(\d+)\s*$',
+    ]
+    for pattern in patterns:
+        match = re.search(pattern, error_lower)
+        if match:
+            tokens = int(match.group(1))
+            if tokens >= 1:
+                return tokens
+    return None
+
+
 def _model_id_matches(candidate_id: str, lookup_model: str) -> bool:
     """Return True if *candidate_id* (from server) matches *lookup_model* (configured).
 

agent/models_dev.py

@@ -153,6 +153,7 @@ PROVIDER_TO_MODELS_DEV: Dict[str, str] = {
     "minimax-cn": "minimax-cn",
     "deepseek": "deepseek",
     "alibaba": "alibaba",
+    "qwen-oauth": "alibaba",
     "copilot": "github-copilot",
     "ai-gateway": "vercel",
     "opencode-zen": "opencode",

agent/prompt_builder.py

@@ -349,6 +349,13 @@ PLATFORM_HINTS = {
         "only — no markdown, no formatting. SMS messages are limited to ~1600 "
         "characters, so be brief and direct."
     ),
+    "bluebubbles": (
+        "You are chatting via iMessage (BlueBubbles). iMessage does not render "
+        "markdown formatting — use plain text. Keep responses concise as they "
+        "appear as text messages. You can send media files natively: include "
+        "MEDIA:/absolute/path/to/file in your response. Images (.jpg, .png, "
+        ".heic) appear as photos and other files arrive as attachments."
+    ),
 }
 
 CONTEXT_FILE_MAX_CHARS = 20_000

agent/rate_limit_tracker.py

@@ -0,0 +1,242 @@
+"""Rate limit tracking for inference API responses.
+
+Captures x-ratelimit-* headers from provider responses and provides
+formatted display for the /usage slash command.  Currently supports
+the Nous Portal header format (also used by OpenRouter and OpenAI-compatible
+APIs that follow the same convention).
+
+Header schema (12 headers total):
+    x-ratelimit-limit-requests          RPM cap
+    x-ratelimit-limit-requests-1h       RPH cap
+    x-ratelimit-limit-tokens            TPM cap
+    x-ratelimit-limit-tokens-1h         TPH cap
+    x-ratelimit-remaining-requests      requests left in minute window
+    x-ratelimit-remaining-requests-1h   requests left in hour window
+    x-ratelimit-remaining-tokens        tokens left in minute window
+    x-ratelimit-remaining-tokens-1h     tokens left in hour window
+    x-ratelimit-reset-requests          seconds until minute request window resets
+    x-ratelimit-reset-requests-1h       seconds until hour request window resets
+    x-ratelimit-reset-tokens            seconds until minute token window resets
+    x-ratelimit-reset-tokens-1h         seconds until hour token window resets
+"""
+
+from __future__ import annotations
+
+import time
+from dataclasses import dataclass, field
+from typing import Any, Dict, Mapping, Optional
+
+
+@dataclass
+class RateLimitBucket:
+    """One rate-limit window (e.g. requests per minute)."""
+
+    limit: int = 0
+    remaining: int = 0
+    reset_seconds: float = 0.0
+    captured_at: float = 0.0  # time.time() when this was captured
+
+    @property
+    def used(self) -> int:
+        return max(0, self.limit - self.remaining)
+
+    @property
+    def usage_pct(self) -> float:
+        if self.limit <= 0:
+            return 0.0
+        return (self.used / self.limit) * 100.0
+
+    @property
+    def remaining_seconds_now(self) -> float:
+        """Estimated seconds remaining until reset, adjusted for elapsed time."""
+        elapsed = time.time() - self.captured_at
+        return max(0.0, self.reset_seconds - elapsed)
+
+
+@dataclass
+class RateLimitState:
+    """Full rate-limit state parsed from response headers."""
+
+    requests_min: RateLimitBucket = field(default_factory=RateLimitBucket)
+    requests_hour: RateLimitBucket = field(default_factory=RateLimitBucket)
+    tokens_min: RateLimitBucket = field(default_factory=RateLimitBucket)
+    tokens_hour: RateLimitBucket = field(default_factory=RateLimitBucket)
+    captured_at: float = 0.0  # when the headers were captured
+    provider: str = ""
+
+    @property
+    def has_data(self) -> bool:
+        return self.captured_at > 0
+
+    @property
+    def age_seconds(self) -> float:
+        if not self.has_data:
+            return float("inf")
+        return time.time() - self.captured_at
+
+
+def _safe_int(value: Any, default: int = 0) -> int:
+    try:
+        return int(float(value))
+    except (TypeError, ValueError):
+        return default
+
+
+def _safe_float(value: Any, default: float = 0.0) -> float:
+    try:
+        return float(value)
+    except (TypeError, ValueError):
+        return default
+
+
+def parse_rate_limit_headers(
+    headers: Mapping[str, str],
+    provider: str = "",
+) -> Optional[RateLimitState]:
+    """Parse x-ratelimit-* headers into a RateLimitState.
+
+    Returns None if no rate limit headers are present.
+    """
+    # Quick check: at least one rate limit header must exist
+    has_any = any(k.lower().startswith("x-ratelimit-") for k in headers)
+    if not has_any:
+        return None
+
+    now = time.time()
+
+    def _bucket(resource: str, suffix: str = "") -> RateLimitBucket:
+        # e.g. resource="requests", suffix="" -> per-minute
+        #      resource="tokens", suffix="-1h" -> per-hour
+        tag = f"{resource}{suffix}"
+        return RateLimitBucket(
+            limit=_safe_int(headers.get(f"x-ratelimit-limit-{tag}")),
+            remaining=_safe_int(headers.get(f"x-ratelimit-remaining-{tag}")),
+            reset_seconds=_safe_float(headers.get(f"x-ratelimit-reset-{tag}")),
+            captured_at=now,
+        )
+
+    return RateLimitState(
+        requests_min=_bucket("requests"),
+        requests_hour=_bucket("requests", "-1h"),
+        tokens_min=_bucket("tokens"),
+        tokens_hour=_bucket("tokens", "-1h"),
+        captured_at=now,
+        provider=provider,
+    )
+
+
+# ── Formatting ──────────────────────────────────────────────────────────
+
+
+def _fmt_count(n: int) -> str:
+    """Human-friendly number: 7999856 -> '8.0M', 33599 -> '33.6K', 799 -> '799'."""
+    if n >= 1_000_000:
+        return f"{n / 1_000_000:.1f}M"
+    if n >= 10_000:
+        return f"{n / 1_000:.1f}K"
+    if n >= 1_000:
+        return f"{n / 1_000:.1f}K"
+    return str(n)
+
+
+def _fmt_seconds(seconds: float) -> str:
+    """Seconds -> human-friendly duration: '58s', '2m 14s', '58m 57s', '1h 2m'."""
+    s = max(0, int(seconds))
+    if s < 60:
+        return f"{s}s"
+    if s < 3600:
+        m, sec = divmod(s, 60)
+        return f"{m}m {sec}s" if sec else f"{m}m"
+    h, remainder = divmod(s, 3600)
+    m = remainder // 60
+    return f"{h}h {m}m" if m else f"{h}h"
+
+
+def _bar(pct: float, width: int = 20) -> str:
+    """ASCII progress bar: [████████░░░░░░░░░░░░] 40%."""
+    filled = int(pct / 100.0 * width)
+    filled = max(0, min(width, filled))
+    empty = width - filled
+    return f"[{'█' * filled}{'░' * empty}]"
+
+
+def _bucket_line(label: str, bucket: RateLimitBucket, label_width: int = 14) -> str:
+    """Format one bucket as a single line."""
+    if bucket.limit <= 0:
+        return f"  {label:<{label_width}}  (no data)"
+
+    pct = bucket.usage_pct
+    used = _fmt_count(bucket.used)
+    limit = _fmt_count(bucket.limit)
+    remaining = _fmt_count(bucket.remaining)
+    reset = _fmt_seconds(bucket.remaining_seconds_now)
+
+    bar = _bar(pct)
+    return f"  {label:<{label_width}} {bar} {pct:5.1f}%  {used}/{limit} used  ({remaining} left, resets in {reset})"
+
+
+def format_rate_limit_display(state: RateLimitState) -> str:
+    """Format rate limit state for terminal/chat display."""
+    if not state.has_data:
+        return "No rate limit data yet — make an API request first."
+
+    age = state.age_seconds
+    if age < 5:
+        freshness = "just now"
+    elif age < 60:
+        freshness = f"{int(age)}s ago"
+    else:
+        freshness = f"{_fmt_seconds(age)} ago"
+
+    provider_label = state.provider.title() if state.provider else "Provider"
+
+    lines = [
+        f"{provider_label} Rate Limits (captured {freshness}):",
+        "",
+        _bucket_line("Requests/min", state.requests_min),
+        _bucket_line("Requests/hr", state.requests_hour),
+        "",
+        _bucket_line("Tokens/min", state.tokens_min),
+        _bucket_line("Tokens/hr", state.tokens_hour),
+    ]
+
+    # Add warnings if any bucket is getting hot
+    warnings = []
+    for label, bucket in [
+        ("requests/min", state.requests_min),
+        ("requests/hr", state.requests_hour),
+        ("tokens/min", state.tokens_min),
+        ("tokens/hr", state.tokens_hour),
+    ]:
+        if bucket.limit > 0 and bucket.usage_pct >= 80:
+            reset = _fmt_seconds(bucket.remaining_seconds_now)
+            warnings.append(f"  ⚠ {label} at {bucket.usage_pct:.0f}% — resets in {reset}")
+
+    if warnings:
+        lines.append("")
+        lines.extend(warnings)
+
+    return "\n".join(lines)
+
+
+def format_rate_limit_compact(state: RateLimitState) -> str:
+    """One-line compact summary for status bars / gateway messages."""
+    if not state.has_data:
+        return "No rate limit data."
+
+    rm = state.requests_min
+    tm = state.tokens_min
+    rh = state.requests_hour
+    th = state.tokens_hour
+
+    parts = []
+    if rm.limit > 0:
+        parts.append(f"RPM: {rm.remaining}/{rm.limit}")
+    if rh.limit > 0:
+        parts.append(f"RPH: {_fmt_count(rh.remaining)}/{_fmt_count(rh.limit)} (resets {_fmt_seconds(rh.remaining_seconds_now)})")
+    if tm.limit > 0:
+        parts.append(f"TPM: {_fmt_count(tm.remaining)}/{_fmt_count(tm.limit)}")
+    if th.limit > 0:
+        parts.append(f"TPH: {_fmt_count(th.remaining)}/{_fmt_count(th.limit)} (resets {_fmt_seconds(th.remaining_seconds_now)})")
+
+    return " | ".join(parts)

agent/subdirectory_hints.py

@@ -159,7 +159,10 @@ class SubdirectoryHintTracker:
 
     def _is_valid_subdir(self, path: Path) -> bool:
         """Check if path is a valid directory to scan for hints."""
-        if not path.is_dir():
+        try:
+            if not path.is_dir():
+                return False
+        except OSError:
             return False
         if path in self._loaded_dirs:
             return False
@@ -172,7 +175,10 @@ class SubdirectoryHintTracker:
         found_hints = []
         for filename in _HINT_FILENAMES:
             hint_path = directory / filename
-            if not hint_path.is_file():
+            try:
+                if not hint_path.is_file():
+                    continue
+            except OSError:
                 continue
             try:
                 content = hint_path.read_text(encoding="utf-8").strip()

cli-config.yaml.example

@@ -48,6 +48,25 @@ model:
   # api_key: "your-key-here"  # Uncomment to set here instead of .env
   base_url: "https://openrouter.ai/api/v1"
 
+  # ── Token limits — two settings, easy to confuse ──────────────────────────
+  #
+  # context_length: TOTAL context window (input + output tokens combined).
+  #   Controls when Hermes compresses history and validates requests.
+  #   Leave unset — Hermes auto-detects the correct value from the provider.
+  #   Set manually only when auto-detection is wrong (e.g. a local server with
+  #   a custom num_ctx, or a proxy that doesn't expose /v1/models).
+  #
+  # context_length: 131072
+  #
+  # max_tokens: OUTPUT cap — maximum tokens the model may generate per response.
+  #   Unrelated to how long your conversation history can be.
+  #   The OpenAI-standard name "max_tokens" is a misnomer; Anthropic's native
+  #   API has since renamed it "max_output_tokens" for clarity.
+  #   Leave unset to use the model's native output ceiling (recommended).
+  #   Set only if you want to deliberately limit individual response length.
+  #
+  # max_tokens: 8192
+
 # =============================================================================
 # OpenRouter Provider Routing (only applies when using OpenRouter)
 # =============================================================================
@@ -117,7 +136,8 @@ terminal:
   timeout: 180
   docker_mount_cwd_to_workspace: false  # SECURITY: off by default. Opt in to mount the launch cwd into Docker /workspace.
   lifetime_seconds: 300
-  # sudo_password: ""  # Enable sudo commands (pipes via sudo -S) - SECURITY WARNING: plaintext!
+  # sudo_password: "hunter2"  # Optional: pipe a sudo password via sudo -S. SECURITY WARNING: plaintext.
+  # sudo_password: ""         # Explicit empty password: try empty and never open the interactive sudo prompt.
 
 # -----------------------------------------------------------------------------
 # OPTION 2: SSH remote execution
@@ -208,13 +228,18 @@ terminal:
 #
 # SECURITY WARNING: Password stored in plaintext!
 #
-# INTERACTIVE PROMPT: If no sudo_password is set and the CLI is running,
+# INTERACTIVE PROMPT: If sudo_password is unset and the CLI is running,
 # you'll be prompted to enter your password when sudo is needed:
 # - 45-second timeout (auto-skips if no input)
 # - Press Enter to skip (command fails gracefully)
 # - Password is hidden while typing
 # - Password is cached for the session
 #
+# EMPTY PASSWORDS: Setting sudo_password to an explicit empty string is different
+# from leaving it unset. Hermes will try an empty password via `sudo -S` and
+# will not open the interactive prompt. This is useful for passwordless sudo,
+# Touch ID sudo setups, and environments where prompting is just noise.
+#
 # ALTERNATIVES:
 # - SSH backend: Configure passwordless sudo on the remote server
 # - Containers: Run as root inside the container (no sudo needed)
@@ -445,6 +470,16 @@ agent:
   # Higher = more room for complex tasks, but costs more tokens
   # Recommended: 20-30 for focused tasks, 50-100 for open exploration
   max_turns: 60
+
+  # Inactivity timeout for gateway agent runs (seconds, 0 = unlimited).
+  # The agent can run indefinitely when actively calling tools or receiving
+  # API responses.  Only fires after the agent has been idle for this duration.
+  # gateway_timeout: 1800
+
+  # Staged warning: send a warning before escalating to full timeout.
+  # Fires once per run when inactivity reaches this threshold (seconds).
+  # Set to 0 to disable the warning.
+  # gateway_timeout_warning: 900
   
   # Enable verbose logging
   verbose: false
@@ -644,10 +679,14 @@ platform_toolsets:
 # Voice Transcription (Speech-to-Text)
 # =============================================================================
 # Automatically transcribe voice messages on messaging platforms.
-# Requires OPENAI_API_KEY in .env (uses OpenAI Whisper API directly).
+# Providers: local (free, faster-whisper) | groq (free tier) | openai (Whisper API) | mistral (Voxtral Transcribe)
+# Set the corresponding API key in .env: GROQ_API_KEY, OPENAI_API_KEY, or MISTRAL_API_KEY.
 stt:
   enabled: true
+  # provider: "local"          # auto-detected if omitted
   model: "whisper-1"  # whisper-1 (cheapest) | gpt-4o-mini-transcribe | gpt-4o-transcribe
+  # mistral:
+  #   model: "voxtral-mini-latest"  # voxtral-mini-latest | voxtral-mini-2602
 
 # =============================================================================
 # Response Pacing (Messaging Platforms)

cli.py

@@ -1546,6 +1546,7 @@ class HermesCLI:
         self._clarify_deadline = 0
         self._sudo_state = None
         self._sudo_deadline = 0
+        self._modal_input_snapshot = None
         self._approval_state = None
         self._approval_deadline = 0
         self._approval_lock = threading.Lock()
@@ -1602,7 +1603,12 @@ class HermesCLI:
         return f"[{('█' * filled) + ('░' * max(0, width - filled))}]"
 
     def _get_status_bar_snapshot(self) -> Dict[str, Any]:
-        model_name = self.model or "unknown"
+        # Prefer the agent's model name — it updates on fallback.
+        # self.model reflects the originally configured model and never
+        # changes mid-session, so the TUI would show a stale name after
+        # _try_activate_fallback() switches provider/model.
+        agent = getattr(self, "agent", None)
+        model_name = (getattr(agent, "model", None) or self.model or "unknown")
         model_short = model_name.split("/")[-1] if "/" in model_name else model_name
         if model_short.endswith(".gguf"):
             model_short = model_short[:-5]
@@ -1628,7 +1634,6 @@ class HermesCLI:
             "compressions": 0,
         }
 
-        agent = getattr(self, "agent", None)
         if not agent:
             return snapshot
 
@@ -4003,59 +4008,7 @@ class HermesCLI:
 
         print("  To change model or provider, use: hermes model")
 
-    def _handle_prompt_command(self, cmd: str):
-        """Handle the /prompt command to view or set system prompt."""
-        parts = cmd.split(maxsplit=1)
-        
-        if len(parts) > 1:
-            # Set new prompt
-            new_prompt = parts[1].strip()
-            
-            if new_prompt.lower() == "clear":
-                self.system_prompt = ""
-                self.agent = None  # Force re-init
-                if save_config_value("agent.system_prompt", ""):
-                    print("(^_^)b System prompt cleared (saved to config)")
-                else:
-                    print("(^_^) System prompt cleared (session only)")
-            else:
-                self.system_prompt = new_prompt
-                self.agent = None  # Force re-init
-                if save_config_value("agent.system_prompt", new_prompt):
-                    print("(^_^)b System prompt set (saved to config)")
-                else:
-                    print("(^_^) System prompt set (session only)")
-                print(f"  \"{new_prompt[:60]}{'...' if len(new_prompt) > 60 else ''}\"")
-        else:
-            # Show current prompt
-            print()
-            print("+" + "-" * 50 + "+")
-            print("|" + " " * 15 + "(^_^) System Prompt" + " " * 15 + "|")
-            print("+" + "-" * 50 + "+")
-            print()
-            if self.system_prompt:
-                # Word wrap the prompt for display
-                words = self.system_prompt.split()
-                lines = []
-                current_line = ""
-                for word in words:
-                    if len(current_line) + len(word) + 1 <= 50:
-                        current_line += (" " if current_line else "") + word
-                    else:
-                        lines.append(current_line)
-                        current_line = word
-                if current_line:
-                    lines.append(current_line)
-                for line in lines:
-                    print(f"  {line}")
-            else:
-                print("  (no custom prompt set - using default)")
-            print()
-            print("  Usage:")
-            print("    /prompt <text>  - Set a custom system prompt")
-            print("    /prompt clear   - Remove custom prompt")
-            print("    /personality    - Use a predefined personality")
-            print()
+
     
 
     @staticmethod
@@ -4555,9 +4508,7 @@ class HermesCLI:
             self._handle_model_switch(cmd_original)
         elif canonical == "provider":
             self._show_model_and_providers()
-        elif canonical == "prompt":
-            # Use original case so prompt text isn't lowercased
-            self._handle_prompt_command(cmd_original)
+
         elif canonical == "personality":
             # Use original case (handler lowercases the personality name itself)
             self._handle_personality_command(cmd_original)
@@ -4668,13 +4619,13 @@ class HermesCLI:
                             if output:
                                 self.console.print(_rich_text_from_ansi(output))
                             else:
-                                ChatConsole().print("[dim]Command returned no output[/]")
+                                self.console.print("[dim]Command returned no output[/]")
                         except subprocess.TimeoutExpired:
-                            ChatConsole().print("[bold red]Quick command timed out (30s)[/]")
+                            self.console.print("[bold red]Quick command timed out (30s)[/]")
                         except Exception as e:
-                            ChatConsole().print(f"[bold red]Quick command error: {e}[/]")
+                            self.console.print(f"[bold red]Quick command error: {e}[/]")
                     else:
-                        ChatConsole().print(f"[bold red]Quick command '{base_cmd}' has no command defined[/]")
+                        self.console.print(f"[bold red]Quick command '{base_cmd}' has no command defined[/]")
                 elif qcmd.get("type") == "alias":
                     target = qcmd.get("target", "").strip()
                     if target:
@@ -4683,9 +4634,9 @@ class HermesCLI:
                         aliased_command = f"{target} {user_args}".strip()
                         return self.process_command(aliased_command)
                     else:
-                        ChatConsole().print(f"[bold red]Quick command '{base_cmd}' has no target defined[/]")
+                        self.console.print(f"[bold red]Quick command '{base_cmd}' has no target defined[/]")
                 else:
-                    ChatConsole().print(f"[bold red]Quick command '{base_cmd}' has unsupported type (supported: 'exec', 'alias')[/]")
+                    self.console.print(f"[bold red]Quick command '{base_cmd}' has unsupported type (supported: 'exec', 'alias')[/]")
             # Check for plugin-registered slash commands
             elif base_cmd.lstrip("/") in _get_plugin_cmd_handler_names():
                 from hermes_cli.plugins import get_plugin_command_handler
@@ -5408,12 +5359,27 @@ class HermesCLI:
             print(f"  ❌ Compression failed: {e}")
 
     def _show_usage(self):
-        """Show cumulative token usage for the current session."""
+        """Show rate limits (if available) and session token usage."""
         if not self.agent:
             print("(._.) No active agent -- send a message first.")
             return
 
         agent = self.agent
+        calls = agent.session_api_calls
+
+        if calls == 0:
+            print("(._.) No API calls made yet in this session.")
+            return
+
+        # ── Rate limits (shown first when available) ────────────────
+        rl_state = agent.get_rate_limit_state()
+        if rl_state and rl_state.has_data:
+            from agent.rate_limit_tracker import format_rate_limit_display
+            print()
+            print(format_rate_limit_display(rl_state))
+            print()
+
+        # ── Session token usage ─────────────────────────────────────
         input_tokens = getattr(agent, "session_input_tokens", 0) or 0
         output_tokens = getattr(agent, "session_output_tokens", 0) or 0
         cache_read_tokens = getattr(agent, "session_cache_read_tokens", 0) or 0
@@ -5421,13 +5387,7 @@ class HermesCLI:
         prompt = agent.session_prompt_tokens
         completion = agent.session_completion_tokens
         total = agent.session_total_tokens
-        calls = agent.session_api_calls
 
-        if calls == 0:
-            print("(._.) No API calls made yet in this session.")
-            return
-
-        # Current context window state
         compressor = agent.context_compressor
         last_prompt = compressor.last_prompt_tokens
         ctx_len = compressor.context_length
@@ -6205,6 +6165,7 @@ class HermesCLI:
         timeout = 45
         response_queue = queue.Queue()
 
+        self._capture_modal_input_snapshot()
         self._sudo_state = {
             "response_queue": response_queue,
         }
@@ -6217,6 +6178,7 @@ class HermesCLI:
                 result = response_queue.get(timeout=1)
                 self._sudo_state = None
                 self._sudo_deadline = 0
+                self._restore_modal_input_snapshot()
                 self._invalidate()
                 if result:
                     _cprint(f"\n{_DIM}  ✓ Password received (cached for session){_RST}")
@@ -6231,6 +6193,7 @@ class HermesCLI:
 
         self._sudo_state = None
         self._sudo_deadline = 0
+        self._restore_modal_input_snapshot()
         self._invalidate()
         _cprint(f"\n{_DIM}  ⏱ Timeout — continuing without sudo{_RST}")
         return ""
@@ -6403,6 +6366,33 @@ class HermesCLI:
     def _secret_capture_callback(self, var_name: str, prompt: str, metadata=None) -> dict:
         return prompt_for_secret(self, var_name, prompt, metadata)
 
+    def _capture_modal_input_snapshot(self) -> None:
+        """Temporarily clear the input buffer and save the user's in-progress draft."""
+        if self._modal_input_snapshot is not None or not getattr(self, "_app", None):
+            return
+        try:
+            buf = self._app.current_buffer
+            self._modal_input_snapshot = {
+                "text": buf.text,
+                "cursor_position": buf.cursor_position,
+            }
+            buf.reset()
+        except Exception:
+            self._modal_input_snapshot = None
+
+    def _restore_modal_input_snapshot(self) -> None:
+        """Restore any draft text that was present before a modal prompt opened."""
+        snapshot = self._modal_input_snapshot
+        self._modal_input_snapshot = None
+        if not snapshot or not getattr(self, "_app", None):
+            return
+        try:
+            buf = self._app.current_buffer
+            buf.text = snapshot.get("text", "")
+            buf.cursor_position = min(snapshot.get("cursor_position", 0), len(buf.text))
+        except Exception:
+            pass
+
     def _submit_secret_response(self, value: str) -> None:
         if not self._secret_state:
             return
@@ -7130,6 +7120,7 @@ class HermesCLI:
         # Sudo password prompt state (similar mechanism to clarify)
         self._sudo_state = None         # dict with response_queue when active
         self._sudo_deadline = 0
+        self._modal_input_snapshot = None
 
         # Dangerous command approval state (similar mechanism to clarify)
         self._approval_state = None     # dict with command, description, choices, selected, response_queue
@@ -7201,7 +7192,6 @@ class HermesCLI:
                 text = event.app.current_buffer.text
                 self._sudo_state["response_queue"].put(text)
                 self._sudo_state = None
-                event.app.current_buffer.reset()
                 event.app.invalidate()
                 return
 
@@ -7406,7 +7396,6 @@ class HermesCLI:
             if self._sudo_state:
                 self._sudo_state["response_queue"].put("")
                 self._sudo_state = None
-                event.app.current_buffer.reset()
                 event.app.invalidate()
                 return
 

cron/scheduler.py

@@ -44,7 +44,7 @@ logger = logging.getLogger(__name__)
 _KNOWN_DELIVERY_PLATFORMS = frozenset({
     "telegram", "discord", "slack", "whatsapp", "signal",
     "matrix", "mattermost", "homeassistant", "dingtalk", "feishu",
-    "wecom", "sms", "email", "webhook",
+    "wecom", "sms", "email", "webhook", "bluebubbles",
 })
 
 from cron.jobs import get_due_jobs, mark_job_run, save_job_output, advance_next_run
@@ -91,7 +91,7 @@ def _resolve_delivery_target(job: dict) -> Optional[dict]:
             }
         # Origin missing (e.g. job created via API/script) — try each
         # platform's home channel as a fallback instead of silently dropping.
-        for platform_name in ("matrix", "telegram", "discord", "slack"):
+        for platform_name in ("matrix", "telegram", "discord", "slack", "bluebubbles"):
             chat_id = os.getenv(f"{platform_name.upper()}_HOME_CHANNEL", "")
             if chat_id:
                 logger.info(
@@ -236,6 +236,7 @@ def _deliver_result(job: dict, content: str, adapters=None, loop=None) -> Option
         "wecom": Platform.WECOM,
         "email": Platform.EMAIL,
         "sms": Platform.SMS,
+        "bluebubbles": Platform.BLUEBUBBLES,
     }
     platform = platform_map.get(platform_name.lower())
     if not platform:

flake.lock

@@ -22,16 +22,16 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1751274312,
-        "narHash": "sha256-/bVBlRpECLVzjV19t5KMdMFWSwKLtb5RyXdjz3LJT+g=",
+        "lastModified": 1775036866,
+        "narHash": "sha256-ZojAnPuCdy657PbTq5V0Y+AHKhZAIwSIT2cb8UgAz/U=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "50ab793786d9de88ee30ec4e4c24fb4236fc2674",
+        "rev": "6201e203d09599479a3b3450ed24fa81537ebc4e",
         "type": "github"
       },
       "original": {
         "owner": "NixOS",
-        "ref": "nixos-24.11",
+        "ref": "nixos-unstable",
         "repo": "nixpkgs",
         "type": "github"
       }

flake.nix

@@ -2,7 +2,7 @@
   description = "Hermes Agent - AI agent framework by Nous Research";
 
   inputs = {
-    nixpkgs.url = "github:NixOS/nixpkgs/nixos-24.11";
+    nixpkgs.url = "github:NixOS/nixpkgs/nixos-unstable";
     flake-parts = {
       url = "github:hercules-ci/flake-parts";
       inputs.nixpkgs-lib.follows = "nixpkgs";

gateway/channel_directory.py

@@ -77,7 +77,7 @@ def build_channel_directory(adapters: Dict[Any, Any]) -> Dict[str, Any]:
             logger.warning("Channel directory: failed to build %s: %s", platform.value, e)
 
     # Telegram, WhatsApp & Signal can't enumerate chats -- pull from session history
-    for plat_name in ("telegram", "whatsapp", "signal", "email", "sms"):
+    for plat_name in ("telegram", "whatsapp", "signal", "email", "sms", "bluebubbles"):
         if plat_name not in platforms:
             platforms[plat_name] = _build_from_sessions(plat_name)
 

gateway/config.py

@@ -63,6 +63,7 @@ class Platform(Enum):
     WEBHOOK = "webhook"
     FEISHU = "feishu"
     WECOM = "wecom"
+    BLUEBUBBLES = "bluebubbles"
 
 
 @dataclass
@@ -287,6 +288,9 @@ class GatewayConfig:
             # WeCom uses extra dict for bot credentials
             elif platform == Platform.WECOM and config.extra.get("bot_id"):
                 connected.append(platform)
+            # BlueBubbles uses extra dict for local server config
+            elif platform == Platform.BLUEBUBBLES and config.extra.get("server_url") and config.extra.get("password"):
+                connected.append(platform)
         return connected
     
     def get_home_channel(self, platform: Platform) -> Optional[HomeChannel]:
@@ -712,6 +716,13 @@ def _apply_env_overrides(config: GatewayConfig) -> None:
             name=os.getenv("DISCORD_HOME_CHANNEL_NAME", "Home"),
         )
     
+    # Reply threading mode for Discord (off/first/all)
+    discord_reply_mode = os.getenv("DISCORD_REPLY_TO_MODE", "").lower()
+    if discord_reply_mode in ("off", "first", "all"):
+        if Platform.DISCORD not in config.platforms:
+            config.platforms[Platform.DISCORD] = PlatformConfig()
+        config.platforms[Platform.DISCORD].reply_to_mode = discord_reply_mode
+    
     # WhatsApp (typically uses different auth mechanism)
     whatsapp_enabled = os.getenv("WHATSAPP_ENABLED", "").lower() in ("true", "1", "yes")
     if whatsapp_enabled:
@@ -941,6 +952,29 @@ def _apply_env_overrides(config: GatewayConfig) -> None:
                 name=os.getenv("WECOM_HOME_CHANNEL_NAME", "Home"),
             )
 
+    # BlueBubbles (iMessage)
+    bluebubbles_server_url = os.getenv("BLUEBUBBLES_SERVER_URL")
+    bluebubbles_password = os.getenv("BLUEBUBBLES_PASSWORD")
+    if bluebubbles_server_url and bluebubbles_password:
+        if Platform.BLUEBUBBLES not in config.platforms:
+            config.platforms[Platform.BLUEBUBBLES] = PlatformConfig()
+        config.platforms[Platform.BLUEBUBBLES].enabled = True
+        config.platforms[Platform.BLUEBUBBLES].extra.update({
+            "server_url": bluebubbles_server_url.rstrip("/"),
+            "password": bluebubbles_password,
+            "webhook_host": os.getenv("BLUEBUBBLES_WEBHOOK_HOST", "127.0.0.1"),
+            "webhook_port": int(os.getenv("BLUEBUBBLES_WEBHOOK_PORT", "8645")),
+            "webhook_path": os.getenv("BLUEBUBBLES_WEBHOOK_PATH", "/bluebubbles-webhook"),
+            "send_read_receipts": os.getenv("BLUEBUBBLES_SEND_READ_RECEIPTS", "true").lower() in ("true", "1", "yes"),
+        })
+    bluebubbles_home = os.getenv("BLUEBUBBLES_HOME_CHANNEL")
+    if bluebubbles_home and Platform.BLUEBUBBLES in config.platforms:
+        config.platforms[Platform.BLUEBUBBLES].home_channel = HomeChannel(
+            platform=Platform.BLUEBUBBLES,
+            chat_id=bluebubbles_home,
+            name=os.getenv("BLUEBUBBLES_HOME_CHANNEL_NAME", "Home"),
+        )
+
     # Session settings
     idle_minutes = os.getenv("SESSION_IDLE_MINUTES")
     if idle_minutes:

gateway/platforms/base.py

@@ -298,6 +298,7 @@ SUPPORTED_DOCUMENT_TYPES = {
     ".pdf": "application/pdf",
     ".md": "text/markdown",
     ".txt": "text/plain",
+    ".log": "text/plain",
     ".zip": "application/zip",
     ".docx": "application/vnd.openxmlformats-officedocument.wordprocessingml.document",
     ".xlsx": "application/vnd.openxmlformats-officedocument.spreadsheetml.sheet",
@@ -407,6 +408,10 @@ class MessageEvent:
     # Auto-loaded skill for topic/channel bindings (e.g., Telegram DM Topics)
     auto_skill: Optional[str] = None
     
+    # Internal flag — set for synthetic events (e.g. background process
+    # completion notifications) that must bypass user authorization checks.
+    internal: bool = False
+
     # Timestamps
     timestamp: datetime = field(default_factory=datetime.now)
     

gateway/platforms/bluebubbles.py

@@ -0,0 +1,828 @@
+"""BlueBubbles iMessage platform adapter.
+
+Uses the local BlueBubbles macOS server for outbound REST sends and inbound
+webhooks.  Supports text messaging, media attachments (images, voice, video,
+documents), tapback reactions, typing indicators, and read receipts.
+
+Architecture based on PR #5869 (benjaminsehl) with inbound attachment
+downloading from PR #4588 (YuhangLin).
+"""
+
+import asyncio
+import json
+import logging
+import os
+import re
+import uuid
+from datetime import datetime
+from typing import Any, Dict, List, Optional
+from urllib.parse import quote
+
+import httpx
+
+from gateway.config import Platform, PlatformConfig
+from gateway.platforms.base import (
+    BasePlatformAdapter,
+    MessageEvent,
+    MessageType,
+    SendResult,
+    cache_image_from_bytes,
+    cache_audio_from_bytes,
+    cache_document_from_bytes,
+)
+
+logger = logging.getLogger(__name__)
+
+# ---------------------------------------------------------------------------
+# Constants
+# ---------------------------------------------------------------------------
+
+DEFAULT_WEBHOOK_HOST = "127.0.0.1"
+DEFAULT_WEBHOOK_PORT = 8645
+DEFAULT_WEBHOOK_PATH = "/bluebubbles-webhook"
+MAX_TEXT_LENGTH = 4000
+
+# Tapback reaction codes (BlueBubbles associatedMessageType values)
+_TAPBACK_ADDED = {
+    2000: "love", 2001: "like", 2002: "dislike",
+    2003: "laugh", 2004: "emphasize", 2005: "question",
+}
+_TAPBACK_REMOVED = {
+    3000: "love", 3001: "like", 3002: "dislike",
+    3003: "laugh", 3004: "emphasize", 3005: "question",
+}
+
+# Webhook event types that carry user messages
+_MESSAGE_EVENTS = {"new-message", "message", "updated-message"}
+
+# Log redaction patterns
+_PHONE_RE = re.compile(r"\+?\d{7,15}")
+_EMAIL_RE = re.compile(r"[\w.+-]+@[\w-]+\.[\w.]+")
+
+
+def _redact(text: str) -> str:
+    """Redact phone numbers and emails from log output."""
+    text = _PHONE_RE.sub("[REDACTED]", text)
+    text = _EMAIL_RE.sub("[REDACTED]", text)
+    return text
+
+
+# ---------------------------------------------------------------------------
+# Helpers
+# ---------------------------------------------------------------------------
+
+def check_bluebubbles_requirements() -> bool:
+    try:
+        import aiohttp  # noqa: F401
+        import httpx as _httpx  # noqa: F401
+    except ImportError:
+        return False
+    return True
+
+
+def _normalize_server_url(raw: str) -> str:
+    value = (raw or "").strip()
+    if not value:
+        return ""
+    if not re.match(r"^https?://", value, flags=re.I):
+        value = f"http://{value}"
+    return value.rstrip("/")
+
+
+def _strip_markdown(text: str) -> str:
+    """Strip common markdown formatting for iMessage plain-text delivery."""
+    text = re.sub(r"\*\*(.+?)\*\*", r"\1", text, flags=re.DOTALL)
+    text = re.sub(r"\*(.+?)\*", r"\1", text, flags=re.DOTALL)
+    text = re.sub(r"__(.+?)__", r"\1", text, flags=re.DOTALL)
+    text = re.sub(r"_(.+?)_", r"\1", text, flags=re.DOTALL)
+    text = re.sub(r"```[a-zA-Z0-9_+-]*\n?", "", text)
+    text = re.sub(r"`(.+?)`", r"\1", text)
+    text = re.sub(r"^#{1,6}\s+", "", text, flags=re.MULTILINE)
+    text = re.sub(r"\[([^\]]+)\]\(([^\)]+)\)", r"\1", text)
+    text = re.sub(r"\n{3,}", "\n\n", text)
+    return text.strip()
+
+
+# ---------------------------------------------------------------------------
+# Adapter
+# ---------------------------------------------------------------------------
+
+class BlueBubblesAdapter(BasePlatformAdapter):
+    platform = Platform.BLUEBUBBLES
+    MAX_MESSAGE_LENGTH = MAX_TEXT_LENGTH
+
+    def __init__(self, config: PlatformConfig):
+        super().__init__(config, Platform.BLUEBUBBLES)
+        extra = config.extra or {}
+        self.server_url = _normalize_server_url(
+            extra.get("server_url") or os.getenv("BLUEBUBBLES_SERVER_URL", "")
+        )
+        self.password = extra.get("password") or os.getenv("BLUEBUBBLES_PASSWORD", "")
+        self.webhook_host = (
+            extra.get("webhook_host")
+            or os.getenv("BLUEBUBBLES_WEBHOOK_HOST", DEFAULT_WEBHOOK_HOST)
+        )
+        self.webhook_port = int(
+            extra.get("webhook_port")
+            or os.getenv("BLUEBUBBLES_WEBHOOK_PORT", str(DEFAULT_WEBHOOK_PORT))
+        )
+        self.webhook_path = (
+            extra.get("webhook_path")
+            or os.getenv("BLUEBUBBLES_WEBHOOK_PATH", DEFAULT_WEBHOOK_PATH)
+        )
+        if not str(self.webhook_path).startswith("/"):
+            self.webhook_path = f"/{self.webhook_path}"
+        self.send_read_receipts = bool(extra.get("send_read_receipts", True))
+        self.client: Optional[httpx.AsyncClient] = None
+        self._runner = None
+        self._private_api_enabled: Optional[bool] = None
+        self._helper_connected: bool = False
+        self._guid_cache: Dict[str, str] = {}
+
+    # ------------------------------------------------------------------
+    # API helpers
+    # ------------------------------------------------------------------
+
+    def _api_url(self, path: str) -> str:
+        sep = "&" if "?" in path else "?"
+        return f"{self.server_url}{path}{sep}password={quote(self.password, safe='')}"
+
+    async def _api_get(self, path: str) -> Dict[str, Any]:
+        assert self.client is not None
+        res = await self.client.get(self._api_url(path))
+        res.raise_for_status()
+        return res.json()
+
+    async def _api_post(self, path: str, payload: Dict[str, Any]) -> Dict[str, Any]:
+        assert self.client is not None
+        res = await self.client.post(self._api_url(path), json=payload)
+        res.raise_for_status()
+        return res.json()
+
+    # ------------------------------------------------------------------
+    # Lifecycle
+    # ------------------------------------------------------------------
+
+    async def connect(self) -> bool:
+        if not self.server_url or not self.password:
+            logger.error(
+                "[bluebubbles] BLUEBUBBLES_SERVER_URL and BLUEBUBBLES_PASSWORD are required"
+            )
+            return False
+        from aiohttp import web
+
+        self.client = httpx.AsyncClient(timeout=30.0)
+        try:
+            await self._api_get("/api/v1/ping")
+            info = await self._api_get("/api/v1/server/info")
+            server_data = (info or {}).get("data", {})
+            self._private_api_enabled = bool(server_data.get("private_api"))
+            self._helper_connected = bool(server_data.get("helper_connected"))
+            logger.info(
+                "[bluebubbles] connected to %s (private_api=%s, helper=%s)",
+                self.server_url,
+                self._private_api_enabled,
+                self._helper_connected,
+            )
+        except Exception as exc:
+            logger.error(
+                "[bluebubbles] cannot reach server at %s: %s", self.server_url, exc
+            )
+            if self.client:
+                await self.client.aclose()
+                self.client = None
+            return False
+
+        app = web.Application()
+        app.router.add_get("/health", lambda _: web.Response(text="ok"))
+        app.router.add_post(self.webhook_path, self._handle_webhook)
+        self._runner = web.AppRunner(app)
+        await self._runner.setup()
+        site = web.TCPSite(self._runner, self.webhook_host, self.webhook_port)
+        await site.start()
+        self._mark_connected()
+        logger.info(
+            "[bluebubbles] webhook listening on http://%s:%s%s",
+            self.webhook_host,
+            self.webhook_port,
+            self.webhook_path,
+        )
+        return True
+
+    async def disconnect(self) -> None:
+        if self.client:
+            await self.client.aclose()
+            self.client = None
+        if self._runner:
+            await self._runner.cleanup()
+            self._runner = None
+        self._mark_disconnected()
+
+    # ------------------------------------------------------------------
+    # Chat GUID resolution
+    # ------------------------------------------------------------------
+
+    async def _resolve_chat_guid(self, target: str) -> Optional[str]:
+        """Resolve an email/phone to a BlueBubbles chat GUID.
+
+        If *target* already contains a semicolon (raw GUID format like
+        ``iMessage;-;user@example.com``), it is returned as-is.  Otherwise
+        the adapter queries the BlueBubbles chat list and matches on
+        ``chatIdentifier`` or participant address.
+        """
+        target = (target or "").strip()
+        if not target:
+            return None
+        # Already a raw GUID
+        if ";" in target:
+            return target
+        if target in self._guid_cache:
+            return self._guid_cache[target]
+        try:
+            payload = await self._api_post(
+                "/api/v1/chat/query",
+                {"limit": 100, "offset": 0, "with": ["participants"]},
+            )
+            for chat in payload.get("data", []) or []:
+                guid = chat.get("guid") or chat.get("chatGuid")
+                identifier = chat.get("chatIdentifier") or chat.get("identifier")
+                if identifier == target:
+                    if guid:
+                        self._guid_cache[target] = guid
+                    return guid
+                for part in chat.get("participants", []) or []:
+                    if (part.get("address") or "").strip() == target and guid:
+                        self._guid_cache[target] = guid
+                        return guid
+        except Exception:
+            pass
+        return None
+
+    async def _create_chat_for_handle(
+        self, address: str, message: str
+    ) -> SendResult:
+        """Create a new chat by sending the first message to *address*."""
+        payload = {
+            "addresses": [address],
+            "message": message,
+            "tempGuid": f"temp-{datetime.utcnow().timestamp()}",
+        }
+        try:
+            res = await self._api_post("/api/v1/chat/new", payload)
+            data = res.get("data") or {}
+            msg_id = data.get("guid") or data.get("messageGuid") or "ok"
+            return SendResult(success=True, message_id=str(msg_id), raw_response=res)
+        except Exception as exc:
+            return SendResult(success=False, error=str(exc))
+
+    # ------------------------------------------------------------------
+    # Text sending
+    # ------------------------------------------------------------------
+
+    async def send(
+        self,
+        chat_id: str,
+        content: str,
+        reply_to: Optional[str] = None,
+        metadata: Optional[Dict[str, Any]] = None,
+    ) -> SendResult:
+        text = _strip_markdown(content or "")
+        if not text:
+            return SendResult(success=False, error="BlueBubbles send requires text")
+        chunks = self.truncate_message(text, max_length=self.MAX_MESSAGE_LENGTH)
+        last = SendResult(success=True)
+        for chunk in chunks:
+            guid = await self._resolve_chat_guid(chat_id)
+            if not guid:
+                # If the target looks like an address, try creating a new chat
+                if self._private_api_enabled and (
+                    "@" in chat_id or re.match(r"^\+\d+", chat_id)
+                ):
+                    return await self._create_chat_for_handle(chat_id, chunk)
+                return SendResult(
+                    success=False,
+                    error=f"BlueBubbles chat not found for target: {chat_id}",
+                )
+            payload: Dict[str, Any] = {
+                "chatGuid": guid,
+                "tempGuid": f"temp-{datetime.utcnow().timestamp()}",
+                "message": chunk,
+            }
+            if reply_to and self._private_api_enabled and self._helper_connected:
+                payload["method"] = "private-api"
+                payload["selectedMessageGuid"] = reply_to
+                payload["partIndex"] = 0
+            try:
+                res = await self._api_post("/api/v1/message/text", payload)
+                data = res.get("data") or {}
+                msg_id = data.get("guid") or data.get("messageGuid") or "ok"
+                last = SendResult(
+                    success=True, message_id=str(msg_id), raw_response=res
+                )
+            except Exception as exc:
+                return SendResult(success=False, error=str(exc))
+        return last
+
+    # ------------------------------------------------------------------
+    # Media sending (outbound)
+    # ------------------------------------------------------------------
+
+    async def _send_attachment(
+        self,
+        chat_id: str,
+        file_path: str,
+        filename: Optional[str] = None,
+        caption: Optional[str] = None,
+        is_audio_message: bool = False,
+    ) -> SendResult:
+        """Send a file attachment via BlueBubbles multipart upload."""
+        if not self.client:
+            return SendResult(success=False, error="Not connected")
+        if not os.path.isfile(file_path):
+            return SendResult(success=False, error=f"File not found: {file_path}")
+
+        guid = await self._resolve_chat_guid(chat_id)
+        if not guid:
+            return SendResult(success=False, error=f"Chat not found: {chat_id}")
+
+        fname = filename or os.path.basename(file_path)
+        try:
+            with open(file_path, "rb") as f:
+                files = {"attachment": (fname, f, "application/octet-stream")}
+                data: Dict[str, str] = {
+                    "chatGuid": guid,
+                    "name": fname,
+                    "tempGuid": uuid.uuid4().hex,
+                }
+                if is_audio_message:
+                    data["isAudioMessage"] = "true"
+                res = await self.client.post(
+                    self._api_url("/api/v1/message/attachment"),
+                    files=files,
+                    data=data,
+                    timeout=120,
+                )
+                res.raise_for_status()
+                result = res.json()
+
+            if caption:
+                await self.send(chat_id, caption)
+
+            if result.get("status") == 200:
+                rdata = result.get("data") or {}
+                msg_id = rdata.get("guid") if isinstance(rdata, dict) else None
+                return SendResult(
+                    success=True, message_id=msg_id, raw_response=result
+                )
+            return SendResult(
+                success=False,
+                error=result.get("message", "Attachment upload failed"),
+            )
+        except Exception as e:
+            return SendResult(success=False, error=str(e))
+
+    async def send_image(
+        self,
+        chat_id: str,
+        image_url: str,
+        caption: Optional[str] = None,
+        reply_to: Optional[str] = None,
+        metadata: Optional[Dict[str, Any]] = None,
+    ) -> SendResult:
+        try:
+            from gateway.platforms.base import cache_image_from_url
+
+            local_path = await cache_image_from_url(image_url)
+            return await self._send_attachment(chat_id, local_path, caption=caption)
+        except Exception:
+            return await super().send_image(chat_id, image_url, caption, reply_to)
+
+    async def send_image_file(
+        self,
+        chat_id: str,
+        image_path: str,
+        caption: Optional[str] = None,
+        reply_to: Optional[str] = None,
+        **kwargs,
+    ) -> SendResult:
+        return await self._send_attachment(chat_id, image_path, caption=caption)
+
+    async def send_voice(
+        self,
+        chat_id: str,
+        audio_path: str,
+        caption: Optional[str] = None,
+        reply_to: Optional[str] = None,
+        **kwargs,
+    ) -> SendResult:
+        return await self._send_attachment(
+            chat_id, audio_path, caption=caption, is_audio_message=True
+        )
+
+    async def send_video(
+        self,
+        chat_id: str,
+        video_path: str,
+        caption: Optional[str] = None,
+        reply_to: Optional[str] = None,
+        **kwargs,
+    ) -> SendResult:
+        return await self._send_attachment(chat_id, video_path, caption=caption)
+
+    async def send_document(
+        self,
+        chat_id: str,
+        file_path: str,
+        caption: Optional[str] = None,
+        file_name: Optional[str] = None,
+        reply_to: Optional[str] = None,
+        **kwargs,
+    ) -> SendResult:
+        return await self._send_attachment(
+            chat_id, file_path, filename=file_name, caption=caption
+        )
+
+    async def send_animation(
+        self,
+        chat_id: str,
+        animation_url: str,
+        caption: Optional[str] = None,
+        reply_to: Optional[str] = None,
+        metadata: Optional[Dict[str, Any]] = None,
+    ) -> SendResult:
+        return await self.send_image(
+            chat_id, animation_url, caption, reply_to, metadata
+        )
+
+    # ------------------------------------------------------------------
+    # Typing indicators
+    # ------------------------------------------------------------------
+
+    async def send_typing(self, chat_id: str, metadata=None) -> None:
+        if not self._private_api_enabled or not self._helper_connected or not self.client:
+            return
+        try:
+            guid = await self._resolve_chat_guid(chat_id)
+            if guid:
+                encoded = quote(guid, safe="")
+                await self.client.post(
+                    self._api_url(f"/api/v1/chat/{encoded}/typing"), timeout=5
+                )
+        except Exception:
+            pass
+
+    async def stop_typing(self, chat_id: str) -> None:
+        if not self._private_api_enabled or not self._helper_connected or not self.client:
+            return
+        try:
+            guid = await self._resolve_chat_guid(chat_id)
+            if guid:
+                encoded = quote(guid, safe="")
+                await self.client.delete(
+                    self._api_url(f"/api/v1/chat/{encoded}/typing"), timeout=5
+                )
+        except Exception:
+            pass
+
+    # ------------------------------------------------------------------
+    # Read receipts
+    # ------------------------------------------------------------------
+
+    async def mark_read(self, chat_id: str) -> bool:
+        if not self._private_api_enabled or not self._helper_connected or not self.client:
+            return False
+        try:
+            guid = await self._resolve_chat_guid(chat_id)
+            if guid:
+                encoded = quote(guid, safe="")
+                await self.client.post(
+                    self._api_url(f"/api/v1/chat/{encoded}/read"), timeout=5
+                )
+                return True
+        except Exception:
+            pass
+        return False
+
+    # ------------------------------------------------------------------
+    # Tapback reactions
+    # ------------------------------------------------------------------
+
+    async def send_reaction(
+        self,
+        chat_id: str,
+        message_guid: str,
+        reaction: str,
+        part_index: int = 0,
+    ) -> SendResult:
+        """Send a tapback reaction (requires Private API helper)."""
+        if not self._private_api_enabled or not self._helper_connected:
+            return SendResult(
+                success=False, error="Private API helper not connected"
+            )
+        guid = await self._resolve_chat_guid(chat_id)
+        if not guid:
+            return SendResult(success=False, error=f"Chat not found: {chat_id}")
+        try:
+            res = await self._api_post(
+                "/api/v1/message/react",
+                {
+                    "chatGuid": guid,
+                    "selectedMessageGuid": message_guid,
+                    "reaction": reaction,
+                    "partIndex": part_index,
+                },
+            )
+            return SendResult(success=True, raw_response=res)
+        except Exception as exc:
+            return SendResult(success=False, error=str(exc))
+
+    # ------------------------------------------------------------------
+    # Chat info
+    # ------------------------------------------------------------------
+
+    async def get_chat_info(self, chat_id: str) -> Dict[str, Any]:
+        is_group = ";+;" in (chat_id or "")
+        info: Dict[str, Any] = {
+            "name": chat_id,
+            "type": "group" if is_group else "dm",
+        }
+        try:
+            guid = await self._resolve_chat_guid(chat_id)
+            if guid:
+                encoded = quote(guid, safe="")
+                res = await self._api_get(
+                    f"/api/v1/chat/{encoded}?with=participants"
+                )
+                data = (res or {}).get("data", {})
+                display_name = (
+                    data.get("displayName")
+                    or data.get("chatIdentifier")
+                    or chat_id
+                )
+                participants = []
+                for p in data.get("participants", []) or []:
+                    addr = (p.get("address") or "").strip()
+                    if addr:
+                        participants.append(addr)
+                info["name"] = display_name
+                if participants:
+                    info["participants"] = participants
+        except Exception:
+            pass
+        return info
+
+    def format_message(self, content: str) -> str:
+        return _strip_markdown(content)
+
+    # ------------------------------------------------------------------
+    # Inbound attachment downloading (from #4588)
+    # ------------------------------------------------------------------
+
+    async def _download_attachment(
+        self, att_guid: str, att_meta: Dict[str, Any]
+    ) -> Optional[str]:
+        """Download an attachment from BlueBubbles and cache it locally.
+
+        Returns the local file path on success, None on failure.
+        """
+        if not self.client:
+            return None
+        try:
+            encoded = quote(att_guid, safe="")
+            resp = await self.client.get(
+                self._api_url(f"/api/v1/attachment/{encoded}/download"),
+                timeout=60,
+                follow_redirects=True,
+            )
+            resp.raise_for_status()
+            data = resp.content
+
+            mime = (att_meta.get("mimeType") or "").lower()
+            transfer_name = att_meta.get("transferName", "")
+
+            if mime.startswith("image/"):
+                ext_map = {
+                    "image/jpeg": ".jpg",
+                    "image/png": ".png",
+                    "image/gif": ".gif",
+                    "image/webp": ".webp",
+                    "image/heic": ".jpg",
+                    "image/heif": ".jpg",
+                    "image/tiff": ".jpg",
+                }
+                ext = ext_map.get(mime, ".jpg")
+                return cache_image_from_bytes(data, ext)
+
+            if mime.startswith("audio/"):
+                ext_map = {
+                    "audio/mp3": ".mp3",
+                    "audio/mpeg": ".mp3",
+                    "audio/ogg": ".ogg",
+                    "audio/wav": ".wav",
+                    "audio/x-caf": ".mp3",
+                    "audio/mp4": ".m4a",
+                    "audio/aac": ".m4a",
+                }
+                ext = ext_map.get(mime, ".mp3")
+                return cache_audio_from_bytes(data, ext)
+
+            # Videos, documents, and everything else
+            filename = transfer_name or f"file_{uuid.uuid4().hex[:8]}"
+            return cache_document_from_bytes(data, filename)
+
+        except Exception as exc:
+            logger.warning(
+                "[bluebubbles] failed to download attachment %s: %s",
+                _redact(att_guid),
+                exc,
+            )
+            return None
+
+    # ------------------------------------------------------------------
+    # Webhook handling
+    # ------------------------------------------------------------------
+
+    def _extract_payload_record(
+        self, payload: Dict[str, Any]
+    ) -> Optional[Dict[str, Any]]:
+        data = payload.get("data")
+        if isinstance(data, dict):
+            return data
+        if isinstance(data, list):
+            for item in data:
+                if isinstance(item, dict):
+                    return item
+        if isinstance(payload.get("message"), dict):
+            return payload.get("message")
+        return payload if isinstance(payload, dict) else None
+
+    @staticmethod
+    def _value(*candidates: Any) -> Optional[str]:
+        for candidate in candidates:
+            if isinstance(candidate, str) and candidate.strip():
+                return candidate.strip()
+        return None
+
+    async def _handle_webhook(self, request):
+        from aiohttp import web
+
+        token = (
+            request.query.get("password")
+            or request.query.get("guid")
+            or request.headers.get("x-password")
+            or request.headers.get("x-guid")
+            or request.headers.get("x-bluebubbles-guid")
+        )
+        if token != self.password:
+            return web.json_response({"error": "unauthorized"}, status=401)
+        try:
+            raw = await request.read()
+            body = raw.decode("utf-8", errors="replace")
+            try:
+                payload = json.loads(body)
+            except Exception:
+                from urllib.parse import parse_qs
+
+                form = parse_qs(body)
+                payload_str = (
+                    form.get("payload")
+                    or form.get("data")
+                    or form.get("message")
+                    or [""]
+                )[0]
+                payload = json.loads(payload_str) if payload_str else {}
+        except Exception as exc:
+            logger.error("[bluebubbles] webhook parse error: %s", exc)
+            return web.json_response({"error": "invalid payload"}, status=400)
+
+        event_type = self._value(payload.get("type"), payload.get("event")) or ""
+        # Only process message events; silently acknowledge everything else
+        if event_type and event_type not in _MESSAGE_EVENTS:
+            return web.Response(text="ok")
+
+        record = self._extract_payload_record(payload) or {}
+        is_from_me = bool(
+            record.get("isFromMe")
+            or record.get("fromMe")
+            or record.get("is_from_me")
+        )
+        if is_from_me:
+            return web.Response(text="ok")
+
+        # Skip tapback reactions delivered as messages
+        assoc_type = record.get("associatedMessageType")
+        if isinstance(assoc_type, int) and assoc_type in {
+            **_TAPBACK_ADDED,
+            **_TAPBACK_REMOVED,
+        }:
+            return web.Response(text="ok")
+
+        text = (
+            self._value(
+                record.get("text"), record.get("message"), record.get("body")
+            )
+            or ""
+        )
+
+        # --- Inbound attachment handling ---
+        attachments = record.get("attachments") or []
+        media_urls: List[str] = []
+        media_types: List[str] = []
+        msg_type = MessageType.TEXT
+
+        for att in attachments:
+            att_guid = att.get("guid", "")
+            if not att_guid:
+                continue
+            cached = await self._download_attachment(att_guid, att)
+            if cached:
+                mime = (att.get("mimeType") or "").lower()
+                media_urls.append(cached)
+                media_types.append(mime)
+                if mime.startswith("image/"):
+                    msg_type = MessageType.PHOTO
+                elif mime.startswith("audio/") or (att.get("uti") or "").endswith(
+                    "caf"
+                ):
+                    msg_type = MessageType.VOICE
+                elif mime.startswith("video/"):
+                    msg_type = MessageType.VIDEO
+                else:
+                    msg_type = MessageType.DOCUMENT
+
+        # With multiple attachments, prefer PHOTO if any images present
+        if len(media_urls) > 1:
+            mime_prefixes = {(m or "").split("/")[0] for m in media_types}
+            if "image" in mime_prefixes:
+                msg_type = MessageType.PHOTO
+
+        if not text and media_urls:
+            text = "(attachment)"
+        # --- End attachment handling ---
+
+        chat_guid = self._value(
+            record.get("chatGuid"),
+            payload.get("chatGuid"),
+            record.get("chat_guid"),
+            payload.get("chat_guid"),
+            payload.get("guid"),
+        )
+        chat_identifier = self._value(
+            record.get("chatIdentifier"),
+            record.get("identifier"),
+            payload.get("chatIdentifier"),
+            payload.get("identifier"),
+        )
+        sender = (
+            self._value(
+                record.get("handle", {}).get("address")
+                if isinstance(record.get("handle"), dict)
+                else None,
+                record.get("sender"),
+                record.get("from"),
+                record.get("address"),
+            )
+            or chat_identifier
+            or chat_guid
+        )
+        if not (chat_guid or chat_identifier) and sender:
+            chat_identifier = sender
+        if not sender or not (chat_guid or chat_identifier) or not text:
+            return web.json_response({"error": "missing message fields"}, status=400)
+
+        session_chat_id = chat_guid or chat_identifier
+        is_group = bool(record.get("isGroup")) or (";+;" in (chat_guid or ""))
+        source = self.build_source(
+            chat_id=session_chat_id,
+            chat_name=chat_identifier or sender,
+            chat_type="group" if is_group else "dm",
+            user_id=sender,
+            user_name=sender,
+            chat_id_alt=chat_identifier,
+        )
+        event = MessageEvent(
+            text=text,
+            message_type=msg_type,
+            source=source,
+            raw_message=payload,
+            message_id=self._value(
+                record.get("guid"),
+                record.get("messageGuid"),
+                record.get("id"),
+            ),
+            reply_to_message_id=self._value(
+                record.get("threadOriginatorGuid"),
+                record.get("associatedMessageGuid"),
+            ),
+            media_urls=media_urls,
+            media_types=media_types,
+        )
+        task = asyncio.create_task(self.handle_message(event))
+        self._background_tasks.add(task)
+        task.add_done_callback(self._background_tasks.discard)
+
+        # Fire-and-forget read receipt
+        if self.send_read_receipts and session_chat_id:
+            asyncio.create_task(self.mark_read(session_chat_id))
+
+        return web.Response(text="ok")

gateway/platforms/discord.py

@@ -455,6 +455,9 @@ class DiscordAdapter(BasePlatformAdapter):
         self._seen_messages: Dict[str, float] = {}
         self._SEEN_TTL = 300   # 5 minutes
         self._SEEN_MAX = 2000  # prune threshold
+        # Reply threading mode: "off" (no replies), "first" (reply on first
+        # chunk only, default), "all" (reply-reference on every chunk).
+        self._reply_to_mode: str = getattr(config, 'reply_to_mode', 'first') or 'first'
 
     async def connect(self) -> bool:
         """Connect to Discord and start receiving events."""
@@ -774,7 +777,7 @@ class DiscordAdapter(BasePlatformAdapter):
             message_ids = []
             reference = None
 
