Merge 11651dad7a into c9a5777739

litellm/responses/litellm_completion_transformation/streaming_iterator.py

@@ -1129,7 +1129,7 @@ class LiteLLMCompletionStreamingIterator(ResponsesAPIStreamingIterator):
                         "id": self._cached_reasoning_item_id,
                         "type": "reasoning",
                         "status": "in_progress",
-                        "summary": None,
+                        "summary": [],
                     }
                 ),
             )

litellm/responses/litellm_completion_transformation/transformation.py

@@ -3436,24 +3436,29 @@ class LiteLLMCompletionResponsesConfig:
                 encrypted_content = LiteLLMCompletionResponsesConfig._encode_thinking_blocks(message)
                 if reasoning_content or encrypted_content:
                     # Only check the first choice for reasoning content
+                    # summary and encrypted_content are reasoning-only extras the shared model does not declare
                     return [
-                        GenericResponseOutputItem(
-                            type="reasoning",
-                            id=mint_reasoning_item_id(),
-                            status=LiteLLMCompletionResponsesConfig._map_chat_completion_finish_reason_to_responses_status(
-                                choice.finish_reason
-                            ),
-                            role="assistant",
-                            content=[
-                                OutputText(
-                                    type="output_text",
-                                    text=text,
-                                    annotations=[],
-                                )
-                                for text in (reasoning_content,)
-                                if text
-                            ],
-                            encrypted_content=encrypted_content,
+                        GenericResponseOutputItem.model_validate(
+                            {
+                                "type": "reasoning",
+                                "id": mint_reasoning_item_id(),
+                                "status": LiteLLMCompletionResponsesConfig._map_chat_completion_finish_reason_to_responses_status(
+                                    choice.finish_reason
+                                ),
+                                "role": "assistant",
+                                "content": [
+                                    OutputText(
+                                        type="output_text",
+                                        text=text,
+                                        annotations=[],
+                                    )
+                                    for text in (reasoning_content,)
+                                    if text
+                                ],
+                                # Required by the Responses schema. The reasoning text rides in content, so it stays empty
+                                "summary": [],
+                                "encrypted_content": encrypted_content,
+                            }
                         )
                     ]
         return []

tests/unit/responses/litellm_completion_transformation/test_litellm_completion_responses.py

@@ -503,6 +503,39 @@ class TestLiteLLMCompletionResponsesConfig:
         assert reasoning_items[0].content[0].text == "counting the primes"
         assert "sig" in reasoning_items[0].encrypted_content
 
+    @pytest.mark.parametrize(
+        "message",
+        [
+            pytest.param(
+                Message(content="No, 1001 = 7 x 11 x 13", role="assistant", reasoning_content="Trying small primes"),
+                id="gemini-thought-text",
+            ),
+            pytest.param(
+                Message(
+                    content="No",
+                    role="assistant",
+                    reasoning_content="",
+                    thinking_blocks=[{"type": "thinking", "thinking": "", "signature": "sig"}],
+                ),
+                id="signature-only-thinking",
+            ),
+        ],
+    )
+    def test_reasoning_item_always_carries_a_summary_list(self, message: Message):
+        """`summary` is required on a Responses reasoning item, and clients that read it
+        (e.g. OpenAI Agents SDK tracing) fail on the missing field"""
+        response: Final = LiteLLMCompletionResponsesConfig.transform_chat_completion_response_to_responses_api_response(
+            request_input="Is 1001 prime?",
+            responses_api_request={},
+            chat_completion_response=ModelResponse(
+                model="gemini-3.8-flash", choices=[Choices(index=0, finish_reason="stop", message=message)]
+            ),
+        )
+
+        reasoning_items: Final = [item for item in response.output if item.type == "reasoning"]
+        assert len(reasoning_items) == 1
+        assert reasoning_items[0].summary == []
+
     def test_transform_chat_completion_response_status_with_stop(self):
         """
         Test that transforming a chat completion response with 'stop' finish_reason

tests/unit/responses/litellm_completion_transformation/test_streaming_iterator_transformation.py

@@ -1173,6 +1173,30 @@ async def test_signature_only_thinking_streams_a_replayable_reasoning_item(sync_
     assert json.loads(reasoning_items[0].encrypted_content)[0]["signature"] == "sig_only"
 
 
+@pytest.mark.parametrize(
+    "stream_chunks",
+    [_reasoning_then_text_chunks, _signed_thinking_and_text_in_one_chunk],
+    ids=["gemini-thought-text", "signature-only-thinking"],
+)
+@pytest.mark.parametrize("sync_mode", [True, False])
+@pytest.mark.asyncio
+async def test_streamed_reasoning_item_carries_a_summary_list_from_added_to_completed(
+    stream_chunks: StreamChunks, sync_mode: bool
+):
+    events: Final = await _collect_events(_build_iterator(stream_chunks()), sync_mode)
+
+    added: Final = next(
+        event.item
+        for event in events
+        if getattr(event, "type", None) == ResponsesAPIStreamEvents.OUTPUT_ITEM_ADDED and event.item.type == "reasoning"
+    )
+    completed: Final = next(
+        event for event in events if getattr(event, "type", None) == ResponsesAPIStreamEvents.RESPONSE_COMPLETED
+    )
+    assert added.summary == []
+    assert [item.summary for item in completed.response.output if item.type == "reasoning"] == [[]]
+
+
 @REASONING_THEN_TEXT_STREAMS
 @pytest.mark.parametrize("sync_mode", [True, False])
 @pytest.mark.asyncio