| 123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265 |
- """Provider reasoning-parameter policy for ``AIAgent``.
- When ``reasoning`` extra_body is safe to send, LM Studio / Ollama / endpoint-profile capability probes,
- ``reasoning_content`` echo families, and strict-API tool-call sanitising.
- Extracted from ``run_agent.py``; every method resolves through ``AIAgent``'s MRO unchanged.
- """
- import logging
- import time
- from agent.lazy_forward import forward as _forward, forward_static as _forward_static
- from agent.message_sanitization import matches_reasoning_echo_family
- from utils import base_url_host_matches
- # Static OpenRouter fallback when the live /v1/models capability cache is cold.
- _OPENROUTER_REASONING_PREFIXES = (
- "deepseek/", "anthropic/", "openai/", "x-ai/", "google/gemini-2", "google/gemma-4",
- "qwen/qwen3", "tencent/hy", "xiaomi/",
- )
- # Hosts whose reasoning gate core owns: a plugin claiming one in ``hostnames`` never replaces it.
- _CORE_REASONING_GATED_HOSTS = ("nousresearch.com", "ai-gateway.vercel.sh", "ollama.com")
- # Probe results cache per (model, base_url). Definitive values cache permanently; an
- # "unknown" value (empty list / None) caches 60s so a transient failure neither sticks
- # for the session nor round-trips every turn.
- _PROBE_TTL_S = 60
- def _cached_probe(agent, cache_attr: str, probe, unknown, definitive):
- """``probe(model, base_url, api_key)`` once per (model, base_url); ``unknown`` is what a raising probe
- yields, ``definitive(value)`` decides permanent vs 60s TTL."""
- cache = getattr(agent, cache_attr, None)
- if cache is None:
- cache = {}
- setattr(agent, cache_attr, cache)
- key = (agent.model, agent.base_url)
- cached = cache.get(key)
- if cached is not None:
- value, ts = cached
- if definitive(value) or (time.monotonic() - ts) < _PROBE_TTL_S:
- return value
- try:
- value = probe(agent.model, agent.base_url, getattr(agent, "api_key", ""))
- except Exception:
- value = unknown
- cache[key] = (value, time.monotonic())
- return value
- def unset_reasoning_default(agent) -> dict | None:
- """Reasoning config for a main-loop request whose ``agent.reasoning_effort`` is unset.
- Asks the active provider profile (``ProviderProfile.default_reasoning_config``; the custom /
- OpenAI-compatible profile answers medium) so a route's own default never silently applies —
- kimi-k3 behind a relay defaults to ``max``, 3x the reasoning tokens of medium. Resolved at
- request time, so ``/model`` and fallback activation re-evaluate it. None keeps the field off
- the wire: a non chat-completions transport, a profile without a default (those decide inside
- ``build_api_kwargs_extras``), a model the catalog / ``model_overrides`` mark
- ``supports_reasoning: false``, or a local Ollama model whose ``/api/show`` lacks ``thinking``.
- """
- if getattr(agent, "api_mode", None) != "chat_completions":
- return None
- provider = str(getattr(agent, "provider", "") or "")
- model = str(getattr(agent, "model", "") or "")
- try:
- from providers import get_provider_profile
- profile = get_provider_profile(provider)
- default = profile.default_reasoning_config(model) if profile is not None else None
- except Exception:
- return None
- if not default:
- return None
- try:
- from agent.models_dev import get_model_capabilities
- caps = get_model_capabilities(provider, model, allow_network=False)
- except Exception:
- caps = None
- if caps is not None and caps.supports_reasoning is False:
- return None
- # ``_ollama_num_ctx`` is only ever set for a server detected as Ollama (agent_init); Ollama
- # 400s ``reasoning_effort`` on a model pulled without the thinking capability.
- if getattr(agent, "_ollama_num_ctx", None) and not agent._ollama_supports_thinking_cached():
- return None
- return dict(default)
- class ReasoningParamsMixin:
- """Reasoning-parameter gating and echo policy (see module docstring)."""
- def _supports_reasoning_extra_body(self) -> bool:
- """True when reasoning extra_body is safe to send: OpenRouter forwards unknown extra_body upstream and
- some routes 400 on ``reasoning``, so gate to known reasoning-capable families and direct Nous Portal."""
- url = self._base_url_lower
- if base_url_host_matches(url, "nousresearch.com") or base_url_host_matches(url, "ai-gateway.vercel.sh"):
- return True
- if (profile := self._reasoning_endpoint_profile()) is not None:
- # A claimed endpoint gates on its own catalog: reasoning goes out iff it declares a payload.