-            if reply_to:
+            if reply_to and self._reply_to_mode != "off":
                 try:
                     ref_msg = await channel.fetch_message(int(reply_to))
                     reference = ref_msg
@@ -782,7 +785,10 @@ class DiscordAdapter(BasePlatformAdapter):
                     logger.debug("Could not fetch reply-to message: %s", e)
 
             for i, chunk in enumerate(chunks):
-                chunk_reference = reference if i == 0 else None
+                if self._reply_to_mode == "all":
+                    chunk_reference = reference
+                else:  # "first" (default) or "off"
+                    chunk_reference = reference if i == 0 else None
                 try:
                     msg = await channel.send(
                         content=chunk,
@@ -1761,8 +1767,9 @@ class DiscordAdapter(BasePlatformAdapter):
             if hasattr(interaction.channel, "guild") and interaction.channel.guild:
                 chat_name = f"{interaction.channel.guild.name} / #{chat_name}"
 
-        # Get channel topic (if available)
-        chat_topic = getattr(interaction.channel, "topic", None)
+        # Get channel topic (if available).
+        # For forum threads, inherit the parent forum's topic.
+        chat_topic = self._get_effective_topic(interaction.channel, is_thread=is_thread)
 
         source = self.build_source(
             chat_id=str(interaction.channel_id),
@@ -1836,6 +1843,10 @@ class DiscordAdapter(BasePlatformAdapter):
 
         chat_name = f"{guild_name} / {thread_name}" if guild_name else thread_name
 
+        # Inherit forum topic when the thread was created inside a forum channel.
+        _chan = getattr(interaction, "channel", None)
+        chat_topic = self._get_effective_topic(_chan, is_thread=True) if _chan else None
+
         source = self.build_source(
             chat_id=thread_id,
             chat_name=chat_name,
@@ -1843,6 +1854,7 @@ class DiscordAdapter(BasePlatformAdapter):
             user_id=str(interaction.user.id),
             user_name=interaction.user.display_name,
             thread_id=thread_id,
+            chat_topic=chat_topic,
         )
 
         event = MessageEvent(
@@ -2128,6 +2140,15 @@ class DiscordAdapter(BasePlatformAdapter):
                 return True
         return False
 
+    def _get_effective_topic(self, channel: Any, is_thread: bool = False) -> Optional[str]:
+        """Return the channel topic, falling back to the parent forum's topic for forum threads."""
+        topic = getattr(channel, "topic", None)
+        if not topic and is_thread:
+            parent = getattr(channel, "parent", None)
+            if parent and self._is_forum_parent(parent):
+                topic = getattr(parent, "topic", None)
+        return topic
+
     def _format_thread_chat_name(self, thread: Any) -> str:
         """Build a readable chat name for thread-like Discord channels, including forum context when available."""
         thread_name = getattr(thread, "name", None) or str(getattr(thread, "id", "thread"))
@@ -2295,8 +2316,10 @@ class DiscordAdapter(BasePlatformAdapter):
             if hasattr(message.channel, "guild") and message.channel.guild:
                 chat_name = f"{message.channel.guild.name} / #{chat_name}"
 
-        # Get channel topic (if available - TextChannels have topics, DMs/threads don't)
-        chat_topic = getattr(message.channel, "topic", None)
+        # Get channel topic (if available - TextChannels have topics, DMs/threads don't).
+        # For threads whose parent is a forum channel, inherit the parent's topic
+        # so forum descriptions (e.g. project instructions) appear in the session context.
+        chat_topic = self._get_effective_topic(message.channel, is_thread=is_thread)
 
         # Build source
         source = self.build_source(
@@ -2359,7 +2382,7 @@ class DiscordAdapter(BasePlatformAdapter):
                         ext or "unknown", content_type,
                     )
                 else:
-                    MAX_DOC_BYTES = 20 * 1024 * 1024
+                    MAX_DOC_BYTES = 32 * 1024 * 1024
                     if att.size and att.size > MAX_DOC_BYTES:
                         logger.warning(
                             "[Discord] Document too large (%s bytes), skipping: %s",
@@ -2383,9 +2406,9 @@ class DiscordAdapter(BasePlatformAdapter):
                             media_urls.append(cached_path)
                             media_types.append(doc_mime)
                             logger.info("[Discord] Cached user document: %s", cached_path)
-                            # Inject text content for .txt/.md files (capped at 100 KB)
+                            # Inject text content for plain-text documents (capped at 100 KB)
                             MAX_TEXT_INJECT_BYTES = 100 * 1024
-                            if ext in (".md", ".txt") and len(raw_bytes) <= MAX_TEXT_INJECT_BYTES:
+                            if ext in (".md", ".txt", ".log") and len(raw_bytes) <= MAX_TEXT_INJECT_BYTES:
                                 try:
                                     text_content = raw_bytes.decode("utf-8")
                                     display_name = att.filename or f"document{ext}"

gateway/platforms/signal.py

@@ -647,7 +647,11 @@ class SignalAdapter(BasePlatformAdapter):
 
         if result is not None:
             self._track_sent_timestamp(result)
-            return SendResult(success=True)
+            # Use the timestamp from the RPC result as a pseudo message_id.
+            # Signal doesn't have real message IDs, but the stream consumer
+            # needs a truthy value to follow its edit→fallback path correctly.
+            _msg_id = str(result.get("timestamp", "")) if isinstance(result, dict) else None
+            return SendResult(success=True, message_id=_msg_id or None)
         return SendResult(success=False, error="RPC send failed")
 
     def _track_sent_timestamp(self, rpc_result) -> None:
@@ -837,6 +841,11 @@ class SignalAdapter(BasePlatformAdapter):
             except asyncio.CancelledError:
                 pass
 
+    async def stop_typing(self, chat_id: str) -> None:
+        """Public interface for stopping typing — called by base adapter's
+        _keep_typing finally block to clean up platform-level typing tasks."""
+        await self._stop_typing_indicator(chat_id)
+
     # ------------------------------------------------------------------
     # Chat Info
     # ------------------------------------------------------------------

gateway/platforms/slack.py

@@ -14,7 +14,7 @@ import logging
 import os
 import re
 import time
-from typing import Dict, Optional, Any
+from typing import Dict, Optional, Any, Tuple
 
 try:
     from slack_bolt.async_app import AsyncApp
@@ -95,6 +95,12 @@ class SlackAdapter(BasePlatformAdapter):
         # respond to ALL subsequent messages in that thread automatically.
         self._mentioned_threads: set = set()
         self._MENTIONED_THREADS_MAX = 5000
+        # Assistant thread metadata keyed by (channel_id, thread_ts). Slack's
+        # AI Assistant lifecycle events can arrive before/alongside message
+        # events, and they carry the user/thread identity needed for stable
+        # session + memory scoping.
+        self._assistant_threads: Dict[Tuple[str, str], Dict[str, str]] = {}
+        self._ASSISTANT_THREADS_MAX = 5000
 
     async def connect(self) -> bool:
         """Connect to Slack via Socket Mode."""
@@ -181,6 +187,14 @@ class SlackAdapter(BasePlatformAdapter):
             async def handle_app_mention(event, say):
                 pass
 
+            @self._app.event("assistant_thread_started")
+            async def handle_assistant_thread_started(event, say):
+                await self._handle_assistant_thread_lifecycle_event(event)
+
+            @self._app.event("assistant_thread_context_changed")
+            async def handle_assistant_thread_context_changed(event, say):
+                await self._handle_assistant_thread_lifecycle_event(event)
+
             # Register slash command handler
             @self._app.command("/hermes")
             async def handle_hermes_command(ack, command):
@@ -755,6 +769,135 @@ class SlackAdapter(BasePlatformAdapter):
 
     # ----- Internal handlers -----
 
+    def _assistant_thread_key(self, channel_id: str, thread_ts: str) -> Optional[Tuple[str, str]]:
+        """Return a stable cache key for Slack assistant thread metadata."""
+        if not channel_id or not thread_ts:
+            return None
+        return (str(channel_id), str(thread_ts))
+
+    def _extract_assistant_thread_metadata(self, event: dict) -> Dict[str, str]:
+        """Extract Slack Assistant thread identity data from an event payload."""
+        assistant_thread = event.get("assistant_thread") or {}
+        context = assistant_thread.get("context") or event.get("context") or {}
+
+        channel_id = (
+            assistant_thread.get("channel_id")
+            or event.get("channel")
+            or context.get("channel_id")
+            or ""
+        )
+        thread_ts = (
+            assistant_thread.get("thread_ts")
+            or event.get("thread_ts")
+            or event.get("message_ts")
+            or ""
+        )
+        user_id = (
+            assistant_thread.get("user_id")
+            or event.get("user")
+            or context.get("user_id")
+            or ""
+        )
+        team_id = (
+            event.get("team")
+            or event.get("team_id")
+            or assistant_thread.get("team_id")
+            or ""
+        )
+        context_channel_id = context.get("channel_id") or ""
+
+        return {
+            "channel_id": str(channel_id) if channel_id else "",
+            "thread_ts": str(thread_ts) if thread_ts else "",
+            "user_id": str(user_id) if user_id else "",
+            "team_id": str(team_id) if team_id else "",
+            "context_channel_id": str(context_channel_id) if context_channel_id else "",
+        }
+
+    def _cache_assistant_thread_metadata(self, metadata: Dict[str, str]) -> None:
+        """Remember assistant thread identity data for later message events."""
+        channel_id = metadata.get("channel_id", "")
+        thread_ts = metadata.get("thread_ts", "")
+        key = self._assistant_thread_key(channel_id, thread_ts)
+        if not key:
+            return
+
+        existing = self._assistant_threads.get(key, {})
+        merged = dict(existing)
+        merged.update({k: v for k, v in metadata.items() if v})
+        self._assistant_threads[key] = merged
+
+        # Evict oldest entries when the cache exceeds the limit
+        if len(self._assistant_threads) > self._ASSISTANT_THREADS_MAX:
+            excess = len(self._assistant_threads) - self._ASSISTANT_THREADS_MAX // 2
+            for old_key in list(self._assistant_threads)[:excess]:
+                del self._assistant_threads[old_key]
+
+        team_id = merged.get("team_id", "")
+        if team_id and channel_id:
+            self._channel_team[channel_id] = team_id
+
+    def _lookup_assistant_thread_metadata(
+        self,
+        event: dict,
+        channel_id: str = "",
+        thread_ts: str = "",
+    ) -> Dict[str, str]:
+        """Load cached assistant-thread metadata that matches the current event."""
+        metadata = self._extract_assistant_thread_metadata(event)
+        if channel_id and not metadata.get("channel_id"):
+            metadata["channel_id"] = channel_id
+        if thread_ts and not metadata.get("thread_ts"):
+            metadata["thread_ts"] = thread_ts
+
+        key = self._assistant_thread_key(
+            metadata.get("channel_id", ""),
+            metadata.get("thread_ts", ""),
+        )
+        cached = self._assistant_threads.get(key, {}) if key else {}
+        if cached:
+            merged = dict(cached)
+            merged.update({k: v for k, v in metadata.items() if v})
+            return merged
+        return metadata
+
+    def _seed_assistant_thread_session(self, metadata: Dict[str, str]) -> None:
+        """Prime the session store so assistant threads get stable user scoping."""
+        session_store = getattr(self, "_session_store", None)
+        if not session_store:
+            return
+
+        channel_id = metadata.get("channel_id", "")
+        thread_ts = metadata.get("thread_ts", "")
+        user_id = metadata.get("user_id", "")
+        if not channel_id or not thread_ts or not user_id:
+            return
+
+        source = self.build_source(
+            chat_id=channel_id,
+            chat_name=channel_id,
+            chat_type="dm",
+            user_id=user_id,
+            thread_id=thread_ts,
+            chat_topic=metadata.get("context_channel_id") or None,
+        )
+
+        try:
+            session_store.get_or_create_session(source)
+        except Exception:
+            logger.debug(
+                "[Slack] Failed to seed assistant thread session for %s/%s",
+                channel_id,
+                thread_ts,
+                exc_info=True,
+            )
+
+    async def _handle_assistant_thread_lifecycle_event(self, event: dict) -> None:
+        """Handle Slack Assistant lifecycle events that carry user/thread identity."""
+        metadata = self._extract_assistant_thread_metadata(event)
+        self._cache_assistant_thread_metadata(metadata)
+        self._seed_assistant_thread_session(metadata)
+
     async def _handle_slack_message(self, event: dict) -> None:
         """Handle an incoming Slack message event."""
         # Dedup: Slack Socket Mode can redeliver events after reconnects (#4777)
@@ -781,10 +924,21 @@ class SlackAdapter(BasePlatformAdapter):
             return
 
         text = event.get("text", "")
-        user_id = event.get("user", "")
         channel_id = event.get("channel", "")
         ts = event.get("ts", "")
-        team_id = event.get("team", "")
+        assistant_meta = self._lookup_assistant_thread_metadata(
+            event,
+            channel_id=channel_id,
+            thread_ts=event.get("thread_ts", ""),
+        )
+        user_id = event.get("user") or assistant_meta.get("user_id", "")
+        if not channel_id:
+            channel_id = assistant_meta.get("channel_id", "")
+        team_id = (
+            event.get("team")
+            or event.get("team_id")
+            or assistant_meta.get("team_id", "")
+        )
 
         # Track which workspace owns this channel
         if team_id and channel_id:
@@ -792,6 +946,8 @@ class SlackAdapter(BasePlatformAdapter):
 
         # Determine if this is a DM or channel message
         channel_type = event.get("channel_type", "")
+        if not channel_type and channel_id.startswith("D"):
+            channel_type = "im"
         is_dm = channel_type == "im"
 
         # Build thread_ts for session keying.
@@ -800,7 +956,7 @@ class SlackAdapter(BasePlatformAdapter):
         # In DMs: only use the real thread_ts — top-level DMs should share
         #   one continuous session, threaded DMs get their own session.
         if is_dm:
-            thread_ts = event.get("thread_ts")  # None for top-level DMs
+            thread_ts = event.get("thread_ts") or assistant_meta.get("thread_ts")  # None for top-level DMs
         else:
             thread_ts = event.get("thread_ts") or ts  # ts fallback for channels
 

gateway/run.py

@@ -184,6 +184,8 @@ if _config_path.exists():
             # Env var from .env takes precedence (already in os.environ).
             if "gateway_timeout" in _agent_cfg and "HERMES_AGENT_TIMEOUT" not in os.environ:
                 os.environ["HERMES_AGENT_TIMEOUT"] = str(_agent_cfg["gateway_timeout"])
+            if "gateway_timeout_warning" in _agent_cfg and "HERMES_AGENT_TIMEOUT_WARNING" not in os.environ:
+                os.environ["HERMES_AGENT_TIMEOUT_WARNING"] = str(_agent_cfg["gateway_timeout_warning"])
         # Timezone: bridge config.yaml → HERMES_TIMEZONE env var.
         # HERMES_TIMEZONE from .env takes precedence (already in os.environ).
         _tz_cfg = _cfg.get("timezone", "")
@@ -1073,6 +1075,7 @@ class GatewayRunner:
                        "MATRIX_ALLOWED_USERS", "DINGTALK_ALLOWED_USERS",
                        "FEISHU_ALLOWED_USERS",
                        "WECOM_ALLOWED_USERS",
+                       "BLUEBUBBLES_ALLOWED_USERS",
                        "GATEWAY_ALLOWED_USERS")
         )
         _allow_all = os.getenv("GATEWAY_ALLOW_ALL_USERS", "").lower() in ("true", "1", "yes") or any(
@@ -1083,7 +1086,8 @@ class GatewayRunner:
                        "SMS_ALLOW_ALL_USERS", "MATTERMOST_ALLOW_ALL_USERS",
                        "MATRIX_ALLOW_ALL_USERS", "DINGTALK_ALLOW_ALL_USERS",
                        "FEISHU_ALLOW_ALL_USERS",
-                       "WECOM_ALLOW_ALL_USERS")
+                       "WECOM_ALLOW_ALL_USERS",
+                       "BLUEBUBBLES_ALLOW_ALL_USERS")
         )
         if not _any_allowlist and not _allow_all:
             logger.warning(
@@ -1654,6 +1658,13 @@ class GatewayRunner:
             adapter.gateway_runner = self  # For cross-platform delivery
             return adapter
 
+        elif platform == Platform.BLUEBUBBLES:
+            from gateway.platforms.bluebubbles import BlueBubblesAdapter, check_bluebubbles_requirements
+            if not check_bluebubbles_requirements():
+                logger.warning("BlueBubbles: aiohttp/httpx missing or BLUEBUBBLES_SERVER_URL/BLUEBUBBLES_PASSWORD not configured")
+                return None
+            return BlueBubblesAdapter(config)
+
         return None
     
     def _is_user_authorized(self, source: SessionSource) -> bool:
@@ -1692,6 +1703,7 @@ class GatewayRunner:
             Platform.DINGTALK: "DINGTALK_ALLOWED_USERS",
             Platform.FEISHU: "FEISHU_ALLOWED_USERS",
             Platform.WECOM: "WECOM_ALLOWED_USERS",
+            Platform.BLUEBUBBLES: "BLUEBUBBLES_ALLOWED_USERS",
         }
         platform_allow_all_map = {
             Platform.TELEGRAM: "TELEGRAM_ALLOW_ALL_USERS",
@@ -1706,6 +1718,7 @@ class GatewayRunner:
             Platform.DINGTALK: "DINGTALK_ALLOW_ALL_USERS",
             Platform.FEISHU: "FEISHU_ALLOW_ALL_USERS",
             Platform.WECOM: "WECOM_ALLOW_ALL_USERS",
+            Platform.BLUEBUBBLES: "BLUEBUBBLES_ALLOW_ALL_USERS",
         }
 
         # Per-platform allow-all flag (e.g., DISCORD_ALLOW_ALL_USERS=true)
@@ -1779,8 +1792,11 @@ class GatewayRunner:
         """
         source = event.source
 
-        # Check if user is authorized
-        if not self._is_user_authorized(source):
+        # Internal events (e.g. background-process completion notifications)
+        # are system-generated and must skip user authorization.
+        if getattr(event, "internal", False):
+            pass
+        elif not self._is_user_authorized(source):
             logger.warning("Unauthorized user: %s (%s) on %s", source.user_id, source.user_name, source.platform.value)
             # In DMs: offer pairing code. In groups: silently ignore.
             if source.chat_type == "dm" and self._get_unauthorized_dm_behavior(source.platform) == "pair":
@@ -5264,19 +5280,28 @@ class GatewayRunner:
 
         agent = self._running_agents.get(session_key)
         if agent and hasattr(agent, "session_total_tokens") and agent.session_api_calls > 0:
-            lines = [
-                "📊 **Session Token Usage**",
-                f"Prompt (input): {agent.session_prompt_tokens:,}",
-                f"Completion (output): {agent.session_completion_tokens:,}",
-                f"Total: {agent.session_total_tokens:,}",
-                f"API calls: {agent.session_api_calls}",
-            ]
+            lines = []
+
+            # Rate limits first (when available from provider headers)
+            rl_state = agent.get_rate_limit_state()
+            if rl_state and rl_state.has_data:
+                from agent.rate_limit_tracker import format_rate_limit_compact
+                lines.append(f"⏱️ **Rate Limits:** {format_rate_limit_compact(rl_state)}")
+                lines.append("")
+
+            # Session token usage
+            lines.append("📊 **Session Token Usage**")
+            lines.append(f"Prompt (input): {agent.session_prompt_tokens:,}")
+            lines.append(f"Completion (output): {agent.session_completion_tokens:,}")
+            lines.append(f"Total: {agent.session_total_tokens:,}")
+            lines.append(f"API calls: {agent.session_api_calls}")
             ctx = agent.context_compressor
             if ctx.last_prompt_tokens:
                 pct = min(100, ctx.last_prompt_tokens / ctx.context_length * 100) if ctx.context_length else 0
                 lines.append(f"Context: {ctx.last_prompt_tokens:,} / {ctx.context_length:,} ({pct:.0f}%)")
             if ctx.compression_count:
                 lines.append(f"Compressions: {ctx.compression_count}")
+
             return "\n".join(lines)
 
         # No running agent -- check session history for a rough count
@@ -5518,7 +5543,7 @@ class GatewayRunner:
         Platform.TELEGRAM, Platform.DISCORD, Platform.SLACK, Platform.WHATSAPP,
         Platform.SIGNAL, Platform.MATTERMOST, Platform.MATRIX,
         Platform.HOMEASSISTANT, Platform.EMAIL, Platform.SMS, Platform.DINGTALK,
-        Platform.FEISHU, Platform.WECOM, Platform.LOCAL,
+        Platform.FEISHU, Platform.WECOM, Platform.BLUEBUBBLES, Platform.LOCAL,
     })
 
     async def _handle_update_command(self, event: MessageEvent) -> str:
@@ -6158,6 +6183,7 @@ class GatewayRunner:
                                 text=synth_text,
                                 message_type=MessageType.TEXT,
                                 source=_source,
+                                internal=True,
                             )
                             logger.info(
                                 "Process %s finished — injecting agent notification for session %s",
@@ -6308,7 +6334,15 @@ class GatewayRunner:
         # Falls back to env vars for backward compatibility.
         # YAML 1.1 parses bare `off` as boolean False — normalise before
         # the `or` chain so it doesn't silently fall through to "all".
-        _raw_tp = user_config.get("display", {}).get("tool_progress")
+        #
+        # Per-platform overrides (display.tool_progress_overrides) take
+        # priority over the global setting — e.g. Signal users can set
+        # tool_progress to "off" while keeping Telegram on "all".
+        _display_cfg = user_config.get("display", {})
+        _overrides = _display_cfg.get("tool_progress_overrides", {})
+        _raw_tp = _overrides.get(platform_key)
+        if _raw_tp is None:
+            _raw_tp = _display_cfg.get("tool_progress")
         if _raw_tp is False:
             _raw_tp = "off"
         progress_mode = (
@@ -6412,6 +6446,18 @@ class GatewayRunner:
             if not adapter:
                 return
 
+            # Skip tool progress for platforms that don't support message
+            # editing (e.g. iMessage/BlueBubbles) — each progress update
+            # would become a separate message bubble, which is noisy.
+            from gateway.platforms.base import BasePlatformAdapter as _BaseAdapter
+            if type(adapter).edit_message is _BaseAdapter.edit_message:
+                while not progress_queue.empty():
+                    try:
+                        progress_queue.get_nowait()
+                    except Exception:
+                        break
+                return
+
             progress_lines = []      # Accumulated tool lines
             progress_msg_id = None   # ID of the progress message to edit
             can_edit = True          # False once an edit fails (platform doesn't support it)
@@ -7106,6 +7152,9 @@ class GatewayRunner:
             # Default 1800s (30 min inactivity).  0 = unlimited.
             _agent_timeout_raw = float(os.getenv("HERMES_AGENT_TIMEOUT", 1800))
             _agent_timeout = _agent_timeout_raw if _agent_timeout_raw > 0 else None
+            _agent_warning_raw = float(os.getenv("HERMES_AGENT_TIMEOUT_WARNING", 900))
+            _agent_warning = _agent_warning_raw if _agent_warning_raw > 0 else None
+            _warning_fired = False
             loop = asyncio.get_event_loop()
             _executor_task = asyncio.ensure_future(
                 loop.run_in_executor(None, run_sync)
@@ -7138,6 +7187,25 @@ class GatewayRunner:
                             _idle_secs = _act.get("seconds_since_activity", 0.0)
                         except Exception:
                             pass
+                    # Staged warning: fire once before escalating to full timeout.
+                    if (not _warning_fired and _agent_warning is not None
+                            and _idle_secs >= _agent_warning):
+                        _warning_fired = True
+                        _warn_adapter = self.adapters.get(source.platform)
+                        if _warn_adapter:
+                            _elapsed_warn = int(_agent_warning // 60) or 1
+                            _remaining_mins = int((_agent_timeout - _agent_warning) // 60) or 1
+                            try:
+                                await _warn_adapter.send(
+                                    source.chat_id,
+                                    f"⚠️ No activity for {_elapsed_warn} min. "
+                                    f"If the agent does not respond soon, it will "
+                                    f"be timed out in {_remaining_mins} min. "
+                                    f"You can continue waiting or use /reset.",
+                                    metadata=_status_thread_metadata,
+                                )
+                            except Exception as _warn_err:
+                                logger.debug("Inactivity warning send error: %s", _warn_err)
                     if _idle_secs >= _agent_timeout:
                         _inactivity_timeout = True
                         break

gateway/session.py

@@ -193,6 +193,7 @@ _PII_SAFE_PLATFORMS = frozenset({
     Platform.WHATSAPP,
     Platform.SIGNAL,
     Platform.TELEGRAM,
+    Platform.BLUEBUBBLES,
 })
 """Platforms where user IDs can be safely redacted (no in-message mention system
 that requires raw IDs).  Discord is excluded because mentions use ``<@user_id>``

gateway/stream_consumer.py

@@ -353,6 +353,17 @@ class GatewayStreamConsumer:
                     self._message_id = result.message_id
                     self._already_sent = True
                     self._last_sent_text = text
+                elif result.success:
+                    # Platform accepted the message but returned no message_id
+                    # (e.g. Signal).  Can't edit without an ID — switch to
+                    # fallback mode: suppress intermediate deltas, send only
+                    # the missing tail once the final response is ready.
+                    self._already_sent = True
+                    self._edit_supported = False
+                    self._fallback_prefix = self._clean_for_display(text)
+                    self._fallback_final_send = True
+                    # Sentinel prevents re-entering this branch on every delta
+                    self._message_id = "__no_edit__"
                 else:
                     # Initial send failed — disable streaming for this session
                     self._edit_supported = False

hermes_cli/auth.py

@@ -67,12 +67,16 @@ DEFAULT_AGENT_KEY_MIN_TTL_SECONDS = 30 * 60  # 30 minutes
 ACCESS_TOKEN_REFRESH_SKEW_SECONDS = 120       # refresh 2 min before expiry
 DEVICE_AUTH_POLL_INTERVAL_CAP_SECONDS = 1     # poll at most every 1s
 DEFAULT_CODEX_BASE_URL = "https://chatgpt.com/backend-api/codex"
+DEFAULT_QWEN_BASE_URL = "https://portal.qwen.ai/v1"
 DEFAULT_GITHUB_MODELS_BASE_URL = "https://api.githubcopilot.com"
 DEFAULT_COPILOT_ACP_BASE_URL = "acp://copilot"
 DEFAULT_GEMINI_BASE_URL = "https://generativelanguage.googleapis.com/v1beta/openai"
 CODEX_OAUTH_CLIENT_ID = "app_EMoamEEZ73f0CkXaXp7hrann"
 CODEX_OAUTH_TOKEN_URL = "https://auth.openai.com/oauth/token"
 CODEX_ACCESS_TOKEN_REFRESH_SKEW_SECONDS = 120
+QWEN_OAUTH_CLIENT_ID = "f0304373b74a44d2b584a3fb70ca9e56"
+QWEN_OAUTH_TOKEN_URL = "https://chat.qwen.ai/api/v1/oauth2/token"
+QWEN_ACCESS_TOKEN_REFRESH_SKEW_SECONDS = 120
 
 
 # =============================================================================
@@ -112,6 +116,12 @@ PROVIDER_REGISTRY: Dict[str, ProviderConfig] = {
         auth_type="oauth_external",
         inference_base_url=DEFAULT_CODEX_BASE_URL,
     ),
+    "qwen-oauth": ProviderConfig(
+        id="qwen-oauth",
+        name="Qwen OAuth",
+        auth_type="oauth_external",
+        inference_base_url=DEFAULT_QWEN_BASE_URL,
+    ),
     "copilot": ProviderConfig(
         id="copilot",
         name="GitHub Copilot",
@@ -240,7 +250,7 @@ PROVIDER_REGISTRY: Dict[str, ProviderConfig] = {
 # Kimi Code Endpoint Detection
 # =============================================================================
 
-# Kimi Code (platform.kimi.ai) issues keys prefixed "sk-kimi-" that only work
+# Kimi Code (kimi.com/code) issues keys prefixed "sk-kimi-" that only work
 # on api.kimi.com/coding/v1.  Legacy keys from platform.moonshot.ai work on
 # api.moonshot.ai/v1 (the default).  Auto-detect when user hasn't set
 # KIMI_BASE_URL explicitly.
@@ -817,6 +827,7 @@ def resolve_provider(
         "github-copilot-acp": "copilot-acp", "copilot-acp-agent": "copilot-acp",
         "aigateway": "ai-gateway", "vercel": "ai-gateway", "vercel-ai-gateway": "ai-gateway",
         "opencode": "opencode-zen", "zen": "opencode-zen",
+        "qwen-portal": "qwen-oauth", "qwen-cli": "qwen-oauth", "qwen-oauth": "qwen-oauth",
         "hf": "huggingface", "hugging-face": "huggingface", "huggingface-hub": "huggingface",
         "go": "opencode-go", "opencode-go-sub": "opencode-go",
         "kilo": "kilocode", "kilo-code": "kilocode", "kilo-gateway": "kilocode",
@@ -946,6 +957,176 @@ def _codex_access_token_is_expiring(access_token: Any, skew_seconds: int) -> boo
     return float(exp) <= (time.time() + max(0, int(skew_seconds)))
 
 
+def _qwen_cli_auth_path() -> Path:
+    return Path.home() / ".qwen" / "oauth_creds.json"
+
+
+def _read_qwen_cli_tokens() -> Dict[str, Any]:
+    auth_path = _qwen_cli_auth_path()
+    if not auth_path.exists():
+        raise AuthError(
+            "Qwen CLI credentials not found. Run 'qwen auth qwen-oauth' first.",
+            provider="qwen-oauth",
+            code="qwen_auth_missing",
+        )
+    try:
+        data = json.loads(auth_path.read_text(encoding="utf-8"))
+    except Exception as exc:
+        raise AuthError(
+            f"Failed to read Qwen CLI credentials from {auth_path}: {exc}",
+            provider="qwen-oauth",
+            code="qwen_auth_read_failed",
+        ) from exc
+    if not isinstance(data, dict):
+        raise AuthError(
+            f"Invalid Qwen CLI credentials in {auth_path}.",
+            provider="qwen-oauth",
+            code="qwen_auth_invalid",
+        )
+    return data
+
+
+def _save_qwen_cli_tokens(tokens: Dict[str, Any]) -> Path:
+    auth_path = _qwen_cli_auth_path()
+    auth_path.parent.mkdir(parents=True, exist_ok=True)
+    tmp_path = auth_path.with_suffix(".tmp")
+    tmp_path.write_text(json.dumps(tokens, indent=2, sort_keys=True) + "\n", encoding="utf-8")
+    os.chmod(tmp_path, stat.S_IRUSR | stat.S_IWUSR)
+    tmp_path.replace(auth_path)
+    return auth_path
+
+
+def _qwen_access_token_is_expiring(expiry_date_ms: Any, skew_seconds: int = QWEN_ACCESS_TOKEN_REFRESH_SKEW_SECONDS) -> bool:
+    try:
+        expiry_ms = int(expiry_date_ms)
+    except Exception:
+        return True
+    return (time.time() + max(0, int(skew_seconds))) * 1000 >= expiry_ms
+
+
+def _refresh_qwen_cli_tokens(tokens: Dict[str, Any], timeout_seconds: float = 20.0) -> Dict[str, Any]:
+    refresh_token = str(tokens.get("refresh_token", "") or "").strip()
+    if not refresh_token:
+        raise AuthError(
+            "Qwen OAuth refresh token missing. Re-run 'qwen auth qwen-oauth'.",
+            provider="qwen-oauth",
+            code="qwen_refresh_token_missing",
+        )
+
+    try:
+        response = httpx.post(
+            QWEN_OAUTH_TOKEN_URL,
+            headers={
+                "Content-Type": "application/x-www-form-urlencoded",
+                "Accept": "application/json",
+            },
+            data={
+                "grant_type": "refresh_token",
+                "refresh_token": refresh_token,
+                "client_id": QWEN_OAUTH_CLIENT_ID,
+            },
+            timeout=timeout_seconds,
+        )
+    except Exception as exc:
+        raise AuthError(
+            f"Qwen OAuth refresh failed: {exc}",
+            provider="qwen-oauth",
+            code="qwen_refresh_failed",
+        ) from exc
+
+    if response.status_code >= 400:
+        body = response.text.strip()
+        raise AuthError(
+            "Qwen OAuth refresh failed. Re-run 'qwen auth qwen-oauth'."
+            + (f" Response: {body}" if body else ""),
+            provider="qwen-oauth",
+            code="qwen_refresh_failed",
+        )
+
+    try:
+        payload = response.json()
+    except Exception as exc:
+        raise AuthError(
+            f"Qwen OAuth refresh returned invalid JSON: {exc}",
+            provider="qwen-oauth",
+            code="qwen_refresh_invalid_json",
+        ) from exc
+
+    if not isinstance(payload, dict) or not str(payload.get("access_token", "") or "").strip():
+        raise AuthError(
+            "Qwen OAuth refresh response missing access_token.",
+            provider="qwen-oauth",
+            code="qwen_refresh_invalid_response",
+        )
+
+    expires_in = payload.get("expires_in")
+    try:
+        expires_in_seconds = int(expires_in)
+    except Exception:
+        expires_in_seconds = 6 * 60 * 60
+
+    refreshed = {
+        "access_token": str(payload.get("access_token", "") or "").strip(),
+        "refresh_token": str(payload.get("refresh_token", refresh_token) or refresh_token).strip(),
+        "token_type": str(payload.get("token_type", tokens.get("token_type", "Bearer")) or "Bearer").strip() or "Bearer",
+        "resource_url": str(payload.get("resource_url", tokens.get("resource_url", "portal.qwen.ai")) or "portal.qwen.ai").strip(),
+        "expiry_date": int(time.time() * 1000) + max(1, expires_in_seconds) * 1000,
+    }
+    _save_qwen_cli_tokens(refreshed)
+    return refreshed
+
+
+def resolve_qwen_runtime_credentials(
+    *,
+    force_refresh: bool = False,
+    refresh_if_expiring: bool = True,
+    refresh_skew_seconds: int = QWEN_ACCESS_TOKEN_REFRESH_SKEW_SECONDS,
+) -> Dict[str, Any]:
+    tokens = _read_qwen_cli_tokens()
+    access_token = str(tokens.get("access_token", "") or "").strip()
+    should_refresh = bool(force_refresh)
+    if not should_refresh and refresh_if_expiring:
+        should_refresh = _qwen_access_token_is_expiring(tokens.get("expiry_date"), refresh_skew_seconds)
+    if should_refresh:
+        tokens = _refresh_qwen_cli_tokens(tokens)
+        access_token = str(tokens.get("access_token", "") or "").strip()
+    if not access_token:
+        raise AuthError(
+            "Qwen OAuth access token missing. Re-run 'qwen auth qwen-oauth'.",
+            provider="qwen-oauth",
+            code="qwen_access_token_missing",
+        )
+
+    base_url = os.getenv("HERMES_QWEN_BASE_URL", "").strip().rstrip("/") or DEFAULT_QWEN_BASE_URL
+    return {
+        "provider": "qwen-oauth",
+        "base_url": base_url,
+        "api_key": access_token,
+        "source": "qwen-cli",
+        "expires_at_ms": tokens.get("expiry_date"),
+        "auth_file": str(_qwen_cli_auth_path()),
+    }
+
+
+def get_qwen_auth_status() -> Dict[str, Any]:
+    auth_path = _qwen_cli_auth_path()
+    try:
+        creds = resolve_qwen_runtime_credentials(refresh_if_expiring=False)
+        return {
+            "logged_in": True,
+            "auth_file": str(auth_path),
+            "source": creds.get("source"),
+            "api_key": creds.get("api_key"),
+            "expires_at_ms": creds.get("expires_at_ms"),
+        }
+    except AuthError as exc:
+        return {
+            "logged_in": False,
+            "auth_file": str(auth_path),
+            "error": str(exc),
+        }
+
+
 # =============================================================================
 # SSH / remote session detection
 # =============================================================================
@@ -2072,6 +2253,8 @@ def get_auth_status(provider_id: Optional[str] = None) -> Dict[str, Any]:
         return get_nous_auth_status()
     if target == "openai-codex":
         return get_codex_auth_status()
+    if target == "qwen-oauth":
+        return get_qwen_auth_status()
     if target == "copilot-acp":
         return get_external_process_provider_status(target)
     # API-key providers
@@ -2834,12 +3017,15 @@ def _login_nous(args, pconfig: ProviderConfig) -> None:
             _save_provider_state(auth_store, "nous", auth_state)
             saved_to = _save_auth_store(auth_store)
 
-        config_path = _update_config_for_provider("nous", inference_base_url)
         print()
         print("Login successful!")
         print(f"  Auth state: {saved_to}")
-        print(f"  Config updated: {config_path} (model.provider=nous)")
 
+        # Resolve model BEFORE writing provider to config.yaml so we never
+        # leave the config in a half-updated state (provider=nous but model
+        # still set to the previous provider's model, e.g. opus from
+        # OpenRouter).  The auth.json active_provider was already set above.
+        selected_model = None
         try:
             runtime_key = auth_state.get("agent_key") or auth_state.get("access_token")
             if not isinstance(runtime_key, str) or not runtime_key:
@@ -2873,9 +3059,6 @@ def _login_nous(args, pconfig: ProviderConfig) -> None:
                     unavailable_models=unavailable_models,
                     portal_url=_portal,
                 )
-                if selected_model:
-                    _save_model_choice(selected_model)
-                    print(f"Default model set to: {selected_model}")
             elif unavailable_models:
                 _url = (_portal or DEFAULT_NOUS_PORTAL_URL).rstrip("/")
                 print("No free models currently available.")
@@ -2887,6 +3070,15 @@ def _login_nous(args, pconfig: ProviderConfig) -> None:
             print()
             print(f"Login succeeded, but could not fetch available models. Reason: {message}")
 
+        # Write provider + model atomically so config is never mismatched.
+        config_path = _update_config_for_provider(
+            "nous", inference_base_url, default_model=selected_model,
+        )
+        if selected_model:
+            _save_model_choice(selected_model)
+            print(f"Default model set to: {selected_model}")
+        print(f"  Config updated: {config_path} (model.provider=nous)")
+
     except KeyboardInterrupt:
         print("\nLogin cancelled.")
         raise SystemExit(130)

hermes_cli/auth_commands.py

@@ -32,7 +32,7 @@ from hermes_constants import OPENROUTER_BASE_URL
 
 
 # Providers that support OAuth login in addition to API keys.
-_OAUTH_CAPABLE_PROVIDERS = {"anthropic", "nous", "openai-codex"}
+_OAUTH_CAPABLE_PROVIDERS = {"anthropic", "nous", "openai-codex", "qwen-oauth"}
 
 
 def _get_custom_provider_names() -> list:
@@ -147,7 +147,7 @@ def auth_add_command(args) -> None:
         if provider.startswith(CUSTOM_POOL_PREFIX):
             requested_type = AUTH_TYPE_API_KEY
         else:
-            requested_type = AUTH_TYPE_OAUTH if provider in {"anthropic", "nous", "openai-codex"} else AUTH_TYPE_API_KEY
+            requested_type = AUTH_TYPE_OAUTH if provider in {"anthropic", "nous", "openai-codex", "qwen-oauth"} else AUTH_TYPE_API_KEY
 
     pool = load_pool(provider)
 
@@ -250,6 +250,26 @@ def auth_add_command(args) -> None:
         print(f'Added {provider} OAuth credential #{len(pool.entries())}: "{entry.label}"')
         return
 
+    if provider == "qwen-oauth":
+        creds = auth_mod.resolve_qwen_runtime_credentials(refresh_if_expiring=False)
+        label = (getattr(args, "label", None) or "").strip() or label_from_token(
+            creds["api_key"],
+            _oauth_default_label(provider, len(pool.entries()) + 1),
+        )
+        entry = PooledCredential(
+            provider=provider,
+            id=uuid.uuid4().hex[:6],
+            label=label,
+            auth_type=AUTH_TYPE_OAUTH,
+            priority=0,
+            source=f"{SOURCE_MANUAL}:qwen_cli",
+            access_token=creds["api_key"],
+            base_url=creds.get("base_url"),
+        )
+        pool.add_entry(entry)
+        print(f'Added {provider} OAuth credential #{len(pool.entries())}: "{entry.label}"')
+        return
+
     raise SystemExit(f"`hermes auth add {provider}` is not implemented for auth type {requested_type} yet.")
 
 

hermes_cli/banner.py

@@ -295,10 +295,16 @@ def _format_context_length(tokens: int) -> str:
     """Format a token count for display (e.g. 128000 → '128K', 1048576 → '1M')."""
     if tokens >= 1_000_000:
         val = tokens / 1_000_000
-        return f"{val:g}M"
+        rounded = round(val)
+        if abs(val - rounded) < 0.05:
+            return f"{rounded}M"
+        return f"{val:.1f}M"
     elif tokens >= 1_000:
         val = tokens / 1_000
-        return f"{val:g}K"
+        rounded = round(val)
+        if abs(val - rounded) < 0.05:
+            return f"{rounded}K"
+        return f"{val:.1f}K"
     return str(tokens)
 
 

hermes_cli/commands.py

@@ -87,8 +87,7 @@ COMMAND_REGISTRY: list[CommandDef] = [
     CommandDef("model", "Switch model for this session", "Configuration", args_hint="[model] [--global]"),
     CommandDef("provider", "Show available providers and current provider",
                "Configuration"),
-    CommandDef("prompt", "View/set custom system prompt", "Configuration",
-               cli_only=True, args_hint="[text]", subcommands=("clear",)),
+
     CommandDef("personality", "Set a predefined personality", "Configuration",
                args_hint="[name]"),
     CommandDef("statusbar", "Toggle the context/model status bar", "Configuration",
@@ -129,7 +128,7 @@ COMMAND_REGISTRY: list[CommandDef] = [
     CommandDef("commands", "Browse all commands and skills (paginated)", "Info",
                gateway_only=True, args_hint="[page]"),
     CommandDef("help", "Show available commands", "Info"),
-    CommandDef("usage", "Show token usage for the current session", "Info"),
+    CommandDef("usage", "Show token usage and rate limits for the current session", "Info"),
     CommandDef("insights", "Show usage insights and analytics", "Info",
                args_hint="[days]"),
     CommandDef("platforms", "Show gateway/messaging platform status", "Info",

hermes_cli/config.py

@@ -39,6 +39,7 @@ _EXTRA_ENV_KEYS = frozenset({
     "DINGTALK_CLIENT_ID", "DINGTALK_CLIENT_SECRET",
     "FEISHU_APP_ID", "FEISHU_APP_SECRET", "FEISHU_ENCRYPT_KEY", "FEISHU_VERIFICATION_TOKEN",
     "WECOM_BOT_ID", "WECOM_SECRET",
+    "BLUEBUBBLES_SERVER_URL", "BLUEBUBBLES_PASSWORD",
     "TERMINAL_ENV", "TERMINAL_SSH_KEY", "TERMINAL_SSH_PORT",
     "WHATSAPP_MODE", "WHATSAPP_ENABLED",
     "MATTERMOST_HOME_CHANNEL", "MATTERMOST_REPLY_MODE",
@@ -157,7 +158,14 @@ def get_project_root() -> Path:
     return Path(__file__).parent.parent.resolve()
 
 def _secure_dir(path):
-    """Set directory to owner-only access (0700). No-op on Windows."""
+    """Set directory to owner-only access (0700). No-op on Windows.
+
+    Skipped in managed mode — the NixOS module sets group-readable
+    permissions (0750) so interactive users in the hermes group can
+    share state with the gateway service.
+    """
+    if is_managed():
+        return
     try:
         os.chmod(path, 0o700)
     except (OSError, NotImplementedError):
@@ -165,7 +173,13 @@ def _secure_dir(path):
 
 
 def _secure_file(path):
-    """Set file to owner-only read/write (0600). No-op on Windows."""
+    """Set file to owner-only read/write (0600). No-op on Windows.
+
+    Skipped in managed mode — the NixOS activation script sets
+    group-readable permissions (0640) on config files.
+    """
+    if is_managed():
+        return
     try:
         if os.path.exists(str(path)):
             os.chmod(path, 0o600)
@@ -217,6 +231,10 @@ DEFAULT_CONFIG = {
         # (force on/off for all models), or a list of model-name substrings
         # to match (e.g. ["gpt", "codex", "gemini", "qwen"]).
         "tool_use_enforcement": "auto",
+        # Staged inactivity warning: send a warning to the user at this
+        # threshold before escalating to a full timeout.  The warning fires
+        # once per run and does not interrupt the agent.  0 = disable warning.
+        "gateway_timeout_warning": 900,
     },
     
     "terminal": {
@@ -379,6 +397,7 @@ DEFAULT_CONFIG = {
         "show_cost": False,       # Show $ cost in the status bar (off by default)
         "skin": "default",
         "tool_progress_command": False,  # Enable /verbose command in messaging gateway
+        "tool_progress_overrides": {},  # Per-platform overrides: {"signal": "off", "telegram": "all"}
         "tool_preview_length": 0,  # Max chars for tool call previews (0 = no limit, show full paths/commands)
     },
 
@@ -413,7 +432,7 @@ DEFAULT_CONFIG = {
     
     "stt": {
         "enabled": True,
-        "provider": "local",  # "local" (free, faster-whisper) | "groq" | "openai" (Whisper API)
+        "provider": "local",  # "local" (free, faster-whisper) | "groq" | "openai" (Whisper API) | "mistral" (Voxtral Transcribe)
         "local": {
             "model": "base",  # tiny, base, small, medium, large-v3
             "language": "",  # auto-detect by default; set to "en", "es", "fr", etc. to force
@@ -421,6 +440,9 @@ DEFAULT_CONFIG = {
         "openai": {
             "model": "whisper-1",  # whisper-1, gpt-4o-mini-transcribe, gpt-4o-transcribe
         },
+        "mistral": {
+            "model": "voxtral-mini-latest",  # voxtral-mini-latest, voxtral-mini-2602
+        },
     },
 
     "voice": {
@@ -547,7 +569,7 @@ DEFAULT_CONFIG = {
     },
 
     # Config schema version - bump this when adding new required fields
-    "_config_version": 12,
+    "_config_version": 13,
 }
 
 # =============================================================================
@@ -724,6 +746,14 @@ OPTIONAL_ENV_VARS = {
         "category": "provider",
         "advanced": True,
     },
+    "HERMES_QWEN_BASE_URL": {
+        "description": "Qwen Portal base URL override (default: https://portal.qwen.ai/v1)",
+        "prompt": "Qwen Portal base URL (leave empty for default)",
+        "url": None,
+        "password": False,
+        "category": "provider",
+        "advanced": True,
+    },
     "OPENCODE_ZEN_API_KEY": {
         "description": "OpenCode Zen API key (pay-as-you-go access to curated models)",
         "prompt": "OpenCode Zen API key",
@@ -975,6 +1005,13 @@ OPTIONAL_ENV_VARS = {
         "password": False,
         "category": "messaging",
     },
+    "DISCORD_REPLY_TO_MODE": {
+        "description": "Discord reply threading mode: 'off' (no reply references), 'first' (reply on first message only, default), 'all' (reply on every chunk)",
+        "prompt": "Discord reply mode (off/first/all)",
+        "url": None,
+        "password": False,
+        "category": "messaging",
+    },
     "SLACK_BOT_TOKEN": {
         "description": "Slack bot token (xoxb-). Get from OAuth & Permissions after installing your app. "
                        "Required scopes: chat:write, app_mentions:read, channels:history, groups:history, "
@@ -1088,6 +1125,27 @@ OPTIONAL_ENV_VARS = {
         "category": "messaging",
         "advanced": True,
     },
+    "BLUEBUBBLES_SERVER_URL": {
+        "description": "BlueBubbles server URL for iMessage integration (e.g. http://192.168.1.10:1234)",
+        "prompt": "BlueBubbles server URL",
+        "url": "https://bluebubbles.app/",
+        "password": False,
+        "category": "messaging",
+    },
+    "BLUEBUBBLES_PASSWORD": {
+        "description": "BlueBubbles server password (from BlueBubbles Server → Settings → API)",
+        "prompt": "BlueBubbles server password",
+        "url": None,
+        "password": True,
+        "category": "messaging",
+    },
+    "BLUEBUBBLES_ALLOWED_USERS": {
+        "description": "Comma-separated iMessage addresses (email or phone) allowed to use the bot",
+        "prompt": "Allowed iMessage addresses (comma-separated)",
+        "url": None,
+        "password": False,
+        "category": "messaging",
+    },
     "GATEWAY_ALLOW_ALL_USERS": {
         "description": "Allow all users to interact with messaging bots (true/false). Default: false.",
         "prompt": "Allow all users (true/false)",
@@ -1159,7 +1217,7 @@ OPTIONAL_ENV_VARS = {
         "category": "setting",
     },
     "SUDO_PASSWORD": {
-        "description": "Sudo password for terminal commands requiring root access",
+        "description": "Sudo password for terminal commands requiring root access; set to an explicit empty string to try empty without prompting",
         "prompt": "Sudo password",
         "url": None,
         "password": True,
@@ -1643,6 +1701,21 @@ def migrate_config(interactive: bool = True, quiet: bool = False) -> Dict[str, A
                         ep = providers_dict[key]
                         print(f"    → {key}: {ep.get('api', '')}")
 
+    # ── Version 12 → 13: clear dead LLM_MODEL / OPENAI_MODEL from .env ──
+    # These env vars were written by the old setup wizard but nothing reads
+    # them anymore (config.yaml is the sole source of truth since March 2026).
+    # Stale entries cause user confusion — see issue report.
+    if current_ver < 13:
+        for dead_var in ("LLM_MODEL", "OPENAI_MODEL"):
+            try:
+                old_val = get_env_value(dead_var)
+                if old_val:
+                    save_env_value(dead_var, "")
+                    if not quiet:
+                        print(f"  ✓ Cleared {dead_var} from .env (no longer used — config.yaml is source of truth)")
+            except Exception:
+                pass
+
     if current_ver < latest_ver and not quiet:
         print(f"Config version: {current_ver} → {latest_ver}")
     

hermes_cli/doctor.py

@@ -812,69 +812,83 @@ def run_doctor(args):
         check_warn("No GITHUB_TOKEN", f"(60 req/hr rate limit — set in {_DHH}/.env for better rates)")
 
     # =========================================================================
-    # Honcho memory
+    # Memory Provider (only check the active provider, if any)
     # =========================================================================
     print()
-    print(color("◆ Honcho Memory", Colors.CYAN, Colors.BOLD))
+    print(color("◆ Memory Provider", Colors.CYAN, Colors.BOLD))
 
+    _active_memory_provider = ""
     try:
-        from plugins.memory.honcho.client import HonchoClientConfig, resolve_config_path
-        hcfg = HonchoClientConfig.from_global_config()
-        _honcho_cfg_path = resolve_config_path()
+        import yaml as _yaml
+        _mem_cfg_path = HERMES_HOME / "config.yaml"
+        if _mem_cfg_path.exists():
+            with open(_mem_cfg_path) as _f:
+                _raw_cfg = _yaml.safe_load(_f) or {}
+            _active_memory_provider = (_raw_cfg.get("memory") or {}).get("provider", "")
+    except Exception:
+        pass
 
-        if not _honcho_cfg_path.exists():
-            check_warn("Honcho config not found", "run: hermes memory setup")
-        elif not hcfg.enabled:
-            check_info(f"Honcho disabled (set enabled: true in {_honcho_cfg_path} to activate)")
-        elif not (hcfg.api_key or hcfg.base_url):
-            check_fail("Honcho API key or base URL not set", "run: hermes memory setup")
-            issues.append("No Honcho API key — run 'hermes memory setup'")
-        else:
-            from plugins.memory.honcho.client import get_honcho_client, reset_honcho_client
-            reset_honcho_client()
-            try:
-                get_honcho_client(hcfg)
-                check_ok(
-                    "Honcho connected",
-                    f"workspace={hcfg.workspace_id} mode={hcfg.recall_mode} freq={hcfg.write_frequency}",
-                )
-            except Exception as _e:
-                check_fail("Honcho connection failed", str(_e))
-                issues.append(f"Honcho unreachable: {_e}")
-    except ImportError:
-        check_warn("honcho-ai not installed", "pip install honcho-ai")
-    except Exception as _e:
-        check_warn("Honcho check failed", str(_e))
+    if not _active_memory_provider:
+        check_ok("Built-in memory active", "(no external provider configured — this is fine)")
+    elif _active_memory_provider == "honcho":
+        try:
+            from plugins.memory.honcho.client import HonchoClientConfig, resolve_config_path
+            hcfg = HonchoClientConfig.from_global_config()
+            _honcho_cfg_path = resolve_config_path()
 
-    # =========================================================================
-    # Mem0 memory
-    # =========================================================================
-    print()
-    print(color("◆ Mem0 Memory", Colors.CYAN, Colors.BOLD))
-
-    try:
-        from plugins.memory.mem0 import _load_config as _load_mem0_config
-        mem0_cfg = _load_mem0_config()
-        mem0_key = mem0_cfg.get("api_key", "")
-        if mem0_key:
-            check_ok("Mem0 API key configured")
-            check_info(f"user_id={mem0_cfg.get('user_id', '?')}  agent_id={mem0_cfg.get('agent_id', '?')}")
-            # Check if mem0.json exists but is missing api_key (the bug we fixed)
-            mem0_json = HERMES_HOME / "mem0.json"
-            if mem0_json.exists():
+            if not _honcho_cfg_path.exists():
+                check_warn("Honcho config not found", "run: hermes memory setup")
+            elif not hcfg.enabled:
+                check_info(f"Honcho disabled (set enabled: true in {_honcho_cfg_path} to activate)")
+            elif not (hcfg.api_key or hcfg.base_url):
+                check_fail("Honcho API key or base URL not set", "run: hermes memory setup")
+                issues.append("No Honcho API key — run 'hermes memory setup'")
+            else:
+                from plugins.memory.honcho.client import get_honcho_client, reset_honcho_client
+                reset_honcho_client()
                 try:
-                    import json as _json
-                    file_cfg = _json.loads(mem0_json.read_text())
-                    if not file_cfg.get("api_key") and mem0_key:
-                        check_info("api_key from .env (not in mem0.json) — this is fine")
-                except Exception:
-                    pass
-        else:
-            check_warn("Mem0 not configured", "(set MEM0_API_KEY in .env or run hermes memory setup)")
-    except ImportError:
-        check_warn("Mem0 plugin not loadable", "(optional)")
-    except Exception as _e:
-        check_warn("Mem0 check failed", str(_e))
+                    get_honcho_client(hcfg)
+                    check_ok(
+                        "Honcho connected",
+                        f"workspace={hcfg.workspace_id} mode={hcfg.recall_mode} freq={hcfg.write_frequency}",
+                    )
+                except Exception as _e:
+                    check_fail("Honcho connection failed", str(_e))
+                    issues.append(f"Honcho unreachable: {_e}")
+        except ImportError:
+            check_fail("honcho-ai not installed", "pip install honcho-ai")
+            issues.append("Honcho is set as memory provider but honcho-ai is not installed")
+        except Exception as _e:
+            check_warn("Honcho check failed", str(_e))
+    elif _active_memory_provider == "mem0":
+        try:
+            from plugins.memory.mem0 import _load_config as _load_mem0_config
+            mem0_cfg = _load_mem0_config()
+            mem0_key = mem0_cfg.get("api_key", "")
+            if mem0_key:
+                check_ok("Mem0 API key configured")
+                check_info(f"user_id={mem0_cfg.get('user_id', '?')}  agent_id={mem0_cfg.get('agent_id', '?')}")
+            else:
+                check_fail("Mem0 API key not set", "(set MEM0_API_KEY in .env or run hermes memory setup)")
+                issues.append("Mem0 is set as memory provider but API key is missing")
+        except ImportError:
+            check_fail("Mem0 plugin not loadable", "pip install mem0ai")
+            issues.append("Mem0 is set as memory provider but mem0ai is not installed")
+        except Exception as _e:
+            check_warn("Mem0 check failed", str(_e))
+    else:
+        # Generic check for other memory providers (openviking, hindsight, etc.)
+        try:
+            from plugins.memory import load_memory_provider
+            _provider = load_memory_provider(_active_memory_provider)
+            if _provider and _provider.is_available():
+                check_ok(f"{_active_memory_provider} provider active")
+            elif _provider:
+                check_warn(f"{_active_memory_provider} configured but not available", "run: hermes memory status")
+            else:
+                check_warn(f"{_active_memory_provider} plugin not found", "run: hermes memory setup")
+        except Exception as _e:
+            check_warn(f"{_active_memory_provider} check failed", str(_e))
 
