fix(extensions): nest extra_body on the any-llm chat path

src/agents/extensions/models/any_llm_model.py

@@ -1363,12 +1363,12 @@ def _consume_background_cleanup_task_result(task: asyncio.Future[Any]) -> None:
 
     def _build_chat_extra_kwargs(self, model_settings: ModelSettings) -> dict[str, Any]:
         extra_kwargs: dict[str, Any] = {}
-        if model_settings.extra_query:
+        if model_settings.extra_query is not None:
             extra_kwargs["extra_query"] = copy(model_settings.extra_query)
-        if model_settings.metadata:
+        if model_settings.metadata is not None:
             extra_kwargs["metadata"] = copy(model_settings.metadata)
-        if isinstance(model_settings.extra_body, dict):
-            extra_kwargs.update(model_settings.extra_body)
+        if model_settings.extra_body is not None:
+            extra_kwargs["extra_body"] = copy(model_settings.extra_body)
         if model_settings.extra_args:
             extra_kwargs.update(model_settings.extra_args)
         return extra_kwargs

tests/models/test_any_llm_model.py

@@ -371,6 +371,47 @@ def __bool__(self) -> bool:
     assert provider.chat_calls[0]["reasoning_effort"] == "low"
 
 
+@pytest.mark.allow_call_model_methods
+@pytest.mark.asyncio
+async def test_any_llm_chat_nests_extra_body_instead_of_flattening(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    provider = FakeAnyLLMProvider(supports_responses=False, chat_response=_chat_completion("Hello"))
+    module, _create_calls = _import_any_llm_module(monkeypatch, provider)
+    model = module.AnyLLMModel(model="openrouter/openai/gpt-5.4-mini")
+    extra_body = {"cached_content": "some_cache", "foo": 123, "temperature": 0.9}
+    settings = ModelSettings(
+        temperature=0.1,
+        extra_body=extra_body,
+        extra_query={},
+        metadata={},
+    )
+
+    await model.get_response(
+        system_instructions=None,
+        input="hi",
+        model_settings=settings,
+        tools=[],
+        output_schema=None,
+        handoffs=[],
+        tracing=ModelTracing.DISABLED,
+        previous_response_id=None,
+        conversation_id=None,
+        prompt=None,
+    )
+
+    call = provider.chat_calls[0]
+    assert call["temperature"] == 0.1
+    assert call["extra_body"] == extra_body
+    assert call["extra_body"] is not extra_body
+    assert call["extra_query"] == {}
+    assert call["metadata"] == {}
+    assert "cached_content" not in call
+    assert "foo" not in call
+    extra_body["foo"] = 999
+    assert call["extra_body"]["foo"] == 123
+
+
 @pytest.mark.allow_call_model_methods
 @pytest.mark.asyncio
 @pytest.mark.parametrize("provider_name", ["gemini", "vertexai"])