fix(litellm): omit temperature for reasoning models

Share the OpenAI-compatible reasoning model detector between the native
OpenAI-compatible provider and LiteLLM-backed providers.

Skip forwarding explicit temperature values through LiteLLM for those
models, including Azure GPT-5 deployments, o-series models, and DeepSeek
reasoning routes.

Limit the LiteLLM Router check to deployments reachable from the default
entrypoint so unrelated model groups do not affect call parameters.

Author: Sanderhoff-alt

hindsight-api-slim/hindsight_api/engine/providers/litellm_llm.py

@@ -24,6 +24,7 @@
 from hindsight_api.config import DEFAULT_LLM_TIMEOUT, ENV_LLM_TIMEOUT
 from hindsight_api.engine.llm_interface import LLMInterface, OutputTooLongError
 from hindsight_api.engine.llm_trace import LLMResponseUsage, stash_response_usage
+from hindsight_api.engine.providers.model_capabilities import supports_openai_compatible_reasoning
 from hindsight_api.engine.response_models import LLMToolCall, LLMToolCallResult, TokenUsage
 from hindsight_api.metrics import get_metrics_collector
 from hindsight_api.worker.stage import set_stage
@@ -159,12 +160,18 @@ def _build_common_kwargs(
         if self._default_headers:
             kwargs.setdefault("extra_headers", dict(self._default_headers))
 
+        if self._should_omit_temperature():
+            kwargs.pop("temperature", None)
+
         # Bedrock service tier: flex (50% cheaper), priority, or reserved
         if self.model.startswith("bedrock/") and self.bedrock_service_tier is not None:
             kwargs["service_tier"] = self.bedrock_service_tier
 
         return kwargs
 
+    def _should_omit_temperature(self) -> bool:
+        return supports_openai_compatible_reasoning(self.model)
+
     # ── per-model output-tokens cap (shared with Router subclass) ────────────
     # Hindsight's defaults (e.g. retain_max_completion_tokens=64000) target
     # high-capacity models. When a configured deployment supports fewer

hindsight-api-slim/hindsight_api/engine/providers/litellm_router_llm.py

@@ -37,6 +37,7 @@
 from typing import Any
 
 from hindsight_api.engine.providers.litellm_llm import LiteLLMLLM
+from hindsight_api.engine.providers.model_capabilities import supports_openai_compatible_reasoning
 
 logger = logging.getLogger(__name__)
 
@@ -94,6 +95,7 @@ def __init__(
         # deployment Router picks. Uses LiteLLM's own per-model registry; unknown
         # models contribute no cap. See LiteLLMLLM._cap_max_completion_tokens.
         self._router_output_cap = self._compute_router_output_cap(config)
+        self._router_omits_temperature = self._config_has_temperature_rejecting_model(config)
 
         logger.info("LiteLLM Router initialized; entrypoint model_name=%r", _ENTRYPOINT_MODEL_NAME)
 
@@ -130,6 +132,48 @@ def _resolve_completion_model(self, response: Any) -> str:
     def _get_model_output_cap(self) -> int | None:
         return self._router_output_cap
 
+    def _should_omit_temperature(self) -> bool:
+        return bool(getattr(self, "_router_omits_temperature", False))
+
+    def _config_has_temperature_rejecting_model(self, config: dict[str, Any]) -> bool:
+        reachable_model_names = self._reachable_model_names(config)
+        for deployment in (config.get("model_list") or []) if isinstance(config, dict) else []:
+            if not isinstance(deployment, dict):
+                continue
+            model_name = deployment.get("model_name")
+            if model_name not in reachable_model_names:
+                continue
+            params = deployment.get("litellm_params") or {}
+            model_str = params.get("model") if isinstance(params, dict) else None
+            if model_str and supports_openai_compatible_reasoning(model_str):
+                return True
+        return False
+
+    def _reachable_model_names(self, config: dict[str, Any]) -> set[str]:
+        reachable = {_ENTRYPOINT_MODEL_NAME}
+        pending = [_ENTRYPOINT_MODEL_NAME]
+        fallback_specs = []
+        for key in ("fallbacks", "context_window_fallbacks"):
+            value = config.get(key) if isinstance(config, dict) else None
+            if isinstance(value, list):
+                fallback_specs.extend(value)
+
+        while pending:
+            source = pending.pop()
+            for spec in fallback_specs:
+                if not isinstance(spec, dict):
+                    continue
+                targets = spec.get(source)
+                if isinstance(targets, str):
+                    targets = [targets]
+                if not isinstance(targets, list):
+                    continue
+                for target in targets:
+                    if isinstance(target, str) and target not in reachable:
+                        reachable.add(target)
+                        pending.append(target)
+        return reachable
+
     def _build_common_kwargs(
         self,
         messages: list[dict[str, Any]],
@@ -144,7 +188,7 @@ def _build_common_kwargs(
         }
         if max_completion_tokens is not None:
             kwargs["max_completion_tokens"] = self._cap_max_completion_tokens(max_completion_tokens)
-        if temperature is not None:
+        if temperature is not None and not self._should_omit_temperature():
             kwargs["temperature"] = temperature
 