     # =========================================================================
     # Profiles

hermes_cli/dump.py

@@ -0,0 +1,337 @@
+"""
+Dump command for hermes CLI.
+
+Outputs a compact, plain-text summary of the user's Hermes setup
+that can be copy-pasted into Discord/GitHub/Telegram for support context.
+No ANSI colors, no checkmarks — just data.
+"""
+
+import json
+import os
+import platform
+import subprocess
+import sys
+from pathlib import Path
+
+from hermes_cli.config import get_hermes_home, get_env_path, get_project_root, load_config
+from hermes_constants import display_hermes_home
+
+
+def _get_git_commit(project_root: Path) -> str:
+    """Return short git commit hash, or '(unknown)'."""
+    try:
+        result = subprocess.run(
+            ["git", "rev-parse", "--short=8", "HEAD"],
+            capture_output=True, text=True, timeout=5,
+            cwd=str(project_root),
+        )
+        if result.returncode == 0:
+            return result.stdout.strip()
+    except Exception:
+        pass
+    return "(unknown)"
+
+
+def _key_present(name: str) -> str:
+    """Return 'set' or 'not set' for an env var."""
+    return "set" if os.getenv(name) else "not set"
+
+
+def _redact(value: str) -> str:
+    """Redact all but first 4 and last 4 chars."""
+    if not value:
+        return ""
+    if len(value) < 12:
+        return "***"
+    return value[:4] + "..." + value[-4:]
+
+
+def _gateway_status() -> str:
+    """Return a short gateway status string."""
+    if sys.platform.startswith("linux"):
+        try:
+            from hermes_cli.gateway import get_service_name
+            svc = get_service_name()
+        except Exception:
+            svc = "hermes-gateway"
+        try:
+            r = subprocess.run(
+                ["systemctl", "--user", "is-active", svc],
+                capture_output=True, text=True, timeout=5,
+            )
+            return "running (systemd)" if r.stdout.strip() == "active" else "stopped"
+        except Exception:
+            return "unknown"
+    elif sys.platform == "darwin":
+        try:
+            from hermes_cli.gateway import get_launchd_label
+            r = subprocess.run(
+                ["launchctl", "list", get_launchd_label()],
+                capture_output=True, text=True, timeout=5,
+            )
+            return "loaded (launchd)" if r.returncode == 0 else "not loaded"
+        except Exception:
+            return "unknown"
+    return "N/A"
+
+
+def _count_skills(hermes_home: Path) -> int:
+    """Count installed skills."""
+    skills_dir = hermes_home / "skills"
+    if not skills_dir.is_dir():
+        return 0
+    count = 0
+    for item in skills_dir.rglob("SKILL.md"):
+        count += 1
+    return count
+
+
+def _count_mcp_servers(config: dict) -> int:
+    """Count configured MCP servers."""
+    mcp = config.get("mcp", {})
+    servers = mcp.get("servers", {})
+    return len(servers)
+
+
+def _cron_summary(hermes_home: Path) -> str:
+    """Return cron jobs summary."""
+    jobs_file = hermes_home / "cron" / "jobs.json"
+    if not jobs_file.exists():
+        return "0"
+    try:
+        with open(jobs_file, encoding="utf-8") as f:
+            data = json.load(f)
+        jobs = data.get("jobs", [])
+        active = sum(1 for j in jobs if j.get("enabled", True))
+        return f"{active} active / {len(jobs)} total"
+    except Exception:
+        return "(error reading)"
+
+
+def _configured_platforms() -> list[str]:
+    """Return list of configured messaging platform names."""
+    checks = {
+        "telegram": "TELEGRAM_BOT_TOKEN",
+        "discord": "DISCORD_BOT_TOKEN",
+        "slack": "SLACK_BOT_TOKEN",
+        "whatsapp": "WHATSAPP_ENABLED",
+        "signal": "SIGNAL_HTTP_URL",
+        "email": "EMAIL_ADDRESS",
+        "sms": "TWILIO_ACCOUNT_SID",
+        "matrix": "MATRIX_HOMESERVER_URL",
+        "mattermost": "MATTERMOST_URL",
+        "homeassistant": "HASS_TOKEN",
+        "dingtalk": "DINGTALK_CLIENT_ID",
+        "feishu": "FEISHU_APP_ID",
+        "wecom": "WECOM_BOT_ID",
+    }
+    return [name for name, env in checks.items() if os.getenv(env)]
+
+
+def _memory_provider(config: dict) -> str:
+    """Return the active memory provider name."""
+    mem = config.get("memory", {})
+    provider = mem.get("provider", "")
+    return provider if provider else "built-in"
+
+
+def _get_model_and_provider(config: dict) -> tuple[str, str]:
+    """Extract model and provider from config."""
+    model_cfg = config.get("model", "")
+    if isinstance(model_cfg, dict):
+        model = model_cfg.get("default") or model_cfg.get("model") or model_cfg.get("name") or "(not set)"
+        provider = model_cfg.get("provider") or "(auto)"
+    elif isinstance(model_cfg, str):
+        model = model_cfg or "(not set)"
+        provider = "(auto)"
+    else:
+        model = "(not set)"
+        provider = "(auto)"
+    return model, provider
+
+
+def _config_overrides(config: dict) -> dict[str, str]:
+    """Find non-default config values worth reporting.
+    
+    Returns a flat dict of dotpath -> value for interesting overrides.
+    """
+    from hermes_cli.config import DEFAULT_CONFIG
+
+    overrides = {}
+
+    # Sections with interesting user-facing overrides
+    interesting_paths = [
+        ("agent", "max_turns"),
+        ("agent", "gateway_timeout"),
+        ("agent", "tool_use_enforcement"),
+        ("terminal", "backend"),
+        ("terminal", "docker_image"),
+        ("terminal", "persistent_shell"),
+        ("browser", "allow_private_urls"),
+        ("compression", "enabled"),
+        ("compression", "threshold"),
+        ("display", "streaming"),
+        ("display", "skin"),
+        ("display", "show_reasoning"),
+        ("smart_model_routing", "enabled"),
+        ("privacy", "redact_pii"),
+        ("tts", "provider"),
+    ]
+
+    for section, key in interesting_paths:
+        default_section = DEFAULT_CONFIG.get(section, {})
+        user_section = config.get(section, {})
+        if not isinstance(default_section, dict) or not isinstance(user_section, dict):
+            continue
+        default_val = default_section.get(key)
+        user_val = user_section.get(key)
+        if user_val is not None and user_val != default_val:
+            overrides[f"{section}.{key}"] = str(user_val)
+
+    # Toolsets (if different from default)
+    default_toolsets = DEFAULT_CONFIG.get("toolsets", [])
+    user_toolsets = config.get("toolsets", [])
+    if user_toolsets != default_toolsets:
+        overrides["toolsets"] = str(user_toolsets)
+
+    # Fallback providers
+    fallbacks = config.get("fallback_providers", [])
+    if fallbacks:
+        overrides["fallback_providers"] = str(fallbacks)
+
+    return overrides
+
+
+def run_dump(args):
+    """Output a compact, copy-pasteable setup summary."""
+    show_keys = getattr(args, "show_keys", False)
+
+    # Load env from .env file so key checks work
+    from dotenv import load_dotenv
+    env_path = get_env_path()
+    if env_path.exists():
+        try:
+            load_dotenv(env_path, encoding="utf-8")
+        except UnicodeDecodeError:
+            load_dotenv(env_path, encoding="latin-1")
+    # Also try project .env as dev fallback
+    load_dotenv(get_project_root() / ".env", override=False, encoding="utf-8")
+
+    project_root = get_project_root()
+    hermes_home = get_hermes_home()
+
+    try:
+        from hermes_cli import __version__, __release_date__
+    except ImportError:
+        __version__ = "(unknown)"
+        __release_date__ = ""
+
+    commit = _get_git_commit(project_root)
+
+    try:
+        config = load_config()
+    except Exception:
+        config = {}
+
+    model, provider = _get_model_and_provider(config)
+
+    # Profile
+    try:
+        from hermes_cli.profiles import get_active_profile_name
+        profile = get_active_profile_name() or "(default)"
+    except Exception:
+        profile = "(default)"
+
+    # Terminal backend
+    terminal_cfg = config.get("terminal", {})
+    backend = terminal_cfg.get("backend", "local")
+
+    # OpenAI SDK version
+    try:
+        import openai
+        openai_ver = openai.__version__
+    except ImportError:
+        openai_ver = "not installed"
+
+    # OS info
+    os_info = f"{platform.system()} {platform.release()} {platform.machine()}"
+
+    lines = []
+    lines.append("--- hermes dump ---")
+    ver_str = f"{__version__}"
+    if __release_date__:
+        ver_str += f" ({__release_date__})"
+    ver_str += f" [{commit}]"
+    lines.append(f"version:          {ver_str}")
+    lines.append(f"os:               {os_info}")
+    lines.append(f"python:           {sys.version.split()[0]}")
+    lines.append(f"openai_sdk:       {openai_ver}")
+    lines.append(f"profile:          {profile}")
+    lines.append(f"hermes_home:      {display_hermes_home()}")
+    lines.append(f"model:            {model}")
+    lines.append(f"provider:         {provider}")
+    lines.append(f"terminal:         {backend}")
+
+    # API keys
+    lines.append("")
+    lines.append("api_keys:")
+    api_keys = [
+        ("OPENROUTER_API_KEY", "openrouter"),
+        ("OPENAI_API_KEY", "openai"),
+        ("ANTHROPIC_API_KEY", "anthropic"),
+        ("ANTHROPIC_TOKEN", "anthropic_token"),
+        ("NOUS_API_KEY", "nous"),
+        ("GLM_API_KEY", "glm/zai"),
+        ("ZAI_API_KEY", "zai"),
+        ("KIMI_API_KEY", "kimi"),
+        ("MINIMAX_API_KEY", "minimax"),
+        ("DEEPSEEK_API_KEY", "deepseek"),
+        ("DASHSCOPE_API_KEY", "dashscope"),
+        ("HF_TOKEN", "huggingface"),
+        ("AI_GATEWAY_API_KEY", "ai_gateway"),
+        ("OPENCODE_ZEN_API_KEY", "opencode_zen"),
+        ("OPENCODE_GO_API_KEY", "opencode_go"),
+        ("KILOCODE_API_KEY", "kilocode"),
+        ("FIRECRAWL_API_KEY", "firecrawl"),
+        ("TAVILY_API_KEY", "tavily"),
+        ("BROWSERBASE_API_KEY", "browserbase"),
+        ("FAL_KEY", "fal"),
+        ("ELEVENLABS_API_KEY", "elevenlabs"),
+        ("GITHUB_TOKEN", "github"),
+    ]
+
+    for env_var, label in api_keys:
+        val = os.getenv(env_var, "")
+        if show_keys and val:
+            display = _redact(val)
+        else:
+            display = "set" if val else "not set"
+        lines.append(f"  {label:<20} {display}")
+
+    # Features summary
+    lines.append("")
+    lines.append("features:")
+
+    toolsets = config.get("toolsets", ["hermes-cli"])
+    lines.append(f"  toolsets:           {', '.join(toolsets) if toolsets else '(default)'}")
+    lines.append(f"  mcp_servers:        {_count_mcp_servers(config)}")
+    lines.append(f"  memory_provider:    {_memory_provider(config)}")
+    lines.append(f"  gateway:            {_gateway_status()}")
+
+    platforms = _configured_platforms()
+    lines.append(f"  platforms:          {', '.join(platforms) if platforms else 'none'}")
+    lines.append(f"  cron_jobs:          {_cron_summary(hermes_home)}")
+    lines.append(f"  skills:             {_count_skills(hermes_home)}")
+
+    # Config overrides (non-default values)
+    overrides = _config_overrides(config)
+    if overrides:
+        lines.append("")
+        lines.append("config_overrides:")
+        for key, val in overrides.items():
+            lines.append(f"  {key}: {val}")
+
+    lines.append("--- end dump ---")
+
+    output = "\n".join(lines)
+    print(output)

hermes_cli/gateway.py

@@ -1588,6 +1588,34 @@ _PLATFORMS = [
              "help": "Chat ID for scheduled results and notifications."},
         ],
     },
+    {
+        "key": "bluebubbles",
+        "label": "BlueBubbles (iMessage)",
+        "emoji": "💬",
+        "token_var": "BLUEBUBBLES_SERVER_URL",
+        "setup_instructions": [
+            "1. Install BlueBubbles on a Mac that will act as your iMessage server:",
+            "   https://bluebubbles.app/",
+            "2. Complete the BlueBubbles setup wizard — sign in with your Apple ID",
+            "3. In BlueBubbles Settings → API, note the Server URL and password",
+            "4. The server URL is typically http://<your-mac-ip>:1234",
+            "5. Hermes connects via the BlueBubbles REST API and receives",
+            "   incoming messages via a local webhook",
+            "6. To authorize users, use DM pairing: hermes pairing generate bluebubbles",
+            "   Share the code — the user sends it via iMessage to get approved",
+        ],
+        "vars": [
+            {"name": "BLUEBUBBLES_SERVER_URL", "prompt": "BlueBubbles server URL (e.g. http://192.168.1.10:1234)", "password": False,
+             "help": "The URL shown in BlueBubbles Settings → API."},
+            {"name": "BLUEBUBBLES_PASSWORD", "prompt": "BlueBubbles server password", "password": True,
+             "help": "The password shown in BlueBubbles Settings → API."},
+            {"name": "BLUEBUBBLES_ALLOWED_USERS", "prompt": "Pre-authorized phone numbers or iMessage IDs (comma-separated, or leave empty for DM pairing)", "password": False,
+             "is_allowlist": True,
+             "help": "Optional — pre-authorize specific users. Leave empty to use DM pairing instead (recommended)."},
+            {"name": "BLUEBUBBLES_HOME_CHANNEL", "prompt": "Home channel (phone number or iMessage ID for cron/notifications, or empty)", "password": False,
+             "help": "Phone number or Apple ID to deliver cron results and notifications to."},
+        ],
+    },
 ]
 
 

hermes_cli/main.py

@@ -918,6 +918,7 @@ def select_provider_and_model(args=None):
         "openrouter": "OpenRouter",
         "nous": "Nous Portal",
         "openai-codex": "OpenAI Codex",
+        "qwen-oauth": "Qwen OAuth",
         "copilot-acp": "GitHub Copilot ACP",
         "copilot": "GitHub Copilot",
         "anthropic": "Anthropic",
@@ -947,6 +948,7 @@ def select_provider_and_model(args=None):
         ("openrouter", "OpenRouter (100+ models, pay-per-use)"),
         ("anthropic", "Anthropic (Claude models — API key or Claude Code)"),
         ("openai-codex", "OpenAI Codex"),
+        ("qwen-oauth", "Qwen OAuth (reuses local Qwen CLI login)"),
         ("copilot", "GitHub Copilot (uses GITHUB_TOKEN or gh auth token)"),
         ("huggingface", "Hugging Face Inference Providers (20+ open models)"),
     ]
@@ -1043,6 +1045,8 @@ def select_provider_and_model(args=None):
         _model_flow_nous(config, current_model, args=args)
     elif selected_provider == "openai-codex":
         _model_flow_openai_codex(config, current_model)
+    elif selected_provider == "qwen-oauth":
+        _model_flow_qwen_oauth(config, current_model)
     elif selected_provider == "copilot-acp":
         _model_flow_copilot_acp(config, current_model)
     elif selected_provider == "copilot":
@@ -1359,6 +1363,56 @@ def _model_flow_openai_codex(config, current_model=""):
 
 
 
+_DEFAULT_QWEN_PORTAL_MODELS = [
+    "qwen3-coder-plus",
+    "qwen3-coder",
+]
+
+
+def _model_flow_qwen_oauth(_config, current_model=""):
+    """Qwen OAuth provider: reuse local Qwen CLI login, then pick model."""
+    from hermes_cli.auth import (
+        get_qwen_auth_status,
+        resolve_qwen_runtime_credentials,
+        _prompt_model_selection,
+        _save_model_choice,
+        _update_config_for_provider,
+        DEFAULT_QWEN_BASE_URL,
+    )
+    from hermes_cli.models import fetch_api_models
+
+    status = get_qwen_auth_status()
+    if not status.get("logged_in"):
+        print("Not logged into Qwen CLI OAuth.")
+        print("Run: qwen auth qwen-oauth")
+        auth_file = status.get("auth_file")
+        if auth_file:
+            print(f"Expected credentials file: {auth_file}")
+        if status.get("error"):
+            print(f"Error: {status.get('error')}")
+        return
+
+    # Try live model discovery, fall back to curated list.
+    models = None
+    try:
+        creds = resolve_qwen_runtime_credentials(refresh_if_expiring=True)
+        models = fetch_api_models(creds["api_key"], creds["base_url"])
+    except Exception:
+        pass
+    if not models:
+        models = list(_DEFAULT_QWEN_PORTAL_MODELS)
+
+    default = current_model or (models[0] if models else "qwen3-coder-plus")
+    selected = _prompt_model_selection(models, current_model=default)
+    if selected:
+        _save_model_choice(selected)
+        _update_config_for_provider("qwen-oauth", DEFAULT_QWEN_BASE_URL)
+        print(f"Default model set to: {selected} (via Qwen OAuth)")
+    else:
+        print("No change.")
+
+
+
 def _model_flow_custom(config):
     """Custom endpoint: collect URL, API key, and model name.
 
@@ -1420,7 +1474,11 @@ def _model_flow_custom(config):
             f"Hermes will still save it."
         )
         if probe.get("suggested_base_url"):
-            print(f"  If this server expects /v1, try base URL: {probe['suggested_base_url']}")
+            suggested = probe["suggested_base_url"]
+            if suggested.endswith("/v1"):
+                print(f"  If this server expects /v1 in the path, try base URL: {suggested}")
+            else:
+                print(f"  If /v1 should not be in the base URL, try: {suggested}")
 
     # Select model — use probe results when available, fall back to manual input
     model_name = ""
@@ -2585,6 +2643,12 @@ def cmd_doctor(args):
     run_doctor(args)
 
 
+def cmd_dump(args):
+    """Dump setup summary for support/debugging."""
+    from hermes_cli.dump import run_dump
+    run_dump(args)
+
+
 def cmd_config(args):
     """Configuration management."""
     from hermes_cli.config import config_command
@@ -4666,6 +4730,22 @@ For more help on a command:
         help="Attempt to fix issues automatically"
     )
     doctor_parser.set_defaults(func=cmd_doctor)
+
+    # =========================================================================
+    # dump command
+    # =========================================================================
+    dump_parser = subparsers.add_parser(
+        "dump",
+        help="Dump setup summary for support/debugging",
+        description="Output a compact, plain-text summary of your Hermes setup "
+                    "that can be copy-pasted into Discord/GitHub for support context"
+    )
+    dump_parser.add_argument(
+        "--show-keys",
+        action="store_true",
+        help="Show redacted API key prefixes (first/last 4 chars) instead of just set/not set"
+    )
+    dump_parser.set_defaults(func=cmd_dump)
     
     # =========================================================================
     # config command

hermes_cli/model_normalize.py

@@ -84,6 +84,7 @@ _PASSTHROUGH_PROVIDERS: frozenset[str] = frozenset({
     "minimax",
     "minimax-cn",
     "alibaba",
+    "qwen-oauth",
     "huggingface",
     "openai-codex",
     "custom",

hermes_cli/model_switch.py

@@ -537,8 +537,11 @@ def switch_model(
                     )
             else:
                 # --- Step c: On aggregator, convert vendor:model to vendor/model ---
+                # Only convert when there's no slash — a slash means the name
+                # is already in vendor/model format and the colon is a variant
+                # tag (:free, :extended, :fast) that must be preserved.
                 colon_pos = raw_input.find(":")
-                if colon_pos > 0 and is_aggregator(current_provider):
+                if colon_pos > 0 and "/" not in raw_input and is_aggregator(current_provider):
                     left = raw_input[:colon_pos].strip().lower()
                     right = raw_input[colon_pos + 1:].strip()
                     if left and right:
@@ -730,6 +733,7 @@ def list_authenticated_providers(
         fetch_models_dev,
         get_provider_info as _mdev_pinfo,
     )
+    from hermes_cli.auth import PROVIDER_REGISTRY
     from hermes_cli.models import OPENROUTER_MODELS, _PROVIDER_MODELS
 
     results: List[dict] = []
@@ -750,9 +754,16 @@ def list_authenticated_providers(
         if not isinstance(pdata, dict):
             continue
 
-        env_vars = pdata.get("env", [])
-        if not isinstance(env_vars, list):
-            continue
+        # Prefer auth.py PROVIDER_REGISTRY for env var names — it's our
+        # source of truth.  models.dev can have wrong mappings (e.g.
+        # minimax-cn → MINIMAX_API_KEY instead of MINIMAX_CN_API_KEY).
+        pconfig = PROVIDER_REGISTRY.get(hermes_id)
+        if pconfig and pconfig.api_key_env_vars:
+            env_vars = list(pconfig.api_key_env_vars)
+        else:
+            env_vars = pdata.get("env", [])
+            if not isinstance(env_vars, list):
+                continue
 
         # Check if any env var is set
         has_creds = any(os.environ.get(ev) for ev in env_vars)

hermes_cli/models.py

@@ -483,6 +483,7 @@ _PROVIDER_LABELS = {
     "ai-gateway": "AI Gateway",
     "kilocode": "Kilo Code",
     "alibaba": "Alibaba Cloud (DashScope)",
+    "qwen-oauth": "Qwen OAuth (Portal)",
     "huggingface": "Hugging Face",
     "custom": "Custom endpoint",
 }
@@ -522,6 +523,7 @@ _PROVIDER_ALIASES = {
     "aliyun": "alibaba",
     "qwen": "alibaba",
     "alibaba-cloud": "alibaba",
+    "qwen-portal": "qwen-oauth",
     "hf": "huggingface",
     "hugging-face": "huggingface",
     "huggingface-hub": "huggingface",
@@ -767,6 +769,7 @@ def list_available_providers() -> list[dict[str, str]]:
         "openrouter", "nous", "openai-codex", "copilot", "copilot-acp",
         "gemini", "huggingface",
         "zai", "kimi-coding", "minimax", "minimax-cn", "kilocode", "anthropic", "alibaba",
+        "qwen-oauth",
         "opencode-zen", "opencode-go",
         "ai-gateway", "deepseek", "custom",
     ]
@@ -1529,7 +1532,7 @@ def probe_api_models(
 
     return {
         "models": None,
-        "probed_url": tried[-1] if tried else normalized.rstrip("/") + "/models",
+        "probed_url": tried[0] if tried else normalized.rstrip("/") + "/models",
         "resolved_base_url": normalized,
         "suggested_base_url": alternate_base if alternate_base != normalized else None,
         "used_fallback": False,

hermes_cli/profiles.py

@@ -102,7 +102,7 @@ _RESERVED_NAMES = frozenset({
 # Hermes subcommands that cannot be used as profile names/aliases
 _HERMES_SUBCOMMANDS = frozenset({
     "chat", "model", "gateway", "setup", "whatsapp", "login", "logout",
-    "status", "cron", "doctor", "config", "pairing", "skills", "tools",
+    "status", "cron", "doctor", "dump", "config", "pairing", "skills", "tools",
     "mcp", "sessions", "insights", "version", "update", "uninstall",
     "profile", "plugins", "honcho", "acp",
 })
@@ -1007,7 +1007,7 @@ _hermes_completion() {
 
     # Top-level subcommands
     if [[ "$COMP_CWORD" == 1 ]]; then
-        local commands="chat model gateway setup status cron doctor config skills tools mcp sessions profile update version"
+        local commands="chat model gateway setup status cron doctor dump config skills tools mcp sessions profile update version"
         COMPREPLY=($(compgen -W "$commands" -- "$cur"))
     fi
 }
@@ -1032,7 +1032,7 @@ _hermes() {
     _arguments \\
         '-p[Profile name]:profile:($profiles)' \\
         '--profile[Profile name]:profile:($profiles)' \\
-        '1:command:(chat model gateway setup status cron doctor config skills tools mcp sessions profile update version)' \\
+        '1:command:(chat model gateway setup status cron doctor dump config skills tools mcp sessions profile update version)' \\
         '*::arg:->args'
 
     case $words[1] in

hermes_cli/providers.py

@@ -58,6 +58,12 @@ HERMES_OVERLAYS: Dict[str, HermesOverlay] = {
         auth_type="oauth_external",
         base_url_override="https://chatgpt.com/backend-api/codex",
     ),
+    "qwen-oauth": HermesOverlay(
+        transport="openai_chat",
+        auth_type="oauth_external",
+        base_url_override="https://portal.qwen.ai/v1",
+        base_url_env_var="HERMES_QWEN_BASE_URL",
+    ),
     "copilot-acp": HermesOverlay(
         transport="codex_responses",
         auth_type="external_process",

hermes_cli/runtime_provider.py

@@ -14,11 +14,13 @@ from agent.credential_pool import CredentialPool, PooledCredential, get_custom_p
 from hermes_cli.auth import (
     AuthError,
     DEFAULT_CODEX_BASE_URL,
+    DEFAULT_QWEN_BASE_URL,
     PROVIDER_REGISTRY,
     format_auth_error,
     resolve_provider,
     resolve_nous_runtime_credentials,
     resolve_codex_runtime_credentials,
+    resolve_qwen_runtime_credentials,
     resolve_api_key_provider_credentials,
     resolve_external_process_provider_credentials,
     has_usable_secret,
@@ -148,6 +150,9 @@ def _resolve_runtime_from_pool_entry(
     if provider == "openai-codex":
         api_mode = "codex_responses"
         base_url = base_url or DEFAULT_CODEX_BASE_URL
+    elif provider == "qwen-oauth":
+        api_mode = "chat_completions"
+        base_url = base_url or DEFAULT_QWEN_BASE_URL
     elif provider == "anthropic":
         api_mode = "anthropic_messages"
         cfg_provider = str(model_cfg.get("provider") or "").strip().lower()
@@ -691,6 +696,24 @@ def resolve_runtime_provider(
             logger.info("Auto-detected Codex provider but credentials failed; "
                         "falling through to next provider.")
 
+    if provider == "qwen-oauth":
+        try:
+            creds = resolve_qwen_runtime_credentials()
+            return {
+                "provider": "qwen-oauth",
+                "api_mode": "chat_completions",
+                "base_url": creds.get("base_url", "").rstrip("/"),
+                "api_key": creds.get("api_key", ""),
+                "source": creds.get("source", "qwen-cli"),
+                "expires_at_ms": creds.get("expires_at_ms"),
+                "requested_provider": requested_provider,
+            }
+        except AuthError:
+            if requested_provider != "auto":
+                raise
+            logger.info("Qwen OAuth credentials failed; "
+                        "falling through to next provider.")
+
     if provider == "copilot-acp":
         creds = resolve_external_process_provider_credentials(provider)
         return {

hermes_cli/setup.py

@@ -2167,6 +2167,71 @@ def _setup_whatsapp():
         print_info("or personal self-chat) and pair via QR code.")
 
 
+def _setup_bluebubbles():
+    """Configure BlueBubbles iMessage gateway."""
+    print_header("BlueBubbles (iMessage)")
+    existing = get_env_value("BLUEBUBBLES_SERVER_URL")
+    if existing:
+        print_info("BlueBubbles: already configured")
+        if not prompt_yes_no("Reconfigure BlueBubbles?", False):
+            return
+
+    print_info("Connects Hermes to iMessage via BlueBubbles — a free, open-source")
+    print_info("macOS server that bridges iMessage to any device.")
+    print_info("   Requires a Mac running BlueBubbles Server v1.0.0+")
+    print_info("   Download: https://bluebubbles.app/")
+    print()
+    print_info("In BlueBubbles Server → Settings → API, note your Server URL and Password.")
+    print()
+
+    server_url = prompt("BlueBubbles server URL (e.g. http://192.168.1.10:1234)")
+    if not server_url:
+        print_warning("Server URL is required — skipping BlueBubbles setup")
+        return
+    save_env_value("BLUEBUBBLES_SERVER_URL", server_url.rstrip("/"))
+
+    password = prompt("BlueBubbles server password", password=True)
+    if not password:
+        print_warning("Password is required — skipping BlueBubbles setup")
+        return
+    save_env_value("BLUEBUBBLES_PASSWORD", password)
+    print_success("BlueBubbles credentials saved")
+
+    print()
+    print_info("🔒 Security: Restrict who can message your bot")
+    print_info("   Use iMessage addresses: email (user@icloud.com) or phone (+15551234567)")
+    print()
+    allowed_users = prompt("Allowed iMessage addresses (comma-separated, leave empty for open access)")
+    if allowed_users:
+        save_env_value("BLUEBUBBLES_ALLOWED_USERS", allowed_users.replace(" ", ""))
+        print_success("BlueBubbles allowlist configured")
+    else:
+        print_info("⚠️  No allowlist set — anyone who can iMessage you can use the bot!")
+
+    print()
+    print_info("📬 Home Channel: phone or email for cron job delivery and notifications.")
+    print_info("   You can also set this later with /set-home in your iMessage chat.")
+    home_channel = prompt("Home channel address (leave empty to set later)")
+    if home_channel:
+        save_env_value("BLUEBUBBLES_HOME_CHANNEL", home_channel)
+
+    print()
+    print_info("Advanced settings (defaults are fine for most setups):")
+    if prompt_yes_no("Configure webhook listener settings?", False):
+        webhook_port = prompt("Webhook listener port (default: 8645)")
+        if webhook_port:
+            try:
+                save_env_value("BLUEBUBBLES_WEBHOOK_PORT", str(int(webhook_port)))
+                print_success(f"Webhook port set to {webhook_port}")
+            except ValueError:
+                print_warning("Invalid port number, using default 8645")
+
+    print()
+    print_info("Requires the BlueBubbles Private API helper for typing indicators,")
+    print_info("read receipts, and tapback reactions. Basic messaging works without it.")
+    print_info("   Install: https://docs.bluebubbles.app/helper-bundle/installation")
+
+
 def _setup_webhooks():
     """Configure webhook integration."""
     print_header("Webhooks")
@@ -2221,6 +2286,7 @@ _GATEWAY_PLATFORMS = [
     ("Matrix", "MATRIX_ACCESS_TOKEN", _setup_matrix),
     ("Mattermost", "MATTERMOST_TOKEN", _setup_mattermost),
     ("WhatsApp", "WHATSAPP_ENABLED", _setup_whatsapp),
+    ("BlueBubbles (iMessage)", "BLUEBUBBLES_SERVER_URL", _setup_bluebubbles),
     ("Webhooks (GitHub, GitLab, etc.)", "WEBHOOK_ENABLED", _setup_webhooks),
 ]
 
@@ -2264,6 +2330,7 @@ def setup_gateway(config: dict):
         or get_env_value("MATRIX_ACCESS_TOKEN")
         or get_env_value("MATRIX_PASSWORD")
         or get_env_value("WHATSAPP_ENABLED")
+        or get_env_value("BLUEBUBBLES_SERVER_URL")
         or get_env_value("WEBHOOK_ENABLED")
     )
     if any_messaging:
@@ -2283,6 +2350,8 @@ def setup_gateway(config: dict):
             missing_home.append("Discord")
         if get_env_value("SLACK_BOT_TOKEN") and not get_env_value("SLACK_HOME_CHANNEL"):
             missing_home.append("Slack")
+        if get_env_value("BLUEBUBBLES_SERVER_URL") and not get_env_value("BLUEBUBBLES_HOME_CHANNEL"):
+            missing_home.append("BlueBubbles")
 
         if missing_home:
             print()
@@ -2453,6 +2522,8 @@ def _get_section_config_summary(config: dict, section_key: str) -> Optional[str]
             platforms.append("WhatsApp")
         if get_env_value("SIGNAL_ACCOUNT"):
             platforms.append("Signal")
+        if get_env_value("BLUEBUBBLES_SERVER_URL"):
+            platforms.append("BlueBubbles")
         if platforms:
             return ", ".join(platforms)
         return None  # No platforms configured — section must run
@@ -2501,9 +2572,120 @@ _OPENCLAW_SCRIPT = (
 )
 
 
+def _load_openclaw_migration_module():
+    """Load the openclaw_to_hermes migration script as a module.
+
+    Returns the loaded module, or None if the script can't be loaded.
+    """
+    if not _OPENCLAW_SCRIPT.exists():
+        return None
+
+    spec = importlib.util.spec_from_file_location(
+        "openclaw_to_hermes", _OPENCLAW_SCRIPT
+    )
+    if spec is None or spec.loader is None:
+        return None
+
+    mod = importlib.util.module_from_spec(spec)
+    # Register in sys.modules so @dataclass can resolve the module
+    # (Python 3.11+ requires this for dynamically loaded modules)
+    import sys as _sys
+    _sys.modules[spec.name] = mod
+    try:
+        spec.loader.exec_module(mod)
+    except Exception:
+        _sys.modules.pop(spec.name, None)
+        raise
+    return mod
+
+
+# Item kinds that represent high-impact changes warranting explicit warnings.
+# Gateway tokens/channels can hijack messaging platforms from the old agent.
+# Config values may have different semantics between OpenClaw and Hermes.
+# Instruction/context files (.md) can contain incompatible setup procedures.
+_HIGH_IMPACT_KIND_KEYWORDS = {
+    "gateway": "⚠ Gateway/messaging — this will configure Hermes to use your OpenClaw messaging channels",
+    "telegram": "⚠ Telegram — this will point Hermes at your OpenClaw Telegram bot",
+    "slack": "⚠ Slack — this will point Hermes at your OpenClaw Slack workspace",
+    "discord": "⚠ Discord — this will point Hermes at your OpenClaw Discord bot",
+    "whatsapp": "⚠ WhatsApp — this will point Hermes at your OpenClaw WhatsApp connection",
+    "config": "⚠ Config values — OpenClaw settings may not map 1:1 to Hermes equivalents",
+    "soul": "⚠ Instruction file — may contain OpenClaw-specific setup/restart procedures",
+    "memory": "⚠ Memory/context file — may reference OpenClaw-specific infrastructure",
+    "context": "⚠ Context file — may contain OpenClaw-specific instructions",
+}
+
+
+def _print_migration_preview(report: dict):
+    """Print a detailed dry-run preview of what migration would do.
+
+    Groups items by category and adds explicit warnings for high-impact
+    changes like gateway token takeover and config value differences.
+    """
+    items = report.get("items", [])
+    if not items:
+        print_info("Nothing to migrate.")
+        return
+
+    migrated_items = [i for i in items if i.get("status") == "migrated"]
+    conflict_items = [i for i in items if i.get("status") == "conflict"]
+    skipped_items = [i for i in items if i.get("status") == "skipped"]
+
+    warnings_shown = set()
+
+    if migrated_items:
+        print(color("  Would import:", Colors.GREEN))
+        for item in migrated_items:
+            kind = item.get("kind", "unknown")
+            dest = item.get("destination", "")
+            if dest:
+                dest_short = str(dest).replace(str(Path.home()), "~")
+                print(f"      {kind:<22s} → {dest_short}")
+            else:
+                print(f"      {kind}")
+
+            # Check for high-impact items and collect warnings
+            kind_lower = kind.lower()
+            dest_lower = str(dest).lower()
+            for keyword, warning in _HIGH_IMPACT_KIND_KEYWORDS.items():
+                if keyword in kind_lower or keyword in dest_lower:
+                    warnings_shown.add(warning)
+        print()
+
+    if conflict_items:
+        print(color("  Would overwrite (conflicts with existing Hermes config):", Colors.YELLOW))
+        for item in conflict_items:
+            kind = item.get("kind", "unknown")
+            reason = item.get("reason", "already exists")
+            print(f"      {kind:<22s}  {reason}")
+        print()
+
+    if skipped_items:
+        print(color("  Would skip:", Colors.DIM))
+        for item in skipped_items:
+            kind = item.get("kind", "unknown")
+            reason = item.get("reason", "")
+            print(f"      {kind:<22s}  {reason}")
+        print()
+
+    # Print collected warnings
+    if warnings_shown:
+        print(color("  ── Warnings ──", Colors.YELLOW))
+        for warning in sorted(warnings_shown):
+            print(color(f"    {warning}", Colors.YELLOW))
+        print()
+        print(color("  Note: OpenClaw config values may have different semantics in Hermes.", Colors.YELLOW))
+        print(color("  For example, OpenClaw's tool_call_execution: \"auto\" ≠ Hermes's yolo mode.", Colors.YELLOW))
+        print(color("  Instruction files (.md) from OpenClaw may contain incompatible procedures.", Colors.YELLOW))
+        print()
+
+
 def _offer_openclaw_migration(hermes_home: Path) -> bool:
     """Detect ~/.openclaw and offer to migrate during first-time setup.
 
+    Runs a dry-run first to show the user exactly what would be imported,
+    overwritten, or taken over. Only executes after explicit confirmation.
+
     Returns True if migration ran successfully, False otherwise.
     """
     openclaw_dir = Path.home() / ".openclaw"
@@ -2516,12 +2698,12 @@ def _offer_openclaw_migration(hermes_home: Path) -> bool:
     print()
     print_header("OpenClaw Installation Detected")
     print_info(f"Found OpenClaw data at {openclaw_dir}")
-    print_info("Hermes can import your settings, memories, skills, and API keys.")
+    print_info("Hermes can preview what would be imported before making any changes.")
     print()
 
-    if not prompt_yes_no("Would you like to import from OpenClaw?", default=True):
+    if not prompt_yes_no("Would you like to see what can be imported?", default=True):
         print_info(
-            "Skipping migration. You can run it later via the openclaw-migration skill."
+            "Skipping migration. You can run it later with: hermes claw migrate --dry-run"
         )
         return False
 
@@ -2530,34 +2712,71 @@ def _offer_openclaw_migration(hermes_home: Path) -> bool:
     if not config_path.exists():
         save_config(load_config())
 
-    # Dynamically load the migration script
+    # Load the migration module
     try:
-        spec = importlib.util.spec_from_file_location(
-            "openclaw_to_hermes", _OPENCLAW_SCRIPT
-        )
-        if spec is None or spec.loader is None:
+        mod = _load_openclaw_migration_module()
+        if mod is None:
             print_warning("Could not load migration script.")
             return False
+    except Exception as e:
+        print_warning(f"Could not load migration script: {e}")
+        logger.debug("OpenClaw migration module load error", exc_info=True)
+        return False
 
-        mod = importlib.util.module_from_spec(spec)
-        # Register in sys.modules so @dataclass can resolve the module
-        # (Python 3.11+ requires this for dynamically loaded modules)
-        import sys as _sys
-        _sys.modules[spec.name] = mod
-        try:
-            spec.loader.exec_module(mod)
-        except Exception:
-            _sys.modules.pop(spec.name, None)
-            raise
-
-        # Run migration with the "full" preset, execute mode, no overwrite
+    # ── Phase 1: Dry-run preview ──
+    try:
         selected = mod.resolve_selected_options(None, None, preset="full")
+        dry_migrator = mod.Migrator(
+            source_root=openclaw_dir.resolve(),
+            target_root=hermes_home.resolve(),
+            execute=False,  # dry-run — no files modified
+            workspace_target=None,
+            overwrite=True,  # show everything including conflicts
+            migrate_secrets=True,
+            output_dir=None,
+            selected_options=selected,
+            preset_name="full",
+        )
+        preview_report = dry_migrator.migrate()
+    except Exception as e:
+        print_warning(f"Migration preview failed: {e}")
+        logger.debug("OpenClaw migration preview error", exc_info=True)
+        return False
+
+    # Display the full preview
+    preview_summary = preview_report.get("summary", {})
+    preview_count = preview_summary.get("migrated", 0)
+
+    if preview_count == 0:
+        print()
+        print_info("Nothing to import from OpenClaw.")
+        return False
+
+    print()
+    print_header(f"Migration Preview — {preview_count} item(s) would be imported")
+    print_info("No changes have been made yet. Review the list below:")
+    print()
+    _print_migration_preview(preview_report)
+
+    # ── Phase 2: Confirm and execute ──
+    if not prompt_yes_no("Proceed with migration?", default=False):
+        print_info(
+            "Migration cancelled. You can run it later with: hermes claw migrate"
+        )
+        print_info(
+            "Use --dry-run to preview again, or --preset minimal for a lighter import."
+        )
+        return False
+
+    # Execute the migration — overwrite=False so existing Hermes configs are
+    # preserved. The user saw the preview; conflicts are skipped by default.
+    try:
         migrator = mod.Migrator(
             source_root=openclaw_dir.resolve(),
             target_root=hermes_home.resolve(),
             execute=True,
             workspace_target=None,
-            overwrite=True,
+            overwrite=False,  # preserve existing Hermes config
             migrate_secrets=True,
             output_dir=None,
             selected_options=selected,
@@ -2569,7 +2788,7 @@ def _offer_openclaw_migration(hermes_home: Path) -> bool:
         logger.debug("OpenClaw migration error", exc_info=True)
         return False
 
-    # Print summary
+    # Print final summary
     summary = report.get("summary", {})
     migrated = summary.get("migrated", 0)
     skipped = summary.get("skipped", 0)
@@ -2580,7 +2799,7 @@ def _offer_openclaw_migration(hermes_home: Path) -> bool:
     if migrated:
         print_success(f"Imported {migrated} item(s) from OpenClaw.")
     if conflicts:
-        print_info(f"Skipped {conflicts} item(s) that already exist in Hermes.")
+        print_info(f"Skipped {conflicts} item(s) that already exist in Hermes (use hermes claw migrate --overwrite to force).")
     if skipped:
         print_info(f"Skipped {skipped} item(s) (not found or unchanged).")
     if errors:

hermes_cli/skills_config.py

@@ -23,6 +23,7 @@ PLATFORMS = {
     "slack":    "💼 Slack",
     "whatsapp": "📱 WhatsApp",
     "signal":   "📡 Signal",
+    "bluebubbles": "💬 BlueBubbles",
     "email":    "📧 Email",
     "homeassistant": "🏠 Home Assistant",
     "mattermost": "💬 Mattermost",

hermes_cli/status.py

@@ -153,12 +153,14 @@ def show_status(args):
     print(color("◆ Auth Providers", Colors.CYAN, Colors.BOLD))
 
     try:
-        from hermes_cli.auth import get_nous_auth_status, get_codex_auth_status
+        from hermes_cli.auth import get_nous_auth_status, get_codex_auth_status, get_qwen_auth_status
         nous_status = get_nous_auth_status()
         codex_status = get_codex_auth_status()
+        qwen_status = get_qwen_auth_status()
     except Exception:
         nous_status = {}
         codex_status = {}
+        qwen_status = {}
 
     nous_logged_in = bool(nous_status.get("logged_in"))
     print(
@@ -189,6 +191,21 @@ def show_status(args):
     if codex_status.get("error") and not codex_logged_in:
         print(f"    Error:      {codex_status.get('error')}")
 
+    qwen_logged_in = bool(qwen_status.get("logged_in"))
+    print(
+        f"  {'Qwen OAuth':<12}  {check_mark(qwen_logged_in)} "
+        f"{'logged in' if qwen_logged_in else 'not logged in (run: qwen auth qwen-oauth)'}"
+    )
+    qwen_auth_file = qwen_status.get("auth_file")
+    if qwen_auth_file:
+        print(f"    Auth file:  {qwen_auth_file}")
+    qwen_exp = qwen_status.get("expires_at_ms")
+    if qwen_exp:
+        from datetime import datetime, timezone
+        print(f"    Access exp: {datetime.fromtimestamp(int(qwen_exp) / 1000, tz=timezone.utc).isoformat()}")
+    if qwen_status.get("error") and not qwen_logged_in:
+        print(f"    Error:      {qwen_status.get('error')}")
+
     # =========================================================================
     # Nous Subscription Features
     # =========================================================================
@@ -285,6 +302,7 @@ def show_status(args):
         "DingTalk": ("DINGTALK_CLIENT_ID", None),
         "Feishu": ("FEISHU_APP_ID", "FEISHU_HOME_CHANNEL"),
         "WeCom": ("WECOM_BOT_ID", "WECOM_HOME_CHANNEL"),
+        "BlueBubbles": ("BLUEBUBBLES_SERVER_URL", "BLUEBUBBLES_HOME_CHANNEL"),
     }
     
     for name, (token_var, home_var) in platforms.items():

hermes_cli/tools_config.py

@@ -126,6 +126,7 @@ PLATFORMS = {
     "slack":    {"label": "💼 Slack",      "default_toolset": "hermes-slack"},
     "whatsapp": {"label": "📱 WhatsApp",   "default_toolset": "hermes-whatsapp"},
     "signal":   {"label": "📡 Signal",     "default_toolset": "hermes-signal"},
+    "bluebubbles": {"label": "💙 BlueBubbles", "default_toolset": "hermes-bluebubbles"},
     "homeassistant": {"label": "🏠 Home Assistant", "default_toolset": "hermes-homeassistant"},
     "email":    {"label": "📧 Email",      "default_toolset": "hermes-email"},
     "matrix":   {"label": "💬 Matrix",     "default_toolset": "hermes-matrix"},

hermes_state.py

@@ -1235,10 +1235,10 @@ class SessionDB:
         self._execute_write(_do)
 
     def delete_session(self, session_id: str) -> bool:
-        """Delete a session, its child sessions, and all their messages.
+        """Delete a session and all its messages.
 
-        Child sessions (subagent runs, compression continuations) are deleted
-        first to satisfy the ``parent_session_id`` foreign key constraint.
+        Child sessions are orphaned (parent_session_id set to NULL) rather
+        than cascade-deleted, so they remain accessible independently.
         Returns True if the session was found and deleted.
         """
         def _do(conn):
@@ -1247,15 +1247,12 @@ class SessionDB:
             )
             if cursor.fetchone()[0] == 0:
                 return False
-            # Delete child sessions first (FK constraint)
-            child_ids = [r[0] for r in conn.execute(
-                "SELECT id FROM sessions WHERE parent_session_id = ?",
+            # Orphan child sessions so FK constraint is satisfied
+            conn.execute(
+                "UPDATE sessions SET parent_session_id = NULL "
+                "WHERE parent_session_id = ?",
                 (session_id,),
-            ).fetchall()]
-            for cid in child_ids:
-                conn.execute("DELETE FROM messages WHERE session_id = ?", (cid,))
-                conn.execute("DELETE FROM sessions WHERE id = ?", (cid,))
-            # Delete the session itself
+            )
             conn.execute("DELETE FROM messages WHERE session_id = ?", (session_id,))
             conn.execute("DELETE FROM sessions WHERE id = ?", (session_id,))
             return True
@@ -1264,9 +1261,9 @@ class SessionDB:
     def prune_sessions(self, older_than_days: int = 90, source: str = None) -> int:
         """Delete sessions older than N days. Returns count of deleted sessions.
 
-        Only prunes ended sessions (not active ones).  Child sessions whose
-        parents are being pruned are deleted first to satisfy the
-        ``parent_session_id`` foreign key constraint.
+        Only prunes ended sessions (not active ones).  Child sessions outside
+        the prune window are orphaned (parent_session_id set to NULL) rather
+        than cascade-deleted.
         """
         cutoff = time.time() - (older_than_days * 86400)
 
@@ -1284,17 +1281,16 @@ class SessionDB:
                 )
             session_ids = set(row["id"] for row in cursor.fetchall())
 
-            # Delete children first whose parents are in the prune set
-            # (avoids FK constraint errors)
-            for sid in list(session_ids):
-                child_ids = [r[0] for r in conn.execute(
-                    "SELECT id FROM sessions WHERE parent_session_id = ?",
-                    (sid,),
-                ).fetchall()]
-                for cid in child_ids:
-                    conn.execute("DELETE FROM messages WHERE session_id = ?", (cid,))
-                    conn.execute("DELETE FROM sessions WHERE id = ?", (cid,))
-                    session_ids.discard(cid)  # don't double-delete
+            if not session_ids:
+                return 0
+
+            # Orphan any sessions whose parent is about to be deleted
+            placeholders = ",".join("?" * len(session_ids))
+            conn.execute(
+                f"UPDATE sessions SET parent_session_id = NULL "
+                f"WHERE parent_session_id IN ({placeholders})",
+                list(session_ids),
+            )
 
             for sid in session_ids:
                 conn.execute("DELETE FROM messages WHERE session_id = ?", (sid,))

nix/nixosModules.nix

@@ -464,7 +464,11 @@
       addToSystemPackages = mkOption {
         type = types.bool;
         default = false;
-        description = "Add hermes CLI to environment.systemPackages.";
+        description = ''
+          Add the hermes CLI to environment.systemPackages and export
+          HERMES_HOME system-wide (via environment.variables) so interactive
+          shells share state with the gateway service.
+        '';
       };
 
       # ── OCI Container (opt-in) ──────────────────────────────────────────
@@ -545,8 +549,12 @@
       })
 
       # ── Host CLI ──────────────────────────────────────────────────────
+      # Add the hermes CLI to system PATH and export HERMES_HOME system-wide
+      # so interactive shells share state (sessions, skills, cron) with the
+      # gateway service instead of creating a separate ~/.hermes/.
       (lib.mkIf cfg.addToSystemPackages {
         environment.systemPackages = [ cfg.package ];
+        environment.variables.HERMES_HOME = "${cfg.stateDir}/.hermes";
       })
 
       # ── Directories ───────────────────────────────────────────────────
@@ -561,7 +569,7 @@
 
       # ── Activation: link config + auth + documents ────────────────────
       {
-        system.activationScripts."hermes-agent-setup" = lib.stringAfter [ "users" "setupSecrets" ] ''
+        system.activationScripts."hermes-agent-setup" = lib.stringAfter ([ "users" ] ++ lib.optional (config.system.activationScripts ? setupSecrets) "setupSecrets") ''
           # Ensure directories exist (activation runs before tmpfiles)
           mkdir -p ${cfg.stateDir}/.hermes
           mkdir -p ${cfg.stateDir}/home
@@ -601,7 +609,7 @@
           # so this is the single source of truth for both native and container mode.
           ${lib.optionalString (cfg.environment != {} || cfg.environmentFiles != []) ''
             ENV_FILE="${cfg.stateDir}/.hermes/.env"
-            install -o ${cfg.user} -g ${cfg.group} -m 0600 /dev/null "$ENV_FILE"
+            install -o ${cfg.user} -g ${cfg.group} -m 0640 /dev/null "$ENV_FILE"
             cat > "$ENV_FILE" <<'HERMES_NIX_ENV_EOF'
 ${envFileContent}
 HERMES_NIX_ENV_EOF

nix/packages.nix

@@ -14,7 +14,7 @@
       };
 
       runtimeDeps = with pkgs; [
-        nodejs_20 ripgrep git openssh ffmpeg
+        nodejs_20 ripgrep git openssh ffmpeg tirith
       ];
 
       runtimePath = pkgs.lib.makeBinPath runtimeDeps;

nix/python.nix

@@ -6,14 +6,68 @@
   uv2nix,
   pyproject-nix,
   pyproject-build-systems,
+  stdenv,
 }:
 let
   workspace = uv2nix.lib.workspace.loadWorkspace { workspaceRoot = ./..; };
+  hacks = callPackage pyproject-nix.build.hacks { };
 
   overlay = workspace.mkPyprojectOverlay {
     sourcePreference = "wheel";
   };
 
+  isAarch64Darwin = stdenv.hostPlatform.system == "aarch64-darwin";
+
+  # Keep the workspace locked through uv2nix, but supply the local voice stack
+  # from nixpkgs so wheel-only transitive artifacts do not break evaluation.
+  mkPrebuiltPassthru = dependencies: {
+    inherit dependencies;
+    optional-dependencies = { };
+    dependency-groups = { };
+  };
+
+  mkPrebuiltOverride = final: from: dependencies:
+    hacks.nixpkgsPrebuilt {
+      inherit from;
+      prev = {
+        nativeBuildInputs = [ final.pyprojectHook ];
+        passthru = mkPrebuiltPassthru dependencies;
+      };
+    };
+
+  pythonPackageOverrides = final: _prev:
+    if isAarch64Darwin then {
+      numpy = mkPrebuiltOverride final python311.pkgs.numpy { };
+
+      av = mkPrebuiltOverride final python311.pkgs.av { };
+
+      humanfriendly = mkPrebuiltOverride final python311.pkgs.humanfriendly { };
+
+      coloredlogs = mkPrebuiltOverride final python311.pkgs.coloredlogs {
+        humanfriendly = [ ];
+      };
+
+      onnxruntime = mkPrebuiltOverride final python311.pkgs.onnxruntime {
+        coloredlogs = [ ];
+        numpy = [ ];
+        packaging = [ ];
+      };
+
+      ctranslate2 = mkPrebuiltOverride final python311.pkgs.ctranslate2 {
+        numpy = [ ];
+        pyyaml = [ ];
+      };
+
+      faster-whisper = mkPrebuiltOverride final python311.pkgs.faster-whisper {
+        av = [ ];
+        ctranslate2 = [ ];
+        huggingface-hub = [ ];
+        onnxruntime = [ ];
+        tokenizers = [ ];
+        tqdm = [ ];
+      };
+    } else {};
+
   pythonSet =
     (callPackage pyproject-nix.build.packages {
       python = python311;
@@ -21,6 +75,7 @@ let
       (lib.composeManyExtensions [
         pyproject-build-systems.overlays.default
         overlay
+        pythonPackageOverrides
       ]);
 in
 pythonSet.mkVirtualEnv "hermes-agent-env" {

optional-skills/migration/openclaw-migration/scripts/openclaw_to_hermes.py

@@ -1803,30 +1803,34 @@ class Migrator:
     def migrate_cron_jobs(self, config: Optional[Dict[str, Any]] = None) -> None:
         config = config or self.load_openclaw_config()
         cron = config.get("cron") or {}
-        if not cron:
-            self.record("cron-jobs", None, None, "skipped", "No cron configuration found")
-            return
-
-        # Archive the full cron config
-        if self.archive_dir and self.execute:
-            self.archive_dir.mkdir(parents=True, exist_ok=True)
-            dest = self.archive_dir / "cron-config.json"
-            dest.write_text(json.dumps(cron, indent=2, ensure_ascii=False) + "\n", encoding="utf-8")
-            self.record("cron-jobs", "openclaw.json cron.*", str(dest), "archived",
-                        "Cron config archived. Use 'hermes cron' to recreate jobs manually.")
-        else:
-            self.record("cron-jobs", "openclaw.json cron.*", "archive/cron-config.json",
-                        "archived", "Would archive cron config")
-
-        # Also check for cron store files
         cron_store = self.source_root / "cron"
+        found_any = False
+
+        # Archive the full cron config when present
+        if cron:
+            found_any = True
+            if self.archive_dir and self.execute:
+                self.archive_dir.mkdir(parents=True, exist_ok=True)
+                dest = self.archive_dir / "cron-config.json"
+                dest.write_text(json.dumps(cron, indent=2, ensure_ascii=False) + "\n", encoding="utf-8")
+                self.record("cron-jobs", "openclaw.json cron.*", str(dest), "archived",
+                            "Cron config archived. Use 'hermes cron' to recreate jobs manually.")
+            else:
+                self.record("cron-jobs", "openclaw.json cron.*", "archive/cron-config.json",
+                            "archived", "Would archive cron config")
+
+        # Also check for cron store files even when config.cron is missing
         if cron_store.is_dir() and self.archive_dir:
+            found_any = True
             dest_cron = self.archive_dir / "cron-store"
             if self.execute:
                 shutil.copytree(cron_store, dest_cron, dirs_exist_ok=True)
             self.record("cron-jobs", str(cron_store), str(dest_cron), "archived",
                         "Cron job store archived")
 
+        if not found_any:
+            self.record("cron-jobs", None, None, "skipped", "No cron configuration found")
+
     # ── Hooks ─────────────────────────────────────────────────
     def migrate_hooks_config(self, config: Optional[Dict[str, Any]] = None) -> None:
         config = config or self.load_openclaw_config()
@@ -2454,6 +2458,15 @@ class Migrator:
                 notes.append(f"- **{item.kind}**: {item.reason}")
             notes.append("")
 
+        has_cron_config_archive = any(
+            i.kind == "cron-jobs" and i.status == "archived" and i.destination and i.destination.endswith("cron-config.json")
+            for i in self.items
+        )
+        has_cron_store_archive = any(
+            i.kind == "cron-jobs" and i.status == "archived" and i.destination and i.destination.endswith("cron-store")
+            for i in self.items
+        )
+
         notes.extend([
             "## IMPORTANT: Archive the OpenClaw Directory",
             "",
@@ -2475,7 +2488,14 @@ class Migrator:
             "- Run `hermes claw cleanup` to archive the OpenClaw directory (prevents state confusion)",
             "- Run `hermes setup` to configure any remaining settings",
             "- Run `hermes mcp list` to verify MCP servers were imported correctly",
-            "- Run `hermes cron` to recreate scheduled tasks (see archive/cron-config.json)",
+        ])
+
+        if has_cron_config_archive:
+            notes.append("- Run `hermes cron` to recreate scheduled tasks (see archive/cron-config.json)")
+        elif has_cron_store_archive:
+            notes.append("- Run `hermes cron` to recreate scheduled tasks (see archived cron-store)")
+
+        notes.extend([
             "- Run `hermes gateway install` if you need the gateway service",
             "- Review `~/.hermes/config.yaml` for any adjustments",
             "",

plugins/memory/hindsight/README.md

@@ -1,11 +1,12 @@
 # Hindsight Memory Provider
 
-Long-term memory with knowledge graph, entity resolution, and multi-strategy retrieval. Supports cloud and local (embedded) modes.
+Long-term memory with knowledge graph, entity resolution, and multi-strategy retrieval. Supports cloud, local embedded, and local external modes.
 
 ## Requirements
 
 - **Cloud:** API key from [ui.hindsight.vectorize.io](https://ui.hindsight.vectorize.io)
-- **Local:** API key for a supported LLM provider (OpenAI, Anthropic, Gemini, Groq, MiniMax, or Ollama). Embeddings and reranking run locally — no additional API keys needed.
+- **Local Embedded:** API key for a supported LLM provider (OpenAI, Anthropic, Gemini, Groq, OpenRouter, MiniMax, Ollama, or any OpenAI-compatible endpoint). Embeddings and reranking run locally — no additional API keys needed.
+- **Local External:** A running Hindsight instance (Docker or self-hosted) reachable over HTTP.
 
