feat(openai): Include support for reasoning blocks when using openai compatible llm gateways

libs/partners/openai/langchain_openai/chat_models/base.py

@@ -163,6 +163,9 @@ def _convert_dict_to_message(_dict: Mapping[str, Any]) -> BaseMessage:
         # Also OpenAI returns None for tool invocations
         content = _dict.get("content", "") or ""
         additional_kwargs: dict = {}
+        # add reasoning blocks if available
+        if reasoning_content := _dict.get("reasoning_content"):
+            additional_kwargs["reasoning_content"] = reasoning_content
         if function_call := _dict.get("function_call"):
             additional_kwargs["function_call"] = dict(function_call)
         tool_calls = []
@@ -377,6 +380,8 @@ def _convert_delta_to_message_chunk(
     if role == "user" or default_class == HumanMessageChunk:
         return HumanMessageChunk(content=content, id=id_)
     if role == "assistant" or default_class == AIMessageChunk:
+        if reasoning_content := _dict.get("reasoning_content"):
+            additional_kwargs["reasoning_content"] = reasoning_content
         return AIMessageChunk(
             content=content,
             additional_kwargs=additional_kwargs,
@@ -1017,6 +1022,9 @@ def _convert_chunk_to_generation_chunk(
     ) -> ChatGenerationChunk | None:
         if chunk.get("type") == "content.delta":  # From beta.chat.completions.stream
             return None
+        # add reasoning blocks if available
+        if reasoning_content := chunk.get("reasoning_content"):
+            chunk["additional_kwargs"]["reasoning_content"] = reasoning_content
         token_usage = chunk.get("usage")
         choices = (
             chunk.get("choices", [])

libs/partners/openai/pyproject.toml

@@ -44,6 +44,7 @@ test = [
     "langchain",
     "langchain-core",
     "langchain-tests",
+    "httpx>=0.28.1",
 ]
 lint = ["ruff>=0.13.1,<0.14.0"]
 dev = ["langchain-core"]

libs/partners/openai/tests/integration_tests/chat_models/test_base.py

@@ -27,7 +27,7 @@
     _validate_tool_call_message,
     magic_function,
 )
-from pydantic import BaseModel, Field, field_validator
+from pydantic import BaseModel, Field, SecretStr, field_validator
 from typing_extensions import TypedDict
 
 from langchain_openai import ChatOpenAI
@@ -1320,3 +1320,30 @@ async def test_schema_parsing_failures_responses_api_async() -> None:
         assert e.response is not None  # type: ignore[attr-defined]
     else:
         raise AssertionError
+
+
+# Test thinking models with LiteLLM proxy
+@pytest.mark.scheduled
+@pytest.mark.skipif(
+    os.environ.get("REASONING_BASE_URL") is None
+    or os.environ.get("REASONING_API_KEY") is None,
+    reason="REASONING_BASE_URL or REASONING_API_KEY is not set",
+)
+def test_with_reasoning_proxy() -> None:
+    """Test reasoning models with proxy."""
+    chat = ChatOpenAI(
+        model="claude-sonnet-4-5-20250929",
+        reasoning_effort="medium",
+        base_url=os.environ["REASONING_BASE_URL"],
+        api_key=SecretStr(os.environ["REASONING_API_KEY"]),
+        max_retries=3,
+        # Disable SSL verification for self-signed certificates
+        http_client=httpx.Client(verify=False),  # noqa: S501
+    )
+    # Using a prompt that will trigger reasoning
+    message = HumanMessage(content="Reason and think about the meaning of life")
+    response = chat.invoke([message])
+    assert isinstance(response, AIMessage)
+    assert isinstance(response.content, str)
+    # Assert that reasoning_content and thinking_blocks are in additional_kwargs
+    assert "reasoning_content" in response.additional_kwargs

libs/partners/openai/tests/unit_tests/chat_models/test_base.py

@@ -281,6 +281,16 @@ def test__convert_dict_to_message_tool_call() -> None:
     assert reverted_message_dict == message
 
 
+def test__convert_dict_to_message_reasoning_blocks() -> None:
+    message = {"role": "assistant", "content": "foo", "reasoning_content": "bar"}
+    result = _convert_dict_to_message(message)
+    expected_output = AIMessage(
+        content="foo",
+        additional_kwargs={"reasoning_content": "bar"},
+    )
+    assert result == expected_output
+
+
 class MockAsyncContextManager:
     def __init__(self, chunk_list: list) -> None:
         self.current_chunk = 0