         # Forward operator-configured default headers as ``extra_headers`` so they
@@ -154,8 +198,9 @@ def _build_common_kwargs(
         # concern, so we inject them here too — mirroring the base provider.
         # ``setdefault`` keeps any explicit per-call ``extra_headers`` authoritative;
         # a per-call copy prevents LiteLLM/downstream from mutating the stored dict.
-        if self._default_headers:
-            kwargs.setdefault("extra_headers", dict(self._default_headers))
+        default_headers = getattr(self, "_default_headers", {})
+        if default_headers:
+            kwargs.setdefault("extra_headers", dict(default_headers))
         return kwargs
 
     async def verify_connection(self) -> None:

hindsight-api-slim/hindsight_api/engine/providers/model_capabilities.py

@@ -0,0 +1,11 @@
+"""Shared provider model capability helpers."""
+
+
+def supports_openai_compatible_reasoning(model: str) -> bool:
+    """Return True for OpenAI-compatible reasoning model names."""
+    model_lower = (model or "").lower()
+    if "deepseek" in model_lower:
+        # DeepSeek v4-flash is the non-thinking route. Treating every
+        # DeepSeek model as reasoning injects unsupported reasoning params.
+        return any(x in model_lower for x in ["v4-pro", "reasoner", "r1", "thinking"])
+    return any(x in model_lower for x in ["gpt-5", "o1", "o3"])

hindsight-api-slim/hindsight_api/engine/providers/openai_compatible_llm.py

@@ -38,6 +38,7 @@
 from hindsight_api.engine.bank_attribution import apply_bank_attribution
 from hindsight_api.engine.llm_interface import LLMInterface, OutputTooLongError, ProviderRateLimitResetError
 from hindsight_api.engine.llm_trace import LLMResponseUsage, stash_response_usage
+from hindsight_api.engine.providers.model_capabilities import supports_openai_compatible_reasoning
 from hindsight_api.engine.response_models import LLMToolCall, LLMToolCallResult, TokenUsage
 from hindsight_api.metrics import get_metrics_collector
 from hindsight_api.worker.stage import set_stage
@@ -594,13 +595,7 @@ async def verify_connection(self) -> None:
 
     def _supports_reasoning_model(self) -> bool:
         """Check if the current model is a reasoning model (o1, o3, GPT-5, DeepSeek)."""
-        model_lower = self.model.lower()
-        if "deepseek" in model_lower:
-            # DeepSeek v4-flash is the non-thinking route. Treating every
-            # DeepSeek model as a reasoning model injects reasoning_effort,
-            # which conflicts with thinking-disabled flash calls.
-            return any(x in model_lower for x in ["v4-pro", "reasoner", "r1", "thinking"])
-        return any(x in model_lower for x in ["gpt-5", "o1", "o3"])
+        return supports_openai_compatible_reasoning(self.model)
 
     def _get_max_reasoning_tokens(self) -> int | None:
         """Get max reasoning tokens for reasoning models."""

hindsight-api-slim/tests/test_llm_extra_body.py

@@ -349,15 +349,15 @@ class StructuredAnswer(BaseModel):
 # ─── LiteLLM ──────────────────────────────────────────────────────────────────
 
 
-def _make_litellm_provider(extra_body=None, default_headers=None):
+def _make_litellm_provider(extra_body=None, default_headers=None, model="gpt-4o"):
     pytest.importorskip("litellm")
     from hindsight_api.engine.providers.litellm_llm import LiteLLMLLM
 
     return LiteLLMLLM(
         provider="litellm",
         api_key="fake-key",
         base_url="",
-        model="gpt-4o",
+        model=model,
         extra_body=extra_body,
         default_headers=default_headers,
     )
@@ -375,6 +375,18 @@ def _fake_litellm_response():
     return resp
 