 ## Setup
 
@@ -21,17 +22,28 @@ hermes config set memory.provider hindsight
 echo "HINDSIGHT_API_KEY=your-key" >> ~/.hermes/.env
 ```
 
-### Cloud Mode
+### Cloud
 
 Connects to the Hindsight Cloud API. Requires an API key from [ui.hindsight.vectorize.io](https://ui.hindsight.vectorize.io).
 
-### Local Mode
+### Local Embedded
 
-Runs an embedded Hindsight server with built-in PostgreSQL. Requires an LLM API key (e.g. Groq, OpenAI, Anthropic) for memory extraction and synthesis. The daemon starts automatically in the background on first use and stops after 5 minutes of inactivity.
+Hermes spins up a local Hindsight daemon with built-in PostgreSQL. Requires an LLM API key for memory extraction and synthesis. The daemon starts automatically in the background on first use and stops after 5 minutes of inactivity.
+
+Supports any OpenAI-compatible LLM endpoint (llama.cpp, vLLM, LM Studio, etc.) — pick `openai_compatible` as the provider and enter the base URL.
 
 Daemon startup logs: `~/.hermes/logs/hindsight-embed.log`
 Daemon runtime logs: `~/.hindsight/profiles/<profile>.log`
 
+To open the Hindsight web UI (local embedded mode only):
+```bash
+hindsight-embed -p hermes ui start
+```
+
+### Local External
+
+Points the plugin at an existing Hindsight instance you're already running (Docker, self-hosted, etc.). No daemon management — just a URL and an optional API key.
+
 ## Config
 
 Config file: `~/.hermes/hindsight/config.json`
@@ -40,39 +52,58 @@ Config file: `~/.hermes/hindsight/config.json`
 
 | Key | Default | Description |
 |-----|---------|-------------|
-| `mode` | `cloud` | `cloud` or `local` |
-| `api_url` | `https://api.hindsight.vectorize.io` | API URL (cloud mode) |
-| `api_url` | `http://localhost:8888` | API URL (local mode, unused — daemon manages its own port) |
+| `mode` | `cloud` | `cloud`, `local_embedded`, or `local_external` |
+| `api_url` | `https://api.hindsight.vectorize.io` | API URL (cloud and local_external modes) |
 
-### Memory
+### Memory Bank
 
 | Key | Default | Description |
 |-----|---------|-------------|
 | `bank_id` | `hermes` | Memory bank name |
-| `budget` | `mid` | Recall thoroughness: `low` / `mid` / `high` |
+| `bank_mission` | — | Reflect mission (identity/framing for reflect reasoning). Applied via Banks API. |
+| `bank_retain_mission` | — | Retain mission (steers what gets extracted). Applied via Banks API. |
+
+### Recall
+
+| Key | Default | Description |
+|-----|---------|-------------|
+| `recall_budget` | `mid` | Recall thoroughness: `low` / `mid` / `high` |
+| `recall_prefetch_method` | `recall` | Auto-recall method: `recall` (raw facts) or `reflect` (LLM synthesis) |
+| `recall_max_tokens` | `4096` | Maximum tokens for recall results |
+| `recall_max_input_chars` | `800` | Maximum input query length for auto-recall |
+| `recall_prompt_preamble` | — | Custom preamble for recalled memories in context |
+| `recall_tags` | — | Tags to filter when searching memories |
+| `recall_tags_match` | `any` | Tag matching mode: `any` / `all` / `any_strict` / `all_strict` |
+| `auto_recall` | `true` | Automatically recall memories before each turn |
+
+### Retain
+
+| Key | Default | Description |
+|-----|---------|-------------|
+| `auto_retain` | `true` | Automatically retain conversation turns |
+| `retain_async` | `true` | Process retain asynchronously on the Hindsight server |
+| `retain_every_n_turns` | `1` | Retain every N turns (1 = every turn) |
+| `retain_context` | `conversation between Hermes Agent and the User` | Context label for retained memories |
+| `tags` | — | Tags applied when storing memories |
 
 ### Integration
 
 | Key | Default | Description |
 |-----|---------|-------------|
 | `memory_mode` | `hybrid` | How memories are integrated into the agent |
-| `prefetch_method` | `recall` | Method for automatic context injection |
 
 **memory_mode:**
 - `hybrid` — automatic context injection + tools available to the LLM
 - `context` — automatic injection only, no tools exposed
 - `tools` — tools only, no automatic injection
 
-**prefetch_method:**
-- `recall` — injects raw memory facts (fast)
-- `reflect` — injects LLM-synthesized summary (slower, more coherent)
-
-### Local Mode LLM
+### Local Embedded LLM
 
 | Key | Default | Description |
 |-----|---------|-------------|
-| `llm_provider` | `openai` | LLM provider: `openai`, `anthropic`, `gemini`, `groq`, `minimax`, `ollama` |
-| `llm_model` | per-provider | Model name (e.g. `gpt-4o-mini`, `openai/gpt-oss-120b`) |
+| `llm_provider` | `openai` | `openai`, `anthropic`, `gemini`, `groq`, `openrouter`, `minimax`, `ollama`, `lmstudio`, `openai_compatible` |
+| `llm_model` | per-provider | Model name (e.g. `gpt-4o-mini`, `qwen/qwen3.5-9b`) |
+| `llm_base_url` | — | Endpoint URL for `openai_compatible` (e.g. `http://192.168.1.10:8080/v1`) |
 
 The LLM API key is stored in `~/.hermes/.env` as `HINDSIGHT_LLM_API_KEY`.
 
@@ -92,7 +123,12 @@ Available in `hybrid` and `tools` memory modes:
 |----------|-------------|
 | `HINDSIGHT_API_KEY` | API key for Hindsight Cloud |
 | `HINDSIGHT_LLM_API_KEY` | LLM API key for local mode |
+| `HINDSIGHT_API_LLM_BASE_URL` | LLM Base URL for local mode (e.g. OpenRouter) |
 | `HINDSIGHT_API_URL` | Override API endpoint |
 | `HINDSIGHT_BANK_ID` | Override bank name |
 | `HINDSIGHT_BUDGET` | Override recall budget |
-| `HINDSIGHT_MODE` | Override mode (`cloud` / `local`) |
+| `HINDSIGHT_MODE` | Override mode (`cloud`, `local_embedded`, `local_external`) |
+
+## Client Version
+
+Requires `hindsight-client >= 0.4.22`. The plugin auto-upgrades on session start if an older version is detected.

plugins/memory/hindsight/__init__.py

@@ -28,21 +28,25 @@ from hermes_constants import get_hermes_home
 from typing import Any, Dict, List
 
 from agent.memory_provider import MemoryProvider
+from hermes_constants import get_hermes_home
 from tools.registry import tool_error
 
 logger = logging.getLogger(__name__)
 
 _DEFAULT_API_URL = "https://api.hindsight.vectorize.io"
 _DEFAULT_LOCAL_URL = "http://localhost:8888"
+_MIN_CLIENT_VERSION = "0.4.22"
 _VALID_BUDGETS = {"low", "mid", "high"}
 _PROVIDER_DEFAULT_MODELS = {
     "openai": "gpt-4o-mini",
     "anthropic": "claude-haiku-4-5",
     "gemini": "gemini-2.5-flash",
     "groq": "openai/gpt-oss-120b",
+    "openrouter": "qwen/qwen3.5-9b",
     "minimax": "MiniMax-M2.7",
     "ollama": "gemma3:12b",
     "lmstudio": "local-model",
+    "openai_compatible": "your-model-name",
 }
 
 
@@ -188,6 +192,7 @@ class HindsightMemoryProvider(MemoryProvider):
         self._bank_id = "hermes"
         self._budget = "mid"
         self._mode = "cloud"
+        self._llm_base_url = ""
         self._memory_mode = "hybrid"  # "context", "tools", or "hybrid"
         self._prefetch_method = "recall"  # "recall" or "reflect"
         self._client = None
@@ -195,6 +200,31 @@ class HindsightMemoryProvider(MemoryProvider):
         self._prefetch_lock = threading.Lock()
         self._prefetch_thread = None
         self._sync_thread = None
+        self._session_id = ""
+
+        # Tags
+        self._tags: list[str] | None = None
+        self._recall_tags: list[str] | None = None
+        self._recall_tags_match = "any"
+
+        # Retain controls
+        self._auto_retain = True
+        self._retain_every_n_turns = 1
+        self._retain_context = "conversation between Hermes Agent and the User"
+        self._turn_counter = 0
+        self._session_turns: list[str] = []  # accumulates ALL turns for the session
+
+        # Recall controls
+        self._auto_recall = True
+        self._recall_max_tokens = 4096
+        self._recall_types: list[str] | None = None
+        self._recall_prompt_preamble = ""
+        self._recall_max_input_chars = 800
+
+        # Bank
+        self._bank_mission = ""
+        self._bank_retain_mission: str | None = None
+        self._retain_async = True
 
     @property
     def name(self) -> str:
@@ -204,7 +234,7 @@ class HindsightMemoryProvider(MemoryProvider):
         try:
             cfg = _load_config()
             mode = cfg.get("mode", "cloud")
-            if mode == "local":
+            if mode in ("local", "local_embedded", "local_external"):
                 return True
             has_key = bool(cfg.get("apiKey") or os.environ.get("HINDSIGHT_API_KEY", ""))
             has_url = bool(cfg.get("api_url") or os.environ.get("HINDSIGHT_API_URL", ""))
@@ -228,68 +258,306 @@ class HindsightMemoryProvider(MemoryProvider):
         existing.update(values)
         config_path.write_text(json.dumps(existing, indent=2))
 
+    def post_setup(self, hermes_home: str, config: dict) -> None:
+        """Custom setup wizard — installs only the deps needed for the selected mode."""
+        import getpass
+        import subprocess
+        import shutil
+        import sys
+        from pathlib import Path
+
+        from hermes_cli.config import save_config
+
+        from hermes_cli.memory_setup import _curses_select
+
+        print("\n  Configuring Hindsight memory:\n")
+
+        # Step 1: Mode selection
+        mode_items = [
+            ("Cloud", "Hindsight Cloud API (lightweight, just needs an API key)"),
+            ("Local Embedded", "Run Hindsight locally (downloads ~200MB, needs LLM key)"),
+            ("Local External", "Connect to an existing Hindsight instance"),
+        ]
+        mode_idx = _curses_select("  Select mode", mode_items, default=0)
+        mode = ["cloud", "local_embedded", "local_external"][mode_idx]
+
+        provider_config: dict = {"mode": mode}
+        env_writes: dict = {}
+
+        # Step 2: Install/upgrade deps for selected mode
+        _MIN_CLIENT_VERSION = "0.4.22"
+        cloud_dep = f"hindsight-client>={_MIN_CLIENT_VERSION}"
+        local_dep = "hindsight-all"
+        if mode == "local_embedded":
+            deps_to_install = [local_dep]
+        elif mode == "local_external":
+            deps_to_install = [cloud_dep]
+        else:
+            deps_to_install = [cloud_dep]
+
+        print(f"\n  Checking dependencies...")
+        uv_path = shutil.which("uv")
+        if not uv_path:
+            print("  ⚠ uv not found — install it: curl -LsSf https://astral.sh/uv/install.sh | sh")
+            print(f"  Then run manually: uv pip install --python {sys.executable} {' '.join(deps_to_install)}")
+        else:
+            try:
+                subprocess.run(
+                    [uv_path, "pip", "install", "--python", sys.executable, "--quiet", "--upgrade"] + deps_to_install,
+                    check=True, timeout=120, capture_output=True,
+                )
+                print(f"  ✓ Dependencies up to date")
+            except Exception as e:
+                print(f"  ⚠ Install failed: {e}")
+                print(f"  Run manually: uv pip install --python {sys.executable} {' '.join(deps_to_install)}")
+
+        # Step 3: Mode-specific config
+        if mode == "cloud":
+            print(f"\n  Get your API key at https://ui.hindsight.vectorize.io\n")
+            existing_key = os.environ.get("HINDSIGHT_API_KEY", "")
+            if existing_key:
+                masked = f"...{existing_key[-4:]}" if len(existing_key) > 4 else "set"
+                sys.stdout.write(f"  API key (current: {masked}, blank to keep): ")
+                sys.stdout.flush()
+                api_key = getpass.getpass(prompt="") if sys.stdin.isatty() else sys.stdin.readline().strip()
+            else:
+                sys.stdout.write("  API key: ")
+                sys.stdout.flush()
+                api_key = getpass.getpass(prompt="") if sys.stdin.isatty() else sys.stdin.readline().strip()
+            if api_key:
+                env_writes["HINDSIGHT_API_KEY"] = api_key
+
+            val = input(f"  API URL [{_DEFAULT_API_URL}]: ").strip()
+            if val:
+                provider_config["api_url"] = val
+
+        elif mode == "local_external":
+            val = input(f"  Hindsight API URL [{_DEFAULT_LOCAL_URL}]: ").strip()
+            provider_config["api_url"] = val or _DEFAULT_LOCAL_URL
+
+            sys.stdout.write("  API key (optional, blank to skip): ")
+            sys.stdout.flush()
+            api_key = getpass.getpass(prompt="") if sys.stdin.isatty() else sys.stdin.readline().strip()
+            if api_key:
+                env_writes["HINDSIGHT_API_KEY"] = api_key
+
+        else:  # local_embedded
+            providers_list = list(_PROVIDER_DEFAULT_MODELS.keys())
+            llm_items = [
+                (p, f"default model: {_PROVIDER_DEFAULT_MODELS[p]}")
+                for p in providers_list
+            ]
+            llm_idx = _curses_select("  Select LLM provider", llm_items, default=0)
+            llm_provider = providers_list[llm_idx]
+
+            provider_config["llm_provider"] = llm_provider
+
+            if llm_provider == "openai_compatible":
+                val = input("  LLM endpoint URL (e.g. http://192.168.1.10:8080/v1): ").strip()
+                if val:
+                    provider_config["llm_base_url"] = val
+            elif llm_provider == "openrouter":
+                provider_config["llm_base_url"] = "https://openrouter.ai/api/v1"
+
+            default_model = _PROVIDER_DEFAULT_MODELS.get(llm_provider, "gpt-4o-mini")
+            val = input(f"  LLM model [{default_model}]: ").strip()
+            provider_config["llm_model"] = val or default_model
+
+            sys.stdout.write("  LLM API key: ")
+            sys.stdout.flush()
+            llm_key = getpass.getpass(prompt="") if sys.stdin.isatty() else sys.stdin.readline().strip()
+            if llm_key:
+                env_writes["HINDSIGHT_LLM_API_KEY"] = llm_key
+
+        # Step 4: Save everything
+        provider_config["bank_id"] = "hermes"
+        provider_config["recall_budget"] = "mid"
+        bank_id = "hermes"
+        config["memory"]["provider"] = "hindsight"
+        save_config(config)
+
+        self.save_config(provider_config, hermes_home)
+
+        if env_writes:
+            env_path = Path(hermes_home) / ".env"
+            env_path.parent.mkdir(parents=True, exist_ok=True)
+            existing_lines = []
+            if env_path.exists():
+                existing_lines = env_path.read_text().splitlines()
+            updated_keys = set()
+            new_lines = []
+            for line in existing_lines:
+                key_match = line.split("=", 1)[0].strip() if "=" in line and not line.startswith("#") else None
+                if key_match and key_match in env_writes:
+                    new_lines.append(f"{key_match}={env_writes[key_match]}")
+                    updated_keys.add(key_match)
+                else:
+                    new_lines.append(line)
+            for k, v in env_writes.items():
+                if k not in updated_keys:
+                    new_lines.append(f"{k}={v}")
+            env_path.write_text("\n".join(new_lines) + "\n")
+
+        print(f"\n  ✓ Hindsight memory configured ({mode} mode)")
+        if env_writes:
+            print(f"  API keys saved to .env")
+        print(f"\n  Start a new session to activate.\n")
+
     def get_config_schema(self):
         return [
-            {"key": "mode", "description": "Cloud API or local embedded mode", "default": "cloud", "choices": ["cloud", "local"]},
-            {"key": "api_url", "description": "Hindsight API URL", "default": _DEFAULT_API_URL, "when": {"mode": "cloud"}},
+            {"key": "mode", "description": "Connection mode", "default": "cloud", "choices": ["cloud", "local_embedded", "local_external"]},
+            # Cloud mode
+            {"key": "api_url", "description": "Hindsight Cloud API URL", "default": _DEFAULT_API_URL, "when": {"mode": "cloud"}},
             {"key": "api_key", "description": "Hindsight Cloud API key", "secret": True, "env_var": "HINDSIGHT_API_KEY", "url": "https://ui.hindsight.vectorize.io", "when": {"mode": "cloud"}},
-            {"key": "llm_provider", "description": "LLM provider for local mode", "default": "openai", "choices": ["openai", "anthropic", "gemini", "groq", "minimax", "ollama"], "when": {"mode": "local"}},
-            {"key": "llm_api_key", "description": "LLM API key for local Hindsight", "secret": True, "env_var": "HINDSIGHT_LLM_API_KEY", "when": {"mode": "local"}},
-            {"key": "llm_model", "description": "LLM model for local mode", "default": "gpt-4o-mini", "default_from": {"field": "llm_provider", "map": _PROVIDER_DEFAULT_MODELS}, "when": {"mode": "local"}},
+            # Local external mode
+            {"key": "api_url", "description": "Hindsight API URL", "default": _DEFAULT_LOCAL_URL, "when": {"mode": "local_external"}},
+            {"key": "api_key", "description": "API key (optional)", "secret": True, "env_var": "HINDSIGHT_API_KEY", "when": {"mode": "local_external"}},
+            # Local embedded mode
+            {"key": "llm_provider", "description": "LLM provider", "default": "openai", "choices": ["openai", "anthropic", "gemini", "groq", "openrouter", "minimax", "ollama", "lmstudio", "openai_compatible"], "when": {"mode": "local_embedded"}},
+            {"key": "llm_base_url", "description": "Endpoint URL (e.g. http://192.168.1.10:8080/v1)", "default": "", "when": {"mode": "local_embedded", "llm_provider": "openai_compatible"}},
+            {"key": "llm_api_key", "description": "LLM API key (optional for openai_compatible)", "secret": True, "env_var": "HINDSIGHT_LLM_API_KEY", "when": {"mode": "local_embedded"}},
+            {"key": "llm_model", "description": "LLM model", "default": "gpt-4o-mini", "default_from": {"field": "llm_provider", "map": _PROVIDER_DEFAULT_MODELS}, "when": {"mode": "local_embedded"}},
             {"key": "bank_id", "description": "Memory bank name", "default": "hermes"},
-            {"key": "budget", "description": "Recall thoroughness", "default": "mid", "choices": ["low", "mid", "high"]},
+            {"key": "bank_mission", "description": "Mission/purpose description for the memory bank"},
+            {"key": "bank_retain_mission", "description": "Custom extraction prompt for memory retention"},
+            {"key": "recall_budget", "description": "Recall thoroughness", "default": "mid", "choices": ["low", "mid", "high"]},
             {"key": "memory_mode", "description": "Memory integration mode", "default": "hybrid", "choices": ["hybrid", "context", "tools"]},
-            {"key": "prefetch_method", "description": "Auto-recall method", "default": "recall", "choices": ["recall", "reflect"]},
+            {"key": "recall_prefetch_method", "description": "Auto-recall method", "default": "recall", "choices": ["recall", "reflect"]},
+            {"key": "tags", "description": "Tags applied when storing memories (comma-separated)", "default": ""},
+            {"key": "recall_tags", "description": "Tags to filter when searching memories (comma-separated)", "default": ""},
+            {"key": "recall_tags_match", "description": "Tag matching mode for recall", "default": "any", "choices": ["any", "all", "any_strict", "all_strict"]},
+            {"key": "auto_recall", "description": "Automatically recall memories before each turn", "default": True},
+            {"key": "auto_retain", "description": "Automatically retain conversation turns", "default": True},
+            {"key": "retain_every_n_turns", "description": "Retain every N turns (1 = every turn)", "default": 1},
+            {"key": "retain_async","description": "Process retain asynchronously on the Hindsight server", "default": True},
+            {"key": "retain_context", "description": "Context label for retained memories", "default": "conversation between Hermes Agent and the User"},
+            {"key": "recall_max_tokens", "description": "Maximum tokens for recall results", "default": 4096},
+            {"key": "recall_max_input_chars", "description": "Maximum input query length for auto-recall", "default": 800},
+            {"key": "recall_prompt_preamble", "description": "Custom preamble for recalled memories in context"},
         ]
 
     def _get_client(self):
         """Return the cached Hindsight client (created once, reused)."""
         if self._client is None:
-            if self._mode == "local":
+            if self._mode == "local_embedded":
                 from hindsight import HindsightEmbedded
-                # Disable __del__ on the class to prevent "attached to a
-                # different loop" errors during GC — we handle cleanup in
-                # shutdown() instead.
                 HindsightEmbedded.__del__ = lambda self: None
-                self._client = HindsightEmbedded(
+                llm_provider = self._config.get("llm_provider", "")
+                if llm_provider in ("openai_compatible", "openrouter"):
+                    llm_provider = "openai"
+                logger.debug("Creating HindsightEmbedded client (profile=%s, provider=%s)",
+                             self._config.get("profile", "hermes"), llm_provider)
+                kwargs = dict(
                     profile=self._config.get("profile", "hermes"),
-                    llm_provider=self._config.get("llm_provider", ""),
-                    llm_api_key=self._config.get("llmApiKey") or os.environ.get("HINDSIGHT_LLM_API_KEY", ""),
+                    llm_provider=llm_provider,
+                    llm_api_key=self._config.get("llmApiKey") or self._config.get("llm_api_key") or os.environ.get("HINDSIGHT_LLM_API_KEY", ""),
                     llm_model=self._config.get("llm_model", ""),
                 )
+                if self._llm_base_url:
+                    kwargs["llm_base_url"] = self._llm_base_url
+                self._client = HindsightEmbedded(**kwargs)
             else:
                 from hindsight_client import Hindsight
                 kwargs = {"base_url": self._api_url, "timeout": 30.0}
                 if self._api_key:
                     kwargs["api_key"] = self._api_key
+                logger.debug("Creating Hindsight cloud client (url=%s, has_key=%s)",
+                             self._api_url, bool(self._api_key))
                 self._client = Hindsight(**kwargs)
         return self._client
 
     def initialize(self, session_id: str, **kwargs) -> None:
+        self._session_id = session_id
+
+        # Check client version and auto-upgrade if needed
+        try:
+            from importlib.metadata import version as pkg_version
+            from packaging.version import Version
+            installed = pkg_version("hindsight-client")
+            if Version(installed) < Version(_MIN_CLIENT_VERSION):
+                logger.warning("hindsight-client %s is outdated (need >=%s), attempting upgrade...",
+                               installed, _MIN_CLIENT_VERSION)
+                import shutil, subprocess, sys
+                uv_path = shutil.which("uv")
+                if uv_path:
+                    try:
+                        subprocess.run(
+                            [uv_path, "pip", "install", "--python", sys.executable,
+                             "--quiet", "--upgrade", f"hindsight-client>={_MIN_CLIENT_VERSION}"],
+                            check=True, timeout=120, capture_output=True,
+                        )
+                        logger.info("hindsight-client upgraded to >=%s", _MIN_CLIENT_VERSION)
+                    except Exception as e:
+                        logger.warning("Auto-upgrade failed: %s. Run: uv pip install 'hindsight-client>=%s'",
+                                       e, _MIN_CLIENT_VERSION)
+                else:
+                    logger.warning("uv not found. Run: pip install 'hindsight-client>=%s'", _MIN_CLIENT_VERSION)
+        except Exception:
+            pass  # packaging not available or other issue — proceed anyway
+
         self._config = _load_config()
         self._mode = self._config.get("mode", "cloud")
-        self._api_key = self._config.get("apiKey") or os.environ.get("HINDSIGHT_API_KEY", "")
-        default_url = _DEFAULT_LOCAL_URL if self._mode == "local" else _DEFAULT_API_URL
+        # "local" is a legacy alias for "local_embedded"
+        if self._mode == "local":
+            self._mode = "local_embedded"
+        self._api_key = self._config.get("apiKey") or self._config.get("api_key") or os.environ.get("HINDSIGHT_API_KEY", "")
+        default_url = _DEFAULT_LOCAL_URL if self._mode in ("local_embedded", "local_external") else _DEFAULT_API_URL
         self._api_url = self._config.get("api_url") or os.environ.get("HINDSIGHT_API_URL", default_url)
+        self._llm_base_url = self._config.get("llm_base_url", "")
 
         banks = self._config.get("banks", {}).get("hermes", {})
         self._bank_id = self._config.get("bank_id") or banks.get("bankId", "hermes")
-        budget = self._config.get("budget") or banks.get("budget", "mid")
+        budget = self._config.get("recall_budget") or self._config.get("budget") or banks.get("budget", "mid")
         self._budget = budget if budget in _VALID_BUDGETS else "mid"
 
         memory_mode = self._config.get("memory_mode", "hybrid")
         self._memory_mode = memory_mode if memory_mode in ("context", "tools", "hybrid") else "hybrid"
 
-        prefetch_method = self._config.get("prefetch_method", "recall")
+        prefetch_method = self._config.get("recall_prefetch_method", "recall")
         self._prefetch_method = prefetch_method if prefetch_method in ("recall", "reflect") else "recall"
 
-        logger.info("Hindsight initialized: mode=%s, api_url=%s, bank=%s, budget=%s, memory_mode=%s, prefetch_method=%s",
-                     self._mode, self._api_url, self._bank_id, self._budget, self._memory_mode, self._prefetch_method)
+        # Bank options
+        self._bank_mission = self._config.get("bank_mission", "")
+        self._bank_retain_mission = self._config.get("bank_retain_mission") or None
+
+        # Tags
+        self._tags = self._config.get("tags") or None
+        self._recall_tags = self._config.get("recall_tags") or None
+        self._recall_tags_match = self._config.get("recall_tags_match", "any")
+
+        # Retain controls
+        self._auto_retain = self._config.get("auto_retain", True)
+        self._retain_every_n_turns = max(1, int(self._config.get("retain_every_n_turns", 1)))
+        self._retain_context = self._config.get("retain_context", "conversation between Hermes Agent and the User")
+
+        # Recall controls
+        self._auto_recall = self._config.get("auto_recall", True)
+        self._recall_max_tokens = int(self._config.get("recall_max_tokens", 4096))
+        self._recall_types = self._config.get("recall_types") or None
+        self._recall_prompt_preamble = self._config.get("recall_prompt_preamble", "")
+        self._recall_max_input_chars = int(self._config.get("recall_max_input_chars", 800))
+        self._retain_async = self._config.get("retain_async", True)
+
+        _client_version = "unknown"
+        try:
+            from importlib.metadata import version as pkg_version
+            _client_version = pkg_version("hindsight-client")
+        except Exception:
+            pass
+        logger.info("Hindsight initialized: mode=%s, api_url=%s, bank=%s, budget=%s, memory_mode=%s, prefetch_method=%s, client=%s",
+                     self._mode, self._api_url, self._bank_id, self._budget, self._memory_mode, self._prefetch_method, _client_version)
+        logger.debug("Hindsight config: auto_retain=%s, auto_recall=%s, retain_every_n=%d, "
+                     "retain_async=%s, retain_context=%s, "
+                     "recall_max_tokens=%d, recall_max_input_chars=%d, tags=%s, recall_tags=%s",
+                     self._auto_retain, self._auto_recall, self._retain_every_n_turns,
+                     self._retain_async, self._retain_context,
+                     self._recall_max_tokens, self._recall_max_input_chars,
+                     self._tags, self._recall_tags)
 
         # For local mode, start the embedded daemon in the background so it
         # doesn't block the chat. Redirect stdout/stderr to a log file to
         # prevent rich startup output from spamming the terminal.
-        if self._mode == "local":
+        if self._mode == "local_embedded":
             def _start_daemon():
                 import traceback
                 log_dir = get_hermes_home() / "logs"
@@ -311,9 +579,12 @@ class HindsightMemoryProvider(MemoryProvider):
                     # If the config changed and the daemon is running, stop it.
                     from pathlib import Path as _Path
                     profile_env = _Path.home() / ".hindsight" / "profiles" / f"{profile}.env"
-                    current_key = self._config.get("llmApiKey") or os.environ.get("HINDSIGHT_LLM_API_KEY", "")
+                    current_key = self._config.get("llm_api_key") or os.environ.get("HINDSIGHT_LLM_API_KEY", "")
                     current_provider = self._config.get("llm_provider", "")
                     current_model = self._config.get("llm_model", "")
+                    current_base_url = self._config.get("llm_base_url") or os.environ.get("HINDSIGHT_API_LLM_BASE_URL", "")
+                    # Map openai_compatible/openrouter → openai for the daemon (OpenAI wire format)
+                    daemon_provider = "openai" if current_provider in ("openai_compatible", "openrouter") else current_provider
 
                     # Read saved profile config
                     saved = {}
@@ -324,20 +595,24 @@ class HindsightMemoryProvider(MemoryProvider):
                                 saved[k.strip()] = v.strip()
 
                     config_changed = (
-                        saved.get("HINDSIGHT_API_LLM_PROVIDER") != current_provider or
+                        saved.get("HINDSIGHT_API_LLM_PROVIDER") != daemon_provider or
                         saved.get("HINDSIGHT_API_LLM_MODEL") != current_model or
-                        saved.get("HINDSIGHT_API_LLM_API_KEY") != current_key
+                        saved.get("HINDSIGHT_API_LLM_API_KEY") != current_key or
+                        saved.get("HINDSIGHT_API_LLM_BASE_URL", "") != current_base_url
                     )
 
                     if config_changed:
                         # Write updated profile .env
                         profile_env.parent.mkdir(parents=True, exist_ok=True)
-                        profile_env.write_text(
-                            f"HINDSIGHT_API_LLM_PROVIDER={current_provider}\n"
+                        env_lines = (
+                            f"HINDSIGHT_API_LLM_PROVIDER={daemon_provider}\n"
                             f"HINDSIGHT_API_LLM_API_KEY={current_key}\n"
                             f"HINDSIGHT_API_LLM_MODEL={current_model}\n"
                             f"HINDSIGHT_API_LOG_LEVEL=info\n"
                         )
+                        if current_base_url:
+                            env_lines += f"HINDSIGHT_API_LLM_BASE_URL={current_base_url}\n"
+                        profile_env.write_text(env_lines)
                         if client._manager.is_running(profile):
                             with open(log_path, "a") as f:
                                 f.write("\n=== Config changed, restarting daemon ===\n")
@@ -378,47 +653,118 @@ class HindsightMemoryProvider(MemoryProvider):
 
     def prefetch(self, query: str, *, session_id: str = "") -> str:
         if self._prefetch_thread and self._prefetch_thread.is_alive():
+            logger.debug("Prefetch: waiting for background thread to complete")
             self._prefetch_thread.join(timeout=3.0)
         with self._prefetch_lock:
             result = self._prefetch_result
             self._prefetch_result = ""
         if not result:
+            logger.debug("Prefetch: no results available")
             return ""
-        return f"## Hindsight Memory\n{result}"
+        logger.debug("Prefetch: returning %d chars of context", len(result))
+        header = self._recall_prompt_preamble or (
+            "# Hindsight Memory (persistent cross-session context)\n"
+            "Use this to answer questions about the user and prior sessions. "
+            "Do not call tools to look up information that is already present here."
+        )
+        return f"{header}\n\n{result}"
 
     def queue_prefetch(self, query: str, *, session_id: str = "") -> None:
         if self._memory_mode == "tools":
+            logger.debug("Prefetch: skipped (tools-only mode)")
             return
+        if not self._auto_recall:
+            logger.debug("Prefetch: skipped (auto_recall disabled)")
+            return
+        # Truncate query to max chars
+        if self._recall_max_input_chars and len(query) > self._recall_max_input_chars:
+            query = query[:self._recall_max_input_chars]
+
         def _run():
             try:
                 client = self._get_client()
                 if self._prefetch_method == "reflect":
+                    logger.debug("Prefetch: calling reflect (bank=%s, query_len=%d)", self._bank_id, len(query))
                     resp = _run_sync(client.areflect(bank_id=self._bank_id, query=query, budget=self._budget))
                     text = resp.text or ""
                 else:
-                    resp = _run_sync(client.arecall(bank_id=self._bank_id, query=query, budget=self._budget))
-                    text = "\n".join(r.text for r in resp.results if r.text) if resp.results else ""
+                    recall_kwargs: dict = {
+                        "bank_id": self._bank_id, "query": query,
+                        "budget": self._budget, "max_tokens": self._recall_max_tokens,
+                    }
+                    if self._recall_tags:
+                        recall_kwargs["tags"] = self._recall_tags
+                        recall_kwargs["tags_match"] = self._recall_tags_match
+                    if self._recall_types:
+                        recall_kwargs["types"] = self._recall_types
+                    logger.debug("Prefetch: calling recall (bank=%s, query_len=%d, budget=%s)",
+                                 self._bank_id, len(query), self._budget)
+                    resp = _run_sync(client.arecall(**recall_kwargs))
+                    num_results = len(resp.results) if resp.results else 0
+                    logger.debug("Prefetch: recall returned %d results", num_results)
+                    text = "\n".join(f"- {r.text}" for r in resp.results if r.text) if resp.results else ""
                 if text:
                     with self._prefetch_lock:
                         self._prefetch_result = text
             except Exception as e:
-                logger.debug("Hindsight prefetch failed: %s", e)
+                logger.debug("Hindsight prefetch failed: %s", e, exc_info=True)
 
         self._prefetch_thread = threading.Thread(target=_run, daemon=True, name="hindsight-prefetch")
         self._prefetch_thread.start()
 
     def sync_turn(self, user_content: str, assistant_content: str, *, session_id: str = "") -> None:
-        """Retain conversation turn in background (non-blocking)."""
-        combined = f"User: {user_content}\nAssistant: {assistant_content}"
+        """Retain conversation turn in background (non-blocking).
+
+        Respects retain_every_n_turns for batching.
+        """
+        if not self._auto_retain:
+            logger.debug("sync_turn: skipped (auto_retain disabled)")
+            return
+
+        from datetime import datetime, timezone
+        now = datetime.now(timezone.utc).isoformat()
+
+        messages = [
+            {"role": "user", "content": user_content, "timestamp": now},
+            {"role": "assistant", "content": assistant_content, "timestamp": now},
+        ]
+
+        turn = json.dumps(messages)
+        self._session_turns.append(turn)
+        self._turn_counter += 1
+
+        # Only retain every N turns
+        if self._turn_counter % self._retain_every_n_turns != 0:
+            logger.debug("sync_turn: buffered turn %d (will retain at turn %d)",
+                         self._turn_counter, self._turn_counter + (self._retain_every_n_turns - self._turn_counter % self._retain_every_n_turns))
+            return
+
+        logger.debug("sync_turn: retaining %d turns, total session content %d chars",
+                     len(self._session_turns), sum(len(t) for t in self._session_turns))
+        # Send the ENTIRE session as a single JSON array (document_id deduplicates).
+        # Each element in _session_turns is a JSON string of that turn's messages.
+        content = "[" + ",".join(self._session_turns) + "]"
 
         def _sync():
             try:
                 client = self._get_client()
-                _run_sync(client.aretain(
-                    bank_id=self._bank_id, content=combined, context="conversation"
+                item: dict = {
+                    "content": content,
+                    "context": self._retain_context,
+                }
+                if self._tags:
+                    item["tags"] = self._tags
+                logger.debug("Hindsight retain: bank=%s, doc=%s, async=%s, content_len=%d, num_turns=%d",
+                             self._bank_id, self._session_id, self._retain_async, len(content), len(self._session_turns))
+                _run_sync(client.aretain_batch(
+                    bank_id=self._bank_id,
+                    items=[item],
+                    document_id=self._session_id,
+                    retain_async=self._retain_async,
                 ))
+                logger.debug("Hindsight retain succeeded")
             except Exception as e:
-                logger.warning("Hindsight sync failed: %s", e)
+                logger.warning("Hindsight sync failed: %s", e, exc_info=True)
 
         if self._sync_thread and self._sync_thread.is_alive():
             self._sync_thread.join(timeout=5.0)
@@ -443,12 +789,18 @@ class HindsightMemoryProvider(MemoryProvider):
                 return tool_error("Missing required parameter: content")
             context = args.get("context")
             try:
-                _run_sync(client.aretain(
-                    bank_id=self._bank_id, content=content, context=context
-                ))
+                retain_kwargs: dict = {
+                    "bank_id": self._bank_id, "content": content, "context": context,
+                }
+                if self._tags:
+                    retain_kwargs["tags"] = self._tags
+                logger.debug("Tool hindsight_retain: bank=%s, content_len=%d, context=%s",
+                             self._bank_id, len(content), context)
+                _run_sync(client.aretain(**retain_kwargs))
+                logger.debug("Tool hindsight_retain: success")
                 return json.dumps({"result": "Memory stored successfully."})
             except Exception as e:
-                logger.warning("hindsight_retain failed: %s", e)
+                logger.warning("hindsight_retain failed: %s", e, exc_info=True)
                 return tool_error(f"Failed to store memory: {e}")
 
         elif tool_name == "hindsight_recall":
@@ -456,15 +808,26 @@ class HindsightMemoryProvider(MemoryProvider):
             if not query:
                 return tool_error("Missing required parameter: query")
             try:
-                resp = _run_sync(client.arecall(
-                    bank_id=self._bank_id, query=query, budget=self._budget
-                ))
+                recall_kwargs: dict = {
+                    "bank_id": self._bank_id, "query": query, "budget": self._budget,
+                    "max_tokens": self._recall_max_tokens,
+                }
+                if self._recall_tags:
+                    recall_kwargs["tags"] = self._recall_tags
+                    recall_kwargs["tags_match"] = self._recall_tags_match
+                if self._recall_types:
+                    recall_kwargs["types"] = self._recall_types
+                logger.debug("Tool hindsight_recall: bank=%s, query_len=%d, budget=%s",
+                             self._bank_id, len(query), self._budget)
+                resp = _run_sync(client.arecall(**recall_kwargs))
+                num_results = len(resp.results) if resp.results else 0
+                logger.debug("Tool hindsight_recall: %d results", num_results)
                 if not resp.results:
                     return json.dumps({"result": "No relevant memories found."})
                 lines = [f"{i}. {r.text}" for i, r in enumerate(resp.results, 1)]
                 return json.dumps({"result": "\n".join(lines)})
             except Exception as e:
-                logger.warning("hindsight_recall failed: %s", e)
+                logger.warning("hindsight_recall failed: %s", e, exc_info=True)
                 return tool_error(f"Failed to search memory: {e}")
 
         elif tool_name == "hindsight_reflect":
@@ -472,24 +835,28 @@ class HindsightMemoryProvider(MemoryProvider):
             if not query:
                 return tool_error("Missing required parameter: query")
             try:
+                logger.debug("Tool hindsight_reflect: bank=%s, query_len=%d, budget=%s",
+                             self._bank_id, len(query), self._budget)
                 resp = _run_sync(client.areflect(
                     bank_id=self._bank_id, query=query, budget=self._budget
                 ))
+                logger.debug("Tool hindsight_reflect: response_len=%d", len(resp.text or ""))
                 return json.dumps({"result": resp.text or "No relevant memories found."})
             except Exception as e:
-                logger.warning("hindsight_reflect failed: %s", e)
+                logger.warning("hindsight_reflect failed: %s", e, exc_info=True)
                 return tool_error(f"Failed to reflect: {e}")
 
         return tool_error(f"Unknown tool: {tool_name}")
 
     def shutdown(self) -> None:
+        logger.debug("Hindsight shutdown: waiting for background threads")
         global _loop, _loop_thread
         for t in (self._prefetch_thread, self._sync_thread):
             if t and t.is_alive():
                 t.join(timeout=5.0)
         if self._client is not None:
             try:
-                if self._mode == "local":
+                if self._mode == "local_embedded":
                     # Use the public close() API. The RuntimeError from
                     # aiohttp's "attached to a different loop" is expected
                     # and harmless — the daemon keeps running independently.

plugins/memory/hindsight/plugin.yaml

@@ -2,9 +2,7 @@ name: hindsight
 version: 1.0.0
 description: "Hindsight — long-term memory with knowledge graph, entity resolution, and multi-strategy retrieval."
 pip_dependencies:
-  - hindsight-client
-  - hindsight-all
-requires_env:
-  - HINDSIGHT_API_KEY
+  - "hindsight-client>=0.4.22"
+requires_env: []
 hooks:
   - on_session_end

pyproject.toml

@@ -62,6 +62,7 @@ mcp = ["mcp>=1.2.0,<2"]
 homeassistant = ["aiohttp>=3.9.0,<4"]
 sms = ["aiohttp>=3.9.0,<4"]
 acp = ["agent-client-protocol>=0.9.0,<1.0"]
+mistral = ["mistralai>=2.3.0,<3"]
 dingtalk = ["dingtalk-stream>=0.1.0,<1"]
 feishu = ["lark-oapi>=1.5.3,<2"]
 rl = [
@@ -94,6 +95,7 @@ all = [
   "hermes-agent[voice]",
   "hermes-agent[dingtalk]",
   "hermes-agent[feishu]",
+  "hermes-agent[mistral]",
 ]
 
 [project.scripts]

run_agent.py

@@ -66,7 +66,7 @@ from model_tools import (
     handle_function_call,
     check_toolset_requirements,
 )
-from tools.terminal_tool import cleanup_vm, get_active_env
+from tools.terminal_tool import cleanup_vm, get_active_env, is_persistent_env
 from tools.tool_result_storage import maybe_persist_tool_result, enforce_turn_budget
 from tools.interrupt import set_interrupt as _set_interrupt
 from tools.browser_tool import cleanup_browser
@@ -77,6 +77,7 @@ from hermes_constants import OPENROUTER_BASE_URL
 # Agent internals extracted to agent/ package for modularity
 from agent.memory_manager import build_memory_context_block
 from agent.retry_utils import jittered_backoff
+from agent.error_classifier import classify_api_error, FailoverReason
 from agent.prompt_builder import (
     DEFAULT_AGENT_IDENTITY, PLATFORM_HINTS,
     MEMORY_GUIDANCE, SESSION_SEARCH_GUIDANCE, SKILLS_GUIDANCE,
@@ -86,6 +87,7 @@ from agent.model_metadata import (
     fetch_model_metadata,
     estimate_tokens_rough, estimate_messages_tokens_rough, estimate_request_tokens_rough,
     get_next_probe_tier, parse_context_limit_from_error,
+    parse_available_output_tokens_from_error,
     save_context_length, is_local_endpoint,
     query_ollama_num_ctx,
 )
@@ -413,6 +415,27 @@ def _strip_budget_warnings_from_history(messages: list) -> None:
 # =========================================================================
 
 
+# =========================================================================
+# Qwen Portal headers — mimics QwenCode CLI for portal.qwen.ai compatibility.
+# Extracted as a module-level helper so both __init__ and
+# _apply_client_headers_for_base_url can share it.
+# =========================================================================
+_QWEN_CODE_VERSION = "0.14.1"
+
+
+def _qwen_portal_headers() -> dict:
+    """Return default HTTP headers required by Qwen Portal API."""
+    import platform as _plat
+
+    _ua = f"QwenCode/{_QWEN_CODE_VERSION} ({_plat.system().lower()}; {_plat.machine()})"
+    return {
+        "User-Agent": _ua,
+        "X-DashScope-CacheControl": "enable",
+        "X-DashScope-UserAgent": _ua,
+        "X-DashScope-AuthType": "qwen-oauth",
+    }
+
+
 class AIAgent:
     """
     AI Agent with tool calling capabilities.
@@ -421,6 +444,13 @@ class AIAgent:
     for AI models that support function calling.
     """
 
+    # ── Class-level context pressure dedup (survives across instances) ──
+    # The gateway creates a new AIAgent per message, so instance-level flags
+    # reset every time.  This dict tracks {session_id: (warn_level, timestamp)}
+    # to suppress duplicate warnings within a cooldown window.
+    _context_pressure_last_warned: dict = {}
+    _CONTEXT_PRESSURE_COOLDOWN = 300  # seconds between re-warning same session
+
     @property
     def base_url(self) -> str:
         return self._base_url
@@ -652,7 +682,8 @@ class AIAgent:
         # Context pressure warnings: notify the USER (not the LLM) as context
         # fills up.  Purely informational — displayed in CLI output and sent via
         # status_callback for gateway platforms.  Does NOT inject into messages.
-        self._context_pressure_warned = False
+        # Tiered: fires at 85% and again at 95% of compaction threshold.
+        self._context_pressure_warned_at = 0.0  # highest tier already shown
 
         # Activity tracking — updated on each API call, tool execution, and
         # stream chunk.  Used by the gateway timeout handler to report what the
@@ -663,6 +694,10 @@ class AIAgent:
         self._current_tool: str | None = None
         self._api_call_count: int = 0
 
+        # Rate limit tracking — updated from x-ratelimit-* response headers
+        # after each API call.  Accessed by /usage slash command.
+        self._rate_limit_state: Optional["RateLimitState"] = None
+
         # Centralized logging — agent.log (INFO+) and errors.log (WARNING+)
         # both live under ~/.hermes/logs/.  Idempotent, so gateway mode
         # (which creates a new AIAgent per message) won't duplicate handlers.
@@ -756,6 +791,8 @@ class AIAgent:
                     client_kwargs["default_headers"] = {
                         "User-Agent": "KimiCLI/1.3",
                     }
+                elif "portal.qwen.ai" in effective_base.lower():
+                    client_kwargs["default_headers"] = _qwen_portal_headers()
             else:
                 # No explicit creds — use the centralized provider router
                 from agent.auxiliary_client import resolve_provider_client
@@ -1664,9 +1701,25 @@ class AIAgent:
         return None
 
     def _cleanup_task_resources(self, task_id: str) -> None:
-        """Clean up VM and browser resources for a given task."""
+        """Clean up VM and browser resources for a given task.
+
+        Skips ``cleanup_vm`` when the active terminal environment is marked
+        persistent (``persistent_filesystem=True``) so that long-lived sandbox
+        containers survive between turns. The idle reaper in
+        ``terminal_tool._cleanup_inactive_envs`` still tears them down once
+        ``terminal.lifetime_seconds`` is exceeded. Non-persistent backends are
+        torn down per-turn as before to prevent resource leakage (the original
+        intent of this hook for the Morph backend, see commit fbd3a2fd).
+        """
         try:
-            cleanup_vm(task_id)
+            if is_persistent_env(task_id):
+                if self.verbose_logging:
+                    logging.debug(
+                        f"Skipping per-turn cleanup_vm for persistent env {task_id}; "
+                        f"idle reaper will handle it."
+                    )
+            else:
+                cleanup_vm(task_id)
         except Exception as e:
             if self.verbose_logging:
                 logging.warning(f"Failed to cleanup VM for task {task_id}: {e}")
@@ -2498,6 +2551,29 @@ class AIAgent:
         self._last_activity_ts = time.time()
         self._last_activity_desc = desc
 
+    def _capture_rate_limits(self, http_response: Any) -> None:
+        """Parse x-ratelimit-* headers from an HTTP response and cache the state.
+
+        Called after each streaming API call.  The httpx Response object is
+        available on the OpenAI SDK Stream via ``stream.response``.
+        """
+        if http_response is None:
+            return
+        headers = getattr(http_response, "headers", None)
+        if not headers:
+            return
+        try:
+            from agent.rate_limit_tracker import parse_rate_limit_headers
+            state = parse_rate_limit_headers(headers, provider=self.provider)
+            if state is not None:
+                self._rate_limit_state = state
+        except Exception:
+            pass  # Never let header parsing break the agent loop
+
+    def get_rate_limit_state(self):
+        """Return the last captured RateLimitState, or None."""
+        return self._rate_limit_state
+
     def get_activity_summary(self) -> dict:
         """Return a snapshot of the agent's current activity for diagnostics.
 
@@ -4080,6 +4156,8 @@ class AIAgent:
             self._client_kwargs["default_headers"] = copilot_default_headers()
         elif "api.kimi.com" in normalized:
             self._client_kwargs["default_headers"] = {"User-Agent": "KimiCLI/1.3"}
+        elif "portal.qwen.ai" in normalized:
+            self._client_kwargs["default_headers"] = _qwen_portal_headers()
         else:
             self._client_kwargs.pop("default_headers", None)
 
@@ -4350,6 +4428,11 @@ class AIAgent:
             self._touch_activity("waiting for provider response (streaming)")
             stream = request_client_holder["client"].chat.completions.create(**stream_kwargs)
 
+            # Capture rate limit headers from the initial HTTP response.
+            # The OpenAI SDK Stream object exposes the underlying httpx
+            # response via .response before any chunks are consumed.
+            self._capture_rate_limits(getattr(stream, "response", None))
+
             content_parts: list = []
             tool_calls_acc: dict = {}
             tool_gen_notified: set = set()
@@ -4703,18 +4786,25 @@ class AIAgent:
                     self._close_request_openai_client(request_client, reason="stream_request_complete")
 
         _stream_stale_timeout_base = float(os.getenv("HERMES_STREAM_STALE_TIMEOUT", 180.0))
-        # Scale the stale timeout for large contexts: slow models (like Opus)
-        # can legitimately think for minutes before producing the first token
-        # when the context is large.  Without this, the stale detector kills
-        # healthy connections during the model's thinking phase, producing
-        # spurious RemoteProtocolError ("peer closed connection").
-        _est_tokens = sum(len(str(v)) for v in api_kwargs.get("messages", [])) // 4
-        if _est_tokens > 100_000:
-            _stream_stale_timeout = max(_stream_stale_timeout_base, 300.0)
-        elif _est_tokens > 50_000:
-            _stream_stale_timeout = max(_stream_stale_timeout_base, 240.0)
+        # Local providers (Ollama, oMLX, llama-cpp) can take 300+ seconds
+        # for prefill on large contexts.  Disable the stale detector unless
+        # the user explicitly set HERMES_STREAM_STALE_TIMEOUT.
+        if _stream_stale_timeout_base == 180.0 and self.base_url and is_local_endpoint(self.base_url):
+            _stream_stale_timeout = float("inf")
+            logger.debug("Local provider detected (%s) — stale stream timeout disabled", self.base_url)
         else:
-            _stream_stale_timeout = _stream_stale_timeout_base
+            # Scale the stale timeout for large contexts: slow models (like Opus)
+            # can legitimately think for minutes before producing the first token
+            # when the context is large.  Without this, the stale detector kills
+            # healthy connections during the model's thinking phase, producing
+            # spurious RemoteProtocolError ("peer closed connection").
+            _est_tokens = sum(len(str(v)) for v in api_kwargs.get("messages", [])) // 4
+            if _est_tokens > 100_000:
+                _stream_stale_timeout = max(_stream_stale_timeout_base, 300.0)
+            elif _est_tokens > 50_000:
+                _stream_stale_timeout = max(_stream_stale_timeout_base, 240.0)
+            else:
+                _stream_stale_timeout = _stream_stale_timeout_base
 
         t = threading.Thread(target=_call, daemon=True)
         t.start()
@@ -4870,7 +4960,7 @@ class AIAgent:
                 effective_key = (fb_client.api_key or resolve_anthropic_token() or "") if fb_provider == "anthropic" else (fb_client.api_key or "")
                 self.api_key = effective_key
                 self._anthropic_api_key = effective_key
-                self._anthropic_base_url = getattr(fb_client, "base_url", None)
+                self._anthropic_base_url = fb_base_url
                 self._anthropic_client = build_anthropic_client(effective_key, self._anthropic_base_url)
                 self._is_anthropic_oauth = _is_oauth_token(effective_key)
                 self.client = None
@@ -4879,9 +4969,21 @@ class AIAgent:
                 # Swap OpenAI client and config in-place
                 self.api_key = fb_client.api_key
                 self.client = fb_client
+                # Preserve provider-specific headers that
+                # resolve_provider_client() may have baked into
+                # fb_client via the default_headers kwarg.  The OpenAI
+                # SDK stores these in _custom_headers.  Without this,
+                # subsequent request-client rebuilds (via
+                # _create_request_openai_client) drop the headers,
+                # causing 403s from providers like Kimi Coding that
+                # require a User-Agent sentinel.
+                fb_headers = getattr(fb_client, "_custom_headers", None)
+                if not fb_headers:
+                    fb_headers = getattr(fb_client, "default_headers", None)
                 self._client_kwargs = {
                     "api_key": fb_client.api_key,
                     "base_url": fb_base_url,
+                    **({"default_headers": dict(fb_headers)} if fb_headers else {}),
                 }
 
             # Re-evaluate prompt caching for the new provider/model
@@ -5226,24 +5328,97 @@ class AIAgent:
         base = (getattr(self, "base_url", "") or "").lower()
         return "dashscope" in base or "aliyuncs" in base or "opencode.ai/zen/go" in base
 
+    def _is_qwen_portal(self) -> bool:
+        """Return True when the base URL targets Qwen Portal."""
+        return "portal.qwen.ai" in self._base_url_lower
+
+    def _qwen_prepare_chat_messages(self, api_messages: list) -> list:
+        prepared = copy.deepcopy(api_messages)
+        if not prepared:
+            return prepared
+
+        for msg in prepared:
+            if not isinstance(msg, dict):
+                continue
+            content = msg.get("content")
+            if isinstance(content, str):
+                msg["content"] = [{"type": "text", "text": content}]
+            elif isinstance(content, list):
+                # Normalize: convert bare strings to text dicts, keep dicts as-is.
+                # deepcopy already created independent copies, no need for dict().
+                normalized_parts = []
+                for part in content:
+                    if isinstance(part, str):
+                        normalized_parts.append({"type": "text", "text": part})
+                    elif isinstance(part, dict):
+                        normalized_parts.append(part)
+                if normalized_parts:
+                    msg["content"] = normalized_parts
+
+        # Inject cache_control on the last part of the system message.
+        for msg in prepared:
+            if isinstance(msg, dict) and msg.get("role") == "system":
+                content = msg.get("content")
+                if isinstance(content, list) and content and isinstance(content[-1], dict):
+                    content[-1]["cache_control"] = {"type": "ephemeral"}
+                break
+
+        return prepared
+
+    def _qwen_prepare_chat_messages_inplace(self, messages: list) -> None:
+        """In-place variant — mutates an already-copied message list."""
+        if not messages:
+            return
+
+        for msg in messages:
+            if not isinstance(msg, dict):
+                continue
+            content = msg.get("content")
+            if isinstance(content, str):
+                msg["content"] = [{"type": "text", "text": content}]
+            elif isinstance(content, list):
+                normalized_parts = []
+                for part in content:
+                    if isinstance(part, str):
+                        normalized_parts.append({"type": "text", "text": part})
+                    elif isinstance(part, dict):
+                        normalized_parts.append(part)
+                if normalized_parts:
+                    msg["content"] = normalized_parts
+
+        for msg in messages:
+            if isinstance(msg, dict) and msg.get("role") == "system":
+                content = msg.get("content")
+                if isinstance(content, list) and content and isinstance(content[-1], dict):
+                    content[-1]["cache_control"] = {"type": "ephemeral"}
+                break
+
     def _build_api_kwargs(self, api_messages: list) -> dict:
         """Build the keyword arguments dict for the active API mode."""
         if self.api_mode == "anthropic_messages":
             from agent.anthropic_adapter import build_anthropic_kwargs
             anthropic_messages = self._prepare_anthropic_messages_for_api(api_messages)
-            # Pass context_length so the adapter can clamp max_tokens if the
-            # user configured a smaller context window than the model's output limit.
+            # Pass context_length (total input+output window) so the adapter can
+            # clamp max_tokens (output cap) when the user configured a smaller
+            # context window than the model's native output limit.
             ctx_len = getattr(self, "context_compressor", None)
             ctx_len = ctx_len.context_length if ctx_len else None
+            # _ephemeral_max_output_tokens is set for one call when the API
+            # returns "max_tokens too large given prompt" — it caps output to
+            # the available window space without touching context_length.
+            ephemeral_out = getattr(self, "_ephemeral_max_output_tokens", None)
+            if ephemeral_out is not None:
+                self._ephemeral_max_output_tokens = None  # consume immediately
             return build_anthropic_kwargs(
                 model=self.model,
                 messages=anthropic_messages,
                 tools=self.tools,
-                max_tokens=self.max_tokens,
+                max_tokens=ephemeral_out if ephemeral_out is not None else self.max_tokens,
                 reasoning_config=self.reasoning_config,
                 is_oauth=self._is_anthropic_oauth,
                 preserve_dots=self._anthropic_preserve_dots(),
                 context_length=ctx_len,
+                base_url=getattr(self, "_anthropic_base_url", None),
             )
 
         if self.api_mode == "codex_responses":
@@ -5337,6 +5512,17 @@ class AIAgent:
                             tool_call.pop("call_id", None)
                             tool_call.pop("response_item_id", None)
 
+        # Qwen portal: normalize content to list-of-dicts, inject cache_control.
+        # Must run AFTER codex sanitization so we transform the final messages.
+        # If sanitization already deepcopied, reuse that copy (in-place).
+        if self._is_qwen_portal():
+            if sanitized_messages is api_messages:
+                # No sanitization was done — we need our own copy.
+                sanitized_messages = self._qwen_prepare_chat_messages(sanitized_messages)
+            else:
+                # Already a deepcopy — transform in place to avoid a second deepcopy.
+                self._qwen_prepare_chat_messages_inplace(sanitized_messages)
+
         # GPT-5 and Codex models respond better to 'developer' than 'system'
         # for instruction-following.  Swap the role at the API boundary so
         # internal message representation stays uniform ("system").
@@ -5369,11 +5555,17 @@ class AIAgent:
             "messages": sanitized_messages,
             "timeout": float(os.getenv("HERMES_API_TIMEOUT", 1800.0)),
         }
+        if self._is_qwen_portal():
+            api_kwargs["metadata"] = {
+                "sessionId": self.session_id or "hermes",
+                "promptId": str(uuid.uuid4()),
+            }
         if self.tools:
             api_kwargs["tools"] = self.tools
 
         if self.max_tokens is not None:
-            api_kwargs.update(self._max_tokens_param(self.max_tokens))
+            if not self._is_qwen_portal():
+                api_kwargs.update(self._max_tokens_param(self.max_tokens))
         elif self._is_openrouter_url() and "claude" in (self.model or "").lower():
             # OpenRouter translates requests to Anthropic's Messages API,
             # which requires max_tokens as a mandatory field.  When we omit
@@ -5438,6 +5630,9 @@ class AIAgent:
             options["num_ctx"] = self._ollama_num_ctx
             extra_body["options"] = options
 
+        if self._is_qwen_portal():
+            extra_body["vl_high_resolution_images"] = True
+
         if extra_body:
             api_kwargs["extra_body"] = extra_body
 
@@ -5753,7 +5948,7 @@ class AIAgent:
                     tools=[memory_tool_def],
                     temperature=0.3,
                     max_tokens=5120,
-                    timeout=30.0,
+                    # timeout resolved from auxiliary.flush_memories.timeout config
                 )
             except RuntimeError:
                 _aux_available = False
@@ -5785,7 +5980,10 @@ class AIAgent:
                     "temperature": 0.3,
                     **self._max_tokens_param(5120),
                 }
-                response = self._ensure_primary_openai_client(reason="flush_memories").chat.completions.create(**api_kwargs, timeout=30.0)
+                from agent.auxiliary_client import _get_task_timeout
+                response = self._ensure_primary_openai_client(reason="flush_memories").chat.completions.create(
+                    **api_kwargs, timeout=_get_task_timeout("flush_memories")
+                )
 
             # Extract tool calls from the response, handling all API formats
             tool_calls = []
@@ -5892,6 +6090,15 @@ class AIAgent:
             except Exception as e:
                 logger.warning("Session DB compression split failed — new session will NOT be indexed: %s", e)
 
+        # Warn on repeated compressions (quality degrades with each pass)
+        _cc = self.context_compressor.compression_count
+        if _cc >= 2:
+            self._vprint(
+                f"{self.log_prefix}⚠️  Session compressed {_cc} times — "
+                f"accuracy may degrade. Consider /new to start fresh.",
+                force=True,
+            )
+
         # Update token estimate after compaction so pressure calculations
         # use the post-compression count, not the stale pre-compression one.
         _compressed_est = (
@@ -5904,12 +6111,16 @@ class AIAgent:
         # Only reset the pressure warning if compression actually brought
         # us below the warning level (85% of threshold).  When compression
         # can't reduce enough (e.g. threshold is very low, or system prompt
-        # alone exceeds the warning level), keep the flag set to prevent
+        # alone exceeds the warning level), keep the tier set to prevent
         # spamming the user with repeated warnings every loop iteration.
         if self.context_compressor.threshold_tokens > 0:
             _post_progress = _compressed_est / self.context_compressor.threshold_tokens
