fix(responses): preserve reasoning before raw dict messages

litellm/completion_extras/litellm_responses_transformation/transformation.py

@@ -1356,6 +1356,15 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
                 elif handle_raw_dict_callback is not None:
                     choice, index = handle_raw_dict_callback(item=raw_item, index=index)
                     if choice is not None:
+                        if pending_reasoning_items:
+                            choice.message = choice.message.model_copy(
+                                update={
+                                    "reasoning_items": _as_chat_reasoning_items(pending_reasoning_items),
+                                    "reasoning_content": reasoning_content,
+                                }
+                            )
+                            pending_reasoning_items = ()  # flush
+                            reasoning_content = None
                         choices.append(choice)
             else:
                 pass  # don't fail request if item in list is not supported

tests/unit/completion_extras/litellm_responses_transformation/test_completion_extras_litellm_responses_transformation_transformation.py

@@ -3881,6 +3881,34 @@ def test_convert_response_output_merges_raw_dict_message_and_function_call() ->
     assert len(choices[0].message.tool_calls) == 1
 
 
+def test_convert_response_output_keeps_reasoning_before_raw_dict_message() -> None:
+    handler: Final = LiteLLMResponsesTransformationHandler()
+    reasoning: Final = ResponseReasoningItem(
+        id="rs_raw_before",
+        summary=[Summary(type="summary_text", text="Checking the forecast.")],
+        type="reasoning",
+        content=None,
+        encrypted_content=None,
+        status=None,
+    )
+    raw_message: Final = {
+        "type": "message",
+        "role": "assistant",
+        "content": [{"type": "output_text", "text": "Let me check.", "annotations": []}],
+    }
+
+    choices: Final = LiteLLMResponsesTransformationHandler._convert_response_output_to_choices(
+        (reasoning, raw_message),
+        handle_raw_dict_callback=handler._handle_raw_dict_response_item,
+    )
+
+    assert len(choices) == 1
+    assert choices[0].message.content == "Let me check."
+    assert choices[0].message.reasoning_content == "Checking the forecast."
+    assert choices[0].message.reasoning_items is not None
+    assert [item["id"] for item in choices[0].message.reasoning_items] == ["rs_raw_before"]
+
+
 def test_convert_tools_to_responses_format_flattens_nested_custom_tool():
     from litellm.completion_extras.litellm_responses_transformation.transformation import (
         LiteLLMResponsesTransformationHandler,