 
+async def _call_litellm_provider(provider, temperature=0.1):
+    provider._acompletion = AsyncMock(return_value=_fake_litellm_response())
+    with patch("hindsight_api.engine.providers.litellm_llm.get_metrics_collector"):
+        await provider.call(
+            messages=[{"role": "user", "content": "hi"}],
+            temperature=temperature,
+            scope="test",
+            max_retries=0,
+        )
+    return provider._acompletion.call_args.kwargs
+
+
 def test_litellm_stores_extra_body():
     provider = _make_litellm_provider(extra_body=EXTRA_BODY)
     assert provider._extra_body == EXTRA_BODY
@@ -406,6 +418,33 @@ async def test_litellm_explicit_param_wins_over_extra_body():
     assert provider._acompletion.call_args.kwargs.get("temperature") == 0.9
 
 
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    ("model", "expected_temperature"),
+    [
+        ("azure/gpt-5.5", None),
+        ("openai/o3-mini", None),
+        ("deepseek/deepseek-reasoner", None),
+        ("deepseek/deepseek-v4-flash", 0.1),
+    ],
+)
+async def test_litellm_temperature_for_reasoning_models(model, expected_temperature):
+    kwargs = await _call_litellm_provider(_make_litellm_provider(model=model))
+
+    if expected_temperature is None:
+        assert "temperature" not in kwargs
+    else:
+        assert kwargs["temperature"] == expected_temperature
+
+
+@pytest.mark.asyncio
+async def test_litellm_gpt5_omits_extra_body_temperature():
+    provider = _make_litellm_provider(extra_body=EXTRA_BODY, model="azure/gpt-5.5")
+    kwargs = await _call_litellm_provider(provider, temperature=None)
+    assert "temperature" not in kwargs
+    assert kwargs.get("top_p") == 0.9
+
+
 def test_litellm_router_forwards_extra_body():
     """The Router subclass forwards extra_body through to the shared LiteLLM base."""
     pytest.importorskip("litellm")

hindsight-api-slim/tests/test_llm_router_provider.py

@@ -70,6 +70,22 @@ def mock_router_response() -> MagicMock:
     return response
 
 
+def _router_config(
+    default_model: str,
+    *,
+    extra_model: str | None = None,
+    extra_name: str = "fallback",
+    fallbacks: list[dict[str, list[str]]] | None = None,
+) -> dict[str, Any]:
+    model_list = [{"model_name": "default", "litellm_params": {"model": default_model, "api_key": "sk"}}]
+    if extra_model:
+        model_list.append({"model_name": extra_name, "litellm_params": {"model": extra_model, "api_key": "sk"}})
+    config: dict[str, Any] = {"model_list": model_list}
+    if fallbacks:
+        config["fallbacks"] = fallbacks
+    return config
+
+
 # --- config parsing ----------------------------------------------------------
 
 
@@ -202,6 +218,20 @@ def _make_router_provider(config: dict[str, Any], mock_router: Any) -> LiteLLMRo
 
 
 class TestRouterCall:
+    async def _call_with_temperature(self, config, mock_router_response):
+        mock_router = MagicMock()
+        mock_router.acompletion = AsyncMock(return_value=mock_router_response)
+        provider = _make_router_provider(config, mock_router)
+        provider._router_omits_temperature = provider._config_has_temperature_rejecting_model(config)
+
+        await provider.call(
+            messages=[{"role": "user", "content": "hi"}],
+            temperature=0.1,
+            max_retries=0,
+        )
+
+        return mock_router.acompletion.await_args.kwargs
+
     @pytest.mark.asyncio
     async def test_plain_text_call_targets_default_entrypoint(self, two_step_config, mock_router_response):
         mock_router = MagicMock()
@@ -314,6 +344,35 @@ async def test_no_cap_when_litellm_registry_has_no_data(self, two_step_config, m
         kwargs = mock_router.acompletion.await_args.kwargs
         assert kwargs["max_completion_tokens"] == 64000
 
+    @pytest.mark.asyncio
+    @pytest.mark.parametrize(
+        ("config", "expected_temperature"),
+        [
+            (_router_config("azure/gpt-5.5"), None),
+            (_router_config("openai/gpt-4o-mini", extra_model="azure/gpt-5.5", extra_name="unused"), 0.1),
+            (
+                _router_config(
+                    "openai/gpt-4o-mini",
+                    extra_model="azure/gpt-5.5",
+                    fallbacks=[{"default": ["fallback"]}],
+                ),
+                None,
+            ),
+        ],
+    )
+    async def test_router_temperature_for_reachable_reasoning_models(
+        self,
+        config,
+        expected_temperature,
+        mock_router_response,
+    ):
+        kwargs = await self._call_with_temperature(config, mock_router_response)
+        assert kwargs["model"] == "default"
+        if expected_temperature is None:
+            assert "temperature" not in kwargs
+        else:
+            assert kwargs["temperature"] == expected_temperature
+
     @pytest.mark.asyncio
     async def test_call_with_tools(self, two_step_config):
         response = MagicMock()