             if _post_progress < 0.85:
-                self._context_pressure_warned = False
+                self._context_pressure_warned_at = 0.0
+                # Clear class-level dedup for this session so a fresh
+                # warning cycle can start if context grows again.
+                _sid = self.session_id or "default"
+                AIAgent._context_pressure_last_warned.pop(_sid, None)
 
         # Clear the file-read dedup cache.  After compression the original
         # read content is summarised away — if the model re-reads the same
@@ -7091,6 +7302,7 @@ class AIAgent:
         length_continue_retries = 0
         truncated_response_prefix = ""
         compression_attempts = 0
+        _turn_exit_reason = "unknown"  # Diagnostic: why the loop ended
         
         # Clear any stale interrupt state at start
         self.clear_interrupt()
@@ -7115,6 +7327,7 @@ class AIAgent:
             # Check for interrupt request (e.g., user sent new message)
             if self._interrupt_requested:
                 interrupted = True
+                _turn_exit_reason = "interrupted_by_user"
                 if not self.quiet_mode:
                     self._safe_print("\n⚡ Breaking out of tool loop due to interrupt...")
                 break
@@ -7123,6 +7336,7 @@ class AIAgent:
             self._api_call_count = api_call_count
             self._touch_activity(f"starting API call #{api_call_count}")
             if not self.iteration_budget.consume():
+                _turn_exit_reason = "budget_exhausted"
                 if not self.quiet_mode:
                     self._safe_print(f"\n⚠️  Iteration budget exhausted ({self.iteration_budget.used}/{self.iteration_budget.max_total} iterations used)")
                 break
@@ -7827,6 +8041,25 @@ class AIAgent:
 
                     status_code = getattr(api_error, "status_code", None)
                     error_context = self._extract_api_error_context(api_error)
+
+                    # ── Classify the error for structured recovery decisions ──
+                    _compressor = getattr(self, "context_compressor", None)
+                    _ctx_len = getattr(_compressor, "context_length", 200000) if _compressor else 200000
+                    classified = classify_api_error(
+                        api_error,
+                        provider=getattr(self, "provider", "") or "",
+                        model=getattr(self, "model", "") or "",
+                        approx_tokens=approx_tokens,
+                        context_length=_ctx_len,
+                        num_messages=len(api_messages) if api_messages else 0,
+                    )
+                    logger.debug(
+                        "Error classified: reason=%s status=%s retryable=%s compress=%s rotate=%s fallback=%s",
+                        classified.reason.value, classified.status_code,
+                        classified.retryable, classified.should_compress,
+                        classified.should_rotate_credential, classified.should_fallback,
+                    )
+
                     recovered_with_pool, has_retried_429 = self._recover_with_credential_pool(
                         status_code=status_code,
                         has_retried_429=has_retried_429,
@@ -7889,27 +8122,24 @@ class AIAgent:
                     # from all messages so the next retry sends no thinking
                     # blocks at all.  One-shot — don't retry infinitely.
                     if (
-                        self.api_mode == "anthropic_messages"
-                        and status_code == 400
+                        classified.reason == FailoverReason.thinking_signature
                         and not thinking_sig_retry_attempted
                     ):
-                        _err_msg_lower = str(api_error).lower()
-                        if "signature" in _err_msg_lower and "thinking" in _err_msg_lower:
-                            thinking_sig_retry_attempted = True
-                            for _m in messages:
-                                if isinstance(_m, dict):
-                                    _m.pop("reasoning_details", None)
-                            self._vprint(
-                                f"{self.log_prefix}⚠️  Thinking block signature invalid — "
-                                f"stripped all thinking blocks, retrying...",
-                                force=True,
-                            )
-                            logging.warning(
-                                "%sThinking block signature recovery: stripped "
-                                "reasoning_details from %d messages",
-                                self.log_prefix, len(messages),
-                            )
-                            continue
+                        thinking_sig_retry_attempted = True
+                        for _m in messages:
+                            if isinstance(_m, dict):
+                                _m.pop("reasoning_details", None)
+                        self._vprint(
+                            f"{self.log_prefix}⚠️  Thinking block signature invalid — "
+                            f"stripped all thinking blocks, retrying...",
+                            force=True,
+                        )
+                        logging.warning(
+                            "%sThinking block signature recovery: stripped "
+                            "reasoning_details from %d messages",
+                            self.log_prefix, len(messages),
+                        )
+                        continue
 
                     retry_count += 1
                     elapsed_time = time.time() - api_start_time
@@ -7966,14 +8196,7 @@ class AIAgent:
                     # is NOT a transient rate limit — retrying or switching
                     # credentials won't help.  Reduce context to 200k (the
                     # standard tier) and compress.
-                    # Only applies to Sonnet — Opus 1M is general access.
-                    _is_long_context_tier_error = (
-                        status_code == 429
-                        and "extra usage" in error_msg
-                        and "long context" in error_msg
-                        and "sonnet" in self.model.lower()
-                    )
-                    if _is_long_context_tier_error:
+                    if classified.reason == FailoverReason.long_context_tier:
                         _reduced_ctx = 200000
                         compressor = self.context_compressor
                         old_ctx = compressor.context_length
@@ -8018,13 +8241,9 @@ class AIAgent:
                     # When a fallback model is configured, switch immediately instead
                     # of burning through retries with exponential backoff -- the
                     # primary provider won't recover within the retry window.
-                    is_rate_limited = (
-                        status_code == 429
-                        or "rate limit" in error_msg
-                        or "too many requests" in error_msg
-                        or "rate_limit" in error_msg
-                        or "usage limit" in error_msg
-                        or "quota" in error_msg
+                    is_rate_limited = classified.reason in (
+                        FailoverReason.rate_limit,
+                        FailoverReason.billing,
                     )
                     if is_rate_limited and self._fallback_index < len(self._fallback_chain):
                         # Don't eagerly fallback if credential pool rotation may
@@ -8040,10 +8259,7 @@ class AIAgent:
                                 continue
 
                     is_payload_too_large = (
-                        status_code == 413
-                        or 'request entity too large' in error_msg
-                        or 'payload too large' in error_msg
-                        or 'error code: 413' in error_msg
+                        classified.reason == FailoverReason.payload_too_large
                     )
 
                     if is_payload_too_large:
@@ -8087,69 +8303,59 @@ class AIAgent:
                             }
 
                     # Check for context-length errors BEFORE generic 4xx handler.
-                    # Local backends (LM Studio, Ollama, llama.cpp) often return
-                    # HTTP 400 with messages like "Context size has been exceeded"
-                    # which must trigger compression, not an immediate abort.
-                    is_context_length_error = any(phrase in error_msg for phrase in [
-                        'context length', 'context size', 'maximum context',
-                        'token limit', 'too many tokens', 'reduce the length',
-                        'exceeds the limit', 'context window',
-                        'request entity too large',  # OpenRouter/Nous 413 safety net
-                        'prompt is too long',  # Anthropic: "prompt is too long: N tokens > M maximum"
-                        'prompt exceeds max length',  # Z.AI / GLM: generic 400 overflow wording
-                    ])
-
-                    # Fallback heuristic: Anthropic sometimes returns a generic
-                    # 400 invalid_request_error with just "Error" as the message
-                    # when the context is too large.  If the error message is very
-                    # short/generic AND the session is large, treat it as a
-                    # probable context-length error and attempt compression rather
-                    # than aborting.  This prevents an infinite failure loop where
-                    # each failed message gets persisted, making the session even
-                    # larger. (#1630)
-                    if not is_context_length_error and status_code == 400:
-                        ctx_len = getattr(getattr(self, 'context_compressor', None), 'context_length', 200000)
-                        is_large_session = approx_tokens > ctx_len * 0.4 or len(api_messages) > 80
-                        is_generic_error = len(error_msg.strip()) < 30  # e.g. just "error"
-                        if is_large_session and is_generic_error:
-                            is_context_length_error = True
-                            self._vprint(
-                                f"{self.log_prefix}⚠️  Generic 400 with large session "
-                                f"(~{approx_tokens:,} tokens, {len(api_messages)} msgs) — "
-                                f"treating as probable context overflow.",
-                                force=True,
-                            )
-
-                    # Server disconnects on large sessions are often caused by
-                    # the request exceeding the provider's context/payload limit
-                    # without a proper HTTP error response.  Treat these as
-                    # context-length errors to trigger compression rather than
-                    # burning through retries that will all fail the same way.
-                    # This breaks the death spiral: disconnect → no token data
-                    # → no compression → bigger session → more disconnects.
-                    # (#2153)
-                    if not is_context_length_error and not status_code:
-                        _is_server_disconnect = (
-                            'server disconnected' in error_msg
-                            or 'peer closed connection' in error_msg
-                            or error_type in ('ReadError', 'RemoteProtocolError', 'ServerDisconnectedError')
-                        )
-                        if _is_server_disconnect:
-                            ctx_len = getattr(getattr(self, 'context_compressor', None), 'context_length', 200000)
-                            _is_large = approx_tokens > ctx_len * 0.6 or len(api_messages) > 200
-                            if _is_large:
-                                is_context_length_error = True
-                                self._vprint(
-                                    f"{self.log_prefix}⚠️  Server disconnected with large session "
-                                    f"(~{approx_tokens:,} tokens, {len(api_messages)} msgs) — "
-                                    f"treating as context-length error, attempting compression.",
-                                    force=True,
-                                )
+                    # The classifier detects context overflow from: explicit error
+                    # messages, generic 400 + large session heuristic (#1630), and
+                    # server disconnect + large session pattern (#2153).
+                    is_context_length_error = (
+                        classified.reason == FailoverReason.context_overflow
+                    )
 
                     if is_context_length_error:
                         compressor = self.context_compressor
                         old_ctx = compressor.context_length
 
+                        # ── Distinguish two very different errors ───────────
+                        # 1. "Prompt too long": the INPUT exceeds the context window.
+                        #    Fix: reduce context_length + compress history.
+                        # 2. "max_tokens too large": input is fine, but
+                        #    input_tokens + requested max_tokens > context_window.
+                        #    Fix: reduce max_tokens (the OUTPUT cap) for this call.
+                        #    Do NOT shrink context_length — the window is unchanged.
+                        #
+                        # Note: max_tokens = output token cap (one response).
+                        #       context_length = total window (input + output combined).
+                        available_out = parse_available_output_tokens_from_error(error_msg)
+                        if available_out is not None:
+                            # Error is purely about the output cap being too large.
+                            # Cap output to the available space and retry without
+                            # touching context_length or triggering compression.
+                            safe_out = max(1, available_out - 64)  # small safety margin
+                            self._ephemeral_max_output_tokens = safe_out
+                            self._vprint(
+                                f"{self.log_prefix}⚠️  Output cap too large for current prompt — "
+                                f"retrying with max_tokens={safe_out:,} "
+                                f"(available_tokens={available_out:,}; context_length unchanged at {old_ctx:,})",
+                                force=True,
+                            )
+                            # Still count against compression_attempts so we don't
+                            # loop forever if the error keeps recurring.
+                            compression_attempts += 1
+                            if compression_attempts > max_compression_attempts:
+                                self._vprint(f"{self.log_prefix}❌ Max compression attempts ({max_compression_attempts}) reached.", force=True)
+                                self._vprint(f"{self.log_prefix}   💡 Try /new to start a fresh conversation, or /compress to retry compression.", force=True)
+                                logging.error(f"{self.log_prefix}Context compression failed after {max_compression_attempts} attempts.")
+                                self._persist_session(messages, conversation_history)
+                                return {
+                                    "messages": messages,
+                                    "completed": False,
+                                    "api_calls": api_call_count,
+                                    "error": f"Context length exceeded: max compression attempts ({max_compression_attempts}) reached.",
+                                    "partial": True
+                                }
+                            restart_with_compressed_messages = True
+                            break
+
+                        # Error is about the INPUT being too large — reduce context_length.
                         # Try to parse the actual limit from the error message
                         parsed_limit = parse_context_limit_from_error(error_msg)
                         if parsed_limit and parsed_limit < old_ctx:
@@ -8216,35 +8422,30 @@ class AIAgent:
                                 "partial": True
                             }
 
-                    # Check for non-retryable client errors (4xx HTTP status codes).
-                    # These indicate a problem with the request itself (bad model ID,
-                    # invalid API key, forbidden, etc.) and will never succeed on retry.
-                    # Note: 413 and context-length errors are excluded — handled above.
-                    # 429 (rate limit) is transient and MUST be retried with backoff.
-                    # 529 (Anthropic overloaded) is also transient.
-                    # Also catch local validation errors (ValueError, TypeError) — these
-                    # are programming bugs, not transient failures.
-                    # Exclude UnicodeEncodeError — it's a ValueError subclass but is
-                    # handled separately by the surrogate sanitization path above.
-                    _RETRYABLE_STATUS_CODES = {413, 429, 529}
+                    # Check for non-retryable client errors.  The classifier
+                    # already accounts for 413, 429, 529 (transient), context
+                    # overflow, and generic-400 heuristics.  Local validation
+                    # errors (ValueError, TypeError) are programming bugs.
                     is_local_validation_error = (
                         isinstance(api_error, (ValueError, TypeError))
                         and not isinstance(api_error, UnicodeEncodeError)
                     )
-                    # Detect generic 400s from Anthropic OAuth (transient server-side failures).
-                    # Real invalid_request_error responses include a descriptive message;
-                    # transient ones contain only "Error" or are empty. (ref: issue #1608)
-                    _err_body = getattr(api_error, "body", None) or {}
-                    _err_message = (_err_body.get("error", {}).get("message", "") if isinstance(_err_body, dict) else "")
-                    _is_generic_400 = (status_code == 400 and _err_message.strip().lower() in ("error", ""))
-                    is_client_status_error = isinstance(status_code, int) and 400 <= status_code < 500 and status_code not in _RETRYABLE_STATUS_CODES and not _is_generic_400
-                    is_client_error = (is_local_validation_error or is_client_status_error or any(phrase in error_msg for phrase in [
-                        'error code: 401', 'error code: 403',
-                        'error code: 404', 'error code: 422',
-                        'is not a valid model', 'invalid model', 'model not found',
-                        'invalid api key', 'invalid_api_key', 'authentication',
-                        'unauthorized', 'forbidden', 'not found',
-                    ])) and not is_context_length_error
+                    is_client_error = (
+                        is_local_validation_error
+                        or (
+                            not classified.retryable
+                            and not classified.should_compress
+                            and classified.reason not in (
+                                FailoverReason.rate_limit,
+                                FailoverReason.billing,
+                                FailoverReason.overloaded,
+                                FailoverReason.context_overflow,
+                                FailoverReason.payload_too_large,
+                                FailoverReason.long_context_tier,
+                                FailoverReason.thinking_signature,
+                            )
+                        )
+                    ) and not is_context_length_error
 
                     if is_client_error:
                         # Try fallback before aborting — a different provider
@@ -8264,7 +8465,7 @@ class AIAgent:
                         self._vprint(f"{self.log_prefix}   🔌 Provider: {_provider}  Model: {_model}", force=True)
                         self._vprint(f"{self.log_prefix}   🌐 Endpoint: {_base}", force=True)
                         # Actionable guidance for common auth errors
-                        if status_code in (401, 403) or "unauthorized" in error_msg or "forbidden" in error_msg or "permission" in error_msg:
+                        if classified.is_auth or classified.reason == FailoverReason.billing:
                             if _provider == "openai-codex" and status_code == 401:
                                 self._vprint(f"{self.log_prefix}   💡 Codex OAuth token was rejected (HTTP 401). Your token may have been", force=True)
                                 self._vprint(f"{self.log_prefix}      refreshed by another client (Codex CLI, VS Code). To fix:", force=True)
@@ -8424,6 +8625,7 @@ class AIAgent:
             
             # If the API call was interrupted, skip response processing
             if interrupted:
+                _turn_exit_reason = "interrupted_during_api_call"
                 break
 
             if restart_with_compressed_messages:
@@ -8443,6 +8645,7 @@ class AIAgent:
             # (e.g. repeated context-length errors that exhausted retry_count),
             # the `response` variable is still None. Break out cleanly.
             if response is None:
+                _turn_exit_reason = "all_retries_exhausted_no_response"
                 print(f"{self.log_prefix}❌ All API retries exhausted with no successful response.")
                 self._persist_session(messages, conversation_history)
                 break
@@ -8849,13 +9052,34 @@ class AIAgent:
                     # compaction fires, not the raw context window.
                     # Does not inject into messages — just prints to CLI output
                     # and fires status_callback for gateway platforms.
+                    # Tiered: 85% (orange) and 95% (red/critical).
                     if _compressor.threshold_tokens > 0:
                         _compaction_progress = _real_tokens / _compressor.threshold_tokens
-                        if _compaction_progress >= 0.85 and not self._context_pressure_warned:
-                            self._context_pressure_warned = True
-                            self._emit_context_pressure(_compaction_progress, _compressor)
+                        # Determine the warning tier for this progress level
+                        _warn_tier = 0.0
+                        if _compaction_progress >= 0.95:
+                            _warn_tier = 0.95
+                        elif _compaction_progress >= 0.85:
+                            _warn_tier = 0.85
+                        if _warn_tier > self._context_pressure_warned_at:
+                            # Class-level dedup: check if this session was already
+                            # warned at this tier within the cooldown window.
+                            _sid = self.session_id or "default"
+                            _last = AIAgent._context_pressure_last_warned.get(_sid)
+                            _now = time.time()
+                            if _last is None or _last[0] < _warn_tier or (_now - _last[1]) >= self._CONTEXT_PRESSURE_COOLDOWN:
+                                self._context_pressure_warned_at = _warn_tier
+                                AIAgent._context_pressure_last_warned[_sid] = (_warn_tier, _now)
+                                self._emit_context_pressure(_compaction_progress, _compressor)
+                                # Evict stale entries (older than 2x cooldown)
+                                _cutoff = _now - self._CONTEXT_PRESSURE_COOLDOWN * 2
+                                AIAgent._context_pressure_last_warned = {
+                                    k: v for k, v in AIAgent._context_pressure_last_warned.items()
+                                    if v[1] > _cutoff
+                                }
 
                     if self.compression_enabled and _compressor.should_compress(_real_tokens):
+                        self._safe_print("  ⟳ compacting context…")
                         messages, active_system_prompt = self._compress_context(
                             messages, system_message,
                             approx_tokens=self.context_compressor.last_prompt_tokens,
@@ -8885,6 +9109,7 @@ class AIAgent:
                         # instead of wasting API calls on retries that won't help.
                         fallback = getattr(self, '_last_content_with_tools', None)
                         if fallback:
+                            _turn_exit_reason = "fallback_prior_turn_content"
                             logger.debug("Empty follow-up after tool calls — using prior turn content as final response")
                             self._last_content_with_tools = None
                             self._empty_content_retries = 0
@@ -8930,8 +9155,28 @@ class AIAgent:
                             self._save_session_log(messages)
                             continue
 
-                        # Exhausted prefill attempts or no structured
-                        # reasoning — fall through to "(empty)" terminal.
+                        # ── Empty response retry (no reasoning) ──────
+                        # Model returned nothing — no content, no
+                        # structured reasoning, no tool calls.  Common
+                        # with open models (transient provider issues,
+                        # rate limits, sampling flukes).  Silently retry
+                        # up to 3 times before giving up.  Skip when
+                        # content has inline <think> tags (model chose
+                        # to reason, just no visible text).
+                        _truly_empty = not final_response.strip()
+                        if _truly_empty and not _has_structured and self._empty_content_retries < 3:
+                            self._empty_content_retries += 1
+                            self._vprint(
+                                f"{self.log_prefix}↻ Empty response (no content or reasoning) "
+                                f"— retrying ({self._empty_content_retries}/3)",
+                                force=True,
+                            )
+                            continue
+
+                        # Exhausted prefill attempts, empty retries, or
+                        # structured reasoning with no content —
+                        # fall through to "(empty)" terminal.
+                        _turn_exit_reason = "empty_response_exhausted"
                         reasoning_text = self._extract_reasoning(assistant_message)
                         assistant_msg = self._build_assistant_message(assistant_message, finish_reason)
                         assistant_msg["content"] = "(empty)"
@@ -8941,7 +9186,7 @@ class AIAgent:
                             reasoning_preview = reasoning_text[:500] + "..." if len(reasoning_text) > 500 else reasoning_text
                             self._vprint(f"{self.log_prefix}ℹ️  Reasoning-only response (no visible content). Reasoning: {reasoning_preview}")
                         else:
-                            self._vprint(f"{self.log_prefix}ℹ️  Empty response (no content or reasoning).")
+                            self._vprint(f"{self.log_prefix}ℹ️  Empty response (no content or reasoning) after 3 retries.")
 
                         final_response = "(empty)"
                         break
@@ -9003,6 +9248,7 @@ class AIAgent:
 
                     messages.append(final_msg)
                     
+                    _turn_exit_reason = f"text_response(finish_reason={finish_reason})"
                     if not self.quiet_mode:
                         self._safe_print(f"🎉 Conversation completed after {api_call_count} OpenAI-compatible API call(s)")
                     break
@@ -9052,6 +9298,7 @@ class AIAgent:
 
                 # If we're near the limit, break to avoid infinite loops
                 if api_call_count >= self.max_iterations - 1:
+                    _turn_exit_reason = f"error_near_max_iterations({error_msg[:80]})"
                     final_response = f"I apologize, but I encountered repeated errors: {error_msg}"
                     # Append as assistant so the history stays valid for
                     # session resume (avoids consecutive user messages).
@@ -9062,6 +9309,7 @@ class AIAgent:
             api_call_count >= self.max_iterations
             or self.iteration_budget.remaining <= 0
         ):
+            _turn_exit_reason = f"max_iterations_reached({api_call_count}/{self.max_iterations})"
             if self.iteration_budget.remaining <= 0 and not self.quiet_mode:
                 print(f"\n⚠️  Iteration budget exhausted ({self.iteration_budget.used}/{self.iteration_budget.max_total} iterations used)")
             final_response = self._handle_max_iterations(messages, api_call_count)
@@ -9078,6 +9326,49 @@ class AIAgent:
         # Persist session to both JSON log and SQLite
         self._persist_session(messages, conversation_history)
 
+        # ── Turn-exit diagnostic log ─────────────────────────────────────
+        # Always logged at INFO so agent.log captures WHY every turn ended.
+        # When the last message is a tool result (agent was mid-work), log
+        # at WARNING — this is the "just stops" scenario users report.
+        _last_msg_role = messages[-1].get("role") if messages else None
+        _last_tool_name = None
+        if _last_msg_role == "tool":
+            # Walk back to find the assistant message with the tool call
+            for _m in reversed(messages):
+                if _m.get("role") == "assistant" and _m.get("tool_calls"):
+                    _tcs = _m["tool_calls"]
+                    if _tcs and isinstance(_tcs[0], dict):
+                        _last_tool_name = _tcs[-1].get("function", {}).get("name")
+                    break
+
+        _turn_tool_count = sum(
+            1 for m in messages
+            if isinstance(m, dict) and m.get("role") == "assistant" and m.get("tool_calls")
+        )
+        _resp_len = len(final_response) if final_response else 0
+        _budget_used = self.iteration_budget.used if self.iteration_budget else 0
+        _budget_max = self.iteration_budget.max_total if self.iteration_budget else 0
+
+        _diag_msg = (
+            "Turn ended: reason=%s model=%s api_calls=%d/%d budget=%d/%d "
+            "tool_turns=%d last_msg_role=%s response_len=%d session=%s"
+        )
+        _diag_args = (
+            _turn_exit_reason, self.model, api_call_count, self.max_iterations,
+            _budget_used, _budget_max,
+            _turn_tool_count, _last_msg_role, _resp_len,
+            self.session_id or "none",
+        )
+
+        if _last_msg_role == "tool" and not interrupted:
+            # Agent was mid-work — this is the "just stops" case.
+            logger.warning(
+                "Turn ended with pending tool result (agent may appear stuck). "
+                + _diag_msg + " last_tool=%s",
+                *_diag_args, _last_tool_name,
+            )
+        else:
+            logger.info(_diag_msg, *_diag_args)
 
         # Plugin hook: post_llm_call
         # Fired once per turn after the tool-calling loop completes.

skills/autonomous-ai-agents/hermes-agent/SKILL.md

@@ -249,7 +249,6 @@ Type these during an interactive chat session.
 /config              Show config (CLI)
 /model [name]        Show or change model
 /provider            Show provider info
-/prompt [text]       View/set system prompt (CLI)
 /personality [name]  Set personality
 /reasoning [level]   Set reasoning (none|low|medium|high|xhigh|show|hide)
 /verbose             Cycle: off → new → all → verbose

tests/agent/test_auxiliary_client.py

@@ -77,6 +77,20 @@ class TestReadCodexAccessToken:
         result = _read_codex_access_token()
         assert result == "tok-123"
 
+    def test_pool_without_selected_entry_falls_back_to_auth_store(self, tmp_path, monkeypatch):
+        hermes_home = tmp_path / "hermes"
+        hermes_home.mkdir(parents=True, exist_ok=True)
+        monkeypatch.setenv("HERMES_HOME", str(hermes_home))
+
+        valid_jwt = "eyJhbGciOiJSUzI1NiJ9.eyJleHAiOjk5OTk5OTk5OTl9.sig"
+        with patch("agent.auxiliary_client._select_pool_entry", return_value=(True, None)), \
+             patch("hermes_cli.auth._read_codex_tokens", return_value={
+                 "tokens": {"access_token": valid_jwt, "refresh_token": "refresh"}
+             }):
+            result = _read_codex_access_token()
+
+        assert result == valid_jwt
+
     def test_missing_returns_none(self, tmp_path, monkeypatch):
         hermes_home = tmp_path / "hermes"
         hermes_home.mkdir(parents=True, exist_ok=True)
@@ -238,6 +252,24 @@ class TestAnthropicOAuthFlag:
         assert mock_build.call_args.args[0] == "sk-ant-oat01-pooled"
 
 
+class TestTryCodex:
+    def test_pool_without_selected_entry_falls_back_to_auth_store(self):
+        with (
+            patch("agent.auxiliary_client._select_pool_entry", return_value=(True, None)),
+            patch("agent.auxiliary_client._read_codex_access_token", return_value="codex-auth-token"),
+            patch("agent.auxiliary_client.OpenAI") as mock_openai,
+        ):
+            mock_openai.return_value = MagicMock()
+            from agent.auxiliary_client import _try_codex
+
+            client, model = _try_codex()
+
+        assert client is not None
+        assert model == "gpt-5.2-codex"
+        assert mock_openai.call_args.kwargs["api_key"] == "codex-auth-token"
+        assert mock_openai.call_args.kwargs["base_url"] == "https://chatgpt.com/backend-api/codex"
+
+
 class TestExpiredCodexFallback:
     """Test that expired Codex tokens don't block the auto chain."""
 
@@ -737,8 +769,8 @@ class TestAuxiliaryPoolAwareness:
         assert client is not None
         assert client.__class__.__name__ == "AnthropicAuxiliaryClient"
 
-    def test_vision_auto_prefers_openrouter_over_active_provider(self, monkeypatch):
-        """OpenRouter is tried before the active provider in vision auto."""
+    def test_vision_auto_prefers_active_provider_over_openrouter(self, monkeypatch):
+        """Active provider is tried before OpenRouter in vision auto."""
         monkeypatch.setenv("OPENROUTER_API_KEY", "or-key")
         monkeypatch.setenv("ANTHROPIC_API_KEY", "***")
 
@@ -746,12 +778,13 @@ class TestAuxiliaryPoolAwareness:
             patch("agent.auxiliary_client._read_nous_auth", return_value=None),
             patch("agent.auxiliary_client._read_main_provider", return_value="anthropic"),
             patch("agent.auxiliary_client._read_main_model", return_value="claude-sonnet-4"),
-            patch("agent.auxiliary_client.OpenAI") as mock_openai,
+            patch("agent.anthropic_adapter.build_anthropic_client", return_value=MagicMock()),
+            patch("agent.anthropic_adapter.resolve_anthropic_token", return_value="***"),
         ):
             provider, client, model = resolve_vision_provider_client()
 
-        # OpenRouter should win over anthropic active provider
-        assert provider == "openrouter"
+        # Active provider should win over OpenRouter
+        assert provider == "anthropic"
 
     def test_vision_auto_uses_named_custom_as_active_provider(self, monkeypatch):
         """Named custom provider works as active provider fallback in vision auto."""

tests/agent/test_context_compressor.py

@@ -324,7 +324,10 @@ class TestCompressWithClient:
         with patch("agent.context_compressor.get_model_context_length", return_value=100000):
             c = ContextCompressor(model="test", quiet_mode=True, protect_first_n=2, protect_last_n=2)
 
-        # Last head message (index 1) is "assistant" → summary should be "user"
+        # Last head message (index 1) is "assistant" → summary should be "user".
+        # With min_tail=3, tail = last 3 messages (indices 5-7).
+        # head_last=assistant, tail_first=assistant → summary_role="user", no collision.
+        # Need 8 messages: min_for_compress = 2+3+1 = 6, must have > 6.
         msgs = [
             {"role": "user", "content": "msg 0"},
             {"role": "assistant", "content": "msg 1"},
@@ -332,6 +335,8 @@ class TestCompressWithClient:
             {"role": "assistant", "content": "msg 3"},
             {"role": "user", "content": "msg 4"},
             {"role": "assistant", "content": "msg 5"},
+            {"role": "user", "content": "msg 6"},
+            {"role": "assistant", "content": "msg 7"},
         ]
         with patch("agent.context_compressor.call_llm", return_value=mock_response):
             result = c.compress(msgs)
@@ -460,8 +465,10 @@ class TestCompressWithClient:
             c = ContextCompressor(model="test", quiet_mode=True, protect_first_n=2, protect_last_n=2)
 
         # Head: [system, user]        → last head = user
-        # Tail: [assistant, user]     → first tail = assistant
+        # Tail: [assistant, user, assistant] → first tail = assistant
         # summary_role="assistant" collides with tail, "user" collides with head → merge
+        # With min_tail=3, tail = last 3 messages (indices 5-7).
+        # Need 8 messages: min_for_compress = 2+3+1 = 6, must have > 6.
         msgs = [
             {"role": "system", "content": "system prompt"},
             {"role": "user", "content": "msg 1"},
@@ -470,6 +477,7 @@ class TestCompressWithClient:
             {"role": "assistant", "content": "msg 4"},   # compressed
             {"role": "assistant", "content": "msg 5"},   # tail start
             {"role": "user", "content": "msg 6"},
+            {"role": "assistant", "content": "msg 7"},
         ]
         with patch("agent.context_compressor.call_llm", return_value=mock_response):
             result = c.compress(msgs)
@@ -481,7 +489,7 @@ class TestCompressWithClient:
             if r1 in ("user", "assistant") and r2 in ("user", "assistant"):
                 assert r1 != r2, f"consecutive {r1} at indices {i-1},{i}"
 
-        # The summary should be merged into the first tail message (assistant)
+        # The summary should be merged into the first tail message (assistant at index 5)
         first_tail = [m for m in result if "msg 5" in (m.get("content") or "")]
         assert len(first_tail) == 1
         assert "summary text" in first_tail[0]["content"]
@@ -496,14 +504,18 @@ class TestCompressWithClient:
         with patch("agent.context_compressor.get_model_context_length", return_value=100000):
             c = ContextCompressor(model="test", quiet_mode=True, protect_first_n=2, protect_last_n=2)
 
-        # Head=assistant, Tail=assistant → summary_role="user", no collision
+        # Head=assistant, Tail=assistant → summary_role="user", no collision.
+        # With min_tail=3, tail = last 3 messages (indices 5-7).
+        # Need 8 messages: min_for_compress = 2+3+1 = 6, must have > 6.
         msgs = [
             {"role": "user", "content": "msg 0"},
             {"role": "assistant", "content": "msg 1"},
             {"role": "user", "content": "msg 2"},
             {"role": "assistant", "content": "msg 3"},
-            {"role": "assistant", "content": "msg 4"},
-            {"role": "user", "content": "msg 5"},
+            {"role": "user", "content": "msg 4"},
+            {"role": "assistant", "content": "msg 5"},
+            {"role": "user", "content": "msg 6"},
+            {"role": "assistant", "content": "msg 7"},
         ]
         with patch("agent.context_compressor.call_llm", return_value=mock_response):
             result = c.compress(msgs)
@@ -600,3 +612,158 @@ class TestSummaryTargetRatio:
         with patch("agent.context_compressor.get_model_context_length", return_value=100_000):
             c = ContextCompressor(model="test", quiet_mode=True)
         assert c.protect_last_n == 20
+
+
+class TestTokenBudgetTailProtection:
+    """Tests for token-budget-based tail protection (PR #6240).
+
+    The core change: tail protection is now based on a token budget rather
+    than a fixed message count.  This prevents large tool outputs from
+    blocking compaction.
+    """
+
+    @pytest.fixture()
+    def budget_compressor(self):
+        """Compressor with known token budget for tail protection tests."""
+        with patch("agent.context_compressor.get_model_context_length", return_value=200_000):
+            c = ContextCompressor(
+                model="test/model",
+                threshold_percent=0.50,  # 100K threshold
+                protect_first_n=2,
+                protect_last_n=20,
+                quiet_mode=True,
+            )
+            return c
+
+    def test_large_tool_outputs_no_longer_block_compaction(self, budget_compressor):
+        """The motivating scenario: 20 messages with large tool outputs should
+        NOT prevent compaction.  With message-count tail protection they would
+        all be protected, leaving nothing to summarize."""
+        c = budget_compressor
+        messages = [
+            {"role": "user", "content": "Start task"},
+            {"role": "assistant", "content": "On it"},
+        ]
+        # Add 20 messages with large tool outputs (~5K chars each ≈ 1250 tokens)
+        for i in range(10):
+            messages.append({
+                "role": "assistant", "content": None,
+                "tool_calls": [{"function": {"name": f"tool_{i}", "arguments": "{}"}}],
+            })
+            messages.append({
+                "role": "tool", "content": "x" * 5000,
+                "tool_call_id": f"call_{i}",
+            })
+        # Add 3 recent small messages
+        messages.append({"role": "user", "content": "What's the status?"})
+        messages.append({"role": "assistant", "content": "Here's what I found..."})
+        messages.append({"role": "user", "content": "Continue"})
+
+        # The tail cut should NOT protect all 20 tool messages
+        head_end = c.protect_first_n
+        cut = c._find_tail_cut_by_tokens(messages, head_end)
+        tail_size = len(messages) - cut
+        # With token budget, the tail should be much smaller than 20+
+        assert tail_size < 20, f"Tail {tail_size} messages — large tool outputs are blocking compaction"
+        # But at least 3 (hard minimum)
+        assert tail_size >= 3
+
+    def test_min_tail_always_3_messages(self, budget_compressor):
+        """Even with a tiny token budget, at least 3 messages are protected."""
+        c = budget_compressor
+        # Override to a tiny budget
+        c.tail_token_budget = 10
+        messages = [
+            {"role": "user", "content": "hello"},
+            {"role": "assistant", "content": "hi"},
+            {"role": "user", "content": "do something"},
+            {"role": "assistant", "content": "working on it"},
+            {"role": "user", "content": "more work"},
+            {"role": "assistant", "content": "done"},
+            {"role": "user", "content": "thanks"},
+        ]
+        head_end = 2
+        cut = c._find_tail_cut_by_tokens(messages, head_end)
+        tail_size = len(messages) - cut
+        assert tail_size >= 3, f"Tail is only {tail_size} messages, min should be 3"
+
+    def test_soft_ceiling_allows_oversized_message(self, budget_compressor):
+        """The 1.5x soft ceiling allows an oversized message to be included
+        rather than splitting it."""
+        c = budget_compressor
+        # Set a small budget — 500 tokens
+        c.tail_token_budget = 500
+        messages = [
+            {"role": "user", "content": "hello"},
+            {"role": "assistant", "content": "hi"},
+            {"role": "user", "content": "read the file"},
+            # This message is ~600 tokens (> budget of 500, but < 1.5x = 750)
+            {"role": "assistant", "content": "a" * 2400},
+            {"role": "user", "content": "short"},
+            {"role": "assistant", "content": "short reply"},
+            {"role": "user", "content": "continue"},
+        ]
+        head_end = 2
+        cut = c._find_tail_cut_by_tokens(messages, head_end)
+        # The oversized message at index 3 should NOT be the cut point
+        # because 1.5x ceiling = 750 tokens and accumulated would be ~610
+        # (short msgs + oversized msg) which is < 750
+        tail_size = len(messages) - cut
+        assert tail_size >= 3
+
+    def test_small_conversation_still_compresses(self, budget_compressor):
+        """With the new min of 8 messages (head=2 + 3 + 1 guard + 2 middle),
+        a small but compressible conversation should still compress."""
+        c = budget_compressor
+        # 9 messages: head(2) + 4 middle + 3 tail = compressible
+        messages = []
+        for i in range(9):
+            role = "user" if i % 2 == 0 else "assistant"
+            messages.append({"role": role, "content": f"Message {i}"})
+
+        # Should not early-return (needs > protect_first_n + 3 + 1 = 6)
+        # Mock the summary generation to avoid real API call
+        with patch.object(c, "_generate_summary", return_value="Summary of conversation"):
+            result = c.compress(messages, current_tokens=90_000)
+        # Should have compressed (fewer messages than original)
+        assert len(result) < len(messages)
+
+    def test_prune_with_token_budget(self, budget_compressor):
+        """_prune_old_tool_results with protect_tail_tokens respects the budget."""
+        c = budget_compressor
+        messages = [
+            {"role": "user", "content": "start"},
+            {"role": "assistant", "content": None,
+             "tool_calls": [{"function": {"name": "read_file", "arguments": '{"path": "big.txt"}'}}]},
+            {"role": "tool", "content": "x" * 10000, "tool_call_id": "c1"},  # ~2500 tokens
+            {"role": "assistant", "content": None,
+             "tool_calls": [{"function": {"name": "read_file", "arguments": '{"path": "small.txt"}'}}]},
+            {"role": "tool", "content": "y" * 10000, "tool_call_id": "c2"},  # ~2500 tokens
+            {"role": "user", "content": "short recent message"},
+            {"role": "assistant", "content": "short reply"},
+        ]
+        # With a 1000-token budget, only the last couple messages should be protected
+        result, pruned = c._prune_old_tool_results(
+            messages, protect_tail_count=2, protect_tail_tokens=1000,
+        )
+        # At least one old tool result should have been pruned
+        assert pruned >= 1
+
+    def test_prune_without_token_budget_uses_message_count(self, budget_compressor):
+        """Without protect_tail_tokens, falls back to message-count behavior."""
+        c = budget_compressor
+        messages = [
+            {"role": "user", "content": "start"},
+            {"role": "assistant", "content": None,
+             "tool_calls": [{"function": {"name": "tool", "arguments": "{}"}}]},
+            {"role": "tool", "content": "x" * 5000, "tool_call_id": "c1"},
+            {"role": "user", "content": "recent"},
+            {"role": "assistant", "content": "reply"},
+        ]
+        # protect_tail_count=3 means last 3 messages protected
+        result, pruned = c._prune_old_tool_results(
+            messages, protect_tail_count=3,
+        )
+        # Tool at index 2 is outside the protected tail (last 3 = indices 2,3,4)
+        # so it might or might not be pruned depending on boundary
+        assert isinstance(pruned, int)

tests/agent/test_credential_pool.py

@@ -214,6 +214,42 @@ def test_exhausted_entry_resets_after_ttl(tmp_path, monkeypatch):
     assert entry.last_status == "ok"
 
 
+def test_exhausted_402_entry_resets_after_one_hour(tmp_path, monkeypatch):
+    """402-exhausted credentials recover after 1 hour, not 24."""
+    monkeypatch.setenv("HERMES_HOME", str(tmp_path / "hermes"))
+    _write_auth_store(
+        tmp_path,
+        {
+            "version": 1,
+            "credential_pool": {
+                "openrouter": [
+                    {
+                        "id": "cred-1",
+                        "label": "primary",
+                        "auth_type": "api_key",
+                        "priority": 0,
+                        "source": "manual",
+                        "access_token": "***",
+                        "base_url": "https://openrouter.ai/api/v1",
+                        "last_status": "exhausted",
+                        "last_status_at": time.time() - 3700,  # ~1h2m ago
+                        "last_error_code": 402,
+                    }
+                ]
+            },
+        },
+    )
+
+    from agent.credential_pool import load_pool
+
+    pool = load_pool("openrouter")
+    entry = pool.select()
+
+    assert entry is not None
+    assert entry.id == "cred-1"
+    assert entry.last_status == "ok"
+
+
 def test_explicit_reset_timestamp_overrides_default_429_ttl(tmp_path, monkeypatch):
     monkeypatch.setenv("HERMES_HOME", str(tmp_path / "hermes"))
     _write_auth_store(

tests/agent/test_error_classifier.py

@@ -0,0 +1,782 @@
+"""Tests for agent.error_classifier — structured API error classification."""
+
+import pytest
+from agent.error_classifier import (
+    ClassifiedError,
+    FailoverReason,
+    classify_api_error,
+    _extract_status_code,
+    _extract_error_body,
+    _extract_error_code,
+    _classify_402,
+)
+
+
+# ── Helper: mock API errors ────────────────────────────────────────────
+
+class MockAPIError(Exception):
+    """Simulates an OpenAI SDK APIStatusError."""
+    def __init__(self, message, status_code=None, body=None):
+        super().__init__(message)
+        self.status_code = status_code
+        self.body = body or {}
+
+
+class MockTransportError(Exception):
+    """Simulates a transport-level error with a specific type name."""
+    pass
+
+
+class ReadTimeout(MockTransportError):
+    pass
+
+
+class ConnectError(MockTransportError):
+    pass
+
+
+class RemoteProtocolError(MockTransportError):
+    pass
+
+
+class ServerDisconnectedError(MockTransportError):
+    pass
+
+
+# ── Test: FailoverReason enum ──────────────────────────────────────────
+
+class TestFailoverReason:
+    def test_all_reasons_have_string_values(self):
+        for reason in FailoverReason:
+            assert isinstance(reason.value, str)
+
+    def test_enum_members_exist(self):
+        expected = {
+            "auth", "auth_permanent", "billing", "rate_limit",
+            "overloaded", "server_error", "timeout",
+            "context_overflow", "payload_too_large",
+            "model_not_found", "format_error",
+            "thinking_signature", "long_context_tier", "unknown",
+        }
+        actual = {r.value for r in FailoverReason}
+        assert expected == actual
+
+
+# ── Test: ClassifiedError ──────────────────────────────────────────────
+
+class TestClassifiedError:
+    def test_is_auth_property(self):
+        e1 = ClassifiedError(reason=FailoverReason.auth)
+        assert e1.is_auth is True
+
+        e2 = ClassifiedError(reason=FailoverReason.auth_permanent)
+        assert e2.is_auth is True
+
+        e3 = ClassifiedError(reason=FailoverReason.billing)
+        assert e3.is_auth is False
+
+    def test_is_transient_property(self):
+        transient_reasons = [
+            FailoverReason.rate_limit,
+            FailoverReason.overloaded,
+            FailoverReason.server_error,
+            FailoverReason.timeout,
+            FailoverReason.unknown,
+        ]
+        for reason in transient_reasons:
+            e = ClassifiedError(reason=reason)
+            assert e.is_transient is True, f"{reason} should be transient"
+
+        non_transient = [
+            FailoverReason.auth,
+            FailoverReason.billing,
+            FailoverReason.model_not_found,
+            FailoverReason.format_error,
+        ]
+        for reason in non_transient:
+            e = ClassifiedError(reason=reason)
+            assert e.is_transient is False, f"{reason} should NOT be transient"
+
+    def test_defaults(self):
+        e = ClassifiedError(reason=FailoverReason.unknown)
+        assert e.retryable is True
+        assert e.should_compress is False
+        assert e.should_rotate_credential is False
+        assert e.should_fallback is False
+        assert e.status_code is None
+        assert e.message == ""
+
+
+# ── Test: Status code extraction ───────────────────────────────────────
+
+class TestExtractStatusCode:
+    def test_from_status_code_attr(self):
+        e = MockAPIError("fail", status_code=429)
+        assert _extract_status_code(e) == 429
+
+    def test_from_status_attr(self):
+        class ErrWithStatus(Exception):
+            status = 503
+        assert _extract_status_code(ErrWithStatus()) == 503
+
+    def test_from_cause_chain(self):
+        inner = MockAPIError("inner", status_code=401)
+        outer = Exception("outer")
+        outer.__cause__ = inner
+        assert _extract_status_code(outer) == 401
+
+    def test_none_when_missing(self):
+        assert _extract_status_code(Exception("generic")) is None
+
+    def test_rejects_non_http_status(self):
+        """Integers outside 100-599 on .status should be ignored."""
+        class ErrWeirdStatus(Exception):
+            status = 42
+        assert _extract_status_code(ErrWeirdStatus()) is None
+
+
+# ── Test: Error body extraction ────────────────────────────────────────
+
+class TestExtractErrorBody:
+    def test_from_body_attr(self):
+        e = MockAPIError("fail", body={"error": {"message": "bad"}})
+        assert _extract_error_body(e) == {"error": {"message": "bad"}}
+
+    def test_empty_when_no_body(self):
+        assert _extract_error_body(Exception("generic")) == {}
+
+
+# ── Test: Error code extraction ────────────────────────────────────────
+
+class TestExtractErrorCode:
+    def test_from_nested_error_code(self):
+        body = {"error": {"code": "rate_limit_exceeded"}}
+        assert _extract_error_code(body) == "rate_limit_exceeded"
+
+    def test_from_nested_error_type(self):
+        body = {"error": {"type": "invalid_request_error"}}
+        assert _extract_error_code(body) == "invalid_request_error"
+
+    def test_from_top_level_code(self):
+        body = {"code": "model_not_found"}
+        assert _extract_error_code(body) == "model_not_found"
+
+    def test_empty_when_no_code(self):
+        assert _extract_error_code({}) == ""
+        assert _extract_error_code({"error": {"message": "oops"}}) == ""
+
+
+# ── Test: 402 disambiguation ───────────────────────────────────────────
+
+class TestClassify402:
+    """The critical 402 billing vs rate_limit disambiguation."""
+
+    def test_billing_exhaustion(self):
+        """Plain 402 = billing."""
+        result = _classify_402(
+            "payment required",
+            lambda reason, **kw: ClassifiedError(reason=reason, **kw),
+        )
+        assert result.reason == FailoverReason.billing
+        assert result.should_rotate_credential is True
+
+    def test_transient_usage_limit(self):
+        """402 with 'usage limit' + 'try again' = rate limit, not billing."""
+        result = _classify_402(
+            "usage limit exceeded. try again in 5 minutes",
+            lambda reason, **kw: ClassifiedError(reason=reason, **kw),
+        )
+        assert result.reason == FailoverReason.rate_limit
+        assert result.should_rotate_credential is True
+
+    def test_quota_with_retry(self):
+        """402 with 'quota' + 'retry' = rate limit."""
+        result = _classify_402(
+            "quota exceeded, please retry after the window resets",
+            lambda reason, **kw: ClassifiedError(reason=reason, **kw),
+        )
+        assert result.reason == FailoverReason.rate_limit
+
+    def test_quota_without_retry(self):
+        """402 with just 'quota' but no transient signal = billing."""
+        result = _classify_402(
+            "quota exceeded",
+            lambda reason, **kw: ClassifiedError(reason=reason, **kw),
+        )
+        assert result.reason == FailoverReason.billing
+
+    def test_insufficient_credits(self):
+        result = _classify_402(
+            "insufficient credits to complete request",
+            lambda reason, **kw: ClassifiedError(reason=reason, **kw),
+        )
+        assert result.reason == FailoverReason.billing
+
+
+# ── Test: Full classification pipeline ─────────────────────────────────
+
+class TestClassifyApiError:
+    """End-to-end classification tests."""
+
+    # ── Auth errors ──
+
+    def test_401_classified_as_auth(self):
+        e = MockAPIError("Unauthorized", status_code=401)
+        result = classify_api_error(e, provider="openrouter")
+        assert result.reason == FailoverReason.auth
+        assert result.should_rotate_credential is True
+        # 401 is non-retryable on its own — credential rotation runs
+        # before the retryability check in the agent loop.
+        assert result.retryable is False
+        assert result.should_fallback is True
+
+    def test_403_classified_as_auth(self):
+        e = MockAPIError("Forbidden", status_code=403)
+        result = classify_api_error(e, provider="anthropic")
+        assert result.reason == FailoverReason.auth
+        assert result.should_fallback is True
+
+    def test_403_key_limit_classified_as_billing(self):
+        """OpenRouter 403 'key limit exceeded' is billing, not auth."""
+        e = MockAPIError("Key limit exceeded for this key", status_code=403)
+        result = classify_api_error(e, provider="openrouter")
+        assert result.reason == FailoverReason.billing
+        assert result.should_rotate_credential is True
+        assert result.should_fallback is True
+
+    def test_403_spending_limit_classified_as_billing(self):
+        e = MockAPIError("spending limit reached", status_code=403)
+        result = classify_api_error(e, provider="openrouter")
+        assert result.reason == FailoverReason.billing
+
+    # ── Billing ──
+
+    def test_402_plain_billing(self):
+        e = MockAPIError("Payment Required", status_code=402)
+        result = classify_api_error(e)
+        assert result.reason == FailoverReason.billing
+        assert result.retryable is False
+
+    def test_402_transient_usage_limit(self):
+        e = MockAPIError("usage limit exceeded, try again later", status_code=402)
+        result = classify_api_error(e)
+        assert result.reason == FailoverReason.rate_limit
+        assert result.retryable is True
+
+    # ── Rate limit ──
+
+    def test_429_rate_limit(self):
+        e = MockAPIError("Too Many Requests", status_code=429)
+        result = classify_api_error(e)
+        assert result.reason == FailoverReason.rate_limit
+        assert result.should_fallback is True
+
+    # ── Server errors ──
+
+    def test_500_server_error(self):
+        e = MockAPIError("Internal Server Error", status_code=500)
+        result = classify_api_error(e)
+        assert result.reason == FailoverReason.server_error
+        assert result.retryable is True
+
+    def test_502_server_error(self):
+        e = MockAPIError("Bad Gateway", status_code=502)
+        result = classify_api_error(e)
+        assert result.reason == FailoverReason.server_error
+
+    def test_503_overloaded(self):
+        e = MockAPIError("Service Unavailable", status_code=503)
+        result = classify_api_error(e)
+        assert result.reason == FailoverReason.overloaded
+
+    def test_529_anthropic_overloaded(self):
+        e = MockAPIError("Overloaded", status_code=529)
+        result = classify_api_error(e)
+        assert result.reason == FailoverReason.overloaded
+
+    # ── Model not found ──
+
+    def test_404_model_not_found(self):
+        e = MockAPIError("model not found", status_code=404)
+        result = classify_api_error(e)
+        assert result.reason == FailoverReason.model_not_found
+        assert result.should_fallback is True
+        assert result.retryable is False
+
+    def test_404_generic(self):
+        e = MockAPIError("Not Found", status_code=404)
+        result = classify_api_error(e)
+        assert result.reason == FailoverReason.model_not_found
+
+    # ── Payload too large ──
+
+    def test_413_payload_too_large(self):
+        e = MockAPIError("Request Entity Too Large", status_code=413)
+        result = classify_api_error(e)
+        assert result.reason == FailoverReason.payload_too_large
+        assert result.should_compress is True
+
+    # ── Context overflow ──
+
+    def test_400_context_length(self):
+        e = MockAPIError("context length exceeded: 250000 > 200000", status_code=400)
+        result = classify_api_error(e)
+        assert result.reason == FailoverReason.context_overflow
+        assert result.should_compress is True
+
+    def test_400_too_many_tokens(self):
+        e = MockAPIError("This model's maximum context is 128000 tokens, too many tokens", status_code=400)
+        result = classify_api_error(e)
+        assert result.reason == FailoverReason.context_overflow
+
+    def test_400_prompt_too_long(self):
+        e = MockAPIError("prompt is too long: 300000 tokens > 200000 maximum", status_code=400)
+        result = classify_api_error(e)
+        assert result.reason == FailoverReason.context_overflow
+
+    def test_400_generic_large_session(self):
+        """Generic 400 with large session → context overflow heuristic."""
+        e = MockAPIError(
+            "Error",
+            status_code=400,
+            body={"error": {"message": "Error"}},
+        )
+        result = classify_api_error(e, approx_tokens=100000, context_length=200000)
+        assert result.reason == FailoverReason.context_overflow
+
+    def test_400_generic_small_session_is_format_error(self):
+        """Generic 400 with small session → format error, not context overflow."""
+        e = MockAPIError(
+            "Error",
+            status_code=400,
+            body={"error": {"message": "Error"}},
+        )
+        result = classify_api_error(e, approx_tokens=1000, context_length=200000)
+        assert result.reason == FailoverReason.format_error
+
+    # ── Server disconnect + large session ──
+
+    def test_disconnect_large_session_context_overflow(self):
+        """Server disconnect with large session → context overflow."""
+        e = Exception("server disconnected without sending complete message")
+        result = classify_api_error(e, approx_tokens=150000, context_length=200000)
+        assert result.reason == FailoverReason.context_overflow
+        assert result.should_compress is True
+
+    def test_disconnect_small_session_timeout(self):
+        """Server disconnect with small session → timeout."""
+        e = Exception("server disconnected without sending complete message")
+        result = classify_api_error(e, approx_tokens=5000, context_length=200000)
+        assert result.reason == FailoverReason.timeout
+
+    # ── Provider-specific: Anthropic thinking signature ──
+
+    def test_anthropic_thinking_signature(self):
+        e = MockAPIError(
+            "thinking block has invalid signature",
+            status_code=400,
+        )
+        result = classify_api_error(e, provider="anthropic")
+        assert result.reason == FailoverReason.thinking_signature
+        assert result.retryable is True
+
+    def test_non_anthropic_400_with_signature_not_classified_as_thinking(self):
+        """400 with 'signature' but from non-Anthropic → format error."""
+        e = MockAPIError("invalid signature", status_code=400)
+        result = classify_api_error(e, provider="openrouter", approx_tokens=0)
+        # Without "thinking" in the message, it shouldn't be thinking_signature
+        assert result.reason != FailoverReason.thinking_signature
+
+    # ── Provider-specific: Anthropic long-context tier ──
+
+    def test_anthropic_long_context_tier(self):
+        e = MockAPIError(
+            "Extra usage is required for long context requests over 200k tokens",
+            status_code=429,
+        )
+        result = classify_api_error(e, provider="anthropic", model="claude-sonnet-4")
+        assert result.reason == FailoverReason.long_context_tier
+        assert result.should_compress is True
+
+    def test_normal_429_not_long_context(self):
+        """Normal 429 without 'extra usage' + 'long context' → rate_limit."""
+        e = MockAPIError("Too Many Requests", status_code=429)
+        result = classify_api_error(e, provider="anthropic")
+        assert result.reason == FailoverReason.rate_limit
+
+    # ── Transport errors ──
+
+    def test_read_timeout(self):
+        e = ReadTimeout("Read timed out")
+        result = classify_api_error(e)
+        assert result.reason == FailoverReason.timeout
+        assert result.retryable is True
+
+    def test_connect_error(self):
+        e = ConnectError("Connection refused")
+        result = classify_api_error(e)
+        assert result.reason == FailoverReason.timeout
+
+    def test_connection_error_builtin(self):
+        e = ConnectionError("Connection reset by peer")
+        result = classify_api_error(e)
