Skip to content

The agent engineering platform.

fix(openai): keep Responses stream items on one block when output_index drifts (#41121)

gethin-langchain4bbbc5a3b0ad
diff --git a/libs/partners/openai/langchain_openai/chat_models/base.py b/libs/partners/openai/langchain_openai/chat_models/base.py
index 739d74da87..bad056a75a 100644
--- a/libs/partners/openai/langchain_openai/chat_models/base.py
+++ b/libs/partners/openai/langchain_openai/chat_models/base.py
@@ -1727,6 +1727,7 @@ class BaseChatOpenAI(BaseChatModel):
current_output_index = -1
current_sub_index = -1
has_reasoning = False
+ item_output_indices: dict[str, int] = {}
for chunk in response:
metadata = headers if is_first_chunk else {}
(
@@ -1743,6 +1744,7 @@ class BaseChatOpenAI(BaseChatModel):
metadata=metadata,
has_reasoning=has_reasoning,
output_version=self.output_version,
+ item_output_indices=item_output_indices,
)
if generation_chunk:
if is_first_chunk and base_generation_info:
@@ -1797,6 +1799,7 @@ class BaseChatOpenAI(BaseChatModel):
current_output_index = -1
current_sub_index = -1
has_reasoning = False
+ item_output_indices: dict[str, int] = {}
async for chunk in _astream_with_chunk_timeout(
response,
self.stream_chunk_timeout,
@@ -1817,6 +1820,7 @@ class BaseChatOpenAI(BaseChatModel):
metadata=metadata,
has_reasoning=has_reasoning,
output_version=self.output_version,
+ item_output_indices=item_output_indices,
)
if generation_chunk:
if is_first_chunk and base_generation_info:
@@ -5372,7 +5376,21 @@ def _convert_responses_chunk_to_generation_chunk(
metadata: dict | None = None,
has_reasoning: bool = False,
output_version: str | None = None,
+ item_output_indices: dict[str, int] | None = None, # item id -> `output_index`
) -> tuple[int, int, int, ChatGenerationChunk | None]:
+ # After several built-in tool calls the API can report a later `output_index` on
+ # an item's `done` events than on its `added` event. Pin each item to the first
+ # `output_index` seen so its events keep landing on the same content block.
+ item_id = getattr(getattr(chunk, "item", None), "id", None) or getattr(
+ chunk, "item_id", None
+ )
+ if (
+ item_output_indices is not None
+ and item_id is not None
+ and (stream_output_index := getattr(chunk, "output_index", None)) is not None
+ ):
+ item_output_indices.setdefault(item_id, stream_output_index)
+
def _advance(output_idx: int, sub_idx: int | None = None) -> None:
"""Advance indexes tracked during streaming.
@@ -5408,6 +5426,8 @@ def _convert_responses_chunk_to_generation_chunk(
the current index accordingly.
"""
nonlocal current_index, current_output_index, current_sub_index
+ if item_output_indices is not None and item_id in item_output_indices:
+ output_idx = item_output_indices[item_id]
if sub_idx is None:
if current_output_index != output_idx:
current_index += 1
diff --git a/libs/partners/openai/tests/unit_tests/chat_models/test_responses_stream.py b/libs/partners/openai/tests/unit_tests/chat_models/test_responses_stream.py
index 49ba658ed0..391954e91d 100644
--- a/libs/partners/openai/tests/unit_tests/chat_models/test_responses_stream.py
+++ b/libs/partners/openai/tests/unit_tests/chat_models/test_responses_stream.py
@@ -845,6 +845,105 @@ def test_responses_reasoning_done_without_encrypted_content_emits_no_chunk() ->
)
+_OUTPUT_INDEX_DRIFT_TEXT = '{"news_items": []}'
+
+# `output_index` on the `done` events is one higher than on the `added` events.
+_output_index_drift_stream = [
+ ResponseOutputItemAddedEvent(
+ item=ResponseReasoningItem(
+ id="rs_123", summary=[], type="reasoning", status=None
+ ),
+ output_index=0,
+ sequence_number=1,
+ type="response.output_item.added",
+ ),
+ ResponseOutputItemDoneEvent(
+ item=ResponseReasoningItem(
+ id="rs_123",
+ summary=[],
+ type="reasoning",
+ encrypted_content="encrypted-content",
+ status=None,
+ ),
+ output_index=1,
+ sequence_number=2,
+ type="response.output_item.done",
+ ),
+ ResponseOutputItemAddedEvent(
+ item=ResponseOutputMessage(
+ id="msg_123",
+ content=[],
+ role="assistant",
+ status="in_progress",
+ type="message",
+ ),
+ output_index=1,
+ sequence_number=3,
+ type="response.output_item.added",
+ ),
+ ResponseTextDeltaEvent(
+ content_index=0,
+ delta=_OUTPUT_INDEX_DRIFT_TEXT,
+ item_id="msg_123",
+ output_index=1,
+ sequence_number=4,
+ logprobs=[],
+ type="response.output_text.delta",
+ ),
+ ResponseTextDoneEvent(
+ content_index=0,
+ item_id="msg_123",
+ output_index=2,
+ sequence_number=5,
+ text=_OUTPUT_INDEX_DRIFT_TEXT,
+ logprobs=[],
+ type="response.output_text.done",
+ ),
+]
+
+
+def test_responses_stream_output_index_drift() -> None:
+ llm = ChatOpenAI(model=MODEL, use_responses_api=True, output_version="responses/v1")
+ mock_client = MagicMock()
+
+ def mock_create(*args: Any, **kwargs: Any) -> MockSyncContextManager:
+ return MockSyncContextManager(_output_index_drift_stream)
+
+ mock_client.responses.create = mock_create
+
+ full: BaseMessageChunk | None = None
+ with patch.object(llm, "root_client", mock_client):
+ for chunk in llm.stream("test"):
+ full = chunk if full is None else full + chunk
+ assert isinstance(full, AIMessageChunk)
+ assert full.text == _OUTPUT_INDEX_DRIFT_TEXT
+ assert [(block["type"], block["index"]) for block in full.content] == [ # type: ignore[index]
+ ("reasoning", 0),
+ ("text", 1),
+ ]
+
+
+async def test_responses_astream_output_index_drift() -> None:
+ llm = ChatOpenAI(model=MODEL, use_responses_api=True, output_version="responses/v1")
+ mock_client = MagicMock()
+
+ async def mock_create(*args: Any, **kwargs: Any) -> MockAsyncContextManager:
+ return MockAsyncContextManager(_output_index_drift_stream)
+
+ mock_client.responses.create = mock_create
+
+ full: BaseMessageChunk | None = None
+ with patch.object(llm, "root_async_client", mock_client):
+ async for chunk in llm.astream("test"):
+ full = chunk if full is None else full + chunk
+ assert isinstance(full, AIMessageChunk)
+ assert full.text == _OUTPUT_INDEX_DRIFT_TEXT
+ assert [(block["type"], block["index"]) for block in full.content] == [ # type: ignore[index]
+ ("reasoning", 0),
+ ("text", 1),
+ ]
+
+
def test_responses_stream_events_v3_emits_reasoning_lifecycle() -> None:
"""v3 streaming emits `content-block-finish` events for reasoning blocks.