fix(openai): preserve streamed encrypted reasoning

[Johannes du Plessis (johannes117)]

ChatOpenAI preserves encrypted reasoning from streamed Responses API calls for stateless replay.

---

OpenAI emits an empty placeholder on response.output_item.added and the final encrypted payload on response.output_item.done. ChatOpenAI now merges the completed payload into streamed output and skips empty payloads during stateless replay.

Made by Open SWE

[open-swe]

✅ Open SWE Review: No issues found

Open SWE reviewed this PR and found no potential bugs to report.

Open in Web • View Open SWE trace

[Will James (wiggzz)]

Review note from the Codex triage assistant: the converter approach looks correct, including same-index aggregation for both output versions and filtering empty encrypted payloads from stateless replay. The current VCR failures appear to be cassette misses caused by the corrected request shape, not product failures: test_mcp_builtin_zdr and both test_mcp_builtin_zdr_v1 cases fall through to network and raise APIConnectionError. Re-recording/updating libs/partners/openai/tests/cassettes/test_mcp_builtin_zdr.yaml.gz should make CI green. One optional coverage improvement would be a single explicit streamed-output -> store=False replay regression, though the existing stream reconstruction plus replay tests already cover the behavior substantively.

[Will James (wiggzz)]

FWIW this is impacting us pretty hard - we started getting 404s from OpenAI earlier because LangChain was not preserving the encrypted reasoning content. We're working around it but look forward to pulling in this patch when it's released.

[Jordan Hayashi (jhhayashi)]

thanks for the quick patch, we also got hit by this today

dropping the reasoning item can leave its function_call replaying an fc_ id with nothing to pair it to, and OpenAI rejects that with an HTTP 400 because the function call is missing its reasoning item. we worked around it by stripping the ID off of any calls between the dropped reasoning item and the next reasoning item

[q734738781]

Confirmed this also affects DeepAgents using LangChain's Codex OAuth (ChatOpenAICodex), with the exact rs* 404 under store=False.

[ccurme (ccurme)]

Released in langchain-openai 1.5.1.

[Will James (wiggzz)]

thanks for the quick patch, we also got hit by this today

dropping the reasoning item can leave its function_call replaying an fc_ id with nothing to pair it to, and OpenAI rejects that with an HTTP 400 because the function call is missing its reasoning item. we worked around it by stripping the ID off of any calls between the dropped reasoning item and the next reasoning item

Jordan Hayashi (@jhhayashi) heads up that if you do that, you're dropping reasoning on subsequent turns, which could decrease performance (e.g. the model will lose track of what it was doing and need to re-reason)