+        assert result.reason == FailoverReason.timeout
+
+    def test_timeout_error_builtin(self):
+        e = TimeoutError("timed out")
+        result = classify_api_error(e)
+        assert result.reason == FailoverReason.timeout
+
+    # ── Error code classification ──
+
+    def test_error_code_resource_exhausted(self):
+        e = MockAPIError(
+            "Resource exhausted",
+            body={"error": {"code": "resource_exhausted", "message": "Too many requests"}},
+        )
+        result = classify_api_error(e)
+        assert result.reason == FailoverReason.rate_limit
+
+    def test_error_code_model_not_found(self):
+        e = MockAPIError(
+            "Model not available",
+            body={"error": {"code": "model_not_found"}},
+        )
+        result = classify_api_error(e)
+        assert result.reason == FailoverReason.model_not_found
+
+    def test_error_code_context_length_exceeded(self):
+        e = MockAPIError(
+            "Context too large",
+            body={"error": {"code": "context_length_exceeded"}},
+        )
+        result = classify_api_error(e)
+        assert result.reason == FailoverReason.context_overflow
+
+    # ── Message-only patterns (no status code) ──
+
+    def test_message_billing_pattern(self):
+        e = Exception("insufficient credits to complete this request")
+        result = classify_api_error(e)
+        assert result.reason == FailoverReason.billing
+
+    def test_message_rate_limit_pattern(self):
+        e = Exception("rate limit reached for this model")
+        result = classify_api_error(e)
+        assert result.reason == FailoverReason.rate_limit
+
+    def test_message_auth_pattern(self):
+        e = Exception("invalid api key provided")
+        result = classify_api_error(e)
+        assert result.reason == FailoverReason.auth
+
+    def test_message_model_not_found_pattern(self):
+        e = Exception("gpt-99 is not a valid model")
+        result = classify_api_error(e)
+        assert result.reason == FailoverReason.model_not_found
+
+    def test_message_context_overflow_pattern(self):
+        e = Exception("maximum context length exceeded")
+        result = classify_api_error(e)
+        assert result.reason == FailoverReason.context_overflow
+
+    # ── Unknown / fallback ──
+
+    def test_generic_exception_is_unknown(self):
+        e = Exception("something weird happened")
+        result = classify_api_error(e)
+        assert result.reason == FailoverReason.unknown
+        assert result.retryable is True
+
+    # ── Format error ──
+
+    def test_400_descriptive_format_error(self):
+        """400 with descriptive message (not context overflow) → format error."""
+        e = MockAPIError(
+            "Invalid value for parameter 'temperature': must be between 0 and 2",
+            status_code=400,
+            body={"error": {"message": "Invalid value for parameter 'temperature': must be between 0 and 2"}},
+        )
+        result = classify_api_error(e, approx_tokens=1000)
+        assert result.reason == FailoverReason.format_error
+        assert result.retryable is False
+
+    def test_422_format_error(self):
+        e = MockAPIError("Unprocessable Entity", status_code=422)
+        result = classify_api_error(e)
+        assert result.reason == FailoverReason.format_error
+        assert result.retryable is False
+
+    def test_400_flat_body_descriptive_not_context_overflow(self):
+        """Responses API flat body with descriptive error + large session → format error.
+
+        The Codex Responses API returns errors in flat body format:
+        {"message": "...", "type": "..."} without an "error" wrapper.
+        A descriptive 400 must NOT be misclassified as context overflow
+        just because the session is large.
+        """
+        e = MockAPIError(
+            "Invalid 'input[index].name': string does not match pattern.",
+            status_code=400,
+            body={"message": "Invalid 'input[index].name': string does not match pattern.",
+                  "type": "invalid_request_error"},
+        )
+        result = classify_api_error(e, approx_tokens=200000, context_length=400000, num_messages=500)
+        assert result.reason == FailoverReason.format_error
+        assert result.retryable is False
+
+    def test_400_flat_body_generic_large_session_still_context_overflow(self):
+        """Flat body with generic 'Error' message + large session → context overflow.
+
+        Regression: the flat-body fallback must not break the existing heuristic
+        for genuinely generic errors from providers that use flat bodies.
+        """
+        e = MockAPIError(
+            "Error",
+            status_code=400,
+            body={"message": "Error"},
+        )
+        result = classify_api_error(e, approx_tokens=100000, context_length=200000)
+        assert result.reason == FailoverReason.context_overflow
+
+    # ── Peer closed + large session ──
+
+    def test_peer_closed_large_session(self):
+        e = Exception("peer closed connection without sending complete message")
+        result = classify_api_error(e, approx_tokens=130000, context_length=200000)
+        assert result.reason == FailoverReason.context_overflow
+
+    # ── Chinese error messages ──
+
+    def test_chinese_context_overflow(self):
+        e = MockAPIError("超过最大长度限制", status_code=400)
+        result = classify_api_error(e)
+        assert result.reason == FailoverReason.context_overflow
+
+    # ── Result metadata ──
+
+    def test_provider_and_model_in_result(self):
+        e = MockAPIError("fail", status_code=500)
+        result = classify_api_error(e, provider="openrouter", model="gpt-5")
+        assert result.provider == "openrouter"
+        assert result.model == "gpt-5"
+        assert result.status_code == 500
+
+    def test_message_extracted(self):
+        e = MockAPIError(
+            "outer",
+            status_code=500,
+            body={"error": {"message": "Internal server error occurred"}},
+        )
+        result = classify_api_error(e)
+        assert result.message == "Internal server error occurred"
+
+
+# ── Test: Adversarial / edge cases (from live testing) ─────────────────
+
+class TestAdversarialEdgeCases:
+    """Edge cases discovered during live testing with real SDK objects."""
+
+    def test_empty_exception_message(self):
+        result = classify_api_error(Exception(""))
+        assert result.reason == FailoverReason.unknown
+        assert result.retryable is True
+
+    def test_500_with_none_body(self):
+        e = MockAPIError("fail", status_code=500, body=None)
+        result = classify_api_error(e)
+        assert result.reason == FailoverReason.server_error
+
+    def test_non_dict_body(self):
+        """Some providers return strings instead of JSON."""
+        class StringBodyError(Exception):
+            status_code = 400
+            body = "just a string"
+        result = classify_api_error(StringBodyError("bad"))
+        assert result.reason == FailoverReason.format_error
+
+    def test_list_body(self):
+        class ListBodyError(Exception):
+            status_code = 500
+            body = [{"error": "something"}]
+        result = classify_api_error(ListBodyError("server error"))
+        assert result.reason == FailoverReason.server_error
+
+    def test_circular_cause_chain(self):
+        """Must not infinite-loop on circular __cause__."""
+        e = Exception("circular")
+        e.__cause__ = e
+        result = classify_api_error(e)
+        assert result.reason == FailoverReason.unknown
+
+    def test_three_level_cause_chain(self):
+        inner = MockAPIError("inner", status_code=429)
+        middle = Exception("middle")
+        middle.__cause__ = inner
+        outer = RuntimeError("outer")
+        outer.__cause__ = middle
+        result = classify_api_error(outer)
+        assert result.status_code == 429
+        assert result.reason == FailoverReason.rate_limit
+
+    def test_400_with_rate_limit_text(self):
+        """Some providers send rate limits as 400 instead of 429."""
+        e = MockAPIError(
+            "rate limit policy",
+            status_code=400,
+            body={"error": {"message": "rate limit exceeded on this model"}},
+        )
+        result = classify_api_error(e, provider="openrouter")
+        assert result.reason == FailoverReason.rate_limit
+
+    def test_400_with_billing_text(self):
+        """Some providers send billing errors as 400."""
+        e = MockAPIError(
+            "billing",
+            status_code=400,
+            body={"error": {"message": "insufficient credits for this request"}},
+        )
+        result = classify_api_error(e)
+        assert result.reason == FailoverReason.billing
+
+    def test_200_with_error_body(self):
+        """200 status with error in body — should be unknown, not crash."""
+        class WeirdSuccess(Exception):
+            status_code = 200
+            body = {"error": {"message": "loading"}}
+        result = classify_api_error(WeirdSuccess("model loading"))
+        assert result.reason == FailoverReason.unknown
+
+    def test_ollama_context_size_exceeded(self):
+        e = MockAPIError(
+            "Error",
+            status_code=400,
+            body={"error": {"message": "context size has been exceeded"}},
+        )
+        result = classify_api_error(e, provider="ollama")
+        assert result.reason == FailoverReason.context_overflow
+
+    def test_connection_refused_error(self):
+        e = ConnectionRefusedError("Connection refused: localhost:11434")
+        result = classify_api_error(e, provider="ollama")
+        assert result.reason == FailoverReason.timeout
+
+    def test_body_message_enrichment(self):
+        """Body message must be included in pattern matching even when
+        str(error) doesn't contain it (OpenAI SDK APIStatusError)."""
+        e = MockAPIError(
+            "Usage limit",  # str(e) = "usage limit"
+            status_code=402,
+            body={"error": {"message": "Usage limit reached, try again in 5 minutes"}},
+        )
+        result = classify_api_error(e)
+        # "try again" is only in body, not in str(e)
+        assert result.reason == FailoverReason.rate_limit
+
+    def test_disconnect_pattern_ordering(self):
+        """Disconnect + large session must beat generic transport catch."""
+        class FakeRemoteProtocol(Exception):
+            pass
+        # Type name isn't in _TRANSPORT_ERROR_TYPES but message has disconnect pattern
+        e = Exception("peer closed connection without sending complete message")
+        result = classify_api_error(e, approx_tokens=150000, context_length=200000)
+        assert result.reason == FailoverReason.context_overflow
+        assert result.should_compress is True
+
+    def test_credit_balance_too_low(self):
+        e = MockAPIError(
+            "Credits low",
+            status_code=402,
+            body={"error": {"message": "Your credit balance is too low"}},
+        )
+        result = classify_api_error(e, provider="anthropic")
+        assert result.reason == FailoverReason.billing
+
+    def test_deepseek_402_chinese(self):
+        """Chinese billing message should still match billing patterns."""
+        # "余额不足" doesn't match English billing patterns, but 402 defaults to billing
+        e = MockAPIError("余额不足", status_code=402)
+        result = classify_api_error(e, provider="deepseek")
+        assert result.reason == FailoverReason.billing
+
+    def test_openrouter_wrapped_context_overflow_in_metadata_raw(self):
+        """OpenRouter wraps provider errors in metadata.raw JSON string."""
+        e = MockAPIError(
+            "Provider returned error",
+            status_code=400,
+            body={
+                "error": {
+                    "message": "Provider returned error",
+                    "code": 400,
+                    "metadata": {
+                        "raw": '{"error":{"message":"context length exceeded: 50000 > 32768"}}'
+                    }
+                }
+            },
+        )
+        result = classify_api_error(e, provider="openrouter", approx_tokens=10000)
+        assert result.reason == FailoverReason.context_overflow
+        assert result.should_compress is True
+
+    def test_openrouter_wrapped_rate_limit_in_metadata_raw(self):
+        e = MockAPIError(
+            "Provider returned error",
+            status_code=400,
+            body={
+                "error": {
+                    "message": "Provider returned error",
+                    "metadata": {
+                        "raw": '{"error":{"message":"Rate limit exceeded. Please retry after 30s."}}'
+                    }
+                }
+            },
+        )
+        result = classify_api_error(e, provider="openrouter")
+        assert result.reason == FailoverReason.rate_limit
+
+    def test_thinking_signature_via_openrouter(self):
+        """Thinking signature errors proxied through OpenRouter must be caught."""
+        e = MockAPIError(
+            "thinking block has invalid signature",
+            status_code=400,
+        )
+        # provider is openrouter, not anthropic — old code missed this
+        result = classify_api_error(e, provider="openrouter", model="anthropic/claude-sonnet-4")
+        assert result.reason == FailoverReason.thinking_signature
+
+    def test_generic_400_large_by_message_count(self):
+        """Many small messages (>80) should trigger context overflow heuristic."""
+        e = MockAPIError(
+            "Error",
+            status_code=400,
+            body={"error": {"message": "Error"}},
+        )
+        # Low token count but high message count
+        result = classify_api_error(
+            e, approx_tokens=5000, context_length=200000, num_messages=100,
+        )
+        assert result.reason == FailoverReason.context_overflow
+
+    def test_disconnect_large_by_message_count(self):
+        """Server disconnect with 200+ messages should trigger context overflow."""
+        e = Exception("server disconnected without sending complete message")
+        result = classify_api_error(
+            e, approx_tokens=5000, context_length=200000, num_messages=250,
+        )
+        assert result.reason == FailoverReason.context_overflow
+
+    def test_openrouter_wrapped_model_not_found_in_metadata_raw(self):
+        e = MockAPIError(
+            "Provider returned error",
+            status_code=400,
+            body={
+                "error": {
+                    "message": "Provider returned error",
+                    "metadata": {
+                        "raw": '{"error":{"message":"The model gpt-99 does not exist"}}'
+                    }
+                }
+            },
+        )
+        result = classify_api_error(e, provider="openrouter")
+        assert result.reason == FailoverReason.model_not_found

tests/agent/test_rate_limit_tracker.py

@@ -0,0 +1,212 @@
+"""Tests for agent.rate_limit_tracker — header parsing and formatting."""
+
+import time
+import pytest
+from agent.rate_limit_tracker import (
+    RateLimitBucket,
+    RateLimitState,
+    parse_rate_limit_headers,
+    format_rate_limit_display,
+    format_rate_limit_compact,
+    _fmt_count,
+    _fmt_seconds,
+    _bar,
+)
+
+
+# ── Sample headers from Nous inference API ──────────────────────────────
+
+NOUS_HEADERS = {
+    "x-ratelimit-limit-requests": "800",
+    "x-ratelimit-limit-requests-1h": "33600",
+    "x-ratelimit-limit-tokens": "8000000",
+    "x-ratelimit-limit-tokens-1h": "336000000",
+    "x-ratelimit-remaining-requests": "795",
+    "x-ratelimit-remaining-requests-1h": "33590",
+    "x-ratelimit-remaining-tokens": "7999500",
+    "x-ratelimit-remaining-tokens-1h": "335999000",
+    "x-ratelimit-reset-requests": "45.5",
+    "x-ratelimit-reset-requests-1h": "3500.0",
+    "x-ratelimit-reset-tokens": "42.3",
+    "x-ratelimit-reset-tokens-1h": "3490.0",
+}
+
+
+class TestParseHeaders:
+    def test_basic_parsing(self):
+        state = parse_rate_limit_headers(NOUS_HEADERS, provider="nous")
+        assert state is not None
+        assert state.provider == "nous"
+        assert state.has_data
+
+        assert state.requests_min.limit == 800
+        assert state.requests_min.remaining == 795
+        assert state.requests_min.reset_seconds == 45.5
+
+        assert state.requests_hour.limit == 33600
+        assert state.requests_hour.remaining == 33590
+
+        assert state.tokens_min.limit == 8000000
+        assert state.tokens_min.remaining == 7999500
+
+        assert state.tokens_hour.limit == 336000000
+        assert state.tokens_hour.remaining == 335999000
+        assert state.tokens_hour.reset_seconds == 3490.0
+
+    def test_no_headers(self):
+        state = parse_rate_limit_headers({})
+        assert state is None
+
+    def test_partial_headers(self):
+        headers = {
+            "x-ratelimit-limit-requests": "100",
+            "x-ratelimit-remaining-requests": "50",
+        }
+        state = parse_rate_limit_headers(headers)
+        assert state is not None
+        assert state.requests_min.limit == 100
+        assert state.requests_min.remaining == 50
+        # Missing fields default to 0
+        assert state.tokens_min.limit == 0
+
+    def test_non_rate_limit_headers_ignored(self):
+        headers = {
+            "content-type": "application/json",
+            "server": "nginx",
+        }
+        state = parse_rate_limit_headers(headers)
+        assert state is None
+
+    def test_malformed_values(self):
+        headers = {
+            "x-ratelimit-limit-requests": "not-a-number",
+            "x-ratelimit-remaining-requests": "",
+            "x-ratelimit-reset-requests": "abc",
+        }
+        state = parse_rate_limit_headers(headers)
+        assert state is not None
+        assert state.requests_min.limit == 0
+        assert state.requests_min.remaining == 0
+        assert state.requests_min.reset_seconds == 0.0
+
+
+class TestBucket:
+    def test_used(self):
+        b = RateLimitBucket(limit=800, remaining=795, reset_seconds=45.0, captured_at=time.time())
+        assert b.used == 5
+
+    def test_usage_pct(self):
+        b = RateLimitBucket(limit=100, remaining=20, reset_seconds=30.0, captured_at=time.time())
+        assert b.usage_pct == pytest.approx(80.0)
+
+    def test_usage_pct_zero_limit(self):
+        b = RateLimitBucket(limit=0, remaining=0)
+        assert b.usage_pct == 0.0
+
+    def test_remaining_seconds_now(self):
+        now = time.time()
+        b = RateLimitBucket(limit=800, remaining=795, reset_seconds=60.0, captured_at=now - 10)
+        # ~50 seconds should remain
+        assert 49 <= b.remaining_seconds_now <= 51
+
+    def test_remaining_seconds_expired(self):
+        b = RateLimitBucket(limit=800, remaining=795, reset_seconds=30.0, captured_at=time.time() - 60)
+        assert b.remaining_seconds_now == 0.0
+
+
+class TestFormatting:
+    def test_fmt_count_millions(self):
+        assert _fmt_count(8000000) == "8.0M"
+        assert _fmt_count(336000000) == "336.0M"
+
+    def test_fmt_count_thousands(self):
+        assert _fmt_count(33600) == "33.6K"
+        assert _fmt_count(1500) == "1.5K"
+
+    def test_fmt_count_small(self):
+        assert _fmt_count(800) == "800"
+        assert _fmt_count(0) == "0"
+
+    def test_fmt_seconds_short(self):
+        assert _fmt_seconds(45) == "45s"
+        assert _fmt_seconds(0) == "0s"
+
+    def test_fmt_seconds_minutes(self):
+        assert _fmt_seconds(125) == "2m 5s"
+        assert _fmt_seconds(120) == "2m"
+
+    def test_fmt_seconds_hours(self):
+        assert _fmt_seconds(3660) == "1h 1m"
+        assert _fmt_seconds(3600) == "1h"
+
+    def test_bar(self):
+        bar = _bar(50.0, width=10)
+        assert bar == "[█████░░░░░]"
+        assert _bar(0.0, width=10) == "[░░░░░░░░░░]"
+        assert _bar(100.0, width=10) == "[██████████]"
+
+    def test_format_display_no_data(self):
+        state = RateLimitState()
+        result = format_rate_limit_display(state)
+        assert "No rate limit data" in result
+
+    def test_format_display_with_data(self):
+        state = parse_rate_limit_headers(NOUS_HEADERS, provider="nous")
+        result = format_rate_limit_display(state)
+        assert "Nous" in result
+        assert "Requests/min" in result
+        assert "Requests/hr" in result
+        assert "Tokens/min" in result
+        assert "Tokens/hr" in result
+        assert "resets in" in result
+
+    def test_format_display_warning_on_high_usage(self):
+        headers = {
+            **NOUS_HEADERS,
+            "x-ratelimit-remaining-requests": "50",  # 750/800 used = 93.75%
+        }
+        state = parse_rate_limit_headers(headers)
+        result = format_rate_limit_display(state)
+        assert "⚠" in result
+
+    def test_format_compact(self):
+        state = parse_rate_limit_headers(NOUS_HEADERS, provider="nous")
+        result = format_rate_limit_compact(state)
+        assert "RPM:" in result
+        assert "RPH:" in result
+        assert "TPM:" in result
+        assert "TPH:" in result
+        assert "resets" in result
+
+    def test_format_compact_no_data(self):
+        state = RateLimitState()
+        result = format_rate_limit_compact(state)
+        assert "No rate limit data" in result
+
+
+class TestAgentIntegration:
+    """Test that AIAgent captures rate limit state correctly."""
+
+    def test_capture_rate_limits_from_headers(self):
+        """Simulate the header capture path without a real API call."""
+        import sys
+        import os
+        # Use a mock httpx-like response
+        class MockResponse:
+            headers = NOUS_HEADERS
+
+        # Import AIAgent minimally
+        from unittest.mock import MagicMock, patch
+
+        # Test the parsing directly
+        state = parse_rate_limit_headers(MockResponse.headers, provider="nous")
+        assert state is not None
+        assert state.requests_min.limit == 800
+        assert state.tokens_hour.limit == 336000000
+
+    def test_capture_rate_limits_none_response(self):
+        """_capture_rate_limits should handle None gracefully."""
+        from agent.rate_limit_tracker import parse_rate_limit_headers
+        # None should not crash
+        result = parse_rate_limit_headers({})
+        assert result is None

tests/agent/test_subdirectory_hints.py

@@ -3,6 +3,7 @@
 import os
 import pytest
 from pathlib import Path
+from unittest.mock import patch
 
 from agent.subdirectory_hints import SubdirectoryHintTracker
 
@@ -189,3 +190,45 @@ class TestSubdirectoryHintTracker:
             "terminal", {"command": "curl https://example.com/frontend/api"}
         )
         assert result is None
+
+
+class TestPermissionErrorHandling:
+    """Regression tests for PermissionError in filesystem checks (ref #6214)."""
+
+    def test_is_valid_subdir_permission_error(self, tmp_path):
+        """_is_valid_subdir should return False when is_dir() raises PermissionError."""
+        tracker = SubdirectoryHintTracker(working_dir=str(tmp_path))
+        restricted = tmp_path / "restricted"
+        restricted.mkdir()
+        with patch.object(Path, "is_dir", side_effect=PermissionError("Permission denied")):
+            assert tracker._is_valid_subdir(restricted) is False
+
+    def test_load_hints_permission_error_on_is_file(self, tmp_path):
+        """_load_hints_for_directory should skip files when is_file() raises PermissionError."""
+        tracker = SubdirectoryHintTracker(working_dir=str(tmp_path))
+        restricted = tmp_path / "restricted"
+        restricted.mkdir()
+        original_is_file = Path.is_file
+        def patched_is_file(self):
+            if "restricted" in str(self):
+                raise PermissionError("Permission denied")
+            return original_is_file(self)
+        with patch.object(Path, "is_file", patched_is_file):
+            result = tracker._load_hints_for_directory(restricted)
+        assert result is None
+
+    def test_check_tool_call_survives_inaccessible_path(self, project):
+        """Full check_tool_call should not crash when a path is inaccessible."""
+        tracker = SubdirectoryHintTracker(working_dir=str(project))
+        original_is_dir = Path.is_dir
+        def patched_is_dir(self):
+            if "backend" in str(self) and "src" not in str(self):
+                raise PermissionError("Permission denied")
+            return original_is_dir(self)
+        with patch.object(Path, "is_dir", patched_is_dir):
+            # Should not raise — gracefully skip the inaccessible directory
+            result = tracker.check_tool_call(
+                "read_file", {"path": str(project / "backend" / "src" / "main.py")}
+            )
+            # Result may be None (backend skipped) — the key point is no crash
+            assert result is None or isinstance(result, str)

tests/cli/test_cli_approval_ui.py

@@ -2,22 +2,65 @@ import queue
 import threading
 import time
 from types import SimpleNamespace
-from unittest.mock import MagicMock
+from unittest.mock import MagicMock, patch
 
+import cli as cli_module
 from cli import HermesCLI
 
 
+class _FakeBuffer:
+    def __init__(self, text="", cursor_position=None):
+        self.text = text
+        self.cursor_position = len(text) if cursor_position is None else cursor_position
+
+    def reset(self, append_to_history=False):
+        self.text = ""
+        self.cursor_position = 0
+
+
 def _make_cli_stub():
     cli = HermesCLI.__new__(HermesCLI)
     cli._approval_state = None
     cli._approval_deadline = 0
     cli._approval_lock = threading.Lock()
+    cli._sudo_state = None
+    cli._sudo_deadline = 0
+    cli._modal_input_snapshot = None
     cli._invalidate = MagicMock()
-    cli._app = SimpleNamespace(invalidate=MagicMock())
+    cli._app = SimpleNamespace(invalidate=MagicMock(), current_buffer=_FakeBuffer())
     return cli
 
 
 class TestCliApprovalUi:
+    def test_sudo_prompt_restores_existing_draft_after_response(self):
+        cli = _make_cli_stub()
+        cli._app.current_buffer = _FakeBuffer("draft command", cursor_position=5)
+        result = {}
+
+        def _run_callback():
+            result["value"] = cli._sudo_password_callback()
+
+        with patch.object(cli_module, "_cprint"):
+            thread = threading.Thread(target=_run_callback, daemon=True)
+            thread.start()
+
+            deadline = time.time() + 2
+            while cli._sudo_state is None and time.time() < deadline:
+                time.sleep(0.01)
+
+            assert cli._sudo_state is not None
+            assert cli._app.current_buffer.text == ""
+
+            cli._app.current_buffer.text = "secret"
+            cli._app.current_buffer.cursor_position = len("secret")
+            cli._sudo_state["response_queue"].put("secret")
+
+            thread.join(timeout=2)
+
+        assert result["value"] == "secret"
+        assert cli._app.current_buffer.text == "draft command"
+        assert cli._app.current_buffer.cursor_position == 5
+
     def test_approval_callback_includes_view_for_long_commands(self):
         cli = _make_cli_stub()
         command = "sudo dd if=/tmp/githubcli-keyring.gpg of=/usr/share/keyrings/githubcli-archive-keyring.gpg bs=4M status=progress"

tests/cli/test_cli_status_bar.py

@@ -41,6 +41,7 @@ def _attach_agent(
         session_completion_tokens=completion_tokens,
         session_total_tokens=total_tokens,
         session_api_calls=api_calls,
+        get_rate_limit_state=lambda: None,
         context_compressor=SimpleNamespace(
             last_prompt_tokens=context_tokens,
             context_length=context_length,

tests/conftest.py

@@ -38,6 +38,8 @@ def _isolate_hermes_home(tmp_path, monkeypatch):
     monkeypatch.delenv("HERMES_SESSION_CHAT_ID", raising=False)
     monkeypatch.delenv("HERMES_SESSION_CHAT_NAME", raising=False)
     monkeypatch.delenv("HERMES_GATEWAY_SESSION", raising=False)
+    # Avoid making real calls during tests if this key is set in the env files
+    monkeypatch.delenv("OPENROUTER_API_KEY", raising=False)
 
 
 @pytest.fixture()

tests/gateway/test_bluebubbles.py

@@ -0,0 +1,361 @@
+"""Tests for the BlueBubbles iMessage gateway adapter."""
+import pytest
+
+from gateway.config import Platform, PlatformConfig
+
+
+def _make_adapter(monkeypatch, **extra):
+    monkeypatch.setenv("BLUEBUBBLES_SERVER_URL", "http://localhost:1234")
+    monkeypatch.setenv("BLUEBUBBLES_PASSWORD", "secret")
+    from gateway.platforms.bluebubbles import BlueBubblesAdapter
+
+    cfg = PlatformConfig(
+        enabled=True,
+        extra={
+            "server_url": "http://localhost:1234",
+            "password": "secret",
+            **extra,
+        },
+    )
+    return BlueBubblesAdapter(cfg)
+
+
+class TestBlueBubblesPlatformEnum:
+    def test_bluebubbles_enum_exists(self):
+        assert Platform.BLUEBUBBLES.value == "bluebubbles"
+
+
+class TestBlueBubblesConfigLoading:
+    def test_apply_env_overrides_bluebubbles(self, monkeypatch):
+        monkeypatch.setenv("BLUEBUBBLES_SERVER_URL", "http://localhost:1234")
+        monkeypatch.setenv("BLUEBUBBLES_PASSWORD", "secret")
+        monkeypatch.setenv("BLUEBUBBLES_WEBHOOK_PORT", "9999")
+        from gateway.config import GatewayConfig, _apply_env_overrides
+
+        config = GatewayConfig()
+        _apply_env_overrides(config)
+        assert Platform.BLUEBUBBLES in config.platforms
+        bc = config.platforms[Platform.BLUEBUBBLES]
+        assert bc.enabled is True
+        assert bc.extra["server_url"] == "http://localhost:1234"
+        assert bc.extra["password"] == "secret"
+        assert bc.extra["webhook_port"] == 9999
+
+    def test_connected_platforms_includes_bluebubbles(self, monkeypatch):
+        monkeypatch.setenv("BLUEBUBBLES_SERVER_URL", "http://localhost:1234")
+        monkeypatch.setenv("BLUEBUBBLES_PASSWORD", "secret")
+        from gateway.config import GatewayConfig, _apply_env_overrides
+
+        config = GatewayConfig()
+        _apply_env_overrides(config)
+        assert Platform.BLUEBUBBLES in config.get_connected_platforms()
+
+    def test_home_channel_set_from_env(self, monkeypatch):
+        monkeypatch.setenv("BLUEBUBBLES_SERVER_URL", "http://localhost:1234")
+        monkeypatch.setenv("BLUEBUBBLES_PASSWORD", "secret")
+        monkeypatch.setenv("BLUEBUBBLES_HOME_CHANNEL", "user@example.com")
+        from gateway.config import GatewayConfig, _apply_env_overrides
+
+        config = GatewayConfig()
+        _apply_env_overrides(config)
+        hc = config.platforms[Platform.BLUEBUBBLES].home_channel
+        assert hc is not None
+        assert hc.chat_id == "user@example.com"
+
+    def test_not_connected_without_password(self, monkeypatch):
+        monkeypatch.setenv("BLUEBUBBLES_SERVER_URL", "http://localhost:1234")
+        monkeypatch.delenv("BLUEBUBBLES_PASSWORD", raising=False)
+        from gateway.config import GatewayConfig, _apply_env_overrides
+
+        config = GatewayConfig()
+        _apply_env_overrides(config)
+        assert Platform.BLUEBUBBLES not in config.get_connected_platforms()
+
+
+class TestBlueBubblesHelpers:
+    def test_check_requirements(self, monkeypatch):
+        monkeypatch.setenv("BLUEBUBBLES_SERVER_URL", "http://localhost:1234")
+        monkeypatch.setenv("BLUEBUBBLES_PASSWORD", "secret")
+        from gateway.platforms.bluebubbles import check_bluebubbles_requirements
+
+        assert check_bluebubbles_requirements() is True
+
+    def test_format_message_strips_markdown(self, monkeypatch):
+        adapter = _make_adapter(monkeypatch)
+        assert adapter.format_message("**Hello** `world`") == "Hello world"
+
+    def test_strip_markdown_headers(self, monkeypatch):
+        adapter = _make_adapter(monkeypatch)
+        assert adapter.format_message("## Heading\ntext") == "Heading\ntext"
+
+    def test_strip_markdown_links(self, monkeypatch):
+        adapter = _make_adapter(monkeypatch)
+        assert adapter.format_message("[click here](http://example.com)") == "click here"
+
+    def test_init_normalizes_webhook_path(self, monkeypatch):
+        adapter = _make_adapter(monkeypatch, webhook_path="bluebubbles-webhook")
+        assert adapter.webhook_path == "/bluebubbles-webhook"
+
+    def test_init_preserves_leading_slash(self, monkeypatch):
+        adapter = _make_adapter(monkeypatch, webhook_path="/my-hook")
+        assert adapter.webhook_path == "/my-hook"
+
+    def test_server_url_normalized(self, monkeypatch):
+        adapter = _make_adapter(monkeypatch, server_url="http://localhost:1234/")
+        assert adapter.server_url == "http://localhost:1234"
+
+    def test_server_url_adds_scheme(self, monkeypatch):
+        adapter = _make_adapter(monkeypatch, server_url="localhost:1234")
+        assert adapter.server_url == "http://localhost:1234"
+
+
+class TestBlueBubblesWebhookParsing:
+    def test_webhook_prefers_chat_guid_over_message_guid(self, monkeypatch):
+        adapter = _make_adapter(monkeypatch)
+        payload = {
+            "guid": "MESSAGE-GUID",
+            "chatGuid": "iMessage;-;user@example.com",
+            "chatIdentifier": "user@example.com",
+        }
+        record = adapter._extract_payload_record(payload) or {}
+        chat_guid = adapter._value(
+            record.get("chatGuid"),
+            payload.get("chatGuid"),
+            record.get("chat_guid"),
+            payload.get("chat_guid"),
+            payload.get("guid"),
+        )
+        assert chat_guid == "iMessage;-;user@example.com"
+
+    def test_webhook_can_fall_back_to_sender_when_chat_fields_missing(self, monkeypatch):
+        adapter = _make_adapter(monkeypatch)
+        payload = {
+            "data": {
+                "guid": "MESSAGE-GUID",
+                "text": "hello",
+                "handle": {"address": "user@example.com"},
+                "isFromMe": False,
+            }
+        }
+        record = adapter._extract_payload_record(payload) or {}
+        chat_guid = adapter._value(
+            record.get("chatGuid"),
+            payload.get("chatGuid"),
+            record.get("chat_guid"),
+            payload.get("chat_guid"),
+            payload.get("guid"),
+        )
+        chat_identifier = adapter._value(
+            record.get("chatIdentifier"),
+            record.get("identifier"),
+            payload.get("chatIdentifier"),
+            payload.get("identifier"),
+        )
+        sender = (
+            adapter._value(
+                record.get("handle", {}).get("address")
+                if isinstance(record.get("handle"), dict)
+                else None,
+                record.get("sender"),
+                record.get("from"),
+                record.get("address"),
+            )
+            or chat_identifier
+            or chat_guid
+        )
+        if not (chat_guid or chat_identifier) and sender:
+            chat_identifier = sender
+        assert chat_identifier == "user@example.com"
+
+    def test_extract_payload_record_accepts_list_data(self, monkeypatch):
+        adapter = _make_adapter(monkeypatch)
+        payload = {
+            "type": "new-message",
+            "data": [
+                {
+                    "text": "hello",
+                    "chatGuid": "iMessage;-;user@example.com",
+                    "chatIdentifier": "user@example.com",
+                }
+            ],
+        }
+        record = adapter._extract_payload_record(payload)
+        assert record == payload["data"][0]
+
+    def test_extract_payload_record_dict_data(self, monkeypatch):
+        adapter = _make_adapter(monkeypatch)
+        payload = {"data": {"text": "hello", "chatGuid": "iMessage;-;+1234"}}
+        record = adapter._extract_payload_record(payload)
+        assert record["text"] == "hello"
+
+    def test_extract_payload_record_fallback_to_message(self, monkeypatch):
+        adapter = _make_adapter(monkeypatch)
+        payload = {"message": {"text": "hello"}}
+        record = adapter._extract_payload_record(payload)
+        assert record["text"] == "hello"
+
+
+class TestBlueBubblesGuidResolution:
+    def test_raw_guid_returned_as_is(self, monkeypatch):
+        """If target already contains ';' it's a raw GUID — return unchanged."""
+        adapter = _make_adapter(monkeypatch)
+        import asyncio
+
+        result = asyncio.get_event_loop().run_until_complete(
+            adapter._resolve_chat_guid("iMessage;-;user@example.com")
+        )
+        assert result == "iMessage;-;user@example.com"
+
+    def test_empty_target_returns_none(self, monkeypatch):
+        adapter = _make_adapter(monkeypatch)
+        import asyncio
+
+        result = asyncio.get_event_loop().run_until_complete(
+            adapter._resolve_chat_guid("")
+        )
+        assert result is None
+
+
+class TestBlueBubblesToolsetIntegration:
+    def test_toolset_exists(self):
+        from toolsets import TOOLSETS
+
+        assert "hermes-bluebubbles" in TOOLSETS
+
+    def test_toolset_in_gateway_composite(self):
+        from toolsets import TOOLSETS
+
+        gateway = TOOLSETS["hermes-gateway"]
+        assert "hermes-bluebubbles" in gateway["includes"]
+
+
+class TestBlueBubblesPromptHint:
+    def test_platform_hint_exists(self):
+        from agent.prompt_builder import PLATFORM_HINTS
+
+        assert "bluebubbles" in PLATFORM_HINTS
+        hint = PLATFORM_HINTS["bluebubbles"]
+        assert "iMessage" in hint
+        assert "plain text" in hint
+
+
+class TestBlueBubblesAttachmentDownload:
+    """Verify _download_attachment routes to the correct cache helper."""
+
+    def test_download_image_uses_image_cache(self, monkeypatch):
+        """Image MIME routes to cache_image_from_bytes."""
+        adapter = _make_adapter(monkeypatch)
+        import asyncio
+        import httpx
+
+        # Mock the HTTP client response
+        class MockResponse:
+            status_code = 200
+            content = b"\x89PNG\r\n\x1a\n"
+
+            def raise_for_status(self):
+                pass
+
+        async def mock_get(*args, **kwargs):
+            return MockResponse()
+
+        adapter.client = type("MockClient", (), {"get": mock_get})()
+
+        cached_path = None
+
+        def mock_cache_image(data, ext):
+            nonlocal cached_path
+            cached_path = f"/tmp/test_image{ext}"
+            return cached_path
+
+        monkeypatch.setattr(
+            "gateway.platforms.bluebubbles.cache_image_from_bytes",
+            mock_cache_image,
+        )
+
+        att_meta = {"mimeType": "image/png", "transferName": "photo.png"}
+        result = asyncio.get_event_loop().run_until_complete(
+            adapter._download_attachment("att-guid-123", att_meta)
+        )
+        assert result == "/tmp/test_image.png"
+
+    def test_download_audio_uses_audio_cache(self, monkeypatch):
+        """Audio MIME routes to cache_audio_from_bytes."""
+        adapter = _make_adapter(monkeypatch)
+        import asyncio
+
+        class MockResponse:
+            status_code = 200
+            content = b"fake-audio-data"
+
+            def raise_for_status(self):
+                pass
+
+        async def mock_get(*args, **kwargs):
+            return MockResponse()
+
+        adapter.client = type("MockClient", (), {"get": mock_get})()
+
+        cached_path = None
+
+        def mock_cache_audio(data, ext):
+            nonlocal cached_path
+            cached_path = f"/tmp/test_audio{ext}"
+            return cached_path
+
+        monkeypatch.setattr(
+            "gateway.platforms.bluebubbles.cache_audio_from_bytes",
+            mock_cache_audio,
+        )
+
+        att_meta = {"mimeType": "audio/mpeg", "transferName": "voice.mp3"}
+        result = asyncio.get_event_loop().run_until_complete(
+            adapter._download_attachment("att-guid-456", att_meta)
+        )
+        assert result == "/tmp/test_audio.mp3"
+
+    def test_download_document_uses_document_cache(self, monkeypatch):
+        """Non-image/audio MIME routes to cache_document_from_bytes."""
+        adapter = _make_adapter(monkeypatch)
+        import asyncio
+
+        class MockResponse:
+            status_code = 200
+            content = b"fake-doc-data"
+
+            def raise_for_status(self):
+                pass
+
+        async def mock_get(*args, **kwargs):
+            return MockResponse()
+
+        adapter.client = type("MockClient", (), {"get": mock_get})()
+
+        cached_path = None
+
+        def mock_cache_doc(data, filename):
+            nonlocal cached_path
+            cached_path = f"/tmp/{filename}"
+            return cached_path
+
+        monkeypatch.setattr(
+            "gateway.platforms.bluebubbles.cache_document_from_bytes",
+            mock_cache_doc,
+        )
+
+        att_meta = {"mimeType": "application/pdf", "transferName": "report.pdf"}
+        result = asyncio.get_event_loop().run_until_complete(
+            adapter._download_attachment("att-guid-789", att_meta)
+        )
+        assert result == "/tmp/report.pdf"
+
+    def test_download_returns_none_without_client(self, monkeypatch):
+        """No client → returns None gracefully."""
+        adapter = _make_adapter(monkeypatch)
+        adapter.client = None
+        import asyncio
+
+        result = asyncio.get_event_loop().run_until_complete(
+            adapter._download_attachment("att-guid", {"mimeType": "image/png"})
+        )
+        assert result is None

tests/gateway/test_discord_document_handling.py

@@ -209,14 +209,31 @@ class TestIncomingDocumentHandling:
         assert "[Content of readme.md]:" in event.text
         assert "# Title" in event.text
 
+    @pytest.mark.asyncio
+    async def test_log_content_injected(self, adapter):
+        """.log file under 100KB should be treated as text/plain and injected."""
+        file_content = b"BLE trace line 1\nBLE trace line 2"
+
+        with _mock_aiohttp_download(file_content):
+            msg = make_message(
+                attachments=[make_attachment(filename="btsnoop_hci.log", content_type="text/plain")],
+                content="please inspect this",
+            )
+            await adapter._handle_message(msg)
+
+        event = adapter.handle_message.call_args[0][0]
+        assert "[Content of btsnoop_hci.log]:" in event.text
+        assert "BLE trace line 1" in event.text
+        assert "please inspect this" in event.text
+
     @pytest.mark.asyncio
     async def test_oversized_document_skipped(self, adapter):
-        """A document over 20MB should be skipped — media_urls stays empty."""
+        """A document over 32MB should be skipped — media_urls stays empty."""
         msg = make_message([
             make_attachment(
                 filename="huge.pdf",
                 content_type="application/pdf",
-                size=25 * 1024 * 1024,
+                size=33 * 1024 * 1024,
             )
         ])
         await adapter._handle_message(msg)
@@ -226,6 +243,24 @@ class TestIncomingDocumentHandling:
         # handler must still be called
         adapter.handle_message.assert_called_once()
 
+    @pytest.mark.asyncio
+    async def test_mid_sized_zip_under_32mb_is_cached(self, adapter):
+        """A 25MB .zip should be accepted now that Discord documents allow up to 32MB."""
+        msg = make_message([
+            make_attachment(
+                filename="bugreport.zip",
+                content_type="application/zip",
+                size=25 * 1024 * 1024,
+            )
+        ])
+
+        with _mock_aiohttp_download(b"PK\x03\x04test"):
+            await adapter._handle_message(msg)
+
+        event = adapter.handle_message.call_args[0][0]
+        assert len(event.media_urls) == 1
+        assert event.media_types == ["application/zip"]
+
     @pytest.mark.asyncio
     async def test_zip_document_cached(self, adapter):
         """A .zip file should be cached as a supported document."""

tests/gateway/test_discord_reply_mode.py

@@ -0,0 +1,277 @@
+"""Tests for Discord reply_to_mode functionality.
+
+Covers the threading behavior control for multi-chunk replies:
+- "off": Never reply-reference to original message
+- "first": Only first chunk uses reply reference (default)
+- "all": All chunks reply-reference the original message
+"""
+import os
+import sys
+from types import SimpleNamespace
+from unittest.mock import MagicMock, AsyncMock, patch
+
+import pytest
+
+from gateway.config import PlatformConfig, GatewayConfig, Platform, _apply_env_overrides
+
+
+def _ensure_discord_mock():
+    """Install a mock discord module when discord.py isn't available."""
+    if "discord" in sys.modules and hasattr(sys.modules["discord"], "__file__"):
+        return
+
+    discord_mod = MagicMock()
+    discord_mod.Intents.default.return_value = MagicMock()
+    discord_mod.Client = MagicMock
+    discord_mod.File = MagicMock
+    discord_mod.DMChannel = type("DMChannel", (), {})
+    discord_mod.Thread = type("Thread", (), {})
+    discord_mod.ForumChannel = type("ForumChannel", (), {})
+    discord_mod.ui = SimpleNamespace(View=object, button=lambda *a, **k: (lambda fn: fn), Button=object)
+    discord_mod.ButtonStyle = SimpleNamespace(success=1, primary=2, secondary=2, danger=3, green=1, grey=2, blurple=2, red=3)
+    discord_mod.Color = SimpleNamespace(orange=lambda: 1, green=lambda: 2, blue=lambda: 3, red=lambda: 4, purple=lambda: 5)
+    discord_mod.Interaction = object
+    discord_mod.Embed = MagicMock
+    discord_mod.app_commands = SimpleNamespace(
+        describe=lambda **kwargs: (lambda fn: fn),
+        choices=lambda **kwargs: (lambda fn: fn),
+        Choice=lambda **kwargs: SimpleNamespace(**kwargs),
+    )
+
+    ext_mod = MagicMock()
+    commands_mod = MagicMock()
+    commands_mod.Bot = MagicMock
+    ext_mod.commands = commands_mod
+
+    sys.modules.setdefault("discord", discord_mod)
+    sys.modules.setdefault("discord.ext", ext_mod)
+    sys.modules.setdefault("discord.ext.commands", commands_mod)
+
+
+_ensure_discord_mock()
+
+from gateway.platforms.discord import DiscordAdapter  # noqa: E402
+
+
+@pytest.fixture()
+def adapter_factory():
+    """Factory to create DiscordAdapter with custom reply_to_mode."""
+    def create(reply_to_mode: str = "first"):
+        config = PlatformConfig(enabled=True, token="test-token", reply_to_mode=reply_to_mode)
+        return DiscordAdapter(config)
+    return create
+
+
+class TestReplyToModeConfig:
+    """Tests for reply_to_mode configuration loading."""
+
+    def test_default_mode_is_first(self, adapter_factory):
+        adapter = adapter_factory()
+        assert adapter._reply_to_mode == "first"
+
+    def test_off_mode(self, adapter_factory):
+        adapter = adapter_factory(reply_to_mode="off")
+        assert adapter._reply_to_mode == "off"
+
+    def test_first_mode(self, adapter_factory):
+        adapter = adapter_factory(reply_to_mode="first")
+        assert adapter._reply_to_mode == "first"
+
+    def test_all_mode(self, adapter_factory):
+        adapter = adapter_factory(reply_to_mode="all")
+        assert adapter._reply_to_mode == "all"
+
+    def test_invalid_mode_stored_as_is(self, adapter_factory):
+        """Invalid modes are stored but send() handles them gracefully."""
+        adapter = adapter_factory(reply_to_mode="invalid")
+        assert adapter._reply_to_mode == "invalid"
+
+    def test_none_mode_defaults_to_first(self):
+        config = PlatformConfig(enabled=True, token="test-token")
+        adapter = DiscordAdapter(config)
+        assert adapter._reply_to_mode == "first"
+
+    def test_empty_string_mode_defaults_to_first(self):
+        config = PlatformConfig(enabled=True, token="test-token", reply_to_mode="")
+        adapter = DiscordAdapter(config)
+        assert adapter._reply_to_mode == "first"
+
+
+def _make_discord_adapter(reply_to_mode: str = "first"):
+    """Create a DiscordAdapter with mocked client and channel for send() tests."""
+    config = PlatformConfig(enabled=True, token="test-token", reply_to_mode=reply_to_mode)
+    adapter = DiscordAdapter(config)
+
+    # Mock the Discord client and channel
+    mock_channel = AsyncMock()
+    ref_message = MagicMock()
+    mock_channel.fetch_message = AsyncMock(return_value=ref_message)
+
+    sent_msg = MagicMock()
+    sent_msg.id = 42
+    mock_channel.send = AsyncMock(return_value=sent_msg)
+
+    mock_client = MagicMock()
+    mock_client.get_channel = MagicMock(return_value=mock_channel)
+
+    adapter._client = mock_client
+    return adapter, mock_channel, ref_message
+
+
+class TestSendWithReplyToMode:
+    """Tests for send() method respecting reply_to_mode."""
+
+    @pytest.mark.asyncio
+    async def test_off_mode_no_reply_reference(self):
+        adapter, channel, ref_msg = _make_discord_adapter("off")
+        adapter.truncate_message = lambda content, max_len: ["chunk1", "chunk2", "chunk3"]
+
+        await adapter.send("12345", "test content", reply_to="999")
+
+        # Should never try to fetch the reference message
+        channel.fetch_message.assert_not_called()
+        # All chunks sent without reference
+        for call in channel.send.call_args_list:
+            assert call.kwargs.get("reference") is None
+
+    @pytest.mark.asyncio
+    async def test_first_mode_only_first_chunk_references(self):
+        adapter, channel, ref_msg = _make_discord_adapter("first")
+        adapter.truncate_message = lambda content, max_len: ["chunk1", "chunk2", "chunk3"]
+
+        await adapter.send("12345", "test content", reply_to="999")
+
+        # Should fetch the reference message
+        channel.fetch_message.assert_called_once_with(999)
+        calls = channel.send.call_args_list
+        assert len(calls) == 3
+        assert calls[0].kwargs.get("reference") is ref_msg
+        assert calls[1].kwargs.get("reference") is None
+        assert calls[2].kwargs.get("reference") is None
+
+    @pytest.mark.asyncio
+    async def test_all_mode_all_chunks_reference(self):
+        adapter, channel, ref_msg = _make_discord_adapter("all")
+        adapter.truncate_message = lambda content, max_len: ["chunk1", "chunk2", "chunk3"]
+
+        await adapter.send("12345", "test content", reply_to="999")
+
+        channel.fetch_message.assert_called_once_with(999)
+        calls = channel.send.call_args_list
+        assert len(calls) == 3
+        for call in calls:
+            assert call.kwargs.get("reference") is ref_msg
+
+    @pytest.mark.asyncio
+    async def test_no_reply_to_param_no_reference(self):
+        adapter, channel, ref_msg = _make_discord_adapter("all")
+        adapter.truncate_message = lambda content, max_len: ["chunk1", "chunk2"]
+
+        await adapter.send("12345", "test content", reply_to=None)
+
+        channel.fetch_message.assert_not_called()
+        for call in channel.send.call_args_list:
+            assert call.kwargs.get("reference") is None
+
+    @pytest.mark.asyncio
+    async def test_single_chunk_respects_first_mode(self):
+        adapter, channel, ref_msg = _make_discord_adapter("first")
+        adapter.truncate_message = lambda content, max_len: ["single chunk"]
+
+        await adapter.send("12345", "test", reply_to="999")
+
+        calls = channel.send.call_args_list
+        assert len(calls) == 1
+        assert calls[0].kwargs.get("reference") is ref_msg
+
+    @pytest.mark.asyncio
+    async def test_single_chunk_off_mode(self):
+        adapter, channel, ref_msg = _make_discord_adapter("off")
+        adapter.truncate_message = lambda content, max_len: ["single chunk"]
+
+        await adapter.send("12345", "test", reply_to="999")
+
+        channel.fetch_message.assert_not_called()
+        calls = channel.send.call_args_list
+        assert len(calls) == 1
+        assert calls[0].kwargs.get("reference") is None
+
+    @pytest.mark.asyncio
+    async def test_invalid_mode_falls_back_to_first_behavior(self):
+        """Invalid mode behaves like 'first' — only first chunk gets reference."""
+        adapter, channel, ref_msg = _make_discord_adapter("banana")
+        adapter.truncate_message = lambda content, max_len: ["chunk1", "chunk2"]
+
+        await adapter.send("12345", "test", reply_to="999")
+
+        calls = channel.send.call_args_list
+        assert len(calls) == 2
+        assert calls[0].kwargs.get("reference") is ref_msg
+        assert calls[1].kwargs.get("reference") is None
+
+
+class TestConfigSerialization:
+    """Tests for reply_to_mode serialization (shared with Telegram)."""
+
+    def test_to_dict_includes_reply_to_mode(self):
+        config = PlatformConfig(enabled=True, token="test", reply_to_mode="all")
+        result = config.to_dict()
+        assert result["reply_to_mode"] == "all"
+
+    def test_from_dict_loads_reply_to_mode(self):
+        data = {"enabled": True, "token": "***", "reply_to_mode": "off"}
+        config = PlatformConfig.from_dict(data)
+        assert config.reply_to_mode == "off"
+
+    def test_from_dict_defaults_to_first(self):
+        data = {"enabled": True, "token": "***"}
+        config = PlatformConfig.from_dict(data)
+        assert config.reply_to_mode == "first"
+
+
+class TestEnvVarOverride:
+    """Tests for DISCORD_REPLY_TO_MODE environment variable override."""
+
+    def _make_config(self):
+        config = GatewayConfig()
+        config.platforms[Platform.DISCORD] = PlatformConfig(enabled=True, token="test")
+        return config
+
+    def test_env_var_sets_off_mode(self):
+        config = self._make_config()
+        with patch.dict(os.environ, {"DISCORD_REPLY_TO_MODE": "off"}, clear=False):
+            _apply_env_overrides(config)
+        assert config.platforms[Platform.DISCORD].reply_to_mode == "off"
+
+    def test_env_var_sets_all_mode(self):
+        config = self._make_config()
+        with patch.dict(os.environ, {"DISCORD_REPLY_TO_MODE": "all"}, clear=False):
+            _apply_env_overrides(config)
+        assert config.platforms[Platform.DISCORD].reply_to_mode == "all"
+
+    def test_env_var_case_insensitive(self):
+        config = self._make_config()
+        with patch.dict(os.environ, {"DISCORD_REPLY_TO_MODE": "ALL"}, clear=False):
+            _apply_env_overrides(config)
+        assert config.platforms[Platform.DISCORD].reply_to_mode == "all"
+
+    def test_env_var_invalid_value_ignored(self):
+        config = self._make_config()
+        with patch.dict(os.environ, {"DISCORD_REPLY_TO_MODE": "banana"}, clear=False):
+            _apply_env_overrides(config)
+        assert config.platforms[Platform.DISCORD].reply_to_mode == "first"
+
+    def test_env_var_empty_value_ignored(self):
+        config = self._make_config()
+        with patch.dict(os.environ, {"DISCORD_REPLY_TO_MODE": ""}, clear=False):
+            _apply_env_overrides(config)
+        assert config.platforms[Platform.DISCORD].reply_to_mode == "first"
+
+    def test_env_var_creates_platform_config_if_missing(self):
+        """DISCORD_REPLY_TO_MODE creates PlatformConfig even without DISCORD_BOT_TOKEN."""
+        config = GatewayConfig()
+        assert Platform.DISCORD not in config.platforms
+        with patch.dict(os.environ, {"DISCORD_REPLY_TO_MODE": "off"}, clear=False):
+            _apply_env_overrides(config)
+        assert Platform.DISCORD in config.platforms
+        assert config.platforms[Platform.DISCORD].reply_to_mode == "off"

tests/gateway/test_gateway_inactivity_timeout.py

@@ -0,0 +1,315 @@
+"""Tests for staged inactivity timeout in gateway agent runs.
+
+Tests cover:
+- Warning fires once when inactivity reaches gateway_timeout_warning threshold
+- Warning does not fire when gateway_timeout is 0 (unlimited)
+- Warning fires only once per run, not on every poll
+- Full timeout still fires at gateway_timeout threshold
+- Warning respects HERMES_AGENT_TIMEOUT_WARNING env var
+- Warning disabled when gateway_timeout_warning is 0
+"""
+
+import concurrent.futures
+import os
+import sys
+import time
+from pathlib import Path
+from unittest.mock import MagicMock, patch
+
+import pytest
+
+sys.path.insert(0, str(Path(__file__).parent.parent.parent))
+
+
+class FakeAgent:
+    """Mock agent with controllable activity summary for timeout tests."""
+
+    def __init__(self, idle_seconds=0.0, activity_desc="tool_call",
+                 current_tool=None, api_call_count=5, max_iterations=90):
+        self._idle_seconds = idle_seconds
+        self._activity_desc = activity_desc
+        self._current_tool = current_tool
+        self._api_call_count = api_call_count
+        self._max_iterations = max_iterations
+        self._interrupted = False
+        self._interrupt_msg = None
+
+    def get_activity_summary(self):
+        return {
+            "last_activity_ts": time.time() - self._idle_seconds,
+            "last_activity_desc": self._activity_desc,
+            "seconds_since_activity": self._idle_seconds,
+            "current_tool": self._current_tool,
+            "api_call_count": self._api_call_count,
+            "max_iterations": self._max_iterations,
+        }
+
+    def interrupt(self, msg):
+        self._interrupted = True
+        self._interrupt_msg = msg
+
+    def run_conversation(self, prompt):
+        return {"final_response": "Done", "messages": []}
+
+
+class SlowFakeAgent(FakeAgent):
+    """Agent that runs for a while, then goes idle."""
+
+    def __init__(self, run_duration=0.5, idle_after=None, **kwargs):
+        super().__init__(**kwargs)
+        self._run_duration = run_duration
+        self._idle_after = idle_after
+        self._start_time = None
+
+    def get_activity_summary(self):
+        summary = super().get_activity_summary()
+        if self._idle_after is not None and self._start_time:
+            elapsed = time.time() - self._start_time
+            if elapsed > self._idle_after:
+                idle_time = elapsed - self._idle_after
+                summary["seconds_since_activity"] = idle_time
+                summary["last_activity_desc"] = "api_call_streaming"
+            else:
+                summary["seconds_since_activity"] = 0.0
+        return summary
+
+    def run_conversation(self, prompt):
+        self._start_time = time.time()
+        time.sleep(self._run_duration)
+        return {"final_response": "Completed after work", "messages": []}
+
+
+class TestStagedInactivityWarning:
+    """Test the staged inactivity warning before full timeout."""
+
+    def test_warning_fires_once_before_timeout(self):
+        """Warning fires when inactivity reaches warning threshold."""
+        agent = SlowFakeAgent(
+            run_duration=10.0,
+            idle_after=0.1,
+            activity_desc="api_call_streaming",
+        )
+
+        _agent_timeout = 20.0
+        _agent_warning = 5.0
+        _POLL_INTERVAL = 0.1
+
+        pool = concurrent.futures.ThreadPoolExecutor(max_workers=1)
+        future = pool.submit(agent.run_conversation, "test prompt")
+        _inactivity_timeout = False
+        _warning_fired = False
+        _warning_send_count = 0
+
+        while True:
+            done, _ = concurrent.futures.wait({future}, timeout=_POLL_INTERVAL)
+            if done:
+                result = future.result()
+                break
+            _idle_secs = 0.0
+            if hasattr(agent, "get_activity_summary"):
+                try:
+                    _act = agent.get_activity_summary()
+                    _idle_secs = _act.get("seconds_since_activity", 0.0)
+                except Exception:
+                    pass
+            if (not _warning_fired and _agent_warning > 0
+                    and _idle_secs >= _agent_warning):
+                _warning_fired = True
+                _warning_send_count += 1
+            if _idle_secs >= _agent_timeout:
+                _inactivity_timeout = True
+                break
+
+        pool.shutdown(wait=False, cancel_futures=True)
+
+        assert _warning_fired
+        assert _warning_send_count == 1
+        assert not _inactivity_timeout
+
+    def test_warning_disabled_when_zero(self):
+        """No warning fires when gateway_timeout_warning is 0."""
+        agent = SlowFakeAgent(
+            run_duration=5.0,
+            idle_after=0.1,
+        )
+
+        _agent_timeout = 20.0
+        _agent_warning = 0.0
+        _POLL_INTERVAL = 0.1
+
+        pool = concurrent.futures.ThreadPoolExecutor(max_workers=1)
+        future = pool.submit(agent.run_conversation, "test")
+        _warning_fired = False
+
+        while True:
+            done, _ = concurrent.futures.wait({future}, timeout=_POLL_INTERVAL)
+            if done:
+                future.result()
+                break
+            _idle_secs = 0.0
+            if hasattr(agent, "get_activity_summary"):
+                try:
+                    _act = agent.get_activity_summary()
+                    _idle_secs = _act.get("seconds_since_activity", 0.0)
+                except Exception:
+                    pass
+            if (not _warning_fired and _agent_warning > 0
+                    and _idle_secs >= _agent_warning):
+                _warning_fired = True
+            if _idle_secs >= _agent_timeout:
+                break
+
+        pool.shutdown(wait=False, cancel_futures=True)
+        assert not _warning_fired
+
+    def test_warning_fires_only_once(self):
+        """Warning fires exactly once even if agent remains idle."""
+        agent = SlowFakeAgent(
+            run_duration=10.0,
+            idle_after=0.05,
+        )
+
+        _agent_timeout = 20.0
+        _agent_warning = 0.2
+        _POLL_INTERVAL = 0.05
+
+        pool = concurrent.futures.ThreadPoolExecutor(max_workers=1)
+        future = pool.submit(agent.run_conversation, "test")
+        _warning_count = 0
+
+        while True:
+            done, _ = concurrent.futures.wait({future}, timeout=_POLL_INTERVAL)
+            if done:
+                future.result()
+                break
+            _idle_secs = 0.0
+            if hasattr(agent, "get_activity_summary"):
+                try:
+                    _act = agent.get_activity_summary()
+                    _idle_secs = _act.get("seconds_since_activity", 0.0)
+                except Exception:
+                    pass
+            if (not _warning_count and _agent_warning > 0
+                    and _idle_secs >= _agent_warning):
+                _warning_count += 1
+            if _idle_secs >= _agent_timeout:
+                break
+
+        pool.shutdown(wait=False, cancel_futures=True)
+        assert _warning_count == 1
+
+    def test_full_timeout_still_fires_after_warning(self):
+        """Full timeout fires even after warning was sent."""
+        agent = SlowFakeAgent(
+            run_duration=15.0,
+            idle_after=0.1,
+            activity_desc="waiting for provider response (streaming)",
+        )
+
+        _agent_timeout = 1.0
+        _agent_warning = 0.3
+        _POLL_INTERVAL = 0.05
+
+        pool = concurrent.futures.ThreadPoolExecutor(max_workers=1)
+        future = pool.submit(agent.run_conversation, "test")
+        _inactivity_timeout = False
+        _warning_fired = False
+
+        while True:
+            done, _ = concurrent.futures.wait({future}, timeout=_POLL_INTERVAL)
+            if done:
+                future.result()
+                break
+            _idle_secs = 0.0
+            if hasattr(agent, "get_activity_summary"):
+                try:
+                    _act = agent.get_activity_summary()
+                    _idle_secs = _act.get("seconds_since_activity", 0.0)
+                except Exception:
+                    pass
+            if (not _warning_fired and _agent_warning > 0
+                    and _idle_secs >= _agent_warning):
+                _warning_fired = True
+            if _idle_secs >= _agent_timeout:
+                _inactivity_timeout = True
+                break
+
+        pool.shutdown(wait=False, cancel_futures=True)
+        assert _warning_fired
+        assert _inactivity_timeout
+
+    def test_warning_env_var_respected(self, monkeypatch):
+        """HERMES_AGENT_TIMEOUT_WARNING env var is parsed correctly."""
+        monkeypatch.setenv("HERMES_AGENT_TIMEOUT_WARNING", "600")
+        _warning = float(os.getenv("HERMES_AGENT_TIMEOUT_WARNING", 900))
+        assert _warning == 600.0
+
+    def test_warning_zero_means_disabled(self, monkeypatch):
+        """HERMES_AGENT_TIMEOUT_WARNING=0 disables the warning."""
+        monkeypatch.setenv("HERMES_AGENT_TIMEOUT_WARNING", "0")
+        _raw = float(os.getenv("HERMES_AGENT_TIMEOUT_WARNING", 900))
+        _warning = _raw if _raw > 0 else None
+        assert _warning is None
+
+    def test_unlimited_timeout_no_warning(self):
+        """When timeout is unlimited (0), no warning fires either."""
+        agent = SlowFakeAgent(
+            run_duration=0.5,
+            idle_after=0.0,
+        )
+
+        _agent_timeout = None
+        _agent_warning = 5.0
+        _POLL_INTERVAL = 0.05
+
+        pool = concurrent.futures.ThreadPoolExecutor(max_workers=1)
+        future = pool.submit(agent.run_conversation, "test")
+
+        result = future.result(timeout=2.0)
+        pool.shutdown(wait=False)
+
+        assert result["final_response"] == "Completed after work"
+
+
+class TestWarningThresholdBelowTimeout:
+    """Test that warning threshold must be less than timeout threshold."""
+
+    def test_warning_at_half_timeout(self):
+        """Warning fires at half the timeout duration."""
+        agent = SlowFakeAgent(
+            run_duration=10.0,
+            idle_after=0.1,
+            activity_desc="receiving stream response",
+        )
+
+        _agent_timeout = 2.0
+        _agent_warning = 1.0
+        _POLL_INTERVAL = 0.05
+
+        pool = concurrent.futures.ThreadPoolExecutor(max_workers=1)
+        future = pool.submit(agent.run_conversation, "test")
+        _warning_fired = False
+        _timeout_fired = False
+
+        while True:
+            done, _ = concurrent.futures.wait({future}, timeout=_POLL_INTERVAL)
+            if done:
+                future.result()
+                break
+            _idle_secs = 0.0
+            if hasattr(agent, "get_activity_summary"):
+                try:
+                    _act = agent.get_activity_summary()
+                    _idle_secs = _act.get("seconds_since_activity", 0.0)
+                except Exception:
+                    pass
+            if (not _warning_fired and _agent_warning > 0
+                    and _idle_secs >= _agent_warning):
+                _warning_fired = True
+            if _idle_secs >= _agent_timeout:
+                _timeout_fired = True
+                break
+
+        pool.shutdown(wait=False, cancel_futures=True)
+        assert _warning_fired
+        assert _timeout_fired

tests/gateway/test_internal_event_bypass_pairing.py

@@ -0,0 +1,226 @@
+"""Tests that internal synthetic events (e.g. background process completion)
+bypass user authorization and do not trigger DM pairing.
+
+Regression test for the bug where ``_run_process_watcher`` with
+``notify_on_complete=True`` injected a ``MessageEvent`` without ``user_id``,
+causing ``_is_user_authorized`` to reject it and the gateway to send a
+pairing code to the chat.
+"""
+
+import asyncio
+from types import SimpleNamespace
+from unittest.mock import AsyncMock, patch
+
+import pytest
+
+from gateway.config import GatewayConfig, Platform
+from gateway.platforms.base import MessageEvent
+from gateway.run import GatewayRunner
+from gateway.session import SessionSource
+
+
+# ---------------------------------------------------------------------------
+# Helpers
+# ---------------------------------------------------------------------------
+
+class _FakeRegistry:
+    """Return pre-canned sessions, then None once exhausted."""
+
+    def __init__(self, sessions):
+        self._sessions = list(sessions)
+
+    def get(self, session_id):
+        if self._sessions:
+            return self._sessions.pop(0)
+        return None
+
+
+def _build_runner(monkeypatch, tmp_path) -> GatewayRunner:
+    """Create a GatewayRunner with notifications set to 'all'."""
+    (tmp_path / "config.yaml").write_text(
+        "display:\n  background_process_notifications: all\n",
+        encoding="utf-8",
+    )
+
+    import gateway.run as gateway_run
+
+    monkeypatch.setattr(gateway_run, "_hermes_home", tmp_path)
+
+    runner = GatewayRunner(GatewayConfig())
+    adapter = SimpleNamespace(send=AsyncMock(), handle_message=AsyncMock())
+    runner.adapters[Platform.DISCORD] = adapter
+    return runner
+
+
+def _watcher_dict_with_notify():
+    return {
+        "session_id": "proc_test_internal",
+        "check_interval": 0,
+        "session_key": "agent:main:discord:dm:123",
+        "platform": "discord",
+        "chat_id": "123",
+        "thread_id": "",
+        "notify_on_complete": True,
+    }
+
+
+# ---------------------------------------------------------------------------
+# Tests
+# ---------------------------------------------------------------------------
+
+@pytest.mark.asyncio
+async def test_notify_on_complete_sets_internal_flag(monkeypatch, tmp_path):
+    """Synthetic completion event must have internal=True."""
+    import tools.process_registry as pr_module
+
+    sessions = [
+        SimpleNamespace(
+            output_buffer="done\n", exited=True, exit_code=0, command="echo test"
+        ),
+    ]
+    monkeypatch.setattr(pr_module, "process_registry", _FakeRegistry(sessions))
+
+    async def _instant_sleep(*_a, **_kw):
+        pass
+    monkeypatch.setattr(asyncio, "sleep", _instant_sleep)
+
+    runner = _build_runner(monkeypatch, tmp_path)
+    adapter = runner.adapters[Platform.DISCORD]
+
+    await runner._run_process_watcher(_watcher_dict_with_notify())
+
+    assert adapter.handle_message.await_count == 1
+    event = adapter.handle_message.await_args.args[0]
+    assert isinstance(event, MessageEvent)
+    assert event.internal is True, "Synthetic completion event must be marked internal"
+
+
+@pytest.mark.asyncio
+async def test_internal_event_bypasses_authorization(monkeypatch, tmp_path):
+    """An internal event should skip _is_user_authorized entirely."""
+    import gateway.run as gateway_run
+
+    monkeypatch.setattr(gateway_run, "_hermes_home", tmp_path)
+    (tmp_path / "config.yaml").write_text("", encoding="utf-8")
+
+    runner = GatewayRunner(GatewayConfig())
+
+    # Create an internal event with no user_id (simulates the bug scenario)
+    source = SessionSource(
+        platform=Platform.DISCORD,
+        chat_id="123",
+        chat_type="dm",
+    )
+    event = MessageEvent(
+        text="[SYSTEM: Background process completed]",
+        source=source,
+        internal=True,
+    )
+
+    # Track if _is_user_authorized is called
+    auth_called = False
+    original_auth = GatewayRunner._is_user_authorized
+
+    def tracking_auth(self, src):
+        nonlocal auth_called
+        auth_called = True
+        return original_auth(self, src)
+
+    monkeypatch.setattr(GatewayRunner, "_is_user_authorized", tracking_auth)
+
+    # _handle_message will proceed past auth check and eventually fail on
+    # downstream logic. We just need to verify auth is skipped.
+    try:
+        await runner._handle_message(event)
+    except Exception:
+        pass  # Expected — downstream code needs more setup
+
+    assert not auth_called, (
+        "_is_user_authorized should NOT be called for internal events"
+    )
+
+
+@pytest.mark.asyncio
+async def test_internal_event_does_not_trigger_pairing(monkeypatch, tmp_path):
+    """An internal event with no user_id must not generate a pairing code."""
+    import gateway.run as gateway_run
+
+    monkeypatch.setattr(gateway_run, "_hermes_home", tmp_path)
+    (tmp_path / "config.yaml").write_text("", encoding="utf-8")
+
+    runner = GatewayRunner(GatewayConfig())
+    # Add adapter so pairing would have somewhere to send
+    adapter = SimpleNamespace(send=AsyncMock())
+    runner.adapters[Platform.DISCORD] = adapter
+
+    source = SessionSource(
+        platform=Platform.DISCORD,
+        chat_id="123",
+        chat_type="dm",  # DM would normally trigger pairing
+    )
+    event = MessageEvent(
+        text="[SYSTEM: Background process completed]",
+        source=source,
+        internal=True,
+    )
+
+    # Track pairing code generation
+    generate_called = False
+    original_generate = runner.pairing_store.generate_code
+
+    def tracking_generate(*args, **kwargs):
+        nonlocal generate_called
+        generate_called = True
+        return original_generate(*args, **kwargs)
+
+    runner.pairing_store.generate_code = tracking_generate
+
+    try:
+        await runner._handle_message(event)
+    except Exception:
+        pass  # Expected — downstream code needs more setup
+
+    assert not generate_called, (
+        "Pairing code should NOT be generated for internal events"
+    )
+
+
+@pytest.mark.asyncio
+async def test_non_internal_event_without_user_triggers_pairing(monkeypatch, tmp_path):
+    """Verify the normal (non-internal) path still triggers pairing for unknown users."""
+    import gateway.run as gateway_run
+
+    monkeypatch.setattr(gateway_run, "_hermes_home", tmp_path)
+    (tmp_path / "config.yaml").write_text("", encoding="utf-8")
+
+    # Clear env vars that could let all users through (loaded by
+    # module-level dotenv in gateway/run.py from the real ~/.hermes/.env).
+    monkeypatch.delenv("DISCORD_ALLOW_ALL_USERS", raising=False)
+    monkeypatch.delenv("DISCORD_ALLOWED_USERS", raising=False)
+    monkeypatch.delenv("GATEWAY_ALLOW_ALL_USERS", raising=False)
+    monkeypatch.delenv("GATEWAY_ALLOWED_USERS", raising=False)
+
+    runner = GatewayRunner(GatewayConfig())
+    adapter = SimpleNamespace(send=AsyncMock())
+    runner.adapters[Platform.DISCORD] = adapter
+
+    source = SessionSource(
+        platform=Platform.DISCORD,
+        chat_id="123",
+        chat_type="dm",
+        user_id="unknown_user_999",
+    )
+    # Normal event (not internal)
+    event = MessageEvent(
+        text="hello",
+        source=source,
+        internal=False,
+    )
+
+    result = await runner._handle_message(event)
+
+    # Should return None (unauthorized) and send pairing message
+    assert result is None
+    assert adapter.send.await_count == 1
+    sent_text = adapter.send.await_args.args[1]
+    assert "don't recognize you" in sent_text

tests/gateway/test_media_download_retry.py

@@ -38,10 +38,11 @@ def _make_timeout_error() -> httpx.TimeoutException:
 # cache_image_from_url (base.py)
 # ---------------------------------------------------------------------------
 
+@patch("tools.url_safety.is_safe_url", return_value=True)
 class TestCacheImageFromUrl:
     """Tests for gateway.platforms.base.cache_image_from_url"""
 
-    def test_success_on_first_attempt(self, tmp_path, monkeypatch):
+    def test_success_on_first_attempt(self, _mock_safe, tmp_path, monkeypatch):
         """A clean 200 response caches the image and returns a path."""
         monkeypatch.setattr("gateway.platforms.base.IMAGE_CACHE_DIR", tmp_path / "img")
 
@@ -65,7 +66,7 @@ class TestCacheImageFromUrl:
         assert path.endswith(".jpg")
         mock_client.get.assert_called_once()
 
-    def test_retries_on_timeout_then_succeeds(self, tmp_path, monkeypatch):
+    def test_retries_on_timeout_then_succeeds(self, _mock_safe, tmp_path, monkeypatch):
         """A timeout on the first attempt is retried; second attempt succeeds."""
         monkeypatch.setattr("gateway.platforms.base.IMAGE_CACHE_DIR", tmp_path / "img")
 
@@ -95,7 +96,7 @@ class TestCacheImageFromUrl:
         assert mock_client.get.call_count == 2
         mock_sleep.assert_called_once()
 
-    def test_retries_on_429_then_succeeds(self, tmp_path, monkeypatch):
+    def test_retries_on_429_then_succeeds(self, _mock_safe, tmp_path, monkeypatch):
         """A 429 response on the first attempt is retried; second attempt succeeds."""
         monkeypatch.setattr("gateway.platforms.base.IMAGE_CACHE_DIR", tmp_path / "img")
 
@@ -122,7 +123,7 @@ class TestCacheImageFromUrl:
         assert path.endswith(".jpg")
         assert mock_client.get.call_count == 2
 
-    def test_raises_after_max_retries_exhausted(self, tmp_path, monkeypatch):
+    def test_raises_after_max_retries_exhausted(self, _mock_safe, tmp_path, monkeypatch):
         """Timeout on every attempt raises after all retries are consumed."""
         monkeypatch.setattr("gateway.platforms.base.IMAGE_CACHE_DIR", tmp_path / "img")
 
@@ -145,7 +146,7 @@ class TestCacheImageFromUrl:
         # 3 total calls: initial + 2 retries
         assert mock_client.get.call_count == 3
 
-    def test_non_retryable_4xx_raises_immediately(self, tmp_path, monkeypatch):
+    def test_non_retryable_4xx_raises_immediately(self, _mock_safe, tmp_path, monkeypatch):
         """A 404 (non-retryable) is raised immediately without any retry."""
         monkeypatch.setattr("gateway.platforms.base.IMAGE_CACHE_DIR", tmp_path / "img")
 
@@ -175,10 +176,11 @@ class TestCacheImageFromUrl:
 # cache_audio_from_url (base.py)
 # ---------------------------------------------------------------------------
 
+@patch("tools.url_safety.is_safe_url", return_value=True)
 class TestCacheAudioFromUrl:
     """Tests for gateway.platforms.base.cache_audio_from_url"""
 
-    def test_success_on_first_attempt(self, tmp_path, monkeypatch):
+    def test_success_on_first_attempt(self, _mock_safe, tmp_path, monkeypatch):
         """A clean 200 response caches the audio and returns a path."""
         monkeypatch.setattr("gateway.platforms.base.AUDIO_CACHE_DIR", tmp_path / "audio")
 
@@ -202,7 +204,7 @@ class TestCacheAudioFromUrl:
         assert path.endswith(".ogg")
         mock_client.get.assert_called_once()
 
-    def test_retries_on_timeout_then_succeeds(self, tmp_path, monkeypatch):
+    def test_retries_on_timeout_then_succeeds(self, _mock_safe, tmp_path, monkeypatch):
         """A timeout on the first attempt is retried; second attempt succeeds."""
         monkeypatch.setattr("gateway.platforms.base.AUDIO_CACHE_DIR", tmp_path / "audio")
 
@@ -232,7 +234,7 @@ class TestCacheAudioFromUrl:
         assert mock_client.get.call_count == 2
         mock_sleep.assert_called_once()
 
-    def test_retries_on_429_then_succeeds(self, tmp_path, monkeypatch):
+    def test_retries_on_429_then_succeeds(self, _mock_safe, tmp_path, monkeypatch):
         """A 429 response on the first attempt is retried; second attempt succeeds."""
         monkeypatch.setattr("gateway.platforms.base.AUDIO_CACHE_DIR", tmp_path / "audio")
 
@@ -259,7 +261,7 @@ class TestCacheAudioFromUrl:
         assert path.endswith(".ogg")
         assert mock_client.get.call_count == 2
 
-    def test_retries_on_500_then_succeeds(self, tmp_path, monkeypatch):
+    def test_retries_on_500_then_succeeds(self, _mock_safe, tmp_path, monkeypatch):
         """A 500 response on the first attempt is retried; second attempt succeeds."""
         monkeypatch.setattr("gateway.platforms.base.AUDIO_CACHE_DIR", tmp_path / "audio")
 
@@ -286,7 +288,7 @@ class TestCacheAudioFromUrl:
         assert path.endswith(".ogg")
         assert mock_client.get.call_count == 2
 
-    def test_raises_after_max_retries_exhausted(self, tmp_path, monkeypatch):
+    def test_raises_after_max_retries_exhausted(self, _mock_safe, tmp_path, monkeypatch):
         """Timeout on every attempt raises after all retries are consumed."""
         monkeypatch.setattr("gateway.platforms.base.AUDIO_CACHE_DIR", tmp_path / "audio")
 
@@ -309,7 +311,7 @@ class TestCacheAudioFromUrl:
         # 3 total calls: initial + 2 retries
         assert mock_client.get.call_count == 3
 
-    def test_non_retryable_4xx_raises_immediately(self, tmp_path, monkeypatch):
+    def test_non_retryable_4xx_raises_immediately(self, _mock_safe, tmp_path, monkeypatch):
         """A 404 (non-retryable) is raised immediately without any retry."""
         monkeypatch.setattr("gateway.platforms.base.AUDIO_CACHE_DIR", tmp_path / "audio")
 

tests/gateway/test_signal.py

@@ -707,3 +707,66 @@ class TestSignalSendDocumentViaHelper:
 
         assert result.success is False
         assert "/nonexistent.pdf" in result.error
+
+
+# ---------------------------------------------------------------------------
+# send() returns message_id from timestamp (#4647)
+# ---------------------------------------------------------------------------
+
+class TestSignalSendReturnsMessageId:
+    """Signal send() must return a timestamp-based message_id so the stream
+    consumer can follow its edit→fallback path correctly."""
+
+    @pytest.mark.asyncio
+    async def test_send_returns_timestamp_as_message_id(self, monkeypatch):
+        adapter = _make_signal_adapter(monkeypatch)
+        mock_rpc, _ = _stub_rpc({"timestamp": 1712345678000})
+        adapter._rpc = mock_rpc
+        adapter._stop_typing_indicator = AsyncMock()
+
+        result = await adapter.send(chat_id="+155****4567", content="hello")
+
+        assert result.success is True
+        assert result.message_id == "1712345678000"
+
+    @pytest.mark.asyncio
+    async def test_send_returns_none_message_id_when_no_timestamp(self, monkeypatch):
+        adapter = _make_signal_adapter(monkeypatch)
+        mock_rpc, _ = _stub_rpc({})  # No timestamp key
+        adapter._rpc = mock_rpc
+        adapter._stop_typing_indicator = AsyncMock()
+
+        result = await adapter.send(chat_id="+155****4567", content="hello")
+
+        assert result.success is True
+        assert result.message_id is None
+
+    @pytest.mark.asyncio
+    async def test_send_returns_none_message_id_for_non_dict(self, monkeypatch):
+        adapter = _make_signal_adapter(monkeypatch)
+        mock_rpc, _ = _stub_rpc("ok")  # Non-dict result
+        adapter._rpc = mock_rpc
+        adapter._stop_typing_indicator = AsyncMock()
+
+        result = await adapter.send(chat_id="+155****4567", content="hello")
+
+        assert result.success is True
+        assert result.message_id is None
+
+
+# ---------------------------------------------------------------------------
+# stop_typing() delegates to _stop_typing_indicator (#4647)
+# ---------------------------------------------------------------------------
+
+class TestSignalStopTyping:
+    """Signal must expose a public stop_typing() so base adapter's
+    _keep_typing finally block can clean up platform-level typing tasks."""
+
+    @pytest.mark.asyncio
+    async def test_stop_typing_calls_private_method(self, monkeypatch):
+        adapter = _make_signal_adapter(monkeypatch)
+        adapter._stop_typing_indicator = AsyncMock()
+
+        await adapter.stop_typing("+155****4567")
+
+        adapter._stop_typing_indicator.assert_awaited_once_with("+155****4567")

tests/gateway/test_slack.py

@@ -96,7 +96,7 @@ class TestAppMentionHandler:
     """Verify that the app_mention event handler is registered."""
 
     def test_app_mention_registered_on_connect(self):
-        """connect() should register both 'message' and 'app_mention' handlers."""
+        """connect() should register message + assistant lifecycle handlers."""
         config = PlatformConfig(enabled=True, token="xoxb-fake")
         adapter = SlackAdapter(config)
 
@@ -145,6 +145,8 @@ class TestAppMentionHandler:
 
         assert "message" in registered_events
         assert "app_mention" in registered_events
+        assert "assistant_thread_started" in registered_events
+        assert "assistant_thread_context_changed" in registered_events
         assert "/hermes" in registered_commands
 
 
@@ -840,6 +842,114 @@ class TestThreadReplyHandling:
         adapter.handle_message.assert_not_called()
 
 
+# ---------------------------------------------------------------------------
+# TestAssistantThreadLifecycle
+# ---------------------------------------------------------------------------
+
+
+class TestAssistantThreadLifecycle:
+    """Slack Assistant lifecycle events should seed session/user context."""
+
+    @pytest.fixture()
+    def mock_session_store(self):
+        store = MagicMock()
+        store._entries = {}
+        store._ensure_loaded = MagicMock()
+        store.config = MagicMock()
+        store.config.group_sessions_per_user = True
+        store.get_or_create_session = MagicMock()
+        return store
+
+    @pytest.fixture()
+    def assistant_adapter(self, mock_session_store):
+        config = PlatformConfig(enabled=True, token="***")
+        a = SlackAdapter(config)
+        a._app = MagicMock()
+        a._app.client = AsyncMock()
+        a._bot_user_id = "U_BOT"
+        a._team_bot_user_ids = {"T_TEAM": "U_BOT"}
+        a._running = True
+        a.handle_message = AsyncMock()
+        a.set_session_store(mock_session_store)
+        return a
+
+    @pytest.mark.asyncio
+    async def test_lifecycle_event_seeds_session_store(self, assistant_adapter, mock_session_store):
+        event = {
+            "type": "assistant_thread_started",
+            "team_id": "T_TEAM",
+            "assistant_thread": {
+                "channel_id": "D123",
+                "thread_ts": "171.000",
+                "user_id": "U_USER",
+                "context": {"channel_id": "C_ORIGIN"},
+            },
+        }
+
+        await assistant_adapter._handle_assistant_thread_lifecycle_event(event)
+
+        assert assistant_adapter._assistant_threads[("D123", "171.000")]["user_id"] == "U_USER"
+        mock_session_store.get_or_create_session.assert_called_once()
+        source = mock_session_store.get_or_create_session.call_args[0][0]
+        assert source.chat_id == "D123"
+        assert source.chat_type == "dm"
+        assert source.user_id == "U_USER"
+        assert source.thread_id == "171.000"
+        assert source.chat_topic == "C_ORIGIN"
+
+    @pytest.mark.asyncio
+    async def test_message_uses_cached_assistant_thread_identity(self, assistant_adapter):
+        assistant_adapter._assistant_threads[("D123", "171.000")] = {
+            "channel_id": "D123",
+            "thread_ts": "171.000",
+            "user_id": "U_USER",
+            "team_id": "T_TEAM",
+        }
+        assistant_adapter._app.client.users_info = AsyncMock(return_value={
+            "user": {"profile": {"display_name": "Tyler"}}
+        })
+        assistant_adapter._app.client.reactions_add = AsyncMock()
+        assistant_adapter._app.client.reactions_remove = AsyncMock()
+
+        event = {
+            "text": "hello from assistant dm",
+            "channel": "D123",
+            "channel_type": "im",
+            "thread_ts": "171.000",
+            "ts": "171.111",
+            "team": "T_TEAM",
+        }
+
+        await assistant_adapter._handle_slack_message(event)
+
+        msg_event = assistant_adapter.handle_message.call_args[0][0]
+        assert msg_event.source.user_id == "U_USER"
+        assert msg_event.source.thread_id == "171.000"
+        assert msg_event.source.user_name == "Tyler"
+
+    def test_assistant_threads_cache_eviction(self, assistant_adapter):
+        """Cache should evict oldest entries when exceeding the size limit."""
+        assistant_adapter._ASSISTANT_THREADS_MAX = 10
+        # Fill to the limit
+        for i in range(10):
+            assistant_adapter._cache_assistant_thread_metadata({
+                "channel_id": f"D{i}",
+                "thread_ts": f"{i}.000",
+                "user_id": f"U{i}",
+            })
+        assert len(assistant_adapter._assistant_threads) == 10
+
+        # Adding one more should trigger eviction (down to max // 2 = 5)
+        assistant_adapter._cache_assistant_thread_metadata({
+            "channel_id": "D999",
+            "thread_ts": "999.000",
+            "user_id": "U999",
+        })
+        assert len(assistant_adapter._assistant_threads) <= 10
+        # The newest entry must survive eviction
+        assert ("D999", "999.000") in assistant_adapter._assistant_threads
+
+
 # ---------------------------------------------------------------------------
 # TestUserNameResolution
 # ---------------------------------------------------------------------------

tests/gateway/test_stream_consumer.py

@@ -383,6 +383,60 @@ class TestSegmentBreakOnToolBoundary:
         sent_texts = [call[1]["content"] for call in adapter.send.call_args_list]
         assert sent_texts == ["Hello ▉", "Next segment"]
 
+    @pytest.mark.asyncio
+    async def test_no_message_id_enters_fallback_mode(self):
+        """Platform returns success but no message_id (Signal) — must not
+        re-send on every delta.  Should enter fallback mode and send only
+        the continuation at finish."""
+        adapter = MagicMock()
+        # First send succeeds but returns no message_id (Signal behavior)
+        send_result_no_id = SimpleNamespace(success=True, message_id=None)
+        # Fallback final send succeeds
+        send_result_final = SimpleNamespace(success=True, message_id="msg_final")
+        adapter.send = AsyncMock(side_effect=[send_result_no_id, send_result_final])
+        adapter.edit_message = AsyncMock(return_value=SimpleNamespace(success=True))
+        adapter.MAX_MESSAGE_LENGTH = 4096
+
+        config = StreamConsumerConfig(edit_interval=0.01, buffer_threshold=5)
+        consumer = GatewayStreamConsumer(adapter, "chat_123", config)
+
+        consumer.on_delta("Hello")
+        task = asyncio.create_task(consumer.run())
+        await asyncio.sleep(0.08)
+        consumer.on_delta(" world, this is a longer response.")
+        await asyncio.sleep(0.08)
+        consumer.finish()
+        await task
+
+        # Should send exactly 2 messages: initial chunk + fallback continuation
+        # NOT one message per delta
+        assert adapter.send.call_count == 2
+        assert consumer.already_sent
+        # edit_message should NOT have been called (no valid message_id to edit)
+        adapter.edit_message.assert_not_called()
+
+    @pytest.mark.asyncio
+    async def test_no_message_id_single_delta_marks_already_sent(self):
+        """When the entire response fits in one delta and platform returns no
+        message_id, already_sent must still be True to prevent the gateway
+        from re-sending the full response."""
+        adapter = MagicMock()
+        send_result = SimpleNamespace(success=True, message_id=None)
+        adapter.send = AsyncMock(return_value=send_result)
+        adapter.MAX_MESSAGE_LENGTH = 4096
+
+        config = StreamConsumerConfig(edit_interval=0.01, buffer_threshold=5)
+        consumer = GatewayStreamConsumer(adapter, "chat_123", config)
+
+        consumer.on_delta("Short response.")
+        consumer.finish()
+
+        await consumer.run()
+
+        assert consumer.already_sent
+        # Only one send call (the initial message)
+        assert adapter.send.call_count == 1
+
     @pytest.mark.asyncio
     async def test_fallback_final_splits_long_continuation_without_dropping_text(self):
         """Long continuation tails should be chunked when fallback final-send runs."""

tests/gateway/test_wecom.py

@@ -4,7 +4,7 @@ import base64
 import os
 from pathlib import Path
 from types import SimpleNamespace
-from unittest.mock import AsyncMock
+from unittest.mock import AsyncMock, patch
 
 import pytest
 
@@ -355,7 +355,8 @@ class TestMediaUpload:
         assert calls[3][1]["chunk_index"] == 2
 
     @pytest.mark.asyncio
-    async def test_download_remote_bytes_rejects_large_content_length(self):
+    @patch("tools.url_safety.is_safe_url", return_value=True)
+    async def test_download_remote_bytes_rejects_large_content_length(self, _mock_safe):
         from gateway.platforms.wecom import WeComAdapter
 
         class FakeResponse:

tests/hermes_cli/test_api_key_providers.py

@@ -628,14 +628,21 @@ class TestHasAnyProviderConfigured:
     def test_claude_code_creds_ignored_on_fresh_install(self, monkeypatch, tmp_path):
         """Claude Code credentials should NOT skip the wizard when Hermes is unconfigured."""
         from hermes_cli import config as config_module
+        from hermes_cli.auth import PROVIDER_REGISTRY
         hermes_home = tmp_path / ".hermes"
         hermes_home.mkdir()
         monkeypatch.setattr(config_module, "get_env_path", lambda: hermes_home / ".env")
         monkeypatch.setattr(config_module, "get_hermes_home", lambda: hermes_home)
         # Clear all provider env vars so earlier checks don't short-circuit
-        for var in ("OPENROUTER_API_KEY", "OPENAI_API_KEY", "ANTHROPIC_API_KEY",
-                     "ANTHROPIC_TOKEN", "OPENAI_BASE_URL"):
+        _all_vars = {"OPENROUTER_API_KEY", "OPENAI_API_KEY", "ANTHROPIC_API_KEY",
+                      "ANTHROPIC_TOKEN", "OPENAI_BASE_URL"}
+        for pconfig in PROVIDER_REGISTRY.values():
+            if pconfig.auth_type == "api_key":
+                _all_vars.update(pconfig.api_key_env_vars)
+        for var in _all_vars:
             monkeypatch.delenv(var, raising=False)
+        # Prevent gh-cli / copilot auth fallback from leaking in
+        monkeypatch.setattr("hermes_cli.auth.get_auth_status", lambda _pid: {})
         # Simulate valid Claude Code credentials
         monkeypatch.setattr(
             "agent.anthropic_adapter.read_claude_code_credentials",
@@ -710,6 +717,7 @@ class TestHasAnyProviderConfigured:
         """config.yaml model dict with empty default and no creds stays false."""
         import yaml
         from hermes_cli import config as config_module
+        from hermes_cli.auth import PROVIDER_REGISTRY
         hermes_home = tmp_path / ".hermes"
         hermes_home.mkdir()
         config_file = hermes_home / "config.yaml"
@@ -719,9 +727,15 @@ class TestHasAnyProviderConfigured:
         monkeypatch.setattr(config_module, "get_env_path", lambda: hermes_home / ".env")
         monkeypatch.setattr(config_module, "get_hermes_home", lambda: hermes_home)
         monkeypatch.setenv("HERMES_HOME", str(hermes_home))
-        for var in ("OPENROUTER_API_KEY", "OPENAI_API_KEY", "ANTHROPIC_API_KEY",
-                     "ANTHROPIC_TOKEN", "OPENAI_BASE_URL"):
+        _all_vars = {"OPENROUTER_API_KEY", "OPENAI_API_KEY", "ANTHROPIC_API_KEY",
+                      "ANTHROPIC_TOKEN", "OPENAI_BASE_URL"}
+        for pconfig in PROVIDER_REGISTRY.values():
+            if pconfig.auth_type == "api_key":
+                _all_vars.update(pconfig.api_key_env_vars)
+        for var in _all_vars:
             monkeypatch.delenv(var, raising=False)
+        # Prevent gh-cli / copilot auth fallback from leaking in
+        monkeypatch.setattr("hermes_cli.auth.get_auth_status", lambda _pid: {})
         from hermes_cli.main import _has_any_provider_configured
         assert _has_any_provider_configured() is False
 
@@ -941,9 +955,10 @@ class TestHuggingFaceModels:
         """Every HF model should have a context length entry."""
         from hermes_cli.models import _PROVIDER_MODELS
         from agent.model_metadata import DEFAULT_CONTEXT_LENGTHS
+        lower_keys = {k.lower() for k in DEFAULT_CONTEXT_LENGTHS}
         hf_models = _PROVIDER_MODELS["huggingface"]
         for model in hf_models:
-            assert model in DEFAULT_CONTEXT_LENGTHS, (
+            assert model.lower() in lower_keys, (
                 f"HF model {model!r} missing from DEFAULT_CONTEXT_LENGTHS"
             )
 

tests/hermes_cli/test_auth_qwen_provider.py

@@ -0,0 +1,399 @@
+"""Tests for Qwen OAuth provider authentication (hermes_cli/auth.py).
+
+Covers: _qwen_cli_auth_path, _read_qwen_cli_tokens, _save_qwen_cli_tokens,
+_qwen_access_token_is_expiring, _refresh_qwen_cli_tokens,
+resolve_qwen_runtime_credentials, get_qwen_auth_status.
+"""
+
+import json
+import os
+import stat
+import time
+from pathlib import Path
+from unittest.mock import MagicMock, patch
+
+import pytest
+
+from hermes_cli.auth import (
+    AuthError,
+    DEFAULT_QWEN_BASE_URL,
+    QWEN_ACCESS_TOKEN_REFRESH_SKEW_SECONDS,
+    _qwen_cli_auth_path,
+    _read_qwen_cli_tokens,
+    _save_qwen_cli_tokens,
+    _qwen_access_token_is_expiring,
+    _refresh_qwen_cli_tokens,
+    resolve_qwen_runtime_credentials,
+    get_qwen_auth_status,
+)
+
+
+# ---------------------------------------------------------------------------
+# Helpers
+# ---------------------------------------------------------------------------
+
+def _make_qwen_tokens(
+    access_token="test-access-token",
+    refresh_token="test-refresh-token",
+    expiry_date=None,
+    **extra,
+):
+    """Create a minimal Qwen CLI OAuth credential dict."""
+    if expiry_date is None:
+        # 1 hour from now in milliseconds
+        expiry_date = int((time.time() + 3600) * 1000)
+    data = {
+        "access_token": access_token,
+        "refresh_token": refresh_token,
+        "token_type": "Bearer",
+        "expiry_date": expiry_date,
+        "resource_url": "portal.qwen.ai",
+    }
+    data.update(extra)
+    return data
+
+
+def _write_qwen_creds(tmp_path, tokens=None):
+    """Write tokens to the Qwen CLI credentials file and return the path."""
+    qwen_dir = tmp_path / ".qwen"
+    qwen_dir.mkdir(parents=True, exist_ok=True)
+    creds_path = qwen_dir / "oauth_creds.json"
+    if tokens is None:
+        tokens = _make_qwen_tokens()
+    creds_path.write_text(json.dumps(tokens), encoding="utf-8")
+    return creds_path
+
+
+@pytest.fixture()
+def qwen_env(tmp_path, monkeypatch):
+    """Redirect _qwen_cli_auth_path to tmp_path/.qwen/oauth_creds.json."""
+    creds_path = tmp_path / ".qwen" / "oauth_creds.json"
+    monkeypatch.setattr(
+        "hermes_cli.auth._qwen_cli_auth_path", lambda: creds_path
+    )
+    return tmp_path
+
+
+# ---------------------------------------------------------------------------
+# _qwen_cli_auth_path
+# ---------------------------------------------------------------------------
+
+def test_qwen_cli_auth_path_returns_expected_location():
+    path = _qwen_cli_auth_path()
+    assert path == Path.home() / ".qwen" / "oauth_creds.json"
+
+
+# ---------------------------------------------------------------------------
+# _read_qwen_cli_tokens
+# ---------------------------------------------------------------------------
+
+def test_read_qwen_cli_tokens_success(qwen_env):
+    tokens = _make_qwen_tokens(access_token="my-access")
+    _write_qwen_creds(qwen_env, tokens)
+    result = _read_qwen_cli_tokens()
+    assert result["access_token"] == "my-access"
+    assert result["refresh_token"] == "test-refresh-token"
+
+
+def test_read_qwen_cli_tokens_missing_file(qwen_env):
+    with pytest.raises(AuthError) as exc:
+        _read_qwen_cli_tokens()
+    assert exc.value.code == "qwen_auth_missing"
+
+
+def test_read_qwen_cli_tokens_invalid_json(qwen_env):
+    creds_path = qwen_env / ".qwen" / "oauth_creds.json"
+    creds_path.parent.mkdir(parents=True, exist_ok=True)
+    creds_path.write_text("not json{{{", encoding="utf-8")
+    with pytest.raises(AuthError) as exc:
+        _read_qwen_cli_tokens()
+    assert exc.value.code == "qwen_auth_read_failed"
+
+
+def test_read_qwen_cli_tokens_non_dict(qwen_env):
+    creds_path = qwen_env / ".qwen" / "oauth_creds.json"
+    creds_path.parent.mkdir(parents=True, exist_ok=True)
+    creds_path.write_text(json.dumps(["a", "b"]), encoding="utf-8")
+    with pytest.raises(AuthError) as exc:
+        _read_qwen_cli_tokens()
+    assert exc.value.code == "qwen_auth_invalid"
+
+
+# ---------------------------------------------------------------------------
+# _save_qwen_cli_tokens
+# ---------------------------------------------------------------------------
+
+def test_save_qwen_cli_tokens_roundtrip(qwen_env):
+    tokens = _make_qwen_tokens(access_token="saved-token")
+    saved_path = _save_qwen_cli_tokens(tokens)
+    assert saved_path.exists()
+    loaded = json.loads(saved_path.read_text(encoding="utf-8"))
+    assert loaded["access_token"] == "saved-token"
+
+
+def test_save_qwen_cli_tokens_creates_parent(qwen_env):
+    tokens = _make_qwen_tokens()
+    saved_path = _save_qwen_cli_tokens(tokens)
+    assert saved_path.parent.exists()
+
+
+def test_save_qwen_cli_tokens_permissions(qwen_env):
+    tokens = _make_qwen_tokens()
+    saved_path = _save_qwen_cli_tokens(tokens)
+    mode = saved_path.stat().st_mode
+    assert mode & stat.S_IRUSR  # owner read
+    assert mode & stat.S_IWUSR  # owner write
+    assert not (mode & stat.S_IRGRP)  # no group read
+    assert not (mode & stat.S_IROTH)  # no other read
+
+
+# ---------------------------------------------------------------------------
+# _qwen_access_token_is_expiring
+# ---------------------------------------------------------------------------
+
+def test_expiring_token_not_expired():
+    # 1 hour from now in milliseconds
+    future_ms = int((time.time() + 3600) * 1000)
+    assert not _qwen_access_token_is_expiring(future_ms)
+
+
+def test_expiring_token_already_expired():
+    # 1 hour ago in milliseconds
+    past_ms = int((time.time() - 3600) * 1000)
+    assert _qwen_access_token_is_expiring(past_ms)
+
+
+def test_expiring_token_within_skew():
+    # Just inside the default skew window
+    near_ms = int((time.time() + QWEN_ACCESS_TOKEN_REFRESH_SKEW_SECONDS - 5) * 1000)
+    assert _qwen_access_token_is_expiring(near_ms)
+
+
+def test_expiring_token_none_returns_true():
+    assert _qwen_access_token_is_expiring(None)
+
+
+def test_expiring_token_non_numeric_returns_true():
+    assert _qwen_access_token_is_expiring("not-a-number")
+
+
+# ---------------------------------------------------------------------------
+# _refresh_qwen_cli_tokens
+# ---------------------------------------------------------------------------
+
+def test_refresh_qwen_cli_tokens_success(qwen_env):
+    tokens = _make_qwen_tokens(refresh_token="old-refresh")
+
+    resp = MagicMock()
+    resp.status_code = 200
+    resp.json.return_value = {
+        "access_token": "new-access",
+        "refresh_token": "new-refresh",
+        "expires_in": 7200,
+    }
+
+    with patch("hermes_cli.auth.httpx") as mock_httpx:
+        mock_httpx.post.return_value = resp
+        result = _refresh_qwen_cli_tokens(tokens)
+
+    assert result["access_token"] == "new-access"
+    assert result["refresh_token"] == "new-refresh"
+    assert "expiry_date" in result
+
+
+def test_refresh_qwen_cli_tokens_preserves_old_refresh_if_not_in_response(qwen_env):
+    tokens = _make_qwen_tokens(refresh_token="keep-me")
+
+    resp = MagicMock()
+    resp.status_code = 200
+    resp.json.return_value = {
+        "access_token": "new-access",
+        # No refresh_token in response — should keep old one
+        "expires_in": 3600,
+    }
+
+    with patch("hermes_cli.auth.httpx") as mock_httpx:
+        mock_httpx.post.return_value = resp
+        result = _refresh_qwen_cli_tokens(tokens)
+
+    assert result["refresh_token"] == "keep-me"
+
+
+def test_refresh_qwen_cli_tokens_missing_refresh_token():
+    tokens = {"access_token": "at", "refresh_token": ""}
+    with pytest.raises(AuthError) as exc:
+        _refresh_qwen_cli_tokens(tokens)
+    assert exc.value.code == "qwen_refresh_token_missing"
+
+
+def test_refresh_qwen_cli_tokens_http_error(qwen_env):
+    tokens = _make_qwen_tokens()
+
+    resp = MagicMock()
+    resp.status_code = 401
+    resp.text = "unauthorized"
+
+    with patch("hermes_cli.auth.httpx") as mock_httpx:
+        mock_httpx.post.return_value = resp
+        with pytest.raises(AuthError) as exc:
+            _refresh_qwen_cli_tokens(tokens)
+    assert exc.value.code == "qwen_refresh_failed"
+
+
+def test_refresh_qwen_cli_tokens_network_error(qwen_env):
+    tokens = _make_qwen_tokens()
+
+    with patch("hermes_cli.auth.httpx") as mock_httpx:
+        mock_httpx.post.side_effect = ConnectionError("timeout")
+        with pytest.raises(AuthError) as exc:
+            _refresh_qwen_cli_tokens(tokens)
+    assert exc.value.code == "qwen_refresh_failed"
+
+
+def test_refresh_qwen_cli_tokens_invalid_json_response(qwen_env):
+    tokens = _make_qwen_tokens()
+
+    resp = MagicMock()
+    resp.status_code = 200
+    resp.json.side_effect = ValueError("bad json")
+
+    with patch("hermes_cli.auth.httpx") as mock_httpx:
+        mock_httpx.post.return_value = resp
+        with pytest.raises(AuthError) as exc:
+            _refresh_qwen_cli_tokens(tokens)
+    assert exc.value.code == "qwen_refresh_invalid_json"
+
+
+def test_refresh_qwen_cli_tokens_missing_access_token_in_response(qwen_env):
+    tokens = _make_qwen_tokens()
+
+    resp = MagicMock()
+    resp.status_code = 200
+    resp.json.return_value = {"something": "but no access_token"}
+
+    with patch("hermes_cli.auth.httpx") as mock_httpx:
+        mock_httpx.post.return_value = resp
+        with pytest.raises(AuthError) as exc:
+            _refresh_qwen_cli_tokens(tokens)
+    assert exc.value.code == "qwen_refresh_invalid_response"
+
+
+def test_refresh_qwen_cli_tokens_default_expires_in(qwen_env):
+    """When expires_in is missing, default to 6 hours."""
+    tokens = _make_qwen_tokens()
+
+    resp = MagicMock()
+    resp.status_code = 200
+    resp.json.return_value = {"access_token": "new"}
+
+    with patch("hermes_cli.auth.httpx") as mock_httpx:
+        mock_httpx.post.return_value = resp
+        result = _refresh_qwen_cli_tokens(tokens)
+
+    # Verify expiry_date is roughly now + 6h (within 60s tolerance)
+    expected_ms = int(time.time() * 1000) + 6 * 60 * 60 * 1000
+    assert abs(result["expiry_date"] - expected_ms) < 60_000
+
+
+def test_refresh_qwen_cli_tokens_saves_to_disk(qwen_env):
+    tokens = _make_qwen_tokens()
+
+    resp = MagicMock()
+    resp.status_code = 200
+    resp.json.return_value = {
+        "access_token": "disk-check",
+        "expires_in": 3600,
+    }
+
+    with patch("hermes_cli.auth.httpx") as mock_httpx:
+        mock_httpx.post.return_value = resp
+        _refresh_qwen_cli_tokens(tokens)
+
+    # Verify it was persisted
+    creds_path = qwen_env / ".qwen" / "oauth_creds.json"
+    assert creds_path.exists()
+    saved = json.loads(creds_path.read_text(encoding="utf-8"))
+    assert saved["access_token"] == "disk-check"
+
+
+# ---------------------------------------------------------------------------
+# resolve_qwen_runtime_credentials
+# ---------------------------------------------------------------------------
+
+def test_resolve_qwen_runtime_credentials_fresh_token(qwen_env):
+    tokens = _make_qwen_tokens(access_token="fresh-at")
+    _write_qwen_creds(qwen_env, tokens)
+
+    creds = resolve_qwen_runtime_credentials(refresh_if_expiring=False)
+    assert creds["provider"] == "qwen-oauth"
+    assert creds["api_key"] == "fresh-at"
+    assert creds["base_url"] == DEFAULT_QWEN_BASE_URL
+    assert creds["source"] == "qwen-cli"
+
+
+def test_resolve_qwen_runtime_credentials_triggers_refresh(qwen_env):
+    # Write an expired token
+    expired_ms = int((time.time() - 3600) * 1000)
+    tokens = _make_qwen_tokens(access_token="old", expiry_date=expired_ms)
+    _write_qwen_creds(qwen_env, tokens)
+
+    refreshed = _make_qwen_tokens(access_token="refreshed-at")
+
+    with patch(
+        "hermes_cli.auth._refresh_qwen_cli_tokens", return_value=refreshed
+    ) as mock_refresh:
+        creds = resolve_qwen_runtime_credentials()
+    mock_refresh.assert_called_once()
+    assert creds["api_key"] == "refreshed-at"
+
+
+def test_resolve_qwen_runtime_credentials_force_refresh(qwen_env):
+    tokens = _make_qwen_tokens(access_token="old-at")
+    _write_qwen_creds(qwen_env, tokens)
+
+    refreshed = _make_qwen_tokens(access_token="force-refreshed")
+
+    with patch(
+        "hermes_cli.auth._refresh_qwen_cli_tokens", return_value=refreshed
+    ) as mock_refresh:
+        creds = resolve_qwen_runtime_credentials(force_refresh=True)
+    mock_refresh.assert_called_once()
+    assert creds["api_key"] == "force-refreshed"
+
+
+def test_resolve_qwen_runtime_credentials_missing_access_token(qwen_env):
+    tokens = _make_qwen_tokens(access_token="")
+    _write_qwen_creds(qwen_env, tokens)
+
+    with pytest.raises(AuthError) as exc:
+        resolve_qwen_runtime_credentials(refresh_if_expiring=False)
+    assert exc.value.code == "qwen_access_token_missing"
+
+
+def test_resolve_qwen_runtime_credentials_base_url_env_override(qwen_env, monkeypatch):
+    tokens = _make_qwen_tokens(access_token="at")
+    _write_qwen_creds(qwen_env, tokens)
+    monkeypatch.setenv("HERMES_QWEN_BASE_URL", "https://custom.qwen.ai/v1")
+
+    creds = resolve_qwen_runtime_credentials(refresh_if_expiring=False)
+    assert creds["base_url"] == "https://custom.qwen.ai/v1"
+
+
+# ---------------------------------------------------------------------------
+# get_qwen_auth_status
+# ---------------------------------------------------------------------------
+
+def test_get_qwen_auth_status_logged_in(qwen_env):
+    tokens = _make_qwen_tokens(access_token="status-at")
+    _write_qwen_creds(qwen_env, tokens)
+
+    status = get_qwen_auth_status()
+    assert status["logged_in"] is True
+    assert status["api_key"] == "status-at"
+
+
+def test_get_qwen_auth_status_not_logged_in(qwen_env):
+    # No credentials file
+    status = get_qwen_auth_status()
+    assert status["logged_in"] is False
+    assert "error" in status

tests/hermes_cli/test_commands.py

@@ -425,8 +425,8 @@ class TestSlashCommandCompleter:
 class TestSubcommands:
     def test_explicit_subcommands_extracted(self):
         """Commands with explicit subcommands on CommandDef are extracted."""
-        assert "/prompt" in SUBCOMMANDS
-        assert "clear" in SUBCOMMANDS["/prompt"]
+        assert "/skills" in SUBCOMMANDS
+        assert "install" in SUBCOMMANDS["/skills"]
 
     def test_reasoning_has_subcommands(self):
         assert "/reasoning" in SUBCOMMANDS

tests/hermes_cli/test_doctor.py

@@ -136,3 +136,73 @@ def test_check_gateway_service_linger_skips_when_service_not_installed(monkeypat
     out = capsys.readouterr().out
     assert out == ""
     assert issues == []
+
+
+# ── Memory provider section (doctor should only check the *active* provider) ──
+
+
+class TestDoctorMemoryProviderSection:
+    """The ◆ Memory Provider section should respect memory.provider config."""
+
+    def _make_hermes_home(self, tmp_path, provider=""):
+        """Create a minimal HERMES_HOME with config.yaml."""
+        home = tmp_path / ".hermes"
+        home.mkdir(parents=True, exist_ok=True)
+        import yaml
+        config = {"memory": {"provider": provider}} if provider else {"memory": {}}
+        (home / "config.yaml").write_text(yaml.dump(config))
+        return home
+
+    def _run_doctor_and_capture(self, monkeypatch, tmp_path, provider=""):
+        """Run doctor and capture stdout."""
+        home = self._make_hermes_home(tmp_path, provider)
+        monkeypatch.setattr(doctor_mod, "HERMES_HOME", home)
+        monkeypatch.setattr(doctor_mod, "PROJECT_ROOT", tmp_path / "project")
+        monkeypatch.setattr(doctor_mod, "_DHH", str(home))
+        (tmp_path / "project").mkdir(exist_ok=True)
+
+        # Stub tool availability (returns empty) so doctor runs past it
+        fake_model_tools = types.SimpleNamespace(
+            check_tool_availability=lambda *a, **kw: ([], []),
+            TOOLSET_REQUIREMENTS={},
+        )
+        monkeypatch.setitem(sys.modules, "model_tools", fake_model_tools)
+
+        # Stub auth checks to avoid real API calls
+        try:
+            from hermes_cli import auth as _auth_mod
+            monkeypatch.setattr(_auth_mod, "get_nous_auth_status", lambda: {})
+            monkeypatch.setattr(_auth_mod, "get_codex_auth_status", lambda: {})
+        except Exception:
+            pass
+
+        import io, contextlib
+        buf = io.StringIO()
+        with contextlib.redirect_stdout(buf):
+            doctor_mod.run_doctor(Namespace(fix=False))
+        return buf.getvalue()
+
+    def test_no_provider_shows_builtin_ok(self, monkeypatch, tmp_path):
+        out = self._run_doctor_and_capture(monkeypatch, tmp_path, provider="")
+        assert "Memory Provider" in out
+        assert "Built-in memory active" in out
+        # Should NOT mention Honcho or Mem0 errors
+        assert "Honcho API key" not in out
+        assert "Mem0" not in out
+
+    def test_honcho_provider_not_installed_shows_fail(self, monkeypatch, tmp_path):
+        # Make honcho import fail
+        monkeypatch.setitem(
+            sys.modules, "plugins.memory.honcho.client", None
+        )
+        out = self._run_doctor_and_capture(monkeypatch, tmp_path, provider="honcho")
+        assert "Memory Provider" in out
+        # Should show failure since honcho is set but not importable
+        assert "Built-in memory active" not in out
+
+    def test_mem0_provider_not_installed_shows_fail(self, monkeypatch, tmp_path):
+        # Make mem0 import fail
+        monkeypatch.setitem(sys.modules, "plugins.memory.mem0", None)
+        out = self._run_doctor_and_capture(monkeypatch, tmp_path, provider="mem0")
+        assert "Memory Provider" in out
+        assert "Built-in memory active" not in out

tests/hermes_cli/test_model_switch_variant_tags.py

@@ -0,0 +1,70 @@
+"""Tests for OpenRouter variant tag preservation in model switching.
+
+Regression test for GitHub PR #6088 / Discord report: OpenRouter model IDs
+with variant suffixes like ``:free``, ``:extended``, ``:fast`` were being
+mangled by the colon-to-slash conversion in model_switch.py Step c.
+
+The fix: Step c now skips colon→slash conversion when the model name already
+contains a forward slash (i.e. is already in ``vendor/model`` format), since
+the colon is a variant tag, not a vendor separator.
+"""
+import pytest
+from unittest.mock import patch
+
+from hermes_cli.model_switch import switch_model
+
+
+# Shared mock context — skip network calls, credential resolution, catalog lookups
+_MOCK_VALIDATION = {"accepted": True, "persist": True, "recognized": True, "message": None}
+
+
+def _run_switch(raw_input: str, current_provider: str = "openrouter") -> str:
+    """Run switch_model with mocked dependencies, return the resolved model name."""
+    with patch("hermes_cli.model_switch.resolve_alias", return_value=None), \
+         patch("hermes_cli.model_switch.list_provider_models", return_value=[]), \
+         patch("hermes_cli.runtime_provider.resolve_runtime_provider",
+               return_value={"api_key": "test", "base_url": "", "api_mode": "chat_completions"}), \
+         patch("hermes_cli.models.validate_requested_model", return_value=_MOCK_VALIDATION), \
+         patch("hermes_cli.model_switch.get_model_info", return_value=None), \
+         patch("hermes_cli.model_switch.get_model_capabilities", return_value=None), \
+         patch("hermes_cli.models.detect_provider_for_model", return_value=None):
+        result = switch_model(
+            raw_input=raw_input,
+            current_provider=current_provider,
+            current_model="anthropic/claude-sonnet-4.6",
+        )
+        assert result.success, f"switch_model failed: {result.error_message}"
+        return result.new_model
+
+
+class TestVariantTagPreservation:
+    """OpenRouter variant tags (:free, :extended, :fast) must survive model switching."""
+
+    @pytest.mark.parametrize("model,expected", [
+        ("nvidia/nemotron-3-super-120b-a12b:free", "nvidia/nemotron-3-super-120b-a12b:free"),
+        ("anthropic/claude-sonnet-4.6:extended", "anthropic/claude-sonnet-4.6:extended"),
+        ("meta-llama/llama-4-maverick:fast", "meta-llama/llama-4-maverick:fast"),
+    ])
+    def test_slash_format_preserves_variant_tag(self, model, expected):
+        """Models already in vendor/model:tag format must not have their tag mangled."""
+        assert _run_switch(model) == expected
+
+    def test_legacy_colon_format_converts_to_slash(self):
+        """Legacy vendor:model (no slash) should still be converted to vendor/model."""
+        result = _run_switch("nvidia:nemotron-3-super-120b-a12b")
+        assert result == "nvidia/nemotron-3-super-120b-a12b"
+
+    def test_legacy_colon_format_with_tag_converts_first_colon_only(self):
+        """vendor:model:free (no slash) → vendor/model:free — first colon becomes slash."""
+        result = _run_switch("nvidia:nemotron-3-super-120b-a12b:free")
+        assert result == "nvidia/nemotron-3-super-120b-a12b:free"
+
+    def test_bare_model_name_unaffected(self):
+        """Bare model names without colons or slashes should work normally."""
+        result = _run_switch("claude-sonnet-4.6")
+        assert result == "anthropic/claude-sonnet-4.6"
+
+    def test_already_correct_slug_no_tag(self):
+        """Standard vendor/model slugs without tags pass through unchanged."""
+        result = _run_switch("anthropic/claude-sonnet-4.6")
+        assert result == "anthropic/claude-sonnet-4.6"

tests/hermes_cli/test_runtime_provider_resolution.py

@@ -143,6 +143,82 @@ def test_resolve_runtime_provider_codex(monkeypatch):
     assert resolved["requested_provider"] == "openai-codex"
 
 
+def test_resolve_runtime_provider_qwen_oauth(monkeypatch):
+    monkeypatch.setattr(rp, "resolve_provider", lambda *a, **k: "qwen-oauth")
+    monkeypatch.setattr(
+        rp,
+        "resolve_qwen_runtime_credentials",
+        lambda: {
+            "provider": "qwen-oauth",
+            "base_url": "https://portal.qwen.ai/v1",
+            "api_key": "qwen-token",
+            "source": "qwen-cli",
+            "expires_at_ms": 1775640710946,
+        },
+    )
+
+    resolved = rp.resolve_runtime_provider(requested="qwen-oauth")
+
+    assert resolved["provider"] == "qwen-oauth"
+    assert resolved["api_mode"] == "chat_completions"
+    assert resolved["base_url"] == "https://portal.qwen.ai/v1"
+    assert resolved["api_key"] == "qwen-token"
+    assert resolved["requested_provider"] == "qwen-oauth"
+
+
+def test_resolve_runtime_provider_uses_qwen_pool_entry(monkeypatch):
+    class _Entry:
+        access_token = "pool-qwen-token"
+        source = "manual:qwen_cli"
+        base_url = "https://portal.qwen.ai/v1"
+
+    class _Pool:
+        def has_credentials(self):
+            return True
+
+        def select(self):
+            return _Entry()
+
+    monkeypatch.setattr(rp, "resolve_provider", lambda *a, **k: "qwen-oauth")
+    monkeypatch.setattr(rp, "load_pool", lambda provider: _Pool())
+    monkeypatch.setattr(rp, "_get_model_config", lambda: {"provider": "qwen-oauth", "default": "coder-model"})
+
+    resolved = rp.resolve_runtime_provider(requested="qwen-oauth")
+
+    assert resolved["provider"] == "qwen-oauth"
+    assert resolved["api_mode"] == "chat_completions"
+    assert resolved["base_url"] == "https://portal.qwen.ai/v1"
+    assert resolved["api_key"] == "pool-qwen-token"
+    assert resolved["source"] == "manual:qwen_cli"
+
+
+def test_resolve_provider_alias_qwen(monkeypatch):
+    monkeypatch.setattr(rp.auth_mod, "_load_auth_store", lambda: {})
+    monkeypatch.delenv("OPENAI_API_KEY", raising=False)
+    monkeypatch.delenv("OPENROUTER_API_KEY", raising=False)
+    assert rp.resolve_provider("qwen-portal") == "qwen-oauth"
+    assert rp.resolve_provider("qwen-cli") == "qwen-oauth"
+
+
+def test_qwen_oauth_auto_fallthrough_on_auth_failure(monkeypatch):
+    """When requested_provider is 'auto' and Qwen creds fail, fall through."""
+    from hermes_cli.auth import AuthError
+
+    monkeypatch.setattr(rp, "resolve_provider", lambda *a, **k: "qwen-oauth")
+    monkeypatch.setattr(
+        rp,
+        "resolve_qwen_runtime_credentials",
+        lambda **kw: (_ for _ in ()).throw(AuthError("stale", provider="qwen-oauth", code="qwen_auth_missing")),
+    )
+    monkeypatch.setattr(rp, "_get_model_config", lambda: {})
+    monkeypatch.setenv("OPENROUTER_API_KEY", "test-or-key")
+
+    # Should NOT raise — falls through to OpenRouter
+    resolved = rp.resolve_runtime_provider(requested="auto")
+    # The fallthrough means it won't be qwen-oauth
+    assert resolved["provider"] != "qwen-oauth"
+
+
 def test_resolve_runtime_provider_ai_gateway(monkeypatch):
     monkeypatch.setattr(rp, "resolve_provider", lambda *a, **k: "ai-gateway")
     monkeypatch.setattr(rp, "_get_model_config", lambda: {})

tests/hermes_cli/test_setup_openclaw_migration.py

@@ -44,7 +44,7 @@ class TestOfferOpenclawMigration:
             assert setup_mod._offer_openclaw_migration(tmp_path / ".hermes") is False
 
     def test_runs_migration_when_user_accepts(self, tmp_path):
-        """Should dynamically load the script and run the Migrator."""
+        """Should run dry-run preview first, then execute after confirmation."""
         openclaw_dir = tmp_path / ".openclaw"
         openclaw_dir.mkdir()
 
@@ -60,6 +60,7 @@ class TestOfferOpenclawMigration:
         fake_migrator = MagicMock()
         fake_migrator.migrate.return_value = {
             "summary": {"migrated": 3, "skipped": 1, "conflict": 0, "error": 0},
+            "items": [{"kind": "config", "status": "migrated", "destination": "/tmp/x"}],
             "output_dir": str(hermes_home / "migration"),
         }
         fake_mod.Migrator = MagicMock(return_value=fake_migrator)
@@ -70,6 +71,7 @@ class TestOfferOpenclawMigration:
         with (
             patch("hermes_cli.setup.Path.home", return_value=tmp_path),
             patch.object(setup_mod, "_OPENCLAW_SCRIPT", script),
+            # Both prompts answered Yes: preview offer + proceed confirmation
             patch.object(setup_mod, "prompt_yes_no", return_value=True),
             patch.object(setup_mod, "get_config_path", return_value=config_path),
             patch("importlib.util.spec_from_file_location") as mock_spec_fn,
@@ -91,13 +93,75 @@ class TestOfferOpenclawMigration:
         fake_mod.resolve_selected_options.assert_called_once_with(
             None, None, preset="full"
         )
-        fake_mod.Migrator.assert_called_once()
-        call_kwargs = fake_mod.Migrator.call_args[1]
-        assert call_kwargs["execute"] is True
-        assert call_kwargs["overwrite"] is True
-        assert call_kwargs["migrate_secrets"] is True
-        assert call_kwargs["preset_name"] == "full"
-        fake_migrator.migrate.assert_called_once()
+        # Migrator called twice: once for dry-run preview, once for execution
+        assert fake_mod.Migrator.call_count == 2
+
+        # First call: dry-run preview (execute=False, overwrite=True to show all)
+        preview_kwargs = fake_mod.Migrator.call_args_list[0][1]
+        assert preview_kwargs["execute"] is False
+        assert preview_kwargs["overwrite"] is True
+        assert preview_kwargs["migrate_secrets"] is True
+        assert preview_kwargs["preset_name"] == "full"
+
+        # Second call: actual execution (execute=True, overwrite=False to preserve)
+        exec_kwargs = fake_mod.Migrator.call_args_list[1][1]
+        assert exec_kwargs["execute"] is True
+        assert exec_kwargs["overwrite"] is False
+        assert exec_kwargs["migrate_secrets"] is True
+        assert exec_kwargs["preset_name"] == "full"
+
+        # migrate() called twice (once per Migrator instance)
+        assert fake_migrator.migrate.call_count == 2
+
+    def test_user_declines_after_preview(self, tmp_path):
+        """Should return False when user sees preview but declines to proceed."""
+        openclaw_dir = tmp_path / ".openclaw"
+        openclaw_dir.mkdir()
+
+        hermes_home = tmp_path / ".hermes"
+        hermes_home.mkdir()
+        config_path = hermes_home / "config.yaml"
+        config_path.write_text("agent:\n  max_turns: 90\n")
+
+        fake_mod = ModuleType("openclaw_to_hermes")
+        fake_mod.resolve_selected_options = MagicMock(return_value={"soul", "memory"})
+        fake_migrator = MagicMock()
+        fake_migrator.migrate.return_value = {
+            "summary": {"migrated": 3, "skipped": 0, "conflict": 0, "error": 0},
+            "items": [{"kind": "config", "status": "migrated", "destination": "/tmp/x"}],
+        }
+        fake_mod.Migrator = MagicMock(return_value=fake_migrator)
+
+        script = tmp_path / "openclaw_to_hermes.py"
+        script.write_text("# placeholder")
+
+        # First prompt (preview): Yes, Second prompt (proceed): No
+        prompt_responses = iter([True, False])
+
+        with (
+            patch("hermes_cli.setup.Path.home", return_value=tmp_path),
+            patch.object(setup_mod, "_OPENCLAW_SCRIPT", script),
+            patch.object(setup_mod, "prompt_yes_no", side_effect=prompt_responses),
+            patch.object(setup_mod, "get_config_path", return_value=config_path),
+            patch("importlib.util.spec_from_file_location") as mock_spec_fn,
+        ):
+            mock_spec = MagicMock()
+            mock_spec.loader = MagicMock()
+            mock_spec_fn.return_value = mock_spec
+
+            def exec_module(mod):
+                mod.resolve_selected_options = fake_mod.resolve_selected_options
+                mod.Migrator = fake_mod.Migrator
+
+            mock_spec.loader.exec_module = exec_module
+
+            result = setup_mod._offer_openclaw_migration(hermes_home)
+
+        assert result is False
+        # Only dry-run Migrator was created, not the execute one
+        assert fake_mod.Migrator.call_count == 1
+        preview_kwargs = fake_mod.Migrator.call_args[1]
+        assert preview_kwargs["execute"] is False
 
     def test_handles_migration_error_gracefully(self, tmp_path):
         """Should catch exceptions and return False."""

tests/hermes_cli/test_tools_config.py

@@ -354,6 +354,14 @@ def test_first_install_nous_auto_configures_managed_defaults(monkeypatch):
         lambda *args, **kwargs: {"web", "image_gen", "tts", "browser"},
     )
     monkeypatch.setattr("hermes_cli.tools_config.save_config", lambda config: None)
+    # Prevent leaked platform tokens (e.g. DISCORD_BOT_TOKEN from gateway.run
+    # import) from adding extra platforms. The loop in tools_command runs
+    # apply_nous_managed_defaults per platform; a second iteration sees values
+    # set by the first as "explicit" and skips them.
+    monkeypatch.setattr(
+        "hermes_cli.tools_config._get_enabled_platforms",
+        lambda: ["cli"],
+    )
     monkeypatch.setattr(
         "hermes_cli.nous_subscription.get_nous_auth_status",
         lambda: {"logged_in": True},

tests/hermes_cli/test_update_gateway_restart.py

@@ -368,6 +368,9 @@ class TestCmdUpdateLaunchdRestart:
         monkeypatch.setattr(
             gateway_cli, "is_macos", lambda: False,
         )
+        monkeypatch.setattr(
+            gateway_cli, "is_linux", lambda: True,
+        )
 
         mock_run.side_effect = _make_run_side_effect(
             commit_count="3",

tests/plugins/memory/test_hindsight_provider.py

@@ -0,0 +1,598 @@
+"""Tests for the Hindsight memory provider plugin.
+
+Tests cover config loading, tool handlers (tags, max_tokens, types),
+prefetch (auto_recall, preamble, query truncation), sync_turn (auto_retain,
+turn counting, tags), and schema completeness.
+"""
+
+import json
+import threading
+from types import SimpleNamespace
+from unittest.mock import AsyncMock, MagicMock, patch
+
+import pytest
+
+from plugins.memory.hindsight import (
+    HindsightMemoryProvider,
+    RECALL_SCHEMA,
+    REFLECT_SCHEMA,
+    RETAIN_SCHEMA,
+    _load_config,
+)
+
+
+# ---------------------------------------------------------------------------
+# Fixtures
+# ---------------------------------------------------------------------------
+
+
+@pytest.fixture(autouse=True)
+def _clean_env(monkeypatch):
+    """Ensure no stale env vars leak between tests."""
+    for key in (
+        "HINDSIGHT_API_KEY", "HINDSIGHT_API_URL", "HINDSIGHT_BANK_ID",
+        "HINDSIGHT_BUDGET", "HINDSIGHT_MODE", "HINDSIGHT_LLM_API_KEY",
+    ):
+        monkeypatch.delenv(key, raising=False)
+
+
+def _make_mock_client():
+    """Create a mock Hindsight client with async methods."""
+    client = MagicMock()
+    client.aretain = AsyncMock()
+    client.arecall = AsyncMock(
+        return_value=SimpleNamespace(
+            results=[
+                SimpleNamespace(text="Memory 1"),
+                SimpleNamespace(text="Memory 2"),
+            ]
+        )
+    )
+    client.areflect = AsyncMock(
+        return_value=SimpleNamespace(text="Synthesized answer")
+    )
+    client.aretain_batch = AsyncMock()
+    client.aclose = AsyncMock()
+    return client
+
+
+@pytest.fixture()
+def provider(tmp_path, monkeypatch):
+    """Create an initialized HindsightMemoryProvider with a mock client."""
+    config = {
+        "mode": "cloud",
+        "apiKey": "test-key",
+        "api_url": "http://localhost:9999",
+        "bank_id": "test-bank",
+        "budget": "mid",
+        "memory_mode": "hybrid",
+    }
+    config_path = tmp_path / "hindsight" / "config.json"
+    config_path.parent.mkdir(parents=True, exist_ok=True)
+    config_path.write_text(json.dumps(config))
+
+    monkeypatch.setattr(
+        "plugins.memory.hindsight.get_hermes_home", lambda: tmp_path
+    )
+
+    p = HindsightMemoryProvider()
+    p.initialize(session_id="test-session", hermes_home=str(tmp_path), platform="cli")
+    p._client = _make_mock_client()
+    return p
+
+
+@pytest.fixture()
+def provider_with_config(tmp_path, monkeypatch):
+    """Create a provider factory that accepts custom config overrides."""
+    def _make(**overrides):
+        config = {
+            "mode": "cloud",
+            "apiKey": "test-key",
+            "api_url": "http://localhost:9999",
+            "bank_id": "test-bank",
+            "budget": "mid",
+            "memory_mode": "hybrid",
+        }
+        config.update(overrides)
+        config_path = tmp_path / "hindsight" / "config.json"
+        config_path.parent.mkdir(parents=True, exist_ok=True)
+        config_path.write_text(json.dumps(config))
+
+        monkeypatch.setattr(
+            "plugins.memory.hindsight.get_hermes_home", lambda: tmp_path
+        )
+
+        p = HindsightMemoryProvider()
+        p.initialize(session_id="test-session", hermes_home=str(tmp_path), platform="cli")
+        p._client = _make_mock_client()
+        return p
+    return _make
+
+
+# ---------------------------------------------------------------------------
+# Schema tests
+# ---------------------------------------------------------------------------
+
+
+class TestSchemas:
+    def test_retain_schema_has_content(self):
+        assert RETAIN_SCHEMA["name"] == "hindsight_retain"
+        assert "content" in RETAIN_SCHEMA["parameters"]["properties"]
+        assert "content" in RETAIN_SCHEMA["parameters"]["required"]
+
+    def test_recall_schema_has_query(self):
+        assert RECALL_SCHEMA["name"] == "hindsight_recall"
+        assert "query" in RECALL_SCHEMA["parameters"]["properties"]
+        assert "query" in RECALL_SCHEMA["parameters"]["required"]
