From 964f8aa309a3dcd26a8e8e554577010294897846 Mon Sep 17 00:00:00 2001 From: ryan-crabbe-berri Date: Wed, 28 Oct 2026 10:33:58 -0700 Subject: [PATCH] fix(responses): open the reasoning summary part only once reasoning text streams A reasoning item with no summary text, such as signature-only thinking, streamed one empty summary part and closed with it while response.completed carried an empty summary list. The part is now announced right before the first reasoning delta, the done events are skipped when there is no text, and the item closes with the same empty summary the final response carries, matching OpenAI --- .../streaming_iterator.py | 45 ++++++++++--------- .../test_litellm_completion_responses.py | 11 ++--- .../test_streaming_iterator_transformation.py | 14 +++++- 3 files changed, 39 insertions(+), 31 deletions(-) diff --git a/litellm/responses/litellm_completion_transformation/streaming_iterator.py b/litellm/responses/litellm_completion_transformation/streaming_iterator.py index b241107d43d..83e7bfc5e60 100644 --- a/litellm/responses/litellm_completion_transformation/streaming_iterator.py +++ b/litellm/responses/litellm_completion_transformation/streaming_iterator.py @@ -177,6 +177,7 @@ class LiteLLMCompletionStreamingIterator(ResponsesAPIStreamingIterator): self._cached_reasoning_item_id: str | None = None self._sent_reasoning_summary_text_done_event: bool = False self._sent_reasoning_summary_part_done_event: bool = False + self._sent_reasoning_summary_part_added_event: bool = False self._reasoning_summary_text: str = "" # -- GENERIC RESPONSE-EVENTS PENDING QUEUE as required by fix -- self._pending_response_events: list[BaseLiteLLMOpenAIResponseObject] = [] @@ -1035,12 +1036,7 @@ class LiteLLMCompletionStreamingIterator(ResponsesAPIStreamingIterator): **{ "id": reasoning_item_id, "type": "reasoning", - "summary": [ - { - "type": "summary_text", - "text": reasoning_content, - } - ], + "summary": [{"type": "summary_text", "text": reasoning_content}] if reasoning_content else [], } ), ) @@ -1157,9 +1153,7 @@ class LiteLLMCompletionStreamingIterator(ResponsesAPIStreamingIterator): ), ) event.__dict__["sequence_number"] = self._sequence_number - self._pending_response_events.extend( - [event, self.create_reasoning_summary_part_added_event(self._cached_reasoning_item_id)] - ) + self._pending_response_events.append(event) return # Tool-first @@ -1186,25 +1180,27 @@ class LiteLLMCompletionStreamingIterator(ResponsesAPIStreamingIterator): ) reasoning_item_id: Final = self._cached_reasoning_item_id self._queue_compaction_item_done() - self._sequence_number += 1 - text_done_event: Final = self.create_reasoning_summary_text_done_event( - reasoning_item_id=reasoning_item_id, - reasoning_content=reasoning_content, - sequence_number=self._sequence_number, - ) - self._sequence_number += 1 - part_done_event: Final = self.create_reasoning_summary_part_done_event( - reasoning_item_id=reasoning_item_id, - reasoning_content=reasoning_content, - sequence_number=self._sequence_number, - ) + if reasoning_content: + self._sequence_number += 1 + text_done_event: Final = self.create_reasoning_summary_text_done_event( + reasoning_item_id=reasoning_item_id, + reasoning_content=reasoning_content, + sequence_number=self._sequence_number, + ) + self._sequence_number += 1 + part_done_event: Final = self.create_reasoning_summary_part_done_event( + reasoning_item_id=reasoning_item_id, + reasoning_content=reasoning_content, + sequence_number=self._sequence_number, + ) + self._pending_response_events.extend([text_done_event, part_done_event]) self._sequence_number += 1 reasoning_output_item_done_event: Final = self.create_reasoning_output_item_done_event( reasoning_item_id=reasoning_item_id, reasoning_content=reasoning_content, sequence_number=self._sequence_number, ) - self._pending_response_events.extend([text_done_event, part_done_event, reasoning_output_item_done_event]) + self._pending_response_events.append(reasoning_output_item_done_event) self._reasoning_done_emitted = True self._reasoning_active = False @@ -1377,6 +1373,11 @@ class LiteLLMCompletionStreamingIterator(ResponsesAPIStreamingIterator): if self._cached_reasoning_item_id