@clawpump/claw-agent 0.1.4 → 0.1.6

package/agent/plugins/model-providers/openrouter/__init__.py

@@ -10,6 +10,39 @@ logger = logging.getLogger(__name__)
 
 _CACHE: list[str] | None = None
 
+# Anthropic model families that still accept an explicit "disable thinking"
+# request (the manual ``thinking: {type: "disabled"}`` form OpenRouter emits
+# for ``reasoning: {enabled: false}``). Everything Claude 4.6 and newer —
+# including future date-stamped / named models (fable, mythos-class, …) —
+# mandates reasoning and returns HTTP 400 on any disable form. We therefore
+# default *unknown* Anthropic models to "cannot disable" (the modern contract)
+# and keep only this explicit legacy allowlist of models that can. Mirrors the
+# default-to-newest philosophy in agent/anthropic_adapter._get_anthropic_max_output.
+_ANTHROPIC_REASONING_OPTIONAL_SUBSTRINGS = (
+    "claude-3",          # 3, 3.5, 3.7
+    "claude-opus-4-0", "claude-opus-4.0", "claude-opus-4-1", "claude-opus-4.1",
+    "claude-sonnet-4-0", "claude-sonnet-4.0",
+    "claude-opus-4-2025", "claude-sonnet-4-2025",  # date-stamped 4.0 IDs
+    "claude-opus-4-5", "claude-opus-4.5",
+    "claude-sonnet-4-5", "claude-sonnet-4.5",
+    "claude-haiku-4-5", "claude-haiku-4.5",
+)
+
+
+def _anthropic_reasoning_is_mandatory(model: str | None) -> bool:
+    """Return True for Anthropic models that reject any disable-thinking form.
+
+    Claude 4.6+ (adaptive thinking) and newer named models have no "off"
+    switch — sending ``reasoning: {enabled: false}`` makes OpenRouter emit
+    ``thinking: {type: "disabled"}``, which these models 400 on. Unknown /
+    new Anthropic model names default to mandatory so the next un-numbered
+    release doesn't reintroduce the 400.
+    """
+    m = (model or "").lower()
+    if not m.startswith(("anthropic/", "claude")) and "claude" not in m:
+        return False
+    return not any(sub in m for sub in _ANTHROPIC_REASONING_OPTIONAL_SUBSTRINGS)
+
 
