fix(plugins): register dynamically-loaded modules in sys.modules before exec

Dashboard plugin API routes (web_server._mount_plugin_api_routes) and
gateway event hooks (gateway.hooks.HookRegistry.discover_and_load) both
loaded Python files via importlib.util.spec_from_file_location +
exec_module without registering the resulting module in sys.modules.

That breaks any plugin or hook handler that uses `from __future__ import
annotations` together with a Pydantic BaseModel / dataclass / anything
that introspects `__module__`: at first request Pydantic tries to
resolve string-form type hints against the defining module's namespace,
can't find it by name, and raises:

  PydanticUserError: TypeAdapter[...] is not fully defined;
  you should define ... and all referenced types,
  then call `.rebuild()` on the instance.

This is what broke the kanban dashboard's 'triage' button — POST
/api/plugins/kanban/tasks validated against CreateTaskBody (a Pydantic
model in a file using `from __future__ import annotations`) and
returned 500 on every click.

The fix, applied symmetrically to both loaders:

  1. Compute module_name once.
  2. Register the module in sys.modules BEFORE exec_module.
  3. On exec_module failure, pop the half-initialized stub so subsequent
     reloads don't pick up broken state.

GETs were unaffected because they don't build a body TypeAdapter, which
is why this only surfaced when users started POSTing.

agent/anthropic_adapter.py

@@ -390,7 +390,16 @@ def build_anthropic_client(api_key: str, base_url: str = None, timeout: float =
         "timeout": Timeout(timeout=float(_read_timeout), connect=10.0),
     }
     if normalized_base_url:
-        kwargs["base_url"] = normalized_base_url
+        # Azure Anthropic endpoints require an ``api-version`` query parameter.
+        # Pass it via default_query so the SDK appends it to every request URL
+        # without corrupting the base_url (appending it directly produces
+        # malformed paths like /anthropic?api-version=.../v1/messages).
+        _is_azure_endpoint = "azure.com" in normalized_base_url.lower()
+        if _is_azure_endpoint and "api-version" not in normalized_base_url:
+            kwargs["base_url"] = normalized_base_url.rstrip("/")
+            kwargs["default_query"] = {"api-version": "2025-04-15"}
+        else:
+            kwargs["base_url"] = normalized_base_url
     common_betas = _common_betas_for_base_url(normalized_base_url)
 
     if _is_kimi_coding_endpoint(base_url):
@@ -1680,9 +1689,9 @@ def build_anthropic_kwargs(
 
     # ── Strip sampling params on 4.7+ ─────────────────────────────────
     # Opus 4.7 rejects any non-default temperature/top_p/top_k with a 400.
-    # Callers (auxiliary_client, flush_memories, etc.) may set these for
-    # older models; drop them here as a safety net so upstream 4.6 → 4.7
-    # migrations don't require coordinated edits everywhere.
+    # Callers (auxiliary_client, etc.) may set these for older models;
+    # drop them here as a safety net so upstream 4.6 → 4.7 migrations
+    # don't require coordinated edits everywhere.
     if _forbids_sampling_params(model):
         for _sampling_key in ("temperature", "top_p", "top_k"):
             kwargs.pop(_sampling_key, None)

agent/auxiliary_client.py

@@ -42,6 +42,7 @@ import time
 from pathlib import Path  # noqa: F401 — used by test mocks
 from types import SimpleNamespace
 from typing import Any, Dict, List, Optional, Tuple
+from urllib.parse import urlparse, parse_qs, urlunparse
 
 from openai import OpenAI
 
@@ -52,6 +53,17 @@ from utils import base_url_host_matches, base_url_hostname, normalize_proxy_env_
 
 logger = logging.getLogger(__name__)
 
+
+def _extract_url_query_params(url: str):
+    """Extract query params from URL, return (clean_url, default_query dict or None)."""
+    parsed = urlparse(url)
+    if parsed.query:
+        clean = urlunparse(parsed._replace(query=""))
+        params = {k: v[0] for k, v in parse_qs(parsed.query).items()}
+        return clean, params
+    return url, None
+
+
 # Module-level flag: only warn once per process about stale OPENAI_BASE_URL.
 _stale_base_url_warned = False
 
@@ -390,7 +402,7 @@ class _CodexCompletionsAdapter:
         # Note: the Codex endpoint (chatgpt.com/backend-api/codex) does NOT
         # support max_output_tokens or temperature — omit to avoid 400 errors.
 
-        # Tools support for flush_memories and similar callers
+        # Tools support for auxiliary callers (e.g. skills_hub) that pass function schemas
         tools = kwargs.get("tools")
         if tools:
             converted = []
@@ -1157,8 +1169,10 @@ def _try_custom_endpoint() -> Tuple[Optional[Any], Optional[str]]:
         return None, None
     model = _read_main_model() or "gpt-4o-mini"
     logger.debug("Auxiliary client: custom endpoint (%s, api_mode=%s)", model, custom_mode or "chat_completions")
+    _clean_base, _dq = _extract_url_query_params(custom_base)
+    _extra = {"default_query": _dq} if _dq else {}
     if custom_mode == "codex_responses":
-        real_client = OpenAI(api_key=custom_key, base_url=custom_base)
+        real_client = OpenAI(api_key=custom_key, base_url=_clean_base, **_extra)
         return CodexAuxiliaryClient(real_client, model), model
     if custom_mode == "anthropic_messages":
         # Third-party Anthropic-compatible gateway (MiniMax, Zhipu GLM,
@@ -1172,12 +1186,12 @@ def _try_custom_endpoint() -> Tuple[Optional[Any], Optional[str]]:
                 "Custom endpoint declares api_mode=anthropic_messages but the "
                 "anthropic SDK is not installed — falling back to OpenAI-wire."
             )
-            return OpenAI(api_key=custom_key, base_url=custom_base), model
+            return OpenAI(api_key=custom_key, base_url=_clean_base, **_extra), model
         return (
             AnthropicAuxiliaryClient(real_client, model, custom_key, custom_base, is_oauth=False),
             model,
         )
-    return OpenAI(api_key=custom_key, base_url=custom_base), model
+    return OpenAI(api_key=custom_key, base_url=_clean_base, **_extra), model
 
 
 def _try_codex() -> Tuple[Optional[Any], Optional[str]]:
@@ -1349,6 +1363,49 @@ def _is_auth_error(exc: Exception) -> bool:
     return "error code: 401" in err_lower or "authenticationerror" in type(exc).__name__.lower()
 
 
+def _is_unsupported_parameter_error(exc: Exception, param: str) -> bool:
+    """Detect provider 400s for an unsupported request parameter.
+
+    Different OpenAI-compatible endpoints phrase the same class of error a few
+    ways: ``Unsupported parameter: X``, ``unsupported_parameter`` with a
+    ``param`` field, ``X is not supported``, ``unknown parameter: X``,
+    ``unrecognized request argument: X``.  We match on both the parameter
+    name and a generic "unsupported/unknown/unrecognized parameter" marker so
+    call sites can reactively retry without the offending key instead of
+    surfacing a noisy auxiliary failure.
+
+    Generalizes the temperature-specific detector that originally shipped
+    with PR #15621 so the same retry strategy can cover ``max_tokens``,
+    ``seed``, ``top_p``, and any future quirk. Credit @nicholasrae (PR #15416)
+    for the generalization pattern.
+    """
+    param_lower = (param or "").lower()
+    if not param_lower:
+        return False
+    err_lower = str(exc).lower()
+    if param_lower not in err_lower:
+        return False
+    return any(marker in err_lower for marker in (
+        "unsupported parameter",
+        "unsupported_parameter",
+        "not supported",
+        "does not support",
+        "unknown parameter",
+        "unrecognized request argument",
+        "unrecognized parameter",
+        "invalid parameter",
+    ))
+
+
+def _is_unsupported_temperature_error(exc: Exception) -> bool:
+    """Back-compat wrapper: detect API errors where the model rejects ``temperature``.
+
+    Delegates to :func:`_is_unsupported_parameter_error`; kept as a separate
+    public symbol because existing tests and call sites import it by name.
+    """
+    return _is_unsupported_parameter_error(exc, "temperature")
+
+
 def _evict_cached_clients(provider: str) -> None:
     """Drop cached auxiliary clients for a provider so fresh creds are used."""
     normalized = _normalize_aux_provider(provider)
@@ -1782,12 +1839,15 @@ def resolve_provider_client(
                 provider,
             )
             extra = {}
+            _clean_base, _dq = _extract_url_query_params(custom_base)
+            if _dq:
+                extra["default_query"] = _dq
             if base_url_host_matches(custom_base, "api.kimi.com"):
                 extra["default_headers"] = {"User-Agent": "claude-code/0.1.0"}
             elif base_url_host_matches(custom_base, "api.githubcopilot.com"):
                 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)
+            client = OpenAI(api_key=custom_key, base_url=_clean_base, **extra)
             client = _wrap_if_needed(client, final_model, custom_base)
             return (_to_async_client(client, final_model) if async_mode
                     else (client, final_model))
@@ -1824,6 +1884,8 @@ def resolve_provider_client(
                     model or custom_entry.get("model") or _read_main_model() or "gpt-4o-mini",
                     provider,
                 )
+                _clean_base2, _dq2 = _extract_url_query_params(custom_base)
+                _extra2 = {"default_query": _dq2} if _dq2 else {}
                 logger.debug(
                     "resolve_provider_client: named custom provider %r (%s, api_mode=%s)",
                     provider, final_model, entry_api_mode or "chat_completions")
@@ -1841,7 +1903,7 @@ def resolve_provider_client(
                             "installed — falling back to OpenAI-wire.",
                             provider,
                         )
-                        client = OpenAI(api_key=custom_key, base_url=custom_base)
+                        client = OpenAI(api_key=custom_key, base_url=_clean_base2, **_extra2)
                         return (_to_async_client(client, final_model) if async_mode
                                 else (client, final_model))
                     sync_anthropic = AnthropicAuxiliaryClient(
@@ -1850,7 +1912,7 @@ def resolve_provider_client(
                     if async_mode:
                         return AsyncAnthropicAuxiliaryClient(sync_anthropic), final_model
                     return sync_anthropic, final_model
-                client = OpenAI(api_key=custom_key, base_url=custom_base)
+                client = OpenAI(api_key=custom_key, base_url=_clean_base2, **_extra2)
                 # codex_responses or inherited auto-detect (via _wrap_if_needed).
                 # _wrap_if_needed reads the closed-over `api_mode` (the task-level
                 # override). Named-provider entry api_mode=codex_responses also
@@ -2760,8 +2822,8 @@ def _build_call_kwargs(
         temperature = fixed_temperature
 
     # Opus 4.7+ rejects any non-default temperature/top_p/top_k — silently
-    # drop here so auxiliary callers that hardcode temperature (e.g. 0.3 on
-    # flush_memories, 0 on structured-JSON extraction) don't 400 the moment
+    # drop here so auxiliary callers that hardcode temperature (e.g. 0 on
+    # structured-JSON extraction) don't 400 the moment
     # the aux model is flipped to 4.7.
     if temperature is not None:
         from agent.anthropic_adapter import _forbids_sampling_params
@@ -2849,7 +2911,7 @@ def call_llm(
 
     Args:
         task: Auxiliary task name ("compression", "vision", "web_extract",
-              "session_search", "skills_hub", "mcp", "flush_memories").
+              "session_search", "skills_hub", "mcp", "title_generation").
               Reads provider:model from config/env. Ignored if provider is set.
         provider: Explicit provider override.
         model: Explicit model override.
@@ -2952,13 +3014,45 @@ def call_llm(
     if _is_anthropic_compat_endpoint(resolved_provider, _client_base):
         kwargs["messages"] = _convert_openai_images_to_anthropic(kwargs["messages"])
 
-    # Handle max_tokens vs max_completion_tokens retry, then payment fallback.
+    # Handle unsupported temperature, max_tokens vs max_completion_tokens retry,
+    # then payment fallback.
     try:
         return _validate_llm_response(
             client.chat.completions.create(**kwargs), task)
     except Exception as first_err:
+        if "temperature" in kwargs and _is_unsupported_temperature_error(first_err):
+            retry_kwargs = dict(kwargs)
+            retry_kwargs.pop("temperature", None)
+            logger.info(
+                "Auxiliary %s: provider rejected temperature; retrying once without it",
+                task or "call",
+            )
+            try:
+                return _validate_llm_response(
+                    client.chat.completions.create(**retry_kwargs), task)
+            except Exception as retry_err:
+                retry_err_str = str(retry_err)
+                # If retry still fails, fall through to the max_tokens /
+                # payment / auth chains below using the temperature-stripped
+                # kwargs.  Re-raise only if the retry hit something those
+                # chains won't handle.
+                if not (
+                    _is_payment_error(retry_err)
+                    or _is_connection_error(retry_err)
+                    or _is_auth_error(retry_err)
+                    or "max_tokens" in retry_err_str
+                    or "unsupported_parameter" in retry_err_str
+                ):
+                    raise
+                first_err = retry_err
+                kwargs = retry_kwargs
+
         err_str = str(first_err)
-        if "max_tokens" in err_str or "unsupported_parameter" in err_str:
+        if max_tokens is not None and (
+            "max_tokens" in err_str
+            or "unsupported_parameter" in err_str
+            or _is_unsupported_parameter_error(first_err, "max_tokens")
+        ):
             kwargs.pop("max_tokens", None)
             kwargs["max_completion_tokens"] = max_tokens
             try:
@@ -3221,8 +3315,35 @@ async def async_call_llm(
         return _validate_llm_response(
             await client.chat.completions.create(**kwargs), task)
     except Exception as first_err:
+        if "temperature" in kwargs and _is_unsupported_temperature_error(first_err):
+            retry_kwargs = dict(kwargs)
+            retry_kwargs.pop("temperature", None)
+            logger.info(
+                "Auxiliary %s (async): provider rejected temperature; retrying once without it",
+                task or "call",
+            )
+            try:
+                return _validate_llm_response(
+                    await client.chat.completions.create(**retry_kwargs), task)
+            except Exception as retry_err:
+                retry_err_str = str(retry_err)
+                if not (
+                    _is_payment_error(retry_err)
+                    or _is_connection_error(retry_err)
+                    or _is_auth_error(retry_err)
+                    or "max_tokens" in retry_err_str
+                    or "unsupported_parameter" in retry_err_str
+                ):
+                    raise
+                first_err = retry_err
+                kwargs = retry_kwargs
+
         err_str = str(first_err)
-        if "max_tokens" in err_str or "unsupported_parameter" in err_str:
+        if max_tokens is not None and (
+            "max_tokens" in err_str
+            or "unsupported_parameter" in err_str
+            or _is_unsupported_parameter_error(first_err, "max_tokens")
+        ):
             kwargs.pop("max_tokens", None)
             kwargs["max_completion_tokens"] = max_tokens
             try:

agent/codex_responses_adapter.py

@@ -44,22 +44,31 @@ _TOOL_CALL_LEAK_PATTERN = re.compile(
 # Multimodal content helpers
 # ---------------------------------------------------------------------------
 
-def _chat_content_to_responses_parts(content: Any) -> List[Dict[str, Any]]:
+def _chat_content_to_responses_parts(content: Any, *, role: str = "user") -> List[Dict[str, Any]]:
     """Convert chat-style multimodal content to Responses API input parts.
 
     Input:  ``[{"type":"text"|"image_url", ...}]`` (native OpenAI Chat format)
-    Output: ``[{"type":"input_text"|"input_image", ...}]`` (Responses format)
+    Output: ``[{"type":"input_text"|"output_text"|"input_image", ...}]`` (Responses format)
+
+    The ``role`` parameter controls the text content type:
+    - ``"user"`` (default) → ``"input_text"``
+    - ``"assistant"`` → ``"output_text"``
+
+    The Responses API rejects ``input_text`` inside assistant messages and
+    ``output_text`` inside user messages, so callers MUST pass the correct
+    role for the message being converted.
 
     Returns an empty list when ``content`` is not a list or contains no
     recognized parts — callers fall back to the string path.
     """
+    text_type = "output_text" if role == "assistant" else "input_text"
     if not isinstance(content, list):
         return []
     converted: List[Dict[str, Any]] = []
     for part in content:
         if isinstance(part, str):
             if part:
-                converted.append({"type": "input_text", "text": part})
+                converted.append({"type": text_type, "text": part})
             continue
         if not isinstance(part, dict):
             continue
@@ -67,7 +76,7 @@ def _chat_content_to_responses_parts(content: Any) -> List[Dict[str, Any]]:
         if ptype in {"text", "input_text", "output_text"}:
             text = part.get("text")
             if isinstance(text, str) and text:
-                converted.append({"type": "input_text", "text": text})
+                converted.append({"type": text_type, "text": text})
             continue
         if ptype in {"image_url", "input_image"}:
             image_ref = part.get("image_url")
@@ -218,6 +227,23 @@ def _responses_tools(tools: Optional[List[Dict[str, Any]]] = None) -> Optional[L
 # Message format conversion
 # ---------------------------------------------------------------------------
 
+_RESPONSE_MESSAGE_STATUSES = {"completed", "incomplete", "in_progress"}
+
+
+def _normalize_responses_message_status(value: Any, *, default: str = "completed") -> str:
+    """Normalize a Responses assistant message status for replay.
+
+    The API accepts completed/incomplete/in_progress on replayed assistant
+    output messages.  Preserve those exactly (modulo case/hyphen spelling) so
+    incomplete Codex continuation turns don't get falsely marked completed.
+    """
+    if isinstance(value, str):
+        status = value.strip().lower().replace("-", "_").replace(" ", "_")
+        if status in _RESPONSE_MESSAGE_STATUSES:
+            return status
+    return default
+
+
 def _chat_messages_to_responses_input(messages: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
     """Convert internal chat-style messages to Responses input items."""
     items: List[Dict[str, Any]] = []
@@ -233,9 +259,10 @@ def _chat_messages_to_responses_input(messages: List[Dict[str, Any]]) -> List[Di
         if role in {"user", "assistant"}:
             content = msg.get("content", "")
             if isinstance(content, list):
-                content_parts = _chat_content_to_responses_parts(content)
+                content_parts = _chat_content_to_responses_parts(content, role=role)
+                text_type = "output_text" if role == "assistant" else "input_text"
                 content_text = "".join(
-                    p.get("text", "") for p in content_parts if p.get("type") == "input_text"
+                    p.get("text", "") for p in content_parts if p.get("type") == text_type
                 )
             else:
                 content_parts = []
@@ -262,7 +289,57 @@ def _chat_messages_to_responses_input(messages: List[Dict[str, Any]]) -> List[Di
                                 seen_item_ids.add(item_id)
                             has_codex_reasoning = True
 
-                if content_parts:
+                # Replay exact assistant message items (with id/phase) from
+                # previous turns so the API can maintain prefix-cache hits.
+                # OpenAI docs: "preserve and resend phase on all assistant
+                # messages — dropping it can degrade performance."
+                codex_message_items = msg.get("codex_message_items")
+                replayed_message_items = 0
+                if isinstance(codex_message_items, list):
+                    for raw_item in codex_message_items:
+                        if not isinstance(raw_item, dict):
+                            continue
+                        if raw_item.get("type") != "message" or raw_item.get("role") != "assistant":
+                            continue
+                        raw_content_parts = raw_item.get("content")
+                        if not isinstance(raw_content_parts, list):
+                            continue
+
+                        normalized_content_parts = []
+                        for part in raw_content_parts:
+                            if not isinstance(part, dict):
+                                continue
+                            part_type = str(part.get("type") or "").strip()
+                            if part_type not in {"output_text", "text"}:
+                                continue
+                            text = part.get("text", "")
+                            if text is None:
+                                text = ""
+                            if not isinstance(text, str):
+                                text = str(text)
+                            normalized_content_parts.append({"type": "output_text", "text": text})
+
+                        if not normalized_content_parts:
+                            continue
+
+                        replay_item = {
+                            "type": "message",
+                            "role": "assistant",
+                            "status": _normalize_responses_message_status(raw_item.get("status")),
+                            "content": normalized_content_parts,
+                        }
+                        item_id = raw_item.get("id")
+                        if isinstance(item_id, str) and item_id.strip():
+                            replay_item["id"] = item_id.strip()
+                        phase = raw_item.get("phase")
+                        if isinstance(phase, str) and phase.strip():
+                            replay_item["phase"] = phase.strip()
+                        items.append(replay_item)
+                        replayed_message_items += 1
+
+                if replayed_message_items > 0:
+                    pass
+                elif content_parts:
                     items.append({"role": "assistant", "content": content_parts})
                 elif content_text.strip():
                     items.append({"role": "assistant", "content": content_text})
@@ -422,6 +499,47 @@ def _preflight_codex_input_items(raw_items: Any) -> List[Dict[str, Any]]:
                 normalized.append(reasoning_item)
             continue
 
+        if item_type == "message":
+            role = item.get("role")
+            if role != "assistant":
+                raise ValueError(f"Codex Responses input[{idx}] message items must have role='assistant'.")
+            content = item.get("content")
+            if not isinstance(content, list):
+                raise ValueError(f"Codex Responses input[{idx}] message item must have content list.")
+            normalized_content = []
+            for part_idx, part in enumerate(content):
+                if not isinstance(part, dict):
+                    raise ValueError(
+                        f"Codex Responses input[{idx}] message content[{part_idx}] must be an object."
+                    )
+                part_type = part.get("type")
+                if part_type not in {"output_text", "text"}:
+                    raise ValueError(
+                        f"Codex Responses input[{idx}] message content[{part_idx}] has unsupported type {part_type!r}."
+                    )
+                text = part.get("text", "")
+                if text is None:
+                    text = ""
+                if not isinstance(text, str):
+                    text = str(text)
+                normalized_content.append({"type": "output_text", "text": text})
+            if not normalized_content:
+                raise ValueError(f"Codex Responses input[{idx}] message item must contain at least one text part.")
+            normalized_item: Dict[str, Any] = {
+                "type": "message",
+                "role": "assistant",
+                "status": _normalize_responses_message_status(item.get("status")),
+                "content": normalized_content,
+            }
+            item_id = item.get("id")
+            if isinstance(item_id, str) and item_id.strip():
+                normalized_item["id"] = item_id.strip()
+            phase = item.get("phase")
+            if isinstance(phase, str) and phase.strip():
+                normalized_item["phase"] = phase.strip()
+            normalized.append(normalized_item)
+            continue
+
         role = item.get("role")
         if role in {"user", "assistant"}:
             content = item.get("content", "")
@@ -429,13 +547,16 @@ def _preflight_codex_input_items(raw_items: Any) -> List[Dict[str, Any]]:
                 content = ""
             if isinstance(content, list):
                 # Multimodal content from ``_chat_messages_to_responses_input``
-                # is already in Responses format (``input_text`` / ``input_image``).
-                # Validate each part and pass through.
+                # is already in Responses format (``input_text`` / ``output_text``
+                # / ``input_image``).  Validate each part and pass through.
+                # Use the correct text type for the role — ``output_text`` for
+                # assistant messages, ``input_text`` for user messages.
+                text_type = "output_text" if role == "assistant" else "input_text"
                 validated: List[Dict[str, Any]] = []
                 for part_idx, part in enumerate(content):
                     if isinstance(part, str):
                         if part:
-                            validated.append({"type": "input_text", "text": part})
+                            validated.append({"type": text_type, "text": part})
                         continue
                     if not isinstance(part, dict):
                         raise ValueError(
@@ -446,7 +567,7 @@ def _preflight_codex_input_items(raw_items: Any) -> List[Dict[str, Any]]:
                         text = part.get("text", "")
                         if not isinstance(text, str):
                             text = str(text or "")
-                        validated.append({"type": "input_text", "text": text})
+                        validated.append({"type": text_type, "text": text})
                     elif ptype in {"input_image", "image_url"}:
                         image_ref = part.get("image_url", "")
                         detail = part.get("detail")
@@ -703,6 +824,7 @@ def _normalize_codex_response(response: Any) -> tuple[Any, str]:
     content_parts: List[str] = []
     reasoning_parts: List[str] = []
     reasoning_items_raw: List[Dict[str, Any]] = []
+    message_items_raw: List[Dict[str, Any]] = []
     tool_calls: List[Any] = []
     has_incomplete_items = response_status in {"queued", "in_progress", "incomplete"}
     saw_commentary_phase = False
@@ -721,6 +843,7 @@ def _normalize_codex_response(response: Any) -> tuple[Any, str]:
 
         if item_type == "message":
             item_phase = getattr(item, "phase", None)
+            normalized_phase = None
             if isinstance(item_phase, str):
                 normalized_phase = item_phase.strip().lower()
                 if normalized_phase in {"commentary", "analysis"}:
@@ -730,6 +853,18 @@ def _normalize_codex_response(response: Any) -> tuple[Any, str]:
             message_text = _extract_responses_message_text(item)
             if message_text:
                 content_parts.append(message_text)
+                raw_message_item: Dict[str, Any] = {
+                    "type": "message",
+                    "role": "assistant",
+                    "status": _normalize_responses_message_status(item_status),
+                    "content": [{"type": "output_text", "text": message_text}],
+                }
+                item_id = getattr(item, "id", None)
+                if isinstance(item_id, str) and item_id:
+                    raw_message_item["id"] = item_id
+                if normalized_phase:
+                    raw_message_item["phase"] = normalized_phase
+                message_items_raw.append(raw_message_item)
         elif item_type == "reasoning":
             reasoning_text = _extract_responses_reasoning_text(item)
             if reasoning_text:
@@ -842,6 +977,7 @@ def _normalize_codex_response(response: Any) -> tuple[Any, str]:
         reasoning_content=None,
         reasoning_details=None,
         codex_reasoning_items=reasoning_items_raw or None,
+        codex_message_items=message_items_raw or None,
     )
 
     if tool_calls:

agent/context_compressor.py

@@ -318,6 +318,13 @@ class ContextCompressor(ContextEngine):
             int(context_length * self.threshold_percent),
             MINIMUM_CONTEXT_LENGTH,
         )
+        # Recalculate token budgets for the new context length so the
+        # compressor stays calibrated after a model switch (e.g. 200K → 32K).
+        target_tokens = int(self.threshold_tokens * self.summary_target_ratio)
+        self.tail_token_budget = target_tokens
+        self.max_summary_tokens = min(
+            int(context_length * 0.05), _SUMMARY_TOKENS_CEILING,
+        )
 
     def __init__(
         self,

agent/model_metadata.py

@@ -106,9 +106,11 @@ _endpoint_model_metadata_cache_time: Dict[str, float] = {}
 _ENDPOINT_MODEL_CACHE_TTL = 300
 
 # Descending tiers for context length probing when the model is unknown.
-# We start at 128K (a safe default for most modern models) and step down
-# on context-length errors until one works.
+# We start at 256K (covers GPT-5.x, many current large-context models) and
+# step down on context-length errors until one works.  Tier[0] is also the
+# default fallback when no detection method succeeds.
 CONTEXT_PROBE_TIERS = [
+    256_000,
     128_000,
     64_000,
     32_000,
@@ -143,10 +145,11 @@ DEFAULT_CONTEXT_LENGTHS = {
     "claude": 200000,
     # OpenAI — GPT-5 family (most have 400k; specific overrides first)
     # Source: https://developers.openai.com/api/docs/models
-    # GPT-5.5 (launched Apr 23 2026). 400k is the fallback for providers we
-    # can't probe live. ChatGPT Codex OAuth actually caps lower (272k as of
-    # Apr 2026) and is resolved via _resolve_codex_oauth_context_length().
-    "gpt-5.5": 400000,
+    # GPT-5.5 (launched Apr 23 2026) is 1.05M on the direct OpenAI API and
+    # ChatGPT Codex OAuth caps it at 272K; both paths resolve via their own
+    # provider-aware branches (_resolve_codex_oauth_context_length + models.dev).
+    # This hardcoded value is only reached when every probe misses.
+    "gpt-5.5": 1050000,
     "gpt-5.4-nano": 400000,           # 400k (not 1.05M like full 5.4)
     "gpt-5.4-mini": 400000,           # 400k (not 1.05M like full 5.4)
     "gpt-5.4": 1050000,               # GPT-5.4, GPT-5.4 Pro (1.05M context)
@@ -162,7 +165,17 @@ DEFAULT_CONTEXT_LENGTHS = {
     "gemma-4-31b": 256000,
     "gemma-3": 131072,
     "gemma": 8192,  # fallback for older gemma models
-    # DeepSeek
+    # DeepSeek — V4 family ships with a 1M context window. The legacy
+    # aliases ``deepseek-chat`` / ``deepseek-reasoner`` are server-side
+    # mapped to the non-thinking / thinking modes of ``deepseek-v4-flash``
+    # and inherit the same 1M window. The ``deepseek`` substring entry
+    # below remains as a 128K fallback for older / unknown DeepSeek model
+    # ids (e.g. via custom endpoints).
+    # https://api-docs.deepseek.com/zh-cn/quick_start/pricing
+    "deepseek-v4-pro": 1_000_000,
+    "deepseek-v4-flash": 1_000_000,
+    "deepseek-chat": 1_000_000,
+    "deepseek-reasoner": 1_000_000,
     "deepseek": 128000,
     # Meta
     "llama": 131072,
@@ -1193,6 +1206,7 @@ def get_model_context_length(
     api_key: str = "",
     config_context_length: int | None = None,
     provider: str = "",
+    custom_providers: list | None = None,
 ) -> int:
     """Get the context length for a model.
 
@@ -1213,6 +1227,23 @@ def get_model_context_length(
     if config_context_length is not None and isinstance(config_context_length, int) and config_context_length > 0:
         return config_context_length
 
+    # 0b. custom_providers per-model override — check before any probe.
+    # This closes the gap where /model switch and display paths used to fall
+    # back to 128K despite the user having a per-model context_length set.
+    # See #15779.
+    if custom_providers and base_url and model:
+        try:
+            from hermes_cli.config import get_custom_provider_context_length
+            cp_ctx = get_custom_provider_context_length(
+                model=model,
+                base_url=base_url,
+                custom_providers=custom_providers,
+            )
+            if cp_ctx:
+                return cp_ctx
+        except Exception:
+            pass  # fall through to probing
+
     # Normalise provider-prefixed model names (e.g. "local:model-name" →
     # "model-name") so cache lookups and server queries use the bare ID that
     # local servers actually know about.  Ollama "model:tag" colons are preserved.
@@ -1352,7 +1383,7 @@ def get_model_context_length(
     # 6. OpenRouter live API metadata (provider-unaware fallback)
     metadata = fetch_model_metadata()
     if model in metadata:
-        return metadata[model].get("context_length", 128000)
+        return metadata[model].get("context_length", DEFAULT_FALLBACK_CONTEXT)
 
     # 8. Hardcoded defaults (fuzzy match — longest key first for specificity)
     # Only check `default_model in model` (is the key a substring of the input).

agent/nous_rate_guard.py

@@ -180,3 +180,145 @@ def format_remaining(seconds: float) -> str:
     h, remainder = divmod(s, 3600)
     m = remainder // 60
     return f"{h}h {m}m" if m else f"{h}h"
+
+
+# Buckets with reset windows shorter than this are treated as transient
+# (upstream jitter, secondary throttling) rather than a genuine quota
+# exhaustion worth a cross-session breaker trip.
+_MIN_RESET_FOR_BREAKER_SECONDS = 60.0
+
+
+def is_genuine_nous_rate_limit(
+    *,
+    headers: Optional[Mapping[str, str]] = None,
+    last_known_state: Optional[Any] = None,
+) -> bool:
+    """Decide whether a 429 from Nous Portal is a real account rate limit.
+
+    Nous Portal multiplexes multiple upstream providers (DeepSeek, Kimi,
+    MiMo, Hermes, ...) behind one endpoint.  A 429 can mean either:
+
+      (a) The caller's own RPM / RPH / TPM / TPH bucket on Nous is
+          exhausted — a genuine rate limit that will last until the
+          bucket resets.
+      (b) The upstream provider is out of capacity for a specific model
+          — transient, clears in seconds, and has nothing to do with
+          the caller's quota on Nous.
+
+    Tripping the cross-session breaker on (b) blocks ALL Nous requests
+    (and all models, since Nous is one provider key) for minutes even
+    though the caller's account is healthy and a different model would
+    have worked.  That's the bug users hit when DeepSeek V4 Pro 429s
+    trigger a breaker that then blocks Kimi 2.6 and MiMo V2.5 Pro.
+
+    We tell the two apart by looking at:
+
+      1. The 429 response's own ``x-ratelimit-*`` headers.  Nous emits
+         the full suite on every response including 429s.  An exhausted
+         bucket (``remaining == 0`` with a reset window >= 60s) is
+         proof of (a).
+      2. The last-known-good rate-limit state captured by
+         ``_capture_rate_limits()`` on the previous successful
+         response.  If any bucket there was already near-exhausted with
+         a substantial reset window, the current 429 is almost
+         certainly (a) continuing from that condition.
+
+    If neither signal fires, we treat the 429 as (b): fail the single
+    request, let the retry loop or model-switch proceed, and do NOT
+    write the cross-session breaker file.
+
+    Returns True when the evidence points at (a).
+    """
+    # Signal 1: current 429 response headers.
+    state = _parse_buckets_from_headers(headers)
+    if _has_exhausted_bucket(state):
+        return True
+
+    # Signal 2: last-known-good state from a recent successful response.
+    # Accepts either a RateLimitState (dataclass from rate_limit_tracker)
+    # or a dict of bucket snapshots.
+    if last_known_state is not None and _has_exhausted_bucket_in_object(last_known_state):
+        return True
+
+    return False
+
+
+def _parse_buckets_from_headers(
+    headers: Optional[Mapping[str, str]],
+) -> dict[str, tuple[Optional[int], Optional[float]]]:
+    """Extract (remaining, reset_seconds) per bucket from x-ratelimit-* headers.
+
+    Returns empty dict when no rate-limit headers are present.
+    """
+    if not headers:
+        return {}
+
+    lowered = {k.lower(): v for k, v in headers.items()}
+    if not any(k.startswith("x-ratelimit-") for k in lowered):
+        return {}
+
+    def _maybe_int(raw: Optional[str]) -> Optional[int]:
+        if raw is None:
+            return None
+        try:
+            return int(float(raw))
+        except (TypeError, ValueError):
+            return None
+
+    def _maybe_float(raw: Optional[str]) -> Optional[float]:
+        if raw is None:
+            return None
+        try:
+            return float(raw)
+        except (TypeError, ValueError):
+            return None
+
+    result: dict[str, tuple[Optional[int], Optional[float]]] = {}
+    for tag in ("requests", "requests-1h", "tokens", "tokens-1h"):
+        remaining = _maybe_int(lowered.get(f"x-ratelimit-remaining-{tag}"))
+        reset = _maybe_float(lowered.get(f"x-ratelimit-reset-{tag}"))
+        if remaining is not None or reset is not None:
+            result[tag] = (remaining, reset)
+    return result
+
+
+def _has_exhausted_bucket(
+    buckets: Mapping[str, tuple[Optional[int], Optional[float]]],
+) -> bool:
+    """Return True when any bucket has remaining == 0 AND a meaningful reset window."""
+    for remaining, reset in buckets.values():
+        if remaining is None or remaining > 0:
+            continue
+        if reset is None:
+            continue
+        if reset >= _MIN_RESET_FOR_BREAKER_SECONDS:
+            return True
+    return False
+
+
+def _has_exhausted_bucket_in_object(state: Any) -> bool:
+    """Check a RateLimitState-like object for an exhausted bucket.
+
+    Accepts the dataclass from ``agent.rate_limit_tracker`` (buckets
+    exposed as attributes ``requests_min``, ``requests_hour``,
+    ``tokens_min``, ``tokens_hour``) and falls back gracefully for any
+    object missing those attributes.
+    """
+    for attr in ("requests_min", "requests_hour", "tokens_min", "tokens_hour"):
+        bucket = getattr(state, attr, None)
+        if bucket is None:
+            continue
+        limit = getattr(bucket, "limit", 0) or 0
+        remaining = getattr(bucket, "remaining", 0) or 0
+        # Prefer the adjusted "remaining_seconds_now" property when present;
+        # fall back to raw reset_seconds.
+        reset = getattr(bucket, "remaining_seconds_now", None)
+        if reset is None:
+            reset = getattr(bucket, "reset_seconds", 0.0) or 0.0
+        if limit <= 0:
+            continue
+        if remaining > 0:
+            continue
+        if reset >= _MIN_RESET_FOR_BREAKER_SECONDS:
+            return True
+    return False

agent/onboarding.py

@@ -0,0 +1,144 @@
+"""
+Contextual first-touch onboarding hints.
+
+Instead of blocking first-run questionnaires, show a one-time hint the *first*
+time a user hits a behavior fork — message-while-running, first long-running
+tool, etc.  Each hint is shown once per install (tracked in ``config.yaml`` under
+``onboarding.seen.<flag>``) and then never again.
+
+Keep this module tiny and dependency-free so both the CLI and gateway can import
+it without pulling in heavy modules.
+"""
+
+from __future__ import annotations
+
+import logging
+from pathlib import Path
+from typing import Any, Mapping, Optional
+
+logger = logging.getLogger(__name__)
+
+
+# -------------------------------------------------------------------------
+# Flag names (stable — used as config.yaml keys under onboarding.seen)
+# -------------------------------------------------------------------------
+
+BUSY_INPUT_FLAG = "busy_input_prompt"
+TOOL_PROGRESS_FLAG = "tool_progress_prompt"
+
+
+# -------------------------------------------------------------------------
+# Hint content
+# -------------------------------------------------------------------------
+
+def busy_input_hint_gateway(mode: str) -> str:
+    """Hint shown the first time a user messages while the agent is busy.
+
+    ``mode`` is the effective busy_input_mode that was just applied, so the
+    message matches reality ("I just interrupted…" vs "I just queued…").
+    """
+    if mode == "queue":
+        return (
+            "💡 First-time tip — I queued your message instead of interrupting. "
+            "Send `/busy interrupt` to make new messages stop the current task "
+            "immediately, or `/busy status` to check. This notice won't appear again."
+        )
+    return (
+        "💡 First-time tip — I just interrupted my current task to answer you. "
+        "Send `/busy queue` to queue follow-ups for after the current task instead, "
+        "or `/busy status` to check. This notice won't appear again."
+    )
+
+
+def busy_input_hint_cli(mode: str) -> str:
+    """CLI version of the busy-input hint (plain text, no markdown)."""
+    if mode == "queue":
+        return (
+            "(tip) Your message was queued for the next turn. "
+            "Use /busy interrupt to make Enter stop the current run instead. "
+            "This tip only shows once."
+        )
+    return (
+        "(tip) Your message interrupted the current run. "
+        "Use /busy queue to queue messages for the next turn instead. "
+        "This tip only shows once."
+    )
+
+
+def tool_progress_hint_gateway() -> str:
+    return (
+        "💡 First-time tip — that tool took a while and I'm streaming every step. "
+        "If the progress messages feel noisy, send `/verbose` to cycle modes "
+        "(all → new → off). This notice won't appear again."
+    )
+
+
+def tool_progress_hint_cli() -> str:
+    return (
+        "(tip) That tool ran for a while. Use /verbose to cycle tool-progress "
+        "display modes (all -> new -> off -> verbose). This tip only shows once."
+    )
+
+
+# -------------------------------------------------------------------------
+# State read / write
+# -------------------------------------------------------------------------
+
+def _get_seen_dict(config: Mapping[str, Any]) -> Mapping[str, Any]:
+    onboarding = config.get("onboarding") if isinstance(config, Mapping) else None
+    if not isinstance(onboarding, Mapping):
+        return {}
+    seen = onboarding.get("seen")
+    return seen if isinstance(seen, Mapping) else {}
+
+
+def is_seen(config: Mapping[str, Any], flag: str) -> bool:
+    """Return True if the user has already been shown this first-touch hint."""
+    return bool(_get_seen_dict(config).get(flag))
+
+
+def mark_seen(config_path: Path, flag: str) -> bool:
+    """Persist ``onboarding.seen.<flag> = True`` to ``config_path``.
+
+    Uses the atomic YAML writer so a concurrent process can't observe a
+    partially-written file.  Returns True on success, False on any error
+    (including the config file being absent — onboarding is best-effort).
+    """
+    try:
+        import yaml
+        from utils import atomic_yaml_write
+    except Exception as e:  # pragma: no cover — dependency issue
+        logger.debug("onboarding: failed to import yaml/utils: %s", e)
+        return False
+
+    try:
+        cfg: dict = {}
+        if config_path.exists():
+            with open(config_path, encoding="utf-8") as f:
+                cfg = yaml.safe_load(f) or {}
+        if not isinstance(cfg.get("onboarding"), dict):
+            cfg["onboarding"] = {}
+        seen = cfg["onboarding"].get("seen")
+        if not isinstance(seen, dict):
+            seen = {}
+            cfg["onboarding"]["seen"] = seen
+        if seen.get(flag) is True:
+            return True  # already marked — nothing to do
+        seen[flag] = True
+        atomic_yaml_write(config_path, cfg)
+        return True
+    except Exception as e:
+        logger.debug("onboarding: failed to mark flag %s: %s", flag, e)
+        return False
+
+
+__all__ = [
+    "BUSY_INPUT_FLAG",
+    "TOOL_PROGRESS_FLAG",
+    "busy_input_hint_gateway",
+    "busy_input_hint_cli",
+    "tool_progress_hint_gateway",
+    "tool_progress_hint_cli",
+    "is_seen",
+    "mark_seen",
+]

agent/prompt_builder.py

@@ -176,6 +176,64 @@ SKILLS_GUIDANCE = (
     "Skills that aren't maintained become liabilities."
 )
 
+KANBAN_GUIDANCE = (
+    "# You are a Kanban worker\n"
+    "You were spawned by the Hermes Kanban dispatcher to execute ONE task from "
+    "the shared board at `~/.hermes/kanban.db`. Your task id is in "
+    "`$HERMES_KANBAN_TASK`; your workspace is `$HERMES_KANBAN_WORKSPACE`. "
+    "The `kanban_*` tools in your schema are your primary coordination surface — "
+    "they write directly to the shared SQLite DB and work regardless of terminal "
+    "backend (local/docker/modal/ssh).\n"
+    "\n"
+    "## Lifecycle\n"
+    "\n"
+    "1. **Orient.** Call `kanban_show()` first (no args — it defaults to your "
+    "task). The response includes title, body, parent-task handoffs (summary + "
+    "metadata), any prior attempts on this task if you're a retry, the full "
+    "comment thread, and a pre-formatted `worker_context` you can treat as "
+    "ground truth.\n"
+    "2. **Work inside the workspace.** `cd $HERMES_KANBAN_WORKSPACE` before "
+    "any file operations. The workspace is yours for this run. Don't modify "
+    "files outside it unless the task explicitly asks.\n"
+    "3. **Heartbeat on long operations.** Call `kanban_heartbeat(note=...)` "
+    "every few minutes during long subprocesses (training, encoding, crawling). "
+    "Skip heartbeats for short tasks.\n"
+    "4. **Block on genuine ambiguity.** If you need a human decision you cannot "
+    "infer (missing credentials, UX choice, paywalled source, peer output you "
+    "need first), call `kanban_block(reason=\"...\")` and stop. Don't guess. "
+    "The user will unblock with context and the dispatcher will respawn you.\n"
+    "5. **Complete with structured handoff.** Call `kanban_complete(summary=..., "
+    "metadata=...)`. `summary` is 1–3 human-readable sentences naming concrete "
+    "artifacts. `metadata` is machine-readable facts "
+    "(`{changed_files: [...], tests_run: N, decisions: [...]}`). Downstream "
+    "workers read both via their own `kanban_show`. Never put secrets / "
+    "tokens / raw PII in either field — run rows are durable forever.\n"
+    "6. **If follow-up work appears, create it; don't do it.** Use "
+    "`kanban_create(title=..., assignee=<right-profile>, parents=[your-task-id])` "
+    "to spawn a child task for the appropriate specialist profile instead of "
+    "scope-creeping into the next thing.\n"
+    "\n"
+    "## Orchestrator mode\n"
+    "\n"
+    "If your task is itself a decomposition task (e.g. a planner profile given "
+    "a high-level goal), use `kanban_create` to fan out into child tasks — one "
+    "per specialist, each with an explicit `assignee` and `parents=[...]` to "
+    "express dependencies. Then `kanban_complete` your own task with a summary "
+    "of the decomposition. Do NOT execute the work yourself; your job is "
+    "routing, not implementation.\n"
+    "\n"
+    "## Do NOT\n"
+    "\n"
+    "- Do not shell out to `hermes kanban <verb>` for board operations. Use "
+    "the `kanban_*` tools — they work across all terminal backends.\n"
+    "- Do not complete a task you didn't actually finish. Block it.\n"
+    "- Do not assign follow-up work to yourself. Assign it to the right "
+    "specialist profile.\n"
+    "- Do not call `delegate_task` as a board substitute. `delegate_task` is "
+    "for short reasoning subtasks inside your own run; board tasks are for "
+    "cross-agent handoffs that outlive one API loop."
+)
+
 TOOL_USE_ENFORCEMENT_GUIDANCE = (
     "# Tool-use enforcement\n"
     "You MUST use your tools to take action — do not describe what you would do "

agent/transports/__init__.py

@@ -23,9 +23,14 @@ def get_transport(api_mode: str):
     This allows gradual migration — call sites can check for None
     and fall back to the legacy code path.
     """
-    if not _REGISTRY:
-        _discover_transports()
     cls = _REGISTRY.get(api_mode)
+    if cls is None:
+        # The registry can be partially populated when a specific transport
+        # module was imported directly (for example chat_completions before
+        # codex).  Discover on misses, not only when the registry is empty, so
+        # test/order-dependent imports do not make valid api_modes unavailable.
+        _discover_transports()
+        cls = _REGISTRY.get(api_mode)
     if cls is None:
         return None
     return cls()

agent/transports/chat_completions.py

@@ -31,15 +31,15 @@ class ChatCompletionsTransport(ProviderTransport):
     def convert_messages(self, messages: List[Dict[str, Any]], **kwargs) -> List[Dict[str, Any]]:
         """Messages are already in OpenAI format — sanitize Codex leaks only.
 
-        Strips Codex Responses API fields (``codex_reasoning_items`` on the
-        message, ``call_id``/``response_item_id`` on tool_calls) that strict
-        chat-completions providers reject with 400/422.
+        Strips Codex Responses API fields (``codex_reasoning_items`` /
+        ``codex_message_items`` on the message, ``call_id``/``response_item_id``
+        on tool_calls) that strict chat-completions providers reject with 400/422.
         """
         needs_sanitize = False
         for msg in messages:
             if not isinstance(msg, dict):
                 continue
-            if "codex_reasoning_items" in msg:
+            if "codex_reasoning_items" in msg or "codex_message_items" in msg:
                 needs_sanitize = True
                 break
             tool_calls = msg.get("tool_calls")
@@ -59,6 +59,7 @@ class ChatCompletionsTransport(ProviderTransport):
             if not isinstance(msg, dict):
                 continue
             msg.pop("codex_reasoning_items", None)
+            msg.pop("codex_message_items", None)
             tool_calls = msg.get("tool_calls")
             if isinstance(tool_calls, list):
                 for tc in tool_calls:

agent/transports/codex.py

@@ -120,6 +120,24 @@ class ResponsesApiTransport(ProviderTransport):
         if request_overrides:
             kwargs.update(request_overrides)
 
+        if is_codex_backend:
+            prompt_cache_key = kwargs.get("prompt_cache_key")
+            cache_scope_id = str(prompt_cache_key or session_id or "").strip()
+            if cache_scope_id:
+                existing_extra_headers = kwargs.get("extra_headers")
+                merged_extra_headers: Dict[str, str] = {}
+                if isinstance(existing_extra_headers, dict):
+                    merged_extra_headers.update(
+                        {
+                            str(key): str(value)
+                            for key, value in existing_extra_headers.items()
+                            if key and value is not None
+                        }
+                    )
+                merged_extra_headers["session_id"] = cache_scope_id
+                merged_extra_headers["x-client-request-id"] = cache_scope_id
+                kwargs["extra_headers"] = merged_extra_headers
+
         max_tokens = params.get("max_tokens")
         if max_tokens is not None and not is_codex_backend:
             kwargs["max_output_tokens"] = max_tokens
@@ -160,6 +178,8 @@ class ResponsesApiTransport(ProviderTransport):
         provider_data = {}
         if msg and hasattr(msg, "codex_reasoning_items") and msg.codex_reasoning_items:
             provider_data["codex_reasoning_items"] = msg.codex_reasoning_items
+        if msg and hasattr(msg, "codex_message_items") and msg.codex_message_items:
+            provider_data["codex_message_items"] = msg.codex_message_items
         if msg and hasattr(msg, "reasoning_details") and msg.reasoning_details:
             provider_data["reasoning_details"] = msg.reasoning_details
 

agent/transports/types.py

@@ -97,7 +97,7 @@ class NormalizedResponse:
     Response-level ``provider_data`` examples:
 
     * Anthropic: ``{"reasoning_details": [...]}``
-    * Codex: ``{"codex_reasoning_items": [...]}``
+    * Codex: ``{"codex_reasoning_items": [...], "codex_message_items": [...]}``
     * Others: ``None``
     """
 
@@ -126,6 +126,11 @@ class NormalizedResponse:
         pd = self.provider_data or {}
         return pd.get("codex_reasoning_items")
 
+    @property
+    def codex_message_items(self):
+        pd = self.provider_data or {}
+        return pd.get("codex_message_items")
+
 
 # ---------------------------------------------------------------------------
 # Factory helpers

cli-config.yaml.example

@@ -790,9 +790,16 @@ code_execution:
 # Supports single tasks and batch mode (default 3 parallel, configurable).
 delegation:
   max_iterations: 50                          # Max tool-calling turns per child (default: 50)
-  # max_concurrent_children: 3                # Max parallel child agents (default: 3)
-  # max_spawn_depth: 1                        # Tree depth cap (1-3, default: 1 = flat). Raise to 2 or 3 to allow orchestrator children to spawn their own workers.
+  # max_concurrent_children: 3                # Max parallel child agents per batch (default: 3, floor: 1, no ceiling).
+                                              # WARNING: values above 10 multiply API cost linearly.
+  # max_spawn_depth: 1                        # Delegation tree depth cap (range: 1-3, default: 1 = flat).
+                                              # Raise to 2 to allow workers to spawn their own subagents.
+                                              # Requires role="orchestrator" on intermediate agents.
   # orchestrator_enabled: true                # Kill switch for role="orchestrator" children (default: true).
+  # subagent_auto_approve: false              # When a subagent hits a dangerous-command approval prompt, auto-deny (default: false)
+                                              # or auto-approve "once" (true) instead of blocking on stdin.
+                                              # The parent TUI owns stdin, so blocking would deadlock; non-interactive resolution is required.
+                                              # Both choices emit a logger.warning audit line. Flip to true only for cron/batch pipelines.
   # inherit_mcp_toolsets: true                # When explicit child toolsets are narrowed, also keep the parent's MCP toolsets (default: true). Set false for strict intersection.
   # model: "google/gemini-3-flash-preview"    # Override model for subagents (empty = inherit parent)
   # provider: "openrouter"                    # Override provider for subagents (empty = inherit parent)
@@ -817,7 +824,9 @@ delegation:
 # Display
 # =============================================================================
 display:
-  # Use compact banner mode
+  # Use compact banner mode (hides the ASCII-art banner, shows a single line).
+  #   true:  Compact single-line banner
+  #   false: Full ASCII banner with tool/skill summary (default)
   compact: false
 
   # Tool progress display level (CLI and gateway)
@@ -831,12 +840,15 @@ display:
   # Gateway-only natural mid-turn assistant updates.
   # When true, completed assistant status messages are sent as separate chat
   # messages. This is independent of tool_progress and gateway streaming.
+  #   true:  Send mid-turn assistant updates as separate messages (default)
+  #   false: Only send the final response
   interim_assistant_messages: true
 
-  # What Enter does when Hermes is already busy in the CLI.
+  # What Enter does when Hermes is already busy (CLI and gateway platforms).
   #   interrupt: Interrupt the current run and redirect Hermes (default)
   #   queue:     Queue your message for the next turn
-  # Ctrl+C always interrupts regardless of this setting.
+  # Ctrl+C (or /stop in gateway) always interrupts regardless of this setting.
+  # Toggle at runtime with /busy_input_mode <interrupt|queue>.
   busy_input_mode: interrupt
 
   # Background process notifications (gateway/messaging only).
@@ -852,17 +864,22 @@ display:
   # Play terminal bell when agent finishes a response.
   # Useful for long-running tasks — your terminal will ding when the agent is done.
   # Works over SSH. Most terminals can be configured to flash the taskbar or play a sound.
+  #   true:  Ring the terminal bell on each response
+  #   false: Silent (default)
   bell_on_complete: false
 
   # Show model reasoning/thinking before each response.
   # When enabled, a dim box shows the model's thought process above the response.
   # Toggle at runtime with /reasoning show or /reasoning hide.
+  #   true:  Show the reasoning box
+  #   false: Hide reasoning (default)
   show_reasoning: false
 
   # Stream tokens to the terminal as they arrive instead of waiting for the
   # full response. The response box opens on first token and text appears
   # line-by-line. Tool calls are still captured silently.
-  # Stream tokens to the terminal in real-time. Disable to wait for full responses.
+  #   true:  Stream tokens as they arrive (default)
+  #   false: Wait for the full response before rendering
   streaming: true
 
   # ───────────────────────────────────────────────────────────────────────────
@@ -872,10 +889,15 @@ display:
   # response box label, and branding text. Change at runtime with /skin <name>.
   #
   # Built-in skins:
-  #   default  — Classic Hermes gold/kawaii
-  #   ares     — Crimson/bronze war-god theme with spinner wings
-  #   mono     — Clean grayscale monochrome
-  #   slate    — Cool blue developer-focused
+  #   default        — Classic Hermes gold/kawaii
+  #   ares           — Crimson/bronze war-god theme with spinner wings
+  #   mono           — Clean grayscale monochrome
+  #   slate          — Cool blue developer-focused
+  #   daylight       — Bright light-mode theme
+  #   warm-lightmode — Warm paper-tone light-mode theme
+  #   poseidon       — Sea-green/teal Olympian theme
+  #   sisyphus       — Earthy stone-and-moss theme
+  #   charizard      — Fiery orange dragon theme
   #
   # Custom skins: drop a YAML file in ~/.hermes/skins/<name>.yaml
   # Schema (all fields optional, missing values inherit from default):

cli.py

@@ -22,6 +22,7 @@ import re
 import concurrent.futures
 import base64
 import atexit
+import errno
 import tempfile
 import time
 import uuid
@@ -416,6 +417,11 @@ def load_cli_config() -> Dict[str, Any]:
             "base_url": "",    # Direct OpenAI-compatible endpoint for subagents
             "api_key": "",     # API key for delegation.base_url (falls back to OPENAI_API_KEY)
         },
+        "onboarding": {
+            # First-touch hint flags (see agent/onboarding.py).  Each hint is
+            # shown once per install then latched here.
+            "seen": {},
+        },
     }
     
     # Track whether the config file explicitly set terminal config.
@@ -3176,7 +3182,14 @@ class HermesCLI:
         # the configured model (e.g. "qwen3.6-plus"), causing 400 errors.
         runtime_model = runtime.get("model")
         if runtime_model and isinstance(runtime_model, str):
-            self.model = runtime_model
+            # Only use runtime model if: model is unset, or model equals provider name
+            should_use_runtime_model = (
+                not self.model or  # No model configured yet
+                self.model == self.provider or  # Model is the provider slug
+                self.model == runtime.get("name")  # Model matches provider display name
+            )
+            if should_use_runtime_model:
+                self.model = runtime_model
 
         # If model is still empty (e.g. user ran `hermes auth add openai-codex`
         # without `hermes model`), fall back to the provider's first catalog
@@ -4311,7 +4324,7 @@ class HermesCLI:
 
         _cprint(f"\n  {_DIM}Tip: Just type your message to chat with Hermes!{_RST}")
         _cprint(f"  {_DIM}Multi-line: Alt+Enter for a new line{_RST}")
-        _cprint(f"  {_DIM}Draft editor: Ctrl+G{_RST}")
+        _cprint(f"  {_DIM}Draft editor: Ctrl+G (Alt+G in VSCode/Cursor){_RST}")
         if _is_termux_environment():
             _cprint(f"  {_DIM}Attach image: /image {_termux_example_image_path()} or start your prompt with a local image path{_RST}\n")
         else:
@@ -4661,10 +4674,6 @@ class HermesCLI:
     def new_session(self, silent=False):
         """Start a fresh session with a new session ID and cleared agent state."""
         if self.agent and self.conversation_history:
-            try:
-                self.agent.flush_memories(self.conversation_history)
-            except (Exception, KeyboardInterrupt):
-                pass
             # Trigger memory extraction on the old session before session_id rotates.
             self.agent.commit_memory_session(self.conversation_history)
             self._notify_session_boundary("on_session_finalize")
@@ -5149,27 +5158,29 @@ class HermesCLI:
         _cprint(f"  ✓ Model switched: {result.new_model}")
         _cprint(f"    Provider: {provider_label}")
 
+        # Context: always resolve via the provider-aware chain so Codex OAuth,
+        # Copilot, and Nous-enforced caps win over the raw models.dev entry
+        # (e.g. gpt-5.5 is 1.05M on openai but 272K on Codex OAuth).
         mi = result.model_info
+        try:
+            from hermes_cli.model_switch import resolve_display_context_length
+            ctx = resolve_display_context_length(
+                result.new_model,
+                result.target_provider,
+                base_url=result.base_url or self.base_url or "",
+                api_key=result.api_key or self.api_key or "",
+                model_info=mi,
+            )
+            if ctx:
+                _cprint(f"    Context: {ctx:,} tokens")
+        except Exception:
+            pass
         if mi:
-            if mi.context_window:
-                _cprint(f"    Context: {mi.context_window:,} tokens")
             if mi.max_output:
                 _cprint(f"    Max output: {mi.max_output:,} tokens")
             if mi.has_cost_data():
                 _cprint(f"    Cost: {mi.format_cost()}")
             _cprint(f"    Capabilities: {mi.format_capabilities()}")
-        else:
-            try:
-                from agent.model_metadata import get_model_context_length
-                ctx = get_model_context_length(
-                    result.new_model,
-                    base_url=result.base_url or self.base_url,
-                    api_key=result.api_key or self.api_key,
-                    provider=result.target_provider,
-                )
-                _cprint(f"    Context: {ctx:,} tokens")
-            except Exception:
-                pass
 
         cache_enabled = (
             (base_url_host_matches(result.base_url or "", "openrouter.ai") and "claude" in result.new_model.lower())
@@ -5270,24 +5281,22 @@ class HermesCLI:
         # Parse --provider and --global flags
         model_input, explicit_provider, persist_global = parse_model_flags(raw_args)
 
+        # Load providers for switch_model (picker path needs them below)
         user_provs = None
         custom_provs = None
+        try:
+            from hermes_cli.config import get_compatible_custom_providers, load_config
+            cfg = load_config()
+            user_provs = cfg.get("providers")
+            custom_provs = get_compatible_custom_providers(cfg)
+        except Exception:
+            pass
 
         # No args at all: open prompt_toolkit-native picker modal
         if not model_input and not explicit_provider:
             model_display = self.model or "unknown"
             provider_display = get_label(self.provider) if self.provider else "unknown"
 
-            user_provs = None
-            custom_provs = None
-            try:
-                from hermes_cli.config import get_compatible_custom_providers, load_config
-                cfg = load_config()
-                user_provs = cfg.get("providers")
-                custom_provs = get_compatible_custom_providers(cfg)
-            except Exception:
-                pass
-
             try:
                 providers = list_authenticated_providers(
                     current_provider=self.provider or "",
@@ -5374,29 +5383,26 @@ class HermesCLI:
         _cprint(f"  ✓ Model switched: {result.new_model}")
         _cprint(f"    Provider: {provider_label}")
 
-        # Rich metadata from models.dev
+        # Context: always resolve via the provider-aware chain so Codex OAuth,
+        # Copilot, and Nous-enforced caps win over the raw models.dev entry
+        # (e.g. gpt-5.5 is 1.05M on openai but 272K on Codex OAuth).
         mi = result.model_info
+        from hermes_cli.model_switch import resolve_display_context_length
+        ctx = resolve_display_context_length(
+            result.new_model,
+            result.target_provider,
+            base_url=result.base_url or self.base_url or "",
+            api_key=result.api_key or self.api_key or "",
+            model_info=mi,
+        )
+        if ctx:
+            _cprint(f"    Context: {ctx:,} tokens")
         if mi:
-            if mi.context_window:
-                _cprint(f"    Context: {mi.context_window:,} tokens")
             if mi.max_output:
                 _cprint(f"    Max output: {mi.max_output:,} tokens")
             if mi.has_cost_data():
                 _cprint(f"    Cost: {mi.format_cost()}")
             _cprint(f"    Capabilities: {mi.format_capabilities()}")
-        else:
-            # Fallback to old context length lookup
-            try:
-                from agent.model_metadata import get_model_context_length
-                ctx = get_model_context_length(
-                    result.new_model,
-                    base_url=result.base_url or self.base_url,
-                    api_key=result.api_key or self.api_key,
-                    provider=result.target_provider,
-                )
-                _cprint(f"    Context: {ctx:,} tokens")
-            except Exception:
-                pass
 
         # Cache notice
         cache_enabled = (
@@ -5812,7 +5818,28 @@ class HermesCLI:
 
         print(f"(._.) Unknown cron command: {subcommand}")
         print("  Available: list, add, edit, pause, resume, run, remove")
-    
+
+    def _handle_kanban_command(self, cmd: str):
+        """Handle the /kanban command — delegate to the shared kanban CLI.
+
+        The string form passed here is the user's full ``/kanban ...``
+        including the leading slash; we strip it and hand the remainder
+        to ``kanban.run_slash`` which returns a single formatted string.
+        """
+        from hermes_cli.kanban import run_slash
+
+        rest = cmd.strip()
+        if rest.startswith("/"):
+            rest = rest.lstrip("/")
+        if rest.startswith("kanban"):
+            rest = rest[len("kanban"):].lstrip()
+        try:
+            output = run_slash(rest)
+        except Exception as exc:  # pragma: no cover - defensive
+            output = f"(._.) kanban error: {exc}"
+        if output:
+            print(output)
+
     def _handle_skills_command(self, cmd: str):
         """Handle /skills slash command — delegates to hermes_cli.skills_hub."""
         from hermes_cli.skills_hub import handle_skills_slash
@@ -6049,6 +6076,8 @@ class HermesCLI:
             self.save_conversation()
         elif canonical == "cron":
             self._handle_cron_command(cmd_original)
+        elif canonical == "kanban":
+            self._handle_kanban_command(cmd_original)
         elif canonical == "skills":
             with self._busy_command(self._slow_command_status(cmd_original)):
                 self._handle_skills_command(cmd_original)
@@ -6123,8 +6152,6 @@ class HermesCLI:
             self._handle_agents_command()
         elif canonical == "background":
             self._handle_background_command(cmd_original)
-        elif canonical == "btw":
-            self._handle_btw_command(cmd_original)
         elif canonical == "queue":
             # Extract prompt after "/queue " or "/q "
             parts = cmd_original.split(None, 1)
@@ -6411,122 +6438,6 @@ class HermesCLI:
         self._background_tasks[task_id] = thread
         thread.start()
 