is None: self._cached_reasoning_item_id = mint_reasoning_item_id() + if not self._sent_reasoning_summary_part_added_event: + self._sent_reasoning_summary_part_added_event = True + self._pending_response_events.append( + self.create_reasoning_summary_part_added_event(self._cached_reasoning_item_id) + ) return ReasoningSummaryTextDeltaEvent( type=ResponsesAPIStreamEvents.REASONING_SUMMARY_TEXT_DELTA, diff --git a/tests/unit/responses/litellm_completion_transformation/test_litellm_completion_responses.py b/tests/unit/responses/litellm_completion_transformation/test_litellm_completion_responses.py index 8a94d475918..3fd4be6f722 100644 --- a/tests/unit/responses/litellm_completion_transformation/test_litellm_completion_responses.py +++ b/tests/unit/responses/litellm_completion_transformation/test_litellm_completion_responses.py @@ -4834,9 +4834,9 @@ class TestEnsureOutputItemContentPartAdded: assert completed_event.type == ResponsesAPIStreamEvents.RESPONSE_COMPLETED assert completed_event.response.status == "completed" - def test_reasoning_item_opens_a_summary_part_instead_of_a_content_part(self): - """Reasoning items announce their summary part, never a content_part.added event.""" - from litellm.types.llms.openai import OutputItemAddedEvent, ResponsePartAddedEvent + 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 iterator = self._make_iterator() chunk = self._make_reasoning_chunk() @@ -4844,11 +4844,8 @@ class TestEnsureOutputItemContentPartAdded: iterator._ensure_output_item_for_chunk(chunk) events = iterator._pending_response_events - assert len(events) == 2 + assert len(events) == 1 assert isinstance(events[0], OutputItemAddedEvent) - assert isinstance(events[1], ResponsePartAddedEvent) - assert events[1].item_id == events[0].item.id - assert events[1].output_index == events[0].output_index assert iterator.sent_content_part_added_event is False def test_only_emits_once(self): diff --git a/tests/unit/responses/litellm_completion_transformation/test_streaming_iterator_transformation.py b/tests/unit/responses/litellm_completion_transformation/test_streaming_iterator_transformation.py index 8c5b620f51c..65bcafef683 100644 --- a/tests/unit/responses/litellm_completion_transformation/test_streaming_iterator_transformation.py +++ b/tests/unit/responses/litellm_completion_transformation/test_streaming_iterator_transformation.py @@ -1245,6 +1245,10 @@ def _split_reasoning_then_text_chunks() -> tuple[ModelResponseStream, ...]: return (_reasoning_chunk("let me "), _reasoning_chunk("think"), _chunk("Hello!", finish_reason="stop")) +def _signature_only_thinking_then_text_chunks() -> tuple[ModelResponseStream, ...]: + return (_signature_only_thinking_chunk("sig_only"), _chunk("4", finish_reason="stop")) + + def _streamed_reasoning_text(chunks: Sequence[ModelResponseStream]) -> str: return "".join(getattr(chunk.choices[0].delta, "reasoning_content", None) or "" for chunk in chunks) @@ -1277,11 +1281,17 @@ def _reasoning_item(items: Sequence[Mapping[str, object]]) -> Mapping[str, objec @pytest.mark.parametrize( "stream_chunks", - [_split_reasoning_then_text_chunks, _reasoning_and_text_in_one_chunk, _signed_thinking_and_text_in_one_chunk], + [ + _split_reasoning_then_text_chunks, + _reasoning_and_text_in_one_chunk, + _signed_thinking_and_text_in_one_chunk, + _signature_only_thinking_then_text_chunks, + ], ids=[ "split_reasoning_then_text_chunks", "reasoning_and_text_in_one_chunk", "signed_thinking_and_text_in_one_chunk", + "signature_only_thinking_then_text_chunks", ], ) @pytest.mark.parametrize("sync_mode", [True, False]) @@ -1298,7 +1308,7 @@ async def test_sdk_stream_readers_can_place_every_reasoning_summary_delta(stream completed_summary: Final = _reasoning_item(completed_payload["response"]["output"])["summary"] assert "".join(part["text"] for part in accumulated) == _streamed_reasoning_text(chunks) assert _reasoning_item(done_items)["summary"] == accumulated - assert "".join(part["text"] for part in completed_summary) == _streamed_reasoning_text(chunks) + assert completed_summary == accumulated TypeAdapter(list[OpenAIReasoningSummaryPart]).validate_python([*accumulated, *completed_summary])