fix(responses): harden reasoning replay edge cases

litellm/responses/litellm_completion_transformation/session_handler.py

@@ -152,9 +152,8 @@ class ResponsesSessionHandler:
             and _response_output
             and _response_output != {}
         ):
-            if (
-                _response_output.get("object") == "response"
-                or "output" in _response_output
+            if ResponsesSessionHandler._is_responses_api_output_response(
+                _response_output
             ):
                 chat_completion_message_history.extend(
                     LiteLLMCompletionResponsesConfig.transform_responses_api_output_to_chat_completion_messages(
@@ -170,6 +169,35 @@ class ResponsesSessionHandler:
                     chat_completion_message_history.append(getattr(choice, "message"))
         return chat_completion_message_history
 
+    @staticmethod
+    def _is_responses_api_output_response(response_output: dict) -> bool:
+        if response_output.get("object") == "response":
+            return True
+        if "choices" in response_output:
+            return False
+        output = response_output.get("output")
+        if not isinstance(output, list):
+            return False
+        responses_output_types = {
+            "reasoning",
+            "message",
+            "function_call",
+            "function_call_output",
+            "web_search_call",
+            "file_search_call",
+            "computer_call",
+            "image_generation_call",
+            "code_interpreter_call",
+            "mcp_call",
+            "custom_tool_call",
+        }
+        for output_item in output:
+            if not isinstance(output_item, dict):
+                continue
+            if output_item.get("type") in responses_output_types:
+                return True
+        return False
+
     @staticmethod
     async def get_proxy_server_request_from_spend_log(
         spend_log: SpendLogsPayload,

litellm/responses/litellm_completion_transformation/transformation.py

@@ -420,6 +420,15 @@ class LiteLLMCompletionResponsesConfig:
             )
             if new_role != "assistant":
                 continue
+            reasoning_content = (
+                LiteLLMCompletionResponsesConfig._get_chat_message_reasoning_content(
+                    new_msg
+                )
+            )
+            LiteLLMCompletionResponsesConfig._set_assistant_reasoning_content(
+                last_msg,
+                reasoning_content,
+            )
             for tool_call in LiteLLMCompletionResponsesConfig._get_tool_calls_list(
                 new_msg
             ):

tests/test_litellm/responses/litellm_completion_transformation/test_reasoning_content_transformation.py

@@ -325,6 +325,58 @@ class TestReasoningContentFinalResponse:
         assert tool_calls[0].get("id") == tool_call_id
 
 
+def test_merge_consecutive_function_call_messages_preserves_reasoning_content():
+    messages = [
+        {
+            "role": "assistant",
+            "content": None,
+            "tool_calls": [
+                {
+                    "id": "call_first",
+                    "type": "function",
+                    "function": {
+                        "name": "first_tool",
+                        "arguments": "{}",
+                    },
+                }
+            ],
+        }
+    ]
+    chat_completion_messages = [
+        {
+            "role": "assistant",
+            "content": None,
+            "reasoning_content": "Need a second tool before answering.",
+            "tool_calls": [
+                {
+                    "id": "call_second",
+                    "type": "function",
+                    "function": {
+                        "name": "second_tool",
+                        "arguments": "{}",
+                    },
+                }
+            ],
+        }
+    ]
+
+    merged = LiteLLMCompletionResponsesConfig._merge_consecutive_function_call_messages(
+        messages=messages,
+        chat_completion_messages=chat_completion_messages,
+        input_item={"type": "function_call", "call_id": "call_second"},
+        existing_tool_call_ids={"call_first"},
+    )
+
+    assert merged is True
+    assert messages[0].get("reasoning_content") == (
+        "Need a second tool before answering."
+    )
+    assert [tool_call.get("id") for tool_call in messages[0].get("tool_calls", [])] == [
+        "call_first",
+        "call_second",
+    ]
+
+
 def test_streaming_chunk_id_raw():
     """Test that streaming chunk IDs are raw (not encoded) to match OpenAI format"""
     chunk = ModelResponseStream(

tests/test_litellm/responses/litellm_completion_transformation/test_session_handler.py

@@ -237,6 +237,52 @@ async def test_session_history_preserves_responses_reasoning_before_tool_call():
     assert tool_calls[0].get("function", {}).get("arguments") == '{"location":"Boston"}'
 
 
+@pytest.mark.asyncio
+async def test_chat_completion_with_output_key_uses_choices_not_responses_output():
+    mock_spend_logs = [
+        {
+            "request_id": "chatcmpl-custom-output",
+            "call_type": "aresponses",
+            "session_id": "session-custom-output",
+            "proxy_server_request": {
+                "input": "Return a custom provider response.",
+                "model": "custom/provider-model",
+            },
+            "response": {
+                "id": "chatcmpl-custom-output",
+                "object": "chat.completion",
+                "model": "custom/provider-model",
+                "output": [{"custom_provider_payload": "not a Responses API item"}],
+                "choices": [
+                    {
+                        "index": 0,
+                        "message": {
+                            "role": "assistant",
+                            "content": "Provider answer from choices.",
+                        },
+                        "finish_reason": "stop",
+                    }
+                ],
+            },
+            "status": "success",
+        }
+    ]
+
+    with patch.object(
+        ResponsesSessionHandler,
+        "get_all_spend_logs_for_previous_response_id",
+        new_callable=AsyncMock,
+    ) as mock_get_spend_logs:
+        mock_get_spend_logs.return_value = mock_spend_logs
+        result = await ResponsesSessionHandler.get_chat_completion_message_history_for_previous_response_id(
+            "chatcmpl-custom-output"
+        )
+
+    messages = result["messages"]
+    assert [message.get("role") for message in messages] == ["user", "assistant"]
+    assert messages[1].get("content") == "Provider answer from choices."
+
+
 @pytest.mark.asyncio
 async def test_previous_response_tool_output_continuation_replays_reasoning_content():
     """