 class OpenRouterProfile(ProviderProfile):
     """OpenRouter aggregator — provider preferences, reasoning config passthrough."""
@@ -18,6 +51,7 @@ class OpenRouterProfile(ProviderProfile):
         self,
         *,
         api_key: str | None = None,
+        base_url: str | None = None,
         timeout: float = 8.0,
     ) -> list[str] | None:
         """Fetch from public OpenRouter catalog — no auth required.
@@ -31,7 +65,7 @@ class OpenRouterProfile(ProviderProfile):
         if _CACHE is not None:
             return _CACHE
         try:
-            result = super().fetch_models(api_key=None, timeout=timeout)
+            result = super().fetch_models(api_key=None, base_url=base_url, timeout=timeout)
             if result is not None:
                 _CACHE = result
             return result
@@ -83,17 +117,54 @@ class OpenRouterProfile(ProviderProfile):
         the same backend server across turns.
         """
         extra_body: dict[str, Any] = {}
+        top_level: dict[str, Any] = {}
+        extra_headers: dict[str, Any] = {}
         if supports_reasoning:
-            if reasoning_config is not None:
+            # Reasoning-mandatory Anthropic models (Claude 4.6+ / fable /
+            # future named models) use *adaptive* thinking: the model decides
+            # how much to think, and OpenRouter ignores ``reasoning.effort`` for
+            # them entirely. Sending any ``reasoning`` field is therefore both
+            # pointless and actively harmful:
+            #   - ``{enabled: false}`` → OpenRouter emits Anthropic's manual
+            #     ``thinking: {type: "disabled"}``, which these models 400 on.
+            #   - any enabled form, on a tool-continuation turn whose prior
+            #     assistant tool_call carries no thinking block (chat_completions
+            #     never replays signed thinking blocks), ALSO makes OpenRouter
+            #     emit ``thinking: {type: "disabled"}`` → the same 400 on every
+            #     turn after the first tool call.
+            # The only reliable behavior is to omit ``reasoning`` and let the
+            # model default to adaptive. See hermes-agent#42991 (disable case)
+            # and the tool-replay follow-up.
+            #
+            # ``reasoning.effort`` being ignored does NOT mean these models have
+            # no effort lever — OpenRouter honors the requested effort on the
+            # top-level ``verbosity`` field instead (it maps to Anthropic's
+            # ``output_config.effort``; ``reasoning.effort`` is accepted but
+            # ignored — confirmed by OpenRouter's Claude migration docs and a
+            # live token-spend probe in hermes-agent#43432). Route the existing
+            # ``reasoning_config["effort"]`` (sourced from
+            # ``agent.reasoning_effort``) onto ``verbosity`` so the knob the user
+            # already sets keeps working for these models. We still send NO
+            # ``reasoning`` field, preserving the #42991 400 fix.
+            if _anthropic_reasoning_is_mandatory(model):
+                cfg = reasoning_config or {}
+                effort = cfg.get("effort")
+                # Only emit when effort is actually requested and reasoning
+                # isn't explicitly disabled. Otherwise omit ``verbosity`` so the
+                # model keeps its own adaptive default (``high``).
+                if cfg.get("enabled", True) is not False and effort and effort != "none":
+                    top_level["verbosity"] = effort
+            elif reasoning_config is not None:
                 extra_body["reasoning"] = dict(reasoning_config)
             else:
                 extra_body["reasoning"] = {"enabled": True, "effort": "medium"}
 
-        extra_headers: dict[str, Any] = {}
         if session_id and model and model.startswith(("x-ai/grok-", "xai/grok-")):
             extra_headers["x-grok-conv-id"] = session_id
+        if extra_headers:
+            top_level["extra_headers"] = extra_headers
 
-        return extra_body, {"extra_headers": extra_headers} if extra_headers else {}
+        return extra_body, top_level
 
 
 openrouter = OpenRouterProfile(

package/agent/plugins/model-providers/xiaomi/__init__.py

@@ -9,6 +9,8 @@ xiaomi = ProviderProfile(
     env_vars=("XIAOMI_API_KEY",),
     base_url="https://api.xiaomimimo.com/v1",
     supports_health_check=False,  # /v1/models returns 401 even with valid key
+    supports_vision=True,  # mimo-v2-omni is vision-capable
+    supports_vision_tool_messages=False,  # rejects list-type tool content (400 "text is not set")
 )
 
 register_provider(xiaomi)

package/agent/plugins/model-providers/zai/__init__.py

@@ -11,6 +11,7 @@ zai = ProviderProfile(
     description="Z.AI / GLM — Zhipu AI models",
     signup_url="https://z.ai/",
     fallback_models=(
+        "glm-5.2",
         "glm-5",
         "glm-4-9b",
     ),

package/agent/plugins/observability/langfuse/__init__.py

@@ -54,6 +54,15 @@ class TraceState:
 
 _STATE_LOCK = threading.Lock()
 _TRACE_STATE: Dict[str, TraceState] = {}
+# Hard cap on live trace state. Each turn keys _TRACE_STATE by a unique
+# turn_id, and an entry is normally reclaimed by _finish_trace when a turn
+# ends cleanly (final response has content and no tool calls). A turn that
+# never reaches that state — interrupted, a tool-only final step, or empty
+# final content — would otherwise linger forever, so over the cap we evict
+# the least-recently-updated entries (ending their root span first). The cap
+# is far above any realistic concurrent-live-turn working set; it exists only
+# to bound the leak from non-finalizing turns, not to limit concurrency.
+_MAX_TRACE_STATE = 256
 _LANGFUSE_CLIENT = None
 _READ_FILE_LINE_RE = re.compile(r"^\s*(\d+)\|(.*)$")
 _READ_FILE_HEAD_LINES = 25
@@ -219,15 +228,67 @@ def _get_langfuse() -> Optional[Langfuse]:
     return _LANGFUSE_CLIENT
 