-    def _handle_btw_command(self, cmd: str):
-        """Handle /btw <question> — ephemeral side question using session context.
-
-        Snapshots the current conversation history, spawns a no-tools agent in
-        a background thread, and prints the answer without persisting anything
-        to the main session.
-        """
-        parts = cmd.strip().split(maxsplit=1)
-        if len(parts) < 2 or not parts[1].strip():
-            _cprint("  Usage: /btw <question>")
-            _cprint("  Example: /btw what module owns session title sanitization?")
-            _cprint("  Answers using session context. No tools, not persisted.")
-            return
-
-        question = parts[1].strip()
-        task_id = f"btw_{datetime.now().strftime('%H%M%S')}_{uuid.uuid4().hex[:6]}"
-
-        if not self._ensure_runtime_credentials():
-            _cprint("  (>_<) Cannot start /btw: no valid credentials.")
-            return
-
-        turn_route = self._resolve_turn_agent_config(question)
-        history_snapshot = list(self.conversation_history)
-
-        preview = question[:60] + ("..." if len(question) > 60 else "")
-        _cprint(f'  💬 /btw: "{preview}"')
-
-        def run_btw():
-            try:
-                btw_agent = AIAgent(
-                    model=turn_route["model"],
-                    api_key=turn_route["runtime"].get("api_key"),
-                    base_url=turn_route["runtime"].get("base_url"),
-                    provider=turn_route["runtime"].get("provider"),
-                    api_mode=turn_route["runtime"].get("api_mode"),
-                    acp_command=turn_route["runtime"].get("command"),
-                    acp_args=turn_route["runtime"].get("args"),
-                    max_iterations=8,
-                    enabled_toolsets=[],
-                    quiet_mode=True,
-                    verbose_logging=False,
-                    session_id=task_id,
-                    platform="cli",
-                    reasoning_config=self.reasoning_config,
-                    service_tier=self.service_tier,
-                    request_overrides=turn_route.get("request_overrides"),
-                    providers_allowed=self._providers_only,
-                    providers_ignored=self._providers_ignore,
-                    providers_order=self._providers_order,
-                    provider_sort=self._provider_sort,
-                    provider_require_parameters=self._provider_require_params,
-                    provider_data_collection=self._provider_data_collection,
-                    fallback_model=self._fallback_model,
-                    session_db=None,
-                    skip_memory=True,
-                    skip_context_files=True,
-                    persist_session=False,
-                )
-
-                btw_prompt = (
-                    "[Ephemeral /btw side question. Answer using the conversation "
-                    "context. No tools available. Be direct and concise.]\n\n"
-                    + question
-                )
-                result = btw_agent.run_conversation(
-                    user_message=btw_prompt,
-                    conversation_history=history_snapshot,
-                    task_id=task_id,
-                )
-
-                response = (result.get("final_response") or "") if result else ""
-                if not response and result and result.get("error"):
-                    response = f"Error: {result['error']}"
-
-                # TUI refresh before printing
-                if self._app:
-                    self._app.invalidate()
-                    time.sleep(0.05)
-                print()
-
-                if response:
-                    try:
-                        from hermes_cli.skin_engine import get_active_skin
-                        _skin = get_active_skin()
-                        _resp_color = _skin.get_color("response_border", "#4F6D4A")
-                    except Exception:
-                        _resp_color = "#4F6D4A"
-
-                    ChatConsole().print(Panel(
-                        _render_final_assistant_content(response, mode=self.final_response_markdown),
-                        title=f"[{_resp_color} bold]⚕ /btw[/]",
-                        title_align="left",
-                        border_style=_resp_color,
-                        box=rich_box.HORIZONTALS,
-                        padding=(1, 4),
-                    ))
-                else:
-                    _cprint("  💬 /btw: (no response)")
-
-                if self.bell_on_complete:
-                    sys.stdout.write("\a")
-                    sys.stdout.flush()
-
-            except Exception as e:
-                if self._app:
-                    self._app.invalidate()
-                    time.sleep(0.05)
-                print()
-                _cprint(f"  ❌ /btw failed: {e}")
-            finally:
-                if self._app:
-                    self._invalidate(min_interval=0)
-
-        thread = threading.Thread(target=run_btw, daemon=True, name=f"btw-{task_id}")
-        thread.start()
-
     @staticmethod
     def _try_launch_chrome_debug(port: int, system: str) -> bool:
         """Try to launch Chrome/Chromium with remote debugging enabled.
@@ -7411,6 +7322,31 @@ class HermesCLI:
                     _cprint(f"  {line}")
                 except Exception:
                     pass
+                # First-touch onboarding: on the first tool in this process
+                # that takes longer than the threshold while we're in the
+                # noisiest progress mode, print a one-time hint about
+                # /verbose.  Latched on self so it fires at most once per
+                # process; persisted to config.yaml so it never fires again
+                # across processes either.
+                try:
+                    if (
+                        not getattr(self, "_long_tool_hint_fired", False)
+                        and self.tool_progress_mode == "all"
+                        and duration >= 30.0
+                    ):
+                        from agent.onboarding import (
+                            TOOL_PROGRESS_FLAG,
+                            is_seen,
+                            mark_seen,
+                            tool_progress_hint_cli,
+                        )
+                        if not is_seen(CLI_CONFIG, TOOL_PROGRESS_FLAG):
+                            self._long_tool_hint_fired = True
+                            _cprint(f"  {_DIM}{tool_progress_hint_cli()}{_RST}")
+                            mark_seen(_hermes_home / "config.yaml", TOOL_PROGRESS_FLAG)
+                            CLI_CONFIG.setdefault("onboarding", {}).setdefault("seen", {})[TOOL_PROGRESS_FLAG] = True
+                except Exception:
+                    pass
             self._invalidate()
             return
         if event_type != "tool.started":
@@ -9294,6 +9230,24 @@ class HermesCLI:
                                          f"agent_running={self._agent_running}\n")
                         except Exception:
                             pass
+                    # First-touch onboarding: on the very first busy-while-running
+                    # event for this install, print a one-line tip explaining the
+                    # /busy knob.  Flag persists to config.yaml and never fires
+                    # again.  Guarded for exceptions so onboarding can't break
+                    # the input loop.
+                    try:
+                        from agent.onboarding import (
+                            BUSY_INPUT_FLAG,
+                            busy_input_hint_cli,
+                            is_seen,
+                            mark_seen,
+                        )
+                        if not is_seen(CLI_CONFIG, BUSY_INPUT_FLAG):
+                            _cprint(f"  {_DIM}{busy_input_hint_cli(self.busy_input_mode)}{_RST}")
+                            mark_seen(_hermes_home / "config.yaml", BUSY_INPUT_FLAG)
+                            CLI_CONFIG.setdefault("onboarding", {}).setdefault("seen", {})[BUSY_INPUT_FLAG] = True
+                    except Exception:
+                        pass
                 else:
                     self._pending_input.put(payload)
                 event.app.current_buffer.reset(append_to_history=True)
@@ -9308,14 +9262,18 @@ class HermesCLI:
             """Ctrl+Enter (c-j) inserts a newline. Most terminals send c-j for Ctrl+Enter."""
             event.current_buffer.insert_text('\n')
 
-        @kb.add(
-            'c-g',
-            filter=Condition(
-                lambda: not self._clarify_state and not self._approval_state and not self._sudo_state and not self._secret_state
-            ),
+        # VSCode/Cursor bind Ctrl+G to "Find Next" at the editor level, so
+        # the keystroke never reaches the embedded terminal. Alt+G is unbound
+        # in those IDEs and arrives here as ('escape', 'g') — register it as
+        # a fallback so the editor handoff works inside Cursor/VSCode too.
+        _editor_filter = Condition(
+            lambda: not self._clarify_state and not self._approval_state and not self._sudo_state and not self._secret_state
         )
+
+        @kb.add('c-g', filter=_editor_filter)
+        @kb.add('escape', 'g', filter=_editor_filter)
         def handle_open_in_editor(event):
-            """Ctrl+G opens the current draft in an external editor."""
+            """Ctrl+G (or Alt+G in VSCode/Cursor) opens the current draft in an external editor."""
             cli_ref._open_external_editor(event.current_buffer)
 
         @kb.add('tab', eager=True)
@@ -9779,6 +9737,11 @@ class HermesCLI:
                 completer=_completer,
             ),
         )
+        # Keep prompt_toolkit on its simple tempfile path. Setting
+        # buffer.tempfile = "prompt.md" triggers its complex-tempfile branch,
+        # which tries to mkdir() the mkdtemp() directory again and raises
+        # EEXIST. The suffix keeps markdown highlighting without that bug.
+        input_area.buffer.tempfile_suffix = '.md'
 
         # Dynamic height: accounts for both explicit newlines AND visual
         # wrapping of long lines so the input area always fits its content.
@@ -10731,6 +10694,8 @@ class HermesCLI:
                 return  # silently suppress
             if isinstance(exc, KeyError) and "is not registered" in str(exc):
                 return  # suppress selector registration failures (#6393)
+            if isinstance(exc, OSError) and getattr(exc, "errno", None) == errno.EIO:
+                return  # suppress I/O errors from broken stdout on interrupt (#13710)
             # Fall back to default handler for everything else
             loop.default_exception_handler(context)
 
@@ -10763,9 +10728,11 @@ class HermesCLI:
         except (EOFError, KeyboardInterrupt, BrokenPipeError):
             pass
         except (KeyError, OSError) as _stdin_err:
-            # Catch selector registration failures from broken stdin (#6393).
-            # This is the fallback for cases that slip past the fstat() guard.
-            if "is not registered" in str(_stdin_err) or "Bad file descriptor" in str(_stdin_err):
+            # Catch selector registration failures from broken stdin (#6393)
+            # and I/O errors from broken stdout during interrupt (#13710).
+            if isinstance(_stdin_err, OSError) and getattr(_stdin_err, "errno", None) == errno.EIO:
+                pass  # suppress broken-stdout I/O errors on interrupt (#13710)
+            elif "is not registered" in str(_stdin_err) or "Bad file descriptor" in str(_stdin_err):
                 print(
                     f"\nError: stdin is not usable ({_stdin_err}).\n"
                     "This can happen with certain Python installations (e.g. uv-managed cPython on macOS).\n"
@@ -10784,12 +10751,6 @@ class HermesCLI:
                     self.agent.interrupt()
                 except Exception:
                     pass
-            # Flush memories before exit (only for substantial conversations)
-            if self.agent and self.conversation_history:
-                try:
-                    self.agent.flush_memories(self.conversation_history)
-                except (Exception, KeyboardInterrupt):
-                    pass
             # Shut down voice recorder (release persistent audio stream)
             if hasattr(self, '_voice_recorder') and self._voice_recorder:
                 try:

cron/jobs.py

@@ -16,7 +16,7 @@ import uuid
 from datetime import datetime, timedelta
 from pathlib import Path
 from hermes_constants import get_hermes_home
-from typing import Optional, Dict, List, Any
+from typing import Optional, Dict, List, Any, Union
 
 logger = logging.getLogger(__name__)
 
@@ -417,6 +417,7 @@ def create_job(
     provider: Optional[str] = None,
     base_url: Optional[str] = None,
     script: Optional[str] = None,
+    context_from: Optional[Union[str, List[str]]] = None,
     enabled_toolsets: Optional[List[str]] = None,
     workdir: Optional[str] = None,
 ) -> Dict[str, Any]:
@@ -438,6 +439,9 @@ def create_job(
         script: Optional path to a Python script whose stdout is injected into the
                 prompt each run.  The script runs before the agent turn, and its output
                 is prepended as context.  Useful for data collection / change detection.
+        context_from: Optional job ID (or list of job IDs) whose most recent output
+                      is injected into the prompt as context before each run.
+                      Useful for chaining cron jobs: job A finds data, job B processes it.
         enabled_toolsets: Optional list of toolset names to restrict the agent to.
                           When set, only tools from these toolsets are loaded, reducing
                           token overhead. When omitted, all default tools are loaded.
@@ -481,6 +485,14 @@ def create_job(
     normalized_toolsets = normalized_toolsets or None
     normalized_workdir = _normalize_workdir(workdir)
 
+    # Normalize context_from: accept str or list of str, store as list or None
+    if isinstance(context_from, str):
+        context_from = [context_from.strip()] if context_from.strip() else None
+    elif isinstance(context_from, list):
+        context_from = [str(j).strip() for j in context_from if str(j).strip()] or None
+    else:
+        context_from = None
+
     label_source = (prompt or (normalized_skills[0] if normalized_skills else None)) or "cron job"
     job = {
         "id": job_id,
@@ -492,6 +504,7 @@ def create_job(
         "provider": normalized_provider,
         "base_url": normalized_base_url,
         "script": normalized_script,
+        "context_from": context_from,
         "schedule": parsed_schedule,
         "schedule_display": parsed_schedule.get("display", schedule),
         "repeat": {

cron/scheduler.py

@@ -671,6 +671,47 @@ def _build_job_prompt(job: dict, prerun_script: Optional[tuple] = None) -> str:
                 f"{prompt}"
             )
 
+    # Inject output from referenced cron jobs as context.
+    context_from = job.get("context_from")
+    if context_from:
+        from cron.jobs import OUTPUT_DIR
+        if isinstance(context_from, str):
+            context_from = [context_from]
+        for source_job_id in context_from:
+            # Guard against path traversal — valid job IDs are 12-char hex strings
+            if not source_job_id or not all(c in "0123456789abcdef" for c in source_job_id):
+                logger.warning("context_from: skipping invalid job_id %r", source_job_id)
+                continue
+            try:
+                job_output_dir = OUTPUT_DIR / source_job_id
+                if not job_output_dir.exists():
+                    continue  # silent skip — no output yet
+                output_files = sorted(
+                    job_output_dir.glob("*.md"),
+                    key=lambda f: f.stat().st_mtime,
+                    reverse=True,
+                )
+                if not output_files:
+                    continue  # silent skip — no output yet
+                latest_output = output_files[0].read_text(encoding="utf-8").strip()
+                # Truncate to 8K characters to avoid prompt bloat
+                _MAX_CONTEXT_CHARS = 8000
+                if len(latest_output) > _MAX_CONTEXT_CHARS:
+                    latest_output = latest_output[:_MAX_CONTEXT_CHARS] + "\n\n[... output truncated ...]"
+                if latest_output:
+                    prompt = (
+                        f"## Output from job '{source_job_id}'\n"
+                        "The following is the most recent output from a preceding "
+                        "cron job. Use it as context for your analysis.\n\n"
+                        f"```\n{latest_output}\n```\n\n"
+                        f"{prompt}"
+                    )
+                else:
+                    continue  # silent skip — empty output
+            except (OSError, PermissionError) as e:
+                logger.warning("context_from: failed to read output for job %r: %s", source_job_id, e)
+                # silent skip — do not pollute the prompt with error messages
+
     # Always prepend cron execution guidance so the agent knows how
     # delivery works and can suppress delivery when appropriate.
     cron_hint = (

docker/entrypoint.sh

@@ -41,6 +41,15 @@ if [ "$(id -u)" = "0" ]; then
             echo "Warning: chown failed (rootless container?) — continuing anyway"
     fi
 
+    # Ensure config.yaml is readable by the hermes runtime user even if it was
+    # edited on the host after initial ownership setup. Must run here (as root)
+    # rather than after the gosu drop, otherwise a non-root caller like
+    # `docker run -u $(id -u):$(id -g)` hits "Operation not permitted" (#15865).
+    if [ -f "$HERMES_HOME/config.yaml" ]; then
+        chown hermes:hermes "$HERMES_HOME/config.yaml" 2>/dev/null || true
+        chmod 640 "$HERMES_HOME/config.yaml" 2>/dev/null || true
+    fi
+
     echo "Dropping root privileges"
     exec gosu hermes "$0" "$@"
 fi
@@ -67,13 +76,6 @@ if [ ! -f "$HERMES_HOME/config.yaml" ]; then
     cp "$INSTALL_DIR/cli-config.yaml.example" "$HERMES_HOME/config.yaml"
 fi
 
-# Ensure the main config file remains accessible to the hermes runtime user
-# even if it was edited on the host after initial ownership setup.
-if [ -f "$HERMES_HOME/config.yaml" ]; then
-    chown hermes:hermes "$HERMES_HOME/config.yaml"
-    chmod 640 "$HERMES_HOME/config.yaml"
-fi
-
 # SOUL.md
 if [ ! -f "$HERMES_HOME/SOUL.md" ]; then
     cp "$INSTALL_DIR/docker/SOUL.md" "$HERMES_HOME/SOUL.md"

gateway/hooks.py

@@ -21,6 +21,7 @@ Errors in hooks are caught and logged but never block the main pipeline.
 
 import asyncio
 import importlib.util
+import sys
 from typing import Any, Callable, Dict, List, Optional
 
 import yaml
@@ -103,16 +104,28 @@ class HookRegistry:
                     print(f"[hooks] Skipping {hook_name}: no events declared", flush=True)
                     continue
 
-                # Dynamically load the handler module
+                # Dynamically load the handler module.
+                # Register in sys.modules BEFORE exec_module so Pydantic /
+                # dataclasses / typing introspection can resolve forward
+                # references (triggered by `from __future__ import annotations`
+                # in the handler). Without this, a handler that declares a
+                # Pydantic BaseModel for webhook/event payloads fails at first
+                # dispatch with "TypeAdapter ... is not fully defined".
+                module_name = f"hermes_hook_{hook_name}"
                 spec = importlib.util.spec_from_file_location(
-                    f"hermes_hook_{hook_name}", handler_path
+                    module_name, handler_path
                 )
                 if spec is None or spec.loader is None:
                     print(f"[hooks] Skipping {hook_name}: could not load handler.py", flush=True)
                     continue
 
                 module = importlib.util.module_from_spec(spec)
-                spec.loader.exec_module(module)
+                sys.modules[module_name] = module
+                try:
+                    spec.loader.exec_module(module)
+                except Exception:
+                    sys.modules.pop(module_name, None)
+                    raise
 
                 handle_fn = getattr(module, "handle", None)
                 if handle_fn is None:

gateway/platforms/api_server.py

@@ -9,6 +9,7 @@ Exposes an HTTP server with endpoints:
 - GET  /v1/models                  — lists hermes-agent as an available model
 - POST /v1/runs                    — start a run, returns run_id immediately (202)
 - GET  /v1/runs/{run_id}/events    — SSE stream of structured lifecycle events
+- POST /v1/runs/{run_id}/stop    — interrupt a running agent
 - GET  /health                     — health check
 - GET  /health/detailed            — rich status for cross-container dashboard probing
 
@@ -586,6 +587,9 @@ class APIServerAdapter(BasePlatformAdapter):
         self._run_streams: Dict[str, "asyncio.Queue[Optional[Dict]]"] = {}
         # Creation timestamps for orphaned-run TTL sweep
         self._run_streams_created: Dict[str, float] = {}
+        # Active run agent/task references for stop support
+        self._active_run_agents: Dict[str, Any] = {}
+        self._active_run_tasks: Dict[str, "asyncio.Task"] = {}
         self._session_db: Optional[Any] = None  # Lazy-init SessionDB for session continuity
 
     @staticmethod
@@ -2441,6 +2445,7 @@ class APIServerAdapter(BasePlatformAdapter):
                     stream_delta_callback=_text_cb,
                     tool_progress_callback=event_cb,
                 )
+                self._active_run_agents[run_id] = agent
                 def _run_sync():
                     r = agent.run_conversation(
                         user_message=user_message,
@@ -2480,8 +2485,11 @@ class APIServerAdapter(BasePlatformAdapter):
                     q.put_nowait(None)
                 except Exception:
                     pass
+                self._active_run_agents.pop(run_id, None)
+                self._active_run_tasks.pop(run_id, None)
 
         task = asyncio.create_task(_run_and_close())
+        self._active_run_tasks[run_id] = task
         try:
             self._background_tasks.add(task)
         except TypeError:
@@ -2540,6 +2548,44 @@ class APIServerAdapter(BasePlatformAdapter):
 
         return response
 
+    async def _handle_stop_run(self, request: "web.Request") -> "web.Response":
+        """POST /v1/runs/{run_id}/stop — interrupt a running agent."""
+        auth_err = self._check_auth(request)
+        if auth_err:
+            return auth_err
+
+        run_id = request.match_info["run_id"]
+        agent = self._active_run_agents.get(run_id)
+        task = self._active_run_tasks.get(run_id)
+
+        if agent is None and task is None:
+            return web.json_response(_openai_error(f"Run not found: {run_id}", code="run_not_found"), status=404)
+
+        if agent is not None:
+            try:
+                agent.interrupt("Stop requested via API")
+            except Exception:
+                pass
+
+        if task is not None and not task.done():
+            task.cancel()
+            # Bounded wait: run_conversation() executes in the default
+            # executor thread which task.cancel() cannot preempt — we rely on
+            # agent.interrupt() above to break the loop. Cap the wait so a
+            # slow/unresponsive interrupt can't hang this handler.
+            try:
+                await asyncio.wait_for(asyncio.shield(task), timeout=5.0)
+            except asyncio.TimeoutError:
+                logger.warning(
+                    "[api_server] stop for run %s timed out after 5s; "
+                    "agent may still be finishing the current step",
+                    run_id,
+                )
+            except (asyncio.CancelledError, Exception):
+                pass
+
+        return web.json_response({"run_id": run_id, "status": "stopping"})
+
     async def _sweep_orphaned_runs(self) -> None:
         """Periodically clean up run streams that were never consumed."""
         while True:
@@ -2554,6 +2600,8 @@ class APIServerAdapter(BasePlatformAdapter):
                 logger.debug("[api_server] sweeping orphaned run %s", run_id)
                 self._run_streams.pop(run_id, None)
                 self._run_streams_created.pop(run_id, None)
+                self._active_run_agents.pop(run_id, None)
+                self._active_run_tasks.pop(run_id, None)
 
     # ------------------------------------------------------------------
     # BasePlatformAdapter interface
@@ -2589,6 +2637,7 @@ class APIServerAdapter(BasePlatformAdapter):
             # Structured event streaming
             self._app.router.add_post("/v1/runs", self._handle_runs)
             self._app.router.add_get("/v1/runs/{run_id}/events", self._handle_run_events)
+            self._app.router.add_post("/v1/runs/{run_id}/stop", self._handle_stop_run)
             # Start background sweep to clean up orphaned (unconsumed) run streams
             sweep_task = asyncio.create_task(self._sweep_orphaned_runs())
             try:

gateway/platforms/base.py

@@ -1025,7 +1025,20 @@ class BasePlatformAdapter(ABC):
         self._post_delivery_callbacks: Dict[str, Any] = {}
         self._expected_cancelled_tasks: set[asyncio.Task] = set()
         self._busy_session_handler: Optional[Callable[[MessageEvent, str], Awaitable[bool]]] = None
-        # Chats where auto-TTS on voice input is disabled (set by /voice off)
+        # Auto-TTS on voice input: ``_auto_tts_default`` is the global default
+        # (``voice.auto_tts`` in config.yaml, pushed by GatewayRunner on connect).
+        # Per-chat overrides live in two sets populated from ``_voice_mode``:
+        #   - ``_auto_tts_enabled_chats``: chat explicitly opted in via ``/voice on``
+        #     or ``/voice tts`` (mode is ``voice_only`` or ``all``). Fires even when
+        #     the global default is False.
+        #   - ``_auto_tts_disabled_chats``: chat explicitly opted out via
+        #     ``/voice off`` (mode is ``off``). Suppresses auto-TTS even when the
+        #     global default is True.
+        # The gate in _process_message() is:
+        #   fire if chat in _auto_tts_enabled_chats
+        #     OR (_auto_tts_default and chat not in _auto_tts_disabled_chats)
+        self._auto_tts_default: bool = False
+        self._auto_tts_enabled_chats: set = set()
         self._auto_tts_disabled_chats: set = set()
         # Chats where typing indicator is paused (e.g. during approval waits).
         # _keep_typing skips send_typing when the chat_id is in this set.
@@ -1047,6 +1060,21 @@ class BasePlatformAdapter(ABC):
     def fatal_error_retryable(self) -> bool:
         return self._fatal_error_retryable
 
+    def _should_auto_tts_for_chat(self, chat_id: str) -> bool:
+        """Whether auto-TTS on voice input should fire for ``chat_id``.
+
+        Decision layers (Issue #16007):
+          1. Explicit ``/voice on`` or ``/voice tts`` → always fire (even if
+             ``voice.auto_tts`` is False).
+          2. Explicit ``/voice off`` → never fire.
+          3. Fall back to the global ``voice.auto_tts`` config default.
+        """
+        if chat_id in self._auto_tts_enabled_chats:
+            return True
+        if chat_id in self._auto_tts_disabled_chats:
+            return False
+        return bool(self._auto_tts_default)
+
     def set_fatal_error_handler(self, handler: Callable[["BasePlatformAdapter"], Awaitable[None] | None]) -> None:
         self._fatal_error_handler = handler
 
@@ -2214,12 +2242,14 @@ class BasePlatformAdapter(ABC):
                     logger.info("[%s] extract_local_files found %d file(s) in response", self.name, len(local_files))
                 
                 # Auto-TTS: if voice message, generate audio FIRST (before sending text)
-                # Skipped when the chat has voice mode disabled (/voice off)
+                # Gated via ``_should_auto_tts_for_chat``: fires when the chat has
+                # an explicit ``/voice on|tts`` opt-in OR when ``voice.auto_tts`` is
+                # True globally and no ``/voice off`` has been issued.
                 _tts_path = None
-                if (event.message_type == MessageType.VOICE
+                if (self._should_auto_tts_for_chat(event.source.chat_id)
+                        and event.message_type == MessageType.VOICE
                         and text_content
-                        and not media_files
-                        and event.source.chat_id not in self._auto_tts_disabled_chats):
+                        and not media_files):
                     try:
                         from tools.tts_tool import text_to_speech_tool, check_tts_requirements
                         if check_tts_requirements():
@@ -2543,6 +2573,9 @@ class BasePlatformAdapter(ABC):
         user_id_alt: Optional[str] = None,
         chat_id_alt: Optional[str] = None,
         is_bot: bool = False,
+        guild_id: Optional[str] = None,
+        parent_chat_id: Optional[str] = None,
+        message_id: Optional[str] = None,
     ) -> SessionSource:
         """Helper to build a SessionSource for this platform."""
         # Normalize empty topic to None
@@ -2560,6 +2593,9 @@ class BasePlatformAdapter(ABC):
             user_id_alt=user_id_alt,
             chat_id_alt=chat_id_alt,
             is_bot=is_bot,
+            guild_id=str(guild_id) if guild_id else None,
+            parent_chat_id=str(parent_chat_id) if parent_chat_id else None,
+            message_id=str(message_id) if message_id else None,
         )
     
     @abstractmethod

gateway/platforms/discord.py

@@ -2315,11 +2315,6 @@ class DiscordAdapter(BasePlatformAdapter):
         async def slash_background(interaction: discord.Interaction, prompt: str):
             await self._run_simple_slash(interaction, f"/background {prompt}", "Background task started~")
 
-        @tree.command(name="btw", description="Ephemeral side question using session context")
-        @discord.app_commands.describe(question="Your side question (no tools, not persisted)")
-        async def slash_btw(interaction: discord.Interaction, question: str):
-            await self._run_simple_slash(interaction, f"/btw {question}")
-
         # ── Auto-register any gateway-available commands not yet on the tree ──
         # This ensures new commands added to COMMAND_REGISTRY in
         # hermes_cli/commands.py automatically appear as Discord slash
@@ -3261,6 +3256,7 @@ class DiscordAdapter(BasePlatformAdapter):
             if auto_thread and not skip_thread and not is_voice_linked_channel and not is_reply_message:
                 thread = await self._auto_create_thread(message)
                 if thread:
+                    parent_channel_id = str(message.channel.id)
                     is_thread = True
                     thread_id = str(thread.id)
                     auto_threaded_channel = thread
@@ -3320,6 +3316,9 @@ class DiscordAdapter(BasePlatformAdapter):
             thread_id=thread_id,
             chat_topic=chat_topic,
             is_bot=getattr(message.author, "bot", False),
+            guild_id=str(message.guild.id) if message.guild else None,
+            parent_chat_id=parent_channel_id,
+            message_id=str(message.id),
         )
 
         # Build media URLs -- download image attachments to local cache so the

gateway/platforms/matrix.py

@@ -532,6 +532,20 @@ class MatrixAdapter(BasePlatformAdapter):
                 )
                 await crypto_store.open()
 
+                # Bind the store to the runtime device_id before any
+                # put_account() runs. PgCryptoStore defaults _device_id
+                # to "" and its crypto_account UPSERT never updates the
+                # device_id column on conflict — so once put_account
+                # writes blank, it stays blank forever. That breaks
+                # every downstream device-scoped olm operation: peer
+                # to-device ciphertext can't find our identity key and
+                # no megolm sessions ever land. Setting _device_id here
+                # (in-memory; the on-disk row may not exist yet) makes
+                # the first put_account write the correct value.
+                # DeviceID is a NewType(str) so plain str works at runtime.
+                if client.device_id:
+                    await crypto_store.put_device_id(client.device_id)
+
                 crypto_state = _CryptoStateStore(state_store, self._joined_rooms)
                 olm = OlmMachine(client, crypto_store, crypto_state)
 

gateway/session.py

@@ -87,6 +87,9 @@ class SessionSource:
     user_id_alt: Optional[str] = None  # Platform-specific stable alt ID (Signal UUID, Feishu union_id)
     chat_id_alt: Optional[str] = None  # Signal group internal ID
     is_bot: bool = False  # True when the message author is a bot/webhook (Discord)
+    guild_id: Optional[str] = None  # Discord guild / Slack workspace / Matrix server scope
+    parent_chat_id: Optional[str] = None  # Parent channel when chat_id refers to a thread
+    message_id: Optional[str] = None  # ID of the triggering message (for pin/reply/react)
     
     @property
     def description(self) -> str:
@@ -124,8 +127,14 @@ class SessionSource:
             d["user_id_alt"] = self.user_id_alt
         if self.chat_id_alt:
             d["chat_id_alt"] = self.chat_id_alt
+        if self.guild_id:
+            d["guild_id"] = self.guild_id
+        if self.parent_chat_id:
+            d["parent_chat_id"] = self.parent_chat_id
+        if self.message_id:
+            d["message_id"] = self.message_id
         return d
-    
+
     @classmethod
     def from_dict(cls, data: Dict[str, Any]) -> "SessionSource":
         return cls(
@@ -139,6 +148,9 @@ class SessionSource:
             chat_topic=data.get("chat_topic"),
             user_id_alt=data.get("user_id_alt"),
             chat_id_alt=data.get("chat_id_alt"),
+            guild_id=data.get("guild_id"),
+            parent_chat_id=data.get("parent_chat_id"),
+            message_id=data.get("message_id"),
         )
     
 
@@ -190,6 +202,31 @@ that requires raw IDs).  Discord is excluded because mentions use ``<@user_id>``
 and the LLM needs the real ID to tag users."""
 
 
+def _discord_tools_loaded() -> bool:
+    """True iff the agent will actually have Discord tools this session.
+
+    Two conditions must hold:
+      1. The `discord` or `discord_admin` toolset is enabled for the
+         Discord platform via `hermes tools` (opt-in, default OFF).
+      2. `DISCORD_BOT_TOKEN` is set — the tool's `check_fn` gates on it
+         at registry time, so the toolset being enabled in config is not
+         enough if the token isn't configured.
+
+    Returns False (safe default — keeps the stale-API disclaimer) on any
+    error so a bad config can't silently promise tools the agent lacks.
+    """
+    if not (os.environ.get("DISCORD_BOT_TOKEN") or "").strip():
+        return False
+    try:
+        from hermes_cli.config import load_config
+        from hermes_cli.tools_config import _get_platform_tools
+        cfg = load_config()
+        enabled = _get_platform_tools(cfg, "discord", include_default_mcp_servers=False)
+        return "discord" in enabled or "discord_admin" in enabled
+    except Exception:
+        return False
+
+
 def build_session_context_prompt(
     context: SessionContext,
     *,
@@ -277,14 +314,33 @@ def build_session_context_prompt(
             "that you can only read messages sent directly to you and respond."
         )
     elif context.source.platform == Platform.DISCORD:
-        lines.append("")
-        lines.append(
-            "**Platform notes:** You are running inside Discord. "
-            "You do NOT have access to Discord-specific APIs — you cannot search "
-            "channel history, pin messages, manage roles, or list server members. "
-            "Do not promise to perform these actions. If the user asks, explain "
-            "that you can only read messages sent directly to you and respond."
-        )
+        # Inject the Discord IDs block only when the agent actually has
+        # Discord tools loaded this session — i.e. the user opted into
+        # `discord` / `discord_admin` via `hermes tools` AND the bot
+        # token is configured.  Otherwise keep the stale-API disclaimer
+        # honest so we never promise tools the agent lacks.
+        if _discord_tools_loaded():
+            src = context.source
+            id_lines = ["", "**Discord IDs (for the `discord` / `discord_admin` tools):**"]
+            if src.guild_id:
+                id_lines.append(f"  - Guild: `{src.guild_id}`")
+            if src.thread_id and src.parent_chat_id:
+                id_lines.append(f"  - Parent channel: `{src.parent_chat_id}`")
+                id_lines.append(f"  - Thread: `{src.thread_id}` (use as `channel_id` for fetch_messages etc.)")
+            else:
+                id_lines.append(f"  - Channel: `{src.chat_id}`")
+            if src.message_id:
+                id_lines.append(f"  - Triggering message: `{src.message_id}`")
+            lines.extend(id_lines)
+        else:
+            lines.append("")
+            lines.append(
+                "**Platform notes:** You are running inside Discord. "
+                "You do NOT have access to Discord-specific APIs — you cannot search "
+                "channel history, pin messages, manage roles, or list server members. "
+                "Do not promise to perform these actions. If the user asks, explain "
+                "that you can only read messages sent directly to you and respond."
+            )
     elif context.source.platform == Platform.BLUEBUBBLES:
         lines.append("")
         lines.append(
@@ -383,11 +439,11 @@ class SessionEntry:
     auto_reset_reason: Optional[str] = None  # "idle" or "daily"
     reset_had_activity: bool = False  # whether the expired session had any messages
     
-    # Set by the background expiry watcher after it successfully flushes
-    # memories for this session.  Persisted to sessions.json so the flag
-    # survives gateway restarts (the old in-memory _pre_flushed_sessions
-    # set was lost on restart, causing redundant re-flushes).
-    memory_flushed: bool = False
+    # Set by the background expiry watcher after it finalizes an expired
+    # session (invoking on_session_finalize hooks and evicting the cached
+    # agent).  Persisted to sessions.json so the flag survives gateway
+    # restarts — prevents redundant finalization runs.
+    expiry_finalized: bool = False
 
     # When True the next call to get_or_create_session() will auto-reset
     # this session (create a new session_id) so the user starts fresh.
@@ -423,7 +479,7 @@ class SessionEntry:
             "last_prompt_tokens": self.last_prompt_tokens,
             "estimated_cost_usd": self.estimated_cost_usd,
             "cost_status": self.cost_status,
-            "memory_flushed": self.memory_flushed,
+            "expiry_finalized": self.expiry_finalized,
             "suspended": self.suspended,
             "resume_pending": self.resume_pending,
             "resume_reason": self.resume_reason,
@@ -475,7 +531,7 @@ class SessionEntry:
             last_prompt_tokens=data.get("last_prompt_tokens", 0),
             estimated_cost_usd=data.get("estimated_cost_usd", 0.0),
             cost_status=data.get("cost_status", "unknown"),
-            memory_flushed=data.get("memory_flushed", False),
+            expiry_finalized=data.get("expiry_finalized", data.get("memory_flushed", False)),
             suspended=data.get("suspended", False),
             resume_pending=data.get("resume_pending", False),
             resume_reason=data.get("resume_reason"),
@@ -1176,6 +1232,7 @@ class SessionStore:
                     reasoning_content=message.get("reasoning_content") if message.get("role") == "assistant" else None,
                     reasoning_details=message.get("reasoning_details") if message.get("role") == "assistant" else None,
                     codex_reasoning_items=message.get("codex_reasoning_items") if message.get("role") == "assistant" else None,
+                    codex_message_items=message.get("codex_message_items") if message.get("role") == "assistant" else None,
                 )
             except Exception as e:
                 logger.debug("Session DB operation failed: %s", e)
@@ -1208,6 +1265,7 @@ class SessionStore:
                         reasoning_content=msg.get("reasoning_content") if role == "assistant" else None,
                         reasoning_details=msg.get("reasoning_details") if role == "assistant" else None,
                         codex_reasoning_items=msg.get("codex_reasoning_items") if role == "assistant" else None,
+                        codex_message_items=msg.get("codex_message_items") if role == "assistant" else None,
                     )
             except Exception as e:
                 logger.debug("Failed to rewrite transcript in DB: %s", e)

hermes_cli/auth.py

@@ -356,6 +356,14 @@ PROVIDER_REGISTRY: Dict[str, ProviderConfig] = {
         api_key_env_vars=(),
         base_url_env_var="BEDROCK_BASE_URL",
     ),
+    "azure-foundry": ProviderConfig(
+        id="azure-foundry",
+        name="Azure Foundry",
+        auth_type="api_key",
+        inference_base_url="",  # User-provided endpoint
+        api_key_env_vars=("AZURE_FOUNDRY_API_KEY",),
+        base_url_env_var="AZURE_FOUNDRY_BASE_URL",
+    ),
 }
 
 
@@ -4236,10 +4244,10 @@ def _login_nous(args, pconfig: ProviderConfig) -> None:
                 )
 
             from hermes_cli.models import (
-                _PROVIDER_MODELS, get_pricing_for_provider,
+                get_curated_nous_model_ids, get_pricing_for_provider,
                 check_nous_free_tier, partition_nous_models_by_tier,
             )
-            model_ids = _PROVIDER_MODELS.get("nous", [])
+            model_ids = get_curated_nous_model_ids()
 
             print()
             unavailable_models: list = []

hermes_cli/azure_detect.py

@@ -0,0 +1,300 @@
+"""Azure Foundry endpoint auto-detection.
+
+Inspect an Azure AI Foundry / Azure OpenAI endpoint to determine:
+  - API transport (OpenAI-style ``chat_completions`` vs
+    Anthropic-style ``anthropic_messages``)
+  - Available models (best effort — Azure does not expose a deployment
+    listing via the inference API key, but Azure OpenAI v1 endpoints
+    return the resource's model catalog via ``GET /models``)
+  - Context length for each discovered/entered model, via the existing
+    :func:`agent.model_metadata.get_model_context_length` resolver.
+
+Rationale:
+
+Azure has no pure-API-key deployment-listing endpoint — per Microsoft,
+deployment enumeration requires ARM management-plane auth.  Azure
+OpenAI v1 endpoints ``{resource}.openai.azure.com/openai/v1`` do return
+a ``/models`` list, but it reflects the resource's *available* models
+rather than the user's *deployed* deployment names.  In practice it is
+still a useful hint — the user picks a familiar model name and we look
+up its context length from the catalog.
+
+The detector never crashes on errors (every HTTP call is wrapped in a
+broad try/except).  Callers get a :class:`DetectionResult` with whatever
+information could be gathered, and fall back to manual entry for the
+rest.
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+import re
+from dataclasses import dataclass, field
+from typing import Optional
+from urllib import request as urllib_request
+from urllib.error import HTTPError, URLError
+from urllib.parse import urlparse, urlunparse
+
+logger = logging.getLogger(__name__)
+
+
+# Default Azure OpenAI ``api-version`` to probe with.  The v1 GA endpoint
+# accepts requests without ``api-version`` entirely, so this is only used
+# as a fallback for pre-v1 resources that still require it.
+_AZURE_OPENAI_PROBE_API_VERSIONS = (
+    "2025-04-01-preview",
+    "2024-10-21",  # oldest GA that supports /models
+)
+
+# Default Azure Anthropic ``api-version``.  Matches the value used by
+# ``agent/anthropic_adapter.py`` when building the Anthropic client.
+_AZURE_ANTHROPIC_API_VERSION = "2025-04-15"
+
+
+@dataclass
+class DetectionResult:
+    """Everything auto-detection could gather from a base URL + API key."""
+
+    #: Detected API transport: ``"chat_completions"``,
+    #: ``"anthropic_messages"``, or ``None`` when detection failed.
+    api_mode: Optional[str] = None
+
+    #: Deployment / model IDs returned by ``/models`` (best effort).
+    #: Empty when the endpoint doesn't expose the list with an API key.
+    models: list[str] = field(default_factory=list)
+
+    #: Lowercased host from the base URL (used for display messages).
+    hostname: str = ""
+
+    #: Human-readable reason the detector chose ``api_mode``.  Useful
+    #: for explaining auto-detection to the user in the wizard.
+    reason: str = ""
+
+    #: ``True`` when ``/models`` returned a valid OpenAI-shaped payload.
+    models_probe_ok: bool = False
+
+    #: ``True`` when the URL was determined to be an Anthropic-style
+    #: endpoint (from path suffix or live probe).
+    is_anthropic: bool = False
+
+
+def _http_get_json(url: str, api_key: str, timeout: float = 6.0) -> tuple[int, Optional[dict]]:
+    """GET a URL with ``api-key`` + ``Authorization`` headers.  Return
+    ``(status_code, parsed_json_or_None)``.  Never raises."""
+    req = urllib_request.Request(url, method="GET")
+    # Azure OpenAI uses ``api-key``.  Some Azure deployments (and
+    # Anthropic-style routes) use ``Authorization: Bearer``.  Send both
+    # so we probe once per URL rather than twice.
+    req.add_header("api-key", api_key)
+    req.add_header("Authorization", f"Bearer {api_key}")
+    req.add_header("User-Agent", "hermes-agent/azure-detect")
+    try:
+        with urllib_request.urlopen(req, timeout=timeout) as resp:
+            body = resp.read()
+            try:
+                return resp.status, json.loads(body.decode("utf-8", errors="replace"))
+            except Exception:
+                return resp.status, None
+    except HTTPError as exc:
+        return exc.code, None
+    except (URLError, TimeoutError, OSError) as exc:
+        logger.debug("azure_detect: GET %s failed: %s", url, exc)
+        return 0, None
+    except Exception as exc:  # pragma: no cover — defensive
+        logger.debug("azure_detect: GET %s unexpected error: %s", url, exc)
+        return 0, None
+
+
+def _strip_trailing_v1(url: str) -> str:
+    """Strip trailing ``/v1`` or ``/v1/`` so we can construct sub-paths."""
+    return re.sub(r"/v1/?$", "", url.rstrip("/"))
+
+
+def _looks_like_anthropic_path(url: str) -> bool:
+    """Return True when the URL's path ends in ``/anthropic`` or
+    contains a ``/anthropic/`` segment.  Used by Azure Foundry
+    resources that route Claude traffic through a dedicated path."""
+    try:
+        parsed = urlparse(url)
+        path = (parsed.path or "").lower().rstrip("/")
+        return path.endswith("/anthropic") or "/anthropic/" in path + "/"
+    except Exception:
+        return False
+
+
+def _extract_model_ids(payload: dict) -> list[str]:
+    """Extract a list of model IDs from an OpenAI-shaped ``/models``
+    response.  Returns ``[]`` on any shape mismatch."""
+    data = payload.get("data") if isinstance(payload, dict) else None
+    if not isinstance(data, list):
+        return []
+    ids: list[str] = []
+    for item in data:
+        if not isinstance(item, dict):
+            continue
+        # OpenAI shape: {"id": "gpt-5.4", "object": "model", ...}
+        mid = item.get("id") or item.get("model") or item.get("name")
+        if isinstance(mid, str) and mid:
+            ids.append(mid)
+    return ids
+
+
+def _probe_openai_models(base_url: str, api_key: str) -> tuple[bool, list[str]]:
+    """Probe ``<base>/models`` for an OpenAI-shaped response.
+
+    Returns ``(ok, models)``.  ``ok`` is True iff the endpoint accepted
+    us as an OpenAI-style caller (200 OK + OpenAI-shaped JSON body).
+    """
+    base_url = base_url.rstrip("/")
+
+    # Azure OpenAI v1: {resource}.openai.azure.com/openai/v1 — no
+    # api-version required for GA paths, so probe without first.
+    candidates = [f"{base_url}/models"]
+    # Fallback: explicit api-version for pre-v1 resources
+    for v in _AZURE_OPENAI_PROBE_API_VERSIONS:
+        candidates.append(f"{base_url}/models?api-version={v}")
+
+    for url in candidates:
+        status, body = _http_get_json(url, api_key)
+        if status == 200 and body is not None:
+            ids = _extract_model_ids(body)
+            if ids:
+                logger.info(
+                    "azure_detect: /models probe OK at %s (%d models)",
+                    url, len(ids),
+                )
+                return True, ids
+            # 200 + empty list still counts as "OpenAI shape, no models
+            # listed" — let the user proceed with manual entry.
+            if isinstance(body, dict) and "data" in body:
+                return True, []
+    return False, []
+
+
+def _probe_anthropic_messages(base_url: str, api_key: str) -> bool:
+    """Send a zero-token request to ``<base>/v1/messages`` and check
+    whether the endpoint at least *recognises* the Anthropic Messages
+    shape (any 4xx that mentions ``messages`` or ``model``, or a 400
+    ``invalid_request`` with an Anthropic error shape).  Never completes
+    a real chat.
+    """
+    base = _strip_trailing_v1(base_url)
+    url = f"{base}/v1/messages?api-version={_AZURE_ANTHROPIC_API_VERSION}"
+    payload = json.dumps({
+        "model": "probe",
+        "max_tokens": 1,
+        "messages": [{"role": "user", "content": "ping"}],
+    }).encode("utf-8")
+    req = urllib_request.Request(url, method="POST", data=payload)
+    req.add_header("api-key", api_key)
+    req.add_header("Authorization", f"Bearer {api_key}")
+    req.add_header("anthropic-version", "2023-06-01")
+    req.add_header("content-type", "application/json")
+    req.add_header("User-Agent", "hermes-agent/azure-detect")
+    try:
+        with urllib_request.urlopen(req, timeout=6.0) as resp:
+            # Should never 200 — "probe" isn't a real deployment.  But
+            # if it does, the endpoint definitely speaks Anthropic.
+            return resp.status < 500
+    except HTTPError as exc:
+        # 4xx with an Anthropic-shaped error body = Anthropic endpoint.
+        try:
+            body = exc.read().decode("utf-8", errors="replace")
+            lowered = body.lower()
+            if "anthropic" in lowered or '"type"' in lowered and '"error"' in lowered:
+                return True
+            # Pre-Azure-v1 Azure Foundry returns a plain 404 for
+            # Anthropic-style calls on non-Anthropic deployments.  A
+            # 400 "model not found" IS Anthropic though.
+            if exc.code == 400 and ("messages" in lowered or "model" in lowered):
+                return True
+            return False
+        except Exception:
+            return False
+    except (URLError, TimeoutError, OSError):
+        return False
+    except Exception:  # pragma: no cover
+        return False
+
+
+def detect(base_url: str, api_key: str) -> DetectionResult:
+    """Inspect an Azure endpoint and describe its transport + models.
+
+    Call this from the wizard before asking the user to pick an API
+    mode manually.  The caller should treat the returned
+    :class:`DetectionResult` as *advisory* — if ``api_mode`` is None,
+    fall back to asking the user.
+    """
+    result = DetectionResult()
+
+    try:
+        parsed = urlparse(base_url)
+        result.hostname = (parsed.hostname or "").lower()
+    except Exception:
+        result.hostname = ""
+
+    # 1. Path sniff.  Azure Foundry exposes Anthropic-style deployments
+    #    under a dedicated ``/anthropic`` path.
+    if _looks_like_anthropic_path(base_url):
+        result.is_anthropic = True
+        result.api_mode = "anthropic_messages"
+        result.reason = "URL path ends in /anthropic → Anthropic Messages API"
+        return result
+
+    # 2. Try the OpenAI-style /models probe.  If this works, the
+    #    endpoint definitely speaks OpenAI wire.
+    ok, models = _probe_openai_models(base_url, api_key)
+    if ok:
+        result.models_probe_ok = True
+        result.models = models
+        result.api_mode = "chat_completions"
+        result.reason = (
+            f"GET /models returned {len(models)} model(s) — OpenAI-style endpoint"
+            if models
+            else "GET /models returned an OpenAI-shaped empty list — OpenAI-style endpoint"
+        )
+        return result
+
+    # 3. Fallback: probe the Anthropic Messages shape.  Slower and more
+    #    intrusive than /models, so only run it when the OpenAI probe
+    #    failed.
+    if _probe_anthropic_messages(base_url, api_key):
+        result.is_anthropic = True
+        result.api_mode = "anthropic_messages"
+        result.reason = "Endpoint accepts Anthropic Messages shape"
+        return result
+
+    # Nothing matched.  Caller falls back to manual selection.
+    result.reason = (
+        "Could not probe endpoint (private network, missing model list, or "
+        "non-standard path) — falling back to manual API-mode selection"
+    )
+    return result
+
+
+def lookup_context_length(model: str, base_url: str, api_key: str) -> Optional[int]:
+    """Thin wrapper around :func:`agent.model_metadata.get_model_context_length`
+    that returns ``None`` when only the fallback default (128k) would
+    fire, so the wizard can distinguish "we actually know this" from
+    "we guessed."""
+    try:
+        from agent.model_metadata import (
+            DEFAULT_FALLBACK_CONTEXT,
+            get_model_context_length,
+        )
+    except Exception:
+        return None
+
+    try:
+        n = get_model_context_length(model, base_url=base_url, api_key=api_key)
+    except Exception as exc:
+        logger.debug("azure_detect: context length lookup failed: %s", exc)
+        return None
+
+    if isinstance(n, int) and n > 0 and n != DEFAULT_FALLBACK_CONTEXT:
+        return n
+    return None
+
+
+__all__ = ["DetectionResult", "detect", "lookup_context_length"]

hermes_cli/commands.py

@@ -84,9 +84,7 @@ COMMAND_REGISTRY: list[CommandDef] = [
     CommandDef("deny", "Deny a pending dangerous command", "Session",
                gateway_only=True),
     CommandDef("background", "Run a prompt in the background", "Session",
-               aliases=("bg",), args_hint="<prompt>"),
-    CommandDef("btw", "Ephemeral side question using session context (no tools, not persisted)", "Session",
-               args_hint="<question>"),
+               aliases=("bg", "btw"), args_hint="<prompt>"),
     CommandDef("agents", "Show active agents and running tasks", "Session",
                aliases=("tasks",)),
     CommandDef("queue", "Queue a prompt for the next turn (doesn't interrupt)", "Session",
@@ -103,7 +101,8 @@ COMMAND_REGISTRY: list[CommandDef] = [
     # Configuration
     CommandDef("config", "Show current configuration", "Configuration",
                cli_only=True),
-    CommandDef("model", "Switch model for this session", "Configuration", args_hint="[model] [--provider name] [--global]"),
+    CommandDef("model", "Switch model for this session", "Configuration",
+               aliases=("provider",), args_hint="[model] [--provider name] [--global]"),
     CommandDef("gquota", "Show Google Gemini Code Assist quota usage", "Info",
                cli_only=True),
 
@@ -141,6 +140,11 @@ COMMAND_REGISTRY: list[CommandDef] = [
     CommandDef("cron", "Manage scheduled tasks", "Tools & Skills",
                cli_only=True, args_hint="[subcommand]",
                subcommands=("list", "add", "create", "edit", "pause", "resume", "run", "remove")),
+    CommandDef("kanban", "Multi-profile collaboration board (tasks, links, comments)",
+               "Tools & Skills", args_hint="[subcommand]",
+               subcommands=("list", "ls", "show", "create", "assign", "link", "unlink",
+                            "claim", "comment", "complete", "block", "unblock", "archive",
+                            "tail", "dispatch", "context", "init", "gc")),
     CommandDef("reload", "Reload .env variables into the running session", "Tools & Skills",
                cli_only=True),
     CommandDef("reload-mcp", "Reload MCP servers from config", "Tools & Skills",

hermes_cli/config.py

@@ -612,14 +612,6 @@ DEFAULT_CONFIG = {
             "timeout": 30,
             "extra_body": {},
         },
-        "flush_memories": {
-            "provider": "auto",
-            "model": "",
-            "base_url": "",
-            "api_key": "",
-            "timeout": 30,
-            "extra_body": {},
-        },
         "title_generation": {
             "provider": "auto",
             "model": "",
@@ -783,6 +775,15 @@ DEFAULT_CONFIG = {
         # warning log if out of range.
         "max_spawn_depth": 1,        # depth cap (1 = flat [default], 2 = orchestrator→leaf, 3 = three-level)
         "orchestrator_enabled": True,  # kill switch for role="orchestrator"
+        # When a subagent hits a dangerous-command approval prompt, the parent's
+        # prompt_toolkit TUI owns stdin — a thread-local input() call from the
+        # subagent worker would deadlock the parent UI. To avoid the deadlock,
+        # subagent threads ALWAYS resolve approvals non-interactively:
+        #   false (default) → auto-deny with a logger.warning audit line (safe)
+        #   true             → auto-approve "once" with a logger.warning audit line
+        # Flip to true only if you trust delegated work to run dangerous cmds
+        # without human review (cron pipelines, batch automation, etc.).
+        "subagent_auto_approve": False,
     },
 
     # Ephemeral prefill messages file — JSON list of {role, content} dicts
@@ -839,7 +840,7 @@ DEFAULT_CONFIG = {
         "auto_thread": True,           # Auto-create threads on @mention in channels (like Slack)
         "reactions": True,             # Add 👀/✅/❌ reactions to messages during processing
         "channel_prompts": {},         # Per-channel ephemeral system prompts (forum parents apply to child threads)
-        # discord_server tool: restrict which actions the agent may call.
+        # discord / discord_admin tools: restrict which actions the agent may call.
         # Default (empty) = all actions allowed (subject to bot privileged intents).
         # Accepts comma-separated string ("list_guilds,list_channels,fetch_messages")
         # or YAML list. Unknown names are dropped with a warning at load time.
@@ -958,6 +959,27 @@ DEFAULT_CONFIG = {
         "backup_count": 3,     # Number of rotated backup files to keep
     },
 
+    # Remotely-hosted model catalog manifest.  When enabled, the CLI fetches
+    # curated model lists for OpenRouter and Nous Portal from this URL,
+    # falling back to the in-repo snapshot on network failure.  Lets us
+    # update model picker lists without shipping a hermes-agent release.
+    # The default URL is served by the docs site GitHub Pages deploy.
+    "model_catalog": {
+        "enabled": True,
+        "url": "https://hermes-agent.nousresearch.com/docs/api/model-catalog.json",
+        # Disk cache TTL in hours.  Beyond this, the CLI refetches on the
+        # next /model or `hermes model` invocation; network failures
+        # silently fall back to the stale cache.
+        "ttl_hours": 24,
+        # Optional per-provider override URLs for third parties that want
+        # to self-host their own curation list using the same schema.
+        # Example:
+        #   providers:
+        #     openrouter:
+        #       url: https://example.com/my-curation.json
+        "providers": {},
+    },
+
     # Network settings — workarounds for connectivity issues.
     "network": {
         # Force IPv4 connections.  On servers with broken or unreachable IPv6,
@@ -994,6 +1016,13 @@ DEFAULT_CONFIG = {
         "min_interval_hours": 24,
     },
 
+    # Contextual first-touch onboarding hints (see agent/onboarding.py).
+    # Each hint is shown once per install and then latched here so it
+    # never fires again.  Users can wipe the section to re-see all hints.
+    "onboarding": {
+        "seen": {},
+    },
+
     # Config schema version - bump this when adding new required fields
     "_config_version": 22,
 }
@@ -1370,6 +1399,21 @@ OPTIONAL_ENV_VARS = {
         "category": "provider",
         "advanced": True,
     },
+    "AZURE_FOUNDRY_API_KEY": {
+        "description": "Azure Foundry API key for custom Azure endpoints",
+        "prompt": "Azure Foundry API Key",
+        "url": "https://ai.azure.com/",
+        "password": True,
+        "category": "provider",
+    },
+    "AZURE_FOUNDRY_BASE_URL": {
+        "description": "Azure Foundry base URL (set via 'hermes model' for endpoint-specific config)",
+        "prompt": "Azure Foundry base URL",
+        "url": None,
+        "password": False,
+        "category": "provider",
+        "advanced": True,
+    },
 
     # ── Tool API keys ──
     "EXA_API_KEY": {
@@ -2205,6 +2249,71 @@ def get_compatible_custom_providers(
     return compatible
 
 
+def get_custom_provider_context_length(
+    model: str,
+    base_url: str,
+    custom_providers: Optional[List[Dict[str, Any]]] = None,
+    config: Optional[Dict[str, Any]] = None,
+) -> Optional[int]:
+    """Look up a per-model ``context_length`` override from ``custom_providers``.
+
+    Matches any entry whose ``base_url`` equals ``base_url`` (trailing-slash
+    insensitive) and returns ``custom_providers[i].models.<model>.context_length``
+    if present and valid.  Returns ``None`` when no override applies.
+
+    This is the single source of truth for custom-provider context overrides,
+    used by:
+      * ``AIAgent.__init__`` (startup resolution)
+      * ``AIAgent.switch_model`` (mid-session ``/model`` switch)
+      * ``hermes_cli.model_switch.resolve_display_context_length`` (``/model`` confirmation display)
+      * ``gateway.run._format_session_info`` (``/info`` display)
+      * ``agent.model_metadata.get_model_context_length`` (when custom_providers is threaded through)
+
+    Before this helper existed, the lookup was duplicated in ``run_agent.py``'s
+    startup path only; every other path (notably ``/model`` switch) fell back
+    to the 128K default.  See #15779.
+    """
+    if not model or not base_url:
+        return None
+    if custom_providers is None:
+        try:
+            custom_providers = get_compatible_custom_providers(config)
+        except Exception:
+            if config is None:
+                return None
+            raw = config.get("custom_providers")
+            custom_providers = raw if isinstance(raw, list) else []
+    if not isinstance(custom_providers, list):
+        return None
+
+    target_url = (base_url or "").rstrip("/")
+    if not target_url:
+        return None
+
+    for entry in custom_providers:
+        if not isinstance(entry, dict):
+            continue
+        entry_url = (entry.get("base_url") or "").rstrip("/")
+        if not entry_url or entry_url != target_url:
+            continue
+        models = entry.get("models")
+        if not isinstance(models, dict):
+            continue
+        model_cfg = models.get(model)
+        if not isinstance(model_cfg, dict):
+            continue
+        raw_ctx = model_cfg.get("context_length")
+        if raw_ctx is None:
+            continue
+        try:
+            ctx = int(raw_ctx)
+        except (TypeError, ValueError):
+            continue
+        if ctx > 0:
+            return ctx
+    return None
+
+
 def check_config_version() -> Tuple[int, int]:
     """
     Check config version.