- from agent.provider_route_hooks import ask_profile
- return ask_profile(profile, "reasoning_payload", None, self.model, None,
- base_url=self.base_url, scope=self) is not None
- if (self.provider or "").strip().lower() == "lmstudio":
- # "off-only" (or absent) means no real reasoning capability.
- return any(opt and opt != "off" for opt in self._lmstudio_reasoning_options_cached())
- if base_url_host_matches(url, "ollama.com"):
- # Ollama Cloud: /api/show capabilities are authoritative.
- return self._ollama_supports_thinking_cached()
- if not self._is_openrouter_url() or base_url_host_matches(url, "api.mistral.ai"):
- return False
- # Live-catalog metadata first (OpenRouter /v1/models supported_parameters) — the static prefix
- # allowlist repeatedly went stale one vendor at a time. Unknown falls back to the static list.
- try:
- from hermes_cli.models_reasoning_caps import openrouter_model_reasoning_capabilities, warm_openrouter_reasoning_caps_async
- caps = openrouter_model_reasoning_capabilities(self.model)
- if caps is None:
- warm_openrouter_reasoning_caps_async() # cache cold — warm in the background, never block
- except Exception:
- caps = None
- if caps is not None:
- return bool(caps.get("supports_reasoning"))
- model = (self.model or "").lower()
- return any(model.startswith(prefix) for prefix in _OPENROUTER_REASONING_PREFIXES)
- def _lmstudio_reasoning_options_cached(self) -> list[str]:
- """LM Studio's published reasoning ``allowed_options`` (gate + clamp so toggle models don't 400 on ``high``)."""
- try:
- from hermes_cli.models_local import lmstudio_model_reasoning_options
- except Exception:
- return []
- return _cached_probe(self, "_lm_reasoning_opts_cache", lmstudio_model_reasoning_options, [], bool)
- def _ollama_supports_thinking_cached(self) -> bool:
- """True only if Ollama's ``/api/show`` declares the ``thinking`` capability."""
- try:
- from hermes_cli.models_local import ollama_model_supports_thinking
- except Exception:
- return False
- return bool(_cached_probe(self, "_ollama_thinking_cache", ollama_model_supports_thinking, None, lambda v: v is not None))
- def _reasoning_endpoint_profile(self):
- """The profile claiming this endpoint (``providers.profile_for_host``) that gates its reasoning, or None.
- Endpoints core gates itself (Nous Portal, Vercel AI Gateway, LM Studio, Ollama Cloud, OpenRouter) keep
- core's gate even when a plugin claims their host for metadata or auxiliary routing."""
- url = self._base_url_lower
- if ((self.provider or "").strip().lower() == "lmstudio" or self._is_openrouter_url()
- or any(base_url_host_matches(url, host) for host in _CORE_REASONING_GATED_HOSTS)):
- return None
- from providers import profile_for_host
- return profile_for_host(str(getattr(self, "base_url", None) or ""))
- def _endpoint_reasoning_payload(self) -> dict | None:
- """``reasoning_payload`` of the profile gating this endpoint's reasoning for the agent's model and
- ``reasoning_config`` (``_reasoning_endpoint_profile``), or None when no profile does."""
- base_url = getattr(self, "base_url", None)
- profile = self._reasoning_endpoint_profile()
- if profile is None:
- return None
- from agent.provider_route_hooks import ask_profile
- return ask_profile(profile, "reasoning_payload", None, self.model, self.reasoning_config,
- base_url=str(base_url), scope=self)
- def _build_assistant_message(self, assistant_message, finish_reason: str) -> dict:
- """``build_assistant_message``; a held turn's row with visible text comes back typed hidden
- (``withhold_held_row``) until ``finish_text_response`` publishes it as the final."""
- from agent.chat_completion_helpers import build_assistant_message
- from agent.stream_delivery import withhold_held_row
- return withhold_held_row(self, build_assistant_message(self, assistant_message, finish_reason))
- def _reasoning_replay_route(self):
- """Reasoning replay decision for the active route (``message_sanitization.reasoning_replay_route``).
- Cached per (api_mode, provider, model, base_url, opt-in, rejected keys) — called many times
- per turn — and recomputed after ``switch_model()`` / fallback activation or a recorded
- field rejection, the only events allowed to change the replayed bytes mid-session.
- """
- from agent.message_sanitization import reasoning_replay_route, rejected_reasoning_carriers
- rejected = rejected_reasoning_carriers(self)
- key = (getattr(self, "api_mode", None), self.provider, self.model, getattr(self, "_base_url_lower", self.base_url),
- self._reasoning_echo_opt_in(), rejected)
- cached = getattr(self, "_reasoning_route_cache", None)
- if cached is not None and cached[0] == key:
- return cached[1]
- route = reasoning_replay_route(key[0], self.provider, self.model, self.base_url,
- echo_opt_in=key[4], rejected=rejected)
- self._reasoning_route_cache = (key, route)
- return route
- def _needs_thinking_reasoning_pad(self) -> bool:
- """True for the must-echo tier: DeepSeek / Kimi / MiMo thinking 400 on a replayed tool-call turn
- without a non-empty ``reasoning_content`` (#15250, #17400), plus a ``model.reasoning_echo`` opt-in
- for gateways proxying them. Never true on a route that does not read the field."""
- return self._reasoning_replay_route().pad
- def _reasoning_echo_opt_in(self) -> bool:
- """``model.reasoning_echo``: treat the *current* provider as must-echo (pad tool-call turns) — for
- gateways proxying DeepSeek/Kimi/MiMo that the host rules miss. Replay itself is on by default;
- fallback activation swaps the flag and ``restore_primary_runtime()`` restores it."""
- return bool(getattr(self, "_reasoning_echo_flag", False))
- @staticmethod
- def _read_reasoning_echo_from_config() -> bool:
- """Read ``model.reasoning_echo`` from config; False on any error."""
- try:
- from hermes_cli.config import load_config_readonly
- return bool((load_config_readonly().get("model") or {}).get("reasoning_echo"))
- except Exception as exc: # health: allow BLE001 -- a config failure must fall back, not fail init
- # Log the class, never the config value or exception traceback.
- logging.getLogger("run_agent").warning(
- "Agent configuration for model.reasoning_echo could not be loaded (%s); "
- "using the default reasoning echo setting. Check the active Hermes Python "
- "environment and dependencies; run 'hermes config check' after "
- "configuration imports work again.",
- type(exc).__name__,
- )
- return False
- # Echo families are host/provider-driven, not model-name-driven: aggregators re-exporting Kimi reject the
- # echo. Rule table: ``message_sanitization._REASONING_ECHO_RULES``. Kimi deliberately passes the raw
- # provider and no model (its rule matches exact provider ids + hosts only).
- def _needs_kimi_tool_reasoning(self) -> bool:
- """True when the current provider is Kimi / Moonshot thinking mode."""
- return matches_reasoning_echo_family("kimi", self.provider, None, self.base_url)
- def _needs_deepseek_tool_reasoning(self) -> bool:
- """True when the current provider is DeepSeek thinking mode (omitting the echo is an HTTP 400).
- DeepSeek V4 thinking mode requires ``reasoning_content`` on every assistant tool-call turn; omitting
- it causes HTTP 400 when the message is replayed in a subsequent API request (#15250).
- """
- return matches_reasoning_echo_family("deepseek", (self.provider or "").lower(), self.model, self.base_url)
- def _needs_mimo_tool_reasoning(self) -> bool:
- """True when the current provider is Xiaomi MiMo thinking mode."""
- return matches_reasoning_echo_family("mimo", (self.provider or "").lower(), self.model, self.base_url)
- _copy_reasoning_content_for_api = _forward("agent.agent_runtime_helpers", "copy_reasoning_content_for_api")
- _reapply_reasoning_echo_for_provider = _forward("agent.agent_runtime_helpers", "reapply_reasoning_echo_for_provider")
- @staticmethod
- def _sanitize_tool_calls_for_strict_api(api_msg: dict, model: str | None = None, *, base_url: str | None = None,
- provider: str | None = None) -> dict:
- """Strip Codex Responses fields from tool_calls for strict Chat Completions APIs (Mistral, Fireworks
- 400/422 on unknown fields). ``extra_content`` (Gemini thought_signature) is kept only on Gemini routes
- (model family, Gemini host or provider). Builds new dicts so the internal history keeps the Codex
- fields for a later fallback."""
- tool_calls = api_msg.get("tool_calls")
- if not isinstance(tool_calls, list):
- return api_msg
- from agent.transports.chat_completions import _model_consumes_thought_signature
- keep_signature = _model_consumes_thought_signature(model, base_url, provider)
- strip = {"call_id", "response_item_id"} | (set() if keep_signature else {"extra_content"})
- api_msg["tool_calls"] = [{k: v for k, v in tc.items() if k not in strip} if isinstance(tc, dict) else tc
- for tc in tool_calls]
- return api_msg
- _sanitize_tool_call_arguments = _forward_static("agent.agent_runtime_helpers", "sanitize_tool_call_arguments")
- def _should_sanitize_tool_calls(self) -> bool:
- """True for every non-Codex API: Codex Responses fields are not Chat Completions schema and 400 elsewhere."""
- return self.api_mode != "codex_responses"
|