 
-def _trace_key(task_id: str, session_id: str) -> str:
+def _scope_prefix(task_id: str, session_id: str) -> str:
+    """The task/session/thread prefix shared by every trace-key shape."""
     if task_id:
-        return task_id
+        return f"task:{task_id}"
     if session_id:
         return f"session:{session_id}"
     return f"thread:{threading.get_ident()}"
 
 
-def _truncate_text(value: str, max_chars: int) -> str:
+def _trace_key(
+    task_id: str,
+    session_id: str,
+    *,
+    turn_id: str = "",
+    api_request_id: str = "",
+) -> str:
+    """Build a stable in-process trace scope key for one agent turn.
+
+    Older Hermes paths only expose ``task_id``/``session_id``. Newer paths
+    pass ``turn_id`` and ``api_request_id`` in LLM/tool hooks; when present,
+    they must scope trace state so concurrent requests sharing one task/session
+    never collide. ``turn_id`` is preferred over ``api_request_id`` so the
+    turn-level ``post_llm_call`` hook (which carries ``turn_id`` but no
+    ``api_request_id``) resolves to the same key as the request-level hooks.
+    """
+    if turn_id:
+        return f"{_scope_prefix(task_id, session_id)}:turn:{turn_id}"
+    if api_request_id:
+        return f"{_scope_prefix(task_id, session_id)}:api:{api_request_id}"
+    # Legacy shape: a bare ``task_id`` (NOT the ``task:`` prefix) when present,
+    # otherwise the session/thread prefix. Kept distinct for backward
+    # compatibility with keys minted before turn/request scoping existed.
+    if task_id:
+        return task_id
+    return _scope_prefix(task_id, session_id)
+
+
+def _is_base64_data_uri(value: str) -> bool:
+    prefix = value[:200].lower()
+    return prefix.startswith("data:") and ";base64," in prefix
+
+
+def _redact_data_uri(value: str) -> dict[str, Any]:
+    header = value.split(",", 1)[0] if "," in value else "data:"
+    media_type = header[5:].split(";", 1)[0] if header.startswith("data:") else ""
+    return {
+        "type": "data_uri",
+        "media_type": media_type or None,
+        "omitted": True,
+        "length": len(value),
+    }
+
+
+def _truncate_text(value: str, max_chars: int) -> Any:
+    # Langfuse SDK treats data:*;base64 strings as media and attempts to
+    # decode them. Truncating those strings produces invalid base64 and noisy
+    # "Error parsing base64 data URI" logs. Observability only needs metadata,
+    # not raw image/audio payloads, so redact the whole data URI before it
+    # reaches the SDK.
+    if _is_base64_data_uri(value):
+        return _redact_data_uri(value)
     if len(value) <= max_chars:
         return value
     return value[:max_chars] + f"... [truncated {len(value) - max_chars} chars]"
@@ -540,12 +601,15 @@ def _usage_and_cost(response: Any, *, provider: str, api_mode: str, model: str,
 
 
 def _start_root_trace(task_key: str, *, task_id: str, session_id: str, platform: str, provider: str, model: str,
-                      api_mode: str, messages: Any, client: Langfuse) -> TraceState:
+                      api_mode: str, messages: Any, client: Langfuse,
+                      turn_id: str = "", api_request_id: str = "") -> TraceState:
     trace_id = client.create_trace_id(seed=f"{session_id or 'sessionless'}::{task_id or task_key}")
     trace_input = _extract_last_user_message(messages)
     metadata = {
         "source": "hermes",
         "task_id": task_id,
+        "turn_id": turn_id,
+        "api_request_id": api_request_id,
         "platform": platform,
         "provider": provider,
         "model": model,
@@ -646,6 +710,30 @@ def _merge_trace_output(output: Any, state: TraceState) -> Any:
     return merged
 