hermes_cli/doctor.py

@@ -320,7 +320,11 @@ def run_doctor(args):
                     known_providers.add("custom:" + name.lower().replace(" ", "-"))
 
             canonical_provider = provider
-            if provider and _resolve_provider_full is not None and provider != "auto":
+            if (
+                provider
+                and _resolve_provider_full is not None
+                and provider not in ("auto", "custom")
+            ):
                 provider_def = _resolve_provider_full(provider, user_providers, custom_providers)
                 canonical_provider = provider_def.id if provider_def is not None else None
 

hermes_cli/fallback_cmd.py

@@ -0,0 +1,361 @@
+"""
+hermes fallback — manage the fallback provider chain.
+
+Fallback providers are tried in order when the primary model fails with
+rate-limit, overload, or connection errors. See:
+https://hermes-agent.nousresearch.com/docs/user-guide/features/fallback-providers
+
+Subcommands:
+  hermes fallback [list]   Show the current fallback chain (default when no subcommand)
+  hermes fallback add      Pick provider + model via the same picker as `hermes model`,
+                           then append the selection to the chain
+  hermes fallback remove   Pick an entry to delete from the chain
+  hermes fallback clear    Remove all fallback entries
+
+Storage: ``fallback_providers`` in ``~/.hermes/config.yaml`` (top-level, list of
+``{provider, model, base_url?, api_mode?}`` dicts).  The legacy single-dict
+``fallback_model`` format is migrated to the new list format on first add.
+"""
+from __future__ import annotations
+
+import copy
+from typing import Any, Dict, List, Optional
+
+
+# ---------------------------------------------------------------------------
+# Helpers
+# ---------------------------------------------------------------------------
+
+def _read_chain(config: Dict[str, Any]) -> List[Dict[str, Any]]:
+    """Return the normalized fallback chain as a list of dicts.
+
+    Accepts both the new list format (``fallback_providers``) and the legacy
+    single-dict format (``fallback_model``).  The returned list is always a
+    fresh copy — callers can mutate without touching the config dict.
+    """
+    chain = config.get("fallback_providers") or []
+    if isinstance(chain, list):
+        result = [dict(e) for e in chain if isinstance(e, dict) and e.get("provider") and e.get("model")]
+        if result:
+            return result
+    legacy = config.get("fallback_model")
+    if isinstance(legacy, dict) and legacy.get("provider") and legacy.get("model"):
+        return [dict(legacy)]
+    if isinstance(legacy, list):
+        return [dict(e) for e in legacy if isinstance(e, dict) and e.get("provider") and e.get("model")]
+    return []
+
+
+def _write_chain(config: Dict[str, Any], chain: List[Dict[str, Any]]) -> None:
+    """Persist the chain to ``fallback_providers`` and clear legacy key."""
+    config["fallback_providers"] = chain
+    # Drop the legacy single-dict key on write so there's only one source of truth.
+    if "fallback_model" in config:
+        config.pop("fallback_model", None)
+
+
+def _format_entry(entry: Dict[str, Any]) -> str:
+    """One-line human-readable rendering of a fallback entry."""
+    provider = entry.get("provider", "?")
+    model = entry.get("model", "?")
+    base = entry.get("base_url")
+    suffix = f"  [{base}]" if base else ""
+    return f"{model}  (via {provider}){suffix}"
+
+
+def _extract_fallback_from_model_cfg(model_cfg: Any) -> Optional[Dict[str, Any]]:
+    """Pull the ``{provider, model, base_url?, api_mode?}`` dict from a ``config["model"]`` snapshot."""
+    if not isinstance(model_cfg, dict):
+        return None
+    provider = (model_cfg.get("provider") or "").strip()
+    # The picker writes the selected model to ``model.default``.
+    model = (model_cfg.get("default") or model_cfg.get("model") or "").strip()
+    if not provider or not model:
+        return None
+    entry: Dict[str, Any] = {"provider": provider, "model": model}
+    base_url = (model_cfg.get("base_url") or "").strip()
+    if base_url:
+        entry["base_url"] = base_url
+    api_mode = (model_cfg.get("api_mode") or "").strip()
+    if api_mode:
+        entry["api_mode"] = api_mode
+    return entry
+
+
+def _snapshot_auth_active_provider() -> Any:
+    """Return the current ``active_provider`` in auth.json, or a sentinel if unavailable."""
+    try:
+        from hermes_cli.auth import _load_auth_store
+        store = _load_auth_store()
+        return store.get("active_provider")
+    except Exception:
+        return None
+
+
+def _restore_auth_active_provider(value: Any) -> None:
+    """Write back a previously snapshotted ``active_provider`` value."""
+    try:
+        from hermes_cli.auth import _auth_store_lock, _load_auth_store, _save_auth_store
+        with _auth_store_lock():
+            store = _load_auth_store()
+            store["active_provider"] = value
+            _save_auth_store(store)
+    except Exception:
+        # Best-effort — if auth.json can't be restored, the user's primary
+        # provider may have been deactivated by the picker.  They can re-run
+        # `hermes model` to fix it.  Don't fail the fallback add.
+        pass
+
+
+# ---------------------------------------------------------------------------
+# Subcommand handlers
+# ---------------------------------------------------------------------------
+
+def cmd_fallback_list(args) -> None:  # noqa: ARG001
+    """Print the current fallback chain."""
+    from hermes_cli.config import load_config
+
+    config = load_config()
+    chain = _read_chain(config)
+
+    print()
+    if not chain:
+        print("  No fallback providers configured.")
+        print()
+        print("  Add one with:  hermes fallback add")
+        print()
+        return
+
+    primary = _describe_primary(config)
+    if primary:
+        print(f"  Primary:   {primary}")
+        print()
+    print(f"  Fallback chain ({len(chain)} {'entry' if len(chain) == 1 else 'entries'}):")
+    for i, entry in enumerate(chain, 1):
+        print(f"    {i}. {_format_entry(entry)}")
+    print()
+    print("  Tried in order when the primary fails (rate-limit, 5xx, connection errors).")
+    print("  Docs: https://hermes-agent.nousresearch.com/docs/user-guide/features/fallback-providers")
+    print()
+
+
+def _describe_primary(config: Dict[str, Any]) -> Optional[str]:
+    """One-line description of the primary model for display purposes."""
+    model_cfg = config.get("model")
+    if isinstance(model_cfg, dict):
+        provider = (model_cfg.get("provider") or "?").strip() or "?"
+        model = (model_cfg.get("default") or model_cfg.get("model") or "?").strip() or "?"
+        return f"{model}  (via {provider})"
+    if isinstance(model_cfg, str) and model_cfg.strip():
+        return model_cfg.strip()
+    return None
+
+
+def cmd_fallback_add(args) -> None:
+    """Launch the same picker as `hermes model`, then append the selection to the chain."""
+    from hermes_cli.main import _require_tty, select_provider_and_model
+    from hermes_cli.config import load_config, save_config
+
+    _require_tty("fallback add")
+
+    # Snapshot BEFORE the picker runs so we can distinguish "user actually
+    # picked something" from "user cancelled" by comparing before/after.
+    before_cfg = load_config()
+    model_before = copy.deepcopy(before_cfg.get("model"))
+    active_provider_before = _snapshot_auth_active_provider()
+
+    print()
+    print("  Adding a fallback provider.  The picker below is the same one used by")
+    print("  `hermes model` — select the provider + model you want as a fallback.")
+    print()
+
+    try:
+        select_provider_and_model(args=args)
+    except SystemExit:
+        # Some provider flows exit on auth failure — restore state and re-raise.
+        _restore_model_cfg(model_before)
+        _restore_auth_active_provider(active_provider_before)
+        raise
+
+    # Read the post-picker state to see what the user selected.
+    after_cfg = load_config()
+    model_after = after_cfg.get("model")
+
+    new_entry = _extract_fallback_from_model_cfg(model_after)
+    if not new_entry:
+        # Picker didn't complete (user cancelled or flow bailed).  Nothing to do.
+        _restore_model_cfg(model_before)
+        _restore_auth_active_provider(active_provider_before)
+        print()
+        print("  No fallback added.")
+        return
+
+    # Picker picked the same thing that's already the primary → nothing changed,
+    # and there's nothing useful to add as a fallback to itself.
+    primary_entry = _extract_fallback_from_model_cfg(model_before)
+    if primary_entry and primary_entry["provider"] == new_entry["provider"] \
+            and primary_entry["model"] == new_entry["model"]:
+        _restore_model_cfg(model_before)
+        _restore_auth_active_provider(active_provider_before)
+        print()
+        print(f"  Selected model matches the current primary ({_format_entry(new_entry)}).")
+        print("  A provider cannot be a fallback for itself — no change.")
+        return
+
+    # Reload the config with the primary restored, then append the new entry
+    # to ``fallback_providers``.  We deliberately re-load (rather than mutating
+    # ``after_cfg``) because the picker may have touched other top-level keys
+    # (custom_providers, providers credentials) that we want to keep.
+    _restore_model_cfg(model_before)
+    _restore_auth_active_provider(active_provider_before)
+
+    final_cfg = load_config()
+    chain = _read_chain(final_cfg)
+
+    # Reject exact-duplicate fallback entries.
+    for existing in chain:
+        if existing.get("provider") == new_entry["provider"] \
+                and existing.get("model") == new_entry["model"]:
+            print()
+            print(f"  {_format_entry(new_entry)} is already in the fallback chain — skipped.")
+            return
+
+    chain.append(new_entry)
+    _write_chain(final_cfg, chain)
+    save_config(final_cfg)
+
+    print()
+    print(f"  Added fallback: {_format_entry(new_entry)}")
+    print(f"  Chain is now {len(chain)} {'entry' if len(chain) == 1 else 'entries'} long.")
+    print()
+    print("  Run `hermes fallback list` to view, or `hermes fallback remove` to delete.")
+
+
+def _restore_model_cfg(model_before: Any) -> None:
+    """Restore ``config["model"]`` to a previously-captured snapshot."""
+    from hermes_cli.config import load_config, save_config
+
+    cfg = load_config()
+    if model_before is None:
+        cfg.pop("model", None)
+    else:
+        cfg["model"] = copy.deepcopy(model_before)
+    save_config(cfg)
+
+
+def cmd_fallback_remove(args) -> None:  # noqa: ARG001
+    """Pick an entry from the chain and remove it."""
+    from hermes_cli.config import load_config, save_config
+
+    config = load_config()
+    chain = _read_chain(config)
+
+    if not chain:
+        print()
+        print("  No fallback providers configured — nothing to remove.")
+        print()
+        return
+
+    choices = [_format_entry(e) for e in chain]
+    choices.append("Cancel")
+
+    try:
+        from hermes_cli.setup import _curses_prompt_choice
+        idx = _curses_prompt_choice("Select a fallback to remove:", choices, 0)
+    except Exception:
+        idx = _numbered_pick("Select a fallback to remove:", choices)
+
+    if idx is None or idx < 0 or idx >= len(chain):
+        print()
+        print("  Cancelled — no change.")
+        return
+
+    removed = chain.pop(idx)
+    _write_chain(config, chain)
+    save_config(config)
+
+    print()
+    print(f"  Removed fallback: {_format_entry(removed)}")
+    if chain:
+        print(f"  Chain is now {len(chain)} {'entry' if len(chain) == 1 else 'entries'} long.")
+    else:
+        print("  Fallback chain is now empty.")
+    print()
+
+
+def cmd_fallback_clear(args) -> None:  # noqa: ARG001
+    """Remove all fallback entries (with confirmation)."""
+    from hermes_cli.config import load_config, save_config
+
+    config = load_config()
+    chain = _read_chain(config)
+
+    if not chain:
+        print()
+        print("  No fallback providers configured — nothing to clear.")
+        print()
+        return
+
+    print()
+    print(f"  Current fallback chain ({len(chain)} {'entry' if len(chain) == 1 else 'entries'}):")
+    for i, entry in enumerate(chain, 1):
+        print(f"    {i}. {_format_entry(entry)}")
+    print()
+    try:
+        resp = input("  Clear all entries? [y/N]: ").strip().lower()
+    except (KeyboardInterrupt, EOFError):
+        print()
+        print("  Cancelled.")
+        return
+    if resp not in ("y", "yes"):
+        print("  Cancelled — no change.")
+        return
+
+    _write_chain(config, [])
+    save_config(config)
+    print()
+    print("  Fallback chain cleared.")
+    print()
+
+
+def _numbered_pick(question: str, choices: List[str]) -> Optional[int]:
+    """Fallback numbered-list picker when curses is unavailable."""
+    print(question)
+    for i, c in enumerate(choices, 1):
+        print(f"  {i}. {c}")
+    print()
+    while True:
+        try:
+            val = input(f"Choice [1-{len(choices)}]: ").strip()
+            if not val:
+                return None
+            idx = int(val) - 1
+            if 0 <= idx < len(choices):
+                return idx
+            print(f"Please enter 1-{len(choices)}")
+        except ValueError:
+            print("Please enter a number")
+        except (KeyboardInterrupt, EOFError):
+            print()
+            return None
+
+
+# ---------------------------------------------------------------------------
+# Dispatch
+# ---------------------------------------------------------------------------
+
+def cmd_fallback(args) -> None:
+    """Top-level dispatcher for ``hermes fallback [subcommand]``."""
+    sub = getattr(args, "fallback_command", None)
+    if sub in (None, "", "list", "ls"):
+        cmd_fallback_list(args)
+    elif sub == "add":
+        cmd_fallback_add(args)
+    elif sub in ("remove", "rm"):
+        cmd_fallback_remove(args)
+    elif sub == "clear":
+        cmd_fallback_clear(args)
+    else:
+        print(f"Unknown fallback subcommand: {sub}")
+        print("Use one of: list, add, remove, clear")
+        raise SystemExit(2)

hermes_cli/hooks.py

@@ -125,6 +125,7 @@ _DEFAULT_PAYLOADS = {
         "task_id": "test-task",
         "tool_call_id": "test-call",
         "result": '{"output": "hello"}',
+        "duration_ms": 42,
     },
     "pre_llm_call": {
         "session_id": "test-session",

hermes_cli/main.py

@@ -839,6 +839,8 @@ def _find_bundled_tui(tui_dir: Path) -> Optional[Path]:
 
 
 def _tui_build_needed(tui_dir: Path) -> bool:
+    if _hermes_ink_bundle_stale(tui_dir):
+        return True
     entry = tui_dir / "dist" / "entry.js"
     if not entry.exists():
         return True
@@ -1026,7 +1028,12 @@ def _make_tui_argv(tui_dir: Path, tui_dev: bool) -> tuple[list[str], Path]:
     return [node, str(root / "dist" / "entry.js")], root
 
 
-def _launch_tui(resume_session_id: Optional[str] = None, tui_dev: bool = False):
+def _launch_tui(
+    resume_session_id: Optional[str] = None,
+    tui_dev: bool = False,
+    model: Optional[str] = None,
+    provider: Optional[str] = None,
+):
     """Replace current process with the TUI."""
     tui_dir = PROJECT_ROOT / "ui-tui"
 
@@ -1036,6 +1043,12 @@ def _launch_tui(resume_session_id: Optional[str] = None, tui_dev: bool = False):
     )
     env.setdefault("HERMES_PYTHON", sys.executable)
     env.setdefault("HERMES_CWD", os.getcwd())
+    if model:
+        env["HERMES_MODEL"] = model
+        env["HERMES_INFERENCE_MODEL"] = model
+    if provider:
+        env["HERMES_TUI_PROVIDER"] = provider
+        env["HERMES_INFERENCE_PROVIDER"] = provider
     # Guarantee an 8GB V8 heap + exposed GC for the TUI. Default node cap is
     # ~1.5–4GB depending on version and can fatal-OOM on long sessions with
     # large transcripts / reasoning blobs. Token-level merge: respect any
@@ -1174,6 +1187,8 @@ def cmd_chat(args):
         _launch_tui(
             getattr(args, "resume", None),
             tui_dev=getattr(args, "tui_dev", False),
+            model=getattr(args, "model", None),
+            provider=getattr(args, "provider", None),
         )
 
     # Import and run the CLI
@@ -1512,6 +1527,83 @@ def select_provider_and_model(args=None):
     all_providers = [(p.slug, p.tui_desc) for p in CANONICAL_PROVIDERS]
 
     def _named_custom_provider_map(cfg) -> dict[str, dict[str, str]]:
+        from hermes_cli.config import read_raw_config
+
+        # Build a lookup of raw (un-expanded) api_key templates keyed by a
+        # stable identity. We intentionally bypass
+        # ``get_compatible_custom_providers(read_raw_config())`` here because
+        # its ``_normalize_custom_provider_entry`` step calls ``urlparse()``
+        # on ``base_url`` and drops any entry whose ``base_url`` is itself an
+        # env-ref template (e.g. ``${NEURALWATT_API_BASE}``). Dropping those
+        # entries is exactly how env-ref preservation fails for the user
+        # config that motivated this fix.
+        raw_api_key_refs: dict[tuple, str] = {}
+        raw_cfg = read_raw_config()
+
+        def _record_raw(
+            name: str,
+            provider_key: str,
+            model: str,
+            api_key: str,
+        ) -> None:
+            template = str(api_key or "").strip()
+            if "${" not in template:
+                return
+            name = str(name or "").strip()
+            provider_key = str(provider_key or "").strip()
+            model = str(model or "").strip()
+            # Index by every plausible identity the loaded (expanded) config
+            # might present: (name), (name, model), (provider_key), and
+            # (provider_key, model). Case-insensitive on name/provider_key so
+            # the loaded entry matches regardless of display casing.
+            if name:
+                raw_api_key_refs.setdefault((name.lower(),), template)
+                raw_api_key_refs.setdefault((name.lower(), model), template)
+            if provider_key:
+                raw_api_key_refs.setdefault((provider_key.lower(),), template)
+                raw_api_key_refs.setdefault(
+                    (provider_key.lower(), model), template
+                )
+
+        raw_list = raw_cfg.get("custom_providers")
+        if isinstance(raw_list, list):
+            for raw_entry in raw_list:
+                if not isinstance(raw_entry, dict):
+                    continue
+                _record_raw(
+                    raw_entry.get("name", ""),
+                    "",
+                    raw_entry.get("model", "")
+                    or raw_entry.get("default_model", ""),
+                    raw_entry.get("api_key", ""),
+                )
+        raw_providers = raw_cfg.get("providers")
+        if isinstance(raw_providers, dict):
+            for raw_key, raw_entry in raw_providers.items():
+                if not isinstance(raw_entry, dict):
+                    continue
+                _record_raw(
+                    raw_entry.get("name", "") or raw_key,
+                    raw_key,
+                    raw_entry.get("model", "")
+                    or raw_entry.get("default_model", ""),
+                    raw_entry.get("api_key", ""),
+                )
+
+        def _lookup_ref(name: str, provider_key: str, model: str) -> str:
+            name_lc = str(name or "").strip().lower()
+            pkey_lc = str(provider_key or "").strip().lower()
+            model = str(model or "").strip()
+            for identity in (
+                (pkey_lc, model),
+                (pkey_lc,),
+                (name_lc, model),
+                (name_lc,),
+            ):
+                if identity[0] and identity in raw_api_key_refs:
+                    return raw_api_key_refs[identity]
+            return ""
+
         custom_provider_map = {}
         for entry in get_compatible_custom_providers(cfg):
             if not isinstance(entry, dict):
@@ -1535,6 +1627,9 @@ def select_provider_and_model(args=None):
                 "model": entry.get("model", ""),
                 "api_mode": entry.get("api_mode", ""),
                 "provider_key": provider_key,
+                "api_key_ref": _lookup_ref(
+                    name, provider_key, entry.get("model", "")
+                ),
             }
         return custom_provider_map
 
@@ -1624,6 +1719,8 @@ def select_provider_and_model(args=None):
         _model_flow_stepfun(config, current_model)
     elif selected_provider == "bedrock":
         _model_flow_bedrock(config, current_model)
+    elif selected_provider == "azure-foundry":
+        _model_flow_azure_foundry(config, current_model)
     elif selected_provider in (
         "gemini",
         "deepseek",
@@ -1707,7 +1804,6 @@ _AUX_TASKS: list[tuple[str, str, str]] = [
     ("session_search",   "Session search",   "past-conversation recall"),
     ("approval",         "Approval",         "smart command approval"),
     ("mcp",              "MCP",              "MCP tool reasoning"),
-    ("flush_memories",   "Flush memories",   "memory consolidation"),
     ("title_generation", "Title generation", "session titles"),
     ("skills_hub",       "Skills hub",       "skills search/install"),
 ]
@@ -2219,13 +2315,13 @@ def _model_flow_nous(config, current_model="", args=None):
     # The live /models endpoint returns hundreds of models; the curated list
     # shows only agentic models users recognize from OpenRouter.
     from hermes_cli.models import (
-        _PROVIDER_MODELS,
+        get_curated_nous_model_ids,
         get_pricing_for_provider,
         check_nous_free_tier,
         partition_nous_models_by_tier,
     )
 
-    model_ids = _PROVIDER_MODELS.get("nous", [])
+    model_ids = get_curated_nous_model_ids()
     if not model_ids:
         print("No curated models available for Nous Portal.")
         return
@@ -2768,6 +2864,19 @@ def _auto_provider_name(base_url: str) -> str:
     return name
 
 
+def _custom_provider_api_key_config_value(provider_info, resolved_api_key=""):
+    """Return the value that should be persisted for a custom provider key."""
+    api_key_ref = str(provider_info.get("api_key_ref", "") or "").strip()
+    if api_key_ref:
+        return api_key_ref
+
+    key_env = str(provider_info.get("key_env", "") or "").strip()
+    if key_env and not str(provider_info.get("api_key", "") or "").strip():
+        return f"${{{key_env}}}"
+
+    return str(resolved_api_key or "").strip()
+
+
 def _save_custom_provider(
     base_url, api_key="", model="", context_length=None, name=None
 ):
@@ -2823,6 +2932,203 @@ def _save_custom_provider(
     print(f'  💾 Saved to custom providers as "{name}" (edit in config.yaml)')
 
 
+def _model_flow_azure_foundry(config, current_model=""):
+    """Azure Foundry provider: configure endpoint, API mode, API key, and model.
+
+    Azure Foundry supports both OpenAI-style (``/v1/chat/completions``) and
+    Anthropic-style (``/v1/messages``) endpoints.  The wizard auto-detects
+    the transport and available models when possible:
+
+    * URLs ending in ``/anthropic`` → Anthropic Messages API.
+    * Successful ``GET <base>/models`` probe → OpenAI-style + populates
+      a picker with the returned deployment / model IDs.
+    * Anthropic Messages probe fallback when ``/models`` fails.
+    * Manual entry when every probe fails (private endpoints, etc.).
+
+    Context lengths for the chosen model are resolved via the standard
+    :func:`agent.model_metadata.get_model_context_length` chain
+    (models.dev, provider metadata, hardcoded family fallbacks).
+    """
+    from hermes_cli.auth import _save_model_choice, deactivate_provider  # noqa: F401
+    from hermes_cli.config import get_env_value, save_env_value, load_config, save_config
+    from hermes_cli import azure_detect
+    import getpass
+
+    # ── Load current Azure Foundry configuration ─────────────────────
+    model_cfg = config.get("model", {})
+    if isinstance(model_cfg, dict) and model_cfg.get("provider") == "azure-foundry":
+        current_base_url = str(model_cfg.get("base_url", "") or "")
+        current_api_mode = str(model_cfg.get("api_mode", "") or "")
+    else:
+        current_base_url = ""
+        current_api_mode = ""
+
+    current_api_key = get_env_value("AZURE_FOUNDRY_API_KEY") or ""
+
+    print()
+    print("Azure Foundry Configuration")
+    print("=" * 50)
+    print()
+    print("Azure Foundry can host models with either OpenAI-style or")
+    print("Anthropic-style API endpoints.  Hermes will probe your")
+    print("endpoint to auto-detect the transport and the deployed")
+    print("models when possible.")
+    print()
+
+    if current_base_url:
+        print(f"  Current endpoint: {current_base_url}")
+    if current_api_mode:
+        _lbl = "OpenAI-style" if current_api_mode == "chat_completions" else "Anthropic-style"
+        print(f"  Current API mode: {_lbl}")
+    if current_api_key:
+        print(f"  Current API key:  {current_api_key[:8]}...")
+    print()
+
+    # ── Step 1: endpoint URL ─────────────────────────────────────────
+    try:
+        base_url = input(
+            f"API endpoint URL [{current_base_url or 'e.g. https://your-resource.openai.azure.com/openai/v1'}]: "
+        ).strip()
+    except (KeyboardInterrupt, EOFError):
+        print("\nCancelled.")
+        return
+
+    effective_url = (base_url or current_base_url).rstrip("/")
+    if not effective_url:
+        print("No endpoint URL provided. Cancelled.")
+        return
+    if not effective_url.startswith(("http://", "https://")):
+        print(f"Invalid URL: {effective_url} (must start with http:// or https://)")
+        return
+
+    # ── Step 2: API key ──────────────────────────────────────────────
+    print()
+    try:
+        api_key = getpass.getpass(
+            f"API key [{current_api_key[:8] + '...' if current_api_key else 'required'}]: "
+        ).strip()
+    except (KeyboardInterrupt, EOFError):
+        print("\nCancelled.")
+        return
+
+    effective_key = api_key or current_api_key
+    if not effective_key:
+        print("No API key provided. Cancelled.")
+        return
+
+    # ── Step 3: auto-detect transport + models ───────────────────────
+    print()
+    print("◐ Probing endpoint to auto-detect transport and models...")
+    detection = azure_detect.detect(effective_url, effective_key)
+
+    discovered_models: list[str] = list(detection.models)
+    api_mode: str = detection.api_mode or ""
+
+    if api_mode:
+        mode_label = "OpenAI-style" if api_mode == "chat_completions" else "Anthropic-style"
+        print(f"✓ Detected API transport: {mode_label}")
+        if detection.reason:
+            print(f"    ({detection.reason})")
+        if discovered_models:
+            print(f"✓ Found {len(discovered_models)} deployed model(s) on this endpoint")
+    else:
+        print(f"⚠ Auto-detection incomplete: {detection.reason}")
+        print()
+        print("Select the API format your Azure Foundry endpoint uses:")
+        print("  1. OpenAI-style  (POST /v1/chat/completions)")
+        print("     For: GPT models, Llama, Mistral, and most open models")
+        print("  2. Anthropic-style  (POST /v1/messages)")
+        print("     For: Claude models deployed via Anthropic API format")
+        try:
+            default_choice = "2" if current_api_mode == "anthropic_messages" else "1"
+            mode_choice = input(f"API format [1/2] ({default_choice}): ").strip() or default_choice
+        except (KeyboardInterrupt, EOFError):
+            print("\nCancelled.")
+            return
+        api_mode = "anthropic_messages" if mode_choice == "2" else "chat_completions"
+
+    # ── Step 4: model name ───────────────────────────────────────────
+    print()
+    effective_model = ""
+    if discovered_models:
+        print("Available models on this endpoint:")
+        for i, mid in enumerate(discovered_models[:30], start=1):
+            print(f"  {i:>2}. {mid}")
+        if len(discovered_models) > 30:
+            print(f"  ... and {len(discovered_models) - 30} more (type name manually if not shown)")
+        print()
+        try:
+            pick = input(
+                f"Pick by number, or type a deployment name [{current_model or discovered_models[0]}]: "
+            ).strip()
+        except (KeyboardInterrupt, EOFError):
+            print("\nCancelled.")
+            return
+        if not pick:
+            effective_model = current_model or discovered_models[0]
+        elif pick.isdigit() and 1 <= int(pick) <= min(len(discovered_models), 30):
+            effective_model = discovered_models[int(pick) - 1]
+        else:
+            effective_model = pick
+    else:
+        try:
+            model_name = input(
+                f"Model / deployment name [{current_model or 'e.g. gpt-5.4, claude-sonnet-4-6'}]: "
+            ).strip()
+        except (KeyboardInterrupt, EOFError):
+            print("\nCancelled.")
+            return
+        effective_model = model_name or current_model
+
+    if not effective_model:
+        print("No model name provided. Cancelled.")
+        return
+
+    # ── Step 5: context-length lookup ────────────────────────────────
+    ctx_len = azure_detect.lookup_context_length(
+        effective_model, effective_url, effective_key,
+    )
+
+    # ── Step 6: persist ──────────────────────────────────────────────
+    save_env_value("AZURE_FOUNDRY_API_KEY", effective_key)
+
+    cfg = load_config()
+    model = cfg.get("model")
+    if not isinstance(model, dict):
+        model = {"default": model} if model else {}
+        cfg["model"] = model
+
+    model["provider"] = "azure-foundry"
+    model["base_url"] = effective_url
+    model["api_mode"] = api_mode
+    model["default"] = effective_model
+    if ctx_len:
+        model["context_length"] = ctx_len
+
+    save_config(cfg)
+    deactivate_provider()
+    config["model"] = dict(model)
+
+    # Clear any conflicting env vars so auxiliary clients don't poison
+    # themselves with a stale OpenAI base URL / key.
+    if get_env_value("OPENAI_BASE_URL"):
+        save_env_value("OPENAI_BASE_URL", "")
+    if get_env_value("OPENAI_API_KEY"):
+        save_env_value("OPENAI_API_KEY", "")
+
+    mode_label = "OpenAI-style" if api_mode == "chat_completions" else "Anthropic-style"
+    print()
+    print("✓ Azure Foundry configured:")
+    print(f"    Endpoint:       {effective_url}")
+    print(f"    API mode:       {mode_label}")
+    print(f"    Model:          {effective_model}")
+    if ctx_len:
+        print(f"    Context length: {ctx_len:,} tokens")
+    else:
+        print("    Context length: not auto-detected (will fall back at runtime)")
+    print()
+
+
 def _remove_custom_provider(config):
     """Let the user remove a saved custom provider from config.yaml."""
     from hermes_cli.config import load_config, save_config
@@ -2909,6 +3215,7 @@ def _model_flow_named_custom(config, provider_info):
     # Resolve key from env var if api_key not set directly
     if not api_key and key_env:
         api_key = os.environ.get(key_env, "")
+    config_api_key = _custom_provider_api_key_config_value(provider_info, api_key)
 
     print(f"  Provider: {name}")
     print(f"  URL:      {base_url}")
@@ -3005,8 +3312,8 @@ def _model_flow_named_custom(config, provider_info):
     else:
         model["provider"] = "custom"
         model["base_url"] = base_url
-        if api_key:
-            model["api_key"] = api_key
+        if config_api_key:
+            model["api_key"] = config_api_key
     # Apply api_mode from custom_providers entry, or clear stale value
     custom_api_mode = provider_info.get("api_mode", "")
     if custom_api_mode:
@@ -3024,15 +3331,15 @@ def _model_flow_named_custom(config, provider_info):
             provider_entry = providers_cfg.get(provider_key)
             if isinstance(provider_entry, dict):
                 provider_entry["default_model"] = model_name
-                if api_key and not str(provider_entry.get("api_key", "") or "").strip():
-                    provider_entry["api_key"] = api_key
+                if config_api_key and not str(provider_entry.get("api_key", "") or "").strip():
+                    provider_entry["api_key"] = config_api_key
                 if key_env and not str(provider_entry.get("key_env", "") or "").strip():
                     provider_entry["key_env"] = key_env
                 cfg["providers"] = providers_cfg
                 save_config(cfg)
     else:
         # Save model name to the custom_providers entry for next time
-        _save_custom_provider(base_url, api_key, model_name)
+        _save_custom_provider(base_url, config_api_key, model_name)
 
     print(f"\n✅ Model set to: {model_name}")
     print(f"   Provider: {name} ({base_url})")
@@ -4473,6 +4780,13 @@ def cmd_webhook(args):
     webhook_command(args)
 
 
+def cmd_kanban(args):
+    """Multi-profile collaboration board."""
+    from hermes_cli.kanban import kanban_command
+
+    return kanban_command(args)
+
+
 def cmd_hooks(args):
     """Shell-hook inspection and management."""
     from hermes_cli.hooks import hooks_command
@@ -5570,6 +5884,54 @@ def _finalize_update_output(state):
             pass
 
 
+def _cmd_update_check():
+    """Implement ``hermes update --check``: fetch and report without installing."""
+    git_dir = PROJECT_ROOT / ".git"
+    if not git_dir.exists():
+        print("✗ Not a git repository — cannot check for updates.")
+        sys.exit(1)
+
+    git_cmd = ["git"]
+    if sys.platform == "win32":
+        git_cmd = ["git", "-c", "windows.appendAtomically=false"]
+
+    print("→ Fetching from origin...")
+    fetch_result = subprocess.run(
+        git_cmd + ["fetch", "origin"],
+        cwd=PROJECT_ROOT,
+        capture_output=True,
+        text=True,
+    )
+    if fetch_result.returncode != 0:
+        stderr = fetch_result.stderr.strip()
+        if "Could not resolve host" in stderr or "unable to access" in stderr:
+            print("✗ Network error — cannot reach the remote repository.")
+        elif "Authentication failed" in stderr or "could not read Username" in stderr:
+            print("✗ Authentication failed — check your git credentials or SSH key.")
+        else:
+            print("✗ Failed to fetch from origin.")
+            if stderr:
+                print(f"  {stderr.splitlines()[0]}")
+        sys.exit(1)
+
+    rev_result = subprocess.run(
+        git_cmd + ["rev-list", "HEAD..origin/main", "--count"],
+        cwd=PROJECT_ROOT,
+        capture_output=True,
+        text=True,
+        check=True,
+    )
+    behind = int(rev_result.stdout.strip())
+
+    if behind == 0:
+        print("✓ Already up to date.")
+    else:
+        commits_word = "commit" if behind == 1 else "commits"
+        print(f"⚕ Update available: {behind} {commits_word} behind origin/main.")
+        from hermes_cli.config import recommended_update_command
+        print(f"  Run '{recommended_update_command()}' to install.")
+
+
 def cmd_update(args):
     """Update Hermes Agent to the latest version.
 
@@ -5583,6 +5945,10 @@ def cmd_update(args):
         managed_error("update Hermes Agent")
         return
 
+    if getattr(args, "check", False):
+        _cmd_update_check()
+        return
+
     gateway_mode = getattr(args, "gateway", False)
 
     # Protect against mid-update terminal disconnects (SIGHUP) and tolerate
@@ -6046,6 +6412,75 @@ def _cmd_update_impl(args, gateway_mode: bool):
             )
             import signal as _signal
 
+            def _wait_for_service_active(
+                scope_cmd_: list, svc_name_: str, timeout: float = 10.0,
+            ) -> bool:
+                """Poll ``systemctl is-active`` until the unit reports active.
+
+                systemd's Stopped -> Started transition after a graceful exit
+                (or a hard restart) is not instantaneous; a one-shot check
+                races that window and falsely reports the unit as down.
+                Poll every 0.5s up to ``timeout`` seconds before giving up.
+                """
+                deadline = _time.monotonic() + max(timeout, 0.5)
+                while True:
+                    try:
+                        _verify = subprocess.run(
+                            scope_cmd_ + ["is-active", svc_name_],
+                            capture_output=True, text=True, timeout=5,
+                        )
+                        if _verify.stdout.strip() == "active":
+                            return True
+                    except (FileNotFoundError, subprocess.TimeoutExpired):
+                        pass
+                    if _time.monotonic() >= deadline:
+                        return False
+                    _time.sleep(0.5)
+
+            def _service_restart_sec(
+                scope_cmd_: list, svc_name_: str, default: float = 0.0,
+            ) -> float:
+                """Read the unit's ``RestartUSec`` (RestartSec) in seconds.
+
+                After a graceful exit-75, systemd waits ``RestartSec`` before
+                respawning the unit.  Callers that poll for ``is-active``
+                must use a timeout >= ``RestartSec`` + transition slack, or
+                they'll give up *during* the cooldown window and wrongly
+                conclude the unit didn't relaunch.
+                """
+                try:
+                    _show = subprocess.run(
+                        scope_cmd_ + [
+                            "show", svc_name_,
+                            "--property=RestartUSec", "--value",
+                        ],
+                        capture_output=True, text=True, timeout=5,
+                    )
+                except (FileNotFoundError, subprocess.TimeoutExpired):
+                    return default
+                raw = (_show.stdout or "").strip()
+                # systemd emits values like "30s", "100ms", "1min 30s", or
+                # "infinity".  Parse conservatively; on any miss return default.
+                if not raw or raw == "infinity":
+                    return default
+                total = 0.0
+                matched = False
+                for part in raw.split():
+                    for _suf, _mult in (
+                        ("ms", 0.001),
+                        ("us", 0.000001),
+                        ("min", 60.0),
+                        ("s", 1.0),
+                    ):
+                        if part.endswith(_suf):
+                            try:
+                                total += float(part[: -len(_suf)]) * _mult
+                                matched = True
+                            except ValueError:
+                                pass
+                            break
+                return total if matched else default
+
             # Drain budget for graceful SIGUSR1 restarts.  The gateway drains
             # for up to ``agent.restart_drain_timeout`` (default 60s) before
             # exiting with code 75; we wait slightly longer so the drain
@@ -6152,14 +6587,23 @@ def _cmd_update_impl(args, gateway_mode: bool):
 
                             if _graceful_ok:
                                 # Gateway exited 75; systemd should relaunch
-                                # via Restart=on-failure.  Verify the new
-                                # process came up.
-                                _time.sleep(3)
-                                verify = subprocess.run(
-                                    scope_cmd + ["is-active", svc_name],
-                                    capture_output=True, text=True, timeout=5,
+                                # via Restart=on-failure.  The unit's
+                                # RestartSec (default 30s on ours) gates the
+                                # respawn — poll past that + slack so we
+                                # don't give up mid-cooldown and falsely
+                                # print "drained but didn't relaunch".  For
+                                # units without RestartSec set we fall back
+                                # to the original 10s budget.
+                                _restart_sec = _service_restart_sec(
+                                    scope_cmd, svc_name, default=0.0,
                                 )
-                                if verify.stdout.strip() == "active":
+                                _post_drain_timeout = max(
+                                    10.0, _restart_sec + 10.0,
+                                )
+                                if _wait_for_service_active(
+                                    scope_cmd, svc_name,
+                                    timeout=_post_drain_timeout,
+                                ):
                                     restarted_services.append(svc_name)
                                     continue
                                 # Process exited but wasn't respawned (older
@@ -6185,14 +6629,9 @@ def _cmd_update_impl(args, gateway_mode: bool):
                                 # Verify the service actually survived the
                                 # restart.  systemctl restart returns 0 even
                                 # if the new process crashes immediately.
-                                _time.sleep(3)
-                                verify = subprocess.run(
-                                    scope_cmd + ["is-active", svc_name],
-                                    capture_output=True,
-                                    text=True,
-                                    timeout=5,
-                                )
-                                if verify.stdout.strip() == "active":
+                                if _wait_for_service_active(
+                                    scope_cmd, svc_name, timeout=10.0,
+                                ):
                                     restarted_services.append(svc_name)
                                 else:
                                     # Retry once — transient startup failures
@@ -6207,14 +6646,9 @@ def _cmd_update_impl(args, gateway_mode: bool):
                                         text=True,
                                         timeout=15,
                                     )
-                                    _time.sleep(3)
-                                    verify2 = subprocess.run(
-                                        scope_cmd + ["is-active", svc_name],
-                                        capture_output=True,
-                                        text=True,
-                                        timeout=5,
-                                    )
-                                    if verify2.stdout.strip() == "active":
+                                    if _wait_for_service_active(
+                                        scope_cmd, svc_name, timeout=10.0,
+                                    ):
                                         restarted_services.append(svc_name)
                                         print(f"  ✓ {svc_name} recovered on retry")
                                     else:
@@ -6796,6 +7230,9 @@ Examples:
     hermes auth remove <p> <t>    Remove pooled credential by index, id, or label
     hermes auth reset <provider>  Clear exhaustion status for a provider
     hermes model                  Select default model
+    hermes fallback [list]        Show fallback provider chain
+    hermes fallback add           Add a fallback provider (same picker as `hermes model`)
+    hermes fallback remove        Remove a fallback provider from the chain
     hermes config                 View configuration
     hermes config edit            Edit config in $EDITOR
     hermes config set model gpt-4 Set a config value
@@ -6821,6 +7258,40 @@ For more help on a command:
     parser.add_argument(
         "--version", "-V", action="store_true", help="Show version and exit"
     )
+    parser.add_argument(
+        "-z",
+        "--oneshot",
+        metavar="PROMPT",
+        default=None,
+        help=(
+            "One-shot mode: send a single prompt and print ONLY the final "
+            "response text to stdout. No banner, no spinner, no tool "
+            "previews, no session_id line. Tools, memory, rules, and "
+            "AGENTS.md in the CWD are loaded as normal; approvals are "
+            "auto-bypassed. Intended for scripts / pipes."
+        ),
+    )
+    # --model / --provider are accepted at the top level so they can pair
+    # with -z without needing the `chat` subcommand.  If neither -z nor a
+    # subcommand consumes them, they fall through harmlessly as None.
+    # Mirrors `hermes chat --model ... --provider ...` semantics.
+    parser.add_argument(
+        "-m",
+        "--model",
+        default=None,
+        help=(
+            "Model override for this invocation (e.g. anthropic/claude-sonnet-4.6). "
+            "Applies to -z/--oneshot and --tui. Also settable via HERMES_INFERENCE_MODEL env var."
+        ),
+    )
+    parser.add_argument(
+        "--provider",
+        default=None,
+        help=(
+            "Provider override for this invocation (e.g. openrouter, anthropic). "
+            "Applies to -z/--oneshot and --tui. Also settable via HERMES_INFERENCE_PROVIDER env var."
+        ),
+    )
     parser.add_argument(
         "--resume",
         "-r",
@@ -7103,6 +7574,42 @@ For more help on a command:
     )
     model_parser.set_defaults(func=cmd_model)
 
+    # =========================================================================
+    # fallback command — manage the fallback provider chain
+    # =========================================================================
+    from hermes_cli.fallback_cmd import cmd_fallback
+
+    fallback_parser = subparsers.add_parser(
+        "fallback",
+        help="Manage fallback providers (tried when the primary model fails)",
+        description=(
+            "Manage the fallback provider chain.  Fallback providers are tried "
+            "in order when the primary model fails with rate-limit, overload, or "
+            "connection errors.  See: "
+            "https://hermes-agent.nousresearch.com/docs/user-guide/features/fallback-providers"
+        ),
+    )
+    fallback_subparsers = fallback_parser.add_subparsers(dest="fallback_command")
+    fallback_subparsers.add_parser(
+        "list",
+        aliases=["ls"],
+        help="Show the current fallback chain (default when no subcommand)",
+    )
+    fallback_subparsers.add_parser(
+        "add",
+        help="Pick a provider + model (same picker as `hermes model`) and append to the chain",
+    )
+    fallback_subparsers.add_parser(
+        "remove",
+        aliases=["rm"],
+        help="Pick an entry to delete from the chain",
+    )
+    fallback_subparsers.add_parser(
+        "clear",
+        help="Remove all fallback entries",
+    )
+    fallback_parser.set_defaults(func=cmd_fallback)
+
     # =========================================================================
     # gateway command
     # =========================================================================
@@ -7273,6 +7780,19 @@ For more help on a command:
     setup_parser.add_argument(
         "--reset", action="store_true", help="Reset configuration to defaults"
     )
+    setup_parser.add_argument(
+        "--reconfigure",
+        action="store_true",
+        help="(Default on existing installs.) Re-run the full wizard, "
+             "showing current values as defaults. Kept for backwards "
+             "compatibility — a bare 'hermes setup' now does this.",
+    )
+    setup_parser.add_argument(
+        "--quick",
+        action="store_true",
+        help="On existing installs: only prompt for items that are missing "
+             "or unset, instead of running the full reconfigure wizard.",
+    )
     setup_parser.set_defaults(func=cmd_setup)
 
     # =========================================================================
@@ -7603,6 +8123,13 @@ For more help on a command:
 
     webhook_parser.set_defaults(func=cmd_webhook)
 
+    # =========================================================================
+    # kanban command — multi-profile collaboration board
+    # =========================================================================
+    from hermes_cli.kanban import build_parser as _build_kanban_parser
+    kanban_parser = _build_kanban_parser(subparsers)
+    kanban_parser.set_defaults(func=cmd_kanban)
+
     # =========================================================================
     # hooks command — shell-hook inspection and management
     # =========================================================================
@@ -8755,6 +9282,12 @@ Examples:
         default=False,
         help="Gateway mode: use file-based IPC for prompts instead of stdin (used internally by /update)",
     )
+    update_parser.add_argument(
+        "--check",
+        action="store_true",
+        default=False,
+        help="Check whether an update is available without installing anything",
+    )
     update_parser.set_defaults(func=cmd_update)
 
     # =========================================================================
@@ -9101,6 +9634,17 @@ Examples:
                 exc_info=True,
             )
 
+    # Handle top-level --oneshot / -z: single-shot mode, stdout = final
+    # response only, nothing else. Bypasses cli.py entirely.
+    if getattr(args, "oneshot", None):
+        from hermes_cli.oneshot import run_oneshot
+
+        sys.exit(run_oneshot(
+            args.oneshot,
+            model=getattr(args, "model", None),
+            provider=getattr(args, "provider", None),
+        ))
+
     # Handle top-level --resume / --continue as shortcut to chat
     if (args.resume or args.continue_last) and args.command is None:
         args.command = "chat"

hermes_cli/model_catalog.py

@@ -0,0 +1,329 @@
+"""Remote model catalog fetcher.
+
+The Hermes docs site hosts a JSON manifest of curated models for providers
+we want to update without shipping a release (currently OpenRouter and
+Nous Portal). This module fetches, validates, and caches that manifest,
+falling back to the in-repo hardcoded lists when the network is unavailable.
+
+Pipeline
+--------
+1. ``get_catalog()`` — returns a parsed manifest dict.
+   - Checks in-process cache (invalidated by TTL).
+   - Reads disk cache at ``~/.hermes/cache/model_catalog.json``.
+   - Fetches the master URL if disk cache is stale or missing.
+   - On any fetch failure, keeps using the stale cache (or empty dict).
+
+2. ``get_curated_openrouter_models()`` / ``get_curated_nous_models()`` —
+   thin accessors returning the shapes existing callers expect. Each
+   falls back to the in-repo hardcoded list on any lookup failure.
+
+Schema (version 1)
+------------------
+::
+
+    {
+      "version": 1,
+      "updated_at": "2026-04-25T22:00:00Z",
+      "metadata": {...},                # free-form
+      "providers": {
+        "openrouter": {
+          "metadata": {...},            # free-form
+          "models": [
+            {"id": "vendor/model", "description": "recommended",
+             "metadata": {...}}          # free-form, model-level
+          ]
+        },
+        "nous": {...}
+      }
+    }
+
+Unknown fields are ignored — extra metadata can be added at either level
+without bumping ``version``. ``version`` bumps are reserved for
+breaking changes (renaming ``providers``, changing ``models`` shape).
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+import os
+import time
+import urllib.error
+import urllib.request
+from pathlib import Path
+from typing import Any
+
+from hermes_cli import __version__ as _HERMES_VERSION
+
+logger = logging.getLogger(__name__)
+
+# ---------------------------------------------------------------------------
+# Constants
+# ---------------------------------------------------------------------------
+
+DEFAULT_CATALOG_URL = (
+    "https://hermes-agent.nousresearch.com/docs/api/model-catalog.json"
+)
+DEFAULT_TTL_HOURS = 24
+DEFAULT_FETCH_TIMEOUT = 8.0
+SUPPORTED_SCHEMA_VERSION = 1
+
+_HERMES_USER_AGENT = f"hermes-cli/{_HERMES_VERSION}"
+
+# In-process cache to avoid repeated disk + parse work across multiple
+# calls within the same session. Invalidated by TTL against the disk file's
+# mtime, so calling code never has to think about this.
+_catalog_cache: dict[str, Any] | None = None
+_catalog_cache_source_mtime: float = 0.0
+
+
+# ---------------------------------------------------------------------------
+# Config
+# ---------------------------------------------------------------------------
+
+
+def _load_catalog_config() -> dict[str, Any]:
+    """Load the ``model_catalog`` config block with defaults filled in."""
+    try:
+        from hermes_cli.config import load_config
+        cfg = load_config() or {}
+    except Exception:
+        cfg = {}
+
+    raw = cfg.get("model_catalog")
+    if not isinstance(raw, dict):
+        raw = {}
+
+    return {
+        "enabled": bool(raw.get("enabled", True)),
+        "url": str(raw.get("url") or DEFAULT_CATALOG_URL),
+        "ttl_hours": float(raw.get("ttl_hours") or DEFAULT_TTL_HOURS),
+        "providers": raw.get("providers") if isinstance(raw.get("providers"), dict) else {},
+    }
+
+
+def _cache_path() -> Path:
+    """Return the disk cache path. Import lazily so tests can monkeypatch home."""
+    from hermes_constants import get_hermes_home
+    return get_hermes_home() / "cache" / "model_catalog.json"
+
+
+# ---------------------------------------------------------------------------
+# Fetch + validate + cache
+# ---------------------------------------------------------------------------
+
+
+def _fetch_manifest(url: str, timeout: float) -> dict[str, Any] | None:
+    """HTTP GET the manifest URL and return a parsed dict, or None on failure."""
+    try:
+        req = urllib.request.Request(
+            url,
+            headers={
+                "Accept": "application/json",
+                "User-Agent": _HERMES_USER_AGENT,
+            },
+        )
+        with urllib.request.urlopen(req, timeout=timeout) as resp:
+            data = json.loads(resp.read().decode())
+    except (urllib.error.URLError, TimeoutError, json.JSONDecodeError, OSError) as exc:
+        logger.info("model catalog fetch failed (%s): %s", url, exc)
+        return None
+    except Exception as exc:  # pragma: no cover — defensive
+        logger.info("model catalog fetch errored (%s): %s", url, exc)
+        return None
+
+    if not _validate_manifest(data):
+        logger.info("model catalog at %s failed schema validation", url)
+        return None
+
+    return data
+
+
+def _validate_manifest(data: Any) -> bool:
+    """Return True when ``data`` matches the minimum manifest shape."""
+    if not isinstance(data, dict):
+        return False
+    version = data.get("version")
+    if not isinstance(version, int) or version > SUPPORTED_SCHEMA_VERSION:
+        # Future schema version we don't understand — refuse rather than
+        # guess. Older schemas (version < 1) aren't supported either.
+        return False
+    providers = data.get("providers")
+    if not isinstance(providers, dict):
+        return False
+    for pname, pblock in providers.items():
+        if not isinstance(pname, str) or not isinstance(pblock, dict):
+            return False
+        models = pblock.get("models")
+        if not isinstance(models, list):
+            return False
+        for m in models:
+            if not isinstance(m, dict):
+                return False
+            if not isinstance(m.get("id"), str) or not m["id"].strip():
+                return False
+    return True
+
+
+def _read_disk_cache() -> tuple[dict[str, Any] | None, float]:
+    """Return ``(data_or_none, mtime)``. mtime is 0 if file is missing."""
+    path = _cache_path()
+    try:
+        mtime = path.stat().st_mtime
+    except (OSError, FileNotFoundError):
+        return (None, 0.0)
+    try:
+        with open(path) as fh:
+            data = json.load(fh)
+    except (OSError, json.JSONDecodeError):
+        return (None, 0.0)
+    if not _validate_manifest(data):
+        return (None, 0.0)
+    return (data, mtime)
+
+
+def _write_disk_cache(data: dict[str, Any]) -> None:
+    path = _cache_path()
+    try:
+        path.parent.mkdir(parents=True, exist_ok=True)
+        tmp = path.with_suffix(path.suffix + ".tmp")
+        with open(tmp, "w") as fh:
+            json.dump(data, fh, indent=2)
+            fh.write("\n")
+        os.replace(tmp, path)
+    except OSError as exc:
+        logger.info("model catalog cache write failed: %s", exc)
+
+
+# ---------------------------------------------------------------------------
+# Public API
+# ---------------------------------------------------------------------------
+
+
+def get_catalog(*, force_refresh: bool = False) -> dict[str, Any]:
+    """Return the parsed model catalog manifest, or an empty dict on failure.
+
+    Callers should treat a missing provider/model as "use the in-repo fallback"
+    — never raise from this function so the CLI keeps working offline.
+    """
+    global _catalog_cache, _catalog_cache_source_mtime
+
+    cfg = _load_catalog_config()
+    if not cfg["enabled"]:
+        return {}
+
+    ttl_seconds = max(0.0, cfg["ttl_hours"] * 3600.0)
+
+    disk_data, disk_mtime = _read_disk_cache()
+    now = time.time()
+    disk_fresh = disk_data is not None and (now - disk_mtime) < ttl_seconds
+
+    # In-process cache hit: disk hasn't changed since we loaded it and still fresh.
+    if (
+        not force_refresh
+        and _catalog_cache is not None
+        and disk_data is not None
+        and disk_mtime == _catalog_cache_source_mtime
+        and disk_fresh
+    ):
+        return _catalog_cache
+
+    # Disk is fresh enough — use it without a network hit.
+    if not force_refresh and disk_fresh and disk_data is not None:
+        _catalog_cache = disk_data
+        _catalog_cache_source_mtime = disk_mtime
+        return disk_data
+
+    # Need to (re)fetch. If it fails, fall back to any stale disk copy.
+    fetched = _fetch_manifest(cfg["url"], DEFAULT_FETCH_TIMEOUT)
+    if fetched is not None:
+        _write_disk_cache(fetched)
+        new_disk_data, new_mtime = _read_disk_cache()
+        if new_disk_data is not None:
+            _catalog_cache = new_disk_data
+            _catalog_cache_source_mtime = new_mtime
+            return new_disk_data
+        _catalog_cache = fetched
+        _catalog_cache_source_mtime = now
+        return fetched
+
+    if disk_data is not None:
+        _catalog_cache = disk_data
+        _catalog_cache_source_mtime = disk_mtime
+        return disk_data
+
+    return {}
+
+
+def _fetch_provider_override(provider: str) -> dict[str, Any] | None:
+    """If ``model_catalog.providers.<name>.url`` is set, fetch that instead."""
+    cfg = _load_catalog_config()
+    if not cfg["enabled"]:
+        return None
+    provider_cfg = cfg["providers"].get(provider)
+    if not isinstance(provider_cfg, dict):
+        return None
+    override_url = provider_cfg.get("url")
+    if not isinstance(override_url, str) or not override_url.strip():
+        return None
+    # Override fetches skip the disk cache because they're usually
+    # third-party self-hosted. Re-request on every call but with a short
+    # timeout so they don't block the picker.
+    return _fetch_manifest(override_url.strip(), DEFAULT_FETCH_TIMEOUT)
+
+
+def _get_provider_block(provider: str) -> dict[str, Any] | None:
+    """Return the provider's manifest block, respecting per-provider overrides."""
+    override = _fetch_provider_override(provider)
+    if override is not None:
+        block = override.get("providers", {}).get(provider)
+        if isinstance(block, dict):
+            return block
+
+    catalog = get_catalog()
+    if not catalog:
+        return None
+    block = catalog.get("providers", {}).get(provider)
+    return block if isinstance(block, dict) else None
+
+
+def get_curated_openrouter_models() -> list[tuple[str, str]] | None:
+    """Return OpenRouter's curated ``[(id, description), ...]`` from the manifest.
+
+    Returns ``None`` when the manifest is unavailable, so callers can fall
+    back to their hardcoded list.
+    """
+    block = _get_provider_block("openrouter")
+    if not block:
+        return None
+    out: list[tuple[str, str]] = []
+    for m in block.get("models", []):
+        mid = str(m.get("id") or "").strip()
+        if not mid:
+            continue
+        desc = str(m.get("description") or "")
+        out.append((mid, desc))
+    return out or None
+
+
+def get_curated_nous_models() -> list[str] | None:
+    """Return Nous Portal's curated list of model ids from the manifest.
+
+    Returns ``None`` when the manifest is unavailable.
+    """
+    block = _get_provider_block("nous")
+    if not block:
+        return None
+    out: list[str] = []
+    for m in block.get("models", []):
+        mid = str(m.get("id") or "").strip()
+        if mid:
+            out.append(mid)
+    return out or None
+
+
+def reset_cache() -> None:
+    """Clear the in-process cache. Used by tests and ``hermes model --refresh``."""
+    global _catalog_cache, _catalog_cache_source_mtime
+    _catalog_cache = None
+    _catalog_cache_source_mtime = 0.0

hermes_cli/model_switch.py

@@ -527,6 +527,49 @@ def _resolve_alias_fallback(
     return None
 
 
+def resolve_display_context_length(
+    model: str,
+    provider: str,
+    base_url: str = "",
+    api_key: str = "",
+    model_info: Optional[ModelInfo] = None,
+    custom_providers: list | None = None,
+) -> Optional[int]:
+    """Resolve the context length to show in /model output.
+
+    models.dev reports per-vendor context (e.g. gpt-5.5 = 1.05M on openai)
+    but provider-enforced limits can be lower (e.g. Codex OAuth caps the
+    same slug at 272k). The authoritative source is
+    ``agent.model_metadata.get_model_context_length`` which already knows
+    about Codex OAuth, Copilot, Nous, and falls back to models.dev for the
+    rest.
+
+    When ``custom_providers`` is provided, per-model ``context_length``
+    overrides from ``custom_providers[].models.<id>.context_length`` are
+    honored — this closes #15779 where ``/model`` switch ignored user-set
+    overrides.
+
+    Prefer the provider-aware value; fall back to ``model_info.context_window``
+    only if the resolver returns nothing.
+    """
+    try:
+        from agent.model_metadata import get_model_context_length
+        ctx = get_model_context_length(
+            model,
+            base_url=base_url or "",
+            api_key=api_key or "",
+            provider=provider or None,
+            custom_providers=custom_providers,
+        )
+        if ctx:
+            return int(ctx)
+    except Exception:
+        pass
+    if model_info is not None and model_info.context_window:
+        return int(model_info.context_window)
+    return None
+
+
 # ---------------------------------------------------------------------------
 # Core model-switching pipeline
 # ---------------------------------------------------------------------------
@@ -795,9 +838,14 @@ def switch_model(
                 requested=current_provider,
                 target_model=new_model,
             )
-            api_key = runtime.get("api_key", "")
-            base_url = runtime.get("base_url", "")
-            api_mode = runtime.get("api_mode", "")
+            # If resolution fell through to "custom" (e.g. named custom provider like
+            # "ollama-launch" that resolve_runtime_provider doesn't know), keep existing
+            # credentials. Otherwise use the resolved values (picks up credential rotation,
+            # base_url adjustments for OpenCode, etc.).
+            if runtime.get("provider") != "custom":
+                api_key = runtime.get("api_key", "")
+                base_url = runtime.get("base_url", "")
+                api_mode = runtime.get("api_mode", "")
         except Exception:
             pass
 
@@ -831,16 +879,31 @@ def switch_model(
             "message": f"Could not validate `{new_model}`: {e}",
         }
 
+    # Override rejection if model is in the user's saved provider config.
+    # API /v1/models may not list cloud/aliased models even though the server supports them.
     if not validation.get("accepted"):
-        msg = validation.get("message", "Invalid model")
-        return ModelSwitchResult(
-            success=False,
-            new_model=new_model,
-            target_provider=target_provider,
-            provider_label=provider_label,
-            is_global=is_global,
-            error_message=msg,
-        )
+        override = False
+        if user_providers:
+            for up in user_providers:
+                if isinstance(up, dict) and up.get("provider") == target_provider:
+                    cfg_models = up.get("models", [])
+                    if new_model in cfg_models or any(
+                        m.get("name") == new_model for m in cfg_models if isinstance(m, dict)
+                    ):
+                        override = True
+                        break
+        if override:
+            validation = {"accepted": True, "persist": True, "recognized": False, "message": validation.get("message", "")}
+        else:
+            msg = validation.get("message", "Invalid model")
+            return ModelSwitchResult(
+                success=False,
+                new_model=new_model,
+                target_provider=target_provider,
+                provider_label=provider_label,
+                is_global=is_global,
+                error_message=msg,
+            )
 
     # Apply auto-correction if validation found a closer match
     if validation.get("corrected_model"):

hermes_cli/models.py

@@ -383,6 +383,9 @@ _PROVIDER_MODELS: dict[str, list[str]] = {
         "us.meta.llama4-maverick-17b-instruct-v1:0",
         "us.meta.llama4-scout-17b-instruct-v1:0",
     ],
+    # Azure Foundry: user-provided endpoint and model.
+    # Empty list because models depend on the endpoint configuration.
+    "azure-foundry": [],
 }
 
 # Vercel AI Gateway: derive the bare-model-id catalog from the curated
@@ -740,6 +743,7 @@ CANONICAL_PROVIDERS: list[ProviderEntry] = [
     ProviderEntry("opencode-zen",   "OpenCode Zen",             "OpenCode Zen (35+ curated models, pay-as-you-go)"),
     ProviderEntry("opencode-go",    "OpenCode Go",              "OpenCode Go (open models, $10/month subscription)"),
     ProviderEntry("bedrock",        "AWS Bedrock",              "AWS Bedrock (Claude, Nova, Llama, DeepSeek — IAM or API key)"),
+    ProviderEntry("azure-foundry",  "Azure Foundry",            "Azure Foundry (OpenAI-style or Anthropic-style endpoint — your Azure AI deployment)"),
 ]
 
 # Derived dicts — used throughout the codebase
@@ -872,7 +876,16 @@ def fetch_openrouter_models(
     if _openrouter_catalog_cache is not None and not force_refresh:
         return list(_openrouter_catalog_cache)
 
-    fallback = list(OPENROUTER_MODELS)
+    # Prefer the remotely-hosted catalog manifest; fall back to the in-repo
+    # snapshot when the manifest is unreachable. Both are curated lists that
+    # drive the picker; the OpenRouter live /v1/models filter (tool support,
+    # free pricing) is applied on top either way.
+    try:
+        from hermes_cli.model_catalog import get_curated_openrouter_models
+        remote = get_curated_openrouter_models()
+    except Exception:
+        remote = None
+    fallback = list(remote) if remote else list(OPENROUTER_MODELS)
     preferred_ids = [mid for mid, _ in fallback]
 
     try:
@@ -925,6 +938,24 @@ def model_ids(*, force_refresh: bool = False) -> list[str]:
     return [mid for mid, _ in fetch_openrouter_models(force_refresh=force_refresh)]
 
 
+def get_curated_nous_model_ids() -> list[str]:
+    """Return the curated Nous Portal model-id list.
+
+    Prefers the remotely-hosted catalog manifest (published under
+    ``website/static/api/model-catalog.json``); falls back to the in-repo
+    snapshot in ``_PROVIDER_MODELS["nous"]`` when the manifest is
+    unreachable. Always returns a list (never None).
+    """
+    try:
+        from hermes_cli.model_catalog import get_curated_nous_models
+        remote = get_curated_nous_models()
+    except Exception:
+        remote = None
+    if remote:
+        return list(remote)
+    return list(_PROVIDER_MODELS.get("nous", []))
+
+
 def _ai_gateway_model_is_free(pricing: Any) -> bool:
     """Return True if an AI Gateway model has $0 input AND output pricing."""
     if not isinstance(pricing, dict):
@@ -1379,27 +1410,93 @@ def curated_models_for_provider(
     return [(m, "") for m in models]
 
 
-def detect_provider_for_model(
+def _provider_keys(provider: str) -> set[str]:
+    key = (provider or "").strip().lower()
+    normalized = normalize_provider(provider)
+    return {k for k in (key, normalized) if k}
+
+
+def _model_in_provider_catalog(name_lower: str, providers: set[str]) -> bool:
+    return any(
+        name_lower == model.lower()
+        for provider in providers
+        for model in _PROVIDER_MODELS.get(provider, [])
+    )
+
+
+_AGGREGATOR_PROVIDERS = frozenset(
+    {"nous", "openrouter", "ai-gateway", "copilot", "kilocode"}
+)
+
+
+def _resolve_static_model_alias(
+    name_lower: str,
+    current_keys: set[str],
+) -> Optional[tuple[str, str]]:
+    """Resolve short aliases (e.g. sonnet/opus) using static catalogs only."""
+    try:
+        from hermes_cli.model_switch import MODEL_ALIASES
+    except Exception:
+        return None
+
+    identity = MODEL_ALIASES.get(name_lower)
+    if identity is None:
+        return None
+
+    vendor = identity.vendor
+    family = identity.family
+
+    def _match(provider: str) -> Optional[str]:
+        models = _PROVIDER_MODELS.get(provider, [])
+        if not models:
+            return None
+        prefix = (
+            f"{vendor}/{family}"
+            if provider in _AGGREGATOR_PROVIDERS
+            else family
+        ).lower()
+        for model in models:
+            if model.lower().startswith(prefix):
+                return model
+        return None
+
+    for provider in current_keys:
+        if matched := _match(provider):
+            return provider, matched
+
+    for provider in _PROVIDER_MODELS:
+        if provider in current_keys or provider in _AGGREGATOR_PROVIDERS:
+            continue
+        if matched := _match(provider):
+            return provider, matched
+
+    for provider in _AGGREGATOR_PROVIDERS:
+        if provider in current_keys and (matched := _match(provider)):
+            return provider, matched
+
+    return None
+
+
+def detect_static_provider_for_model(
     model_name: str,
     current_provider: str,
 ) -> Optional[tuple[str, str]]:
-    """Auto-detect the best provider for a model name.
+    """Auto-detect a provider from static catalogs only.
 
-    Returns ``(provider_id, model_name)`` — the model name may be remapped
-    (e.g. bare ``deepseek-chat`` → ``deepseek/deepseek-chat`` for OpenRouter).
+    Returns ``(provider_id, model_name)``. The model name may be remapped
+    when a static alias or bare provider name resolves to a catalog default.
     Returns ``None`` when no confident match is found.
-
-    Priority:
-    0. Bare provider name → switch to that provider's default model
-    1. Direct provider with credentials (highest)
-    2. Direct provider without credentials → remap to OpenRouter slug
-    3. OpenRouter catalog match
     """
     name = (model_name or "").strip()
     if not name:
         return None
 
     name_lower = name.lower()
+    current_keys = _provider_keys(current_provider)
+
+    alias_match = _resolve_static_model_alias(name_lower, current_keys)
+    if alias_match:
+        return alias_match
 
     # --- Step 0: bare provider name typed as model ---
     # If someone types `/model nous` or `/model anthropic`, treat it as a
@@ -1412,64 +1509,49 @@ def detect_provider_for_model(
         if (
             resolved_provider in _PROVIDER_LABELS
             and default_models
-            and resolved_provider != normalize_provider(current_provider)
+            and resolved_provider not in current_keys
         ):
             return (resolved_provider, default_models[0])
 
     # Aggregators list other providers' models — never auto-switch TO them
-    _AGGREGATORS = {"nous", "openrouter", "ai-gateway", "copilot", "kilocode"}
-
     # If the model belongs to the current provider's catalog, don't suggest switching
-    current_models = _PROVIDER_MODELS.get(current_provider, [])
-    if any(name_lower == m.lower() for m in current_models):
+    if _model_in_provider_catalog(name_lower, current_keys):
         return None
 
     # --- Step 1: check static provider catalogs for a direct match ---
-    direct_match: Optional[str] = None
     for pid, models in _PROVIDER_MODELS.items():
-        if pid == current_provider or pid in _AGGREGATORS:
+        if pid in current_keys or pid in _AGGREGATOR_PROVIDERS:
             continue
         if any(name_lower == m.lower() for m in models):
-            direct_match = pid
-            break
+            return (pid, name)
 
-    if direct_match:
-        # Check if we have credentials for this provider — env vars,
-        # credential pool, or auth store entries.
-        has_creds = False
-        try:
-            from hermes_cli.auth import PROVIDER_REGISTRY
-            pconfig = PROVIDER_REGISTRY.get(direct_match)
-            if pconfig:
-                for env_var in pconfig.api_key_env_vars:
-                    if os.getenv(env_var, "").strip():
-                        has_creds = True
-                        break
-        except Exception:
-            pass
-        # Also check credential pool and auth store — covers OAuth,
-        # Claude Code tokens, and other non-env-var credentials (#10300).
-        if not has_creds:
-            try:
-                from agent.credential_pool import load_pool
-                pool = load_pool(direct_match)
-                if pool.has_credentials():
-                    has_creds = True
-            except Exception:
-                pass
-        if not has_creds:
-            try:
-                from hermes_cli.auth import _load_auth_store
-                store = _load_auth_store()
-                if direct_match in store.get("providers", {}) or direct_match in store.get("credential_pool", {}):
-                    has_creds = True
-            except Exception:
-                pass
+    return None
 
-        # Always return the direct provider match.  If credentials are
-        # missing, the client init will give a clear error rather than
-        # silently routing through the wrong provider (#10300).
-        return (direct_match, name)
+
+def detect_provider_for_model(
+    model_name: str,
+    current_provider: str,
+) -> Optional[tuple[str, str]]:
+    """Auto-detect the best provider for a model name.
+
+    Returns ``(provider_id, model_name)`` — the model name may be remapped
+    (e.g. bare ``deepseek-chat`` → ``deepseek/deepseek-chat`` for OpenRouter).
+    Returns ``None`` when no confident match is found.
+
+    Priority:
+    0. Bare provider name → switch to that provider's default model
+    1. Direct provider static catalog match
+    2. OpenRouter catalog match
+    """
+    name = (model_name or "").strip()
+    if not name:
+        return None
+
+    static_match = detect_static_provider_for_model(name, current_provider)
+    if static_match:
+        return static_match
+    if _model_in_provider_catalog(name.lower(), _provider_keys(current_provider)):
+        return None
 
     # --- Step 2: check OpenRouter catalog ---
     # First try exact match (handles provider/model format)
@@ -2571,8 +2653,8 @@ def validate_requested_model(
                 )
 
             return {
-                "accepted": False,
-                "persist": False,
+                "accepted": True,
+                "persist": True,
                 "recognized": False,
                 "message": message,
             }

hermes_cli/oneshot.py

@@ -0,0 +1,202 @@
+"""Oneshot (-z) mode: send a prompt, get the final content block, exit.
+
+Bypasses cli.py entirely.  No banner, no spinner, no session_id line,
+no stderr chatter.  Just the agent's final text to stdout.
+
+Toolsets = whatever the user has configured for "cli" in `hermes tools`.
+Rules / memory / AGENTS.md / preloaded skills = same as a normal chat turn.
+Approvals = auto-bypassed (HERMES_YOLO_MODE=1 is set for the call).
+Working directory = the user's CWD (AGENTS.md etc. resolve from there as usual).
+
+Model / provider selection mirrors `hermes chat`:
+    - Both optional. If omitted, use the user's configured default.
+    - If both given, pair them exactly as given.
+    - If only --model given, auto-detect the provider that serves it.
+    - If only --provider given, error out (ambiguous — caller must pick a model).
+
+Env var fallbacks (used when the corresponding arg is not passed):
+    - HERMES_INFERENCE_MODEL
+    - HERMES_INFERENCE_PROVIDER  (already read by resolve_runtime_provider)
+"""
+
+from __future__ import annotations
+
+import logging
+import os
+import sys
+from contextlib import redirect_stderr, redirect_stdout
+from typing import Optional
+
+
+def run_oneshot(
+    prompt: str,
+    model: Optional[str] = None,
+    provider: Optional[str] = None,
+) -> int:
+    """Execute a single prompt and print only the final content block.
+
+    Args:
+        prompt: The user message to send.
+        model: Optional model override. Falls back to HERMES_INFERENCE_MODEL
+            env var, then config.yaml's model.default / model.model.
+        provider: Optional provider override. Falls back to
+            HERMES_INFERENCE_PROVIDER env var, then config.yaml's model.provider,
+            then "auto".
+
+    Returns the exit code.  Caller should sys.exit() with the return.
+    """
+    # Silence every stdlib logger for the duration.  AIAgent, tools, and
+    # provider adapters all log to stderr through the root logger; file
+    # handlers added by setup_logging() keep working (they're attached to
+    # the root logger's handler list, not affected by level), but no
+    # bytes reach the terminal.
+    logging.disable(logging.CRITICAL)
+
+    # --provider without --model is ambiguous: carrying the user's configured
+    # model across to a different provider is usually wrong (that provider may
+    # not host it), and silently picking the provider's catalog default hides
+    # the mismatch.  Require the caller to be explicit.  Validate BEFORE the
+    # stderr redirect so the message actually reaches the terminal.
+    env_model_early = os.getenv("HERMES_INFERENCE_MODEL", "").strip()
+    if provider and not ((model or "").strip() or env_model_early):
+        sys.stderr.write(
+            "hermes -z: --provider requires --model (or HERMES_INFERENCE_MODEL). "
+            "Pass both explicitly, or neither to use your configured defaults.\n"
+        )
+        return 2
+
+    # Auto-approve any shell / tool approvals.  Non-interactive by
+    # definition — a prompt would hang forever.
+    os.environ["HERMES_YOLO_MODE"] = "1"
+    os.environ["HERMES_ACCEPT_HOOKS"] = "1"
+
+    # Redirect stderr AND stdout to devnull for the entire call tree.
+    # We'll print the final response to the real stdout at the end.
+    real_stdout = sys.stdout
+    devnull = open(os.devnull, "w")
+
+    try:
+        with redirect_stdout(devnull), redirect_stderr(devnull):
+            response = _run_agent(prompt, model=model, provider=provider)
+    finally:
+        try:
+            devnull.close()
+        except Exception:
+            pass
+
+    if response:
+        real_stdout.write(response)
+        if not response.endswith("\n"):
+            real_stdout.write("\n")
+        real_stdout.flush()
+    return 0
+
+
+def _run_agent(
+    prompt: str,
+    model: Optional[str] = None,
+    provider: Optional[str] = None,
+) -> str:
+    """Build an AIAgent exactly like a normal CLI chat turn would, then
+    run a single conversation.  Returns the final response string."""
+    # Imports are local so they don't run when hermes is invoked for
+    # other commands (keeps top-level CLI startup cheap).
+    from hermes_cli.config import load_config
+    from hermes_cli.models import detect_provider_for_model
+    from hermes_cli.runtime_provider import resolve_runtime_provider
+    from hermes_cli.tools_config import _get_platform_tools
+    from run_agent import AIAgent
+
+    cfg = load_config()
+
+    # Resolve effective model: explicit arg → env var → config.
+    model_cfg = cfg.get("model") or {}
+    if isinstance(model_cfg, str):
+        cfg_model = model_cfg
+    else:
+        cfg_model = model_cfg.get("default") or model_cfg.get("model") or ""
+
+    env_model = os.getenv("HERMES_INFERENCE_MODEL", "").strip()
+    effective_model = (model or "").strip() or env_model or cfg_model
+
+    # Resolve effective provider: explicit arg → (auto-detect from model if
+    # model was explicit) → env / config (handled inside resolve_runtime_provider).
+    #
+    # When --model is given without --provider, auto-detect the provider that
+    # serves that model — same semantic as `/model <name>` in an interactive
+    # session.  Without this, resolve_runtime_provider() would fall back to
+    # the user's configured default provider, which may not host the model
+    # the caller just asked for.
+    effective_provider = (provider or "").strip() or None
+    if effective_provider is None and (model or env_model):
+        # Only auto-detect when the model was explicitly requested via arg or
+        # env var (not when it came from config — that's the "use my defaults"
+        # path and the configured provider is already correct).
+        explicit_model = (model or "").strip() or env_model
+        if explicit_model:
+            cfg_provider = ""
+            if isinstance(model_cfg, dict):
+                cfg_provider = str(model_cfg.get("provider") or "").strip().lower()
+            current_provider = (
+                cfg_provider
+                or os.getenv("HERMES_INFERENCE_PROVIDER", "").strip().lower()
+                or "auto"
+            )
+            detected = detect_provider_for_model(explicit_model, current_provider)
+            if detected:
+                effective_provider, effective_model = detected
+
+    runtime = resolve_runtime_provider(
+        requested=effective_provider,
+        target_model=effective_model or None,
+    )
+
+    # Pull in whatever toolsets the user has enabled for "cli".
+    # sorted() gives stable ordering; set→list for AIAgent's signature.
+    toolsets_list = sorted(_get_platform_tools(cfg, "cli"))
+
+    agent = AIAgent(
+        api_key=runtime.get("api_key"),
+        base_url=runtime.get("base_url"),
+        provider=runtime.get("provider"),
+        api_mode=runtime.get("api_mode"),
+        model=effective_model,
+        enabled_toolsets=toolsets_list,
+        quiet_mode=True,
+        platform="cli",
+        credential_pool=runtime.get("credential_pool"),
+        # Interactive callbacks are intentionally NOT wired beyond this
+        # one.  In oneshot mode there's no user sitting at a terminal:
+        #   - clarify  → returns a synthetic "pick a default" instruction
+        #                so the agent continues instead of stalling on
+        #                the tool's built-in "not available" error
+        #   - sudo password prompt → terminal_tool gates on
+        #                HERMES_INTERACTIVE which we never set
+        #   - shell-hook approval → auto-approved via HERMES_ACCEPT_HOOKS=1
+        #                (set above); also falls back to deny on non-tty
+        #   - dangerous-command approval → bypassed via HERMES_YOLO_MODE=1
+        #   - skill secret capture → returns gracefully when no callback set
+        clarify_callback=_oneshot_clarify_callback,
+    )
+
+    # Belt-and-braces: make sure AIAgent doesn't invoke any streaming
+    # display callbacks that would bypass our stdout capture.
+    agent.suppress_status_output = True
+    agent.stream_delta_callback = None
+    agent.tool_gen_callback = None
+
+    return agent.chat(prompt) or ""
+
+
+def _oneshot_clarify_callback(question: str, choices=None) -> str:
+    """Clarify is disabled in oneshot mode — tell the agent to pick a
+    default and proceed instead of stalling or erroring."""
+    if choices:
+        return (
+            f"[oneshot mode: no user available. Pick the best option from "
+            f"{choices} using your own judgment and continue.]"
+        )
+    return (
+        "[oneshot mode: no user available. Make the most reasonable "
+        "assumption you can and continue.]"
+    )

hermes_cli/providers.py

@@ -167,6 +167,12 @@ HERMES_OVERLAYS: Dict[str, HermesOverlay] = {
         transport="openai_chat",
         base_url_env_var="OLLAMA_BASE_URL",
     ),
+    # Azure Foundry: supports both OpenAI-style and Anthropic-style endpoints.
+    # The transport is determined at runtime from config.yaml model.api_mode.
+    "azure-foundry": HermesOverlay(
+        transport="openai_chat",  # default; overridden by api_mode in config
+        base_url_env_var="AZURE_FOUNDRY_BASE_URL",
+    ),
 }
 
 

hermes_cli/runtime_provider.py

@@ -221,6 +221,19 @@ def _resolve_runtime_from_pool_entry(
     elif provider == "copilot":
         api_mode = _copilot_runtime_api_mode(model_cfg, getattr(entry, "runtime_api_key", ""))
         base_url = base_url or PROVIDER_REGISTRY["copilot"].inference_base_url
+    elif provider == "azure-foundry":
+        # Azure Foundry: read api_mode and base_url from config
+        cfg_provider = str(model_cfg.get("provider") or "").strip().lower()
+        if cfg_provider == "azure-foundry":
+            cfg_base_url = str(model_cfg.get("base_url") or "").strip().rstrip("/")
+            if cfg_base_url:
+                base_url = cfg_base_url
+            configured_mode = _parse_api_mode(model_cfg.get("api_mode"))
+            if configured_mode:
+                api_mode = configured_mode
+        # For Anthropic-style endpoints, strip /v1 suffix
+        if api_mode == "anthropic_messages":
+            base_url = re.sub(r"/v1/?$", "", base_url)
     else:
         configured_provider = str(model_cfg.get("provider") or "").strip().lower()
         # Honour model.base_url from config.yaml when the configured provider
@@ -589,6 +602,71 @@ def _resolve_openrouter_runtime(
     }
 
 
+def _resolve_azure_foundry_runtime(
+    *,
+    requested_provider: str,
+    model_cfg: Dict[str, Any],
+    explicit_api_key: Optional[str] = None,
+    explicit_base_url: Optional[str] = None,
+) -> Dict[str, Any]:
+    """Resolve an Azure Foundry runtime entry.
+
+    Reads ``model.base_url`` + ``model.api_mode`` from config.yaml (or
+    explicit overrides), pulls the API key from ``.env`` / env var, and
+    strips a trailing ``/v1`` for Anthropic-style endpoints because the
+    Anthropic SDK appends ``/v1/messages`` internally.
+
+    Raises :class:`AuthError` when required values are missing.
+    """
+    explicit_api_key = str(explicit_api_key or "").strip()
+    explicit_base_url_clean = str(explicit_base_url or "").strip().rstrip("/")
+
+    cfg_provider = str(model_cfg.get("provider") or "").strip().lower()
+    cfg_base_url = ""
+    cfg_api_mode = "chat_completions"
+    if cfg_provider == "azure-foundry":
+        cfg_base_url = str(model_cfg.get("base_url") or "").strip().rstrip("/")
+        cfg_api_mode = _parse_api_mode(model_cfg.get("api_mode")) or "chat_completions"
+
+    env_base_url = os.getenv("AZURE_FOUNDRY_BASE_URL", "").strip().rstrip("/")
+    base_url = explicit_base_url_clean or cfg_base_url or env_base_url
+    if not base_url:
+        raise AuthError(
+            "Azure Foundry requires a base URL. Set it via 'hermes model' or "
+            "the AZURE_FOUNDRY_BASE_URL environment variable."
+        )
+
+    api_key = explicit_api_key
+    if not api_key:
+        try:
+            from hermes_cli.config import get_env_value
+            api_key = get_env_value("AZURE_FOUNDRY_API_KEY") or ""
+        except Exception:
+            api_key = ""
+    if not api_key:
+        api_key = os.getenv("AZURE_FOUNDRY_API_KEY", "").strip()
+    if not api_key:
+        raise AuthError(
+            "Azure Foundry requires an API key. Set AZURE_FOUNDRY_API_KEY in "
+            "~/.hermes/.env or run 'hermes model' to configure."
+        )
+
+    # Anthropic SDK appends /v1/messages itself, so strip any trailing /v1
+    # we inherited from the configured base_url to avoid double-/v1 paths.
+    if cfg_api_mode == "anthropic_messages":
+        base_url = re.sub(r"/v1/?$", "", base_url)
+
+    source = "explicit" if (explicit_api_key or explicit_base_url) else "config"
+    return {
+        "provider": "azure-foundry",
+        "api_mode": cfg_api_mode,
+        "base_url": base_url,
+        "api_key": api_key,
+        "source": source,
+        "requested_provider": requested_provider,
+    }
+
+
 def _resolve_explicit_runtime(
     *,
     provider: str,
@@ -678,6 +756,15 @@ def _resolve_explicit_runtime(
             "requested_provider": requested_provider,
         }
 
+    # Azure Foundry: user-configured endpoint with selectable API mode
+    if provider == "azure-foundry":
+        return _resolve_azure_foundry_runtime(
+            requested_provider=requested_provider,
+            model_cfg=model_cfg,
+            explicit_api_key=explicit_api_key,
+            explicit_base_url=explicit_base_url,
+        )
+
     pconfig = PROVIDER_REGISTRY.get(provider)
     if pconfig and pconfig.auth_type == "api_key":
         env_url = ""
@@ -746,6 +833,40 @@ def resolve_runtime_provider(
     """
     requested_provider = resolve_requested_provider(requested)
 
+    # Azure Anthropic short-circuit: when explicitly targeting an Azure endpoint
+    # with provider="anthropic", bypass _resolve_named_custom_runtime (which would
+    # return provider="custom" with chat_completions api_mode and no valid key).
+    # Instead, use the Azure key directly with anthropic_messages api_mode.
+    _eff_base = (explicit_base_url or "").strip()
+    if requested_provider == "anthropic" and "azure.com" in _eff_base:
+        _azure_key = (
+            (explicit_api_key or "").strip()
+            or os.getenv("AZURE_ANTHROPIC_KEY", "").strip()
+            or os.getenv("ANTHROPIC_API_KEY", "").strip()
+        )
+        return {
+            "provider": "anthropic",
+            "api_mode": "anthropic_messages",
+            "base_url": _eff_base.rstrip("/"),
+            "api_key": _azure_key,
+            "source": "azure-explicit",
+            "requested_provider": requested_provider,
+        }
+
+    # Azure Foundry: user-configured endpoint with selectable API mode
+    # (OpenAI-style chat_completions or Anthropic-style anthropic_messages).
+    # Resolve before the custom-runtime / pool / generic paths so Azure
+    # config is always picked up from model.base_url + model.api_mode,
+    # regardless of whether the caller passed explicit_* args.
+    if requested_provider == "azure-foundry":
+        azure_runtime = _resolve_azure_foundry_runtime(
+            requested_provider=requested_provider,
+            model_cfg=_get_model_config(),
+            explicit_api_key=explicit_api_key,
+            explicit_base_url=explicit_base_url,
+        )
+        return azure_runtime
+
     custom_runtime = _resolve_named_custom_runtime(
         requested_provider=requested_provider,
         explicit_api_key=explicit_api_key,
@@ -924,13 +1045,6 @@ def resolve_runtime_provider(
 
     # Anthropic (native Messages API)
     if provider == "anthropic":
-        from agent.anthropic_adapter import resolve_anthropic_token
-        token = resolve_anthropic_token()
-        if not token:
-            raise AuthError(
-                "No Anthropic credentials found. Set ANTHROPIC_TOKEN or ANTHROPIC_API_KEY, "
-                "run 'claude setup-token', or authenticate with 'claude /login'."
-            )
         # Allow base URL override from config.yaml model.base_url, but only
         # when the configured provider is anthropic — otherwise a non-Anthropic
         # base_url (e.g. Codex endpoint) would leak into Anthropic requests.
@@ -939,6 +1053,33 @@ def resolve_runtime_provider(
         if cfg_provider == "anthropic":
             cfg_base_url = (model_cfg.get("base_url") or "").strip().rstrip("/")
         base_url = cfg_base_url or "https://api.anthropic.com"
+
+        # For Azure AI Foundry endpoints, use ANTHROPIC_API_KEY directly —
+        # Claude Code OAuth tokens (sk-ant-oat01) are not accepted by Azure.
+        # Azure keys don't start with "sk-ant-" so resolve_anthropic_token()
+        # would find the Claude Code OAuth token first (priority 3) and return
+        # that instead, causing 401s. Detect Azure endpoints and use the env
+        # key directly to bypass the OAuth priority chain.
+        _is_azure_endpoint = "azure.com" in base_url.lower() or (
+            cfg_base_url and "azure.com" in cfg_base_url.lower()
+        )
+        if _is_azure_endpoint:
+            token = (
+                os.getenv("AZURE_ANTHROPIC_KEY", "").strip()
+                or os.getenv("ANTHROPIC_API_KEY", "").strip()
+            )
+            if not token:
+                raise AuthError(
+                    "No Azure Anthropic API key found. Set AZURE_ANTHROPIC_KEY or ANTHROPIC_API_KEY."
+                )
+        else:
+            from agent.anthropic_adapter import resolve_anthropic_token
+            token = resolve_anthropic_token()
+            if not token:
+                raise AuthError(
+                    "No Anthropic credentials found. Set ANTHROPIC_TOKEN or ANTHROPIC_API_KEY, "
+                    "run 'claude setup-token', or authenticate with 'claude /login'."
+                )
         return {
             "provider": "anthropic",
             "api_mode": "anthropic_messages",

hermes_cli/setup.py

@@ -2863,17 +2863,6 @@ SETUP_SECTIONS = [
     ("agent", "Agent Settings", setup_agent_settings),
 ]
 
-# The returning-user menu intentionally omits standalone TTS because model setup
-# already includes TTS selection and tools setup covers the rest of the provider
-# configuration. Keep this list in the same order as the visible menu entries.
-RETURNING_USER_MENU_SECTION_KEYS = [
-    "model",
-    "terminal",
-    "gateway",
-    "tools",
-    "agent",
-]
-
 
 def run_setup_wizard(args):
     """Run the interactive setup wizard.
@@ -2898,6 +2887,9 @@ def run_setup_wizard(args):
         save_config(copy.deepcopy(DEFAULT_CONFIG))
         print_success("Configuration reset to defaults.")
 
+    reconfigure_requested = bool(getattr(args, "reconfigure", False))
+    quick_requested = bool(getattr(args, "quick", False))
+
     config = load_config()
     hermes_home = get_hermes_home()
 
@@ -2989,50 +2981,36 @@ def run_setup_wizard(args):
     migration_ran = False
 
     if is_existing:
-        # ── Returning User Menu ──
-        print()
-        print_header("Welcome Back!")
-        print_success("You already have Hermes configured.")
-        print()
-
-        menu_choices = [
-            "Quick Setup - configure missing items only",
-            "Full Setup - reconfigure everything",
-            "Model & Provider",
-            "Terminal Backend",
-            "Messaging Platforms (Gateway)",
-            "Tools",
-            "Agent Settings",
-            "Exit",
-        ]
-        choice = prompt_choice("What would you like to do?", menu_choices, 0)
-
-        if choice == 0:
-            # Quick setup
+        # Existing install — default is the full-wizard reconfigure flow.
+        # Every prompt shows the current value as its default, so pressing
+        # Enter keeps it.  Opt into `--quick` for the narrow "just fill in
+        # missing items" flow (useful after a partial OpenClaw migration
+        # or when a required API key got cleared).
+        if quick_requested:
             _run_quick_setup(config, hermes_home)
             return
-        elif choice == 1:
-            # Full setup — fall through to run all sections
-            pass
-        elif choice == 7:
-            print_info("Exiting. Run 'hermes setup' again when ready.")
-            return
-        elif 2 <= choice <= 6:
-            # Individual section — map by key, not by position.
-            # SETUP_SECTIONS includes TTS but the returning-user menu skips it,
-            # so positional indexing (choice - 2) would dispatch the wrong section.
-            section_key = RETURNING_USER_MENU_SECTION_KEYS[choice - 2]
-            section = next((s for s in SETUP_SECTIONS if s[0] == section_key), None)
-            if section:
-                _, label, func = section
-                func(config)
-                save_config(config)
-                _print_setup_summary(config, hermes_home)
-            return
+
+        print()
+        print_header("Reconfigure")
+        print_success("You already have Hermes configured.")
+        print_info("Running the full wizard — each prompt shows your current value.")
+        print_info("Press Enter to keep it, or type a new value to change it.")
+        print_info("")
+        print_info("Tip: jump straight to a section with 'hermes setup model|terminal|")
+        print_info("     gateway|tools|agent', or fill only missing items with --quick.")
+        # Fall through to the "Full Setup — run all sections" block below.
+        # --reconfigure is now the default on existing installs; the flag
+        # is preserved for backwards compatibility but is a no-op here.
     else:
         # ── First-Time Setup ──
         print()
 
+        # --reconfigure / --quick on a fresh install are meaningless — fall
+        # through to the normal first-time flow.
+        if reconfigure_requested or quick_requested:
+            print_info("No existing configuration found — running first-time setup.")
+            print()
+
         # Offer OpenClaw migration before configuration begins
         migration_ran = _offer_openclaw_migration(hermes_home)
         if migration_ran:

hermes_cli/tips.py

@@ -10,8 +10,7 @@ import random
 
 TIPS = [
     # --- Slash Commands ---
-    "/btw <question> asks a quick side question without tools or history — great for clarifications.",
-    "/background <prompt> runs a task in a separate session while your current one stays free.",
+    "/background <prompt> (alias /bg or /btw) runs a task in a separate session while your current one stays free.",
     "/branch forks the current session so you can explore a different direction without losing progress.",
     "/compress manually compresses conversation context when things get long.",
     "/rollback lists filesystem checkpoints — restore files the agent modified to any prior state.",

hermes_cli/tools_config.py

@@ -68,25 +68,58 @@ CONFIGURABLE_TOOLSETS = [
     ("rl",              "🧪 RL Training",               "Tinker-Atropos training tools"),
     ("homeassistant",    "🏠 Home Assistant",           "smart home device control"),
     ("spotify",          "🎵 Spotify",                  "playback, search, playlists, library"),
+    ("discord",         "💬 Discord (read/participate)", "fetch messages, search members, create thread"),
+    ("discord_admin",   "🛡️  Discord Server Admin",    "list channels/roles, pin, assign roles"),
 ]
 
 # Toolsets that are OFF by default for new installs.
 # They're still in _HERMES_CORE_TOOLS (available at runtime if enabled),
 # but the setup checklist won't pre-select them for first-time users.
-_DEFAULT_OFF_TOOLSETS = {"moa", "homeassistant", "rl", "spotify"}
+_DEFAULT_OFF_TOOLSETS = {"moa", "homeassistant", "rl", "spotify", "discord", "discord_admin"}
+
+# Platform-scoped toolsets: only appear in the `hermes tools` checklist for
+# these platforms, and only resolve/save for these platforms.  A toolset
+# absent from this map is available on every platform (current behaviour).
+#
+# Use this for tools whose APIs only make sense on one platform (Discord
+# server admin, Slack workspace admin, etc.).  Keeps every other platform's
+# checklist from filling up with irrelevant toggles.
+_TOOLSET_PLATFORM_RESTRICTIONS: Dict[str, Set[str]] = {
+    "discord": {"discord"},
+    "discord_admin": {"discord"},
+}
+
+
+def _toolset_allowed_for_platform(ts_key: str, platform: str) -> bool:
+    """Return True if ``ts_key`` is configurable on ``platform``.
+
+    Toolsets without a restriction entry are allowed everywhere (the default).
+    """
+    allowed = _TOOLSET_PLATFORM_RESTRICTIONS.get(ts_key)
+    return allowed is None or platform in allowed
 
 
 def _get_effective_configurable_toolsets():
     """Return CONFIGURABLE_TOOLSETS + any plugin-provided toolsets.
 
     Plugin toolsets are appended at the end so they appear after the
-    built-in toolsets in the TUI checklist.
+    built-in toolsets in the TUI checklist. A plugin whose toolset key
+    already appears in ``CONFIGURABLE_TOOLSETS`` is skipped — bundled
+    plugins (e.g. ``plugins/spotify``) share their toolset key with the
+    built-in entry, and we want the built-in label/description to win.
+    Without the dedupe, ``hermes tools`` → "reconfigure existing" would
+    list the same toolset twice.
     """
     result = list(CONFIGURABLE_TOOLSETS)
+    seen = {ts_key for ts_key, _, _ in result}
     try:
         from hermes_cli.plugins import discover_plugins, get_plugin_toolsets
         discover_plugins()  # idempotent — ensures plugins are loaded
-        result.extend(get_plugin_toolsets())
+        for entry in get_plugin_toolsets():
+            if entry[0] in seen:
+                continue
+            seen.add(entry[0])
+            result.append(entry)
     except Exception:
         pass
     return result
@@ -591,7 +624,7 @@ def _get_platform_tools(
     include_default_mcp_servers: bool = True,
 ) -> Set[str]:
     """Resolve which individual toolset names are enabled for a platform."""
-    from toolsets import resolve_toolset
+    from toolsets import resolve_toolset, TOOLSETS
 
     platform_toolsets = config.get("platform_toolsets") or {}
     toolset_names = platform_toolsets.get(platform)
@@ -605,6 +638,8 @@ def _get_platform_tools(
     toolset_names = [str(ts) for ts in toolset_names]
 
     configurable_keys = {ts_key for ts_key, _, _ in CONFIGURABLE_TOOLSETS}
+    plugin_ts_keys = _get_plugin_toolset_keys()
+    platform_default_keys = {p["default_toolset"] for p in PLATFORMS.values()}
 
     # If the saved list contains any configurable keys directly, the user
     # has explicitly configured this platform — use direct membership.
@@ -614,7 +649,10 @@ def _get_platform_tools(
     has_explicit_config = any(ts in configurable_keys for ts in toolset_names)
 
     if has_explicit_config:
-        enabled_toolsets = {ts for ts in toolset_names if ts in configurable_keys}
+        enabled_toolsets = {
+            ts for ts in toolset_names
+            if ts in configurable_keys and _toolset_allowed_for_platform(ts, platform)
+        }
     else:
         # No explicit config — fall back to resolving composite toolset names
         # (e.g. "hermes-cli") to individual tool names and reverse-mapping.
@@ -624,14 +662,52 @@ def _get_platform_tools(
 
         enabled_toolsets = set()
         for ts_key, _, _ in CONFIGURABLE_TOOLSETS:
+            if not _toolset_allowed_for_platform(ts_key, platform):
+                continue
             ts_tools = set(resolve_toolset(ts_key))
             if ts_tools and ts_tools.issubset(all_tool_names):
                 enabled_toolsets.add(ts_key)
+
         default_off = set(_DEFAULT_OFF_TOOLSETS)
-        if platform in default_off:
+        # Legacy safety: if the platform's own name matches a default-off
+        # toolset (e.g. `homeassistant` platform + `homeassistant` toolset),
+        # keep that toolset enabled on first install.  Skip this dodge for
+        # platform-restricted toolsets — those are always opt-in even on
+        # their own platform (e.g. `discord` + `discord` should stay OFF).
+        if platform in default_off and platform not in _TOOLSET_PLATFORM_RESTRICTIONS:
             default_off.remove(platform)
         enabled_toolsets -= default_off
 
+    # Recover non-configurable platform toolsets (e.g. discord, feishu_doc,
+    # feishu_drive).  These are part of the platform's default composite but
+    # absent from CONFIGURABLE_TOOLSETS, so they can't appear in the TUI
+    # checklist or in a user-saved config.  Must run in BOTH branches —
+    # otherwise saving via `hermes tools` (which flips has_explicit_config
+    # to True) silently drops them.
+    platform_tool_universe = set(resolve_toolset(PLATFORMS[platform]["default_toolset"]))
+    configurable_tool_universe = set()
+    for ck in configurable_keys:
+        configurable_tool_universe.update(resolve_toolset(ck))
+    claimed = set()
+    for ts_key in enabled_toolsets:
+        claimed.update(resolve_toolset(ts_key))
+    skip = configurable_keys | plugin_ts_keys | platform_default_keys
+    skip |= {k for k in TOOLSETS if k.startswith("hermes-")}
+    skip |= set(_DEFAULT_OFF_TOOLSETS) - {platform}
+    for ts_key, ts_def in TOOLSETS.items():
+        if ts_key in skip:
+            continue
+        if ts_def.get("includes"):
+            continue
+        ts_tools = set(resolve_toolset(ts_key))
+        if not ts_tools or not ts_tools.issubset(platform_tool_universe):
+            continue
+        if ts_tools.issubset(configurable_tool_universe):
+            continue
+        if not ts_tools.issubset(claimed):
+            enabled_toolsets.add(ts_key)
+            claimed.update(ts_tools)
+
     # Plugin toolsets: enabled by default unless explicitly disabled, or
     # unless the toolset is in _DEFAULT_OFF_TOOLSETS (e.g. spotify —
     # shipped as a bundled plugin but user must opt in via `hermes tools`
@@ -639,7 +715,6 @@ def _get_platform_tools(
     # A plugin toolset is "known" for a platform once `hermes tools`
     # has been saved for that platform (tracked via known_plugin_toolsets).
     # Unknown plugins default to enabled; known-but-absent = disabled.
-    plugin_ts_keys = _get_plugin_toolset_keys()
     if plugin_ts_keys:
         known_map = config.get("known_plugin_toolsets", {})
         known_for_platform = set(known_map.get(platform, []))
@@ -657,7 +732,6 @@ def _get_platform_tools(
 
     # Preserve any explicit non-configurable toolset entries (for example,
     # custom toolsets or MCP server names saved in platform_toolsets).
-    platform_default_keys = {p["default_toolset"] for p in PLATFORMS.values()}
     explicit_passthrough = {
         ts
         for ts in toolset_names
@@ -703,6 +777,14 @@ def _save_platform_tools(config: dict, platform: str, enabled_toolset_keys: Set[
     """
     config.setdefault("platform_toolsets", {})
 
+    # Drop platform-scoped toolsets that don't apply here.  Prevents the
+    # "Configure all platforms" checklist (or a hand-edited config.yaml)
+    # from turning on, say, the `discord` toolset for Telegram.
+    enabled_toolset_keys = {
+        ts for ts in enabled_toolset_keys
+        if _toolset_allowed_for_platform(ts, platform)
+    }
+
     # Get the set of all configurable toolset keys (built-in + plugin)
     configurable_keys = {ts_key for ts_key, _, _ in CONFIGURABLE_TOOLSETS}
     plugin_keys = _get_plugin_toolset_keys()
@@ -717,6 +799,7 @@ def _save_platform_tools(config: dict, platform: str, enabled_toolset_keys: Set[
     existing_toolsets = config.get("platform_toolsets", {}).get(platform, [])
     if not isinstance(existing_toolsets, list):
         existing_toolsets = []
+    existing_toolsets = [str(ts) for ts in existing_toolsets]
 
     # Preserve any entries that are NOT configurable toolsets and NOT platform
     # defaults (i.e. only MCP server names should be preserved)
@@ -724,6 +807,11 @@ def _save_platform_tools(config: dict, platform: str, enabled_toolset_keys: Set[
         entry for entry in existing_toolsets
         if entry not in configurable_keys and entry not in platform_default_keys
     }
+    # Opening `hermes tools` is the user's opt-in to reconfigure tools, so treat
+    # saving from the picker as consent to clear the "no_mcp" sentinel. The
+    # picker has no checkbox for no_mcp, so without this users who once set it
+    # by hand could never re-enable MCP servers through the UI.
+    preserved_entries.discard("no_mcp")
 
     # Merge preserved entries with new enabled toolsets
     config["platform_toolsets"][platform] = sorted(enabled_toolset_keys | preserved_entries)
@@ -831,7 +919,7 @@ def _estimate_tool_tokens() -> Dict[str, int]:
     return _tool_token_cache
 
 
-def _prompt_toolset_checklist(platform_label: str, enabled: Set[str]) -> Set[str]:
+def _prompt_toolset_checklist(platform_label: str, enabled: Set[str], platform: str = "cli") -> Set[str]:
     """Multi-select checklist of toolsets. Returns set of selected toolset keys."""
     from hermes_cli.curses_ui import curses_checklist
     from toolsets import resolve_toolset
@@ -839,7 +927,12 @@ def _prompt_toolset_checklist(platform_label: str, enabled: Set[str]) -> Set[str
     # Pre-compute per-tool token counts (cached after first call).
     tool_tokens = _estimate_tool_tokens()
 
-    effective = _get_effective_configurable_toolsets()
+    effective_all = _get_effective_configurable_toolsets()
+    # Drop platform-scoped toolsets that don't apply to this platform.
+    effective = [
+        (k, l, d) for (k, l, d) in effective_all
+        if _toolset_allowed_for_platform(k, platform)
+    ]
 
     labels = []
     for ts_key, ts_label, ts_desc in effective:
@@ -1753,7 +1846,7 @@ def tools_command(args=None, first_install: bool = False, config: dict = None):
             checklist_preselected = current_enabled - _DEFAULT_OFF_TOOLSETS
 
             # Show checklist
-            new_enabled = _prompt_toolset_checklist(pinfo["label"], checklist_preselected)
+            new_enabled = _prompt_toolset_checklist(pinfo["label"], checklist_preselected, pkey)
 
             added = new_enabled - current_enabled
             removed = current_enabled - new_enabled
@@ -2109,7 +2202,11 @@ def _apply_mcp_change(config: dict, targets: List[str], action: str) -> Set[str]
 
 def _print_tools_list(enabled_toolsets: set, mcp_servers: dict, platform: str = "cli"):
     """Print a summary of enabled/disabled toolsets and MCP tool filters."""
-    effective = _get_effective_configurable_toolsets()
+    effective_all = _get_effective_configurable_toolsets()
+    effective = [
+        (k, l, d) for (k, l, d) in effective_all
+        if _toolset_allowed_for_platform(k, platform)
+    ]
     builtin_keys = {ts_key for ts_key, _, _ in CONFIGURABLE_TOOLSETS}
 
     print(f"Built-in toolsets ({platform}):")
@@ -2175,6 +2272,20 @@ def tools_disable_enable_command(args):
             _print_error(f"Unknown toolset '{name}'")
         toolset_targets = [t for t in toolset_targets if t in valid_toolsets]
 
+    # Reject platform-scoped toolsets on platforms that don't allow them.
+    restricted_targets = [
+        t for t in toolset_targets
+        if not _toolset_allowed_for_platform(t, platform)
+    ]
+    if restricted_targets:
+        for name in restricted_targets:
+            allowed = sorted(_TOOLSET_PLATFORM_RESTRICTIONS.get(name) or set())
+            _print_error(
+                f"Toolset '{name}' is not available on platform '{platform}' "
+                f"(only: {', '.join(allowed)})"
+            )
+        toolset_targets = [t for t in toolset_targets if t not in restricted_targets]
+
     if toolset_targets:
         _apply_toolset_change(config, platform, toolset_targets, action)
 

hermes_cli/web_server.py

@@ -3103,13 +3103,23 @@ def _mount_plugin_api_routes():
             _log.warning("Plugin %s declares api=%s but file not found", plugin["name"], api_file_name)
             continue
         try:
-            spec = importlib.util.spec_from_file_location(
-                f"hermes_dashboard_plugin_{plugin['name']}", api_path,
-            )
+            module_name = f"hermes_dashboard_plugin_{plugin['name']}"
+            spec = importlib.util.spec_from_file_location(module_name, api_path)
             if spec is None or spec.loader is None:
                 continue
             mod = importlib.util.module_from_spec(spec)
-            spec.loader.exec_module(mod)
+            # Register in sys.modules BEFORE exec_module so pydantic/FastAPI
+            # can resolve forward references (e.g. models defined in a file
+            # that uses `from __future__ import annotations`). Without this,
+            # TypeAdapter lazy-build fails at first request with
+            # "is not fully defined" because the module namespace isn't
+            # reachable by name for string-annotation resolution.
+            sys.modules[module_name] = mod
+            try:
+                spec.loader.exec_module(mod)
+            except Exception:
+                sys.modules.pop(module_name, None)
+                raise
             router = getattr(mod, "router", None)
             if router is None:
                 _log.warning("Plugin %s api file has no 'router' attribute", plugin["name"])

hermes_state.py

@@ -31,7 +31,7 @@ T = TypeVar("T")
 
 DEFAULT_DB_PATH = get_hermes_home() / "state.db"
 
-SCHEMA_VERSION = 8
+SCHEMA_VERSION = 9
 
 SCHEMA_SQL = """
 CREATE TABLE IF NOT EXISTS schema_version (
@@ -83,7 +83,8 @@ CREATE TABLE IF NOT EXISTS messages (
     reasoning TEXT,
     reasoning_content TEXT,
     reasoning_details TEXT,
-    codex_reasoning_items TEXT
+    codex_reasoning_items TEXT,
+    codex_message_items TEXT
 );
 
 CREATE TABLE IF NOT EXISTS state_meta (
@@ -356,6 +357,15 @@ class SessionDB:
                 except sqlite3.OperationalError:
                     pass  # Column already exists
                 cursor.execute("UPDATE schema_version SET version = 8")
+            if current_version < 9:
+                # v9: preserve replayable Codex assistant message ids/phases so
+                # follow-up turns can rebuild Responses API message items instead
+                # of flattening everything to plain assistant text.
+                try:
+                    cursor.execute('ALTER TABLE messages ADD COLUMN "codex_message_items" TEXT')
+                except sqlite3.OperationalError:
+                    pass  # Column already exists
+                cursor.execute("UPDATE schema_version SET version = 9")
 
         # Unique title index — always ensure it exists (safe to run after migrations
         # since the title column is guaranteed to exist at this point)
@@ -956,6 +966,7 @@ class SessionDB:
         reasoning_content: str = None,
         reasoning_details: Any = None,
         codex_reasoning_items: Any = None,
+        codex_message_items: Any = None,
     ) -> int:
         """
         Append a message to a session. Returns the message row ID.
@@ -972,6 +983,10 @@ class SessionDB:
             json.dumps(codex_reasoning_items)
             if codex_reasoning_items else None
         )
+        codex_message_items_json = (
+            json.dumps(codex_message_items)
+            if codex_message_items else None
+        )
         tool_calls_json = json.dumps(tool_calls) if tool_calls else None
 
         # Pre-compute tool call count
@@ -983,8 +998,9 @@ class SessionDB:
             cursor = conn.execute(
                 """INSERT INTO messages (session_id, role, content, tool_call_id,
                    tool_calls, tool_name, timestamp, token_count, finish_reason,
-                   reasoning, reasoning_content, reasoning_details, codex_reasoning_items)
-                   VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)""",
+                   reasoning, reasoning_content, reasoning_details, codex_reasoning_items,
+                   codex_message_items)
+                   VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)""",
                 (
                     session_id,
                     role,
@@ -999,6 +1015,7 @@ class SessionDB:
                     reasoning_content,
                     reasoning_details_json,
                     codex_items_json,
+                    codex_message_items_json,
                 ),
             )
             msg_id = cursor.lastrowid
@@ -1112,7 +1129,8 @@ class SessionDB:
         with self._lock:
             cursor = self._conn.execute(
                 "SELECT role, content, tool_call_id, tool_calls, tool_name, "
-                "reasoning, reasoning_content, reasoning_details, codex_reasoning_items "
+                "reasoning, reasoning_content, reasoning_details, codex_reasoning_items, "
+                "codex_message_items "
                 "FROM messages WHERE session_id = ? ORDER BY timestamp, id",
                 (session_id,),
             )
@@ -1150,6 +1168,12 @@ class SessionDB:
                     except (json.JSONDecodeError, TypeError):
                         logger.warning("Failed to deserialize codex_reasoning_items, falling back to None")
                         msg["codex_reasoning_items"] = None
+                if row["codex_message_items"]:
+                    try:
+                        msg["codex_message_items"] = json.loads(row["codex_message_items"])
+                    except (json.JSONDecodeError, TypeError):
+                        logger.warning("Failed to deserialize codex_message_items, falling back to None")
+                        msg["codex_message_items"] = None
             messages.append(msg)
         return messages
 

model_tools.py

@@ -24,6 +24,7 @@ import json
 import asyncio
 import logging
 import threading
+import time
 from typing import Dict, Any, List, Optional, Tuple
 
 from tools.registry import discover_builtin_tools, registry
@@ -288,30 +289,34 @@ def get_tool_definitions(
                 filtered_tools[i] = {"type": "function", "function": dynamic_schema}
                 break
 
-    # Rebuild discord_server schema based on the bot's privileged intents
-    # (detected from GET /applications/@me) and the user's action allowlist
-    # in config.  Hides actions the bot's intents don't support so the
-    # model never attempts them, and annotates fetch_messages when the
+    # Rebuild discord / discord_admin schemas based on the bot's privileged
+    # intents (detected from GET /applications/@me) and the user's action
+    # allowlist in config.  Hides actions the bot's intents don't support so
+    # the model never attempts them, and annotates fetch_messages when the
     # MESSAGE_CONTENT intent is missing.
-    if "discord_server" in available_tool_names:
-        try:
-            from tools.discord_tool import get_dynamic_schema
-            dynamic = get_dynamic_schema()
-        except Exception:  # pragma: no cover — defensive, fall back to static
-            dynamic = None
-        if dynamic is None:
-            # Tool filtered out entirely (empty allowlist or detection disabled
-            # the only remaining actions).  Drop it from the schema list.
-            filtered_tools = [
-                t for t in filtered_tools
-                if t.get("function", {}).get("name") != "discord_server"
-            ]
-            available_tool_names.discard("discord_server")
-        else:
-            for i, td in enumerate(filtered_tools):
-                if td.get("function", {}).get("name") == "discord_server":
-                    filtered_tools[i] = {"type": "function", "function": dynamic}
-                    break
+    _discord_schema_fns = {
+        "discord": "get_dynamic_schema_core",
+        "discord_admin": "get_dynamic_schema_admin",
+    }
+    for discord_tool_name in _discord_schema_fns:
+        if discord_tool_name in available_tool_names:
+            try:
+                from tools import discord_tool as _dt
+                schema_fn = getattr(_dt, _discord_schema_fns[discord_tool_name])
+                dynamic = schema_fn()
+            except Exception:
+                dynamic = None
+            if dynamic is None:
+                filtered_tools = [
+                    t for t in filtered_tools
+                    if t.get("function", {}).get("name") != discord_tool_name
+                ]
+                available_tool_names.discard(discord_tool_name)
+            else:
+                for i, td in enumerate(filtered_tools):
+                    if td.get("function", {}).get("name") == discord_tool_name:
+                        filtered_tools[i] = {"type": "function", "function": dynamic}
+                        break
 
     # Strip web tool cross-references from browser_navigate description when
     # web_search / web_extract are not available.  The static schema says
@@ -563,6 +568,14 @@ def handle_function_call(
             except Exception:
                 pass  # file_tools may not be loaded yet
 
+        # Measure tool dispatch latency so post_tool_call and
+        # transform_tool_result hooks can observe per-tool duration.
+        # Inspired by Claude Code 2.1.119, which added ``duration_ms`` to
+        # PostToolUse hook inputs so plugin authors can build latency
+        # dashboards, budget alerts, and regression canaries without having
+        # to wrap every tool manually.  We use monotonic() so the value is
+        # unaffected by wall-clock adjustments during the call.
+        _dispatch_start = time.monotonic()
         if function_name == "execute_code":
             # Prefer the caller-provided list so subagents can't overwrite
             # the parent's tool set via the process-global.
@@ -578,6 +591,7 @@ def handle_function_call(
                 task_id=task_id,
                 user_task=user_task,
             )
+        duration_ms = int((time.monotonic() - _dispatch_start) * 1000)
 
         try:
             from hermes_cli.plugins import invoke_hook
@@ -589,6 +603,7 @@ def handle_function_call(
                 task_id=task_id or "",
                 session_id=session_id or "",
                 tool_call_id=tool_call_id or "",
+                duration_ms=duration_ms,
             )
         except Exception:
             pass
@@ -609,6 +624,7 @@ def handle_function_call(
                 task_id=task_id or "",
                 session_id=session_id or "",
                 tool_call_id=tool_call_id or "",
+                duration_ms=duration_ms,
             )
             for hook_result in hook_results:
                 if isinstance(hook_result, str):

nix/lib.nix

@@ -156,7 +156,7 @@
       for entry in "''${ENTRIES[@]}"; do
         IFS=":" read -r ATTR FOLDER NIX_FILE <<< "$entry"
         echo "==> .#$ATTR ($FOLDER -> $NIX_FILE)"
-        OUTPUT=$(nix build ".#$ATTR.npmDeps" --no-link --print-build-logs 2>&1)
+        OUTPUT=$(nix build ".#$ATTR.npmDeps" --no-link --rebuild --print-build-logs 2>&1)
         STATUS=$?
         if [ "$STATUS" -eq 0 ]; then
           echo "    ok"

nix/web.nix

@@ -4,7 +4,7 @@ let
   src = ../web;
   npmDeps = pkgs.fetchNpmDeps {
     inherit src;
-    hash = "sha256-TS/vrCHbdvXkPcAPxImKzAd2pdDCrKlgYZkXBMQ+TEg=";
+    hash = "sha256-4Z8KQ69QhO83X6zff+5urWBv6MME686MhTTMdwSl65o=";
   };
 
   npm = hermesNpmLib.mkNpmPassthru { folder = "web"; attr = "web"; pname = "hermes-web"; };

plugins/example-dashboard/dashboard/dist/index.js

@@ -91,4 +91,29 @@
 
   // Register this plugin — the dashboard picks it up automatically.
   window.__HERMES_PLUGINS__.register("example", ExamplePage);
+
+  // ─────────────────────────────────────────────────────────────────────
+  // Page-scoped slot demo: inject a small banner at the top of /sessions.
+  //
+  // Built-in pages expose named slots (<page>:top, <page>:bottom) that
+  // plugins can populate without overriding the whole route. The
+  // manifest lists the slots we use in its `slots` array so the shell
+  // knows to render <PluginSlot name="sessions:top" /> there.
+  // ─────────────────────────────────────────────────────────────────────
+  function SessionsTopBanner() {
+    return React.createElement(Card, {
+      className: "border-dashed",
+    },
+      React.createElement(CardContent, { className: "flex items-center gap-3 py-2" },
+        React.createElement(Badge, { variant: "outline" }, "Example"),
+        React.createElement("span", {
+          className: "text-xs text-muted-foreground",
+        }, "This banner was injected into the Sessions page by the example plugin via the ",
+          React.createElement("code", { className: "font-courier" }, "sessions:top"),
+          " slot."),
+      ),
+    );
+  }
+
+  window.__HERMES_PLUGINS__.registerSlot("example", "sessions:top", SessionsTopBanner);
 })();

plugins/example-dashboard/dashboard/manifest.json

@@ -8,6 +8,7 @@
     "path": "/example",
     "position": "after:skills"
   },
+  "slots": ["sessions:top"],
   "entry": "dist/index.js",
   "api": "plugin_api.py"
 }

plugins/kanban/dashboard/dist/style.css

@@ -0,0 +1,752 @@
+/*
+ * Hermes Kanban — dashboard plugin styles.
+ *
+ * All colors reference theme CSS vars so the board reskins with the
+ * active dashboard theme. No hardcoded palette.
+ */
+
+.hermes-kanban {
+  width: 100%;
+}
+
+/* ---- Columns layout -------------------------------------------------- */
+
+.hermes-kanban-columns {
+  display: grid;
+  grid-template-columns: repeat(auto-fit, minmax(260px, 1fr));
+  gap: 0.75rem;
+  align-items: start;
+}
+
+.hermes-kanban-column {
+  display: flex;
+  flex-direction: column;
+  background: color-mix(in srgb, var(--color-card) 85%, transparent);
+  border: 1px solid var(--color-border);
+  border-radius: var(--radius);
+  padding: 0.5rem;
+  min-height: 200px;
+  max-height: calc(100vh - 220px);
+  transition: border-color 120ms ease, background-color 120ms ease;
+}
+
+.hermes-kanban-column--drop {
+  border-color: var(--color-ring);
+  background: color-mix(in srgb, var(--color-ring) 8%, var(--color-card));
+}
+
+.hermes-kanban-column-header {
+  display: flex;
+  align-items: center;
+  gap: 0.5rem;
+  padding: 0.25rem 0.25rem 0.35rem;
+  font-weight: 600;
+  font-size: 0.85rem;
+  color: var(--color-foreground);
+}
+
+.hermes-kanban-column-label {
+  flex: 1;
+  letter-spacing: 0.01em;
+}
+
+.hermes-kanban-column-count {
+  font-variant-numeric: tabular-nums;
+  color: var(--color-muted-foreground);
+  font-size: 0.75rem;
+  font-weight: 500;
+}
+
+.hermes-kanban-column-add {
+  appearance: none;
+  background: transparent;
+  border: 1px solid var(--color-border);
+  color: var(--color-foreground);
+  border-radius: var(--radius-sm, 0.25rem);
+  width: 22px;
+  height: 22px;
+  line-height: 1;
+  font-size: 1rem;
+  cursor: pointer;
+}
+.hermes-kanban-column-add:hover {
+  background: color-mix(in srgb, var(--color-foreground) 8%, transparent);
+}
+
+.hermes-kanban-column-sub {
+  padding: 0 0.25rem 0.5rem;
+  font-size: 0.7rem;
+  color: var(--color-muted-foreground);
+  border-bottom: 1px solid color-mix(in srgb, var(--color-border) 60%, transparent);
+  margin-bottom: 0.5rem;
+}
+
+.hermes-kanban-column-body {
+  display: flex;
+  flex-direction: column;
+  gap: 0.45rem;
+  overflow-y: auto;
+  padding-right: 0.1rem;
+}
+
+.hermes-kanban-empty {
+  padding: 1.5rem 0.5rem;
+  text-align: center;
+  font-size: 0.75rem;
+  color: var(--color-muted-foreground);
+  border: 1px dashed color-mix(in srgb, var(--color-border) 70%, transparent);
+  border-radius: var(--radius-sm, 0.25rem);
+}
+
+/* ---- Status dots ----------------------------------------------------- */
+
+.hermes-kanban-dot {
+  display: inline-block;
+  width: 0.5rem;
+  height: 0.5rem;
+  border-radius: 999px;
+  background: var(--color-muted-foreground);
+}
+.hermes-kanban-dot-triage   { background: #b47dd6; }  /* lilac — fresh/unspecified */
+.hermes-kanban-dot-todo     { background: var(--color-muted-foreground); }
+.hermes-kanban-dot-ready    { background: #d4b348; }  /* amber */
+.hermes-kanban-dot-running  { background: #3fb97d; }  /* green */
+.hermes-kanban-dot-blocked  { background: var(--color-destructive, #d14a4a); }
+.hermes-kanban-dot-done     { background: #4a8cd1; }  /* blue */
+.hermes-kanban-dot-archived { background: var(--color-border); }
+
+/* ---- Progress pill (N/M child tasks done) --------------------------- */
+
+.hermes-kanban-progress {
+  font-family: var(--font-mono, ui-monospace, monospace);
+  font-size: 0.62rem;
+  padding: 0.05rem 0.35rem;
+  border-radius: 999px;
+  background: color-mix(in srgb, var(--color-foreground) 8%, transparent);
+  border: 1px solid color-mix(in srgb, var(--color-border) 80%, transparent);
+  color: var(--color-muted-foreground);
+  letter-spacing: 0.02em;
+}
+.hermes-kanban-progress--full {
+  background: color-mix(in srgb, #3fb97d 22%, transparent);
+  border-color: color-mix(in srgb, #3fb97d 45%, transparent);
+  color: var(--color-foreground);
+}
+
+/* ---- Lanes (per-profile sub-grouping inside Running) ---------------- */
+
+.hermes-kanban-lane {
+  display: flex;
+  flex-direction: column;
+  gap: 0.35rem;
+  padding: 0.25rem 0 0.35rem;
+  border-top: 1px dashed color-mix(in srgb, var(--color-border) 70%, transparent);
+}
+.hermes-kanban-lane:first-child {
+  border-top: 0;
+  padding-top: 0;
+}
+.hermes-kanban-lane-head {
+  display: flex;
+  align-items: center;
+  gap: 0.4rem;
+  font-size: 0.65rem;
+  text-transform: uppercase;
+  letter-spacing: 0.08em;
+  color: var(--color-muted-foreground);
+  padding: 0 0.1rem;
+}
+.hermes-kanban-lane-name {
+  font-weight: 600;
+  font-family: var(--font-mono, ui-monospace, monospace);
+}
+.hermes-kanban-lane-count {
+  margin-left: auto;
+  font-variant-numeric: tabular-nums;
+}
+
+/* ---- Card ------------------------------------------------------------ */
+
+.hermes-kanban-card {
+  cursor: grab;
+  transition: transform 100ms ease, box-shadow 100ms ease;
+}
+.hermes-kanban-card:hover {
+  box-shadow: 0 1px 0 0 var(--color-ring) inset, 0 0 0 1px var(--color-ring) inset;
+}
+.hermes-kanban-card:active {
+  cursor: grabbing;
+  transform: scale(0.995);
+}
+
+.hermes-kanban-card-content {
+  padding: 0.5rem 0.6rem !important;
+  display: flex;
+  flex-direction: column;
+  gap: 0.3rem;
+}
+
+.hermes-kanban-card-row {
+  display: flex;
+  align-items: center;
+  gap: 0.35rem;
+  flex-wrap: wrap;
+}
+
+.hermes-kanban-card-id {
+  font-family: var(--font-mono, ui-monospace, monospace);
+  font-size: 0.65rem;
+  color: var(--color-muted-foreground);
+  letter-spacing: 0.03em;
+}
+
+.hermes-kanban-card-title {
+  font-size: 0.85rem;
+  font-weight: 500;
+  line-height: 1.3;
+  color: var(--color-foreground);
+  word-break: break-word;
+}
+
+.hermes-kanban-card-meta {
+  font-size: 0.7rem;
+  color: var(--color-muted-foreground);
+  gap: 0.55rem;
+}
+
+.hermes-kanban-priority {
+  font-size: 0.6rem !important;
+  padding: 0.05rem 0.3rem !important;
+  background: color-mix(in srgb, var(--color-ring) 18%, transparent);
+  color: var(--color-foreground);
+  border: 1px solid color-mix(in srgb, var(--color-ring) 40%, transparent);
+}
+
+.hermes-kanban-tag {
+  font-size: 0.6rem !important;
+  padding: 0.05rem 0.3rem !important;
+}
+
+.hermes-kanban-assignee {
+  font-weight: 500;
+  color: color-mix(in srgb, var(--color-foreground) 80%, var(--color-muted-foreground));
+}
+.hermes-kanban-unassigned {
+  font-style: italic;
+}
+.hermes-kanban-ago {
+  margin-left: auto;
+}
+
+/* ---- Inline create --------------------------------------------------- */
+
+.hermes-kanban-inline-create {
+  display: flex;
+  flex-direction: column;
+  gap: 0.35rem;
+  padding: 0.5rem;
+  margin-bottom: 0.5rem;
+  background: color-mix(in srgb, var(--color-card) 70%, transparent);
+  border: 1px dashed var(--color-border);
+  border-radius: var(--radius-sm, 0.25rem);
+}
+
+/* ---- Drawer (task detail side panel) --------------------------------- */
+
+.hermes-kanban-drawer-shade {
+  position: fixed;
+  inset: 0;
+  background: rgba(0, 0, 0, 0.45);
+  z-index: 60;
+  display: flex;
+  justify-content: flex-end;
+}
+
+.hermes-kanban-drawer {
+  width: min(480px, 92vw);
+  height: 100vh;
+  background: var(--color-card);
+  border-left: 1px solid var(--color-border);
+  display: flex;
+  flex-direction: column;
+  box-shadow: -4px 0 18px rgba(0, 0, 0, 0.35);
+  animation: hermes-kanban-drawer-in 180ms ease-out;
+}
+
+@keyframes hermes-kanban-drawer-in {
+  from { transform: translateX(100%); opacity: 0.3; }
+  to   { transform: translateX(0);    opacity: 1; }
+}
+
+.hermes-kanban-drawer-head {
+  display: flex;
+  align-items: center;
+  justify-content: space-between;
+  padding: 0.6rem 0.8rem;
+  border-bottom: 1px solid var(--color-border);
+  font-family: var(--font-mono, ui-monospace, monospace);
+}
+
+.hermes-kanban-drawer-close {
+  appearance: none;
+  background: transparent;
+  border: 0;
+  color: var(--color-muted-foreground);
+  font-size: 1.25rem;
+  line-height: 1;
+  cursor: pointer;
+  padding: 0 0.25rem;
+}
+.hermes-kanban-drawer-close:hover { color: var(--color-foreground); }
+
+.hermes-kanban-drawer-body {
+  flex: 1;
+  overflow-y: auto;
+  padding: 0.9rem;
+  display: flex;
+  flex-direction: column;
+  gap: 0.85rem;
+}
+
+.hermes-kanban-drawer-title {
+  display: flex;
+  align-items: center;
+  gap: 0.5rem;
+  font-size: 1rem;
+  font-weight: 600;
+}
+
+.hermes-kanban-drawer-meta {
+  display: flex;
+  flex-direction: column;
+  gap: 0.15rem;
+  padding: 0.5rem 0.6rem;
+  background: color-mix(in srgb, var(--color-foreground) 4%, transparent);
+  border: 1px solid var(--color-border);
+  border-radius: var(--radius-sm, 0.25rem);
+}
+
+.hermes-kanban-meta-row {
+  display: flex;
+  gap: 0.5rem;
+  font-size: 0.72rem;
+}
+.hermes-kanban-meta-label {
+  width: 92px;
+  color: var(--color-muted-foreground);
+}
+.hermes-kanban-meta-value {
+  color: var(--color-foreground);
+  word-break: break-word;
+}
+
+.hermes-kanban-actions {
+  display: flex;
+  flex-wrap: wrap;
+  gap: 0.3rem;
+}
+
+.hermes-kanban-section {
+  display: flex;
+  flex-direction: column;
+  gap: 0.35rem;
+}
+
+.hermes-kanban-section-head {
+  font-size: 0.72rem;
+  font-weight: 600;
+  text-transform: uppercase;
+  letter-spacing: 0.07em;
+  color: var(--color-muted-foreground);
+}
+
+.hermes-kanban-pre {
+  margin: 0;
+  padding: 0.45rem 0.55rem;
+  white-space: pre-wrap;
+  word-break: break-word;
+  background: color-mix(in srgb, var(--color-foreground) 4%, transparent);
+  border: 1px solid var(--color-border);
+  border-radius: var(--radius-sm, 0.25rem);
+  font-family: var(--font-mono, ui-monospace, monospace);
+  font-size: 0.72rem;
+  color: var(--color-foreground);
+}
+
+.hermes-kanban-comment {
+  border-left: 2px solid color-mix(in srgb, var(--color-ring) 35%, transparent);
+  padding-left: 0.5rem;
+  display: flex;
+  flex-direction: column;
+  gap: 0.2rem;
+}
+
+.hermes-kanban-comment-head {
+  display: flex;
+  gap: 0.5rem;
+  font-size: 0.7rem;
+}
+.hermes-kanban-comment-author {
+  font-weight: 600;
+  color: var(--color-foreground);
+}
+.hermes-kanban-comment-ago {
+  color: var(--color-muted-foreground);
+}
+
+.hermes-kanban-event {
+  display: flex;
+  gap: 0.5rem;
+  font-size: 0.7rem;
+  color: var(--color-muted-foreground);
+  font-family: var(--font-mono, ui-monospace, monospace);
+}
+.hermes-kanban-event-kind {
+  color: var(--color-foreground);
+  min-width: 6rem;
+}
+.hermes-kanban-event-payload {
+  color: var(--color-muted-foreground);
+  overflow: hidden;
+  text-overflow: ellipsis;
+  white-space: nowrap;
+  max-width: 280px;
+}
+
+.hermes-kanban-drawer-comment-row {
+  display: flex;
+  gap: 0.4rem;
+  padding: 0.55rem 0.75rem;
+  border-top: 1px solid var(--color-border);
+  background: color-mix(in srgb, var(--color-card) 90%, transparent);
+}
+
+.hermes-kanban-count {
+  display: inline-flex;
+  gap: 0.2rem;
+  align-items: center;
+}
+
+/* ---- Selection chrome ----------------------------------------------- */
+
+.hermes-kanban-card--selected :where(.hermes-kanban-card-content) {
+  box-shadow: 0 0 0 2px var(--color-ring) inset,
+              0 0 0 1px var(--color-ring) inset;
+  background: color-mix(in srgb, var(--color-ring) 6%, var(--color-card));
+}
+
+.hermes-kanban-card-check {
+  width: 0.85rem;
+  height: 0.85rem;
+  margin: 0;
+  cursor: pointer;
+  accent-color: var(--color-ring);
+}
+
+/* ---- Bulk action bar ------------------------------------------------ */
+
+.hermes-kanban-bulk {
+  display: flex;
+  align-items: center;
+  gap: 0.5rem;
+  padding: 0.4rem 0.75rem;
+  background: color-mix(in srgb, var(--color-ring) 10%, var(--color-card));
+  border: 1px solid color-mix(in srgb, var(--color-ring) 40%, var(--color-border));
+  border-radius: var(--radius-sm, 0.25rem);
+  flex-wrap: wrap;
+}
+.hermes-kanban-bulk-count {
+  font-weight: 600;
+  font-size: 0.75rem;
+  padding-right: 0.25rem;
+}
+.hermes-kanban-bulk-btn {
+  height: 1.7rem !important;
+  padding: 0 0.5rem !important;
+  font-size: 0.7rem !important;
+  border: 1px solid var(--color-border);
+  cursor: pointer;
+}
+.hermes-kanban-bulk-btn:hover {
+  background: color-mix(in srgb, var(--color-foreground) 8%, transparent);
+}
+.hermes-kanban-bulk-reassign {
+  display: flex;
+  align-items: center;
+  gap: 0.25rem;
+  padding-left: 0.5rem;
+  border-left: 1px solid color-mix(in srgb, var(--color-border) 70%, transparent);
+}
+
+/* ---- Dependency editor chips --------------------------------------- */
+
+.hermes-kanban-deps-row {
+  display: flex;
+  align-items: center;
+  gap: 0.5rem;
+  margin-bottom: 0.4rem;
+}
+.hermes-kanban-deps-label {
+  font-size: 0.68rem;
+  text-transform: uppercase;
+  letter-spacing: 0.08em;
+  color: var(--color-muted-foreground);
+  min-width: 4rem;
+}
+.hermes-kanban-deps-chips {
+  display: flex;
+  gap: 0.3rem;
+  flex-wrap: wrap;
+  flex: 1;
+}
+.hermes-kanban-deps-empty {
+  font-size: 0.7rem;
+  color: var(--color-muted-foreground);
+  font-style: italic;
+}
+.hermes-kanban-dep-chip {
+  display: inline-flex;
+  align-items: center;
+  gap: 0.15rem;
+  padding: 0.1rem 0.35rem;
+  background: color-mix(in srgb, var(--color-foreground) 6%, transparent);
+  border: 1px solid var(--color-border);
+  border-radius: var(--radius-sm, 0.25rem);
+  font-family: var(--font-mono, ui-monospace, monospace);
+  font-size: 0.68rem;
+  color: var(--color-foreground);
+}
+.hermes-kanban-dep-chip-x {
+  appearance: none;
+  background: transparent;
+  border: 0;
+  color: var(--color-muted-foreground);
+  cursor: pointer;
+  font-size: 0.85rem;
+  line-height: 1;
+  padding: 0 0.15rem;
+}
+.hermes-kanban-dep-chip-x:hover { color: var(--color-destructive, #d14a4a); }
+
+/* ---- Inline edit affordances --------------------------------------- */
+
+.hermes-kanban-editable {
+  cursor: pointer;
+  border-bottom: 1px dotted color-mix(in srgb, var(--color-border) 80%, transparent);
+}
+.hermes-kanban-editable:hover {
+  color: var(--color-foreground);
+  border-bottom-color: var(--color-ring);
+}
+
+.hermes-kanban-drawer-title-text {
+  cursor: pointer;
+}
+.hermes-kanban-drawer-title-text:hover {
+  text-decoration: underline;
+  text-decoration-color: var(--color-ring);
+  text-decoration-style: dotted;
+  text-underline-offset: 3px;
+}
+
+.hermes-kanban-edit-row {
+  display: flex;
+  align-items: center;
+  gap: 0.35rem;
+  width: 100%;
+}
+
+.hermes-kanban-section-head-row {
+  display: flex;
+  align-items: center;
+  justify-content: space-between;
+  gap: 0.5rem;
+}
+.hermes-kanban-edit-link {
+  appearance: none;
+  background: transparent;
+  border: 0;
+  color: var(--color-muted-foreground);
+  font-size: 0.7rem;
+  text-transform: uppercase;
+  letter-spacing: 0.05em;
+  cursor: pointer;
+  padding: 0;
+}
+.hermes-kanban-edit-link:hover { color: var(--color-ring); }
+
+.hermes-kanban-textarea {
+  width: 100%;
+  min-height: 8rem;
+  background: var(--color-card);
+  color: var(--color-foreground);
+  border: 1px solid var(--color-border);
+  border-radius: var(--radius-sm, 0.25rem);
+  padding: 0.5rem 0.6rem;
+  font-family: var(--font-mono, ui-monospace, monospace);
+  font-size: 0.8rem;
+  line-height: 1.5;
+  resize: vertical;
+}
+.hermes-kanban-textarea:focus {
+  outline: none;
+  border-color: var(--color-ring);
+  box-shadow: 0 0 0 2px color-mix(in srgb, var(--color-ring) 30%, transparent);
+}
+
+/* ---- Markdown rendering -------------------------------------------- */
+
+.hermes-kanban-md {
+  font-size: 0.8rem;
+  line-height: 1.55;
+  color: var(--color-foreground);
+}
+.hermes-kanban-md p  { margin: 0.25rem 0; }
+.hermes-kanban-md h1,
+.hermes-kanban-md h2,
+.hermes-kanban-md h3,
+.hermes-kanban-md h4 {
+  margin: 0.6rem 0 0.2rem;
+  line-height: 1.25;
+}
+.hermes-kanban-md h1 { font-size: 1.05rem; }
+.hermes-kanban-md h2 { font-size: 0.95rem; }
+.hermes-kanban-md h3 { font-size: 0.88rem; }
+.hermes-kanban-md h4 { font-size: 0.82rem; }
+.hermes-kanban-md ul {
+  margin: 0.25rem 0 0.25rem 1.1rem;
+  padding: 0;
+}
+.hermes-kanban-md li { margin: 0.1rem 0; }
+.hermes-kanban-md a {
+  color: var(--color-ring);
+  text-decoration: underline;
+}
+.hermes-kanban-md code {
+  font-family: var(--font-mono, ui-monospace, monospace);
+  font-size: 0.75rem;
+  padding: 0.05rem 0.3rem;
+  background: color-mix(in srgb, var(--color-foreground) 8%, transparent);
+  border-radius: 3px;
+}
+.hermes-kanban-md-code {
+  margin: 0.35rem 0;
+  padding: 0.5rem 0.6rem;
+  background: color-mix(in srgb, var(--color-foreground) 5%, transparent);
+  border: 1px solid var(--color-border);
+  border-radius: var(--radius-sm, 0.25rem);
+  overflow-x: auto;
+}
+.hermes-kanban-md-code code {
+  background: transparent;
+  padding: 0;
+  font-size: 0.75rem;
+  white-space: pre;
+}
+.hermes-kanban-md strong { font-weight: 600; }
+
+/* ---- Touch-drag proxy ---------------------------------------------- */
+
+.hermes-kanban-touch-proxy {
+  pointer-events: none;
+  opacity: 0.85;
+  box-shadow: 0 8px 20px rgba(0, 0, 0, 0.35);
+  transform: scale(1.02);
+  transition: none;
+}
+
+
+/* ---- Staleness tiers ------------------------------------------------ */
+
+.hermes-kanban-card--stale-amber :where(.hermes-kanban-card-content) {
+  box-shadow: 0 0 0 1px #d4b34888 inset;
+}
+.hermes-kanban-card--stale-amber:hover :where(.hermes-kanban-card-content) {
+  box-shadow: 0 0 0 2px #d4b348 inset;
+}
+.hermes-kanban-card--stale-red :where(.hermes-kanban-card-content) {
+  box-shadow: 0 0 0 1px var(--color-destructive, #d14a4a) inset,
+              0 0 8px color-mix(in srgb, var(--color-destructive, #d14a4a) 30%, transparent);
+}
+.hermes-kanban-card--stale-red:hover :where(.hermes-kanban-card-content) {
+  box-shadow: 0 0 0 2px var(--color-destructive, #d14a4a) inset,
+              0 0 10px color-mix(in srgb, var(--color-destructive, #d14a4a) 45%, transparent);
+}
+
+/* ---- Worker log pane ------------------------------------------------ */
+
+.hermes-kanban-log {
+  max-height: 340px;
+  overflow: auto;
+  white-space: pre;
+  font-size: 0.7rem;
+  line-height: 1.45;
+}
+
+
+/* ---- Run history (per-attempt log in the drawer) ------------------- */
+
+.hermes-kanban-run {
+  border-left: 2px solid var(--color-border);
+  padding: 0.35rem 0.5rem;
+  margin-bottom: 0.4rem;
+  background: color-mix(in srgb, var(--color-foreground) 3%, transparent);
+  border-radius: var(--radius-sm, 0.25rem);
+}
+.hermes-kanban-run--active  { border-left-color: #3fb97d; }
+.hermes-kanban-run--completed { border-left-color: #4a8cd1; }
+.hermes-kanban-run--ended { border-left-color: #6b7280; }  /* generic fallback when outcome is unset */
+.hermes-kanban-run--blocked { border-left-color: var(--color-destructive, #d14a4a); }
+.hermes-kanban-run--crashed,
+.hermes-kanban-run--timed_out,
+.hermes-kanban-run--gave_up,
+.hermes-kanban-run--spawn_failed {
+  border-left-color: var(--color-destructive, #d14a4a);
+  background: color-mix(in srgb, var(--color-destructive, #d14a4a) 6%, transparent);
+}
+.hermes-kanban-run--reclaimed { border-left-color: #d4b348; }
+
+.hermes-kanban-run-head {
+  display: flex;
+  align-items: center;
+  gap: 0.6rem;
+  font-size: 0.7rem;
+}
+.hermes-kanban-run-outcome {
+  font-family: var(--font-mono, ui-monospace, monospace);
+  font-weight: 600;
+  text-transform: uppercase;
+  letter-spacing: 0.05em;
+  color: var(--color-foreground);
+}
+.hermes-kanban-run-profile {
+  color: var(--color-muted-foreground);
+}
+.hermes-kanban-run-elapsed {
+  font-variant-numeric: tabular-nums;
+  color: var(--color-muted-foreground);
+}
+.hermes-kanban-run-ago {
+  margin-left: auto;
+  color: var(--color-muted-foreground);
+}
+.hermes-kanban-run-summary {
+  font-size: 0.75rem;
+  padding: 0.2rem 0 0;
+  color: var(--color-foreground);
+}
+.hermes-kanban-run-error {
+  font-size: 0.7rem;
+  color: var(--color-destructive, #d14a4a);
+  padding: 0.15rem 0 0;
+  font-family: var(--font-mono, ui-monospace, monospace);
+}
+.hermes-kanban-run-meta {
+  display: block;
+  font-size: 0.65rem;
+  padding: 0.15rem 0 0;
+  color: var(--color-muted-foreground);
+  white-space: pre-wrap;
+  word-break: break-word;
+  font-family: var(--font-mono, ui-monospace, monospace);
+}

plugins/kanban/dashboard/manifest.json

@@ -0,0 +1,14 @@
+{
+  "name": "kanban",
+  "label": "Kanban",
+  "description": "Multi-agent collaboration board — drag-drop cards across columns, read comment threads, see which profile is running what",
+  "icon": "Package",
+  "version": "1.0.0",
+  "tab": {
+    "path": "/kanban",
+    "position": "after:skills"
+  },
+  "entry": "dist/index.js",
+  "css": "dist/style.css",
+  "api": "plugin_api.py"
+}

plugins/kanban/dashboard/plugin_api.py

@@ -0,0 +1,830 @@
+"""Kanban dashboard plugin — backend API routes.
+
+Mounted at /api/plugins/kanban/ by the dashboard plugin system.
+
+This layer is intentionally thin: every handler is a small wrapper around
+``hermes_cli.kanban_db`` or a direct SQL query. Writes use the same code
+paths the CLI and gateway ``/kanban`` command use, so the three surfaces
+cannot drift.
+
+Live updates arrive via the ``/events`` WebSocket, which tails the
+append-only ``task_events`` table on a short poll interval (WAL mode lets
+reads run alongside the dispatcher's IMMEDIATE write transactions).
+
+Security note
+-------------
+The dashboard's HTTP auth middleware (``web_server.auth_middleware``)
+explicitly skips ``/api/plugins/`` — plugin routes are unauthenticated by
+design because the dashboard binds to localhost by default. For the
+WebSocket we still require the session token as a ``?token=`` query
+parameter (browsers cannot set the ``Authorization`` header on an upgrade
+request), matching the established pattern used by the in-browser PTY
+bridge in ``hermes_cli/web_server.py``. If you run the dashboard with
+``--host 0.0.0.0``, every plugin route — kanban included — becomes
+reachable from the network. Don't do that on a shared host.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import hmac
+import json
+import logging
+import sqlite3
+import time
+from dataclasses import asdict
+from typing import Any, Optional
+
+from fastapi import APIRouter, HTTPException, Query, WebSocket, WebSocketDisconnect, status as http_status
+from pydantic import BaseModel, Field
+
+from hermes_cli import kanban_db
+
+log = logging.getLogger(__name__)
+
+router = APIRouter()
+
+
+# ---------------------------------------------------------------------------
+# Auth helper — WebSocket only (HTTP routes live behind the dashboard's
+# existing plugin-bypass; this is documented above).
+# ---------------------------------------------------------------------------
+
+def _check_ws_token(provided: Optional[str]) -> bool:
+    """Constant-time compare against the dashboard session token.
+
+    Imported lazily so the plugin still loads in test contexts where the
+    dashboard web_server module isn't importable (e.g. the bare-FastAPI
+    test harness).
+    """
+    if not provided:
+        return False
+    try:
+        from hermes_cli import web_server as _ws
+    except Exception:
+        # No dashboard context (tests). Accept so the tail loop is still
+        # testable; in production the dashboard module always imports
+        # cleanly because it's the caller.
+        return True
+    expected = getattr(_ws, "_SESSION_TOKEN", None)
+    if not expected:
+        return True
+    return hmac.compare_digest(str(provided), str(expected))
+
+
+def _conn():
+    """Open a kanban_db connection, creating the schema on first use.
+
+    Every handler that mutates the DB goes through this so the plugin
+    self-heals on a fresh install (no user-visible "no such table"
+    error if somebody hits POST /tasks before GET /board).
+    ``init_db`` is idempotent.
+    """
+    try:
+        kanban_db.init_db()
+    except Exception as exc:
+        log.warning("kanban init_db failed: %s", exc)
+    return kanban_db.connect()
+
+
+# ---------------------------------------------------------------------------
+# Serialization helpers
+# ---------------------------------------------------------------------------
+
+# Columns shown by the dashboard, in left-to-right order. "archived" is
+# available via a filter toggle rather than a visible column.
+BOARD_COLUMNS: list[str] = [
+    "triage", "todo", "ready", "running", "blocked", "done",
+]
+
+
+def _task_dict(task: kanban_db.Task) -> dict[str, Any]:
+    d = asdict(task)
+    # Add derived age metrics so the UI can colour stale cards without
+    # computing deltas client-side.
+    d["age"] = kanban_db.task_age(task)
+    # Keep body short on list endpoints; full body comes from /tasks/:id.
+    return d
+
+
+def _event_dict(event: kanban_db.Event) -> dict[str, Any]:
+    return {
+        "id": event.id,
+        "task_id": event.task_id,
+        "kind": event.kind,
+        "payload": event.payload,
+        "created_at": event.created_at,
+        "run_id": event.run_id,
+    }
+
+
+def _comment_dict(c: kanban_db.Comment) -> dict[str, Any]:
+    return {
+        "id": c.id,
+        "task_id": c.task_id,
+        "author": c.author,
+        "body": c.body,
+        "created_at": c.created_at,
+    }
+
+
+def _run_dict(r: kanban_db.Run) -> dict[str, Any]:
+    """Serialise a Run for the drawer's Run history section."""
+    return {
+        "id": r.id,
+        "task_id": r.task_id,
+        "profile": r.profile,
+        "step_key": r.step_key,
+        "status": r.status,
+        "claim_lock": r.claim_lock,
+        "claim_expires": r.claim_expires,
+        "worker_pid": r.worker_pid,
+        "max_runtime_seconds": r.max_runtime_seconds,
+        "last_heartbeat_at": r.last_heartbeat_at,
+        "started_at": r.started_at,
+        "ended_at": r.ended_at,
+        "outcome": r.outcome,
+        "summary": r.summary,
+        "metadata": r.metadata,
+        "error": r.error,
+    }
+
+
+def _links_for(conn: sqlite3.Connection, task_id: str) -> dict[str, list[str]]:
+    """Return {'parents': [...], 'children': [...]} for a task."""
+    parents = [
+        r["parent_id"]
+        for r in conn.execute(
+            "SELECT parent_id FROM task_links WHERE child_id = ? ORDER BY parent_id",
+            (task_id,),
+        )
+    ]
+    children = [
+        r["child_id"]
+        for r in conn.execute(
+            "SELECT child_id FROM task_links WHERE parent_id = ? ORDER BY child_id",
+            (task_id,),
+        )
+    ]
+    return {"parents": parents, "children": children}
+
+
+# ---------------------------------------------------------------------------
+# GET /board
+# ---------------------------------------------------------------------------
+
+@router.get("/board")
+def get_board(
+    tenant: Optional[str] = Query(None, description="Filter to a single tenant"),
+    include_archived: bool = Query(False),
+):
+    """Return the full board grouped by status column.
+
+    ``_conn()`` auto-initializes ``kanban.db`` on first call so a fresh
+    install doesn't surface a "failed to load" error on the plugin tab.
+    """
+    conn = _conn()
+    try:
+        tasks = kanban_db.list_tasks(
+            conn, tenant=tenant, include_archived=include_archived
+        )
+        # Pre-fetch link counts per task (cheap: one query).
+        link_counts: dict[str, dict[str, int]] = {}
+        for row in conn.execute(
+            "SELECT parent_id, child_id FROM task_links"
+        ).fetchall():
+            link_counts.setdefault(row["parent_id"], {"parents": 0, "children": 0})[
+                "children"
+            ] += 1
+            link_counts.setdefault(row["child_id"], {"parents": 0, "children": 0})[
+                "parents"
+            ] += 1
+
+        # Comment + event counts (both cheap aggregates).
+        comment_counts: dict[str, int] = {
+            r["task_id"]: r["n"]
+            for r in conn.execute(
+                "SELECT task_id, COUNT(*) AS n FROM task_comments GROUP BY task_id"
+            )
+        }
+
+        # Progress rollup: for each parent, how many children are done / total.
+        # One pass over task_links joined with child status — cheaper than
+        # N per-task queries and the plugin uses it to render "N/M".
+        progress: dict[str, dict[str, int]] = {}
+        for row in conn.execute(
+            "SELECT l.parent_id AS pid, t.status AS cstatus "
+            "FROM task_links l JOIN tasks t ON t.id = l.child_id"
+        ).fetchall():
+            p = progress.setdefault(row["pid"], {"done": 0, "total": 0})
+            p["total"] += 1
+            if row["cstatus"] == "done":
+                p["done"] += 1
+
+        latest_event_id = conn.execute(
+            "SELECT COALESCE(MAX(id), 0) AS m FROM task_events"
+        ).fetchone()["m"]
+
+        columns: dict[str, list[dict]] = {c: [] for c in BOARD_COLUMNS}
+        if include_archived:
+            columns["archived"] = []
+
+        for t in tasks:
+            d = _task_dict(t)
+            d["link_counts"] = link_counts.get(t.id, {"parents": 0, "children": 0})
+            d["comment_count"] = comment_counts.get(t.id, 0)
+            d["progress"] = progress.get(t.id)  # None when the task has no children
+            col = t.status if t.status in columns else "todo"
+            columns[col].append(d)
+
+        # Stable per-column ordering already applied by list_tasks
+        # (priority DESC, created_at ASC), keep as-is.
+
+        # List of known tenants for the UI filter dropdown.
+        tenants = [
+            r["tenant"]
+            for r in conn.execute(
+                "SELECT DISTINCT tenant FROM tasks WHERE tenant IS NOT NULL ORDER BY tenant"
+            )
+        ]
+        # List of distinct assignees for the lane-by-profile sub-grouping.
+        assignees = [
+            r["assignee"]
+            for r in conn.execute(
+                "SELECT DISTINCT assignee FROM tasks WHERE assignee IS NOT NULL "
+                "AND status != 'archived' ORDER BY assignee"
+            )
+        ]
+
+        return {
+            "columns": [
+                {"name": name, "tasks": columns[name]} for name in columns.keys()
+            ],
+            "tenants": tenants,
+            "assignees": assignees,
+            "latest_event_id": int(latest_event_id),
+            "now": int(time.time()),
+        }
+    finally:
+        conn.close()
+
+
+# ---------------------------------------------------------------------------
+# GET /tasks/:id
+# ---------------------------------------------------------------------------
+
+@router.get("/tasks/{task_id}")
+def get_task(task_id: str):
+    conn = _conn()
+    try:
+        task = kanban_db.get_task(conn, task_id)
+        if task is None:
+            raise HTTPException(status_code=404, detail=f"task {task_id} not found")
+        return {
+            "task": _task_dict(task),
+            "comments": [_comment_dict(c) for c in kanban_db.list_comments(conn, task_id)],
+            "events": [_event_dict(e) for e in kanban_db.list_events(conn, task_id)],
+            "links": _links_for(conn, task_id),
+            "runs": [_run_dict(r) for r in kanban_db.list_runs(conn, task_id)],
+        }
+    finally:
+        conn.close()
+
+
+# ---------------------------------------------------------------------------
+# POST /tasks
+# ---------------------------------------------------------------------------
+
+class CreateTaskBody(BaseModel):
+    title: str
+    body: Optional[str] = None
+    assignee: Optional[str] = None
+    tenant: Optional[str] = None
+    priority: int = 0
+    workspace_kind: str = "scratch"
+    workspace_path: Optional[str] = None
+    parents: list[str] = Field(default_factory=list)
+    triage: bool = False
+    idempotency_key: Optional[str] = None
+    max_runtime_seconds: Optional[int] = None
+    skills: Optional[list[str]] = None
+
+
+@router.post("/tasks")
+def create_task(payload: CreateTaskBody):
+    conn = _conn()
+    try:
+        task_id = kanban_db.create_task(
+            conn,
+            title=payload.title,
+            body=payload.body,
+            assignee=payload.assignee,
+            created_by="dashboard",
+            workspace_kind=payload.workspace_kind,
+            workspace_path=payload.workspace_path,
+            tenant=payload.tenant,
+            priority=payload.priority,
+            parents=payload.parents,
+            triage=payload.triage,
+            idempotency_key=payload.idempotency_key,
+            max_runtime_seconds=payload.max_runtime_seconds,
+            skills=payload.skills,
+        )
+        task = kanban_db.get_task(conn, task_id)
+        return {"task": _task_dict(task) if task else None}
+    except ValueError as e:
+        raise HTTPException(status_code=400, detail=str(e))
+    finally:
+        conn.close()
+
+
+# ---------------------------------------------------------------------------
+# PATCH /tasks/:id  (status / assignee / priority / title / body)
+# ---------------------------------------------------------------------------
+
+class UpdateTaskBody(BaseModel):
+    status: Optional[str] = None
+    assignee: Optional[str] = None
+    priority: Optional[int] = None
+    title: Optional[str] = None
+    body: Optional[str] = None
+    result: Optional[str] = None
+    block_reason: Optional[str] = None
+    # Structured handoff fields — forwarded to complete_task when status
+    # transitions to 'done'. Dashboard parity with ``hermes kanban
+    # complete --summary ... --metadata ...``.
+    summary: Optional[str] = None
+    metadata: Optional[dict] = None
+
+
+@router.patch("/tasks/{task_id}")
+def update_task(task_id: str, payload: UpdateTaskBody):
+    conn = _conn()
+    try:
+        task = kanban_db.get_task(conn, task_id)
+        if task is None:
+            raise HTTPException(status_code=404, detail=f"task {task_id} not found")
+
+        # --- assignee ----------------------------------------------------
+        if payload.assignee is not None:
+            try:
+                ok = kanban_db.assign_task(
+                    conn, task_id, payload.assignee or None,
+                )
+            except RuntimeError as e:
+                raise HTTPException(status_code=409, detail=str(e))
+            if not ok:
+                raise HTTPException(status_code=404, detail="task not found")
+
+        # --- status -------------------------------------------------------
+        if payload.status is not None:
+            s = payload.status
+            ok = True
+            if s == "done":
+                ok = kanban_db.complete_task(
+                    conn, task_id,
+                    result=payload.result,
+                    summary=payload.summary,
+                    metadata=payload.metadata,
+                )
+            elif s == "blocked":
+                ok = kanban_db.block_task(conn, task_id, reason=payload.block_reason)
+            elif s == "ready":
+                # Re-open a blocked task, or just an explicit status set.
+                current = kanban_db.get_task(conn, task_id)
+                if current and current.status == "blocked":
+                    ok = kanban_db.unblock_task(conn, task_id)
+                else:
+                    # Direct status write for drag-drop (todo -> ready etc).
+                    ok = _set_status_direct(conn, task_id, "ready")
+            elif s == "archived":
+                ok = kanban_db.archive_task(conn, task_id)
+            elif s in ("todo", "running", "triage"):
+                ok = _set_status_direct(conn, task_id, s)
+            else:
+                raise HTTPException(status_code=400, detail=f"unknown status: {s}")
+            if not ok:
+                raise HTTPException(
+                    status_code=409,
+                    detail=f"status transition to {s!r} not valid from current state",
+                )
+
+        # --- priority -----------------------------------------------------
+        if payload.priority is not None:
+            with kanban_db.write_txn(conn):
+                conn.execute(
+                    "UPDATE tasks SET priority = ? WHERE id = ?",
+                    (int(payload.priority), task_id),
+                )
+                conn.execute(
+                    "INSERT INTO task_events (task_id, kind, payload, created_at) "
+                    "VALUES (?, 'reprioritized', ?, ?)",
+                    (task_id, json.dumps({"priority": int(payload.priority)}),
+                     int(time.time())),
+                )
+
+        # --- title / body -------------------------------------------------
+        if payload.title is not None or payload.body is not None:
+            with kanban_db.write_txn(conn):
+                sets, vals = [], []
+                if payload.title is not None:
+                    if not payload.title.strip():
+                        raise HTTPException(status_code=400, detail="title cannot be empty")
+                    sets.append("title = ?")
+                    vals.append(payload.title.strip())
+                if payload.body is not None:
+                    sets.append("body = ?")
+                    vals.append(payload.body)
+                vals.append(task_id)
+                conn.execute(
+                    f"UPDATE tasks SET {', '.join(sets)} WHERE id = ?", vals,
+                )
+                conn.execute(
+                    "INSERT INTO task_events (task_id, kind, payload, created_at) "
+                    "VALUES (?, 'edited', NULL, ?)",
+                    (task_id, int(time.time())),
+                )
+
+        updated = kanban_db.get_task(conn, task_id)
+        return {"task": _task_dict(updated) if updated else None}
+    finally:
+        conn.close()
+
+
+def _set_status_direct(
+    conn: sqlite3.Connection, task_id: str, new_status: str,
+) -> bool:
+    """Direct status write for drag-drop moves that aren't covered by the
+    structured complete/block/unblock/archive verbs (e.g. todo<->ready,
+    running<->ready). Appends a ``status`` event row for the live feed.
+
+    When this transitions OFF ``running`` to anything other than the
+    terminal verbs above (which own their own run closing), we close the
+    active run with outcome='reclaimed' so attempt history isn't
+    orphaned. ``running -> ready`` via drag-drop is the common case
+    (user yanking a stuck worker back to the queue).
+    """
+    with kanban_db.write_txn(conn):
+        # Snapshot current state so we know whether to close a run.
+        prev = conn.execute(
+            "SELECT status, current_run_id FROM tasks WHERE id = ?",
+            (task_id,),
+        ).fetchone()
+        if prev is None:
+            return False
+        was_running = prev["status"] == "running"
+
+        cur = conn.execute(
+            "UPDATE tasks SET status = ?, "
+            "  claim_lock = CASE WHEN ? = 'running' THEN claim_lock ELSE NULL END, "
+            "  claim_expires = CASE WHEN ? = 'running' THEN claim_expires ELSE NULL END, "
+            "  worker_pid = CASE WHEN ? = 'running' THEN worker_pid ELSE NULL END "
+            "WHERE id = ?",
+            (new_status, new_status, new_status, new_status, task_id),
+        )
+        if cur.rowcount != 1:
+            return False
+        run_id = None
+        if was_running and new_status != "running" and prev["current_run_id"]:
+            run_id = kanban_db._end_run(
+                conn, task_id,
+                outcome="reclaimed", status="reclaimed",
+                summary=f"status changed to {new_status} (dashboard/direct)",
+            )
+        conn.execute(
+            "INSERT INTO task_events (task_id, run_id, kind, payload, created_at) "
+            "VALUES (?, ?, 'status', ?, ?)",
+            (task_id, run_id, json.dumps({"status": new_status}), int(time.time())),
+        )
+    # If we re-opened something, children may have gone stale.
+    if new_status in ("done", "ready"):
+        kanban_db.recompute_ready(conn)
+    return True
+
+
+# ---------------------------------------------------------------------------
+# Comments
+# ---------------------------------------------------------------------------
+
+class CommentBody(BaseModel):
+    body: str
+    author: Optional[str] = "dashboard"
+
+
+@router.post("/tasks/{task_id}/comments")
+def add_comment(task_id: str, payload: CommentBody):
+    if not payload.body.strip():
+        raise HTTPException(status_code=400, detail="body is required")
+    conn = _conn()
+    try:
+        if kanban_db.get_task(conn, task_id) is None:
+            raise HTTPException(status_code=404, detail=f"task {task_id} not found")
+        kanban_db.add_comment(
+            conn, task_id, author=payload.author or "dashboard", body=payload.body,
+        )
+        return {"ok": True}
+    finally:
+        conn.close()
+
+
+# ---------------------------------------------------------------------------
+# Links
+# ---------------------------------------------------------------------------
+
+class LinkBody(BaseModel):
+    parent_id: str
+    child_id: str
+
+
+@router.post("/links")
+def add_link(payload: LinkBody):
+    conn = _conn()
+    try:
+        kanban_db.link_tasks(conn, payload.parent_id, payload.child_id)
+        return {"ok": True}
+    except ValueError as e:
+        raise HTTPException(status_code=400, detail=str(e))
+    finally:
+        conn.close()
+
+
+@router.delete("/links")
+def delete_link(parent_id: str = Query(...), child_id: str = Query(...)):
+    conn = _conn()
+    try:
+        ok = kanban_db.unlink_tasks(conn, parent_id, child_id)
+        return {"ok": bool(ok)}
+    finally:
+        conn.close()
+
+
+# ---------------------------------------------------------------------------
+# Bulk actions (multi-select on the board)
+# ---------------------------------------------------------------------------
+
+class BulkTaskBody(BaseModel):
+    ids: list[str]
+    status: Optional[str] = None
+    assignee: Optional[str] = None  # "" or None = unassign
+    priority: Optional[int] = None
+    archive: bool = False
+
+
+@router.post("/tasks/bulk")
+def bulk_update(payload: BulkTaskBody):
+    """Apply the same patch to every id in ``payload.ids``.
+
+    This is an *independent* iteration — per-task failures don't abort
+    siblings. Returns per-id outcome so the UI can surface partials.
+    """
+    ids = [i for i in (payload.ids or []) if i]
+    if not ids:
+        raise HTTPException(status_code=400, detail="ids is required")
+    results: list[dict] = []
+    conn = _conn()
+    try:
+        for tid in ids:
+            entry: dict[str, Any] = {"id": tid, "ok": True}
+            try:
+                task = kanban_db.get_task(conn, tid)
+                if task is None:
+                    entry.update(ok=False, error="not found")
+                    results.append(entry)
+                    continue
+                if payload.archive:
+                    if not kanban_db.archive_task(conn, tid):
+                        entry.update(ok=False, error="archive refused")
+                if payload.status is not None and not payload.archive:
+                    s = payload.status
+                    if s == "done":
+                        ok = kanban_db.complete_task(conn, tid)
+                    elif s == "blocked":
+                        ok = kanban_db.block_task(conn, tid)
+                    elif s == "ready":
+                        cur = kanban_db.get_task(conn, tid)
+                        if cur and cur.status == "blocked":
+                            ok = kanban_db.unblock_task(conn, tid)
+                        else:
+                            ok = _set_status_direct(conn, tid, "ready")
+                    elif s in ("todo", "running", "triage"):
+                        ok = _set_status_direct(conn, tid, s)
+                    else:
+                        entry.update(ok=False, error=f"unknown status {s!r}")
+                        results.append(entry)
+                        continue
+                    if not ok:
+                        entry.update(ok=False, error=f"transition to {s!r} refused")
+                if payload.assignee is not None:
+                    try:
+                        if not kanban_db.assign_task(
+                            conn, tid, payload.assignee or None,
+                        ):
+                            entry.update(ok=False, error="assign refused")
+                    except RuntimeError as e:
+                        entry.update(ok=False, error=str(e))
+                if payload.priority is not None:
+                    with kanban_db.write_txn(conn):
+                        conn.execute(
+                            "UPDATE tasks SET priority = ? WHERE id = ?",
+                            (int(payload.priority), tid),
+                        )
+                        conn.execute(
+                            "INSERT INTO task_events (task_id, kind, payload, created_at) "
+                            "VALUES (?, 'reprioritized', ?, ?)",
+                            (tid, json.dumps({"priority": int(payload.priority)}),
+                             int(time.time())),
+                        )
+            except Exception as e:  # defensive — one bad id shouldn't kill the batch
+                entry.update(ok=False, error=str(e))
+            results.append(entry)
+        return {"results": results}
+    finally:
+        conn.close()
+
+
+# ---------------------------------------------------------------------------
+# Plugin config (read dashboard.kanban.* defaults from config.yaml)
+# ---------------------------------------------------------------------------
+
+@router.get("/config")
+def get_config():
+    """Return kanban dashboard preferences from ~/.hermes/config.yaml.
+
+    Reads the ``dashboard.kanban`` section if present; defaults otherwise.
+    Used by the UI to pre-select tenant filters, toggle markdown rendering,
+    or set column-width preferences without a round-trip per page load.
+    """
+    try:
+        from hermes_cli.config import load_config
+        cfg = load_config() or {}
+    except Exception:
+        cfg = {}
+    dash_cfg = (cfg.get("dashboard") or {})
+    # dashboard.kanban may itself be a dict; fall back to {}.
+    k_cfg = dash_cfg.get("kanban") or {}
+    return {
+        "default_tenant": k_cfg.get("default_tenant") or "",
+        "lane_by_profile": bool(k_cfg.get("lane_by_profile", True)),
+        "include_archived_by_default": bool(k_cfg.get("include_archived_by_default", False)),
+        "render_markdown": bool(k_cfg.get("render_markdown", True)),
+    }
+
+
+# ---------------------------------------------------------------------------
+# Stats (per-profile / per-status counts + oldest-ready age)
+# ---------------------------------------------------------------------------
+
+@router.get("/stats")
+def get_stats():
+    """Per-status + per-assignee counts + oldest-ready age.
+
+    Designed for the dashboard HUD and for router profiles that need to
+    answer "is this specialist overloaded?" without scanning the whole
+    board themselves.
+    """
+    conn = _conn()
+    try:
+        return kanban_db.board_stats(conn)
+    finally:
+        conn.close()
+
+
+@router.get("/assignees")
+def get_assignees():
+    """Known profiles + per-profile task counts.
+
+    Returns the union of ``~/.hermes/profiles/*`` on disk and every
+    distinct assignee currently used on the board. The dashboard uses
+    this to populate its assignee dropdown so a freshly-created profile
+    appears in the picker before it's been given any task.
+    """
+    conn = _conn()
+    try:
+        return {"assignees": kanban_db.known_assignees(conn)}
+    finally:
+        conn.close()
+
+
+# ---------------------------------------------------------------------------
+# Worker log (read-only; file written by _default_spawn)
+# ---------------------------------------------------------------------------
+
+@router.get("/tasks/{task_id}/log")
+def get_task_log(task_id: str, tail: Optional[int] = Query(None, ge=1, le=2_000_000)):
+    """Return the worker's stdout/stderr log.
+
+    ``tail`` caps the response size (bytes) so the dashboard drawer
+    doesn't paginate megabytes into the browser. Returns 404 if the task
+    has never spawned. The on-disk log is rotated at 2 MiB per
+    ``_rotate_worker_log`` — a single ``.log.1`` is kept, no further
+    generations, so disk usage per task is bounded at ~4 MiB.
+    """
+    conn = _conn()
+    try:
+        task = kanban_db.get_task(conn, task_id)
+    finally:
+        conn.close()
+    if task is None:
+        raise HTTPException(status_code=404, detail=f"task {task_id} not found")
+    content = kanban_db.read_worker_log(task_id, tail_bytes=tail)
+    log_path = kanban_db.worker_log_path(task_id)
+    size = log_path.stat().st_size if log_path.exists() else 0
+    return {
+        "task_id": task_id,
+        "path": str(log_path),
+        "exists": content is not None,
+        "size_bytes": size,
+        "content": content or "",
+        # Truncated when the on-disk file was larger than the tail cap.
+        "truncated": bool(tail and size > tail),
+    }
+
+
+# ---------------------------------------------------------------------------
+# Dispatch nudge (optional quick-path so the UI doesn't wait 60 s)
+# ---------------------------------------------------------------------------
+
+@router.post("/dispatch")
+def dispatch(dry_run: bool = Query(False), max_n: int = Query(8, alias="max")):
+    conn = _conn()
+    try:
+        result = kanban_db.dispatch_once(
+            conn, dry_run=dry_run, max_spawn=max_n,
+        )
+        # DispatchResult is a dataclass.
+        try:
+            return asdict(result)
+        except TypeError:
+            return {"result": str(result)}
+    finally:
+        conn.close()
+
+
+# ---------------------------------------------------------------------------
+# WebSocket: /events?since=<event_id>
+# ---------------------------------------------------------------------------
+
+# Poll interval for the event tail loop. SQLite WAL + 300 ms polling is
+# the simplest and most robust approach; it adds a fraction of a percent
+# of CPU and has no shared state to synchronize across workers.
+_EVENT_POLL_SECONDS = 0.3
+
+
+@router.websocket("/events")
+async def stream_events(ws: WebSocket):
+    # Enforce the dashboard session token as a query param — browsers can't
+    # set Authorization on a WS upgrade. This matches how the PTY bridge
+    # authenticates in hermes_cli/web_server.py.
+    token = ws.query_params.get("token")
+    if not _check_ws_token(token):
+        await ws.close(code=http_status.WS_1008_POLICY_VIOLATION)
+        return
+    await ws.accept()
+    try:
+        since_raw = ws.query_params.get("since", "0")
+        try:
+            cursor = int(since_raw)
+        except ValueError:
+            cursor = 0
+
+        def _fetch_new(cursor_val: int) -> tuple[int, list[dict]]:
+            conn = kanban_db.connect()
+            try:
+                rows = conn.execute(
+                    "SELECT id, task_id, run_id, kind, payload, created_at "
+                    "FROM task_events WHERE id > ? ORDER BY id ASC LIMIT 200",
+                    (cursor_val,),
+                ).fetchall()
+                out: list[dict] = []
+                new_cursor = cursor_val
+                for r in rows:
+                    try:
+                        payload = json.loads(r["payload"]) if r["payload"] else None
+                    except Exception:
+                        payload = None
+                    out.append({
+                        "id": r["id"],
+                        "task_id": r["task_id"],
+                        "run_id": r["run_id"],
+                        "kind": r["kind"],
+                        "payload": payload,
+                        "created_at": r["created_at"],
+                    })
+                    new_cursor = r["id"]
+                return new_cursor, out
+            finally:
+                conn.close()
+
+        while True:
+            cursor, events = await asyncio.to_thread(_fetch_new, cursor)
+            if events:
+                await ws.send_json({"events": events, "cursor": cursor})
+            await asyncio.sleep(_EVENT_POLL_SECONDS)
+    except WebSocketDisconnect:
+        return
+    except Exception as exc:  # defensive: never crash the dashboard worker
+        log.warning("Kanban event stream error: %s", exc)
+        try:
+            await ws.close()
+        except Exception:
+            pass

plugins/kanban/systemd/hermes-kanban-dispatcher.service

@@ -0,0 +1,17 @@
+[Unit]
+Description=Hermes Kanban dispatcher (hermes kanban daemon)
+Documentation=https://hermes-agent.nousresearch.com/docs/user-guide/features/kanban
+After=network.target
+
+[Service]
+Type=simple
+ExecStart=/usr/bin/env hermes kanban daemon --interval 60 --pidfile %t/hermes-kanban-dispatcher.pid
+Restart=on-failure
+RestartSec=5
+# Log to the journal via stdout/stderr; the dispatcher also writes per-task
+# worker output to $HERMES_HOME/kanban/logs/<task>.log.
+StandardOutput=journal
+StandardError=journal
+
+[Install]
+WantedBy=default.target

plugins/memory/openviking/__init__.py

@@ -43,7 +43,7 @@ _TIMEOUT = 30.0
 # ---------------------------------------------------------------------------
 # Process-level atexit safety net — ensures pending sessions are committed
 # even if shutdown_memory_provider is never called (e.g. gateway crash,
-# SIGKILL, or exception in _async_flush_memories preventing shutdown).
+# SIGKILL, or exception in the session expiry watcher preventing shutdown).
 # ---------------------------------------------------------------------------
 _last_active_provider: Optional["OpenVikingMemoryProvider"] = None
 

pyproject.toml

@@ -78,6 +78,15 @@ termux = [
 ]
 dingtalk = ["dingtalk-stream>=0.20,<1", "alibabacloud-dingtalk>=2.0.0", "qrcode>=7.0,<8"]
 feishu = ["lark-oapi>=1.5.3,<2", "qrcode>=7.0,<8"]
+google = [
+  # Required by the google-workspace skill (Gmail, Calendar, Drive, Contacts,
+  # Sheets, Docs).  Declared here so packagers (Nix, Homebrew) ship them with
+  # the [all] extra and users don't hit runtime `pip install` paths that fail
+  # in environments without pip (e.g. Nix-managed Python).
+  "google-api-python-client>=2.100,<3",
+  "google-auth-oauthlib>=1.0,<2",
+  "google-auth-httplib2>=0.2,<1",
+]
 # `hermes dashboard` (localhost SPA + API).  Not in core to keep the default install lean.
 web = ["fastapi>=0.104.0,<1", "uvicorn[standard]>=0.24.0,<1"]
 rl = [
@@ -110,6 +119,7 @@ all = [
   "hermes-agent[voice]",
   "hermes-agent[dingtalk]",
   "hermes-agent[feishu]",
+  "hermes-agent[google]",
   "hermes-agent[mistral]",
   "hermes-agent[bedrock]",
   "hermes-agent[web]",

run_agent.py

@@ -40,6 +40,7 @@ from types import SimpleNamespace
 import urllib.request
 import uuid
 from typing import List, Dict, Any, Optional
+from urllib.parse import urlparse, parse_qs, urlunparse
 from openai import OpenAI
 import fire
 from datetime import datetime
@@ -85,6 +86,7 @@ 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,
+    KANBAN_GUIDANCE,
     build_nous_subscription_prompt,
 )
 from agent.model_metadata import (
@@ -891,7 +893,6 @@ class AIAgent:
         checkpoints_enabled: bool = False,
         checkpoint_max_snapshots: int = 50,
         pass_session_id: bool = False,
-        persist_session: bool = True,
     ):
         """
         Initialize the AI Agent.
@@ -963,7 +964,6 @@ class AIAgent:
         self.background_review_callback = None  # Optional sync callback for gateway delivery
         self.skip_context_files = skip_context_files
         self.pass_session_id = pass_session_id
-        self.persist_session = persist_session
         self._credential_pool = credential_pool
         self.log_prefix_chars = log_prefix_chars
         self.log_prefix = f"{log_prefix} " if log_prefix else ""
@@ -1033,12 +1033,16 @@ class AIAgent:
         # surface.
         # When api_mode was explicitly provided, respect it — the user
         # knows what their endpoint supports (#10473).
+        # Exception: Azure OpenAI serves gpt-5.x on /chat/completions and
+        # does NOT support the Responses API — skip the upgrade for Azure
+        # (openai.azure.com), even though it looks OpenAI-compatible.
         if (
             api_mode is None
             and self.api_mode == "chat_completions"
             and self.provider != "copilot-acp"
             and not str(self.base_url or "").lower().startswith("acp://copilot")
             and not str(self.base_url or "").lower().startswith("acp+tcp://")
+            and not self._is_azure_openai_url()
             and (
                 self._is_direct_openai_url()
                 or self._provider_model_requires_responses_api(
@@ -1314,7 +1318,22 @@ class AIAgent:
             if api_key and base_url:
                 # Explicit credentials from CLI/gateway — construct directly.
                 # The runtime provider resolver already handled auth for us.
-                client_kwargs = {"api_key": api_key, "base_url": base_url}
+                # Extract query params (e.g. Azure api-version) from base_url
+                # and pass via default_query to prevent loss during SDK URL
+                # joining (httpx drops query string when joining paths).
+                _parsed_url = urlparse(base_url)
+                if _parsed_url.query:
+                    _clean_url = urlunparse(_parsed_url._replace(query=""))
+                    _query_params = {
+                        k: v[0] for k, v in parse_qs(_parsed_url.query).items()
+                    }
+                    client_kwargs = {
+                        "api_key": api_key,
+                        "base_url": _clean_url,
+                        "default_query": _query_params,
+                    }
+                else:
+                    client_kwargs = {"api_key": api_key, "base_url": base_url}
                 if _provider_timeout is not None:
                     client_kwargs["timeout"] = _provider_timeout
                 if self.provider == "copilot-acp":
@@ -1578,7 +1597,6 @@ class AIAgent:
         self._memory_enabled = False
         self._user_profile_enabled = False
         self._memory_nudge_interval = 10
-        self._memory_flush_min_turns = 6
         self._turns_since_memory = 0
         self._iters_since_skill = 0
         if not skip_memory:
@@ -1587,7 +1605,6 @@ class AIAgent:
                 self._memory_enabled = mem_config.get("memory_enabled", False)
                 self._user_profile_enabled = mem_config.get("user_profile_enabled", False)
                 self._memory_nudge_interval = int(mem_config.get("nudge_interval", 10))
-                self._memory_flush_min_turns = int(mem_config.get("flush_min_turns", 6))
                 if self._memory_enabled or self._user_profile_enabled:
                     from tools.memory_tool import MemoryStore
                     self._memory_store = MemoryStore(
@@ -1767,43 +1784,64 @@ class AIAgent:
         # Store for reuse in switch_model (so config override persists across model switches)
         self._config_context_length = _config_context_length
 
+        # Resolve custom_providers list once for reuse below (startup
+        # context-length override and plugin context-engine init).
+        try:
+            from hermes_cli.config import get_compatible_custom_providers
+            _custom_providers = get_compatible_custom_providers(_agent_cfg)
+        except Exception:
+            _custom_providers = _agent_cfg.get("custom_providers")
+            if not isinstance(_custom_providers, list):
+                _custom_providers = []
+
         # Check custom_providers per-model context_length
-        if _config_context_length is None:
+        if _config_context_length is None and _custom_providers:
             try:
-                from hermes_cli.config import get_compatible_custom_providers
-                _custom_providers = get_compatible_custom_providers(_agent_cfg)
+                from hermes_cli.config import get_custom_provider_context_length
+                _cp_ctx_resolved = get_custom_provider_context_length(
+                    model=self.model,
+                    base_url=self.base_url,
+                    custom_providers=_custom_providers,
+                )
+                if _cp_ctx_resolved:
+                    _config_context_length = int(_cp_ctx_resolved)
             except Exception:
-                _custom_providers = _agent_cfg.get("custom_providers")
-                if not isinstance(_custom_providers, list):
-                    _custom_providers = []
-            for _cp_entry in _custom_providers:
-                if not isinstance(_cp_entry, dict):
-                    continue
-                _cp_url = (_cp_entry.get("base_url") or "").rstrip("/")
-                if _cp_url and _cp_url == self.base_url.rstrip("/"):
-                    _cp_models = _cp_entry.get("models", {})
-                    if isinstance(_cp_models, dict):
-                        _cp_model_cfg = _cp_models.get(self.model, {})
-                        if isinstance(_cp_model_cfg, dict):
-                            _cp_ctx = _cp_model_cfg.get("context_length")
-                            if _cp_ctx is not None:
-                                try:
-                                    _config_context_length = int(_cp_ctx)
-                                except (TypeError, ValueError):
-                                    logger.warning(
-                                        "Invalid context_length for model %r in "
-                                        "custom_providers: %r — must be a plain "
-                                        "integer (e.g. 256000, not '256K'). "
-                                        "Falling back to auto-detection.",
-                                        self.model, _cp_ctx,
-                                    )
-                                    print(
-                                        f"\n⚠ Invalid context_length for model {self.model!r} in custom_providers: {_cp_ctx!r}\n"
-                                        f"  Must be a plain integer (e.g. 256000, not '256K').\n"
-                                        f"  Falling back to auto-detected context window.\n",
-                                        file=sys.stderr,
-                                    )
-                    break
+                _cp_ctx_resolved = None
+
+            # Surface a clear warning if the user set a context_length but it
+            # wasn't a valid positive int — the helper silently skips those.
+            if _config_context_length is None:
+                _target = self.base_url.rstrip("/") if self.base_url else ""
+                for _cp_entry in _custom_providers:
+                    if not isinstance(_cp_entry, dict):
+                        continue
+                    _cp_url = (_cp_entry.get("base_url") or "").rstrip("/")
+                    if _target and _cp_url == _target:
+                        _cp_models = _cp_entry.get("models", {})
+                        if isinstance(_cp_models, dict):
+                            _cp_model_cfg = _cp_models.get(self.model, {})
+                            if isinstance(_cp_model_cfg, dict):
+                                _cp_ctx = _cp_model_cfg.get("context_length")
+                                if _cp_ctx is not None:
+                                    try:
+                                        _parsed = int(_cp_ctx)
+                                        if _parsed <= 0:
+                                            raise ValueError
+                                    except (TypeError, ValueError):
+                                        logger.warning(
+                                            "Invalid context_length for model %r in "
+                                            "custom_providers: %r — must be a positive "
+                                            "integer (e.g. 256000, not '256K'). "
+                                            "Falling back to auto-detection.",
+                                            self.model, _cp_ctx,
+                                        )
+                                        print(
+                                            f"\n⚠ Invalid context_length for model {self.model!r} in custom_providers: {_cp_ctx!r}\n"
+                                            f"  Must be a positive integer (e.g. 256000, not '256K').\n"
+                                            f"  Falling back to auto-detected context window.\n",
+                                            file=sys.stderr,
+                                        )
+                        break
         
         # Select context engine: config-driven (like memory providers).
         # 1. Check config.yaml context.engine setting
@@ -1853,6 +1891,7 @@ class AIAgent:
                 api_key=getattr(self, "api_key", ""),
                 config_context_length=_config_context_length,
                 provider=self.provider,
+                custom_providers=_custom_providers,
             )
             self.context_compressor.update_model(
                 model=self.model,
@@ -2143,12 +2182,23 @@ class AIAgent:
         # ── Update context compressor ──
         if hasattr(self, "context_compressor") and self.context_compressor:
             from agent.model_metadata import get_model_context_length
+            # Re-read custom_providers from live config so per-model
+            # context_length overrides are honored when switching to a
+            # custom provider mid-session (closes #15779).
+            _sm_custom_providers = None
+            try:
+                from hermes_cli.config import load_config, get_compatible_custom_providers
+                _sm_cfg = load_config()
+                _sm_custom_providers = get_compatible_custom_providers(_sm_cfg)
+            except Exception:
+                _sm_custom_providers = None
             new_context_length = get_model_context_length(
                 self.model,
                 base_url=self.base_url,
                 api_key=self.api_key,
                 provider=self.provider,
                 config_context_length=getattr(self, "_config_context_length", None),
+                custom_providers=_sm_custom_providers,
             )
             self.context_compressor.update_model(
                 model=self.model,
@@ -2399,6 +2449,7 @@ class AIAgent:
                 base_url=aux_base_url,
                 api_key=aux_api_key,
                 config_context_length=getattr(self, "_aux_compression_context_length_config", None),
+                provider=getattr(self, "provider", ""),
             )
 
             # Hard floor: the auxiliary compression model must have at least
@@ -2425,6 +2476,11 @@ class AIAgent:
                 # compression actually works this session.  The hard floor
                 # above guarantees aux_context >= MINIMUM_CONTEXT_LENGTH,
                 # so the new threshold is always >= 64K.
+                #
+                # The compression summariser sends a single user-role
+                # prompt (no system prompt, no tools) to the aux model, so
+                # new_threshold == aux_context is safe: the request is
+                # the raw messages plus a small summarisation instruction.
                 old_threshold = threshold
                 new_threshold = aux_context
                 self.context_compressor.threshold_tokens = new_threshold
@@ -2500,6 +2556,22 @@ class AIAgent:
             )
         return hostname == "api.openai.com"
 
+    def _is_azure_openai_url(self, base_url: str = None) -> bool:
+        """Return True when a base URL targets Azure OpenAI.
+
+        Azure OpenAI exposes an OpenAI-compatible endpoint at
+        ``{resource}.openai.azure.com/openai/v1`` that accepts the
+        standard ``openai`` Python client.  Unlike api.openai.com it
+        does NOT support the Responses API — gpt-5.x models are served
+        on the regular ``/chat/completions`` path — so routing decisions
+        must treat Azure separately from direct OpenAI.
+        """
+        if base_url is not None:
+            url = str(base_url).lower()
+        else:
+            url = getattr(self, "_base_url_lower", "") or ""
+        return "openai.azure.com" in url
+
     def _resolved_api_call_timeout(self) -> float:
         """Resolve the effective per-call request timeout in seconds.
 
@@ -2671,12 +2743,14 @@ class AIAgent:
 
     def _max_tokens_param(self, value: int) -> dict:
         """Return the correct max tokens kwarg for the current provider.
-        
+
         OpenAI's newer models (gpt-4o, o-series, gpt-5+) require
-        'max_completion_tokens'. OpenRouter, local models, and older
+        'max_completion_tokens'. Azure OpenAI also requires
+        'max_completion_tokens' for gpt-5.x models served via the
+        OpenAI-compatible endpoint. OpenRouter, local models, and older
         OpenAI models use 'max_tokens'.
         """
-        if self._is_direct_openai_url():
+        if self._is_direct_openai_url() or self._is_azure_openai_url():
             return {"max_completion_tokens": value}
         return {"max_tokens": value}
 
@@ -3034,13 +3108,28 @@ class AIAgent:
     )
 
     _SKILL_REVIEW_PROMPT = (
-        "Review the conversation above and consider saving or updating a skill if appropriate.\n\n"
-        "Focus on: was a non-trivial approach used to complete a task that required trial "
-        "and error, or changing course due to experiential findings along the way, or did "
-        "the user expect or desire a different method or outcome?\n\n"
-        "If a relevant skill already exists, update it with what you learned. "
-        "Otherwise, create a new skill if the approach is reusable.\n"
-        "If nothing is worth saving, just say 'Nothing to save.' and stop."
+        "Review the conversation above and consider whether a skill should be saved or updated.\n\n"
+        "Work in this order — do not skip steps:\n\n"
+        "1. SURVEY the existing skill landscape first. Call skills_list to see what you "
+        "have. If anything looks potentially relevant, skill_view it before deciding. "
+        "You are looking for the CLASS of task that just happened, not the exact task. "
+        "Example: a successful Tauri build is in the class \"desktop app build "
+        "troubleshooting\", not \"fix my specific Tauri error today\".\n\n"
+        "2. THINK CLASS-FIRST. What general pattern of task did the user just complete? "
+        "What conditions will trigger this pattern again? Describe the class in one "
+        "sentence before looking at what to save.\n\n"
+        "3. PREFER GENERALIZING AN EXISTING SKILL over creating a new one. If a skill "
+        "already covers the class — even partially — update it (skill_manage patch) "
+        "with the new insight. Broaden its \"when to use\" trigger if needed.\n\n"
+        "4. ONLY CREATE A NEW SKILL when no existing skill reasonably covers the class. "
+        "When you create one, name and scope it at the class level "
+        "(\"react-i18n-setup\", not \"add-i18n-to-my-dashboard-app\"). The trigger "
+        "section must describe the class of situations, not this one session.\n\n"
+        "5. If you notice two existing skills that overlap, note it in your response "
+        "so a future review can consolidate them. Do not consolidate now unless the "
+        "overlap is obvious and low-risk.\n\n"
+        "Only act when something is genuinely worth saving. "
+        "If nothing stands out, just say 'Nothing to save.' and stop."
     )
 
     _COMBINED_REVIEW_PROMPT = (
@@ -3050,9 +3139,16 @@ class AIAgent:
         "about how you should behave, their work style, or ways they want you to operate? "
         "If so, save using the memory tool.\n\n"
         "**Skills**: Was a non-trivial approach used to complete a task that required trial "
-        "and error, or changing course due to experiential findings along the way, or did "
-        "the user expect or desire a different method or outcome? If a relevant skill "
-        "already exists, update it. Otherwise, create a new one if the approach is reusable.\n\n"
+        "and error, changing course due to experiential findings, or a different method "
+        "or outcome than the user expected? If so, work in this order:\n"
+        "  a. SURVEY existing skills first (skills_list, then skill_view on candidates).\n"
+        "  b. Identify the CLASS of task, not the specific task "
+        "(\"desktop app build troubleshooting\", not \"fix my Tauri error\").\n"
+        "  c. PREFER UPDATING/GENERALIZING an existing skill that covers the class.\n"
+        "  d. ONLY CREATE A NEW SKILL if no existing one covers the class. Scope at "
+        "the class level, not this one session.\n"
+        "  e. If you notice overlapping skills during the survey, note it so a future "
+        "review can consolidate them.\n\n"
         "Only act if there's something genuinely worth saving. "
         "If nothing stands out, just say 'Nothing to save.' and stop."
     )
@@ -3150,12 +3246,25 @@ class AIAgent:
                 with open(os.devnull, "w") as _devnull, \
                      contextlib.redirect_stdout(_devnull), \
                      contextlib.redirect_stderr(_devnull):
+                    # Inherit the parent agent's live runtime (provider, model,
+                    # base_url, api_key, api_mode) so the fork uses the exact
+                    # same credentials the main turn is using.  Without this,
+                    # AIAgent.__init__ re-runs auto-resolution from env vars,
+                    # which fails for OAuth-only providers, session-scoped
+                    # creds, or credential-pool setups where the resolver can't
+                    # reconstruct auth from scratch -- producing the spurious
+                    # "No LLM provider configured" warning at end of turn.
+                    _parent_runtime = self._current_main_runtime()
                     review_agent = AIAgent(
                         model=self.model,
                         max_iterations=8,
                         quiet_mode=True,
                         platform=self.platform,
                         provider=self.provider,
+                        api_mode=_parent_runtime.get("api_mode") or None,
+                        base_url=_parent_runtime.get("base_url") or None,
+                        api_key=_parent_runtime.get("api_key") or None,
+                        credential_pool=getattr(self, "_credential_pool", None),
                         parent_session_id=self.session_id,
                     )
                     review_agent._memory_write_origin = "background_review"
@@ -3256,10 +3365,7 @@ class AIAgent:
         """Save session state to both JSON log and SQLite on any exit path.
 
         Ensures conversations are never lost, even on errors or early returns.
-        Skipped when ``persist_session=False`` (ephemeral helper flows).
         """
-        if not self.persist_session:
-            return
         self._apply_persist_user_message_override(messages)
         self._session_messages = messages
         self._save_session_log(messages)
@@ -3309,6 +3415,7 @@ class AIAgent:
                     reasoning_content=msg.get("reasoning_content") if role == "assistant" else None,
                     reasoning_details=msg.get("reasoning_details") if role == "assistant" else None,
                     codex_reasoning_items=msg.get("codex_reasoning_items") if role == "assistant" else None,
+                    codex_message_items=msg.get("codex_message_items") if role == "assistant" else None,
                 )
             self._last_flushed_db_idx = len(messages)
         except Exception as e:
@@ -4391,6 +4498,12 @@ class AIAgent:
             tool_guidance.append(SESSION_SEARCH_GUIDANCE)
         if "skill_manage" in self.valid_tool_names:
             tool_guidance.append(SKILLS_GUIDANCE)
+        # Kanban worker/orchestrator lifecycle — only present when the
+        # dispatcher spawned this process (kanban_show check_fn gates on
+        # HERMES_KANBAN_TASK env var). Normal chat sessions never see
+        # this block.
+        if "kanban_show" in self.valid_tool_names:
+            tool_guidance.append(KANBAN_GUIDANCE)
         if tool_guidance:
             prompt_parts.append(" ".join(tool_guidance))
 
@@ -5137,6 +5250,8 @@ class AIAgent:
         # response.incomplete instead of response.completed).
         self._codex_streamed_text_parts: list = []
         for attempt in range(max_stream_retries + 1):
+            if self._interrupt_requested:
+                raise InterruptedError("Agent interrupted before Codex stream retry")
             collected_output_items: list = []
             try:
                 with active_client.responses.stream(**api_kwargs) as stream:
@@ -5431,6 +5546,11 @@ class AIAgent:
         # Other anthropic_messages providers (MiniMax, Alibaba, etc.) use their own keys.
         if self.provider != "anthropic":
             return False
+        # Azure endpoints use static API keys — OAuth token rotation doesn't apply.
+        # Refreshing would pick up ~/.claude/.credentials.json OAuth token and break auth.
+        _base = getattr(self, "_anthropic_base_url", "") or ""
+        if "azure.com" in _base:
+            return False
 
         try:
             from agent.anthropic_adapter import resolve_anthropic_token, build_anthropic_client
@@ -6306,6 +6426,14 @@ class AIAgent:
 
             try:
                 for _stream_attempt in range(_max_stream_retries + 1):
+                    # Check for interrupt before each retry attempt.  Without
+                    # this, /stop closes the HTTP connection (outer poll loop),
+                    # but the retry loop opens a FRESH connection — negating the
+                    # interrupt entirely.  On slow providers (ollama-cloud) each
+                    # retry can block for the full stream-read timeout (120s+),
+                    # causing multi-minute delays between /stop and response.
+                    if self._interrupt_requested:
+                        raise InterruptedError("Agent interrupted before stream retry")
                     try:
                         if self.api_mode == "anthropic_messages":
                             self._try_refresh_anthropic_client_credentials()
@@ -6779,10 +6907,15 @@ class AIAgent:
             # Determine api_mode from provider / base URL / model
             fb_api_mode = "chat_completions"
             fb_base_url = str(fb_client.base_url)
+            _fb_is_azure = self._is_azure_openai_url(fb_base_url)
             if fb_provider == "openai-codex":
                 fb_api_mode = "codex_responses"
             elif fb_provider == "anthropic" or fb_base_url.rstrip("/").lower().endswith("/anthropic"):
                 fb_api_mode = "anthropic_messages"
+            elif _fb_is_azure:
+                # Azure OpenAI serves gpt-5.x on /chat/completions — does NOT
+                # support the Responses API. Stay on chat_completions.
+                fb_api_mode = "chat_completions"
             elif self._is_direct_openai_url(fb_base_url):
                 fb_api_mode = "codex_responses"
             elif self._provider_model_requires_responses_api(
@@ -7655,6 +7788,13 @@ class AIAgent:
         if codex_items:
             msg["codex_reasoning_items"] = codex_items
 
+        # Codex Responses API: preserve exact assistant message items (with
+        # id/phase) so follow-up turns can replay structured items instead of
+        # flattening to plain text. This is required for prefix cache hits.
+        codex_message_items = getattr(assistant_message, "codex_message_items", None)
+        if codex_message_items:
+            msg["codex_message_items"] = codex_message_items
+
         if assistant_message.tool_calls:
             tool_calls = []
             for tool_call in assistant_message.tool_calls:
@@ -7740,25 +7880,53 @@ class AIAgent:
         if source_msg.get("role") != "assistant":
             return
 
-        explicit_reasoning = source_msg.get("reasoning_content")
-        if isinstance(explicit_reasoning, str):
-            api_msg["reasoning_content"] = explicit_reasoning
+        # 1. Explicit reasoning_content already set — preserve it verbatim
+        # (includes DeepSeek/Kimi's own empty-string placeholder written at
+        # creation time, and any valid reasoning content from the same provider).
+        existing = source_msg.get("reasoning_content")
+        if isinstance(existing, str):
+            api_msg["reasoning_content"] = existing
             return
 
+        # 2. Healthy session: promote 'reasoning' field to 'reasoning_content'
+        # for providers that use the internal 'reasoning' key.
+        # This must happen BEFORE the DeepSeek/Kimi tool-call check so that
+        # genuine reasoning content is not overwritten by the empty-string
+        # fallback (#15812 regression in PR #15478).
         normalized_reasoning = source_msg.get("reasoning")
         if isinstance(normalized_reasoning, str) and normalized_reasoning:
             api_msg["reasoning_content"] = normalized_reasoning
             return
 
-        # Providers that require an echoed reasoning_content on every
-        # assistant tool-call turn. Detection logic lives in the per-provider
-        # helpers so both the creation path (_build_assistant_message) and
-        # this replay path stay in sync.
-        if source_msg.get("tool_calls") and (
+        # 3. DeepSeek / Kimi thinking mode: tool-call turns that lack
+        # reasoning_content are "poisoned history" — a prior provider (MiniMax,
+        # etc.) left them empty. DeepSeek returns HTTP 400 if reasoning_content
+        # is absent on replay; inject "" to satisfy the provider's requirement
+        # without forwarding any cross-provider reasoning content.
+        needs_empty_reasoning = (
+            source_msg.get("tool_calls")
+            and (
+                self._needs_kimi_tool_reasoning()
+                or self._needs_deepseek_tool_reasoning()
+            )
+        )
+        if needs_empty_reasoning:
+            api_msg["reasoning_content"] = ""
+            return
+
+        # 4. DeepSeek / Kimi thinking mode: all assistant messages need
+        # reasoning_content. Inject "" to satisfy the provider's requirement
+        # when no explicit reasoning content is present.
+        if (
             self._needs_kimi_tool_reasoning()
             or self._needs_deepseek_tool_reasoning()
         ):
             api_msg["reasoning_content"] = ""
+            return
+
+        # 5. reasoning_content was present but not a string (e.g. None after
+        # context compaction).  Don't pass null to the API.
+        api_msg.pop("reasoning_content", None)
 
     @staticmethod
     def _sanitize_tool_calls_for_strict_api(api_msg: dict) -> dict:
@@ -7910,251 +8078,6 @@ class AIAgent:
         """
         return self.api_mode != "codex_responses"
 
-    def flush_memories(self, messages: list = None, min_turns: int = None):
-        """Give the model one turn to persist memories before context is lost.
-
-        Called before compression, session reset, or CLI exit. Injects a flush
-        message, makes one API call, executes any memory tool calls, then
-        strips all flush artifacts from the message list.
-
-        Args:
-            messages: The current conversation messages. If None, uses
-                      self._session_messages (last run_conversation state).
-            min_turns: Minimum user turns required to trigger the flush.
-                       None = use config value (flush_min_turns).
-                       0 = always flush (used for compression).
-        """
-        if self._memory_flush_min_turns == 0 and min_turns is None:
-            return
-        if "memory" not in self.valid_tool_names or not self._memory_store:
-            return
-        effective_min = min_turns if min_turns is not None else self._memory_flush_min_turns
-        if self._user_turn_count < effective_min:
-            return
-
-        if messages is None:
-            messages = getattr(self, '_session_messages', None)
-        if not messages or len(messages) < 3:
-            return
-
-        flush_content = (
-            "[System: The session is being compressed. "
-            "Save anything worth remembering — prioritize user preferences, "
-            "corrections, and recurring patterns over task-specific details.]"
-        )
-        _sentinel = f"__flush_{id(self)}_{time.monotonic()}"
-        flush_msg = {"role": "user", "content": flush_content, "_flush_sentinel": _sentinel}
-        messages.append(flush_msg)
-
-        try:
-            # Build API messages for the flush call
-            _needs_sanitize = self._should_sanitize_tool_calls()
-            api_messages = []
-            for msg in messages:
-                api_msg = msg.copy()
-                self._copy_reasoning_content_for_api(msg, api_msg)
-                api_msg.pop("reasoning", None)
-                api_msg.pop("finish_reason", None)
-                api_msg.pop("_flush_sentinel", None)
-                api_msg.pop("_thinking_prefill", None)
-                if _needs_sanitize:
-                    self._sanitize_tool_calls_for_strict_api(api_msg)
-                api_messages.append(api_msg)
-
-            if self._cached_system_prompt:
-                api_messages = [{"role": "system", "content": self._cached_system_prompt}] + api_messages
-
-            # Make one API call with only the memory tool available
-            memory_tool_def = None
-            for t in (self.tools or []):
-                if t.get("function", {}).get("name") == "memory":
-                    memory_tool_def = t
-                    break
-
-            if not memory_tool_def:
-                messages.pop()  # remove flush msg
-                return
-
-            # Use auxiliary client for the flush call when available --
-            # it's cheaper and avoids Codex Responses API incompatibility.
-            from agent.auxiliary_client import (
-                call_llm as _call_llm,
-                _fixed_temperature_for_model,
-                OMIT_TEMPERATURE,
-            )
-            _aux_available = True
-            # Kimi models manage temperature server-side — omit it entirely.
-            # Other models with a fixed contract get that value; everyone else
-            # gets the historical 0.3 default.
-            _fixed_temp = _fixed_temperature_for_model(self.model, self.base_url)
-            _omit_temperature = _fixed_temp is OMIT_TEMPERATURE
-            if _omit_temperature:
-                _flush_temperature = None
-            elif _fixed_temp is not None:
-                _flush_temperature = _fixed_temp
-            else:
-                _flush_temperature = 0.3
-            aux_error = None
-            try:
-                response = _call_llm(
-                    task="flush_memories",
-                    messages=api_messages,
-                    tools=[memory_tool_def],
-                    temperature=_flush_temperature,
-                    max_tokens=5120,
-                    # timeout resolved from auxiliary.flush_memories.timeout config
-                )
-            except Exception as e:
-                aux_error = e
-                _aux_available = False
-                response = None
-
-            if not _aux_available and self.api_mode == "codex_responses":
-                # No auxiliary client -- use the Codex Responses path directly
-                codex_kwargs = self._build_api_kwargs(api_messages)
-                _ct_flush = self._get_transport()
-                if _ct_flush is not None:
-                    codex_kwargs["tools"] = _ct_flush.convert_tools([memory_tool_def])
-                elif not codex_kwargs.get("tools"):
-                    codex_kwargs["tools"] = [memory_tool_def]
-                if _flush_temperature is not None:
-                    codex_kwargs["temperature"] = _flush_temperature
-                else:
-                    codex_kwargs.pop("temperature", None)
-                if "max_output_tokens" in codex_kwargs:
-                    codex_kwargs["max_output_tokens"] = 5120
-                response = self._run_codex_stream(codex_kwargs)
-            elif not _aux_available and self.api_mode == "anthropic_messages":
-                # Native Anthropic — use the transport for kwargs
-                _tflush = self._get_transport()
-                ant_kwargs = _tflush.build_kwargs(
-                    model=self.model, messages=api_messages,
-                    tools=[memory_tool_def], max_tokens=5120,
-                    reasoning_config=None,
-                    preserve_dots=self._anthropic_preserve_dots(),
-                )
-                response = self._anthropic_messages_create(ant_kwargs)
-            elif not _aux_available:
-                api_kwargs = {
-                    "model": self.model,
-                    "messages": api_messages,
-                    "tools": [memory_tool_def],
-                    **self._max_tokens_param(5120),
-                }
-                if _flush_temperature is not None:
-                    api_kwargs["temperature"] = _flush_temperature
-                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")
-                )
-
-            if aux_error is not None:
-                logger.warning("Auxiliary memory flush failed; used fallback path: %s", aux_error)
-                self._emit_auxiliary_failure("memory flush", aux_error)
-
-            def _openai_tool_calls(resp):
-                if resp is not None and hasattr(resp, "choices") and resp.choices:
-                    msg = getattr(resp.choices[0], "message", None)
-                    calls = getattr(msg, "tool_calls", None)
-                    if calls:
-                        return calls
-                return []
-
-            def _codex_output_tool_calls(resp):
-                calls = []
-                for item in getattr(resp, "output", []) or []:
-                    if getattr(item, "type", None) == "function_call":
-                        calls.append(SimpleNamespace(
-                            id=getattr(item, "call_id", None),
-                            type="function",
-                            function=SimpleNamespace(
-                                name=getattr(item, "name", ""),
-                                arguments=getattr(item, "arguments", "{}"),
-                            ),
-                        ))
-                return calls
-
-            # Extract tool calls from the response, handling all API formats
-            tool_calls = []
-            if self.api_mode == "codex_responses" and not _aux_available:
-                _ct_flush = self._get_transport()
-                _cnr_flush = _ct_flush.normalize_response(response) if _ct_flush is not None else None
-                if _cnr_flush and _cnr_flush.tool_calls:
-                    tool_calls = [
-                        SimpleNamespace(
-                            id=tc.id, type="function",
-                            function=SimpleNamespace(name=tc.name, arguments=tc.arguments),
-                        ) for tc in _cnr_flush.tool_calls
-                    ]
-                else:
-                    tool_calls = _codex_output_tool_calls(response)
-            elif self.api_mode == "anthropic_messages" and not _aux_available:
-                _tfn = self._get_transport()
-                _flush_result = _tfn.normalize_response(response, strip_tool_prefix=self._is_anthropic_oauth)
-                if _flush_result and _flush_result.tool_calls:
-                    tool_calls = [
-                        SimpleNamespace(
-                            id=tc.id, type="function",
-                            function=SimpleNamespace(name=tc.name, arguments=tc.arguments),
-                        ) for tc in _flush_result.tool_calls
-                    ]
-            elif self.api_mode in ("chat_completions", "bedrock_converse"):
-                # chat_completions / bedrock — normalize through transport
-                _tfn = self._get_transport()
-                _flush_result = _tfn.normalize_response(response) if _tfn is not None else None
-                if _flush_result and _flush_result.tool_calls:
-                    tool_calls = _flush_result.tool_calls
-                else:
-                    tool_calls = _openai_tool_calls(response)
-            elif _aux_available and hasattr(response, "choices") and response.choices:
-                # Auxiliary client returned OpenAI-shaped response while main
-                # api_mode is codex/anthropic — extract tool_calls from .choices
-                tool_calls = _openai_tool_calls(response)
-
-            for tc in tool_calls:
-                if tc.function.name == "memory":
-                    try:
-                        args = json.loads(tc.function.arguments)
-                        flush_target = args.get("target", "memory")
-                        from tools.memory_tool import memory_tool as _memory_tool
-                        _memory_tool(
-                            action=args.get("action"),
-                            target=flush_target,
-                            content=args.get("content"),
-                            old_text=args.get("old_text"),
-                            store=self._memory_store,
-                        )
-                        if self._memory_manager and args.get("action") in ("add", "replace"):
-                            try:
-                                self._memory_manager.on_memory_write(
-                                    args.get("action", ""),
-                                    flush_target,
-                                    args.get("content", ""),
-                                    metadata=self._build_memory_write_metadata(
-                                        write_origin="memory_flush",
-                                        execution_context="flush_memories",
-                                    ),
-                                )
-                            except Exception:
-                                pass
-                        if not self.quiet_mode:
-                            print(f"  🧠 Memory flush: saved to {args.get('target', 'memory')}")
-                    except Exception as e:
-                        logger.warning("Memory flush tool call failed: %s", e)
-                        self._emit_auxiliary_failure("memory flush tool", e)
-        except Exception as e:
-            logger.warning("Memory flush API call failed: %s", e)
-            self._emit_auxiliary_failure("memory flush", e)
-        finally:
-            # Strip flush artifacts: remove everything from the flush message onward.
-            # Use sentinel marker instead of identity check for robustness.
-            while messages and messages[-1].get("_flush_sentinel") != _sentinel:
-                messages.pop()
-                if not messages:
-                    break
-            if messages and messages[-1].get("_flush_sentinel") == _sentinel:
-                messages.pop()
-
     def _compress_context(self, messages: list, system_message: str, *, approx_tokens: int = None, task_id: str = "default", focus_topic: str = None) -> tuple:
         """Compress conversation context and split the session in SQLite.
 
@@ -8173,8 +8096,6 @@ class AIAgent:
             f"{approx_tokens:,}" if approx_tokens else "unknown", self.model,
             focus_topic,
         )
-        # Pre-compression memory flush: let the model save memories before they're lost
-        self.flush_memories(messages, min_turns=0)
 
         # Notify external memory provider before compression discards context
         if self._memory_manager:
@@ -11126,36 +11047,69 @@ class AIAgent:
                                 continue
 
                     # ── Nous Portal: record rate limit & skip retries ─────
-                    # When Nous returns a 429, record the reset time to a
-                    # shared file so ALL sessions (cron, gateway, auxiliary)
-                    # know not to pile on.  Then skip further retries —
-                    # each one burns another RPH request and deepens the
-                    # rate limit hole.  The retry loop's top-of-iteration
-                    # guard will catch this on the next pass and try
-                    # fallback or bail with a clear message.
+                    # When Nous returns a 429 that is a genuine account-
+                    # level rate limit, record the reset time to a shared
+                    # file so ALL sessions (cron, gateway, auxiliary) know
+                    # not to pile on, then skip further retries -- each
+                    # one burns another RPH request and deepens the hole.
+                    # The retry loop's top-of-iteration guard will catch
+                    # this on the next pass and try fallback or bail.
+                    #
+                    # IMPORTANT: Nous Portal multiplexes multiple upstream
+                    # providers (DeepSeek, Kimi, MiMo, Hermes).  A 429 can
+                    # also mean an UPSTREAM provider is out of capacity
+                    # for one specific model -- transient, clears in
+                    # seconds, nothing to do with the caller's quota.
+                    # Tripping the cross-session breaker on that would
+                    # block every Nous model for minutes.  We use
+                    # ``is_genuine_nous_rate_limit`` to tell the two
+                    # apart via the 429's own x-ratelimit-* headers and
+                    # the last-known-good state captured on the previous
+                    # successful response.
                     if (
                         is_rate_limited
                         and self.provider == "nous"
                         and classified.reason == FailoverReason.rate_limit
                         and not recovered_with_pool
                     ):
+                        _genuine_nous_rate_limit = False
                         try:
-                            from agent.nous_rate_guard import record_nous_rate_limit
+                            from agent.nous_rate_guard import (
+                                is_genuine_nous_rate_limit,
+                                record_nous_rate_limit,
+                            )
                             _err_resp = getattr(api_error, "response", None)
                             _err_hdrs = (
                                 getattr(_err_resp, "headers", None)
                                 if _err_resp else None
                             )
-                            record_nous_rate_limit(
+                            _genuine_nous_rate_limit = is_genuine_nous_rate_limit(
                                 headers=_err_hdrs,
-                                error_context=error_context,
+                                last_known_state=self._rate_limit_state,
                             )
+                            if _genuine_nous_rate_limit:
+                                record_nous_rate_limit(
+                                    headers=_err_hdrs,
+                                    error_context=error_context,
+                                )
+                            else:
+                                logging.info(
+                                    "Nous 429 looks like upstream capacity "
+                                    "(no exhausted bucket in headers or "
+                                    "last-known state) -- not tripping "
+                                    "cross-session breaker."
+                                )
                         except Exception:
                             pass
-                        # Skip straight to max_retries — the top-of-loop
-                        # guard will handle fallback or bail cleanly.
-                        retry_count = max_retries
-                        continue
+                        if _genuine_nous_rate_limit:
+                            # Skip straight to max_retries -- the
+                            # top-of-loop guard will handle fallback or
+                            # bail cleanly.
+                            retry_count = max_retries
+                            continue
+                        # Upstream capacity 429: fall through to normal
+                        # retry logic.  A different model (or the same
+                        # model a moment later) will typically succeed.
 
                     is_payload_too_large = (
                         classified.reason == FailoverReason.payload_too_large
@@ -11757,16 +11711,26 @@ class AIAgent:
                     interim_has_content = bool((interim_msg.get("content") or "").strip())
                     interim_has_reasoning = bool(interim_msg.get("reasoning", "").strip()) if isinstance(interim_msg.get("reasoning"), str) else False
                     interim_has_codex_reasoning = bool(interim_msg.get("codex_reasoning_items"))
+                    interim_has_codex_message_items = bool(interim_msg.get("codex_message_items"))
 
-                    if interim_has_content or interim_has_reasoning or interim_has_codex_reasoning:
+                    if (
+                        interim_has_content
+                        or interim_has_reasoning
+                        or interim_has_codex_reasoning
+                        or interim_has_codex_message_items
+                    ):
                         last_msg = messages[-1] if messages else None
                         # Duplicate detection: two consecutive incomplete assistant
                         # messages with identical content AND reasoning are collapsed.
-                        # For reasoning-only messages (codex_reasoning_items differ but
-                        # visible content/reasoning are both empty), we also compare
-                        # the encrypted items to avoid silently dropping new state.
+                        # For provider-state-only changes (encrypted reasoning
+                        # items or replayable message ids/phases/statuses differ
+                        # while visible content/reasoning are unchanged), compare
+                        # those opaque payloads too so we don't silently drop the
+                        # newer continuation state.
                         last_codex_items = last_msg.get("codex_reasoning_items") if isinstance(last_msg, dict) else None
                         interim_codex_items = interim_msg.get("codex_reasoning_items")
+                        last_codex_message_items = last_msg.get("codex_message_items") if isinstance(last_msg, dict) else None
+                        interim_codex_message_items = interim_msg.get("codex_message_items")
                         duplicate_interim = (
                             isinstance(last_msg, dict)
                             and last_msg.get("role") == "assistant"
@@ -11774,6 +11738,7 @@ class AIAgent:
                             and (last_msg.get("content") or "") == (interim_msg.get("content") or "")
                             and (last_msg.get("reasoning") or "") == (interim_msg.get("reasoning") or "")
                             and last_codex_items == interim_codex_items
+                            and last_codex_message_items == interim_codex_message_items
                         )
                         if not duplicate_interim:
                             messages.append(interim_msg)

scripts/build_model_catalog.py

@@ -0,0 +1,95 @@
+#!/usr/bin/env python3
+"""Build the Hermes Model Catalog — a centralized JSON manifest of curated models.
+
+This script reads the in-repo hardcoded curated lists (``OPENROUTER_MODELS``,
+``_PROVIDER_MODELS["nous"]``) and writes them to a JSON manifest that the
+Hermes CLI fetches at runtime. Publishing the catalog through the docs site
+lets maintainers update model lists without shipping a Hermes release.
+
+The runtime fetcher falls back to the same in-repo hardcoded lists if the
+manifest is unreachable, so this script is a convenience for keeping the
+manifest in sync — not a source of truth.
+
+Usage::
+
+    python scripts/build_model_catalog.py
+
+Output: ``website/static/api/model-catalog.json``
+
+Live URL (after ``deploy-site.yml`` runs on merge to main):
+``https://hermes-agent.nousresearch.com/docs/api/model-catalog.json``
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import sys
+from datetime import datetime, timezone
+
+REPO_ROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+sys.path.insert(0, REPO_ROOT)
+
+# Ensure HERMES_HOME is set for imports that touch it at module level.
+os.environ.setdefault("HERMES_HOME", os.path.join(os.path.expanduser("~"), ".hermes"))
+
+from hermes_cli.models import OPENROUTER_MODELS, _PROVIDER_MODELS  # noqa: E402
+
+OUTPUT_PATH = os.path.join(REPO_ROOT, "website", "static", "api", "model-catalog.json")
+CATALOG_VERSION = 1
+
+
+def build_catalog() -> dict:
+    return {
+        "version": CATALOG_VERSION,
+        "updated_at": datetime.now(timezone.utc).strftime("%Y-%m-%dT%H:%M:%SZ"),
+        "metadata": {
+            "source": "hermes-agent repo",
+            "docs": "https://hermes-agent.nousresearch.com/docs/reference/model-catalog",
+        },
+        "providers": {
+            "openrouter": {
+                "metadata": {
+                    "display_name": "OpenRouter",
+                    "note": (
+                        "Descriptions drive picker badges. Live /api/v1/models "
+                        "filters curated ids by tool-calling support and free pricing."
+                    ),
+                },
+                "models": [
+                    {"id": mid, "description": desc}
+                    for mid, desc in OPENROUTER_MODELS
+                ],
+            },
+            "nous": {
+                "metadata": {
+                    "display_name": "Nous Portal",
+                    "note": (
+                        "Free-tier gating is determined live via Portal pricing "
+                        "(partition_nous_models_by_tier), not this manifest."
+                    ),
+                },
+                "models": [
+                    {"id": mid}
+                    for mid in _PROVIDER_MODELS.get("nous", [])
+                ],
+            },
+        },
+    }
+
+
+def main() -> int:
+    catalog = build_catalog()
+    os.makedirs(os.path.dirname(OUTPUT_PATH), exist_ok=True)
+    with open(OUTPUT_PATH, "w") as fh:
+        json.dump(catalog, fh, indent=2)
+        fh.write("\n")
+
+    print(f"Wrote {OUTPUT_PATH}")
+    for provider, block in catalog["providers"].items():
+        print(f"  {provider}: {len(block['models'])} models")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

scripts/install.sh

@@ -29,10 +29,25 @@ BOLD='\033[1m'
 REPO_URL_SSH="git@github.com:NousResearch/hermes-agent.git"
 REPO_URL_HTTPS="https://github.com/NousResearch/hermes-agent.git"
 HERMES_HOME="${HERMES_HOME:-$HOME/.hermes}"
-INSTALL_DIR="${HERMES_INSTALL_DIR:-$HERMES_HOME/hermes-agent}"
+# INSTALL_DIR is resolved AFTER arg parsing and OS detection so we can pick an
+# FHS-style layout for root installs.  Track whether the user gave us an
+# explicit directory — if so we never override it.
+if [ -n "${HERMES_INSTALL_DIR:-}" ]; then
+    INSTALL_DIR="$HERMES_INSTALL_DIR"
+    INSTALL_DIR_EXPLICIT=true
+else
+    INSTALL_DIR=""
+    INSTALL_DIR_EXPLICIT=false
+fi
 PYTHON_VERSION="3.11"
 NODE_VERSION="22"
 
+# FHS-style root install layout (set by resolve_install_layout when applicable):
+#   code at /usr/local/lib/hermes-agent, command at /usr/local/bin/hermes,
+#   data still at /root/.hermes (HERMES_HOME).  Matches Claude Code / Codex CLI
+#   and keeps Docker bind-mounted /root/ volumes lean.
+ROOT_FHS_LAYOUT=false
+
 # Options
 USE_VENV=true
 RUN_SETUP=true
@@ -64,6 +79,7 @@ while [[ $# -gt 0 ]]; do
             ;;
         --dir)
             INSTALL_DIR="$2"
+            INSTALL_DIR_EXPLICIT=true
             shift 2
             ;;
         --hermes-home)
@@ -79,9 +95,20 @@ while [[ $# -gt 0 ]]; do
             echo "  --no-venv      Don't create virtual environment"
             echo "  --skip-setup   Skip interactive setup wizard"
             echo "  --branch NAME  Git branch to install (default: main)"
-            echo "  --dir PATH     Installation directory (default: ~/.hermes/hermes-agent)"
+            echo "  --dir PATH     Installation directory"
+            echo "                   default (non-root):  ~/.hermes/hermes-agent"
+            echo "                   default (root, Linux): /usr/local/lib/hermes-agent"
             echo "  --hermes-home PATH  Data directory (default: ~/.hermes, or \$HERMES_HOME)"
             echo "  -h, --help     Show this help"
+            echo ""
+            echo "Notes:"
+            echo "  When running as root on Linux, Hermes installs the code under"
+            echo "  /usr/local/lib/hermes-agent and links the command into"
+            echo "  /usr/local/bin/hermes (FHS layout — matches Claude Code / Codex CLI)."
+            echo "  Data, config, sessions, and logs still live in \$HERMES_HOME"
+            echo "  (default /root/.hermes).  This keeps Docker bind-mounted volumes"
+            echo "  small and ensures the command is on PATH for all shells."
+            echo "  Existing installs at \$HERMES_HOME/hermes-agent are preserved in-place."
             exit 0
             ;;
         *)
@@ -163,9 +190,60 @@ is_termux() {
     [ -n "${TERMUX_VERSION:-}" ] || [[ "${PREFIX:-}" == *"com.termux/files/usr"* ]]
 }
 
+# Decide where the repo checkout + venv live, and where the `hermes` command
+# symlink goes.  Called after detect_os so $OS/$DISTRO are known.
+#
+# Defaults:
+#   - Non-root, any OS:       INSTALL_DIR = $HERMES_HOME/hermes-agent
+#                             command link in $HOME/.local/bin
+#   - Termux (any uid):       INSTALL_DIR = $HERMES_HOME/hermes-agent
+#                             command link in $PREFIX/bin (already on PATH)
+#   - Root on Linux (new):    INSTALL_DIR = /usr/local/lib/hermes-agent
+#                             command link in /usr/local/bin
+#                             (unless a legacy install already exists at
+#                              $HERMES_HOME/hermes-agent — then preserve it)
+#
+# Always no-op when the user set --dir or $HERMES_INSTALL_DIR.
+resolve_install_layout() {
+    if [ "$INSTALL_DIR_EXPLICIT" = true ]; then
+        log_info "Install directory: $INSTALL_DIR (explicit)"
+        return 0
+    fi
+
+    # Termux: package manager manages /data/data/..., keep code in HERMES_HOME.
+    if is_termux; then
+        INSTALL_DIR="$HERMES_HOME/hermes-agent"
+        return 0
+    fi
+
+    # Root on Linux: prefer FHS layout unless a legacy install already exists.
+    # macOS root installs keep the legacy layout because /usr/local/ on macOS
+    # is Homebrew territory and we don't want to fight that.
+    if [ "$OS" = "linux" ] && [ "$(id -u)" -eq 0 ]; then
+        if [ -d "$HERMES_HOME/hermes-agent/.git" ]; then
+            INSTALL_DIR="$HERMES_HOME/hermes-agent"
+            log_info "Existing install detected at $INSTALL_DIR — keeping legacy layout"
+            log_info "  (new root installs use /usr/local/lib/hermes-agent)"
+            return 0
+        fi
+        INSTALL_DIR="/usr/local/lib/hermes-agent"
+        ROOT_FHS_LAYOUT=true
+        log_info "Root install on Linux — using FHS layout"
+        log_info "  Code:    $INSTALL_DIR"
+        log_info "  Command: /usr/local/bin/hermes"
+        log_info "  Data:    $HERMES_HOME (unchanged)"
+        return 0
+    fi
+
+    # Default: non-root, non-Termux → legacy user-scoped layout.
+    INSTALL_DIR="$HERMES_HOME/hermes-agent"
+}
+
 get_command_link_dir() {
     if is_termux && [ -n "${PREFIX:-}" ]; then
         echo "$PREFIX/bin"
+    elif [ "$ROOT_FHS_LAYOUT" = true ]; then
+        echo "/usr/local/bin"
     else
         echo "$HOME/.local/bin"
     fi
@@ -174,6 +252,8 @@ get_command_link_dir() {
 get_command_link_display_dir() {
     if is_termux && [ -n "${PREFIX:-}" ]; then
         echo '$PREFIX/bin'
+    elif [ "$ROOT_FHS_LAYOUT" = true ]; then
+        echo '/usr/local/bin'
     else
         echo '~/.local/bin'
     fi
@@ -975,6 +1055,14 @@ setup_path() {
         return 0
     fi
 
+    # FHS layout: /usr/local/bin is on PATH for every standard shell, nothing to inject.
+    if [ "$ROOT_FHS_LAYOUT" = true ]; then
+        export PATH="$command_link_dir:$PATH"
+        log_info "/usr/local/bin is already on PATH for all shells"
+        log_success "hermes command ready"
+        return 0
+    fi
+
     # Check if ~/.local/bin is on PATH; if not, add it to shell config.
     # Detect the user's actual login shell (not the shell running this script,
     # which is always bash when piped from curl).
@@ -1339,12 +1427,12 @@ print_success() {
     echo ""
 
     # Show file locations
-    echo -e "${CYAN}${BOLD}📁 Your files (all in ~/.hermes/):${NC}"
+    echo -e "${CYAN}${BOLD}📁 Your files:${NC}"
     echo ""
-    echo -e "   ${YELLOW}Config:${NC}    ~/.hermes/config.yaml"
-    echo -e "   ${YELLOW}API Keys:${NC}  ~/.hermes/.env"
-    echo -e "   ${YELLOW}Data:${NC}      ~/.hermes/cron/, sessions/, logs/"
-    echo -e "   ${YELLOW}Code:${NC}      ~/.hermes/hermes-agent/"
+    echo -e "   ${YELLOW}Config:${NC}    $HERMES_HOME/config.yaml"
+    echo -e "   ${YELLOW}API Keys:${NC}  $HERMES_HOME/.env"
+    echo -e "   ${YELLOW}Data:${NC}      $HERMES_HOME/cron/, sessions/, logs/"
+    echo -e "   ${YELLOW}Code:${NC}      $INSTALL_DIR"
     echo ""
 
     echo -e "${CYAN}─────────────────────────────────────────────────────────${NC}"
@@ -1364,6 +1452,9 @@ print_success() {
     if [ "$DISTRO" = "termux" ]; then
         echo -e "${YELLOW}⚡ 'hermes' was linked into $(get_command_link_display_dir), which is already on PATH in Termux.${NC}"
         echo ""
+    elif [ "$ROOT_FHS_LAYOUT" = true ]; then
+        echo -e "${YELLOW}⚡ 'hermes' was linked into /usr/local/bin and is ready to use — no shell reload needed.${NC}"
+        echo ""
     else
         echo -e "${YELLOW}⚡ Reload your shell to use 'hermes' command:${NC}"
         echo ""
@@ -1415,6 +1506,7 @@ main() {
     print_banner
 
     detect_os
+    resolve_install_layout
     install_uv
     check_python
     check_git

scripts/release.py

@@ -43,6 +43,7 @@ AUTHOR_MAP = {
     "teknium1@gmail.com": "teknium1",
     "teknium@nousresearch.com": "teknium1",
     "127238744+teknium1@users.noreply.github.com": "teknium1",
+    "focusflow.app.help@gmail.com": "yes999zc",
     "343873859@qq.com": "DrStrangerUJN",
     "uzmpsk.dilekakbas@gmail.com": "dlkakbs",
     "jefferson@heimdallstrategy.com": "Mind-Dragon",
@@ -51,6 +52,7 @@ AUTHOR_MAP = {
     "web3blind@users.noreply.github.com": "web3blind",
     "julia@alexland.us": "alexg0bot",
     "1060770+benjaminsehl@users.noreply.github.com": "benjaminsehl",
+    "nerijusn76@gmail.com": "Nerijusas",
     # contributors (from noreply pattern)
     "david.vv@icloud.com": "davidvv",
     "wangqiang@wangqiangdeMac-mini.local": "xiaoqiang243",
@@ -67,6 +69,9 @@ AUTHOR_MAP = {
     "kshitijk4poor@gmail.com": "kshitijk4poor",
     "keira.voss94@gmail.com": "keiravoss94",
     "16443023+stablegenius49@users.noreply.github.com": "stablegenius49",
+    "fqsy1416@gmail.com": "EKKOLearnAI",
+    "simbamax99@gmail.com": "simbam99",
+    "iris@growthpillars.co": "irispillars",
     "185121704+stablegenius49@users.noreply.github.com": "stablegenius49",
     "101283333+batuhankocyigit@users.noreply.github.com": "batuhankocyigit",
     "255305877+ismell0992-afk@users.noreply.github.com": "ismell0992-afk",
@@ -91,6 +96,8 @@ AUTHOR_MAP = {
     "104278804+Sertug17@users.noreply.github.com": "Sertug17",
     "112503481+caentzminger@users.noreply.github.com": "caentzminger",
     "258577966+voidborne-d@users.noreply.github.com": "voidborne-d",
+    "liusway405@gmail.com": "voidborne-d",
+    "xydarcher@uestc.edu.cn": "Readon",
     "sir_even@icloud.com": "sirEven",
     "36056348+sirEven@users.noreply.github.com": "sirEven",
     "70424851+insecurejezza@users.noreply.github.com": "insecurejezza",
@@ -113,6 +120,7 @@ AUTHOR_MAP = {
     "30841158+n-WN@users.noreply.github.com": "n-WN",
     "tsuijinglei@gmail.com": "hiddenpuppy",
     "jerome@clawwork.ai": "HiddenPuppy",
+    "jerome.benoit@sap.com": "jerome-benoit",
     "wysie@users.noreply.github.com": "Wysie",
     "leoyuan0099@gmail.com": "keyuyuan",
     "bxzt2006@163.com": "Only-Code-A",
@@ -173,6 +181,10 @@ AUTHOR_MAP = {
     "jaisehgal11299@gmail.com": "jaisup",
     "percydikec@gmail.com": "PercyDikec",
     "noonou7@gmail.com": "HenkDz",
+    # Azure Foundry salvage (PRs #9029, #4599, #10086, #8766)
+    "tech@smartlogics.net": "TechPrototyper",
+    "637186+HangGlidersRule@users.noreply.github.com": "HangGlidersRule",
+    "pein892@gmail.com": "pein892",
     "dean.kerr@gmail.com": "deankerr",
     "socrates1024@gmail.com": "socrates1024",
     "seanalt555@gmail.com": "Salt-555",
@@ -407,6 +419,7 @@ AUTHOR_MAP = {
     "105142614+VTRiot@users.noreply.github.com": "VTRiot",
     "vivien000812@gmail.com": "iamagenius00",
     "89228157+Feranmi10@users.noreply.github.com": "Feranmi10",
+    "oluwadareferanmi11@gmail.com": "Feranmi10",
     "simon@gtcl.us": "simon-gtcl",
     "suzukaze.haduki@gmail.com": "houko",
     "cliff@cigii.com": "cgarwood82",
@@ -501,6 +514,8 @@ AUTHOR_MAP = {
     "codex@openai.invalid": "teknium1",
     "screenmachine@gmail.com": "teknium1",
     "chenzeshi@live.com": "chen1749144759",
+    "mor.aleksandr@yahoo.com": "MorAlekss",
+    "ash@users.noreply.github.com": "ash",
 }
 
 

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

@@ -281,7 +281,6 @@ Type these during an interactive chat session.
 ### Utility
 ```
 /branch (/fork)      Branch the current session
-/btw                 Ephemeral side question (doesn't interrupt main task)
 /fast                Toggle priority/fast processing
 /browser             Open CDP browser connection
 /history             Show conversation history (CLI)

skills/devops/kanban-orchestrator/SKILL.md

@@ -0,0 +1,152 @@
+---
+name: kanban-orchestrator
+description: Decomposition playbook + specialist-roster conventions + anti-temptation rules for an orchestrator profile routing work through Kanban. The "don't do the work yourself" rule and the basic lifecycle are auto-injected into every kanban worker's system prompt; this skill is the deeper playbook when you're specifically playing the orchestrator role.
+version: 2.0.0
+metadata:
+  hermes:
+    tags: [kanban, multi-agent, orchestration, routing]
+    related_skills: [kanban-worker]
+---
+
+# Kanban Orchestrator — Decomposition Playbook
+
+> The **core worker lifecycle** (including the `kanban_create` fan-out pattern and the "decompose, don't execute" rule) is auto-injected into every kanban process via the `KANBAN_GUIDANCE` system-prompt block. This skill is the deeper playbook when you're an orchestrator profile whose whole job is routing.
+
+## When to use the board (vs. just doing the work)
+
+Create Kanban tasks when any of these are true:
+
+1. **Multiple specialists are needed.** Research + analysis + writing is three profiles.
+2. **The work should survive a crash or restart.** Long-running, recurring, or important.
+3. **The user might want to interject.** Human-in-the-loop at any step.
+4. **Multiple subtasks can run in parallel.** Fan-out for speed.
+5. **Review / iteration is expected.** A reviewer profile loops on drafter output.
+6. **The audit trail matters.** Board rows persist in SQLite forever.
+
+If *none* of those apply — it's a small one-shot reasoning task — use `delegate_task` instead or answer the user directly.
+
+## The anti-temptation rules
+
+Your job description says "route, don't execute." The rules that enforce that:
+
+- **Do not execute the work yourself.** Your restricted toolset usually doesn't even include terminal/file/code/web for implementation. If you find yourself "just fixing this quickly" — stop and create a task for the right specialist.
+- **For any concrete task, create a Kanban task and assign it.** Every single time.
+- **If no specialist fits, ask the user which profile to create.** Do not default to doing it yourself under "close enough."
+- **Decompose, route, and summarize — that's the whole job.**
+
+## The standard specialist roster (convention)
+
+Unless the user's setup has customized profiles, assume these exist. Adjust to whatever the user actually has — ask if you're unsure.
+
+| Profile | Does | Typical workspace |
+|---|---|---|
+| `researcher` | Reads sources, gathers facts, writes findings | `scratch` |
+| `analyst` | Synthesizes, ranks, de-dupes. Consumes multiple `researcher` outputs | `scratch` |
+| `writer` | Drafts prose in the user's voice | `scratch` or `dir:` into their Obsidian vault |
+| `reviewer` | Reads output, leaves findings, gates approval | `scratch` |
+| `backend-eng` | Writes server-side code | `worktree` |
+| `frontend-eng` | Writes client-side code | `worktree` |
+| `ops` | Runs scripts, manages services, handles deployments | `dir:` into ops scripts repo |
+| `pm` | Writes specs, acceptance criteria | `scratch` |
+
+## Decomposition playbook
+
+### Step 1 — Understand the goal
+
+Ask clarifying questions if the goal is ambiguous. Cheap to ask; expensive to spawn the wrong fleet.
+
+### Step 2 — Sketch the task graph
+
+Before creating anything, draft the graph out loud (in your response to the user). Example for "Analyze whether we should migrate to Postgres":
+
+```
+T1  researcher        research: Postgres cost vs current
+T2  researcher        research: Postgres performance vs current
+T3  analyst           synthesize migration recommendation       parents: T1, T2
+T4  writer            draft decision memo                       parents: T3
+```
+
+Show this to the user. Let them correct it before you create anything.
+
+### Step 3 — Create tasks and link
+
+```python
+t1 = kanban_create(
+    title="research: Postgres cost vs current",
+    assignee="researcher",
+    body="Compare estimated infrastructure costs, migration costs, and ongoing ops costs over a 3-year window. Sources: AWS/GCP pricing, team time estimates, current Postgres bills from peers.",
+    tenant=os.environ.get("HERMES_TENANT"),
+)["task_id"]
+
+t2 = kanban_create(
+    title="research: Postgres performance vs current",
+    assignee="researcher",
+    body="Compare query latency, throughput, and scaling characteristics at our expected data volume (~500GB, 10k QPS peak). Sources: benchmark papers, public case studies, pgbench results if easy.",
+)["task_id"]
+
+t3 = kanban_create(
+    title="synthesize migration recommendation",
+    assignee="analyst",
+    body="Read the findings from T1 (cost) and T2 (performance). Produce a 1-page recommendation with explicit trade-offs and a go/no-go call.",
+    parents=[t1, t2],
+)["task_id"]
+
+t4 = kanban_create(
+    title="draft decision memo",
+    assignee="writer",
+    body="Turn the analyst's recommendation into a 2-page memo for the CTO. Match the tone of previous decision memos in the team's knowledge base.",
+    parents=[t3],
+)["task_id"]
+```
+
+`parents=[...]` gates promotion — children stay in `todo` until every parent reaches `done`, then auto-promote to `ready`. No manual coordination needed; the dispatcher and dependency engine handle it.
+
+### Step 4 — Complete your own task
+
+If you were spawned as a task yourself (e.g. `planner` profile was assigned `T0: "investigate Postgres migration"`), mark it done with a summary of what you created:
+
+```python
+kanban_complete(
+    summary="decomposed into T1-T4: 2 researchers parallel, 1 analyst on their outputs, 1 writer on the recommendation",
+    metadata={
+        "task_graph": {
+            "T1": {"assignee": "researcher", "parents": []},
+            "T2": {"assignee": "researcher", "parents": []},
+            "T3": {"assignee": "analyst", "parents": ["T1", "T2"]},
+            "T4": {"assignee": "writer", "parents": ["T3"]},
+        },
+    },
+)
+```
+
+### Step 5 — Report back to the user
+
+Tell them what you created in plain prose:
+
+> I've queued 4 tasks:
+> - **T1** (researcher): cost comparison
+> - **T2** (researcher): performance comparison, in parallel with T1
+> - **T3** (analyst): synthesizes T1 + T2 into a recommendation
+> - **T4** (writer): turns T3 into a CTO memo
+>
+> The dispatcher will pick up T1 and T2 now. T3 starts when both finish. You'll get a gateway ping when T4 completes. Use the dashboard or `hermes kanban tail <id>` to follow along.
+
+## Common patterns
+
+**Fan-out + fan-in (research → synthesize):** N `researcher` tasks with no parents, one `analyst` task with all of them as parents.
+
+**Pipeline with gates:** `pm → backend-eng → reviewer`. Each stage's `parents=[previous_task]`. Reviewer blocks or completes; if reviewer blocks, the operator unblocks with feedback and respawns.
+
+**Same-profile queue:** 50 tasks, all assigned to `translator`, no dependencies between them. Dispatcher serializes — translator processes them in priority order, accumulating experience in their own memory.
+
+**Human-in-the-loop:** Any task can `kanban_block()` to wait for input. Dispatcher respawns after `/unblock`. The comment thread carries the full context.
+
+## Pitfalls
+
+**Reassignment vs. new task.** If a reviewer blocks with "needs changes," create a NEW task linked from the reviewer's task — don't re-run the same task with a stern look. The new task is assigned to the original implementer profile.
+
+**Argument order for links.** `kanban_link(parent_id=..., child_id=...)` — parent first. Mixing them up demotes the wrong task to `todo`.
+
+**Don't pre-create the whole graph if the shape depends on intermediate findings.** If T3's structure depends on what T1 and T2 find, let T3 exist as a "synthesize findings" task whose own first step is to read parent handoffs and plan the rest. Orchestrators can spawn orchestrators.
+
+**Tenant inheritance.** If `HERMES_TENANT` is set in your env, pass `tenant=os.environ.get("HERMES_TENANT")` on every `kanban_create` call so child tasks stay in the same namespace.

skills/devops/kanban-worker/SKILL.md

@@ -0,0 +1,134 @@
+---
+name: kanban-worker
+description: Pitfalls, examples, and edge cases for Hermes Kanban workers. The lifecycle itself is auto-injected into every worker's system prompt as KANBAN_GUIDANCE (from agent/prompt_builder.py); this skill is what you load when you want deeper detail on specific scenarios.
+version: 2.0.0
+metadata:
+  hermes:
+    tags: [kanban, multi-agent, collaboration, workflow, pitfalls]
+    related_skills: [kanban-orchestrator]
+---
+
+# Kanban Worker — Pitfalls and Examples
+
+> You're seeing this skill because the Hermes Kanban dispatcher spawned you as a worker with `--skills kanban-worker` — it's loaded automatically for every dispatched worker. The **lifecycle** (6 steps: orient → work → heartbeat → block/complete) also lives in the `KANBAN_GUIDANCE` block that's auto-injected into your system prompt. This skill is the deeper detail: good handoff shapes, retry diagnostics, edge cases.
+
+## Workspace handling
+
+Your workspace kind determines how you should behave inside `$HERMES_KANBAN_WORKSPACE`:
+
+| Kind | What it is | How to work |
+|---|---|---|
+| `scratch` | Fresh tmp dir, yours alone | Read/write freely; it gets GC'd when the task is archived. |
+| `dir:<path>` | Shared persistent directory | Other runs will read what you write. Treat it like long-lived state. Path is guaranteed absolute (the kernel rejects relative paths). |
+| `worktree` | Git worktree at the resolved path | If `.git` doesn't exist, run `git worktree add <path> <branch>` from the main repo first, then cd and work normally. Commit work here. |
+
+## Tenant isolation
+
+If `$HERMES_TENANT` is set, the task belongs to a tenant namespace. When reading or writing persistent memory, prefix memory entries with the tenant so context doesn't leak across tenants:
+
+- Good: `business-a: Acme is our biggest customer`
+- Bad (leaks): `Acme is our biggest customer`
+
+## Good summary + metadata shapes
+
+The `kanban_complete(summary=..., metadata=...)` handoff is how downstream workers read what you did. Patterns that work:
+
+**Coding task:**
+```python
+kanban_complete(
+    summary="shipped rate limiter — token bucket, keys on user_id with IP fallback, 14 tests pass",
+    metadata={
+        "changed_files": ["rate_limiter.py", "tests/test_rate_limiter.py"],
+        "tests_run": 14,
+        "tests_passed": 14,
+        "decisions": ["user_id primary, IP fallback for unauthenticated requests"],
+    },
+)
+```
+
+**Research task:**
+```python
+kanban_complete(
+    summary="3 competing libraries reviewed; vLLM wins on throughput, SGLang on latency, Tensorrt-LLM on memory efficiency",
+    metadata={
+        "sources_read": 12,
+        "recommendation": "vLLM",
+        "benchmarks": {"vllm": 1.0, "sglang": 0.87, "trtllm": 0.72},
+    },
+)
+```
+
+**Review task:**
+```python
+kanban_complete(
+    summary="reviewed PR #123; 2 blocking issues found (SQL injection in /search, missing CSRF on /settings)",
+    metadata={
+        "pr_number": 123,
+        "findings": [
+            {"severity": "critical", "file": "api/search.py", "line": 42, "issue": "raw SQL concat"},
+            {"severity": "high", "file": "api/settings.py", "issue": "missing CSRF middleware"},
+        ],
+        "approved": False,
+    },
+)
+```
+
+Shape `metadata` so downstream parsers (reviewers, aggregators, schedulers) can use it without re-reading your prose.
+
+## Block reasons that get answered fast
+
+Bad: `"stuck"` — the human has no context.
+
+Good: one sentence naming the specific decision you need. Leave longer context as a comment instead.
+
+```python
+kanban_comment(
+    task_id=os.environ["HERMES_KANBAN_TASK"],
+    body="Full context: I have user IPs from Cloudflare headers but some users are behind NATs with thousands of peers. Keying on IP alone causes false positives.",
+)
+kanban_block(reason="Rate limit key choice: IP (simple, NAT-unsafe) or user_id (requires auth, skips anonymous endpoints)?")
+```
+
+The block message is what appears in the dashboard / gateway notifier. The comment is the deeper context a human reads when they open the task.
+
+## Heartbeats worth sending
+
+Good heartbeats name progress: `"epoch 12/50, loss 0.31"`, `"scanned 1.2M/2.4M rows"`, `"uploaded 47/120 videos"`.
+
+Bad heartbeats: `"still working"`, empty notes, sub-second intervals. Every few minutes max; skip entirely for tasks under ~2 minutes.
+
+## Retry scenarios
+
+If you open the task and `kanban_show` returns `runs: [...]` with one or more closed runs, you're a retry. The prior runs' `outcome` / `summary` / `error` tell you what didn't work. Don't repeat that path. Typical retry diagnostics:
+
+- `outcome: "timed_out"` — the previous attempt hit `max_runtime_seconds`. You may need to chunk the work or shorten it.
+- `outcome: "crashed"` — OOM or segfault. Reduce memory footprint.
+- `outcome: "spawn_failed"` + `error: "..."` — usually a profile config issue (missing credential, bad PATH). Ask the human via `kanban_block` instead of retrying blindly.
+- `outcome: "reclaimed"` + `summary: "task archived..."` — operator archived the task out from under the previous run; you probably shouldn't be running at all, check status carefully.
+- `outcome: "blocked"` — a previous attempt blocked; the unblock comment should be in the thread by now.
+
+## Do NOT
+
+- Call `delegate_task` as a substitute for `kanban_create`. `delegate_task` is for short reasoning subtasks inside YOUR run; `kanban_create` is for cross-agent handoffs that outlive one API loop.
+- Modify files outside `$HERMES_KANBAN_WORKSPACE` unless the task body says to.
+- Create follow-up tasks assigned to yourself — assign to the right specialist.
+- Complete a task you didn't actually finish. Block it instead.
+
+## Pitfalls
+
+**Task state can change between dispatch and your startup.** Between when the dispatcher claimed and when your process actually booted, the task may have been blocked, reassigned, or archived. Always `kanban_show` first. If it reports `blocked` or `archived`, stop — you shouldn't be running.
+
+**Workspace may have stale artifacts.** Especially `dir:` and `worktree` workspaces can have files from previous runs. Read the comment thread — it usually explains why you're running again and what state the workspace is in.
+
+**Don't rely on the CLI when the guidance is available.** The `kanban_*` tools work across all terminal backends (Docker, Modal, SSH). `hermes kanban <verb>` from your terminal tool will fail in containerized backends because the CLI isn't installed there. When in doubt, use the tool.
+
+## CLI fallback (for scripting)
+
+Every tool has a CLI equivalent for human operators and scripts:
+- `kanban_show` ↔ `hermes kanban show <id> --json`
+- `kanban_complete` ↔ `hermes kanban complete <id> --summary "..." --metadata '{...}'`
+- `kanban_block` ↔ `hermes kanban block <id> "reason"`
+- `kanban_create` ↔ `hermes kanban create "title" --assignee <profile> [--parent <id>]`
+- etc.
+
+Use the tools from inside an agent; the CLI exists for the human at the terminal.

skills/mlops/inference/obliteratus/SKILL.md

@@ -17,6 +17,13 @@ Remove refusal behaviors (guardrails) from open-weight LLMs without retraining o
 
 **License warning:** OBLITERATUS is AGPL-3.0. NEVER import it as a Python library. Always invoke via CLI (`obliteratus` command) or subprocess. This keeps Hermes Agent's MIT license clean.
 
+## Video Guide
+
+Walkthrough of OBLITERATUS used by a Hermes agent to abliterate Gemma:
+https://www.youtube.com/watch?v=8fG9BrNTeHs ("OBLITERATUS: An AI Agent Removed Gemma 4's Safety Guardrails")
+
+Useful when the user wants a visual overview of the end-to-end workflow before running it themselves.
+
 ## When to Use This Skill
 
 Trigger when the user:

skills/productivity/google-workspace/scripts/_hermes_home.py

@@ -0,0 +1,42 @@
+"""Resolve HERMES_HOME for standalone skill scripts.
+
+Skill scripts may run outside the Hermes process (e.g. system Python,
+nix env, CI) where ``hermes_constants`` is not importable.  This module
+provides the same ``get_hermes_home()`` and ``display_hermes_home()``
+contracts as ``hermes_constants`` without requiring it on ``sys.path``.
+
+When ``hermes_constants`` IS available it is used directly so that any
+future enhancements (profile resolution, Docker detection, etc.) are
+picked up automatically.  The fallback path replicates the core logic
+from ``hermes_constants.py`` using only the stdlib.
+
+All scripts under ``google-workspace/scripts/`` should import from here
+instead of duplicating the ``HERMES_HOME = Path(os.getenv(...))`` pattern.
+"""
+
+from __future__ import annotations
+
+import os
+from pathlib import Path
+
+try:
+    from hermes_constants import display_hermes_home as display_hermes_home
+    from hermes_constants import get_hermes_home as get_hermes_home
+except (ModuleNotFoundError, ImportError):
+
+    def get_hermes_home() -> Path:
+        """Return the Hermes home directory (default: ~/.hermes).
+
+        Mirrors ``hermes_constants.get_hermes_home()``."""
+        val = os.environ.get("HERMES_HOME", "").strip()
+        return Path(val) if val else Path.home() / ".hermes"
+
+    def display_hermes_home() -> str:
+        """Return a user-friendly ``~/``-shortened display string.
+
+        Mirrors ``hermes_constants.display_hermes_home()``."""
+        home = get_hermes_home()
+        try:
+            return "~/" + str(home.relative_to(Path.home()))
+        except ValueError:
+            return str(home)

skills/productivity/google-workspace/scripts/google_api.py

@@ -31,7 +31,14 @@ from datetime import datetime, timedelta, timezone
 from email.mime.text import MIMEText
 from pathlib import Path
 
-HERMES_HOME = Path(os.getenv("HERMES_HOME", Path.home() / ".hermes"))
+# Ensure sibling modules (_hermes_home) are importable when run standalone.
+_SCRIPTS_DIR = str(Path(__file__).resolve().parent)
+if _SCRIPTS_DIR not in sys.path:
+    sys.path.insert(0, _SCRIPTS_DIR)
+
+from _hermes_home import get_hermes_home
+
+HERMES_HOME = get_hermes_home()
 TOKEN_PATH = HERMES_HOME / "google_token.json"
 CLIENT_SECRET_PATH = HERMES_HOME / "google_client_secret.json"
 

skills/productivity/google-workspace/scripts/gws_bridge.py

@@ -10,9 +10,12 @@ import sys
 from datetime import datetime, timezone
 from pathlib import Path
 
+# Ensure sibling modules (_hermes_home) are importable when run standalone.
+_SCRIPTS_DIR = str(Path(__file__).resolve().parent)
+if _SCRIPTS_DIR not in sys.path:
+    sys.path.insert(0, _SCRIPTS_DIR)
 
-def get_hermes_home() -> Path:
-    return Path(os.environ.get("HERMES_HOME", Path.home() / ".hermes"))
+from _hermes_home import get_hermes_home
 
 
 def get_token_path() -> Path:

skills/productivity/google-workspace/scripts/setup.py

@@ -21,6 +21,8 @@ Agent workflow:
   6. Run --check to verify. Done.
 """
 
+from __future__ import annotations  # allow PEP 604 `X | None` on Python 3.9+
+
 import argparse
 import json
 import os
@@ -28,13 +30,12 @@ import subprocess
 import sys
 from pathlib import Path
 
-try:
-    from hermes_constants import display_hermes_home, get_hermes_home
-except ModuleNotFoundError:
-    HERMES_AGENT_ROOT = Path(__file__).resolve().parents[4]
-    if HERMES_AGENT_ROOT.exists():
-        sys.path.insert(0, str(HERMES_AGENT_ROOT))
-    from hermes_constants import display_hermes_home, get_hermes_home
+# Ensure sibling modules (_hermes_home) are importable when run standalone.
+_SCRIPTS_DIR = str(Path(__file__).resolve().parent)
+if _SCRIPTS_DIR not in sys.path:
+    sys.path.insert(0, _SCRIPTS_DIR)
+
+from _hermes_home import display_hermes_home, get_hermes_home
 
 HERMES_HOME = get_hermes_home()
 TOKEN_PATH = HERMES_HOME / "google_token.json"
@@ -111,7 +112,11 @@ def install_deps():
         return True
     except subprocess.CalledProcessError as e:
         print(f"ERROR: Failed to install dependencies: {e}")
-        print(f"Try manually: {sys.executable} -m pip install {' '.join(REQUIRED_PACKAGES)}")
+        print(
+            "On environments without pip (e.g. Nix), install the optional extra instead:"
+        )
+        print("  pip install 'hermes-agent[google]'")
+        print(f"Or manually: {sys.executable} -m pip install {' '.join(REQUIRED_PACKAGES)}")
         return False
 
 

tests/agent/test_auxiliary_named_custom_providers.py

@@ -386,7 +386,7 @@ class TestProvidersDictApiModeAnthropicMessages:
                 },
             },
             "auxiliary": {
-                "flush_memories": {
+                "compression": {
                     "provider": "myrelay",
                     "model": "claude-sonnet-4.6",
                 },
@@ -399,11 +399,11 @@ class TestProvidersDictApiModeAnthropicMessages:
             AnthropicAuxiliaryClient,
             AsyncAnthropicAuxiliaryClient,
         )
-        async_client, async_model = get_async_text_auxiliary_client("flush_memories")
+        async_client, async_model = get_async_text_auxiliary_client("compression")
         assert isinstance(async_client, AsyncAnthropicAuxiliaryClient)
         assert async_model == "claude-sonnet-4.6"
 
-        sync_client, sync_model = get_text_auxiliary_client("flush_memories")
+        sync_client, sync_model = get_text_auxiliary_client("compression")
         assert isinstance(sync_client, AnthropicAuxiliaryClient)
         assert sync_model == "claude-sonnet-4.6"
 

tests/agent/test_context_compressor.py

@@ -847,6 +847,32 @@ class TestTokenBudgetTailProtection:
         assert isinstance(pruned, int)
 
 
+class TestUpdateModelBudgets:
+    """Regression: update_model() must recalculate token budgets."""
+
+    def test_tail_budget_recalculated(self):
+        """tail_token_budget must change after switching to a different context length."""
+        from unittest.mock import patch
+        with patch("agent.context_compressor.get_model_context_length", return_value=200_000):
+            comp = ContextCompressor("model-a", threshold_percent=0.50, quiet_mode=True)
+        old_tail = comp.tail_token_budget
+        old_max_summary = comp.max_summary_tokens
+
+        comp.update_model("model-b", context_length=32_000)
+        assert comp.tail_token_budget != old_tail, "tail_token_budget should change"
+        assert comp.tail_token_budget < old_tail, "smaller context → smaller budget"
+        assert comp.max_summary_tokens != old_max_summary, "max_summary_tokens should change"
+
+    def test_budgets_proportional(self):
+        """Budgets should be proportional to context_length after update."""
+        from unittest.mock import patch
+        with patch("agent.context_compressor.get_model_context_length", return_value=100_000):
+            comp = ContextCompressor("model-a", threshold_percent=0.50, quiet_mode=True)
+        comp.update_model("model-b", context_length=10_000)
+        assert comp.tail_token_budget == int(comp.threshold_tokens * comp.summary_target_ratio)
+        assert comp.max_summary_tokens == min(int(10_000 * 0.05), 4000)
+
+
 class TestTruncateToolCallArgsJson:
     """Regression tests for #11762.
 

tests/agent/test_model_metadata.py

@@ -192,6 +192,43 @@ class TestDefaultContextLengths:
                     f"{model_id}: expected {expected_ctx}, got {actual}"
                 )
 
+    def test_deepseek_v4_models_1m_context(self):
+        from agent.model_metadata import get_model_context_length
+        from unittest.mock import patch as mock_patch
+
+        expected_keys = {
+            "deepseek-v4-pro": 1_000_000,
+            "deepseek-v4-flash": 1_000_000,
+            "deepseek-chat": 1_000_000,
+            "deepseek-reasoner": 1_000_000,
+        }
+        for key, value in expected_keys.items():
+            assert key in DEFAULT_CONTEXT_LENGTHS, f"{key} missing"
+            assert DEFAULT_CONTEXT_LENGTHS[key] == value, (
+                f"{key} should be {value}, got {DEFAULT_CONTEXT_LENGTHS[key]}"
+            )
+
+        # Longest-first substring matching must resolve both the bare V4
+        # ids (native DeepSeek) and the vendor-prefixed forms (OpenRouter
+        # / Nous Portal) to 1M without probing down to the legacy 128K
+        # ``deepseek`` substring fallback.
+        with mock_patch("agent.model_metadata.fetch_model_metadata", return_value={}), \
+             mock_patch("agent.model_metadata.fetch_endpoint_model_metadata", return_value={}), \
+             mock_patch("agent.model_metadata.get_cached_context_length", return_value=None):
+            cases = [
+                ("deepseek-v4-pro", 1_000_000),
+                ("deepseek-v4-flash", 1_000_000),
+                ("deepseek/deepseek-v4-pro", 1_000_000),
+                ("deepseek/deepseek-v4-flash", 1_000_000),
+                ("deepseek-chat", 1_000_000),
+                ("deepseek-reasoner", 1_000_000),
+            ]
+            for model_id, expected_ctx in cases:
+                actual = get_model_context_length(model_id)
+                assert actual == expected_ctx, (
+                    f"{model_id}: expected {expected_ctx}, got {actual}"
+                )
+
     def test_all_values_positive(self):
         for key, value in DEFAULT_CONTEXT_LENGTHS.items():
             assert value > 0, f"{key} has non-positive context length"
@@ -303,7 +340,9 @@ class TestCodexOAuthContextLength:
         from agent.model_metadata import get_model_context_length
 
         # OpenRouter — should hit its own catalog path first; when mocked
-        # empty, falls through to hardcoded DEFAULT_CONTEXT_LENGTHS (400k).
+        # empty, falls through to hardcoded DEFAULT_CONTEXT_LENGTHS (1.05M,
+        # matching the real direct-API value — Codex OAuth's 272k cap is
+        # provider-specific and must not leak here).
         with patch("agent.model_metadata.fetch_model_metadata", return_value={}), \
              patch("agent.model_metadata.fetch_endpoint_model_metadata", return_value={}), \
              patch("agent.model_metadata.get_cached_context_length", return_value=None), \
@@ -314,7 +353,7 @@ class TestCodexOAuthContextLength:
                 api_key="",
                 provider="openrouter",
             )
-        assert ctx == 400_000, (
+        assert ctx == 1_050_000, (
             f"Non-Codex gpt-5.5 resolved to {ctx}; Codex 272k override "
             "leaked outside openai-codex provider"
         )
@@ -459,9 +498,10 @@ class TestGetModelContextLength:
 
     @patch("agent.model_metadata.fetch_model_metadata")
     def test_api_missing_context_length_key(self, mock_fetch):
-        """Model in API but without context_length → defaults to 128000."""
+        """Model in API but without context_length → defaults to the top
+        probe tier (currently 256K)."""
         mock_fetch.return_value = {"test/model": {"name": "Test"}}
-        assert get_model_context_length("test/model") == 128000
+        assert get_model_context_length("test/model") == CONTEXT_PROBE_TIERS[0]
 
     @patch("agent.model_metadata.fetch_model_metadata")
     def test_cache_takes_priority_over_api(self, mock_fetch, tmp_path):
@@ -814,14 +854,17 @@ class TestContextProbeTiers:
         for i in range(len(CONTEXT_PROBE_TIERS) - 1):
             assert CONTEXT_PROBE_TIERS[i] > CONTEXT_PROBE_TIERS[i + 1]
 
-    def test_first_tier_is_128k(self):
-        assert CONTEXT_PROBE_TIERS[0] == 128_000
+    def test_first_tier_is_256k(self):
+        assert CONTEXT_PROBE_TIERS[0] == 256_000
 
     def test_last_tier_is_8k(self):
         assert CONTEXT_PROBE_TIERS[-1] == 8_000
 
 
 class TestGetNextProbeTier:
+    def test_from_256k(self):
+        assert get_next_probe_tier(256_000) == 128_000
+
     def test_from_128k(self):
         assert get_next_probe_tier(128_000) == 64_000
 
@@ -841,8 +884,8 @@ class TestGetNextProbeTier:
         assert get_next_probe_tier(100_000) == 64_000
 
     def test_above_max_tier(self):
-        """Value above 128K should return 128K."""
-        assert get_next_probe_tier(500_000) == 128_000
+        """Value above 256K should return 256K."""
+        assert get_next_probe_tier(500_000) == 256_000
 
     def test_zero_returns_none(self):
         assert get_next_probe_tier(0) is None

tests/agent/test_nous_rate_guard.py

@@ -251,3 +251,141 @@ class TestAuxiliaryClientIntegration:
         monkeypatch.setattr(aux, "_read_nous_auth", lambda: None)
         result = aux._try_nous()
         assert result == (None, None)
+
+
+class TestIsGenuineNousRateLimit:
+    """Tell a real account-level 429 apart from an upstream-capacity 429.
+
+    Nous Portal multiplexes upstreams (DeepSeek, Kimi, MiMo, Hermes).
+    A 429 from an upstream out of capacity should NOT trip the
+    cross-session breaker; a real user-quota 429 should.
+    """
+
+    def test_exhausted_hourly_bucket_in_429_headers_is_genuine(self):
+        from agent.nous_rate_guard import is_genuine_nous_rate_limit
+
+        headers = {
+            "x-ratelimit-limit-requests-1h": "800",
+            "x-ratelimit-remaining-requests-1h": "0",
+            "x-ratelimit-reset-requests-1h": "3100",
+            "x-ratelimit-limit-requests": "200",
+            "x-ratelimit-remaining-requests": "198",
+            "x-ratelimit-reset-requests": "40",
+        }
+        assert is_genuine_nous_rate_limit(headers=headers) is True
+
+    def test_exhausted_tokens_bucket_is_genuine(self):
+        from agent.nous_rate_guard import is_genuine_nous_rate_limit
+
+        headers = {
+            "x-ratelimit-limit-tokens": "800000",
+            "x-ratelimit-remaining-tokens": "0",
+            "x-ratelimit-reset-tokens": "45",  # < 60s threshold -> not genuine
+            "x-ratelimit-limit-tokens-1h": "8000000",
+            "x-ratelimit-remaining-tokens-1h": "0",
+            "x-ratelimit-reset-tokens-1h": "1800",  # >= 60s threshold -> genuine
+        }
+        assert is_genuine_nous_rate_limit(headers=headers) is True
+
+    def test_healthy_headers_on_429_are_upstream_capacity(self):
+        # Classic upstream-capacity symptom: Nous edge reports plenty of
+        # headroom on every bucket, but returns 429 anyway because
+        # upstream (DeepSeek / Kimi / ...) is out of capacity.
+        from agent.nous_rate_guard import is_genuine_nous_rate_limit
+
+        headers = {
+            "x-ratelimit-limit-requests": "200",
+            "x-ratelimit-remaining-requests": "198",
+            "x-ratelimit-reset-requests": "40",
+            "x-ratelimit-limit-requests-1h": "800",
+            "x-ratelimit-remaining-requests-1h": "750",
+            "x-ratelimit-reset-requests-1h": "3100",
+            "x-ratelimit-limit-tokens": "800000",
+            "x-ratelimit-remaining-tokens": "790000",
+            "x-ratelimit-reset-tokens": "40",
+            "x-ratelimit-limit-tokens-1h": "8000000",
+            "x-ratelimit-remaining-tokens-1h": "7800000",
+            "x-ratelimit-reset-tokens-1h": "3100",
+        }
+        assert is_genuine_nous_rate_limit(headers=headers) is False
+
+    def test_bare_429_with_no_headers_is_upstream(self):
+        from agent.nous_rate_guard import is_genuine_nous_rate_limit
+
+        assert is_genuine_nous_rate_limit(headers=None) is False
+        assert is_genuine_nous_rate_limit(headers={}) is False
+        assert is_genuine_nous_rate_limit(
+            headers={"content-type": "application/json"}
+        ) is False
+
+    def test_exhausted_bucket_with_short_reset_is_not_genuine(self):
+        # remaining == 0 but reset in < 60s: almost certainly a
+        # secondary per-minute throttle that will clear immediately --
+        # not worth tripping the cross-session breaker.
+        from agent.nous_rate_guard import is_genuine_nous_rate_limit
+
+        headers = {
+            "x-ratelimit-limit-requests": "200",
+            "x-ratelimit-remaining-requests": "0",
+            "x-ratelimit-reset-requests": "30",
+        }
+        assert is_genuine_nous_rate_limit(headers=headers) is False
+
+    def test_last_known_state_with_exhausted_bucket_triggers_genuine(self):
+        # Headers on the 429 lack rate-limit info, but the previous
+        # successful response already showed the hourly bucket
+        # exhausted -- the 429 is almost certainly that limit
+        # continuing.
+        from agent.nous_rate_guard import is_genuine_nous_rate_limit
+        from agent.rate_limit_tracker import parse_rate_limit_headers
+
+        prior_headers = {
+            "x-ratelimit-limit-requests-1h": "800",
+            "x-ratelimit-remaining-requests-1h": "0",
+            "x-ratelimit-reset-requests-1h": "2000",
+            "x-ratelimit-limit-requests": "200",
+            "x-ratelimit-remaining-requests": "100",
+            "x-ratelimit-reset-requests": "30",
+            "x-ratelimit-limit-tokens": "800000",
+            "x-ratelimit-remaining-tokens": "700000",
+            "x-ratelimit-reset-tokens": "30",
+            "x-ratelimit-limit-tokens-1h": "8000000",
+            "x-ratelimit-remaining-tokens-1h": "7000000",
+            "x-ratelimit-reset-tokens-1h": "2000",
+        }
+        last_state = parse_rate_limit_headers(prior_headers, provider="nous")
+        assert is_genuine_nous_rate_limit(
+            headers=None, last_known_state=last_state
+        ) is True
+
+    def test_last_known_state_all_healthy_stays_upstream(self):
+        # Prior state was healthy; bare 429 arrives; should be treated
+        # as upstream capacity.
+        from agent.nous_rate_guard import is_genuine_nous_rate_limit
+        from agent.rate_limit_tracker import parse_rate_limit_headers
+
+        prior_headers = {
+            "x-ratelimit-limit-requests-1h": "800",
+            "x-ratelimit-remaining-requests-1h": "750",
+            "x-ratelimit-reset-requests-1h": "2000",
+            "x-ratelimit-limit-requests": "200",
+            "x-ratelimit-remaining-requests": "180",
+            "x-ratelimit-reset-requests": "30",
+            "x-ratelimit-limit-tokens": "800000",
+            "x-ratelimit-remaining-tokens": "790000",
+            "x-ratelimit-reset-tokens": "30",
+            "x-ratelimit-limit-tokens-1h": "8000000",
+            "x-ratelimit-remaining-tokens-1h": "7900000",
+            "x-ratelimit-reset-tokens-1h": "2000",
+        }
+        last_state = parse_rate_limit_headers(prior_headers, provider="nous")
+        assert is_genuine_nous_rate_limit(
+            headers=None, last_known_state=last_state
+        ) is False
+
+    def test_none_last_state_and_no_headers_is_upstream(self):
+        from agent.nous_rate_guard import is_genuine_nous_rate_limit
+
+        assert is_genuine_nous_rate_limit(
+            headers=None, last_known_state=None
+        ) is False

tests/agent/test_onboarding.py

@@ -0,0 +1,164 @@
+"""Tests for agent/onboarding.py — contextual first-touch hint helpers."""
+
+from __future__ import annotations
+
+import yaml
+import pytest
+
+from agent.onboarding import (
+    BUSY_INPUT_FLAG,
+    TOOL_PROGRESS_FLAG,
+    busy_input_hint_cli,
+    busy_input_hint_gateway,
+    is_seen,
+    mark_seen,
+    tool_progress_hint_cli,
+    tool_progress_hint_gateway,
+)
+
+
+class TestIsSeen:
+    def test_empty_config_unseen(self):
+        assert is_seen({}, BUSY_INPUT_FLAG) is False
+
+    def test_missing_onboarding_unseen(self):
+        assert is_seen({"display": {}}, BUSY_INPUT_FLAG) is False
+
+    def test_onboarding_not_dict_unseen(self):
+        assert is_seen({"onboarding": "nope"}, BUSY_INPUT_FLAG) is False
+
+    def test_seen_dict_missing_flag(self):
+        assert is_seen({"onboarding": {"seen": {}}}, BUSY_INPUT_FLAG) is False
+
+    def test_seen_flag_true(self):
+        cfg = {"onboarding": {"seen": {BUSY_INPUT_FLAG: True}}}
+        assert is_seen(cfg, BUSY_INPUT_FLAG) is True
+
+    def test_seen_flag_falsy(self):
+        cfg = {"onboarding": {"seen": {BUSY_INPUT_FLAG: False}}}
+        assert is_seen(cfg, BUSY_INPUT_FLAG) is False
+
+    def test_other_flags_isolated(self):
+        cfg = {"onboarding": {"seen": {BUSY_INPUT_FLAG: True}}}
+        assert is_seen(cfg, TOOL_PROGRESS_FLAG) is False
+
+
+class TestMarkSeen:
+    def test_creates_missing_file_and_sets_flag(self, tmp_path):
+        cfg_path = tmp_path / "config.yaml"
+        assert mark_seen(cfg_path, BUSY_INPUT_FLAG) is True
+
+        loaded = yaml.safe_load(cfg_path.read_text())
+        assert loaded["onboarding"]["seen"][BUSY_INPUT_FLAG] is True
+
+    def test_preserves_other_config(self, tmp_path):
+        cfg_path = tmp_path / "config.yaml"
+        cfg_path.write_text(yaml.safe_dump({
+            "model": {"default": "claude-sonnet-4.6"},
+            "display": {"skin": "default"},
+        }))
+
+        assert mark_seen(cfg_path, BUSY_INPUT_FLAG) is True
+        loaded = yaml.safe_load(cfg_path.read_text())
+
+        assert loaded["model"]["default"] == "claude-sonnet-4.6"
+        assert loaded["display"]["skin"] == "default"
+        assert loaded["onboarding"]["seen"][BUSY_INPUT_FLAG] is True
+
+    def test_preserves_other_seen_flags(self, tmp_path):
+        cfg_path = tmp_path / "config.yaml"
+        cfg_path.write_text(yaml.safe_dump({
+            "onboarding": {"seen": {TOOL_PROGRESS_FLAG: True}},
+        }))
+
+        assert mark_seen(cfg_path, BUSY_INPUT_FLAG) is True
+        loaded = yaml.safe_load(cfg_path.read_text())
+
+        assert loaded["onboarding"]["seen"][TOOL_PROGRESS_FLAG] is True
+        assert loaded["onboarding"]["seen"][BUSY_INPUT_FLAG] is True
+
+    def test_idempotent(self, tmp_path):
+        cfg_path = tmp_path / "config.yaml"
+        mark_seen(cfg_path, BUSY_INPUT_FLAG)
+        first = cfg_path.read_text()
+
+        # Second call must be a no-op on-disk content (file may be touched,
+        # but the YAML contents should be identical).
+        mark_seen(cfg_path, BUSY_INPUT_FLAG)
+        second = cfg_path.read_text()
+
+        assert yaml.safe_load(first) == yaml.safe_load(second)
+
+    def test_handles_non_dict_onboarding(self, tmp_path):
+        cfg_path = tmp_path / "config.yaml"
+        cfg_path.write_text(yaml.safe_dump({"onboarding": "corrupted"}))
+
+        assert mark_seen(cfg_path, BUSY_INPUT_FLAG) is True
+        loaded = yaml.safe_load(cfg_path.read_text())
+        assert loaded["onboarding"]["seen"][BUSY_INPUT_FLAG] is True
+
+    def test_handles_non_dict_seen(self, tmp_path):
+        cfg_path = tmp_path / "config.yaml"
+        cfg_path.write_text(yaml.safe_dump({"onboarding": {"seen": "corrupted"}}))
+
+        assert mark_seen(cfg_path, BUSY_INPUT_FLAG) is True
+        loaded = yaml.safe_load(cfg_path.read_text())
+        assert loaded["onboarding"]["seen"][BUSY_INPUT_FLAG] is True
+
+
+class TestHintMessages:
+    def test_busy_input_hint_gateway_interrupt(self):
+        msg = busy_input_hint_gateway("interrupt")
+        assert "/busy queue" in msg
+        assert "interrupted" in msg.lower()
+
+    def test_busy_input_hint_gateway_queue(self):
+        msg = busy_input_hint_gateway("queue")
+        assert "/busy interrupt" in msg
+        assert "queued" in msg.lower()
+
+    def test_busy_input_hint_cli_interrupt(self):
+        msg = busy_input_hint_cli("interrupt")
+        assert "/busy queue" in msg
+
+    def test_busy_input_hint_cli_queue(self):
+        msg = busy_input_hint_cli("queue")
+        assert "/busy interrupt" in msg
+
+    def test_tool_progress_hints_mention_verbose(self):
+        assert "/verbose" in tool_progress_hint_gateway()
+        assert "/verbose" in tool_progress_hint_cli()
+
+    def test_hints_are_not_empty(self):
+        for hint in (
+            busy_input_hint_gateway("queue"),
+            busy_input_hint_gateway("interrupt"),
+            busy_input_hint_cli("queue"),
+            busy_input_hint_cli("interrupt"),
+            tool_progress_hint_gateway(),
+            tool_progress_hint_cli(),
+        ):
+            assert hint.strip()
+
+
+class TestRoundTrip:
+    """After mark_seen, is_seen on the re-loaded config must return True."""
+
+    def test_mark_then_is_seen(self, tmp_path):
+        cfg_path = tmp_path / "config.yaml"
+
+        assert mark_seen(cfg_path, BUSY_INPUT_FLAG) is True
+        loaded = yaml.safe_load(cfg_path.read_text())
+
+        assert is_seen(loaded, BUSY_INPUT_FLAG) is True
+        assert is_seen(loaded, TOOL_PROGRESS_FLAG) is False
+
+    def test_mark_both_flags_independently(self, tmp_path):
+        cfg_path = tmp_path / "config.yaml"
+
+        mark_seen(cfg_path, BUSY_INPUT_FLAG)
+        mark_seen(cfg_path, TOOL_PROGRESS_FLAG)
+        loaded = yaml.safe_load(cfg_path.read_text())
+
+        assert is_seen(loaded, BUSY_INPUT_FLAG) is True
+        assert is_seen(loaded, TOOL_PROGRESS_FLAG) is True

tests/agent/test_unsupported_parameter_retry.py

@@ -0,0 +1,201 @@
+"""Regression tests for the generic unsupported-parameter detector in
+``agent.auxiliary_client``.
+
+The original temperature-specific detector (PR #15621) was generalized so the
+same reactive-retry strategy covers any provider that rejects an arbitrary
+request parameter — ``max_tokens``, ``seed``, ``top_p``, future quirks — not
+just ``temperature``. Credit @nicholasrae (PR #15416) for the generalization
+pattern.
+
+These tests lock in:
+  * ``_is_unsupported_parameter_error(exc, param)`` across common phrasings
+  * the back-compat wrapper ``_is_unsupported_temperature_error`` still works
+  * the max_tokens retry branch no longer pops a key that was never set
+    (``max_tokens is None`` gate)
+  * the max_tokens retry branch matches via the generic helper on top of the
+    legacy ``"max_tokens"`` / ``"unsupported_parameter"`` substring checks
+"""
+
+from unittest.mock import patch, MagicMock, AsyncMock
+
+import pytest
+
+from agent.auxiliary_client import (
+    call_llm,
+    async_call_llm,
+    _is_unsupported_parameter_error,
+    _is_unsupported_temperature_error,
+)
+
+
+class TestIsUnsupportedParameterError:
+    """The generic detector must match real provider phrasings for any param."""
+
+    @pytest.mark.parametrize("param,message", [
+        # temperature phrasings (regression coverage via the generic API)
+        ("temperature", "HTTP 400: Unsupported parameter: temperature"),
+        ("temperature", "Error code: 400 - {'error': {'code': 'unsupported_parameter', 'param': 'temperature'}}"),
+        ("temperature", "this model does not support temperature"),
+        # max_tokens phrasings
+        ("max_tokens", "HTTP 400: Unsupported parameter: max_tokens"),
+        ("max_tokens", "Unknown parameter: max_tokens — use max_completion_tokens"),
+        ("max_tokens", "Invalid parameter: max_tokens is not supported"),
+        # arbitrary future params
+        ("seed", "HTTP 400: unrecognized parameter: seed"),
+        ("top_p", "Error: top_p is not supported for this model"),
+    ])
+    def test_matches_real_provider_messages(self, param, message):
+        assert _is_unsupported_parameter_error(RuntimeError(message), param) is True
+
+    @pytest.mark.parametrize("param,message", [
+        # Param not mentioned at all
+        ("temperature", "HTTP 400: max_tokens is too large"),
+        # Param mentioned but not flagged as unsupported
+        ("temperature", "temperature must be between 0 and 2"),
+        # Totally unrelated 400
+        ("max_tokens", "Rate limit exceeded"),
+        # Connection-level errors
+        ("temperature", "Connection reset by peer"),
+    ])
+    def test_does_not_match_unrelated_errors(self, param, message):
+        assert _is_unsupported_parameter_error(RuntimeError(message), param) is False
+
+    def test_empty_param_returns_false(self):
+        assert _is_unsupported_parameter_error(
+            RuntimeError("HTTP 400: Unsupported parameter: temperature"), ""
+        ) is False
+
+    def test_temperature_wrapper_delegates_to_generic(self):
+        """Back-compat: ``_is_unsupported_temperature_error`` still routes through."""
+        msg = "HTTP 400: Unsupported parameter: temperature"
+        assert _is_unsupported_temperature_error(RuntimeError(msg)) is True
+        # And the unrelated-case still holds
+        assert _is_unsupported_temperature_error(
+            RuntimeError("max_tokens is too large")) is False
+
+
+def _dummy_response():
+    """Sentinel — real code calls ``_validate_llm_response`` which we patch out."""
+    return {"ok": True}
+
+
+class TestMaxTokensRetryHardening:
+    """The max_tokens retry branch now (a) gates on ``max_tokens is not None``
+    and (b) also matches the generic phrasings via the helper.
+    """
+
+    def test_sync_max_tokens_retry_skipped_when_max_tokens_is_none(self):
+        """No max_tokens kwarg → must not pop/retry even if the error mentions it.
+
+        Before the hardening, ``kwargs.pop("max_tokens", None)`` was safe but
+        ``kwargs["max_completion_tokens"] = max_tokens`` would set a None
+        value and hit the provider again. The gate skips the whole branch.
+        """
+        client = MagicMock()
+        client.base_url = "https://api.openai.com/v1"
+        err = RuntimeError("HTTP 400: Unsupported parameter: max_tokens")
+        client.chat.completions.create.side_effect = err
+
+        with (
+            patch("agent.auxiliary_client._resolve_task_provider_model",
+                  return_value=("openai-codex", "gpt-5.5", None, None, None)),
+            patch("agent.auxiliary_client._get_cached_client",
+                  return_value=(client, "gpt-5.5")),
+            patch("agent.auxiliary_client._validate_llm_response",
+                  side_effect=lambda resp, _task: resp),
+        ):
+            with pytest.raises(RuntimeError):
+                call_llm(
+                    task="session_search",
+                    messages=[{"role": "user", "content": "hi"}],
+                    temperature=0.3,
+                    # max_tokens omitted on purpose
+                )
+
+        # Only the initial attempt — no retry because the gate blocked it
+        assert client.chat.completions.create.call_count == 1
+
+    def test_sync_max_tokens_retry_matches_generic_phrasing(self):
+        """A 400 saying "Unknown parameter: max_tokens" (not the legacy
+        substring ``"max_tokens"`` bare + no ``unsupported_parameter`` token)
+        now triggers the retry via the generic helper.
+        """
+        client = MagicMock()
+        client.base_url = "https://api.openai.com/v1"
+        err = RuntimeError("Unknown parameter: max_tokens")
+        response = _dummy_response()
+        client.chat.completions.create.side_effect = [err, response]
+
+        with (
+            patch("agent.auxiliary_client._resolve_task_provider_model",
+                  return_value=("openai-codex", "gpt-5.5", None, None, None)),
+            patch("agent.auxiliary_client._get_cached_client",
+                  return_value=(client, "gpt-5.5")),
+            patch("agent.auxiliary_client._validate_llm_response",
+                  side_effect=lambda resp, _task: resp),
+        ):
+            result = call_llm(
+                task="session_search",
+                messages=[{"role": "user", "content": "hi"}],
+                temperature=0.3,
+                max_tokens=512,
+            )
+
+        assert result is response
+        assert client.chat.completions.create.call_count == 2
+        second_call = client.chat.completions.create.call_args_list[1]
+        assert "max_tokens" not in second_call.kwargs
+        assert second_call.kwargs["max_completion_tokens"] == 512
+
+    @pytest.mark.asyncio
+    async def test_async_max_tokens_retry_skipped_when_max_tokens_is_none(self):
+        client = MagicMock()
+        client.base_url = "https://api.openai.com/v1"
+        err = RuntimeError("HTTP 400: Unsupported parameter: max_tokens")
+        client.chat.completions.create = AsyncMock(side_effect=err)
+
+        with (
+            patch("agent.auxiliary_client._resolve_task_provider_model",
+                  return_value=("openai-codex", "gpt-5.5", None, None, None)),
+            patch("agent.auxiliary_client._get_cached_client",
+                  return_value=(client, "gpt-5.5")),
+            patch("agent.auxiliary_client._validate_llm_response",
+                  side_effect=lambda resp, _task: resp),
+        ):
+            with pytest.raises(RuntimeError):
+                await async_call_llm(
+                    task="session_search",
+                    messages=[{"role": "user", "content": "hi"}],
+                    temperature=0.3,
+                )
+
+        assert client.chat.completions.create.call_count == 1
+
+    @pytest.mark.asyncio
+    async def test_async_max_tokens_retry_matches_generic_phrasing(self):
+        client = MagicMock()
+        client.base_url = "https://api.openai.com/v1"
+        err = RuntimeError("Unknown parameter: max_tokens")
+        response = _dummy_response()
+        client.chat.completions.create = AsyncMock(side_effect=[err, response])
+
+        with (
+            patch("agent.auxiliary_client._resolve_task_provider_model",
+                  return_value=("openai-codex", "gpt-5.5", None, None, None)),
+            patch("agent.auxiliary_client._get_cached_client",
+                  return_value=(client, "gpt-5.5")),
+            patch("agent.auxiliary_client._validate_llm_response",
+                  side_effect=lambda resp, _task: resp),
+        ):
+            result = await async_call_llm(
+                task="session_search",
+                messages=[{"role": "user", "content": "hi"}],
+                temperature=0.3,
+                max_tokens=512,
+            )
+
+        assert result is response
+        assert client.chat.completions.create.await_count == 2
+        second_call = client.chat.completions.create.call_args_list[1]
+        assert "max_tokens" not in second_call.kwargs
+        assert second_call.kwargs["max_completion_tokens"] == 512

tests/agent/test_unsupported_temperature_retry.py

@@ -0,0 +1,237 @@
+"""Regression tests for the universal "unsupported temperature" retry in
+``agent.auxiliary_client``.
+
+Auxiliary callers (context compression, session search,
+web extract summarisation, etc.) hardcode ``temperature=0.3`` for historical
+reasons. Several provider/model combinations reject ``temperature`` with a
+400:
+
+  * OpenAI Responses (gpt-5/o-series reasoning models)
+  * Copilot Responses (reasoning models)
+  * OpenRouter reasoning models (gpt-5.5, some anthropic via OAI-compat)
+  * Anthropic Opus 4.7+ via OpenAI-compat endpoints
+  * Kimi/Moonshot (server-managed)
+
+``_fixed_temperature_for_model`` catches Kimi up front, and
+``build_chat_completion_kwargs`` drops temperature for Anthropic Opus 4.7+,
+but the same backend can accept ``temperature`` for some models and reject
+it for others (for example gpt-5.4 accepts but gpt-5.5 rejects on the same
+endpoint). An allow/deny-list is not maintainable across providers.
+
+The universal fix is reactive: when a call returns an
+``Unsupported parameter: temperature`` 400, retry once without temperature.
+These tests lock in that behaviour for both sync and async paths.
+"""
+
+from unittest.mock import patch, MagicMock, AsyncMock
+
+import pytest
+
+from agent.auxiliary_client import (
+    call_llm,
+    async_call_llm,
+    _is_unsupported_temperature_error,
+)
+
+
+class TestIsUnsupportedTemperatureError:
+    """The detector must match the phrasings providers actually return."""
+
+    @pytest.mark.parametrize("message", [
+        # OpenAI / Codex Responses
+        "HTTP 400: Unsupported parameter: temperature",
+        "Error code: 400 - {'error': {'message': \"Unsupported parameter: 'temperature'\"}}",
+        # Copilot / OpenAI error-code form
+        "Error code: 400 - {'error': {'code': 'unsupported_parameter', 'param': 'temperature'}}",
+        # OpenRouter-style
+        "Provider returned error: temperature is not supported for this model",
+        "this model does not support temperature",
+        # Anthropic-style via OAI-compat
+        "temperature: unknown parameter",
+        # Some gateways
+        "unrecognized request argument supplied: temperature",
+    ])
+    def test_matches_real_provider_messages(self, message):
+        assert _is_unsupported_temperature_error(RuntimeError(message)) is True
+
+    @pytest.mark.parametrize("message", [
+        # Unrelated 400s must NOT trigger a silent-retry
+        "HTTP 400: Invalid value: 'tool'. Supported values are: 'assistant'...",
+        "max_tokens is too large for this model",
+        "Rate limit exceeded",
+        "Connection reset by peer",
+        # Temperature value error is a different class of problem
+        "temperature must be between 0 and 2",
+    ])
+    def test_does_not_match_unrelated_errors(self, message):
+        assert _is_unsupported_temperature_error(RuntimeError(message)) is False
+
+
+def _dummy_response():
+    # The real code calls _validate_llm_response which inspects
+    # response.choices[0].message.  The tests here patch that out, so
+    # any sentinel object is fine.
+    return {"ok": True}
+
+
+class TestCallLlmUnsupportedTemperatureRetry:
+    """``call_llm`` retries once without temperature and returns on success."""
+
+    def _setup(self, first_exc):
+        client = MagicMock()
+        client.base_url = "https://api.openai.com/v1"
+        client.chat.completions.create.side_effect = [first_exc, _dummy_response()]
+        return client
+
+    @pytest.mark.parametrize("error_message", [
+        "HTTP 400: Unsupported parameter: temperature",
+        "Error code: 400 - {'error': {'code': 'unsupported_parameter', 'param': 'temperature'}}",
+        "Provider error: this model does not support temperature",
+    ])
+    def test_retries_once_without_temperature(self, error_message):
+        client = self._setup(RuntimeError(error_message))
+
+        with (
+            patch("agent.auxiliary_client._resolve_task_provider_model",
+                  return_value=("openai-codex", "gpt-5.5", None, None, None)),
+            patch("agent.auxiliary_client._get_cached_client",
+                  return_value=(client, "gpt-5.5")),
+            patch("agent.auxiliary_client._validate_llm_response",
+                  side_effect=lambda resp, _task: resp),
+        ):
+            result = call_llm(
+                task="compression",
+                messages=[{"role": "user", "content": "remember this"}],
+                temperature=0.3,
+                max_tokens=500,
+            )
+
+        assert result == {"ok": True}
+        assert client.chat.completions.create.call_count == 2
+        first_kwargs = client.chat.completions.create.call_args_list[0].kwargs
+        retry_kwargs = client.chat.completions.create.call_args_list[1].kwargs
+        assert first_kwargs["temperature"] == 0.3
+        assert "temperature" not in retry_kwargs
+        # other kwargs preserved
+        assert retry_kwargs["max_tokens"] == 500
+
+    def test_non_temperature_400_does_not_retry_as_temperature(self):
+        """Unrelated 400s (e.g. bad tool role) must not silently drop temp."""
+        client = MagicMock()
+        client.base_url = "https://api.openai.com/v1"
+        non_temp_err = RuntimeError(
+            "HTTP 400: Invalid value: 'tool'. Supported values are: 'assistant'..."
+        )
+        client.chat.completions.create.side_effect = non_temp_err
+
+        with (
+            patch("agent.auxiliary_client._resolve_task_provider_model",
+                  return_value=("openai-codex", "gpt-5.5", None, None, None)),
+            patch("agent.auxiliary_client._get_cached_client",
+                  return_value=(client, "gpt-5.5")),
+            patch("agent.auxiliary_client._validate_llm_response",
+                  side_effect=lambda resp, _task: resp),
+            patch("agent.auxiliary_client._try_payment_fallback",
+                  return_value=None),
+        ):
+            with pytest.raises(RuntimeError, match="Invalid value"):
+                call_llm(
+                    task="compression",
+                    messages=[{"role": "user", "content": "x"}],
+                    temperature=0.3,
+                    max_tokens=500,
+                )
+        # Should NOT have retried (non-temperature 400 doesn't match)
+        assert client.chat.completions.create.call_count == 1
+
+    def test_no_retry_when_temperature_not_in_kwargs(self):
+        """If caller didn't send temperature, don't invent a temperature-retry."""
+        client = MagicMock()
+        client.base_url = "https://api.openai.com/v1"
+        # Provider complains about temperature even though we didn't send it.
+        # (Pathological but possible with misleading error text.)  The guard
+        # ``"temperature" in kwargs`` must prevent an unnecessary retry.
+        err = RuntimeError("HTTP 400: Unsupported parameter: temperature")
+        client.chat.completions.create.side_effect = err
+
+        with (
+            patch("agent.auxiliary_client._resolve_task_provider_model",
+                  return_value=("openai-codex", "gpt-5.5", None, None, None)),
+            patch("agent.auxiliary_client._get_cached_client",
+                  return_value=(client, "gpt-5.5")),
+            patch("agent.auxiliary_client._validate_llm_response",
+                  side_effect=lambda resp, _task: resp),
+            patch("agent.auxiliary_client._try_payment_fallback",
+                  return_value=None),
+        ):
+            with pytest.raises(RuntimeError):
+                call_llm(
+                    task="compression",
+                    messages=[{"role": "user", "content": "x"}],
+                    temperature=None,  # explicit: no temperature sent
+                    max_tokens=500,
+                )
+        assert client.chat.completions.create.call_count == 1
+
+
+class TestAsyncCallLlmUnsupportedTemperatureRetry:
+    """``async_call_llm`` mirror of the sync retry semantics."""
+
+    @pytest.mark.asyncio
+    async def test_async_retries_once_without_temperature(self):
+        client = MagicMock()
+        client.base_url = "https://api.openai.com/v1"
+        client.chat.completions.create = AsyncMock(side_effect=[
+            RuntimeError("HTTP 400: Unsupported parameter: temperature"),
+            _dummy_response(),
+        ])
+
+        with (
+            patch("agent.auxiliary_client._resolve_task_provider_model",
+                  return_value=("openai-codex", "gpt-5.5", None, None, None)),
+            patch("agent.auxiliary_client._get_cached_client",
+                  return_value=(client, "gpt-5.5")),
+            patch("agent.auxiliary_client._validate_llm_response",
+                  side_effect=lambda resp, _task: resp),
+        ):
+            result = await async_call_llm(
+                task="session_search",
+                messages=[{"role": "user", "content": "query"}],
+                temperature=0.3,
+                max_tokens=500,
+            )
+
+        assert result == {"ok": True}
+        assert client.chat.completions.create.await_count == 2
+        first_kwargs = client.chat.completions.create.call_args_list[0].kwargs
+        retry_kwargs = client.chat.completions.create.call_args_list[1].kwargs
+        assert first_kwargs["temperature"] == 0.3
+        assert "temperature" not in retry_kwargs
+        assert retry_kwargs["max_tokens"] == 500
+
+    @pytest.mark.asyncio
+    async def test_async_non_temperature_400_does_not_retry(self):
+        client = MagicMock()
+        client.base_url = "https://api.openai.com/v1"
+        client.chat.completions.create = AsyncMock(
+            side_effect=RuntimeError("HTTP 400: Invalid value: 'tool'"),
+        )
+
+        with (
+            patch("agent.auxiliary_client._resolve_task_provider_model",
+                  return_value=("openai-codex", "gpt-5.5", None, None, None)),
+            patch("agent.auxiliary_client._get_cached_client",
+                  return_value=(client, "gpt-5.5")),
+            patch("agent.auxiliary_client._validate_llm_response",
+                  side_effect=lambda resp, _task: resp),
+            patch("agent.auxiliary_client._try_payment_fallback",
+                  return_value=None),
+        ):
+            with pytest.raises(RuntimeError, match="Invalid value"):
+                await async_call_llm(
+                    task="session_search",
+                    messages=[{"role": "user", "content": "x"}],
+                    temperature=0.3,
+                    max_tokens=500,
+                )
+        assert client.chat.completions.create.await_count == 1

tests/agent/transports/test_chat_completions.py

@@ -33,15 +33,18 @@ class TestChatCompletionsBasic:
     def test_convert_messages_strips_codex_fields(self, transport):
         msgs = [
             {"role": "assistant", "content": "ok", "codex_reasoning_items": [{"id": "rs_1"}],
+             "codex_message_items": [{"id": "msg_1", "type": "message"}],
              "tool_calls": [{"id": "call_1", "call_id": "call_1", "response_item_id": "fc_1",
                             "type": "function", "function": {"name": "t", "arguments": "{}"}}]},
         ]
         result = transport.convert_messages(msgs)
         assert "codex_reasoning_items" not in result[0]
+        assert "codex_message_items" not in result[0]
         assert "call_id" not in result[0]["tool_calls"][0]
         assert "response_item_id" not in result[0]["tool_calls"][0]
         # Original list untouched (deepcopy-on-demand)
         assert "codex_reasoning_items" in msgs[0]
+        assert "codex_message_items" in msgs[0]
 
 
 class TestChatCompletionsBuildKwargs:

tests/agent/transports/test_codex_transport.py

@@ -194,6 +194,36 @@ class TestCodexNormalizeResponse:
         assert nr.content == "Hello world"
         assert nr.finish_reason == "stop"
 
+    def test_message_items_preserved_in_provider_data(self, transport):
+        """Codex assistant message item ids/phases must survive transport normalization."""
+        r = SimpleNamespace(
+            output=[
+                SimpleNamespace(
+                    type="message",
+                    role="assistant",
+                    id="msg_abc",
+                    phase="final_answer",
+                    content=[SimpleNamespace(type="output_text", text="Hello world")],
+                    status="completed",
+                ),
+            ],
+            status="completed",
+            incomplete_details=None,
+            usage=SimpleNamespace(input_tokens=10, output_tokens=5,
+                                  input_tokens_details=None, output_tokens_details=None),
+        )
+        nr = transport.normalize_response(r)
+        assert nr.codex_message_items == [
+            {
+                "type": "message",
+                "role": "assistant",
+                "status": "completed",
+                "content": [{"type": "output_text", "text": "Hello world"}],
+                "id": "msg_abc",
+                "phase": "final_answer",
+            }
+        ]
+
     def test_tool_call_response(self, transport):
         """Normalize a Codex response with tool calls."""
         r = SimpleNamespace(

tests/agent/transports/test_transport.py

@@ -60,6 +60,13 @@ class TestTransportRegistry:
         assert t is not None
         assert t.api_mode == "anthropic_messages"
 
+    def test_discovers_missing_transport_when_registry_partially_populated(self):
+        """Importing one transport directly must not hide other valid api_modes."""
+        import agent.transports.chat_completions  # noqa: F401
+        t = get_transport("codex_responses")
+        assert t is not None
+        assert t.api_mode == "codex_responses"
+
     def test_register_and_get(self):
         class DummyTransport(ProviderTransport):
             @property

tests/agent/transports/test_types.py

@@ -270,3 +270,15 @@ class TestNormalizedResponseBackwardCompat:
     def test_codex_reasoning_items_none_when_absent(self):
         nr = NormalizedResponse(content="hi", tool_calls=None, finish_reason="stop")
         assert nr.codex_reasoning_items is None
+
+    def test_codex_message_items_from_provider_data(self):
+        items = [{"id": "msg_1", "type": "message"}]
+        nr = NormalizedResponse(
+            content="hi", tool_calls=None, finish_reason="stop",
+            provider_data={"codex_message_items": items},
+        )
+        assert nr.codex_message_items == items
+
+    def test_codex_message_items_none_when_absent(self):
+        nr = NormalizedResponse(content="hi", tool_calls=None, finish_reason="stop")
+        assert nr.codex_message_items is None

tests/cli/test_cli_new_session.py

@@ -33,7 +33,6 @@ class _FakeAgent:
         self._todo_store.write(
             [{"id": "t1", "content": "unfinished task", "status": "in_progress"}]
         )
-        self.flush_memories = MagicMock()
         self.commit_memory_session = MagicMock()
         self._invalidate_system_prompt = MagicMock()
 
@@ -157,7 +156,6 @@ def test_new_command_creates_real_fresh_session_and_resets_agent_state(tmp_path)
     assert cli.agent._todo_store.read() == []
     assert cli.session_start > old_session_start
     assert cli.agent.session_start == cli.session_start
-    cli.agent.flush_memories.assert_called_once_with([{"role": "user", "content": "hello"}])
     cli.agent._invalidate_system_prompt.assert_called_once()
 
 

tests/cron/test_cron_context_from.py

@@ -0,0 +1,390 @@
+"""Tests for cron job context_from feature (issue #5439 Option C)."""
+
+import sys
+from pathlib import Path
+
+import pytest
+
+sys.path.insert(0, str(Path(__file__).parent.parent.parent))
+
+
+@pytest.fixture
+def cron_env(tmp_path, monkeypatch):
+    """Isolated cron environment with temp HERMES_HOME."""
+    hermes_home = tmp_path / ".hermes"
+    hermes_home.mkdir()
+    (hermes_home / "cron").mkdir()
+    (hermes_home / "cron" / "output").mkdir()
+    monkeypatch.setenv("HERMES_HOME", str(hermes_home))
+
+    import cron.jobs as jobs_mod
+    monkeypatch.setattr(jobs_mod, "HERMES_DIR", hermes_home)
+    monkeypatch.setattr(jobs_mod, "CRON_DIR", hermes_home / "cron")
+    monkeypatch.setattr(jobs_mod, "JOBS_FILE", hermes_home / "cron" / "jobs.json")
+    monkeypatch.setattr(jobs_mod, "OUTPUT_DIR", hermes_home / "cron" / "output")
+
+    return hermes_home
+
+
+class TestJobContextFromField:
+    """Test that context_from is stored and retrieved correctly."""
+
+    def test_create_job_with_context_from_string(self, cron_env):
+        from cron.jobs import create_job, get_job
+
+        job_a = create_job(prompt="Find news", schedule="every 1h")
+        job_b = create_job(
+            prompt="Summarize findings",
+            schedule="every 2h",
+            context_from=job_a["id"],
+        )
+
+        assert job_b["context_from"] == [job_a["id"]]
+        loaded = get_job(job_b["id"])
+        assert loaded["context_from"] == [job_a["id"]]
+
+    def test_create_job_with_context_from_list(self, cron_env):
+        from cron.jobs import create_job, get_job
+
+        job_a = create_job(prompt="Find news", schedule="every 1h")
+        job_b = create_job(prompt="Find weather", schedule="every 1h")
+        job_c = create_job(
+            prompt="Summarize everything",
+            schedule="every 2h",
+            context_from=[job_a["id"], job_b["id"]],
+        )
+
+        assert job_c["context_from"] == [job_a["id"], job_b["id"]]
+
+    def test_create_job_without_context_from(self, cron_env):
+        from cron.jobs import create_job
+
+        job = create_job(prompt="Hello", schedule="every 1h")
+        assert job.get("context_from") is None
+
+    def test_context_from_empty_string_normalized_to_none(self, cron_env):
+        from cron.jobs import create_job
+
+        job = create_job(prompt="Hello", schedule="every 1h", context_from="")
+        assert job.get("context_from") is None
+
+    def test_context_from_empty_list_normalized_to_none(self, cron_env):
+        from cron.jobs import create_job
+
+        job = create_job(prompt="Hello", schedule="every 1h", context_from=[])
+        assert job.get("context_from") is None
+
+
+class TestBuildJobPromptContextFrom:
+    """Test that _build_job_prompt() injects context from referenced jobs."""
+
+    def test_injects_latest_output(self, cron_env):
+        from cron.jobs import create_job, OUTPUT_DIR
+        from cron.scheduler import _build_job_prompt
+
+        job_a = create_job(prompt="Find news", schedule="every 1h")
+
+        # Записываем output для job_a
+        output_dir = OUTPUT_DIR / job_a["id"]
+        output_dir.mkdir(parents=True, exist_ok=True)
+        (output_dir / "2026-04-22_10-00-00.md").write_text(
+            "Today's top story: AI is everywhere.", encoding="utf-8"
+        )
+
+        job_b = create_job(
+            prompt="Summarize the news",
+            schedule="every 2h",
+            context_from=job_a["id"],
+        )
+
+        prompt = _build_job_prompt(job_b)
+        assert "Today's top story: AI is everywhere." in prompt
+        assert f"Output from job '{job_a['id']}'" in prompt
+
+    def test_uses_most_recent_output(self, cron_env):
+        from cron.jobs import create_job, OUTPUT_DIR
+        from cron.scheduler import _build_job_prompt
+        import time
+
+        job_a = create_job(prompt="Find news", schedule="every 1h")
+        output_dir = OUTPUT_DIR / job_a["id"]
+        output_dir.mkdir(parents=True, exist_ok=True)
+
+        old_file = output_dir / "2026-04-22_08-00-00.md"
+        old_file.write_text("Old output", encoding="utf-8")
+        time.sleep(0.01)
+        new_file = output_dir / "2026-04-22_10-00-00.md"
+        new_file.write_text("New output", encoding="utf-8")
+
+        job_b = create_job(
+            prompt="Summarize", schedule="every 2h", context_from=job_a["id"]
+        )
+        prompt = _build_job_prompt(job_b)
+        assert "New output" in prompt
+        assert "Old output" not in prompt
+
+    def test_graceful_when_no_output_yet(self, cron_env):
+        from cron.jobs import create_job
+        from cron.scheduler import _build_job_prompt
+
+        job_a = create_job(prompt="Find news", schedule="every 1h")
+        job_b = create_job(
+            prompt="Summarize", schedule="every 2h", context_from=job_a["id"]
+        )
+
+        # job_a never ran — output dir does not exist
+        # expect silent skip: no placeholder injected, base prompt intact
+        prompt = _build_job_prompt(job_b)
+        assert "no output" not in prompt.lower()
+        assert "not found" not in prompt.lower()
+        assert "Summarize" in prompt
+
+    def test_injects_multiple_context_jobs(self, cron_env):
+        from cron.jobs import create_job, OUTPUT_DIR
+        from cron.scheduler import _build_job_prompt
+
+        job_a = create_job(prompt="Find news", schedule="every 1h")
+        job_b = create_job(prompt="Find weather", schedule="every 1h")
+
+        for job, content in [(job_a, "News: AI boom"), (job_b, "Weather: Sunny")]:
+            out_dir = OUTPUT_DIR / job["id"]
+            out_dir.mkdir(parents=True, exist_ok=True)
+            (out_dir / "2026-04-22_10-00-00.md").write_text(content, encoding="utf-8")
+
+        job_c = create_job(
+            prompt="Daily briefing",
+            schedule="every 2h",
+            context_from=[job_a["id"], job_b["id"]],
+        )
+        prompt = _build_job_prompt(job_c)
+        assert "News: AI boom" in prompt
+        assert "Weather: Sunny" in prompt
+
+    def test_context_injected_before_prompt(self, cron_env):
+        """Context should appear before the job's own prompt."""
+        from cron.jobs import create_job, OUTPUT_DIR
+        from cron.scheduler import _build_job_prompt
+
+        job_a = create_job(prompt="Find data", schedule="every 1h")
+        out_dir = OUTPUT_DIR / job_a["id"]
+        out_dir.mkdir(parents=True, exist_ok=True)
+        (out_dir / "2026-04-22_10-00-00.md").write_text("Context data", encoding="utf-8")
+
+        job_b = create_job(
+            prompt="Process the data above",
+            schedule="every 2h",
+            context_from=job_a["id"],
+        )
+        prompt = _build_job_prompt(job_b)
+        context_pos = prompt.find("Context data")
+        prompt_pos = prompt.find("Process the data above")
+        assert context_pos < prompt_pos
+
+    def test_output_truncated_at_8k_chars(self, cron_env):
+        """Output longer than 8000 chars should be truncated."""
+        from cron.jobs import create_job, OUTPUT_DIR
+        from cron.scheduler import _build_job_prompt
+
+        job_a = create_job(prompt="Find data", schedule="every 1h")
+        out_dir = OUTPUT_DIR / job_a["id"]
+        out_dir.mkdir(parents=True, exist_ok=True)
+        big_output = "x" * 10000
+        (out_dir / "2026-04-22_10-00-00.md").write_text(big_output, encoding="utf-8")
+
+        job_b = create_job(
+            prompt="Process", schedule="every 2h", context_from=job_a["id"]
+        )
+        prompt = _build_job_prompt(job_b)
+        assert "truncated" in prompt
+        assert "x" * 10000 not in prompt
+
+    def test_graceful_when_file_deleted_between_listing_and_reading(self, cron_env):
+        """Job should not crash if output file is deleted mid-read."""
+        from cron.jobs import create_job, OUTPUT_DIR
+        from cron.scheduler import _build_job_prompt
+        from unittest.mock import patch
+
+        job_a = create_job(prompt="Find data", schedule="every 1h")
+        out_dir = OUTPUT_DIR / job_a["id"]
+        out_dir.mkdir(parents=True, exist_ok=True)
+        (out_dir / "2026-04-22_10-00-00.md").write_text("Some output", encoding="utf-8")
+
+        job_b = create_job(
+            prompt="Process", schedule="every 2h", context_from=job_a["id"]
+        )
+
+        # Simulate file deleted between glob() and read_text()
+        original_read = Path.read_text
+        def mock_read_text(self, *args, **kwargs):
+            if self.suffix == ".md":
+                raise FileNotFoundError("file deleted mid-read")
+            return original_read(self, *args, **kwargs)
+
+        with patch.object(Path, "read_text", mock_read_text):
+            prompt = _build_job_prompt(job_b)
+
+        # Job should not crash, prompt should still contain the base prompt
+        assert "Process" in prompt
+
+    def test_graceful_when_permission_error(self, cron_env):
+        """Job should not crash if output directory is not readable."""
+        from cron.jobs import create_job, OUTPUT_DIR
+        from cron.scheduler import _build_job_prompt
+        from unittest.mock import patch
+
+        job_a = create_job(prompt="Find data", schedule="every 1h")
+        out_dir = OUTPUT_DIR / job_a["id"]
+        out_dir.mkdir(parents=True, exist_ok=True)
+        (out_dir / "2026-04-22_10-00-00.md").write_text("Some output", encoding="utf-8")
+
+        job_b = create_job(
+            prompt="Process", schedule="every 2h", context_from=job_a["id"]
+        )
+
+        # Simulate permission error on read
+        original_read = Path.read_text
+        def mock_read_text(self, *args, **kwargs):
+            if self.suffix == ".md":
+                raise PermissionError("permission denied")
+            return original_read(self, *args, **kwargs)
+
+        with patch.object(Path, "read_text", mock_read_text):
+            prompt = _build_job_prompt(job_b)
+
+        # Job should not crash, prompt should still contain the base prompt
+        assert "Process" in prompt
+
+    def test_invalid_job_id_skipped(self, cron_env):
+        """context_from with path traversal job_id should be skipped."""
+        from cron.jobs import create_job
+        from cron.scheduler import _build_job_prompt
+
+        job = create_job(prompt="Process", schedule="every 2h")
+        # Manually inject invalid context_from (simulating tampered jobs.json)
+        job["context_from"] = ["../../../etc/passwd"]
+        prompt = _build_job_prompt(job)
+        # Should not crash and should not inject anything malicious
+        assert "Process" in prompt
+        assert "etc/passwd" not in prompt
+
+
+
+class TestUpdateContextFrom:
+    """Verify the cronjob tool's `update` action wires context_from through.
+
+    Without this, the create-path stores the field but users can never modify
+    or clear it via the tool (schema promises "pass an empty array to clear").
+    """
+
+    def test_update_adds_context_from_to_existing_job(self, cron_env):
+        from cron.jobs import create_job, get_job
+        from tools.cronjob_tools import cronjob
+        import json
+
+        job_a = create_job(prompt="Find news", schedule="every 1h")
+        job_b = create_job(prompt="Summarize", schedule="every 2h")
+        assert job_b.get("context_from") is None
+
+        result = json.loads(cronjob(
+            action="update",
+            job_id=job_b["id"],
+            context_from=job_a["id"],
+        ))
+        assert result["success"] is True
+
+        reloaded = get_job(job_b["id"])
+        assert reloaded["context_from"] == [job_a["id"]]
+
+    def test_update_changes_context_from_reference(self, cron_env):
+        from cron.jobs import create_job, get_job
+        from tools.cronjob_tools import cronjob
+        import json
+
+        job_a = create_job(prompt="Find news", schedule="every 1h")
+        job_a2 = create_job(prompt="Find weather", schedule="every 1h")
+        job_b = create_job(
+            prompt="Summarize", schedule="every 2h", context_from=job_a["id"],
+        )
+        assert job_b["context_from"] == [job_a["id"]]
+
+        result = json.loads(cronjob(
+            action="update",
+            job_id=job_b["id"],
+            context_from=[job_a2["id"]],
+        ))
+        assert result["success"] is True
+        assert get_job(job_b["id"])["context_from"] == [job_a2["id"]]
+
+    def test_update_clears_context_from_with_empty_list(self, cron_env):
+        from cron.jobs import create_job, get_job
+        from tools.cronjob_tools import cronjob
+        import json
+
+        job_a = create_job(prompt="Find news", schedule="every 1h")
+        job_b = create_job(
+            prompt="Summarize", schedule="every 2h", context_from=job_a["id"],
+        )
+        assert get_job(job_b["id"])["context_from"] == [job_a["id"]]
+
+        result = json.loads(cronjob(
+            action="update",
+            job_id=job_b["id"],
+            context_from=[],
+        ))
+        assert result["success"] is True
+        assert get_job(job_b["id"])["context_from"] is None
+
+    def test_update_clears_context_from_with_empty_string(self, cron_env):
+        from cron.jobs import create_job, get_job
+        from tools.cronjob_tools import cronjob
+        import json
+
+        job_a = create_job(prompt="Find news", schedule="every 1h")
+        job_b = create_job(
+            prompt="Summarize", schedule="every 2h", context_from=job_a["id"],
+        )
+
+        result = json.loads(cronjob(
+            action="update",
+            job_id=job_b["id"],
+            context_from="",
+        ))
+        assert result["success"] is True
+        assert get_job(job_b["id"])["context_from"] is None
+
+    def test_update_rejects_unknown_job_reference(self, cron_env):
+        from cron.jobs import create_job
+        from tools.cronjob_tools import cronjob
+        import json
+
+        job_b = create_job(prompt="Summarize", schedule="every 2h")
+
+        result = json.loads(cronjob(
+            action="update",
+            job_id=job_b["id"],
+            context_from=["deadbeef0000"],
+        ))
+        assert result["success"] is False
+        assert "not found" in result["error"]
+
+    def test_update_preserves_context_from_when_not_passed(self, cron_env):
+        """Updating other fields must not clobber context_from."""
+        from cron.jobs import create_job, get_job
+        from tools.cronjob_tools import cronjob
+        import json
+
+        job_a = create_job(prompt="Find news", schedule="every 1h")
+        job_b = create_job(
+            prompt="Summarize", schedule="every 2h", context_from=job_a["id"],
+        )
+
+        # Update an unrelated field
+        result = json.loads(cronjob(
+            action="update",
+            job_id=job_b["id"],
+            prompt="Summarize v2",
+        ))
+        assert result["success"] is True
+        reloaded = get_job(job_b["id"])
+        assert reloaded["prompt"] == "Summarize v2"
+        assert reloaded["context_from"] == [job_a["id"]]

tests/e2e/conftest.py

@@ -346,6 +346,7 @@ def make_discord_message(
 
     return SimpleNamespace(
         id=message_id, content=content, author=author, channel=channel,
+        guild=getattr(channel, "guild", None),
         mentions=mentions, attachments=attachments,
         type=getattr(discord, "MessageType", SimpleNamespace()).default,
         reference=None, created_at=datetime.now(timezone.utc),

tests/gateway/test_api_server_runs.py

@@ -0,0 +1,365 @@
+"""Tests for /v1/runs endpoints: start, events, and stop.
+
+Covers:
+- POST /v1/runs — start a run (202)
+- GET /v1/runs/{run_id}/events — SSE event stream
+- POST /v1/runs/{run_id}/stop — interrupt a running agent
+- Auth, error handling, and cleanup
+"""
+
+import asyncio
+import json
+import threading
+import time as _time
+from unittest.mock import AsyncMock, MagicMock, patch
+
+import pytest
+from aiohttp import web
+from aiohttp.test_utils import TestClient, TestServer
+
+from gateway.config import PlatformConfig
+from gateway.platforms.api_server import (
+    APIServerAdapter,
+    cors_middleware,
+    security_headers_middleware,
+)
+
+
+# ---------------------------------------------------------------------------
+# Helpers
+# ---------------------------------------------------------------------------
+
+
+def _make_adapter(api_key: str = "") -> APIServerAdapter:
+    """Create an adapter with optional API key."""
+    extra = {}
+    if api_key:
+        extra["key"] = api_key
+    config = PlatformConfig(enabled=True, extra=extra)
+    adapter = APIServerAdapter(config)
+    return adapter
+
+
+def _create_runs_app(adapter: APIServerAdapter) -> web.Application:
+    """Create an aiohttp app with /v1/runs routes registered."""
+    mws = [mw for mw in (cors_middleware, security_headers_middleware) if mw is not None]
+    app = web.Application(middlewares=mws)
+    app["api_server_adapter"] = adapter
+    app.router.add_post("/v1/runs", adapter._handle_runs)
+    app.router.add_get("/v1/runs/{run_id}/events", adapter._handle_run_events)
+    app.router.add_post("/v1/runs/{run_id}/stop", adapter._handle_stop_run)
+    return app
+
+
+def _make_slow_agent(**kwargs):
+    """Create a mock agent that blocks in run_conversation until interrupted.
+
+    Returns (mock_agent, agent_ready_event, interrupt_event) where
+    agent_ready_event is set once run_conversation starts, and
+    interrupt_event is set when interrupt() is called.
+    """
+    ready = threading.Event()
+    interrupted = threading.Event()
+
+    mock_agent = MagicMock()
+
+    def _do_interrupt(message=None):
+        interrupted.set()
+
+    mock_agent.interrupt = MagicMock(side_effect=_do_interrupt)
+
+    def _slow_run(user_message=None, conversation_history=None, task_id=None):
+        ready.set()
+        # Block until interrupt() is called
+        interrupted.wait(timeout=10)
+        return {"final_response": "interrupted"}
+
+    mock_agent.run_conversation.side_effect = _slow_run
+    mock_agent.session_prompt_tokens = 0
+    mock_agent.session_completion_tokens = 0
+    mock_agent.session_total_tokens = 0
+
+    return mock_agent, ready, interrupted
+
+
+@pytest.fixture
+def adapter():
+    return _make_adapter()
+
+
+@pytest.fixture
+def auth_adapter():
+    return _make_adapter(api_key="sk-secret")
+
+
+# ---------------------------------------------------------------------------
+# POST /v1/runs — start a run
+# ---------------------------------------------------------------------------
+
+
+class TestStartRun:
+    @pytest.mark.asyncio
+    async def test_start_returns_202(self, adapter):
+        app = _create_runs_app(adapter)
+        async with TestClient(TestServer(app)) as cli:
+            with patch.object(adapter, "_create_agent") as mock_create:
+                mock_agent = MagicMock()
+                mock_agent.run_conversation.return_value = {"final_response": "done"}
+                mock_agent.session_prompt_tokens = 10
+                mock_agent.session_completion_tokens = 5
+                mock_agent.session_total_tokens = 15
+                mock_create.return_value = mock_agent
+
+                resp = await cli.post("/v1/runs", json={"input": "hello"})
+                assert resp.status == 202
+                data = await resp.json()
+                assert data["status"] == "started"
+                assert data["run_id"].startswith("run_")
+
+    @pytest.mark.asyncio
+    async def test_start_invalid_json_returns_400(self, adapter):
+        app = _create_runs_app(adapter)
+        async with TestClient(TestServer(app)) as cli:
+            resp = await cli.post(
+                "/v1/runs",
+                data="not json",
+                headers={"Content-Type": "application/json"},
+            )
+        assert resp.status == 400
+
+    @pytest.mark.asyncio
+    async def test_start_missing_input_returns_400(self, adapter):
+        app = _create_runs_app(adapter)
+        async with TestClient(TestServer(app)) as cli:
+            resp = await cli.post("/v1/runs", json={"model": "test"})
+            assert resp.status == 400
+            data = await resp.json()
+            assert "input" in data["error"]["message"]
+
+    @pytest.mark.asyncio
+    async def test_start_empty_input_returns_400(self, adapter):
+        app = _create_runs_app(adapter)
+        async with TestClient(TestServer(app)) as cli:
+            resp = await cli.post("/v1/runs", json={"input": ""})
+        assert resp.status == 400
+
+    @pytest.mark.asyncio
+    async def test_start_requires_auth(self, auth_adapter):
+        app = _create_runs_app(auth_adapter)
+        async with TestClient(TestServer(app)) as cli:
+            resp = await cli.post("/v1/runs", json={"input": "hello"})
+        assert resp.status == 401
+
+    @pytest.mark.asyncio
+    async def test_start_with_valid_auth(self, auth_adapter):
+        app = _create_runs_app(auth_adapter)
+        async with TestClient(TestServer(app)) as cli:
+            with patch.object(auth_adapter, "_create_agent") as mock_create:
+                mock_agent = MagicMock()
+                mock_agent.run_conversation.return_value = {"final_response": "ok"}
+                mock_agent.session_prompt_tokens = 0
+                mock_agent.session_completion_tokens = 0
+                mock_agent.session_total_tokens = 0
+                mock_create.return_value = mock_agent
+
+                resp = await cli.post(
+                    "/v1/runs",
+                    json={"input": "hello"},
+                    headers={"Authorization": "Bearer sk-secret"},
+                )
+                assert resp.status == 202
+
+
+# ---------------------------------------------------------------------------
+# GET /v1/runs/{run_id}/events — SSE event stream
+# ---------------------------------------------------------------------------
+
+
+class TestRunEvents:
+    @pytest.mark.asyncio
+    async def test_events_stream_returns_completed(self, adapter):
+        """Events stream should receive run.completed when agent finishes."""
+        app = _create_runs_app(adapter)
+        async with TestClient(TestServer(app)) as cli:
+            with patch.object(adapter, "_create_agent") as mock_create:
+                mock_agent = MagicMock()
+                mock_agent.run_conversation.return_value = {"final_response": "Hello!"}
+                mock_agent.session_prompt_tokens = 10
+                mock_agent.session_completion_tokens = 5
+                mock_agent.session_total_tokens = 15
+                mock_create.return_value = mock_agent
+
+                # Start run
+                resp = await cli.post("/v1/runs", json={"input": "hello"})
+                assert resp.status == 202
+                data = await resp.json()
+                run_id = data["run_id"]
+
+                # Subscribe to events
+                events_resp = await cli.get(f"/v1/runs/{run_id}/events")
+                assert events_resp.status == 200
+                body = await events_resp.text()
+
+                # Should contain run.completed
+                assert "run.completed" in body
+                assert "Hello!" in body
+
+    @pytest.mark.asyncio
+    async def test_events_not_found_returns_404(self, adapter):
+        app = _create_runs_app(adapter)
+        async with TestClient(TestServer(app)) as cli:
+            resp = await cli.get("/v1/runs/run_nonexistent/events")
+        assert resp.status == 404
+
+    @pytest.mark.asyncio
+    async def test_events_requires_auth(self, auth_adapter):
+        app = _create_runs_app(auth_adapter)
+        async with TestClient(TestServer(app)) as cli:
+            resp = await cli.get("/v1/runs/run_any/events")
+        assert resp.status == 401
+
+
+# ---------------------------------------------------------------------------
+# POST /v1/runs/{run_id}/stop — interrupt a running agent
+# ---------------------------------------------------------------------------
+
+
+class TestStopRun:
+    @pytest.mark.asyncio
+    async def test_stop_running_agent(self, adapter):
+        """Stop should interrupt the agent and cancel the task."""
+        app = _create_runs_app(adapter)
+        async with TestClient(TestServer(app)) as cli:
+            with patch.object(adapter, "_create_agent") as mock_create:
+                mock_agent, agent_ready, _ = _make_slow_agent()
+                mock_create.return_value = mock_agent
+
+                # Start run
+                resp = await cli.post("/v1/runs", json={"input": "hello"})
+                assert resp.status == 202
+                data = await resp.json()
+                run_id = data["run_id"]
+
+                # Wait for agent to start running in the thread
+                agent_ready.wait(timeout=3.0)
+                await asyncio.sleep(0.1)
+
+                # Verify agent ref is stored
+                assert run_id in adapter._active_run_agents
+
+                # Stop the run
+                stop_resp = await cli.post(f"/v1/runs/{run_id}/stop")
+                assert stop_resp.status == 200
+                stop_data = await stop_resp.json()
+                assert stop_data["run_id"] == run_id
+                assert stop_data["status"] == "stopping"
+
+                # Agent interrupt should have been called
+                mock_agent.interrupt.assert_called_once_with("Stop requested via API")
+
+                # Refs should be cleaned up
+                await asyncio.sleep(0.5)
+                assert run_id not in adapter._active_run_agents
+                assert run_id not in adapter._active_run_tasks
+
+    @pytest.mark.asyncio
+    async def test_stop_nonexistent_run_returns_404(self, adapter):
+        app = _create_runs_app(adapter)
+        async with TestClient(TestServer(app)) as cli:
+            resp = await cli.post("/v1/runs/run_nonexistent/stop")
+        assert resp.status == 404
+
+    @pytest.mark.asyncio
+    async def test_stop_requires_auth(self, auth_adapter):
+        app = _create_runs_app(auth_adapter)
+        async with TestClient(TestServer(app)) as cli:
+            resp = await cli.post("/v1/runs/run_any/stop")
+        assert resp.status == 401
+
+    @pytest.mark.asyncio
+    async def test_stop_already_completed_run_returns_404(self, adapter):
+        """Stopping a run that already finished should return 404 (refs cleaned up)."""
+        app = _create_runs_app(adapter)
+        async with TestClient(TestServer(app)) as cli:
+            with patch.object(adapter, "_create_agent") as mock_create:
+                mock_agent = MagicMock()
+                mock_agent.run_conversation.return_value = {"final_response": "done"}
+                mock_agent.session_prompt_tokens = 0
+                mock_agent.session_completion_tokens = 0
+                mock_agent.session_total_tokens = 0
+                mock_create.return_value = mock_agent
+
+                # Start and wait for completion
+                resp = await cli.post("/v1/runs", json={"input": "hello"})
+                assert resp.status == 202
+                data = await resp.json()
+                run_id = data["run_id"]
+
+                await asyncio.sleep(0.3)
+
+                # Run should be done, refs cleaned up
+                assert run_id not in adapter._active_run_agents
+
+                # Stop should return 404
+                stop_resp = await cli.post(f"/v1/runs/{run_id}/stop")
+                assert stop_resp.status == 404
+
+    @pytest.mark.asyncio
+    async def test_stop_interrupt_exception_does_not_crash(self, adapter):
+        """If agent.interrupt() raises, stop should still succeed."""
+        app = _create_runs_app(adapter)
+        async with TestClient(TestServer(app)) as cli:
+            with patch.object(adapter, "_create_agent") as mock_create:
+                mock_agent, agent_ready, _ = _make_slow_agent()
+                # Override the interrupt side_effect to raise
+                mock_agent.interrupt = MagicMock(side_effect=RuntimeError("interrupt failed"))
+                mock_create.return_value = mock_agent
+
+                resp = await cli.post("/v1/runs", json={"input": "hello"})
+                assert resp.status == 202
+                data = await resp.json()
+                run_id = data["run_id"]
+
+                agent_ready.wait(timeout=3.0)
+                await asyncio.sleep(0.1)
+
+                stop_resp = await cli.post(f"/v1/runs/{run_id}/stop")
+                assert stop_resp.status == 200
+                stop_data = await stop_resp.json()
+                assert stop_data["status"] == "stopping"
+
+    @pytest.mark.asyncio
+    async def test_stop_sends_sentinel_to_events_stream(self, adapter):
+        """After stop, the events stream should close."""
+        app = _create_runs_app(adapter)
+        async with TestClient(TestServer(app)) as cli:
+            with patch.object(adapter, "_create_agent") as mock_create:
+                mock_agent, agent_ready, _ = _make_slow_agent()
+                mock_create.return_value = mock_agent
+
+                # Start run
+                resp = await cli.post("/v1/runs", json={"input": "hello"})
+                assert resp.status == 202
+                data = await resp.json()
+                run_id = data["run_id"]
+
+                agent_ready.wait(timeout=3.0)
+                await asyncio.sleep(0.1)
+
+                # Subscribe to events in background
+                events_task = asyncio.ensure_future(
+                    cli.get(f"/v1/runs/{run_id}/events")
+                )
+
+                await asyncio.sleep(0.1)
+
+                # Stop the run
+                stop_resp = await cli.post(f"/v1/runs/{run_id}/stop")
+                assert stop_resp.status == 200
+
+                # Events stream should close
+                events_resp = await asyncio.wait_for(events_task, timeout=5.0)
+                assert events_resp.status == 200
+                body = await events_resp.text()
+                # Stream should have received run.failed and closed
+                assert "run.failed" in body or "stream closed" in body

tests/gateway/test_async_memory_flush.py

@@ -1,249 +0,0 @@
-"""Tests for proactive memory flush on session expiry.
-
-Verifies that:
-1. _is_session_expired() works from a SessionEntry alone (no source needed)
-2. The sync callback is no longer called in get_or_create_session
-3. memory_flushed flag persists across save/load cycles (prevents restart re-flush)
-4. The background watcher can detect expired sessions
-"""
-
-import pytest
-from datetime import datetime, timedelta
-from pathlib import Path
-from unittest.mock import patch, MagicMock
-
-from gateway.config import Platform, GatewayConfig, SessionResetPolicy
-from gateway.session import SessionSource, SessionStore, SessionEntry
-
-
-@pytest.fixture()
-def idle_store(tmp_path):
-    """SessionStore with a 60-minute idle reset policy."""
-    config = GatewayConfig(
-        default_reset_policy=SessionResetPolicy(mode="idle", idle_minutes=60),
-    )
-    with patch("gateway.session.SessionStore._ensure_loaded"):
-        s = SessionStore(sessions_dir=tmp_path, config=config)
-    s._db = None
-    s._loaded = True
-    return s
-
-
-@pytest.fixture()
-def no_reset_store(tmp_path):
-    """SessionStore with no reset policy (mode=none)."""
-    config = GatewayConfig(
-        default_reset_policy=SessionResetPolicy(mode="none"),
-    )
-    with patch("gateway.session.SessionStore._ensure_loaded"):
-        s = SessionStore(sessions_dir=tmp_path, config=config)
-    s._db = None
-    s._loaded = True
-    return s
-
-
-class TestIsSessionExpired:
-    """_is_session_expired should detect expiry from entry alone."""
-
-    def test_idle_session_expired(self, idle_store):
-        entry = SessionEntry(
-            session_key="agent:main:telegram:dm",
-            session_id="sid_1",
-            created_at=datetime.now() - timedelta(hours=3),
-            updated_at=datetime.now() - timedelta(minutes=120),
-            platform=Platform.TELEGRAM,
-            chat_type="dm",
-        )
-        assert idle_store._is_session_expired(entry) is True
-
-    def test_active_session_not_expired(self, idle_store):
-        entry = SessionEntry(
-            session_key="agent:main:telegram:dm",
-            session_id="sid_2",
-            created_at=datetime.now() - timedelta(hours=1),
-            updated_at=datetime.now() - timedelta(minutes=10),
-            platform=Platform.TELEGRAM,
-            chat_type="dm",
-        )
-        assert idle_store._is_session_expired(entry) is False
-
-    def test_none_mode_never_expires(self, no_reset_store):
-        entry = SessionEntry(
-            session_key="agent:main:telegram:dm",
-            session_id="sid_3",
-            created_at=datetime.now() - timedelta(days=30),
-            updated_at=datetime.now() - timedelta(days=30),
-            platform=Platform.TELEGRAM,
-            chat_type="dm",
-        )
-        assert no_reset_store._is_session_expired(entry) is False
-
-    def test_active_processes_prevent_expiry(self, idle_store):
-        """Sessions with active background processes should never expire."""
-        idle_store._has_active_processes_fn = lambda key: True
-        entry = SessionEntry(
-            session_key="agent:main:telegram:dm",
-            session_id="sid_4",
-            created_at=datetime.now() - timedelta(hours=5),
-            updated_at=datetime.now() - timedelta(hours=5),
-            platform=Platform.TELEGRAM,
-            chat_type="dm",
-        )
-        assert idle_store._is_session_expired(entry) is False
-
-    def test_daily_mode_expired(self, tmp_path):
-        """Daily mode should expire sessions from before today's reset hour."""
-        config = GatewayConfig(
-            default_reset_policy=SessionResetPolicy(mode="daily", at_hour=4),
-        )
-        with patch("gateway.session.SessionStore._ensure_loaded"):
-            store = SessionStore(sessions_dir=tmp_path, config=config)
-        store._db = None
-        store._loaded = True
-
-        entry = SessionEntry(
-            session_key="agent:main:telegram:dm",
-            session_id="sid_5",
-            created_at=datetime.now() - timedelta(days=2),
-            updated_at=datetime.now() - timedelta(days=2),
-            platform=Platform.TELEGRAM,
-            chat_type="dm",
-        )
-        assert store._is_session_expired(entry) is True
-
-
-class TestGetOrCreateSessionNoCallback:
-    """get_or_create_session should NOT call a sync flush callback."""
-
-    def test_auto_reset_creates_new_session_after_flush(self, idle_store):
-        """When a flushed session auto-resets, a new session_id is created."""
-        source = SessionSource(
-            platform=Platform.TELEGRAM,
-            chat_id="123",
-            chat_type="dm",
-        )
-        # Create initial session
-        entry1 = idle_store.get_or_create_session(source)
-        old_sid = entry1.session_id
-
-        # Simulate the watcher having flushed it
-        entry1.memory_flushed = True
-
-        # Simulate the session going idle
-        entry1.updated_at = datetime.now() - timedelta(minutes=120)
-        idle_store._save()
-
-        # Next call should auto-reset
-        entry2 = idle_store.get_or_create_session(source)
-        assert entry2.session_id != old_sid
-        assert entry2.was_auto_reset is True
-        # New session starts with memory_flushed=False
-        assert entry2.memory_flushed is False
-
-    def test_no_sync_callback_invoked(self, idle_store):
-        """No synchronous callback should block during auto-reset."""
-        source = SessionSource(
-            platform=Platform.TELEGRAM,
-            chat_id="123",
-            chat_type="dm",
-        )
-        entry1 = idle_store.get_or_create_session(source)
-        entry1.updated_at = datetime.now() - timedelta(minutes=120)
-        idle_store._save()
-
-        # Verify no _on_auto_reset attribute
-        assert not hasattr(idle_store, '_on_auto_reset')
-
-        # This should NOT block (no sync LLM call)
-        entry2 = idle_store.get_or_create_session(source)
-        assert entry2.was_auto_reset is True
-
-
-class TestMemoryFlushedFlag:
-    """The memory_flushed flag on SessionEntry prevents double-flushing."""
-
-    def test_defaults_to_false(self):
-        entry = SessionEntry(
-            session_key="agent:main:telegram:dm:123",
-            session_id="sid_new",
-            created_at=datetime.now(),
-            updated_at=datetime.now(),
-            platform=Platform.TELEGRAM,
-            chat_type="dm",
-        )
-        assert entry.memory_flushed is False
-
-    def test_persists_through_save_load(self, idle_store):
-        """memory_flushed=True must survive a save/load cycle (simulates restart)."""
-        key = "agent:main:discord:thread:789"
-        entry = SessionEntry(
-            session_key=key,
-            session_id="sid_flushed",
-            created_at=datetime.now() - timedelta(hours=5),
-            updated_at=datetime.now() - timedelta(hours=5),
-            platform=Platform.DISCORD,
-            chat_type="thread",
-            memory_flushed=True,
-        )
-        idle_store._entries[key] = entry
-        idle_store._save()
-
-        # Simulate restart: clear in-memory state, reload from disk
-        idle_store._entries.clear()
-        idle_store._loaded = False
-        idle_store._ensure_loaded()
-
-        reloaded = idle_store._entries[key]
-        assert reloaded.memory_flushed is True
-
-    def test_unflushed_entry_survives_restart_as_unflushed(self, idle_store):
-        """An entry without memory_flushed stays False after reload."""
-        key = "agent:main:telegram:dm:456"
-        entry = SessionEntry(
-            session_key=key,
-            session_id="sid_not_flushed",
-            created_at=datetime.now() - timedelta(hours=2),
-            updated_at=datetime.now() - timedelta(hours=2),
-            platform=Platform.TELEGRAM,
-            chat_type="dm",
-        )
-        idle_store._entries[key] = entry
-        idle_store._save()
-
-        idle_store._entries.clear()
-        idle_store._loaded = False
-        idle_store._ensure_loaded()
-
-        reloaded = idle_store._entries[key]
-        assert reloaded.memory_flushed is False
-
-    def test_roundtrip_to_dict_from_dict(self):
-        """to_dict/from_dict must preserve memory_flushed."""
-        entry = SessionEntry(
-            session_key="agent:main:telegram:dm:999",
-            session_id="sid_rt",
-            created_at=datetime.now(),
-            updated_at=datetime.now(),
-            platform=Platform.TELEGRAM,
-            chat_type="dm",
-            memory_flushed=True,
-        )
-        d = entry.to_dict()
-        assert d["memory_flushed"] is True
-
-        restored = SessionEntry.from_dict(d)
-        assert restored.memory_flushed is True
-
-    def test_legacy_entry_without_field_defaults_false(self):
-        """Old sessions.json entries missing memory_flushed should default to False."""
-        data = {
-            "session_key": "agent:main:telegram:dm:legacy",
-            "session_id": "sid_legacy",
-            "created_at": datetime.now().isoformat(),
-            "updated_at": datetime.now().isoformat(),
-            "platform": "telegram",
-            "chat_type": "dm",
-            # no memory_flushed key
-        }
-        entry = SessionEntry.from_dict(data)
-        assert entry.memory_flushed is False

tests/gateway/test_busy_session_ack.py

@@ -349,3 +349,121 @@ class TestBusySessionAck:
 
         result = await runner._handle_active_session_busy_message(event, sk)
         assert result is False  # not handled, let default path try
+
+
+class TestBusySessionOnboardingHint:
+    """First-touch hint appended to the busy-ack the first time it fires."""
+
+    @pytest.mark.asyncio
+    async def test_first_busy_ack_appends_interrupt_hint(self, tmp_path, monkeypatch):
+        """First busy-while-running message gets an extra hint about /busy."""
+        import gateway.run as _gr
+
+        monkeypatch.setattr(_gr, "_hermes_home", tmp_path)
+        # mark_seen imports utils.atomic_yaml_write; make sure it resolves
+        # against a writable dir by pointing _hermes_home at tmp_path.
+        monkeypatch.setattr(_gr, "_load_gateway_config", lambda: {})
+
+        runner, _sentinel = _make_runner()
+        runner._busy_input_mode = "interrupt"
+        adapter = _make_adapter()
+
+        event = _make_event(text="ping")
+        sk = build_session_key(event.source)
+
+        agent = MagicMock()
+        agent.get_activity_summary.return_value = {
+            "api_call_count": 3, "max_iterations": 60,
+            "current_tool": None, "last_activity_ts": time.time(),
+            "last_activity_desc": "api", "seconds_since_activity": 0.1,
+        }
+        runner._running_agents[sk] = agent
+        runner._running_agents_ts[sk] = time.time() - 5
+        runner.adapters[event.source.platform] = adapter
+
+        await runner._handle_active_session_busy_message(event, sk)
+
+        call_kwargs = adapter._send_with_retry.call_args
+        content = call_kwargs.kwargs.get("content", "")
+
+        # Normal ack body
+        assert "Interrupting" in content
+        # First-touch hint appended
+        assert "First-time tip" in content
+        assert "/busy queue" in content
+
+        # The flag is now persisted to tmp_path/config.yaml
+        import yaml
+        cfg = yaml.safe_load((tmp_path / "config.yaml").read_text())
+        assert cfg["onboarding"]["seen"]["busy_input_prompt"] is True
+
+    @pytest.mark.asyncio
+    async def test_second_busy_ack_omits_hint(self, tmp_path, monkeypatch):
+        """Once the flag is marked, the hint never appears again."""
+        import gateway.run as _gr
+        import yaml
+
+        monkeypatch.setattr(_gr, "_hermes_home", tmp_path)
+        # Pre-populate the config so is_seen() returns True from the start.
+        (tmp_path / "config.yaml").write_text(yaml.safe_dump({
+            "onboarding": {"seen": {"busy_input_prompt": True}},
+        }))
+        monkeypatch.setattr(
+            _gr, "_load_gateway_config",
+            lambda: yaml.safe_load((tmp_path / "config.yaml").read_text()),
+        )
+
+        runner, _sentinel = _make_runner()
+        runner._busy_input_mode = "interrupt"
+        adapter = _make_adapter()
+
+        event = _make_event(text="ping again")
+        sk = build_session_key(event.source)
+
+        agent = MagicMock()
+        agent.get_activity_summary.return_value = {
+            "api_call_count": 3, "max_iterations": 60,
+            "current_tool": None, "last_activity_ts": time.time(),
+            "last_activity_desc": "api", "seconds_since_activity": 0.1,
+        }
+        runner._running_agents[sk] = agent
+        runner._running_agents_ts[sk] = time.time() - 5
+        runner.adapters[event.source.platform] = adapter
+
+        await runner._handle_active_session_busy_message(event, sk)
+
+        call_kwargs = adapter._send_with_retry.call_args
+        content = call_kwargs.kwargs.get("content", "")
+
+        assert "Interrupting" in content
+        assert "First-time tip" not in content
+        assert "/busy queue" not in content
+
+    @pytest.mark.asyncio
+    async def test_queue_mode_hint_points_to_interrupt(self, tmp_path, monkeypatch):
+        """In queue mode the hint should suggest /busy interrupt, not /busy queue."""
+        import gateway.run as _gr
+
+        monkeypatch.setattr(_gr, "_hermes_home", tmp_path)
+        monkeypatch.setattr(_gr, "_load_gateway_config", lambda: {})
+
+        runner, _sentinel = _make_runner()
+        runner._busy_input_mode = "queue"
+        adapter = _make_adapter()
+
+        event = _make_event(text="queue me")
+        sk = build_session_key(event.source)
+        runner.adapters[event.source.platform] = adapter
+
+        agent = MagicMock()
+        runner._running_agents[sk] = agent
+
+        with patch("gateway.run.merge_pending_message_event"):
+            await runner._handle_active_session_busy_message(event, sk)
+
+        content = adapter._send_with_retry.call_args.kwargs.get("content", "")
+        assert "Queued for the next turn" in content
+        assert "First-time tip" in content
+        assert "/busy interrupt" in content
+        # Must NOT tell the user to /busy queue when they're already on queue.
+        assert "/busy queue" not in content

tests/gateway/test_flush_memory_stale_guard.py

@@ -1,240 +0,0 @@
-"""Tests for memory flush stale-overwrite prevention (#2670).
-
-Verifies that:
-1. Cron sessions are skipped (no flush for headless cron runs)
-2. Current memory state is injected into the flush prompt so the
-   flush agent can see what's already saved and avoid overwrites
-3. The flush still works normally when memory files don't exist
-"""
-
-import sys
-import types
-import pytest
-from pathlib import Path
-from unittest.mock import MagicMock, patch, call
-
-
-@pytest.fixture(autouse=True)
-def _mock_dotenv(monkeypatch):
-    """gateway.run imports dotenv at module level; stub it so tests run without the package."""
-    fake = types.ModuleType("dotenv")
-    fake.load_dotenv = lambda *a, **kw: None
-    monkeypatch.setitem(sys.modules, "dotenv", fake)
-
-
-def _make_runner():
-    from gateway.run import GatewayRunner
-
-    runner = object.__new__(GatewayRunner)
-    runner._honcho_managers = {}
-    runner._honcho_configs = {}
-    runner._running_agents = {}
-    runner._pending_messages = {}
-    runner._pending_approvals = {}
-    runner.adapters = {}
-    runner.hooks = MagicMock()
-    runner.session_store = MagicMock()
-    return runner
-
-
-_TRANSCRIPT_4_MSGS = [
-    {"role": "user", "content": "hello"},
-    {"role": "assistant", "content": "hi there"},
-    {"role": "user", "content": "remember my name is Alice"},
-    {"role": "assistant", "content": "Got it, Alice!"},
-]
-
-
-class TestCronSessionBypass:
-    """Cron sessions should never trigger a memory flush."""
-
-    def test_cron_session_skipped(self):
-        runner = _make_runner()
-        runner._flush_memories_for_session("cron_job123_20260323_120000")
-        # session_store.load_transcript should never be called
-        runner.session_store.load_transcript.assert_not_called()
-
-    def test_cron_session_with_prefix_skipped(self):
-        """Cron sessions with different prefixes are still skipped."""
-        runner = _make_runner()
-        runner._flush_memories_for_session("cron_daily_20260323")
-        runner.session_store.load_transcript.assert_not_called()
-
-    def test_non_cron_session_proceeds(self):
-        """Non-cron sessions should still attempt the flush."""
-        runner = _make_runner()
-        runner.session_store.load_transcript.return_value = []
-        runner._flush_memories_for_session("session_abc123")
-        runner.session_store.load_transcript.assert_called_once_with("session_abc123")
-
-
-def _make_flush_context(monkeypatch, memory_dir=None):
-    """Return (runner, tmp_agent, fake_run_agent) with run_agent mocked in sys.modules."""
-    tmp_agent = MagicMock()
-    fake_run_agent = types.ModuleType("run_agent")
-    fake_run_agent.AIAgent = MagicMock(return_value=tmp_agent)
-    monkeypatch.setitem(sys.modules, "run_agent", fake_run_agent)
-
-    runner = _make_runner()
-    runner.session_store.load_transcript.return_value = _TRANSCRIPT_4_MSGS
-    return runner, tmp_agent, memory_dir
-
-
-class TestMemoryInjection:
-    """The flush prompt should include current memory state from disk."""
-
-    def test_memory_content_injected_into_flush_prompt(self, tmp_path, monkeypatch):
-        """When memory files exist, their content appears in the flush prompt."""
-        memory_dir = tmp_path / "memories"
-        memory_dir.mkdir()
-        (memory_dir / "MEMORY.md").write_text("Agent knows Python\n§\nUser prefers dark mode")
-        (memory_dir / "USER.md").write_text("Name: Alice\n§\nTimezone: PST")
-
-        runner, tmp_agent, _ = _make_flush_context(monkeypatch, memory_dir)
-
-        with (
-            patch("gateway.run._resolve_runtime_agent_kwargs", return_value={"api_key": "k"}),
-            patch("gateway.run._resolve_gateway_model", return_value="test-model"),
-            patch.dict("sys.modules", {"tools.memory_tool": MagicMock(get_memory_dir=lambda: memory_dir)}),
-        ):
-            runner._flush_memories_for_session("session_123")
-
-        tmp_agent.run_conversation.assert_called_once()
-        flush_prompt = tmp_agent.run_conversation.call_args.kwargs.get("user_message", "")
-
-        assert "Agent knows Python" in flush_prompt
-        assert "User prefers dark mode" in flush_prompt
-        assert "Name: Alice" in flush_prompt
-        assert "Timezone: PST" in flush_prompt
-        assert "Do NOT overwrite or remove entries" in flush_prompt
-        assert "current live state of memory" in flush_prompt
-
-    def test_flush_works_without_memory_files(self, tmp_path, monkeypatch):
-        """When no memory files exist, flush still runs without the guard."""
-        empty_dir = tmp_path / "no_memories"
-        empty_dir.mkdir()
-
-        runner, tmp_agent, _ = _make_flush_context(monkeypatch)
-
-        with (
-            patch("gateway.run._resolve_runtime_agent_kwargs", return_value={"api_key": "k"}),
-            patch("gateway.run._resolve_gateway_model", return_value="test-model"),
-            patch.dict("sys.modules", {"tools.memory_tool": MagicMock(get_memory_dir=lambda: empty_dir)}),
-        ):
-            runner._flush_memories_for_session("session_456")
-
-        tmp_agent.run_conversation.assert_called_once()
-        flush_prompt = tmp_agent.run_conversation.call_args.kwargs.get("user_message", "")
-        assert "Do NOT overwrite or remove entries" not in flush_prompt
-        assert "Review the conversation above" in flush_prompt
-
-    def test_empty_memory_files_no_injection(self, tmp_path, monkeypatch):
-        """Empty memory files should not trigger the guard section."""
-        memory_dir = tmp_path / "memories"
-        memory_dir.mkdir()
-        (memory_dir / "MEMORY.md").write_text("")
-        (memory_dir / "USER.md").write_text("  \n  ")  # whitespace only
-
-        runner, tmp_agent, _ = _make_flush_context(monkeypatch)
-
-        with (
-            patch("gateway.run._resolve_runtime_agent_kwargs", return_value={"api_key": "k"}),
-            patch("gateway.run._resolve_gateway_model", return_value="test-model"),
-            patch.dict("sys.modules", {"tools.memory_tool": MagicMock(get_memory_dir=lambda: memory_dir)}),
-        ):
-            runner._flush_memories_for_session("session_789")
-
-        tmp_agent.run_conversation.assert_called_once()
-        flush_prompt = tmp_agent.run_conversation.call_args.kwargs.get("user_message", "")
-        assert "current live state of memory" not in flush_prompt
-
-
-class TestFlushAgentSilenced:
-    """The flush agent must not produce any terminal output."""
-
-    def test_print_fn_set_to_noop(self, tmp_path, monkeypatch):
-        """_print_fn on the flush agent must be a no-op so tool output never leaks."""
-        runner = _make_runner()
-        runner.session_store.load_transcript.return_value = _TRANSCRIPT_4_MSGS
-
-        captured_agent = {}
-
-        def _fake_ai_agent(*args, **kwargs):
-            agent = MagicMock()
-            captured_agent["instance"] = agent
-            return agent
-
-        fake_run_agent = types.ModuleType("run_agent")
-        fake_run_agent.AIAgent = _fake_ai_agent
-        monkeypatch.setitem(sys.modules, "run_agent", fake_run_agent)
-
-        with (
-            patch("gateway.run._resolve_runtime_agent_kwargs", return_value={"api_key": "k"}),
-            patch("gateway.run._resolve_gateway_model", return_value="test-model"),
-            patch.dict("sys.modules", {"tools.memory_tool": MagicMock(get_memory_dir=lambda: tmp_path)}),
-        ):
-            runner._flush_memories_for_session("session_silent")
-
-        agent = captured_agent["instance"]
-        assert agent._print_fn is not None, "_print_fn should be overridden to suppress output"
-        # Confirm it is callable and produces no output (no exception)
-        agent._print_fn("should be silenced")
-
-    def test_kawaii_spinner_respects_print_fn(self):
-        """KawaiiSpinner must route all output through print_fn when supplied."""
-        from agent.display import KawaiiSpinner
-
-        written = []
-        spinner = KawaiiSpinner("test", print_fn=lambda *a, **kw: written.append(a))
-        spinner._write("hello")
-        assert written == [("hello",)], "spinner should route through print_fn"
-
-        # A no-op print_fn must produce no output to stdout
-        import io, sys
-        buf = io.StringIO()
-        old_stdout = sys.stdout
-        sys.stdout = buf
-        try:
-            silent_spinner = KawaiiSpinner("silent", print_fn=lambda *a, **kw: None)
-            silent_spinner._write("should not appear")
-            silent_spinner.stop("done")
-        finally:
-            sys.stdout = old_stdout
-        assert buf.getvalue() == "", "no-op print_fn spinner must not write to stdout"
-
-    def test_flush_agent_closes_resources_after_run(self, monkeypatch):
-        """Memory flush should close temporary agent resources after the turn."""
-        runner, tmp_agent, _ = _make_flush_context(monkeypatch)
-        tmp_agent.shutdown_memory_provider = MagicMock()
-        tmp_agent.close = MagicMock()
-
-        with (
-            patch("gateway.run._resolve_runtime_agent_kwargs", return_value={"api_key": "k"}),
-            patch("gateway.run._resolve_gateway_model", return_value="test-model"),
-            patch.dict("sys.modules", {"tools.memory_tool": MagicMock(get_memory_dir=lambda: Path("/nonexistent"))}),
-        ):
-            runner._flush_memories_for_session("session_cleanup")
-
-        tmp_agent.shutdown_memory_provider.assert_called_once()
-        tmp_agent.close.assert_called_once()
-
-
-class TestFlushPromptStructure:
-    """Verify the flush prompt retains its core instructions."""
-
-    def test_core_instructions_present(self, monkeypatch):
-        """The flush prompt should still contain the original guidance."""
-        runner, tmp_agent, _ = _make_flush_context(monkeypatch)
-
-        with (
-            patch("gateway.run._resolve_runtime_agent_kwargs", return_value={"api_key": "k"}),
-            patch("gateway.run._resolve_gateway_model", return_value="test-model"),
-            patch.dict("sys.modules", {"tools.memory_tool": MagicMock(get_memory_dir=lambda: Path("/nonexistent"))}),
-        ):
-            runner._flush_memories_for_session("session_struct")
-
-        flush_prompt = tmp_agent.run_conversation.call_args.kwargs.get("user_message", "")
-        assert "automatically reset" in flush_prompt
-        assert "Save any important facts" in flush_prompt
-        assert "consider saving it as a skill" in flush_prompt
-        assert "Do NOT respond to the user" in flush_prompt

tests/gateway/test_matrix.py

@@ -197,10 +197,14 @@ def _make_fake_mautrix():
             self.account_id = account_id
             self.pickle_key = pickle_key
             self.db = db
+            self._device_id = ""
 
         async def open(self):
             pass
 
+        async def put_device_id(self, device_id):
+            self._device_id = device_id
+
     mautrix_crypto_store_asyncpg.PgCryptoStore = PgCryptoStore
 
     # --- mautrix.util ---

tests/gateway/test_reasoning_command.py

@@ -33,6 +33,7 @@ def _make_runner():
     runner._ephemeral_system_prompt = ""
     runner._prefill_messages = []
     runner._reasoning_config = None
+    runner._session_reasoning_overrides = {}
     runner._show_reasoning = False
     runner._provider_routing = {}
     runner._fallback_model = None
@@ -76,6 +77,10 @@ class TestReasoningCommand:
         source = inspect.getsource(gateway_run.GatewayRunner._handle_message)
         assert '"reasoning"' in source
 
+    def test_parse_reasoning_command_args_accepts_ascii_and_smart_global_flags(self):
+        assert gateway_run.GatewayRunner._parse_reasoning_command_args("high --global") == ("high", True)
+        assert gateway_run.GatewayRunner._parse_reasoning_command_args("—global xhigh") == ("xhigh", True)
+
     @pytest.mark.asyncio
     async def test_reasoning_command_reloads_current_state_from_config(self, tmp_path, monkeypatch):
         hermes_home = tmp_path / "hermes"
@@ -111,13 +116,90 @@ class TestReasoningCommand:
         runner = _make_runner()
         runner._reasoning_config = {"enabled": True, "effort": "medium"}
 
-        result = await runner._handle_reasoning_command(_make_event("/reasoning low"))
+        result = await runner._handle_reasoning_command(_make_event("/reasoning low --global"))
 
         saved = yaml.safe_load(config_path.read_text(encoding="utf-8"))
         assert saved["agent"]["reasoning_effort"] == "low"
         assert runner._reasoning_config == {"enabled": True, "effort": "low"}
         assert "takes effect on next message" in result
 
+    @pytest.mark.asyncio
+    async def test_handle_reasoning_command_defaults_to_session_only(self, tmp_path, monkeypatch):
+        hermes_home = tmp_path / "hermes"
+        hermes_home.mkdir()
+        config_path = hermes_home / "config.yaml"
+        config_path.write_text("agent:\n  reasoning_effort: medium\n", encoding="utf-8")
+
+        monkeypatch.setattr(gateway_run, "_hermes_home", hermes_home)
+
+        runner = _make_runner()
+        event = _make_event("/reasoning high")
+        session_key = runner._session_key_for_source(event.source)
+
+        result = await runner._handle_reasoning_command(event)
+
+        saved = yaml.safe_load(config_path.read_text(encoding="utf-8"))
+        assert saved["agent"]["reasoning_effort"] == "medium"
+        assert runner._session_reasoning_overrides[session_key] == {"enabled": True, "effort": "high"}
+        assert runner._reasoning_config == {"enabled": True, "effort": "high"}
+        assert "session only" in result
+
+    @pytest.mark.asyncio
+    async def test_reasoning_global_clears_existing_session_override(self, tmp_path, monkeypatch):
+        hermes_home = tmp_path / "hermes"
+        hermes_home.mkdir()
+        config_path = hermes_home / "config.yaml"
+        config_path.write_text("agent:\n  reasoning_effort: medium\n", encoding="utf-8")
+
+        monkeypatch.setattr(gateway_run, "_hermes_home", hermes_home)
+
+        runner = _make_runner()
+        event = _make_event("/reasoning low --global")
+        session_key = runner._session_key_for_source(event.source)
+        runner._session_reasoning_overrides[session_key] = {"enabled": True, "effort": "xhigh"}
+
+        result = await runner._handle_reasoning_command(event)
+
+        saved = yaml.safe_load(config_path.read_text(encoding="utf-8"))
+        assert saved["agent"]["reasoning_effort"] == "low"
+        assert session_key not in runner._session_reasoning_overrides
+        assert "saved to config" in result
+
+    @pytest.mark.asyncio
+    async def test_reasoning_reset_clears_session_override_without_config_write(self, tmp_path, monkeypatch):
+        hermes_home = tmp_path / "hermes"
+        hermes_home.mkdir()
+        config_path = hermes_home / "config.yaml"
+        config_path.write_text("agent:\n  reasoning_effort: medium\n", encoding="utf-8")
+
+        monkeypatch.setattr(gateway_run, "_hermes_home", hermes_home)
+
+        runner = _make_runner()
+        event = _make_event("/reasoning reset")
+        session_key = runner._session_key_for_source(event.source)
+        runner._session_reasoning_overrides[session_key] = {"enabled": True, "effort": "xhigh"}
+
+        result = await runner._handle_reasoning_command(event)
+
+        saved = yaml.safe_load(config_path.read_text(encoding="utf-8"))
+        assert saved["agent"]["reasoning_effort"] == "medium"
+        assert session_key not in runner._session_reasoning_overrides
+        assert "cleared" in result
+
+    def test_resolve_session_reasoning_prefers_session_override(self, tmp_path, monkeypatch):
+        hermes_home = tmp_path / "hermes"
+        hermes_home.mkdir()
+        (hermes_home / "config.yaml").write_text("agent:\n  reasoning_effort: low\n", encoding="utf-8")
+
+        monkeypatch.setattr(gateway_run, "_hermes_home", hermes_home)
+
+        runner = _make_runner()
+        source = _make_event("/reasoning").source
+        session_key = runner._session_key_for_source(source)
+        runner._session_reasoning_overrides[session_key] = {"enabled": True, "effort": "xhigh"}
+
+        assert runner._resolve_session_reasoning_config(source=source) == {"enabled": True, "effort": "xhigh"}
+
     def test_run_agent_reloads_reasoning_config_per_message(self, tmp_path, monkeypatch):
         hermes_home = tmp_path / "hermes"
         hermes_home.mkdir()
@@ -167,6 +249,56 @@ class TestReasoningCommand:
         assert _CapturingAgent.last_init is not None
         assert _CapturingAgent.last_init["reasoning_config"] == {"enabled": True, "effort": "low"}
 
+    def test_run_agent_prefers_session_reasoning_override(self, tmp_path, monkeypatch):
+        hermes_home = tmp_path / "hermes"
+        hermes_home.mkdir()
+        (hermes_home / "config.yaml").write_text("agent:\n  reasoning_effort: low\n", encoding="utf-8")
+
+        monkeypatch.setattr(gateway_run, "_hermes_home", hermes_home)
+        monkeypatch.setattr(gateway_run, "_env_path", hermes_home / ".env")
+        monkeypatch.setattr(gateway_run, "load_dotenv", lambda *args, **kwargs: None)
+        monkeypatch.setattr(
+            gateway_run,
+            "_resolve_runtime_agent_kwargs",
+            lambda: {
+                "provider": "openrouter",
+                "api_mode": "chat_completions",
+                "base_url": "https://openrouter.ai/api/v1",
+                "api_key": "***",
+            },
+        )
+        fake_run_agent = types.ModuleType("run_agent")
+        fake_run_agent.AIAgent = _CapturingAgent
+        monkeypatch.setitem(sys.modules, "run_agent", fake_run_agent)
+
+        _CapturingAgent.last_init = None
+        runner = _make_runner()
+        session_key = "agent:main:local:dm"
+        runner._session_reasoning_overrides[session_key] = {"enabled": True, "effort": "high"}
+
+        source = SessionSource(
+            platform=Platform.LOCAL,
+            chat_id="cli",
+            chat_name="CLI",
+            chat_type="dm",
+            user_id="user-1",
+        )
+
+        result = asyncio.run(
+            runner._run_agent(
+                message="ping",
+                context_prompt="",
+                history=[],
+                source=source,
+                session_id="session-1",
+                session_key=session_key,
+            )
+        )
+
+        assert result["final_response"] == "ok"
+        assert _CapturingAgent.last_init is not None
+        assert _CapturingAgent.last_init["reasoning_config"] == {"enabled": True, "effort": "high"}
+
     def test_run_agent_includes_enabled_mcp_servers_in_gateway_toolsets(self, tmp_path, monkeypatch):
         hermes_home = tmp_path / "hermes"
         hermes_home.mkdir()

tests/gateway/test_resume_command.py

@@ -4,7 +4,7 @@ Tests the _handle_resume_command handler (switch to a previously-named session)
 across gateway messenger platforms.
 """
 
-from unittest.mock import MagicMock, AsyncMock
+from unittest.mock import MagicMock
 
 import pytest
 
@@ -53,9 +53,6 @@ def _make_runner(session_db=None, current_session_id="current_session_001",
     mock_store.switch_session.return_value = mock_session_entry
     runner.session_store = mock_store
 
-    # Stub out memory flushing
-    runner._async_flush_memories = AsyncMock()
-
     return runner
 
 
@@ -179,6 +176,40 @@ class TestHandleResumeCommand:
         assert call_args[0][1] == "sess_v2"
         db.close()
 
+    @pytest.mark.asyncio
+    async def test_resume_follows_compression_continuation(self, tmp_path):
+        """Gateway /resume should reopen the live descendant after compression."""
+        from hermes_state import SessionDB
+
+        db = SessionDB(db_path=tmp_path / "state.db")
+        db.create_session("compressed_root", "telegram")
+        db.set_session_title("compressed_root", "Compressed Work")
+        db.end_session("compressed_root", "compression")
+        db.create_session("compressed_child", "telegram", parent_session_id="compressed_root")
+        db.append_message("compressed_child", "user", "hello from continuation")
+        db.create_session("current_session_001", "telegram")
+
+        event = _make_event(text="/resume Compressed Work")
+        runner = _make_runner(
+            session_db=db,
+            current_session_id="current_session_001",
+            event=event,
+        )
+        runner.session_store.load_transcript.side_effect = (
+            lambda session_id: [{"role": "user", "content": "hello from continuation"}]
+            if session_id == "compressed_child"
+            else []
+        )
+
+        result = await runner._handle_resume_command(event)
+
+        assert "Resumed session" in result
+        assert "(1 message)" in result
+        call_args = runner.session_store.switch_session.call_args
+        assert call_args[0][1] == "compressed_child"
+        runner.session_store.load_transcript.assert_called_with("compressed_child")
+        db.close()
+
     @pytest.mark.asyncio
     async def test_resume_clears_running_agent(self, tmp_path):
         """Switching sessions clears any cached running agent."""
@@ -199,28 +230,3 @@ class TestHandleResumeCommand:
 
         assert real_key not in runner._running_agents
         db.close()
-
-    @pytest.mark.asyncio
-    async def test_resume_flushes_memories(self, tmp_path):
-        """Resume should flush memories from the current session before switching."""
-        from hermes_state import SessionDB
-
-        db = SessionDB(db_path=tmp_path / "state.db")
-        db.create_session("old_session", "telegram")
-        db.set_session_title("old_session", "Old Work")
-        db.create_session("current_session_001", "telegram")
-
-        event = _make_event(text="/resume Old Work")
-        runner = _make_runner(
-            session_db=db,
-            current_session_id="current_session_001",
-            event=event,
-        )
-
-        await runner._handle_resume_command(event)
-
-        runner._async_flush_memories.assert_called_once_with(
-            "current_session_001",
-            "agent:main:telegram:dm:67890",
-        )
-        db.close()

tests/gateway/test_run_progress_interrupt.py

@@ -0,0 +1,215 @@
+"""Tests for interrupt-aware tool-progress suppression in gateway.
+
+When a user sends `stop` while the agent is executing a batch of parallel
+tool calls, the gateway's progress_callback should stop queuing 🔍 bubbles
+and the drain loop should drop any already-queued events.  Without this
+guard, the stop acknowledgement appears first but is followed by a trail
+of tool-progress bubbles for calls that were already parsed from the LLM
+response — making the interrupt feel ignored.
+"""
+
+import asyncio
+import importlib
+import sys
+import time
+import types
+from types import SimpleNamespace
+
+import pytest
+
+from gateway.config import Platform, PlatformConfig
+from gateway.platforms.base import BasePlatformAdapter, SendResult
+from gateway.session import SessionSource
+
+
+class ProgressCaptureAdapter(BasePlatformAdapter):
+    def __init__(self, platform=Platform.TELEGRAM):
+        super().__init__(PlatformConfig(enabled=True, token="***"), platform)
+        self.sent = []
+        self.edits = []
+        self.typing = []
+
+    async def connect(self) -> bool:
+        return True
+
+    async def disconnect(self) -> None:
+        return None
+
+    async def send(self, chat_id, content, reply_to=None, metadata=None) -> SendResult:
+        self.sent.append({"chat_id": chat_id, "content": content})
+        return SendResult(success=True, message_id="progress-1")
+
+    async def edit_message(self, chat_id, message_id, content) -> SendResult:
+        self.edits.append({"message_id": message_id, "content": content})
+        return SendResult(success=True, message_id=message_id)
+
+    async def send_typing(self, chat_id, metadata=None) -> None:
+        self.typing.append(chat_id)
+
+    async def stop_typing(self, chat_id) -> None:
+        return None
+
+    async def get_chat_info(self, chat_id: str):
+        return {"id": chat_id}
+
+
+class PreInterruptAgent:
+    """Fires tool-progress events BEFORE the interrupt lands.
+
+    These should render normally.  Baseline for comparison with the
+    interrupted case — proves the harness renders events when no
+    interrupt is active.
+    """
+
+    def __init__(self, **kwargs):
+        self.tool_progress_callback = kwargs.get("tool_progress_callback")
+        self.tools = []
+        self._interrupt_requested = False
+
+    @property
+    def is_interrupted(self) -> bool:
+        return self._interrupt_requested
+
+    def run_conversation(self, message, conversation_history=None, task_id=None):
+        self.tool_progress_callback("tool.started", "web_search", "first search", {})
+        time.sleep(0.35)  # let the drain loop process
+        return {"final_response": "done", "messages": [], "api_calls": 1}
+
+
+class InterruptedAgent:
+    """Fires tool.started events AFTER interrupt — all should be suppressed.
+
+    Mirrors the failure mode in the bug report: LLM returned N parallel
+    web_search calls, interrupt flag flipped, remaining events still
+    rendered as bubbles.  With the fix, none of these should appear.
+    """
+
+    def __init__(self, **kwargs):
+        self.tool_progress_callback = kwargs.get("tool_progress_callback")
+        self.tools = []
+        # Start already interrupted — simulates stop having already landed
+        # by the time the agent batch starts firing tool.started events.
+        self._interrupt_requested = True
+
+    @property
+    def is_interrupted(self) -> bool:
+        return self._interrupt_requested
+
+    def run_conversation(self, message, conversation_history=None, task_id=None):
+        # Parallel tool batch — in production these come from one LLM
+        # response with 5 tool_calls.  All are post-interrupt.
+        self.tool_progress_callback("tool.started", "web_search", "cognee hermes", {})
+        self.tool_progress_callback("tool.started", "web_search", "McBee deer hunting", {})
+        self.tool_progress_callback("tool.started", "web_search", "kuzu graph db", {})
+        self.tool_progress_callback("tool.started", "web_search", "moonshot kimi api", {})
+        self.tool_progress_callback("tool.started", "web_search", "platform.moonshot.cn", {})
+        time.sleep(0.35)  # let the drain loop attempt to process the queue
+        return {"final_response": "interrupted", "messages": [], "api_calls": 1}
+
+
+def _make_runner(adapter):
+    gateway_run = importlib.import_module("gateway.run")
+    GatewayRunner = gateway_run.GatewayRunner
+
+    runner = object.__new__(GatewayRunner)
+    runner.adapters = {adapter.platform: adapter}
+    runner._voice_mode = {}
+    runner._prefill_messages = []
+    runner._ephemeral_system_prompt = ""
+    runner._reasoning_config = None
+    runner._provider_routing = {}
+    runner._fallback_model = None
+    runner._session_db = None
+    runner._running_agents = {}
+    runner._session_run_generation = {}
+    runner.hooks = SimpleNamespace(loaded_hooks=False)
+    runner.config = SimpleNamespace(
+        thread_sessions_per_user=False,
+        group_sessions_per_user=False,
+        stt_enabled=False,
+    )
+    return runner
+
+
+async def _run_once(monkeypatch, tmp_path, agent_cls, session_id):
+    monkeypatch.setenv("HERMES_TOOL_PROGRESS_MODE", "all")
+
+    fake_dotenv = types.ModuleType("dotenv")
+    fake_dotenv.load_dotenv = lambda *args, **kwargs: None
+    monkeypatch.setitem(sys.modules, "dotenv", fake_dotenv)
+
+    fake_run_agent = types.ModuleType("run_agent")
+    fake_run_agent.AIAgent = agent_cls
+    monkeypatch.setitem(sys.modules, "run_agent", fake_run_agent)
+
+    adapter = ProgressCaptureAdapter()
+    runner = _make_runner(adapter)
+    gateway_run = importlib.import_module("gateway.run")
+    monkeypatch.setattr(gateway_run, "_hermes_home", tmp_path)
+    monkeypatch.setattr(
+        gateway_run,
+        "_resolve_runtime_agent_kwargs",
+        lambda: {"api_key": "fake"},
+    )
+    source = SessionSource(
+        platform=Platform.TELEGRAM,
+        chat_id="-1001",
+        chat_type="group",
+        thread_id="17585",
+    )
+    result = await runner._run_agent(
+        message="hi",
+        context_prompt="",
+        history=[],
+        source=source,
+        session_id=session_id,
+        session_key="agent:main:telegram:group:-1001:17585",
+    )
+    return adapter, result
+
+
+@pytest.mark.asyncio
+async def test_baseline_non_interrupted_agent_renders_progress(monkeypatch, tmp_path):
+    """Sanity check: when is_interrupted is False, tool-progress renders normally."""
+    adapter, result = await _run_once(monkeypatch, tmp_path, PreInterruptAgent, "sess-baseline")
+    assert result["final_response"] == "done"
+    rendered = " ".join(c["content"] for c in adapter.sent) + " " + " ".join(
+        c["content"] for c in adapter.edits
+    )
+    assert "first search" in rendered, (
+        "baseline agent should render its tool-progress event — "
+        "if this fails the test harness is broken, not the fix"
+    )
+
+
+@pytest.mark.asyncio
+async def test_progress_suppressed_when_agent_is_interrupted(monkeypatch, tmp_path):
+    """Post-interrupt tool.started events must not render as bubbles.
+
+    This is Bug B from the screenshot: user sends `stop`, agent acks with
+    ⚡ Interrupting, but 5 more 🔍 web_search bubbles still render because
+    their tool.started events were already parsed from the LLM response.
+    With the fix, progress_callback and the drain loop both check
+    is_interrupted and skip these events.
+    """
+    adapter, result = await _run_once(
+        monkeypatch, tmp_path, InterruptedAgent, "sess-interrupted"
+    )
+    assert result["final_response"] == "interrupted"
+
+    rendered = " ".join(c["content"] for c in adapter.sent) + " " + " ".join(
+        c["content"] for c in adapter.edits
+    )
+
+    # None of the post-interrupt queries should appear.
+    for leaked_query in (
+        "cognee hermes",
+        "McBee deer hunting",
+        "kuzu graph db",
+        "moonshot kimi api",
+        "platform.moonshot.cn",
+    ):
+        assert leaked_query not in rendered, (
+            f"event '{leaked_query}' leaked into the UI after interrupt — "
+            f"progress_callback / drain loop is not checking is_interrupted"
+        )

tests/gateway/test_running_agent_session_toggles.py

@@ -165,3 +165,26 @@ async def test_reasoning_rejected_mid_run():
     assert result is not None
     assert "can't run mid-turn" in result
     assert "/reasoning" in result
+
+
+@pytest.mark.asyncio
+async def test_btw_dispatches_mid_run():
+    """/btw mid-run must dispatch to /background's handler, not hit the catch-all.
+
+    /btw is an alias of /background (see hermes_cli/commands.py). Typing
+    /btw mid-turn must spawn a parallel background task — that's the whole
+    point of the command. Before the mid-turn bypass was added for
+    /background, /btw fell through to the "Agent is running — wait or
+    /stop first" catch-all, making it useless in exactly the scenario it
+    was designed for. The alias and the bypass together make it work.
+    """
+    runner = _make_runner()
+    runner._handle_background_command = AsyncMock(
+        return_value='🚀 Background task started: "what module owns titles?"'
+    )
+
+    result = await runner._handle_message(_make_event("/btw what module owns titles?"))
+
+    runner._handle_background_command.assert_awaited_once()
+    assert result is not None
+    assert "can't run mid-turn" not in result

tests/gateway/test_session_boundary_hooks.py

@@ -177,8 +177,8 @@ async def test_idle_expiry_fires_finalize_hook(mock_invoke_hook):
     its reset policy (idle timeout, scheduled reset), it must fire
     ``on_session_finalize`` so plugin providers get the same final-pass
     extraction opportunity they'd get from /new or CLI shutdown.  Before
-    the fix, the expiry path flushed memories and evicted the agent but
-    silently skipped the hook.
+    the fix, the expiry path evicted the agent but silently skipped the
+    hook.
     """
     from datetime import datetime, timedelta
 
@@ -200,7 +200,7 @@ async def test_idle_expiry_fires_finalize_hook(mock_invoke_hook):
         platform=Platform.TELEGRAM,
         chat_type="dm",
     )
-    expired_entry.memory_flushed = False
+    expired_entry.expiry_finalized = False
 
     runner.session_store = MagicMock()
     runner.session_store._ensure_loaded = MagicMock()
@@ -211,24 +211,24 @@ async def test_idle_expiry_fires_finalize_hook(mock_invoke_hook):
     runner.session_store._lock.__exit__ = MagicMock(return_value=None)
     runner.session_store._save = MagicMock()
 
-    runner._async_flush_memories = AsyncMock()
     runner._evict_cached_agent = MagicMock()
     runner._cleanup_agent_resources = MagicMock()
     runner._sweep_idle_cached_agents = MagicMock(return_value=0)
 
     # The watcher starts with `await asyncio.sleep(60)` and loops while
-    # `self._running`. Patch sleep so the 60s initial delay is instant, then
-    # flip `_running` false inside the flush call so the loop exits cleanly
-    # after one pass.
+    # `self._running`.  Patch sleep so the 60s initial delay is instant, and
+    # make the expiry hook invocation flip `_running` false so the loop
+    # exits cleanly after one pass.
     _orig_sleep = __import__("asyncio").sleep
 
     async def _fast_sleep(_):
         await _orig_sleep(0)
 
-    async def _flush_and_stop(session_id, key):
-        runner._running = False  # terminate the loop after this iteration
+    def _hook_and_stop(*a, **kw):
+        runner._running = False
+        return None
 
-    runner._async_flush_memories = AsyncMock(side_effect=_flush_and_stop)
+    mock_invoke_hook.side_effect = _hook_and_stop
 
     with patch("gateway.run.asyncio.sleep", side_effect=_fast_sleep):
         await runner._session_expiry_watcher(interval=0)

tests/gateway/test_session_boundary_security_state.py

@@ -1,7 +1,7 @@
 """Regression tests for approval-state cleanup on session boundaries."""
 
 from datetime import datetime
-from unittest.mock import AsyncMock, MagicMock
+from unittest.mock import MagicMock
 
 import pytest
 
@@ -72,7 +72,6 @@ def _make_resume_runner():
     runner = object.__new__(GatewayRunner)
     runner.adapters = {}
     runner._background_tasks = set()
-    runner._async_flush_memories = AsyncMock()
     runner._running_agents = {}
     runner._running_agents_ts = {}
     runner._busy_ack_ts = {}

tests/gateway/test_session_info.py

@@ -58,7 +58,7 @@ class TestFormatSessionInfo:
                                   {"provider": "", "base_url": "", "api_key": ""})
         with p1, p2, p3:
             info = runner._format_session_info()
-        assert "128K" in info
+        assert "256K" in info
         assert "model.context_length" in info
 
     def test_local_endpoint_shown(self, runner, tmp_path):

tests/gateway/test_session_model_override_routing.py

@@ -54,6 +54,7 @@ def _make_runner():
     runner._background_tasks = set()
     runner._session_db = None
     runner._session_model_overrides = {}
+    runner._session_reasoning_overrides = {}
     runner._pending_model_notes = {}
     runner._pending_approvals = {}
     runner._agent_cache = {}
@@ -102,6 +103,7 @@ def test_run_agent_prefers_session_override_over_global_runtime(monkeypatch):
     )
     session_key = "agent:main:local:dm"
     runner._session_model_overrides[session_key] = _codex_override()
+    runner._session_reasoning_overrides[session_key] = {"enabled": True, "effort": "high"}
 
     result = asyncio.run(
         runner._run_agent(
@@ -121,6 +123,7 @@ def test_run_agent_prefers_session_override_over_global_runtime(monkeypatch):
     assert _CapturingAgent.last_init["api_mode"] == "codex_responses"
     assert _CapturingAgent.last_init["base_url"] == "https://chatgpt.com/backend-api/codex"
     assert _CapturingAgent.last_init["api_key"] == "***"
+    assert _CapturingAgent.last_init["reasoning_config"] == {"enabled": True, "effort": "high"}
 
 
 @pytest.mark.asyncio
@@ -149,6 +152,7 @@ async def test_background_task_prefers_session_override_over_global_runtime(monk
     )
     session_key = runner._session_key_for_source(source)
     runner._session_model_overrides[session_key] = _codex_override()
+    runner._session_reasoning_overrides[session_key] = {"enabled": True, "effort": "high"}
 
     await runner._run_background_task("say hello", source, "bg_test")
 
@@ -158,3 +162,4 @@ async def test_background_task_prefers_session_override_over_global_runtime(monk
     assert _CapturingAgent.last_init["api_mode"] == "codex_responses"
     assert _CapturingAgent.last_init["base_url"] == "https://chatgpt.com/backend-api/codex"
     assert _CapturingAgent.last_init["api_key"] == "***"
+    assert _CapturingAgent.last_init["reasoning_config"] == {"enabled": True, "effort": "high"}

tests/gateway/test_session_model_reset.py

@@ -1,4 +1,4 @@
-"""Tests that /new (and its /reset alias) clears the session-scoped model override."""
+"""Tests that /new (and its /reset alias) clears session-scoped overrides."""
 from datetime import datetime
 from types import SimpleNamespace
 from unittest.mock import AsyncMock, MagicMock
@@ -37,6 +37,7 @@ def _make_runner():
     runner._voice_mode = {}
     runner.hooks = SimpleNamespace(emit=AsyncMock(), loaded_hooks=False)
     runner._session_model_overrides = {}
+    runner._session_reasoning_overrides = {}
     runner._pending_model_notes = {}
     runner._background_tasks = set()
 
@@ -75,14 +76,16 @@ async def test_new_command_clears_session_model_override():
     runner._session_model_overrides[session_key] = {
         "model": "gpt-4o",
         "provider": "openai",
-        "api_key": "sk-test",
+        "api_key": "***",
         "base_url": "",
         "api_mode": "openai",
     }
+    runner._session_reasoning_overrides[session_key] = {"enabled": True, "effort": "high"}
 
     await runner._handle_reset_command(_make_event("/new"))
 
     assert session_key not in runner._session_model_overrides
+    assert session_key not in runner._session_reasoning_overrides
 
 
 @pytest.mark.asyncio
@@ -92,10 +95,12 @@ async def test_new_command_no_override_is_noop():
     session_key = build_session_key(_make_source())
 
     assert session_key not in runner._session_model_overrides
+    assert session_key not in runner._session_reasoning_overrides
 
     await runner._handle_reset_command(_make_event("/new"))
 
     assert session_key not in runner._session_model_overrides
+    assert session_key not in runner._session_reasoning_overrides
 
 
 @pytest.mark.asyncio
@@ -115,12 +120,16 @@ async def test_new_command_only_clears_own_session():
     runner._session_model_overrides[other_key] = {
         "model": "claude-sonnet-4-6",
         "provider": "anthropic",
-        "api_key": "sk-ant-test",
+        "api_key": "***",
         "base_url": "",
         "api_mode": "anthropic",
     }
+    runner._session_reasoning_overrides[session_key] = {"enabled": True, "effort": "high"}
+    runner._session_reasoning_overrides[other_key] = {"enabled": True, "effort": "low"}
 
     await runner._handle_reset_command(_make_event("/new"))
 
     assert session_key not in runner._session_model_overrides
     assert other_key in runner._session_model_overrides
+    assert session_key not in runner._session_reasoning_overrides
+    assert other_key in runner._session_reasoning_overrides

tests/gateway/test_voice_command.py

@@ -177,6 +177,53 @@ class TestHandleVoiceCommand:
 
         assert adapter._auto_tts_disabled_chats == {"123"}
 
+    def test_sync_populates_enabled_chats_from_voice_modes(self, runner):
+        """Issue #16007: sync also restores per-chat /voice on|tts opt-ins.
+
+        The adapter's ``_auto_tts_enabled_chats`` must mirror chats whose
+        persisted voice_mode is ``voice_only`` or ``all`` — without this,
+        ``/voice on`` was relying on a "not in disabled set" default that
+        silently enabled auto-TTS for every chat.
+        """
+        from gateway.config import Platform
+        runner._voice_mode = {
+            "telegram:off_chat": "off",
+            "telegram:on_chat": "voice_only",
+            "telegram:tts_chat": "all",
+            "slack:999": "voice_only",  # wrong platform, must be ignored
+        }
+        adapter = SimpleNamespace(
+            _auto_tts_default=False,
+            _auto_tts_disabled_chats=set(),
+            _auto_tts_enabled_chats=set(),
+            platform=Platform.TELEGRAM,
+        )
+
+        runner._sync_voice_mode_state_to_adapter(adapter)
+
+        assert adapter._auto_tts_disabled_chats == {"off_chat"}
+        assert adapter._auto_tts_enabled_chats == {"on_chat", "tts_chat"}
+
+    def test_sync_pushes_config_default_onto_adapter(self, runner, monkeypatch):
+        """Issue #16007: ``voice.auto_tts`` must propagate to ``_auto_tts_default``."""
+        from gateway.config import Platform
+
+        fake_cfg = {"voice": {"auto_tts": True}}
+        monkeypatch.setattr(
+            "hermes_cli.config.load_config",
+            lambda: fake_cfg,
+        )
+        adapter = SimpleNamespace(
+            _auto_tts_default=False,
+            _auto_tts_disabled_chats=set(),
+            _auto_tts_enabled_chats=set(),
+            platform=Platform.TELEGRAM,
+        )
+
+        runner._sync_voice_mode_state_to_adapter(adapter)
+
+        assert adapter._auto_tts_default is True
+
     def test_restart_restores_voice_off_state(self, runner, tmp_path):
         from gateway.config import Platform
         runner._VOICE_MODE_PATH.write_text(json.dumps({"telegram:123": "off"}))
@@ -2706,3 +2753,56 @@ class TestUDPKeepalive:
             mock_conn.send_packet.assert_called_with(b'\xf8\xff\xfe')
         finally:
             DiscordAdapter._KEEPALIVE_INTERVAL = original_interval
+
+
+# =====================================================================
+# BasePlatformAdapter._should_auto_tts_for_chat — gate for auto-TTS
+# on voice input. Regression test for Issue #16007.
+# =====================================================================
+
+class TestShouldAutoTtsForChat:
+    """Three-layer gate: per-chat enable > per-chat disable > config default."""
+
+    def _make_adapter(self, *, default: bool, enabled=(), disabled=()):
+        """Build a bare adapter with only the attrs the gate reads."""
+        adapter = SimpleNamespace(
+            _auto_tts_default=default,
+            _auto_tts_enabled_chats=set(enabled),
+            _auto_tts_disabled_chats=set(disabled),
+        )
+        # Bind the unbound method — _should_auto_tts_for_chat only reads the
+        # three attrs above via ``self.``, so an unbound call works.
+        from gateway.platforms.base import BasePlatformAdapter
+        return BasePlatformAdapter._should_auto_tts_for_chat, adapter
+
+    def test_default_false_no_override_suppresses(self):
+        """Issue #16007: voice.auto_tts=False and no per-chat state → no TTS."""
+        fn, adapter = self._make_adapter(default=False)
+        assert fn(adapter, "chat1") is False
+
+    def test_default_true_no_override_fires(self):
+        fn, adapter = self._make_adapter(default=True)
+        assert fn(adapter, "chat1") is True
+
+    def test_explicit_enable_overrides_false_default(self):
+        """``/voice on`` with config auto_tts=False still fires."""
+        fn, adapter = self._make_adapter(default=False, enabled={"chat1"})
+        assert fn(adapter, "chat1") is True
+
+    def test_explicit_disable_overrides_true_default(self):
+        """``/voice off`` with config auto_tts=True still suppresses."""
+        fn, adapter = self._make_adapter(default=True, disabled={"chat1"})
+        assert fn(adapter, "chat1") is False
+
+    def test_enabled_wins_over_disabled(self):
+        """An explicit enable beats an explicit disable (enable takes priority)."""
+        fn, adapter = self._make_adapter(
+            default=False, enabled={"chat1"}, disabled={"chat1"}
+        )
+        assert fn(adapter, "chat1") is True
+
+    def test_per_chat_isolation(self):
+        """Enable for chat1 doesn't leak to chat2."""
+        fn, adapter = self._make_adapter(default=False, enabled={"chat1"})
+        assert fn(adapter, "chat1") is True
+        assert fn(adapter, "chat2") is False

tests/hermes_cli/test_apply_model_switch_result_context.py

@@ -0,0 +1,152 @@
+"""Regression test for the `/model` picker confirmation display.
+
+Bug (April 2026): after choosing a model from the interactive `/model` picker,
+``HermesCLI._apply_model_switch_result()`` printed ``ModelInfo.context_window``
+straight from models.dev, which always reports the vendor-wide value (e.g.
+gpt-5.5 = 1,050,000 on ``openai``). That ignored provider-specific caps — in
+particular, ChatGPT Codex OAuth enforces 272K on the same slug. The sibling
+``_handle_model_switch()`` (typed ``/model <name>``) was already fixed to use
+``resolve_display_context_length()``; the picker path was missed, causing
+"sometimes 1M, sometimes 272K" for the same model across sibling UI paths.
+
+Fix: both display paths now go through ``resolve_display_context_length()``.
+"""
+from __future__ import annotations
+
+from unittest.mock import patch
+
+from hermes_cli.model_switch import ModelSwitchResult
+
+
+class _FakeModelInfo:
+    context_window = 1_050_000
+    max_output = 0
+
+    def has_cost_data(self):
+        return False
+
+    def format_capabilities(self):
+        return ""
+
+
+class _StubCLI:
+    """Minimum attrs ``_apply_model_switch_result`` reads on ``self``."""
+    agent = None
+    model = ""
+    provider = ""
+    requested_provider = ""
+    api_key = ""
+    _explicit_api_key = ""
+    base_url = ""
+    _explicit_base_url = ""
+    api_mode = ""
+    _pending_model_switch_note = ""
+
+
+def _run_display(monkeypatch, result):
+    import cli as cli_mod
+
+    captured: list[str] = []
+    monkeypatch.setattr(cli_mod, "_cprint", lambda s, *a, **k: captured.append(str(s)))
+    # Avoid writing to ~/.hermes/config.yaml during the test.
+    monkeypatch.setattr(cli_mod, "save_config_value", lambda *a, **k: None)
+    cli_mod.HermesCLI._apply_model_switch_result(_StubCLI(), result, False)
+    return captured
+
+
+def test_picker_path_uses_provider_aware_context_on_codex(monkeypatch):
+    """``_apply_model_switch_result`` must prefer the provider-aware resolver
+    (272K on Codex) over the raw models.dev value (1.05M for gpt-5.5).
+    """
+    result = ModelSwitchResult(
+        success=True,
+        new_model="gpt-5.5",
+        target_provider="openai-codex",
+        provider_changed=True,
+        api_key="",
+        base_url="https://chatgpt.com/backend-api/codex",
+        api_mode="codex_responses",
+        warning_message="",
+        provider_label="ChatGPT Codex",
+        resolved_via_alias=False,
+        capabilities=None,
+        model_info=_FakeModelInfo(),  # models.dev says 1.05M
+        is_global=False,
+    )
+    with patch(
+        "agent.model_metadata.get_model_context_length",
+        return_value=272_000,
+    ):
+        lines = _run_display(monkeypatch, result)
+
+    ctx_line = next((l for l in lines if "Context:" in l), "")
+    assert "272,000" in ctx_line, (
+        f"picker-path display must show Codex's 272K cap, got: {ctx_line!r}"
+    )
+    assert "1,050,000" not in ctx_line, (
+        f"picker-path display leaked models.dev's 1.05M for Codex: {ctx_line!r}"
+    )
+
+
+def test_picker_path_shows_vendor_value_when_no_provider_cap(monkeypatch):
+    """On providers with no enforced cap (e.g. OpenRouter), the picker path
+    should surface the real 1.05M context for gpt-5.5 — resolver and models.dev
+    agree here.
+    """
+    result = ModelSwitchResult(
+        success=True,
+        new_model="openai/gpt-5.5",
+        target_provider="openrouter",
+        provider_changed=True,
+        api_key="",
+        base_url="https://openrouter.ai/api/v1",
+        api_mode="chat_completions",
+        warning_message="",
+        provider_label="OpenRouter",
+        resolved_via_alias=False,
+        capabilities=None,
+        model_info=_FakeModelInfo(),
+        is_global=False,
+    )
+    with patch(
+        "agent.model_metadata.get_model_context_length",
+        return_value=1_050_000,
+    ):
+        lines = _run_display(monkeypatch, result)
+
+    ctx_line = next((l for l in lines if "Context:" in l), "")
+    assert "1,050,000" in ctx_line, (
+        f"OpenRouter gpt-5.5 should show 1.05M context, got: {ctx_line!r}"
+    )
+
+
+def test_picker_path_falls_back_to_model_info_when_resolver_empty(monkeypatch):
+    """If ``get_model_context_length`` returns nothing (rare — truly unknown
+    endpoint), the display still surfaces ``ModelInfo.context_window`` so the
+    user sees *something* rather than a silent blank.
+    """
+    result = ModelSwitchResult(
+        success=True,
+        new_model="some-model",
+        target_provider="some-provider",
+        provider_changed=True,
+        api_key="",
+        base_url="",
+        api_mode="chat_completions",
+        warning_message="",
+        provider_label="Some Provider",
+        resolved_via_alias=False,
+        capabilities=None,
+        model_info=_FakeModelInfo(),  # context_window = 1_050_000
+        is_global=False,
+    )
+    with patch(
+        "agent.model_metadata.get_model_context_length",
+        return_value=None,
+    ):
+        lines = _run_display(monkeypatch, result)
+
+    ctx_line = next((l for l in lines if "Context:" in l), "")
+    assert "1,050,000" in ctx_line, (
+        f"resolver-empty path should fall back to ModelInfo, got: {ctx_line!r}"
+    )