+
+    def test_reflect_schema_has_query(self):
+        assert REFLECT_SCHEMA["name"] == "hindsight_reflect"
+        assert "query" in REFLECT_SCHEMA["parameters"]["properties"]
+
+    def test_get_tool_schemas_returns_three(self, provider):
+        schemas = provider.get_tool_schemas()
+        assert len(schemas) == 3
+        names = {s["name"] for s in schemas}
+        assert names == {"hindsight_retain", "hindsight_recall", "hindsight_reflect"}
+
+    def test_context_mode_returns_no_tools(self, provider_with_config):
+        p = provider_with_config(memory_mode="context")
+        assert p.get_tool_schemas() == []
+
+
+# ---------------------------------------------------------------------------
+# Config tests
+# ---------------------------------------------------------------------------
+
+
+class TestConfig:
+    def test_default_values(self, provider):
+        assert provider._auto_retain is True
+        assert provider._auto_recall is True
+        assert provider._retain_every_n_turns == 1
+        assert provider._recall_max_tokens == 4096
+        assert provider._recall_max_input_chars == 800
+        assert provider._tags is None
+        assert provider._recall_tags is None
+        assert provider._bank_mission == ""
+        assert provider._bank_retain_mission is None
+        assert provider._retain_context == "conversation between Hermes Agent and the User"
+
+    def test_custom_config_values(self, provider_with_config):
+        p = provider_with_config(
+            tags=["tag1", "tag2"],
+            recall_tags=["recall-tag"],
+            recall_tags_match="all",
+            auto_retain=False,
+            auto_recall=False,
+            retain_every_n_turns=3,
+            retain_context="custom-ctx",
+            bank_retain_mission="Extract key facts",
+            recall_max_tokens=2048,
+            recall_types=["world", "experience"],
+            recall_prompt_preamble="Custom preamble:",
+            recall_max_input_chars=500,
+            bank_mission="Test agent mission",
+        )
+        assert p._tags == ["tag1", "tag2"]
+        assert p._recall_tags == ["recall-tag"]
+        assert p._recall_tags_match == "all"
+        assert p._auto_retain is False
+        assert p._auto_recall is False
+        assert p._retain_every_n_turns == 3
+        assert p._retain_context == "custom-ctx"
+        assert p._bank_retain_mission == "Extract key facts"
+        assert p._recall_max_tokens == 2048
+        assert p._recall_types == ["world", "experience"]
+        assert p._recall_prompt_preamble == "Custom preamble:"
+        assert p._recall_max_input_chars == 500
+        assert p._bank_mission == "Test agent mission"
+
+    def test_config_from_env_fallback(self, tmp_path, monkeypatch):
+        """When no config file exists, falls back to env vars."""
+        monkeypatch.setattr(
+            "plugins.memory.hindsight.get_hermes_home",
+            lambda: tmp_path / "nonexistent",
+        )
+        monkeypatch.setenv("HINDSIGHT_MODE", "cloud")
+        monkeypatch.setenv("HINDSIGHT_API_KEY", "env-key")
+        monkeypatch.setenv("HINDSIGHT_BANK_ID", "env-bank")
+        monkeypatch.setenv("HINDSIGHT_BUDGET", "high")
+
+        cfg = _load_config()
+        assert cfg["apiKey"] == "env-key"
+        assert cfg["banks"]["hermes"]["bankId"] == "env-bank"
+        assert cfg["banks"]["hermes"]["budget"] == "high"
+
+
+# ---------------------------------------------------------------------------
+# Tool handler tests
+# ---------------------------------------------------------------------------
+
+
+class TestToolHandlers:
+    def test_retain_success(self, provider):
+        result = json.loads(provider.handle_tool_call(
+            "hindsight_retain", {"content": "user likes dark mode"}
+        ))
+        assert result["result"] == "Memory stored successfully."
+        provider._client.aretain.assert_called_once()
+        call_kwargs = provider._client.aretain.call_args.kwargs
+        assert call_kwargs["bank_id"] == "test-bank"
+        assert call_kwargs["content"] == "user likes dark mode"
+
+    def test_retain_with_tags(self, provider_with_config):
+        p = provider_with_config(tags=["pref", "ui"])
+        p.handle_tool_call("hindsight_retain", {"content": "likes dark mode"})
+        call_kwargs = p._client.aretain.call_args.kwargs
+        assert call_kwargs["tags"] == ["pref", "ui"]
+
+    def test_retain_without_tags(self, provider):
+        provider.handle_tool_call("hindsight_retain", {"content": "hello"})
+        call_kwargs = provider._client.aretain.call_args.kwargs
+        assert "tags" not in call_kwargs
+
+    def test_retain_missing_content(self, provider):
+        result = json.loads(provider.handle_tool_call(
+            "hindsight_retain", {}
+        ))
+        assert "error" in result
+
+    def test_recall_success(self, provider):
+        result = json.loads(provider.handle_tool_call(
+            "hindsight_recall", {"query": "dark mode"}
+        ))
+        assert "Memory 1" in result["result"]
+        assert "Memory 2" in result["result"]
+
+    def test_recall_passes_max_tokens(self, provider_with_config):
+        p = provider_with_config(recall_max_tokens=2048)
+        p.handle_tool_call("hindsight_recall", {"query": "test"})
+        call_kwargs = p._client.arecall.call_args.kwargs
+        assert call_kwargs["max_tokens"] == 2048
+
+    def test_recall_passes_tags(self, provider_with_config):
+        p = provider_with_config(recall_tags=["tag1"], recall_tags_match="all")
+        p.handle_tool_call("hindsight_recall", {"query": "test"})
+        call_kwargs = p._client.arecall.call_args.kwargs
+        assert call_kwargs["tags"] == ["tag1"]
+        assert call_kwargs["tags_match"] == "all"
+
+    def test_recall_passes_types(self, provider_with_config):
+        p = provider_with_config(recall_types=["world", "experience"])
+        p.handle_tool_call("hindsight_recall", {"query": "test"})
+        call_kwargs = p._client.arecall.call_args.kwargs
+        assert call_kwargs["types"] == ["world", "experience"]
+
+    def test_recall_no_results(self, provider):
+        provider._client.arecall.return_value = SimpleNamespace(results=[])
+        result = json.loads(provider.handle_tool_call(
+            "hindsight_recall", {"query": "test"}
+        ))
+        assert result["result"] == "No relevant memories found."
+
+    def test_recall_missing_query(self, provider):
+        result = json.loads(provider.handle_tool_call(
+            "hindsight_recall", {}
+        ))
+        assert "error" in result
+
+    def test_reflect_success(self, provider):
+        result = json.loads(provider.handle_tool_call(
+            "hindsight_reflect", {"query": "summarize"}
+        ))
+        assert result["result"] == "Synthesized answer"
+
+    def test_reflect_missing_query(self, provider):
+        result = json.loads(provider.handle_tool_call(
+            "hindsight_reflect", {}
+        ))
+        assert "error" in result
+
+    def test_unknown_tool(self, provider):
+        result = json.loads(provider.handle_tool_call(
+            "hindsight_unknown", {}
+        ))
+        assert "error" in result
+
+    def test_retain_error_handling(self, provider):
+        provider._client.aretain.side_effect = RuntimeError("connection failed")
+        result = json.loads(provider.handle_tool_call(
+            "hindsight_retain", {"content": "test"}
+        ))
+        assert "error" in result
+        assert "connection failed" in result["error"]
+
+    def test_recall_error_handling(self, provider):
+        provider._client.arecall.side_effect = RuntimeError("timeout")
+        result = json.loads(provider.handle_tool_call(
+            "hindsight_recall", {"query": "test"}
+        ))
+        assert "error" in result
+
+
+# ---------------------------------------------------------------------------
+# Prefetch tests
+# ---------------------------------------------------------------------------
+
+
+class TestPrefetch:
+    def test_prefetch_returns_empty_when_no_result(self, provider):
+        assert provider.prefetch("test") == ""
+
+    def test_prefetch_default_preamble(self, provider):
+        provider._prefetch_result = "- some memory"
+        result = provider.prefetch("test")
+        assert "Hindsight Memory" in result
+        assert "- some memory" in result
+
+    def test_prefetch_custom_preamble(self, provider_with_config):
+        p = provider_with_config(recall_prompt_preamble="Custom header:")
+        p._prefetch_result = "- memory line"
+        result = p.prefetch("test")
+        assert result.startswith("Custom header:")
+        assert "- memory line" in result
+
+    def test_queue_prefetch_skipped_in_tools_mode(self, provider_with_config):
+        p = provider_with_config(memory_mode="tools")
+        p.queue_prefetch("test")
+        # Should not start a thread
+        assert p._prefetch_thread is None
+
+    def test_queue_prefetch_skipped_when_auto_recall_off(self, provider_with_config):
+        p = provider_with_config(auto_recall=False)
+        p.queue_prefetch("test")
+        assert p._prefetch_thread is None
+
+    def test_queue_prefetch_truncates_query(self, provider_with_config):
+        p = provider_with_config(recall_max_input_chars=10)
+        # Mock _run_sync to capture the query
+        original_query = None
+
+        def _capture_recall(**kwargs):
+            nonlocal original_query
+            original_query = kwargs.get("query", "")
+            return SimpleNamespace(results=[])
+
+        p._client.arecall = AsyncMock(side_effect=_capture_recall)
+
+        long_query = "a" * 100
+        p.queue_prefetch(long_query)
+        if p._prefetch_thread:
+            p._prefetch_thread.join(timeout=5.0)
+
+        # The query passed to arecall should be truncated
+        if original_query is not None:
+            assert len(original_query) <= 10
+
+    def test_queue_prefetch_passes_recall_params(self, provider_with_config):
+        p = provider_with_config(
+            recall_tags=["t1"],
+            recall_tags_match="all",
+            recall_max_tokens=1024,
+            recall_types=["world"],
+        )
+        p.queue_prefetch("test query")
+        if p._prefetch_thread:
+            p._prefetch_thread.join(timeout=5.0)
+
+        call_kwargs = p._client.arecall.call_args.kwargs
+        assert call_kwargs["max_tokens"] == 1024
+        assert call_kwargs["tags"] == ["t1"]
+        assert call_kwargs["tags_match"] == "all"
+        assert call_kwargs["types"] == ["world"]
+
+
+# ---------------------------------------------------------------------------
+# sync_turn tests
+# ---------------------------------------------------------------------------
+
+
+class TestSyncTurn:
+    def _get_retain_kwargs(self, provider):
+        """Helper to get the kwargs from the aretain_batch call."""
+        return provider._client.aretain_batch.call_args.kwargs
+
+    def _get_retain_content(self, provider):
+        """Helper to get the raw content string from the first item."""
+        kwargs = self._get_retain_kwargs(provider)
+        return kwargs["items"][0]["content"]
+
+    def _get_retain_messages(self, provider):
+        """Helper to parse the first turn's messages from retained content.
+
+        Content is a JSON array of turns: [[msgs...], [msgs...], ...]
+        For single-turn tests, returns the first turn's messages.
+        """
+        content = self._get_retain_content(provider)
+        turns = json.loads(content)
+        return turns[0] if len(turns) == 1 else turns
+
+    def test_sync_turn_retains(self, provider):
+        provider.sync_turn("hello", "hi there")
+        if provider._sync_thread:
+            provider._sync_thread.join(timeout=5.0)
+        provider._client.aretain_batch.assert_called_once()
+        messages = self._get_retain_messages(provider)
+        assert len(messages) == 2
+        assert messages[0]["role"] == "user"
+        assert messages[0]["content"] == "hello"
+        assert "timestamp" in messages[0]
+        assert messages[1]["role"] == "assistant"
+        assert messages[1]["content"] == "hi there"
+        assert "timestamp" in messages[1]
+
+    def test_sync_turn_skipped_when_auto_retain_off(self, provider_with_config):
+        p = provider_with_config(auto_retain=False)
+        p.sync_turn("hello", "hi")
+        assert p._sync_thread is None
+        p._client.aretain_batch.assert_not_called()
+
+    def test_sync_turn_with_tags(self, provider_with_config):
+        p = provider_with_config(tags=["conv", "session1"])
+        p.sync_turn("hello", "hi")
+        if p._sync_thread:
+            p._sync_thread.join(timeout=5.0)
+        item = p._client.aretain_batch.call_args.kwargs["items"][0]
+        assert item["tags"] == ["conv", "session1"]
+
+    def test_sync_turn_uses_aretain_batch(self, provider):
+        """sync_turn should use aretain_batch with retain_async."""
+        provider.sync_turn("hello", "hi")
+        if provider._sync_thread:
+            provider._sync_thread.join(timeout=5.0)
+        provider._client.aretain_batch.assert_called_once()
+        call_kwargs = provider._client.aretain_batch.call_args.kwargs
+        assert call_kwargs["document_id"] == "test-session"
+        assert call_kwargs["retain_async"] is True
+        assert len(call_kwargs["items"]) == 1
+        assert call_kwargs["items"][0]["context"] == "conversation between Hermes Agent and the User"
+
+    def test_sync_turn_custom_context(self, provider_with_config):
+        p = provider_with_config(retain_context="my-agent")
+        p.sync_turn("hello", "hi")
+        if p._sync_thread:
+            p._sync_thread.join(timeout=5.0)
+        item = p._client.aretain_batch.call_args.kwargs["items"][0]
+        assert item["context"] == "my-agent"
+
+    def test_sync_turn_every_n_turns(self, provider_with_config):
+        """With retain_every_n_turns=3, only retains on every 3rd turn."""
+        p = provider_with_config(retain_every_n_turns=3)
+
+        p.sync_turn("turn1-user", "turn1-asst")
+        assert p._sync_thread is None  # not retained yet
+
+        p.sync_turn("turn2-user", "turn2-asst")
+        assert p._sync_thread is None  # not retained yet
+
+        p.sync_turn("turn3-user", "turn3-asst")
+        assert p._sync_thread is not None  # retained!
+        p._sync_thread.join(timeout=5.0)
+
+        p._client.aretain_batch.assert_called_once()
+        content = p._client.aretain_batch.call_args.kwargs["items"][0]["content"]
+        # Should contain all 3 turns
+        assert "turn1-user" in content
+        assert "turn2-user" in content
+        assert "turn3-user" in content
+
+    def test_sync_turn_accumulates_full_session(self, provider_with_config):
+        """Each retain sends the ENTIRE session, not just the latest batch."""
+        p = provider_with_config(retain_every_n_turns=2)
+
+        p.sync_turn("turn1-user", "turn1-asst")
+        p.sync_turn("turn2-user", "turn2-asst")
+        if p._sync_thread:
+            p._sync_thread.join(timeout=5.0)
+
+        p._client.aretain_batch.reset_mock()
+
+        p.sync_turn("turn3-user", "turn3-asst")
+        p.sync_turn("turn4-user", "turn4-asst")
+        if p._sync_thread:
+            p._sync_thread.join(timeout=5.0)
+
+        content = p._client.aretain_batch.call_args.kwargs["items"][0]["content"]
+        # Should contain ALL turns from the session
+        assert "turn1-user" in content
+        assert "turn2-user" in content
+        assert "turn3-user" in content
+        assert "turn4-user" in content
+
+    def test_sync_turn_passes_document_id(self, provider):
+        """sync_turn should pass session_id as document_id for dedup."""
+        provider.sync_turn("hello", "hi")
+        if provider._sync_thread:
+            provider._sync_thread.join(timeout=5.0)
+        call_kwargs = provider._client.aretain_batch.call_args.kwargs
+        assert call_kwargs["document_id"] == "test-session"
+
+    def test_sync_turn_error_does_not_raise(self, provider):
+        """Errors in sync_turn should be swallowed (non-blocking)."""
+        provider._client.aretain_batch.side_effect = RuntimeError("network error")
+        provider.sync_turn("hello", "hi")
+        if provider._sync_thread:
+            provider._sync_thread.join(timeout=5.0)
+        # Should not raise
+
+
+# ---------------------------------------------------------------------------
+# System prompt tests
+# ---------------------------------------------------------------------------
+
+
+class TestSystemPrompt:
+    def test_hybrid_mode_prompt(self, provider):
+        block = provider.system_prompt_block()
+        assert "Hindsight Memory" in block
+        assert "hindsight_recall" in block
+        assert "automatically injected" in block
+
+    def test_context_mode_prompt(self, provider_with_config):
+        p = provider_with_config(memory_mode="context")
+        block = p.system_prompt_block()
+        assert "context mode" in block
+        assert "hindsight_recall" not in block
+
+    def test_tools_mode_prompt(self, provider_with_config):
+        p = provider_with_config(memory_mode="tools")
+        block = p.system_prompt_block()
+        assert "tools mode" in block
+        assert "hindsight_recall" in block
+
+
+# ---------------------------------------------------------------------------
+# Config schema tests
+# ---------------------------------------------------------------------------
+
+
+class TestConfigSchema:
+    def test_schema_has_all_new_fields(self, provider):
+        schema = provider.get_config_schema()
+        keys = {f["key"] for f in schema}
+        expected_keys = {
+            "mode", "api_url", "api_key", "llm_provider", "llm_api_key",
+            "llm_model", "bank_id", "bank_mission", "bank_retain_mission",
+            "recall_budget", "memory_mode", "recall_prefetch_method",
+            "tags", "recall_tags", "recall_tags_match",
+            "auto_recall", "auto_retain",
+            "retain_every_n_turns", "retain_async",
+            "retain_context",
+            "recall_max_tokens", "recall_max_input_chars",
+            "recall_prompt_preamble",
+        }
+        assert expected_keys.issubset(keys), f"Missing: {expected_keys - keys}"
+
+
+# ---------------------------------------------------------------------------
+# Availability tests
+# ---------------------------------------------------------------------------
+
+
+class TestAvailability:
+    def test_available_with_api_key(self, tmp_path, monkeypatch):
+        monkeypatch.setattr(
+            "plugins.memory.hindsight.get_hermes_home",
+            lambda: tmp_path / "nonexistent",
+        )
+        monkeypatch.setenv("HINDSIGHT_API_KEY", "test-key")
+        p = HindsightMemoryProvider()
+        assert p.is_available()
+
+    def test_not_available_without_config(self, tmp_path, monkeypatch):
+        monkeypatch.setattr(
+            "plugins.memory.hindsight.get_hermes_home",
+            lambda: tmp_path / "nonexistent",
+        )
+        p = HindsightMemoryProvider()
+        assert not p.is_available()
+
+    def test_available_in_local_mode(self, tmp_path, monkeypatch):
+        monkeypatch.setattr(
+            "plugins.memory.hindsight.get_hermes_home",
+            lambda: tmp_path / "nonexistent",
+        )
+        monkeypatch.setenv("HINDSIGHT_MODE", "local")
+        p = HindsightMemoryProvider()
+        assert p.is_available()

tests/run_agent/test_context_pressure.py

@@ -150,8 +150,8 @@ def agent():
 class TestContextPressureFlags:
     """Context pressure warning flag tracking on AIAgent."""
 
-    def test_flag_initialized_false(self, agent):
-        assert agent._context_pressure_warned is False
+    def test_flag_initialized_zero(self, agent):
+        assert agent._context_pressure_warned_at == 0.0
 
     def test_emit_calls_status_callback(self, agent):
         """status_callback should be invoked with event type and message."""
@@ -210,7 +210,7 @@ class TestContextPressureFlags:
 
     def test_flag_reset_on_compression(self, agent):
         """After _compress_context, context pressure flag should reset."""
-        agent._context_pressure_warned = True
+        agent._context_pressure_warned_at = 0.85
         agent.compression_enabled = True
 
         agent.context_compressor = MagicMock()
@@ -219,6 +219,7 @@ class TestContextPressureFlags:
         ]
         agent.context_compressor.context_length = 200_000
         agent.context_compressor.threshold_tokens = 100_000
+        agent.context_compressor.compression_count = 1
 
         agent._todo_store = MagicMock()
         agent._todo_store.format_for_injection.return_value = None
@@ -233,7 +234,7 @@ class TestContextPressureFlags:
         ]
         agent._compress_context(messages, "system prompt")
 
-        assert agent._context_pressure_warned is False
+        assert agent._context_pressure_warned_at == 0.0
 
     def test_emit_callback_error_handled(self, agent):
         """If status_callback raises, it should be caught gracefully."""
@@ -246,3 +247,115 @@ class TestContextPressureFlags:
 
         # Should not raise
         agent._emit_context_pressure(0.85, compressor)
+
+    def test_tiered_reemits_at_95(self, agent):
+        """Warning fires at 85%, then fires again when crossing 95%."""
+        agent._context_pressure_warned_at = 0.85
+        # Simulate crossing 95%: the tier (0.95) > warned_at (0.85)
+        assert 0.95 > agent._context_pressure_warned_at
+        # After emission at 95%, the tier should update
+        agent._context_pressure_warned_at = 0.95
+        assert agent._context_pressure_warned_at == 0.95
+
+    def test_tiered_no_double_emit_at_same_level(self, agent):
+        """Once warned at 85%, further 85%+ readings don't re-warn."""
+        agent._context_pressure_warned_at = 0.85
+        # At 88%, tier is 0.85, which is NOT > warned_at (0.85)
+        _warn_tier = 0.85 if 0.88 >= 0.85 else 0.0
+        assert not (_warn_tier > agent._context_pressure_warned_at)
+
+    def test_flag_not_reset_when_compression_insufficient(self, agent):
+        """When compression can't drop below 85%, keep the flag set."""
+        agent._context_pressure_warned_at = 0.85
+        agent.compression_enabled = True
+
+        agent.context_compressor = MagicMock()
+        agent.context_compressor.compress.return_value = [
+            {"role": "user", "content": "Summary of conversation so far."}
+        ]
+        agent.context_compressor.context_length = 200
+        # Use a small threshold so the tiny compressed output still
+        # represents >= 85% of it (prevents flag reset).
+        agent.context_compressor.threshold_tokens = 10
+        agent.context_compressor.compression_count = 1
+        agent.context_compressor.last_prompt_tokens = 0
+
+        agent._todo_store = MagicMock()
+        agent._todo_store.format_for_injection.return_value = None
+        agent._build_system_prompt = MagicMock(return_value="system prompt")
+        agent._cached_system_prompt = "old system prompt"
+        agent._session_db = None
+
+        messages = [
+            {"role": "user", "content": "hello"},
+            {"role": "assistant", "content": "hi there"},
+        ]
+        agent._compress_context(messages, "system prompt")
+
+        # Post-compression is ~90% of threshold — flag should NOT reset
+        assert agent._context_pressure_warned_at == 0.85
+
+
+class TestContextPressureGatewayDedup:
+    """Class-level dedup prevents warning spam across AIAgent instances."""
+
+    def setup_method(self):
+        """Clear class-level dedup state between tests."""
+        AIAgent._context_pressure_last_warned.clear()
+
+    def test_second_instance_within_cooldown_suppressed(self):
+        """Same session, same tier, within cooldown — should be suppressed."""
+        import time
+        sid = "test_session_dedup"
+        # Simulate first warning
+        AIAgent._context_pressure_last_warned[sid] = (0.85, time.time())
+        # Second instance checking same tier within cooldown
+        _last = AIAgent._context_pressure_last_warned.get(sid)
+        _should_warn = _last is None or _last[0] < 0.85 or (time.time() - _last[1]) >= AIAgent._CONTEXT_PRESSURE_COOLDOWN
+        assert not _should_warn
+
+    def test_higher_tier_fires_despite_cooldown(self):
+        """Same session, higher tier — should fire even within cooldown."""
+        import time
+        sid = "test_session_tier"
+        AIAgent._context_pressure_last_warned[sid] = (0.85, time.time())
+        _last = AIAgent._context_pressure_last_warned.get(sid)
+        # 0.95 > 0.85 stored tier → should warn
+        _should_warn = _last is None or _last[0] < 0.95 or (time.time() - _last[1]) >= AIAgent._CONTEXT_PRESSURE_COOLDOWN
+        assert _should_warn
+
+    def test_warning_fires_after_cooldown_expires(self):
+        """Same session, same tier, after cooldown — should fire again."""
+        import time
+        sid = "test_session_expired"
+        # Set a timestamp far in the past
+        AIAgent._context_pressure_last_warned[sid] = (0.85, time.time() - AIAgent._CONTEXT_PRESSURE_COOLDOWN - 1)
+        _last = AIAgent._context_pressure_last_warned.get(sid)
+        _should_warn = _last is None or _last[0] < 0.85 or (time.time() - _last[1]) >= AIAgent._CONTEXT_PRESSURE_COOLDOWN
+        assert _should_warn
+
+    def test_compression_clears_dedup(self):
+        """After compression drops below 85%, dedup entry should be cleared."""
+        import time
+        sid = "test_session_clear"
+        AIAgent._context_pressure_last_warned[sid] = (0.85, time.time())
+        assert sid in AIAgent._context_pressure_last_warned
+        # Simulate what _compress_context does on reset
+        AIAgent._context_pressure_last_warned.pop(sid, None)
+        assert sid not in AIAgent._context_pressure_last_warned
+
+    def test_eviction_removes_stale_entries(self):
+        """Stale entries older than 2x cooldown should be evicted."""
+        import time
+        _now = time.time()
+        AIAgent._context_pressure_last_warned = {
+            "fresh": (0.85, _now),
+            "stale": (0.85, _now - AIAgent._CONTEXT_PRESSURE_COOLDOWN * 3),
+        }
+        _cutoff = _now - AIAgent._CONTEXT_PRESSURE_COOLDOWN * 2
+        AIAgent._context_pressure_last_warned = {
+            k: v for k, v in AIAgent._context_pressure_last_warned.items()
+            if v[1] > _cutoff
+        }
+        assert "fresh" in AIAgent._context_pressure_last_warned
+        assert "stale" not in AIAgent._context_pressure_last_warned

tests/run_agent/test_flush_memories_codex.py

@@ -91,6 +91,61 @@ def _chat_response_with_memory_call():
     )
 
 
+class TestFlushMemoriesRespectsConfigTimeout:
+    """flush_memories() must NOT hardcode timeout=30.0 — it should defer
+    to the config value via auxiliary.flush_memories.timeout."""
+
+    def test_auxiliary_path_omits_explicit_timeout(self, monkeypatch):
+        """When calling _call_llm, timeout should NOT be passed so that
+        _get_task_timeout('flush_memories') reads from config."""
+        agent = _make_agent(monkeypatch, api_mode="chat_completions", provider="openrouter")
+
+        mock_response = _chat_response_with_memory_call()
+
+        with patch("agent.auxiliary_client.call_llm", return_value=mock_response) as mock_call:
+            messages = [
+                {"role": "user", "content": "Hello"},
+                {"role": "assistant", "content": "Hi"},
+                {"role": "user", "content": "Note this"},
+            ]
+            with patch("tools.memory_tool.memory_tool", return_value="Saved."):
+                agent.flush_memories(messages)
+
+        mock_call.assert_called_once()
+        call_kwargs = mock_call.call_args
+        # timeout must NOT be explicitly passed (so _get_task_timeout resolves it)
+        assert "timeout" not in call_kwargs.kwargs, (
+            "flush_memories should not pass explicit timeout to _call_llm; "
+            "let _get_task_timeout('flush_memories') resolve from config"
+        )
+
+    def test_fallback_path_uses_config_timeout(self, monkeypatch):
+        """When auxiliary client is unavailable and we fall back to direct
+        OpenAI client, timeout should come from _get_task_timeout, not hardcoded."""
+        agent = _make_agent(monkeypatch, api_mode="chat_completions", provider="openrouter")
+        agent.client = MagicMock()
+        agent.client.chat.completions.create.return_value = _chat_response_with_memory_call()
+
+        custom_timeout = 180.0
+
+        with patch("agent.auxiliary_client.call_llm", side_effect=RuntimeError("no provider")), \
+             patch("agent.auxiliary_client._get_task_timeout", return_value=custom_timeout) as mock_gtt, \
+             patch("tools.memory_tool.memory_tool", return_value="Saved."):
+            messages = [
+                {"role": "user", "content": "Hello"},
+                {"role": "assistant", "content": "Hi"},
+                {"role": "user", "content": "Save this"},
+            ]
+            agent.flush_memories(messages)
+
+        mock_gtt.assert_called_once_with("flush_memories")
+        agent.client.chat.completions.create.assert_called_once()
+        call_kwargs = agent.client.chat.completions.create.call_args
+        assert call_kwargs.kwargs.get("timeout") == custom_timeout, (
+            f"Expected timeout={custom_timeout} from config, got {call_kwargs.kwargs.get('timeout')}"
+        )
+
+
 class TestFlushMemoriesUsesAuxiliaryClient:
     """When an auxiliary client is available, flush_memories should use it
     instead of self.client -- especially critical in Codex mode."""

tests/run_agent/test_run_agent.py

@@ -872,6 +872,52 @@ class TestBuildApiKwargs:
         kwargs = agent._build_api_kwargs(messages)
         assert kwargs["max_tokens"] == 4096
 
+    def test_qwen_portal_formats_messages_and_metadata(self, agent):
+        agent.base_url = "https://portal.qwen.ai/v1"
+        agent._base_url_lower = agent.base_url.lower()
+        agent.session_id = "sess-123"
+        messages = [
+            {"role": "system", "content": "You are helpful"},
+            {"role": "assistant", "content": "Got it"},
+            {"role": "user", "content": "hi"},
+        ]
+        kwargs = agent._build_api_kwargs(messages)
+        assert kwargs["metadata"]["sessionId"] == "sess-123"
+        assert kwargs["extra_body"]["vl_high_resolution_images"] is True
+        assert isinstance(kwargs["messages"][0]["content"], list)
+        assert kwargs["messages"][0]["content"][0]["cache_control"] == {"type": "ephemeral"}
+        assert kwargs["messages"][2]["content"][0]["text"] == "hi"
+
+    def test_qwen_portal_normalizes_bare_string_content_parts(self, agent):
+        agent.base_url = "https://portal.qwen.ai/v1"
+        agent._base_url_lower = agent.base_url.lower()
+        messages = [
+            {"role": "system", "content": [{"type": "text", "text": "system"}]},
+            {"role": "user", "content": ["hello", {"type": "text", "text": "world"}]},
+        ]
+        kwargs = agent._build_api_kwargs(messages)
+        user_content = kwargs["messages"][1]["content"]
+        assert user_content[0] == {"type": "text", "text": "hello"}
+        assert user_content[1] == {"type": "text", "text": "world"}
+
+    def test_qwen_portal_no_system_message(self, agent):
+        agent.base_url = "https://portal.qwen.ai/v1"
+        agent._base_url_lower = agent.base_url.lower()
+        messages = [{"role": "user", "content": "hi"}]
+        kwargs = agent._build_api_kwargs(messages)
+        # Should not crash even without a system message
+        assert kwargs["messages"][0]["content"][0]["text"] == "hi"
+        assert "cache_control" not in kwargs["messages"][0]["content"][0]
+
+    def test_qwen_portal_omits_max_tokens(self, agent):
+        agent.base_url = "https://portal.qwen.ai/v1"
+        agent._base_url_lower = agent.base_url.lower()
+        agent.max_tokens = 4096
+        messages = [{"role": "system", "content": "sys"}, {"role": "user", "content": "hi"}]
+        kwargs = agent._build_api_kwargs(messages)
+        assert "max_tokens" not in kwargs
+        assert "max_completion_tokens" not in kwargs
+
 
 class TestBuildAssistantMessage:
     def test_basic_message(self, agent):
@@ -1622,12 +1668,15 @@ class TestRunConversation:
             if roles[i] == "assistant" and roles[i + 1] == "assistant":
                 raise AssertionError("Consecutive assistant messages found in history")
 
-    def test_truly_empty_response_accepted_without_retry(self, agent):
-        """Truly empty response (no content, no reasoning) should still complete with (empty)."""
+    def test_truly_empty_response_retries_3_times_then_empty(self, agent):
+        """Truly empty response (no content, no reasoning) retries 3 times then falls through to (empty)."""
         self._setup_agent(agent)
         agent.base_url = "http://127.0.0.1:1234/v1"
         empty_resp = _mock_response(content=None, finish_reason="stop")
-        agent.client.chat.completions.create.side_effect = [empty_resp]
+        # 4 responses: 1 original + 3 nudge retries, all empty
+        agent.client.chat.completions.create.side_effect = [
+            empty_resp, empty_resp, empty_resp, empty_resp,
+        ]
         with (
             patch.object(agent, "_persist_session"),
             patch.object(agent, "_save_trajectory"),
@@ -1636,7 +1685,28 @@ class TestRunConversation:
             result = agent.run_conversation("answer me")
         assert result["completed"] is True
         assert result["final_response"] == "(empty)"
-        assert result["api_calls"] == 1  # no retries
+        assert result["api_calls"] == 4  # 1 original + 3 retries
+
+    def test_truly_empty_response_succeeds_on_nudge(self, agent):
+        """Model produces content after being nudged for empty response."""
+        self._setup_agent(agent)
+        agent.base_url = "http://127.0.0.1:1234/v1"
+        empty_resp = _mock_response(content=None, finish_reason="stop")
+        content_resp = _mock_response(
+            content="Here is the actual answer.",
+            finish_reason="stop",
+        )
+        # 1 empty response, then model produces content on nudge
+        agent.client.chat.completions.create.side_effect = [empty_resp, content_resp]
+        with (
+            patch.object(agent, "_persist_session"),
+            patch.object(agent, "_save_trajectory"),
+            patch.object(agent, "_cleanup_task_resources"),
+        ):
+            result = agent.run_conversation("answer me")
+        assert result["completed"] is True
+        assert result["final_response"] == "Here is the actual answer."
+        assert result["api_calls"] == 2  # 1 original + 1 nudge retry
 
     def test_nous_401_refreshes_after_remint_and_retries(self, agent):
         self._setup_agent(agent)

tests/skills/test_openclaw_migration.py

@@ -658,6 +658,47 @@ def test_workspace_agents_records_skip_when_missing(tmp_path: Path):
     assert wa_items[0]["status"] == "skipped"
 
 
+def test_cron_store_is_archived_without_config_cron_section(tmp_path: Path):
+    """Bug fix: archive cron store even when openclaw.json has no top-level cron config."""
+    mod = load_module()
+    source = tmp_path / ".openclaw"
+    target = tmp_path / ".hermes"
+    output_dir = target / "migration-report"
+    source.mkdir()
+    target.mkdir()
+
+    (source / "openclaw.json").write_text(json.dumps({"channels": {}}), encoding="utf-8")
+    (source / "cron").mkdir(parents=True)
+    (source / "cron" / "jobs.json").write_text(
+        json.dumps({"version": 1, "jobs": [{"id": "job-1", "name": "demo"}]}),
+        encoding="utf-8",
+    )
+
+    migrator = mod.Migrator(
+        source_root=source,
+        target_root=target,
+        execute=True,
+        workspace_target=None,
+        overwrite=False,
+        migrate_secrets=False,
+        output_dir=output_dir,
+        selected_options={"cron-jobs"},
+    )
+    report = migrator.migrate()
+
+    cron_items = [item for item in report["items"] if item["kind"] == "cron-jobs"]
+    archived_store = next(
+        (item for item in cron_items if item["destination"] and item["destination"].endswith("archive/cron-store")),
+        None,
+    )
+    assert archived_store is not None
+    assert Path(archived_store["destination"]).joinpath("jobs.json").exists()
+
+    notes_text = (output_dir / "MIGRATION_NOTES.md").read_text(encoding="utf-8")
+    assert "Run `hermes cron` to recreate scheduled tasks" in notes_text
+    assert "archive/cron-config.json" not in notes_text
+
+
 def test_skill_installs_cleanly_under_skills_guard():
     skills_guard = load_skills_guard()
     result = skills_guard.scan_skill(

tests/test_ctx_halving_fix.py

@@ -0,0 +1,319 @@
+"""Tests for the context-halving bugfix.
+
+Background
+----------
+When the API returns "max_tokens too large given prompt" (input is fine,
+but input_tokens + requested max_tokens > context_window), the old code
+incorrectly halved context_length via get_next_probe_tier().
+
+The fix introduces:
+  * parse_available_output_tokens_from_error() — detects this specific
+    error class and returns the available output token budget.
+  * _ephemeral_max_output_tokens on AIAgent — a one-shot override that
+    caps the output for one retry without touching context_length.
+
+Naming note
+-----------
+  max_tokens     = OUTPUT token cap (a single response).
+  context_length = TOTAL context window (input + output combined).
+These are different and the old code conflated them; the fix keeps them
+separate.
+"""
+
+import sys
+import os
+from unittest.mock import MagicMock, patch, PropertyMock
+
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), ".."))
+
+import pytest
+
+
+# ---------------------------------------------------------------------------
+# parse_available_output_tokens_from_error — unit tests
+# ---------------------------------------------------------------------------
+
+class TestParseAvailableOutputTokens:
+    """Pure-function tests; no I/O required."""
+
+    def _parse(self, msg):
+        from agent.model_metadata import parse_available_output_tokens_from_error
+        return parse_available_output_tokens_from_error(msg)
+
+    # ── Should detect and extract ────────────────────────────────────────
+
+    def test_anthropic_canonical_format(self):
+        """Canonical Anthropic error: max_tokens: X > context_window: Y - input_tokens: Z = available_tokens: W"""
+        msg = (
+            "max_tokens: 32768 > context_window: 200000 "
+            "- input_tokens: 190000 = available_tokens: 10000"
+        )
+        assert self._parse(msg) == 10000
+
+    def test_anthropic_format_large_numbers(self):
+        msg = (
+            "max_tokens: 128000 > context_window: 200000 "
+            "- input_tokens: 180000 = available_tokens: 20000"
+        )
+        assert self._parse(msg) == 20000
+
+    def test_available_tokens_variant_spacing(self):
+        """Handles extra spaces around the colon."""
+        msg = "max_tokens: 32768 > 200000 available_tokens : 5000"
+        assert self._parse(msg) == 5000
+
+    def test_available_tokens_natural_language(self):
+        """'available tokens: N' wording (no underscore)."""
+        msg = "max_tokens must be at most 10000 given your prompt (available tokens: 10000)"
+        assert self._parse(msg) == 10000
+
+    def test_single_token_available(self):
+        """Edge case: only 1 token left."""
+        msg = "max_tokens: 9999 > context_window: 10000 - input_tokens: 9999 = available_tokens: 1"
+        assert self._parse(msg) == 1
+
+    # ── Should NOT detect (returns None) ─────────────────────────────────
+
+    def test_prompt_too_long_is_not_output_cap_error(self):
+        """'prompt is too long' errors must NOT be caught — they need context halving."""
+        msg = "prompt is too long: 205000 tokens > 200000 maximum"
+        assert self._parse(msg) is None
+
+    def test_generic_context_window_exceeded(self):
+        """Generic context window errors without available_tokens should not match."""
+        msg = "context window exceeded: maximum is 32768 tokens"
+        assert self._parse(msg) is None
+
+    def test_context_length_exceeded(self):
+        msg = "context_length_exceeded: prompt has 131073 tokens, limit is 131072"
+        assert self._parse(msg) is None
+
+    def test_no_max_tokens_keyword(self):
+        """Error not related to max_tokens at all."""
+        msg = "invalid_api_key: the API key is invalid"
+        assert self._parse(msg) is None
+
+    def test_empty_string(self):
+        assert self._parse("") is None
+
+    def test_rate_limit_error(self):
+        msg = "rate_limit_error: too many requests per minute"
+        assert self._parse(msg) is None
+
+
+# ---------------------------------------------------------------------------
+# build_anthropic_kwargs — output cap clamping
+# ---------------------------------------------------------------------------
+
+class TestBuildAnthropicKwargsClamping:
+    """The context_length clamp only fires when output ceiling > window.
+    For standard Anthropic models (output ceiling < window) it must not fire.
+    """
+
+    def _build(self, model, max_tokens=None, context_length=None):
+        from agent.anthropic_adapter import build_anthropic_kwargs
+        return build_anthropic_kwargs(
+            model=model,
+            messages=[{"role": "user", "content": "hi"}],
+            tools=None,
+            max_tokens=max_tokens,
+            reasoning_config=None,
+            context_length=context_length,
+        )
+
+    def test_no_clamping_when_output_ceiling_fits_in_window(self):
+        """Opus 4.6 native output (128K) < context window (200K) — no clamping."""
+        kwargs = self._build("claude-opus-4-6", context_length=200_000)
+        assert kwargs["max_tokens"] == 128_000
+
+    def test_clamping_fires_for_tiny_custom_window(self):
+        """When context_length is 8K (local model), output cap is clamped to 7999."""
+        kwargs = self._build("claude-opus-4-6", context_length=8_000)
+        assert kwargs["max_tokens"] == 7_999
+
+    def test_explicit_max_tokens_respected_when_within_window(self):
+        """Explicit max_tokens smaller than window passes through unchanged."""
+        kwargs = self._build("claude-opus-4-6", max_tokens=4096, context_length=200_000)
+        assert kwargs["max_tokens"] == 4096
+
+    def test_explicit_max_tokens_clamped_when_exceeds_window(self):
+        """Explicit max_tokens larger than a small window is clamped."""
+        kwargs = self._build("claude-opus-4-6", max_tokens=32_768, context_length=16_000)
+        assert kwargs["max_tokens"] == 15_999
+
+    def test_no_context_length_uses_native_ceiling(self):
+        """Without context_length the native output ceiling is used directly."""
+        kwargs = self._build("claude-sonnet-4-6")
+        assert kwargs["max_tokens"] == 64_000
+
+
+# ---------------------------------------------------------------------------
+# Ephemeral max_tokens mechanism — _build_api_kwargs
+# ---------------------------------------------------------------------------
+
+class TestEphemeralMaxOutputTokens:
+    """_build_api_kwargs consumes _ephemeral_max_output_tokens exactly once
+    and falls back to self.max_tokens on subsequent calls.
+    """
+
+    def _make_agent(self):
+        """Return a minimal AIAgent with api_mode='anthropic_messages' and
+        a stubbed context_compressor, bypassing full __init__ cost."""
+        from run_agent import AIAgent
+        agent = object.__new__(AIAgent)
+        # Minimal attributes used by _build_api_kwargs
+        agent.api_mode = "anthropic_messages"
+        agent.model = "claude-opus-4-6"
+        agent.tools = []
+        agent.max_tokens = None
+        agent.reasoning_config = None
+        agent._is_anthropic_oauth = False
+        agent._ephemeral_max_output_tokens = None
+
+        compressor = MagicMock()
+        compressor.context_length = 200_000
+        agent.context_compressor = compressor
+
+        # Stub out the internal message-preparation helper
+        agent._prepare_anthropic_messages_for_api = MagicMock(
+            return_value=[{"role": "user", "content": "hi"}]
+        )
+        agent._anthropic_preserve_dots = MagicMock(return_value=False)
+        return agent
+
+    def test_ephemeral_override_is_used_on_first_call(self):
+        """When _ephemeral_max_output_tokens is set, it overrides self.max_tokens."""
+        agent = self._make_agent()
+        agent._ephemeral_max_output_tokens = 5_000
+
+        kwargs = agent._build_api_kwargs([{"role": "user", "content": "hi"}])
+        assert kwargs["max_tokens"] == 5_000
+
+    def test_ephemeral_override_is_consumed_after_one_call(self):
+        """After one call the ephemeral override is cleared to None."""
+        agent = self._make_agent()
+        agent._ephemeral_max_output_tokens = 5_000
+
+        agent._build_api_kwargs([{"role": "user", "content": "hi"}])
+        assert agent._ephemeral_max_output_tokens is None
+
+    def test_subsequent_call_uses_self_max_tokens(self):
+        """A second _build_api_kwargs call uses the normal max_tokens path."""
+        agent = self._make_agent()
+        agent._ephemeral_max_output_tokens = 5_000
+        agent.max_tokens = None  # will resolve to native ceiling (128K for Opus 4.6)
+
+        agent._build_api_kwargs([{"role": "user", "content": "hi"}])
+        # Second call — ephemeral is gone
+        kwargs2 = agent._build_api_kwargs([{"role": "user", "content": "hi"}])
+        assert kwargs2["max_tokens"] == 128_000  # Opus 4.6 native ceiling
+
+    def test_no_ephemeral_uses_self_max_tokens_directly(self):
+        """Without an ephemeral override, self.max_tokens is used normally."""
+        agent = self._make_agent()
+        agent.max_tokens = 8_192
+
+        kwargs = agent._build_api_kwargs([{"role": "user", "content": "hi"}])
+        assert kwargs["max_tokens"] == 8_192
+
+
+# ---------------------------------------------------------------------------
+# Integration: error handler does NOT halve context_length for output-cap errors
+# ---------------------------------------------------------------------------
+
+class TestContextNotHalvedOnOutputCapError:
+    """When the API returns 'max_tokens too large given prompt', the handler
+    must set _ephemeral_max_output_tokens and NOT modify context_length.
+    """
+
+    def _make_agent_with_compressor(self, context_length=200_000):
+        from run_agent import AIAgent
+        from agent.context_compressor import ContextCompressor
+
+        agent = object.__new__(AIAgent)
+        agent.api_mode = "anthropic_messages"
+        agent.model = "claude-opus-4-6"
+        agent.base_url = "https://api.anthropic.com"
+        agent.tools = []
+        agent.max_tokens = None
+        agent.reasoning_config = None
+        agent._is_anthropic_oauth = False
+        agent._ephemeral_max_output_tokens = None
+        agent.log_prefix = ""
+        agent.quiet_mode = True
+        agent.verbose_logging = False
+
+        compressor = MagicMock(spec=ContextCompressor)
+        compressor.context_length = context_length
+        compressor.threshold_percent = 0.75
+        agent.context_compressor = compressor
+
+        agent._prepare_anthropic_messages_for_api = MagicMock(
+            return_value=[{"role": "user", "content": "hi"}]
+        )
+        agent._anthropic_preserve_dots = MagicMock(return_value=False)
+        agent._vprint = MagicMock()
+        return agent
+
+    def test_output_cap_error_sets_ephemeral_not_context_length(self):
+        """On 'max_tokens too large' error, _ephemeral_max_output_tokens is set
+        and compressor.context_length is left unchanged."""
+        from agent.model_metadata import parse_available_output_tokens_from_error
+        from agent.model_metadata import get_next_probe_tier
+
+        error_msg = (
+            "max_tokens: 128000 > context_window: 200000 "
+            "- input_tokens: 180000 = available_tokens: 20000"
+        )
+
+        # Simulate the handler logic from run_agent.py
+        agent = self._make_agent_with_compressor(context_length=200_000)
+        old_ctx = agent.context_compressor.context_length
+
+        available_out = parse_available_output_tokens_from_error(error_msg)
+        assert available_out == 20_000, "parser must detect the error"
+
+        # The fix: set ephemeral, skip context_length modification
+        agent._ephemeral_max_output_tokens = max(1, available_out - 64)
+
+        # context_length must be untouched
+        assert agent.context_compressor.context_length == old_ctx
+        assert agent._ephemeral_max_output_tokens == 19_936
+
+    def test_prompt_too_long_still_triggers_probe_tier(self):
+        """Genuine prompt-too-long errors must still use get_next_probe_tier."""
+        from agent.model_metadata import parse_available_output_tokens_from_error
+        from agent.model_metadata import get_next_probe_tier
+
+        error_msg = "prompt is too long: 205000 tokens > 200000 maximum"
+
+        available_out = parse_available_output_tokens_from_error(error_msg)
+        assert available_out is None, "prompt-too-long must not be caught by output-cap parser"
+
+        # The old halving path is still used for this class of error
+        new_ctx = get_next_probe_tier(200_000)
+        assert new_ctx == 128_000
+
+    def test_output_cap_error_safety_margin(self):
+        """The ephemeral value includes a 64-token safety margin below available_out."""
+        from agent.model_metadata import parse_available_output_tokens_from_error
+
+        error_msg = (
+            "max_tokens: 32768 > context_window: 200000 "
+            "- input_tokens: 190000 = available_tokens: 10000"
+        )
+        available_out = parse_available_output_tokens_from_error(error_msg)
+        safe_out = max(1, available_out - 64)
+        assert safe_out == 9_936
+
+    def test_safety_margin_never_goes_below_one(self):
+        """When available_out is very small, safe_out must be at least 1."""
+        from agent.model_metadata import parse_available_output_tokens_from_error
+
+        error_msg = (
+            "max_tokens: 10 > context_window: 200000 "
+            "- input_tokens: 199990 = available_tokens: 1"
+        )
+        available_out = parse_available_output_tokens_from_error(error_msg)
+        safe_out = max(1, available_out - 64)
+        assert safe_out == 1

tests/test_hermes_state.py

@@ -663,6 +663,84 @@ class TestPruneSessions:
         assert db.get_session("old_cli") is None
         assert db.get_session("old_tg") is not None
 
+    def test_prune_with_multilevel_chain(self, db):
+        """Pruning old sessions orphans newer children instead of crashing on FK."""
+        old_ts = time.time() - 200 * 86400
+        recent_ts = time.time() - 10 * 86400
+
+        # Chain: A (old) -> B (old) -> C (recent) -> D (recent)
+        db.create_session(session_id="A", source="cli")
+        db.end_session("A", end_reason="compressed")
+        db.create_session(session_id="B", source="cli", parent_session_id="A")
+        db.end_session("B", end_reason="compressed")
+        db.create_session(session_id="C", source="cli", parent_session_id="B")
+        db.end_session("C", end_reason="compressed")
+        db.create_session(session_id="D", source="cli", parent_session_id="C")
+        db.end_session("D", end_reason="done")
+
+        # Backdate A and B to be old; C and D stay recent
+        for sid, ts in [("A", old_ts), ("B", old_ts), ("C", recent_ts), ("D", recent_ts)]:
+            db._conn.execute(
+                "UPDATE sessions SET started_at = ? WHERE id = ?", (ts, sid)
+            )
+        db._conn.commit()
+
+        # Should not raise IntegrityError
+        pruned = db.prune_sessions(older_than_days=90)
+        assert pruned == 2  # only A and B
+        assert db.get_session("A") is None
+        assert db.get_session("B") is None
+        # C and D survive, C is orphaned (parent_session_id NULL)
+        c = db.get_session("C")
+        assert c is not None
+        assert c["parent_session_id"] is None
+        d = db.get_session("D")
+        assert d is not None
+        assert d["parent_session_id"] == "C"
+
+    def test_prune_entire_old_chain(self, db):
+        """All sessions in a chain are old — entire chain is pruned."""
+        old_ts = time.time() - 200 * 86400
+
+        db.create_session(session_id="X", source="cli")
+        db.end_session("X", end_reason="compressed")
+        db.create_session(session_id="Y", source="cli", parent_session_id="X")
+        db.end_session("Y", end_reason="compressed")
+        db.create_session(session_id="Z", source="cli", parent_session_id="Y")
+        db.end_session("Z", end_reason="done")
+
+        for sid in ("X", "Y", "Z"):
+            db._conn.execute(
+                "UPDATE sessions SET started_at = ? WHERE id = ?", (old_ts, sid)
+            )
+        db._conn.commit()
+
+        pruned = db.prune_sessions(older_than_days=90)
+        assert pruned == 3
+        for sid in ("X", "Y", "Z"):
+            assert db.get_session(sid) is None
+
+
+class TestDeleteSessionOrphansChildren:
+    def test_delete_orphans_children(self, db):
+        """Deleting a parent session orphans its children."""
+        db.create_session(session_id="parent", source="cli")
+        db.create_session(session_id="child", source="cli", parent_session_id="parent")
+        db.create_session(session_id="grandchild", source="cli", parent_session_id="child")
+
+        # Should not raise IntegrityError
+        result = db.delete_session("parent")
+        assert result is True
+        assert db.get_session("parent") is None
+        # Child is orphaned, not deleted
+        child = db.get_session("child")
+        assert child is not None
+        assert child["parent_session_id"] is None
+        # Grandchild is untouched
+        grandchild = db.get_session("grandchild")
+        assert grandchild is not None
+        assert grandchild["parent_session_id"] == "child"
+
 
 # =========================================================================
 # Schema and WAL mode

tests/tools/test_base_environment.py

@@ -0,0 +1,174 @@
+"""Tests for BaseEnvironment unified execution model.
+
+Tests _wrap_command(), _extract_cwd_from_output(), _embed_stdin_heredoc(),
+init_session() failure handling, and the CWD marker contract.
+"""
+
+import uuid
+from unittest.mock import MagicMock
+
+from tools.environments.base import BaseEnvironment, _cwd_marker
+
+
+class _TestableEnv(BaseEnvironment):
+    """Concrete subclass for testing base class methods."""
+
+    def __init__(self, cwd="/tmp", timeout=10):
+        super().__init__(cwd=cwd, timeout=timeout)
+
+    def _run_bash(self, cmd_string, *, login=False, timeout=120, stdin_data=None):
+        raise NotImplementedError("Use mock")
+
+    def cleanup(self):
+        pass
+
+
+class TestWrapCommand:
+    def test_basic_shape(self):
+        env = _TestableEnv()
+        env._snapshot_ready = True
+        wrapped = env._wrap_command("echo hello", "/tmp")
+
+        assert "source" in wrapped
+        assert "cd /tmp" in wrapped or "cd '/tmp'" in wrapped
+        assert "eval 'echo hello'" in wrapped
+        assert "__hermes_ec=$?" in wrapped
+        assert "export -p >" in wrapped
+        assert "pwd -P >" in wrapped
+        assert env._cwd_marker in wrapped
+        assert "exit $__hermes_ec" in wrapped
+
+    def test_no_snapshot_skips_source(self):
+        env = _TestableEnv()
+        env._snapshot_ready = False
+        wrapped = env._wrap_command("echo hello", "/tmp")
+
+        assert "source" not in wrapped
+
+    def test_single_quote_escaping(self):
+        env = _TestableEnv()
+        env._snapshot_ready = True
+        wrapped = env._wrap_command("echo 'hello world'", "/tmp")
+
+        assert "eval 'echo '\\''hello world'\\'''" in wrapped
+
+    def test_tilde_not_quoted(self):
+        env = _TestableEnv()
+        env._snapshot_ready = True
+        wrapped = env._wrap_command("ls", "~")
+
+        assert "cd ~" in wrapped
+        assert "cd '~'" not in wrapped
+
+    def test_cd_failure_exit_126(self):
+        env = _TestableEnv()
+        env._snapshot_ready = True
+        wrapped = env._wrap_command("ls", "/nonexistent")
+
+        assert "exit 126" in wrapped
+
+
+class TestExtractCwdFromOutput:
+    def test_happy_path(self):
+        env = _TestableEnv()
+        marker = env._cwd_marker
+        result = {
+            "output": f"hello\n{marker}/home/user{marker}\n",
+        }
+        env._extract_cwd_from_output(result)
+
+        assert env.cwd == "/home/user"
+        assert marker not in result["output"]
+
+    def test_missing_marker(self):
+        env = _TestableEnv()
+        result = {"output": "hello world\n"}
+        env._extract_cwd_from_output(result)
+
+        assert env.cwd == "/tmp"  # unchanged
+
+    def test_marker_in_command_output(self):
+        """If the marker appears in command output AND as the real marker,
+        rfind grabs the last (real) one."""
+        env = _TestableEnv()
+        marker = env._cwd_marker
+        result = {
+            "output": f"user typed {marker} in their output\nreal output\n{marker}/correct/path{marker}\n",
+        }
+        env._extract_cwd_from_output(result)
+
+        assert env.cwd == "/correct/path"
+
+    def test_output_cleaned(self):
+        env = _TestableEnv()
+        marker = env._cwd_marker
+        result = {
+            "output": f"hello\n{marker}/tmp{marker}\n",
+        }
+        env._extract_cwd_from_output(result)
+
+        assert "hello" in result["output"]
+        assert marker not in result["output"]
+
+
+class TestEmbedStdinHeredoc:
+    def test_heredoc_format(self):
+        result = BaseEnvironment._embed_stdin_heredoc("cat", "hello world")
+
+        assert result.startswith("cat << '")
+        assert "hello world" in result
+        assert "HERMES_STDIN_" in result
+
+    def test_unique_delimiter_each_call(self):
+        r1 = BaseEnvironment._embed_stdin_heredoc("cat", "data")
+        r2 = BaseEnvironment._embed_stdin_heredoc("cat", "data")
+
+        # Extract delimiters
+        d1 = r1.split("'")[1]
+        d2 = r2.split("'")[1]
+        assert d1 != d2  # UUID-based, should be unique
+
+
+class TestInitSessionFailure:
+    def test_snapshot_ready_false_on_failure(self):
+        env = _TestableEnv()
+
+        def failing_run_bash(*args, **kwargs):
+            raise RuntimeError("bash not found")
+
+        env._run_bash = failing_run_bash
+        env.init_session()
+
+        assert env._snapshot_ready is False
+
+    def test_login_flag_when_snapshot_not_ready(self):
+        """When _snapshot_ready=False, execute() should pass login=True to _run_bash."""
+        env = _TestableEnv()
+        env._snapshot_ready = False
+
+        calls = []
+        def mock_run_bash(cmd, *, login=False, timeout=120, stdin_data=None):
+            calls.append({"login": login})
+            # Return a mock process handle
+            mock = MagicMock()
+            mock.poll.return_value = 0
+            mock.returncode = 0
+            mock.stdout = iter([])
+            return mock
+
+        env._run_bash = mock_run_bash
+        env.execute("echo test")
+
+        assert len(calls) == 1
+        assert calls[0]["login"] is True
+
+
+class TestCwdMarker:
+    def test_marker_contains_session_id(self):
+        env = _TestableEnv()
+        assert env._session_id in env._cwd_marker
+
+    def test_unique_per_instance(self):
+        env1 = _TestableEnv()
+        env2 = _TestableEnv()
+        assert env1._cwd_marker != env2._cwd_marker

tests/tools/test_browser_camofox_persistence.py

@@ -16,6 +16,7 @@ from tools.browser_camofox import (
     _managed_persistence_enabled,
     camofox_close,
     camofox_navigate,
+    camofox_soft_cleanup,
     check_camofox_available,
     cleanup_all_camofox_sessions,
     get_vnc_url,
@@ -240,3 +241,50 @@ class TestVncUrlDiscovery:
 
         assert result["vnc_url"] == "http://localhost:6080"
         assert "vnc_hint" in result
+
+
+class TestCamofoxSoftCleanup:
+    """camofox_soft_cleanup drops local state only when managed persistence is on."""
+
+    def test_returns_true_and_drops_session_when_enabled(self, tmp_path, monkeypatch):
+        monkeypatch.setenv("HERMES_HOME", str(tmp_path))
+        monkeypatch.setenv("CAMOFOX_URL", "http://localhost:9377")
+
+        with _enable_persistence():
+            _get_session("task-1")
+            result = camofox_soft_cleanup("task-1")
+
+        assert result is True
+        # Session should have been dropped from in-memory store
+        import tools.browser_camofox as mod
+        with mod._sessions_lock:
+            assert "task-1" not in mod._sessions
+
+    def test_returns_false_when_disabled(self, tmp_path, monkeypatch):
+        monkeypatch.setenv("HERMES_HOME", str(tmp_path))
+        monkeypatch.setenv("CAMOFOX_URL", "http://localhost:9377")
+
+        _get_session("task-1")
+        config = {"browser": {"camofox": {"managed_persistence": False}}}
+        with patch("tools.browser_camofox.load_config", return_value=config):
+            result = camofox_soft_cleanup("task-1")
+
+        assert result is False
+        # Session should still be present — not dropped
+        import tools.browser_camofox as mod
+        with mod._sessions_lock:
+            assert "task-1" in mod._sessions
+
+    def test_does_not_call_server_delete(self, tmp_path, monkeypatch):
+        """Soft cleanup must never hit the Camofox /sessions DELETE endpoint."""
+        monkeypatch.setenv("HERMES_HOME", str(tmp_path))
+        monkeypatch.setenv("CAMOFOX_URL", "http://localhost:9377")
+
+        with (
+            _enable_persistence(),
+            patch("tools.browser_camofox.requests.delete") as mock_delete,
+        ):
+            _get_session("task-1")
+            camofox_soft_cleanup("task-1")
+
+        mock_delete.assert_not_called()

tests/tools/test_browser_camofox_state.py

@@ -63,4 +63,4 @@ class TestCamofoxConfigDefaults:
         from hermes_cli.config import DEFAULT_CONFIG
 
         # managed_persistence is auto-merged by _deep_merge, no version bump needed
-        assert DEFAULT_CONFIG["_config_version"] == 12
+        assert DEFAULT_CONFIG["_config_version"] == 13

tests/tools/test_browser_cleanup.py

@@ -65,6 +65,62 @@ class TestBrowserCleanup:
         mock_stop.assert_called_once_with("task-1")
         mock_run.assert_called_once_with("task-1", "close", [], timeout=10)
 
+    def test_cleanup_camofox_managed_persistence_skips_close(self):
+        """When camofox mode + managed persistence, soft_cleanup fires instead of close."""
+        browser_tool = self.browser_tool
+        browser_tool._active_sessions["task-1"] = {
+            "session_name": "sess-1",
+            "bb_session_id": None,
+        }
+        browser_tool._session_last_activity["task-1"] = 123.0
+
+        with (
+            patch("tools.browser_tool._is_camofox_mode", return_value=True),
+            patch("tools.browser_tool._maybe_stop_recording") as mock_stop,
+            patch(
+                "tools.browser_tool._run_browser_command",
+                return_value={"success": True},
+            ),
+            patch("tools.browser_tool.os.path.exists", return_value=False),
+            patch(
+                "tools.browser_camofox.camofox_soft_cleanup",
+                return_value=True,
+            ) as mock_soft,
+            patch("tools.browser_camofox.camofox_close") as mock_close,
+        ):
+            browser_tool.cleanup_browser("task-1")
+
+        mock_soft.assert_called_once_with("task-1")
+        mock_close.assert_not_called()
+
+    def test_cleanup_camofox_no_persistence_calls_close(self):
+        """When camofox mode but managed persistence is off, camofox_close fires."""
+        browser_tool = self.browser_tool
+        browser_tool._active_sessions["task-1"] = {
+            "session_name": "sess-1",
+            "bb_session_id": None,
+        }
+        browser_tool._session_last_activity["task-1"] = 123.0
+
+        with (
+            patch("tools.browser_tool._is_camofox_mode", return_value=True),
+            patch("tools.browser_tool._maybe_stop_recording") as mock_stop,
+            patch(
+                "tools.browser_tool._run_browser_command",
+                return_value={"success": True},
+            ),
+            patch("tools.browser_tool.os.path.exists", return_value=False),
+            patch(
+                "tools.browser_camofox.camofox_soft_cleanup",
+                return_value=False,
+            ) as mock_soft,
+            patch("tools.browser_camofox.camofox_close") as mock_close,
+        ):
+            browser_tool.cleanup_browser("task-1")
+
+        mock_soft.assert_called_once_with("task-1")
+        mock_close.assert_called_once_with("task-1")
+
     def test_emergency_cleanup_clears_all_tracking_state(self):
         browser_tool = self.browser_tool
         browser_tool._cleanup_done = False