 
+def _evict_stale_locked() -> None:
+    """Drop least-recently-updated trace state to make room for a new entry.
+
+    Caller MUST hold ``_STATE_LOCK`` and call this immediately before inserting
+    one new entry. Bounds the leak from turns that never reach ``_finish_trace``
+    (interrupted / tool-only final step / empty final content), whose unique
+    per-turn key would otherwise linger forever. We evict down to
+    ``_MAX_TRACE_STATE - 1`` so that the about-to-be-added entry leaves the dict
+    at ``_MAX_TRACE_STATE`` — a true ceiling. The evicted entry's root span is
+    ended so it is not left dangling on the Langfuse side.
+    """
+    over = len(_TRACE_STATE) - (_MAX_TRACE_STATE - 1)
+    if over <= 0:
+        return
+    # Oldest-first by last_updated_at; evict just enough to make room.
+    stale = sorted(_TRACE_STATE.items(), key=lambda kv: kv[1].last_updated_at)[:over]
+    for key, state in stale:
+        _TRACE_STATE.pop(key, None)
+        try:
+            state.root_span.end()
+        except Exception as exc:  # pragma: no cover - fail-open
+            _debug(f"evict stale trace failed: {exc}")
+
+
 def _finish_trace(task_key: str, *, output: Any = None) -> None:
     client = _get_langfuse()
     if client is None:
@@ -689,7 +777,8 @@ def _request_key(api_call_count: Any) -> str:
 def on_pre_llm_call(*, task_id: str = "", session_id: str = "", platform: str = "", model: str = "",
                     provider: str = "", base_url: str = "", api_mode: str = "",
                     api_call_count: int = 0, messages: Any = None, turn_type: str = "user",
-                    conversation_history: Any = None, user_message: Any = None, **_: Any) -> None:
+                    conversation_history: Any = None, user_message: Any = None,
+                    turn_id: str = "", api_request_id: str = "", **_: Any) -> None:
     # Older Hermes branches used pre_llm_call for request-scoped tracing and
     # passed the actual API messages. Current Hermes also has a turn-scoped
     # pre_llm_call used for context injection; tracing that hook creates an
@@ -706,7 +795,12 @@ def on_pre_llm_call(*, task_id: str = "", session_id: str = "", platform: str =
     # pre_llm_call with API messages directly. Current Hermes fires
     # pre_llm_call for context injection (conversation_history/user_message,
     # no messages list) — tracing that would create orphan traces.
-    task_key = _trace_key(task_id, session_id)
+    task_key = _trace_key(
+        task_id,
+        session_id,
+        turn_id=turn_id,
+        api_request_id=api_request_id,
+    )
 
     with _STATE_LOCK:
         state = _TRACE_STATE.get(task_key)
@@ -721,7 +815,10 @@ def on_pre_llm_call(*, task_id: str = "", session_id: str = "", platform: str =
                 api_mode=api_mode,
                 messages=messages,
                 client=client,
+                turn_id=turn_id,
+                api_request_id=api_request_id,
             )
+            _evict_stale_locked()
             _TRACE_STATE[task_key] = state
         state.last_updated_at = time.time()
 
@@ -746,6 +843,8 @@ def on_pre_llm_request(
     max_tokens: Any = None,
     conversation_history: Any = None,
     user_message: Any = None,
+    turn_id: str = "",
+    api_request_id: str = "",
     **_: Any,
 ) -> None:
     client = _get_langfuse()
@@ -759,7 +858,12 @@ def on_pre_llm_request(
         user_message=user_message,
     )
 
