fix(eval): strip restricted temperature for default-only models in mocks

packages/uipath/src/uipath/eval/mocks/_structured_output.py

@@ -28,6 +28,31 @@
 logger = logging.getLogger(__name__)
 
 
+def _model_supports_only_default_temperature(model: str | None) -> bool:
+    """Return whether ``model`` rejects any non-default ``temperature``.
+
+    The gpt-5 family and the o-series reasoning models only accept the default
+    temperature (1) on the LLM Gateway and return HTTP 400 for any explicit
+    value, including ``0``.
+    """
+    name = (model or "").lower()
+    return name.startswith(("gpt-5", "o1", "o3", "o4"))
+
+
+def _normalize_completion_kwargs(completion_kwargs: dict[str, Any]) -> dict[str, Any]:
+    """Strip parameters a model rejects before calling the gateway.
+
+    Models that only accept the default temperature reject an explicit
+    ``temperature``, so it is dropped here rather than forwarded verbatim
+    (SRE-607465 / PC-4769).
+    """
+    if "temperature" not in completion_kwargs:
+        return completion_kwargs
+    if not _model_supports_only_default_temperature(completion_kwargs.get("model")):
+        return completion_kwargs
+    return {k: v for k, v in completion_kwargs.items() if k != "temperature"}
+
+
 def _inline_defs(
     schema: dict[str, Any],
 ) -> tuple[dict[str, Any], dict[str, Any]]:
@@ -248,6 +273,7 @@ async def generate_structured_output(
     completion_kwargs: dict[str, Any],
 ) -> Any:
     """Generate structured output using the strategy for the requested model."""
+    completion_kwargs = _normalize_completion_kwargs(completion_kwargs)
     strategy = _strategy_for_model(completion_kwargs.get("model"))
     return await strategy.generate(
         llm,

packages/uipath/tests/cli/eval/mocks/test_structured_output.py

@@ -293,3 +293,45 @@ async def test_openai_models_prefer_response_format():
     assert result == {"a": 1}
     assert len(llm.calls) == 1
     assert "response_format" in llm.calls[0]
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    "model",
+    [
+        "gpt-5-2025-08-07",
+        "gpt-5-mini",
+        "o1-2024-12-17",
+        "o3-mini-2025-01-31",
+    ],
+)
+async def test_default_only_temperature_models_drop_temperature(model: str):
+    # Models like gpt-5 / o-series only accept the default temperature (1); the
+    # LLM gateway returns HTTP 400 when temperature=0 is forwarded (SRE-607465 /
+    # PC-4769). The mock path must strip the restricted temperature before
+    # calling chat_completions.
+    llm = _FakeLLM([_response(SimpleNamespace(content='{"a": 1}', tool_calls=None))])
+    await generate_structured_output(
+        llm,
+        [{"role": "user", "content": "x"}],
+        schema={"type": "object"},
+        response_format_name="OutputSchema",
+        description="d",
+        completion_kwargs={"model": model, "temperature": 0},
+    )
+    assert "temperature" not in llm.calls[0]
+    assert llm.calls[0]["model"] == model
+
+
+@pytest.mark.asyncio
+async def test_standard_models_keep_temperature():
+    llm = _FakeLLM([_response(SimpleNamespace(content='{"a": 1}', tool_calls=None))])
+    await generate_structured_output(
+        llm,
+        [{"role": "user", "content": "x"}],
+        schema={"type": "object"},
+        response_format_name="OutputSchema",
+        description="d",
+        completion_kwargs={"model": "gpt-4.1-mini-2025-04-14", "temperature": 0},
+    )
+    assert llm.calls[0]["temperature"] == 0