Merge a1a73197ce into f82e4c80ac

litellm/responses/litellm_completion_transformation/streaming_iterator.py

@@ -58,6 +58,7 @@ from litellm.types.llms.openai import (
     ResponseIncompleteEvent,
     ResponseInProgressEvent,
     ResponseInputParam,
+    ResponsePartAddedEvent,
     ResponsesAPIOptionalRequestParams,
     ResponsesAPIResponse,
     ResponsesAPIStreamEvents,
@@ -1188,20 +1189,26 @@ class LiteLLMCompletionStreamingIterator(ResponsesAPIStreamingIterator):
                 self._cached_reasoning_item_id = mint_reasoning_item_id()
             self._reasoning_item_id = self._cached_reasoning_item_id
 
-            self._pending_response_events.append(
-                OutputItemAddedEvent(
-                    type=ResponsesAPIStreamEvents.OUTPUT_ITEM_ADDED,
-                    output_index=self._reasoning_index(),
-                    item=BaseLiteLLMOpenAIResponseObject(
-                        **{
-                            "id": self._cached_reasoning_item_id,
-                            "type": "reasoning",
-                            "status": "in_progress",
-                            "summary": None,
-                        }
-                    ),
-                )
+            event = OutputItemAddedEvent(
+                type=ResponsesAPIStreamEvents.OUTPUT_ITEM_ADDED,
+                output_index=self._reasoning_index(),
+                item=BaseLiteLLMOpenAIResponseObject(
+                    **{
+                        "id": self._cached_reasoning_item_id,
+                        "type": "reasoning",
+                        "status": "in_progress",
+                        "summary": [],
+                    }
+                ),
             )
+            part_added_event: Final = ResponsePartAddedEvent(
+                type=ResponsesAPIStreamEvents.RESPONSE_PART_ADDED,
+                item_id=self._cached_reasoning_item_id,
+                output_index=self._reasoning_index(),
+                sequence_number=SEQUENCE_NUMBER_STAMPED_ON_EMISSION,
+                part={"type": "summary_text", "text": ""},
+            )
+            self._pending_response_events.extend([event, part_added_event])
             return
 
         # Tool-first

litellm/responses/streaming_iterator.py

@@ -2048,6 +2048,7 @@ def build_synthetic_response_events(
                         item_id=item_id,
                         output_index=output_index,
                         summary_index=summary_index,
+                        sequence_number=sequence_number,
                         part={**summary_payload, "text": ""},
                     )
                 )

litellm/types/llms/openai.py

@@ -1832,8 +1832,9 @@ class ResponsePartAddedEvent(BaseResponsesStreamEvent):
     type: Literal[ResponsesAPIStreamEvents.RESPONSE_PART_ADDED]
     item_id: str
     output_index: int
-    summary_index: int = 0
+    sequence_number: int
     part: dict
+    summary_index: int = 0
 
 
 class ReasoningSummaryTextDeltaEvent(BaseResponsesStreamEvent):

tests/unit/responses/litellm_completion_transformation/test_litellm_completion_responses.py

@@ -4840,8 +4840,8 @@ class TestEnsureOutputItemContentPartAdded:
         assert completed_event.response.status == "completed"
 
     def test_reasoning_item_does_not_emit_content_part_added(self):
-        """Reasoning items should not get a content_part.added event."""
-        from litellm.types.llms.openai import OutputItemAddedEvent
+        """Reasoning items get a reasoning_summary_part.added event, not a content_part.added one."""
+        from litellm.types.llms.openai import OutputItemAddedEvent, ResponsePartAddedEvent
 
         iterator = self._make_iterator()
         chunk = self._make_reasoning_chunk()
@@ -4849,8 +4849,9 @@ class TestEnsureOutputItemContentPartAdded:
         iterator._ensure_output_item_for_chunk(chunk)
 
         events = iterator._pending_response_events
-        assert len(events) == 1
+        assert len(events) == 2
         assert isinstance(events[0], OutputItemAddedEvent)
+        assert isinstance(events[1], ResponsePartAddedEvent)
         assert iterator.sent_content_part_added_event is False
 
     def test_only_emits_once(self):

tests/unit/responses/litellm_completion_transformation/test_streaming_iterator_transformation.py

@@ -1517,6 +1517,35 @@ async def test_plain_text_stream_announces_exactly_one_message_item(sync_mode: b
             assert event.item_id == message_item_adds[0].item.id
 
 
+@pytest.mark.parametrize("sync_mode", [True, False])
+@pytest.mark.asyncio
+async def test_reasoning_item_opens_with_empty_summary_and_a_summary_part(sync_mode: bool):
+    iterator: Final = _build_iterator(
+        [_reasoning_chunk("let me "), _reasoning_chunk("think"), _chunk("hi", finish_reason="stop")]
+    )
+    events: Final = await _collect_events(iterator, sync_mode)
+
+    types: Final = [event.type for event in events]
+    item_added: Final = next(
+        e for e in events if e.type == ResponsesAPIStreamEvents.OUTPUT_ITEM_ADDED and e.item.type == "reasoning"
+    )
+    part_added: Final = next(e for e in events if e.type == ResponsesAPIStreamEvents.RESPONSE_PART_ADDED)
+    first_delta: Final = types.index(ResponsesAPIStreamEvents.REASONING_SUMMARY_TEXT_DELTA)
+
+    assert item_added.item.summary == []
+    assert (
+        types.index(ResponsesAPIStreamEvents.OUTPUT_ITEM_ADDED)
+        < types.index(ResponsesAPIStreamEvents.RESPONSE_PART_ADDED)
+        < first_delta
+    )
+    assert types.count(ResponsesAPIStreamEvents.RESPONSE_PART_ADDED) == 1
+    assert part_added.item_id == item_added.item.id
+    assert part_added.output_index == item_added.output_index
+    assert part_added.summary_index == 0
+    assert part_added.part == {"type": "summary_text", "text": ""}
+    assert _serialized_sequence_numbers(events) == list(range(len(events)))
+
+
 @pytest.mark.parametrize("sync_mode", [True, False])
 @pytest.mark.asyncio
 async def test_anthropic_tool_search_call_streams_as_tool_search_call_item(sync_mode: bool):