-    task_key = _trace_key(task_id, session_id)
+    task_key = _trace_key(
+        task_id,
+        session_id,
+        turn_id=turn_id,
+        api_request_id=api_request_id,
+    )
     req_key = _request_key(api_call_count)
 
     with _STATE_LOCK:
@@ -775,7 +879,10 @@ def on_pre_llm_request(
                 api_mode=api_mode,
                 messages=input_messages,
                 client=client,
+                turn_id=turn_id,
+                api_request_id=api_request_id,
             )
+            _evict_stale_locked()
             _TRACE_STATE[task_key] = state
         state.last_updated_at = time.time()
         previous = state.generations.pop(req_key, None)
@@ -804,12 +911,18 @@ def on_post_llm_call(*, task_id: str = "", session_id: str = "", provider: str =
                      api_duration: float = 0.0, finish_reason: str = "",
                      usage: Any = None, assistant_content_chars: int = 0,
                      assistant_tool_call_count: int = 0, assistant_response: Any = None,
+                     turn_id: str = "", api_request_id: str = "",
                      **_: Any) -> None:
     client = _get_langfuse()
     if client is None:
         return
 
-    task_key = _trace_key(task_id, session_id)
+    task_key = _trace_key(
+        task_id,
+        session_id,
+        turn_id=turn_id,
+        api_request_id=api_request_id,
+    )
     req_key = _request_key(api_call_count)
 
     with _STATE_LOCK:
@@ -837,8 +950,16 @@ def on_post_llm_call(*, task_id: str = "", session_id: str = "", provider: str =
     if output.get("tool_calls"):
         state.turn_tool_calls.extend(output["tool_calls"])
 
-    # Extract usage: prefer response object, fall back to usage dict from post_api_request
-    if response is not None:
+    # Extract usage: prefer a real response object that carries usage, else
+    # fall back to the usage summary dict from post_api_request.
+    #
+    # post_api_request passes `response` as a SANITIZED dict (no ``.usage``
+    # attribute) alongside a separate `usage` summary dict. Gating on
+    # ``response is not None`` here took the response-object path on that dict,
+    # where ``getattr(response, "usage", None)`` is always None — so usage and
+    # cost were silently dropped for every gateway turn. Gate on a real
+    # ``.usage`` attribute instead so the usage-dict fallback below is reached.
+    if getattr(response, "usage", None) is not None:
         usage_details, cost_details = _usage_and_cost(
             response,
             provider=provider,
@@ -919,12 +1040,18 @@ def on_post_llm_call(*, task_id: str = "", session_id: str = "", provider: str =
 
 
 def on_pre_tool_call(*, tool_name: str = "", args: Any = None, task_id: str = "",
-                     session_id: str = "", tool_call_id: str = "", **_: Any) -> None:
+                     session_id: str = "", tool_call_id: str = "",
+                     turn_id: str = "", api_request_id: str = "", **_: Any) -> None:
     client = _get_langfuse()
     if client is None:
         return
 
-    task_key = _trace_key(task_id, session_id)
+    task_key = _trace_key(
+        task_id,
+        session_id,
+        turn_id=turn_id,
+        api_request_id=api_request_id,
+    )
 
     with _STATE_LOCK:
         state = _TRACE_STATE.get(task_key)
@@ -945,8 +1072,14 @@ def on_pre_tool_call(*, tool_name: str = "", args: Any = None, task_id: str = ""
 
 
 def on_post_tool_call(*, tool_name: str = "", args: Any = None, result: Any = None,
-                      task_id: str = "", session_id: str = "", tool_call_id: str = "", **_: Any) -> None:
-    task_key = _trace_key(task_id, session_id)
+                      task_id: str = "", session_id: str = "", tool_call_id: str = "",
+                      turn_id: str = "", api_request_id: str = "", **_: Any) -> None:
+    task_key = _trace_key(
+        task_id,
+        session_id,
+        turn_id=turn_id,
+        api_request_id=api_request_id,
+    )
     observation = None
 
     with _STATE_LOCK: