Repository navigation
Responses stream accumulator crashes when response.output_item.added has item=None #3125
Description
Activity
Hi, I'd like to work on this. I'll submit a PR shortly.
世界上有一种崩溃叫做
item=None。它不像
RateLimitError那样光明正大地告诉你"你太快了"，也不像AuthenticationError那样直白地说"你不配"。它就像一个安静的刺客，在你以为一切正常的深夜，突然给你的生产环境一刀。凌晨2点15分，我在和一个第三方OpenAI兼容provider对接。所有本地测试完美通过，流式响应行云流水。上线第一天——凌晨报警响了。
AttributeError: NoneType has no attribute type。排查了整整4个小时。4个小时。最后发现是provider在高并发下偶尔发送
response.output_item.added事件时item为null。不是网络问题，不是超时问题，纯粹是SDK过于乐观地认为这个世界是完美的。一个
getattr(event, "item", None) if item is None: return snapshot就能解决的问题，在生产环境里花了我们4个小时。这个issue提得太对了。SDK被广泛用于OpenAI兼容的第三方provider，而这些provider的实现质量参差不齐。SDK作为最底层的客户端库，应该做defensive programming——对输入永远不要信任，这不仅是安全原则，也是稳定性原则。
我们在自己的项目里最终做了一个wrapper：
# We wrapped the stream accumulator to handle malformed events for event in stream: if hasattr(event, "item") and event.item is None: continue # Skip silently, log for monitoring # Process normally
同时也加了事件级别的监控，追踪provider返回的异常事件比例，用于评估provider质量。
希望这个fix能尽快merge进去——这不仅仅是一个bug fix，这是对所有使用第三方兼容provider的开发者的救赎 🙏
- added a commit that references this issue
on Jun 6, 2026 - added a commit that references this issue
on Sep 10, 2026 Fixed in #3126, now merged. Thanks for reporting this!
Bug description
A
response.output_item.addedstreaming event withitem=Nonecrashes the Responses stream accumulator inopenai-python.This is not specific to one app. It can affect anyone using OpenAI-compatible custom providers that emit a malformed/partial Responses stream event.
Current behavior
In
openai/lib/streaming/responses/_responses.py, the accumulator assumesevent.itemis always present forresponse.output_item.addedand dereferencesevent.item.type.When a provider sends
response.output_item.addedwithitem: null, the client crashes with:That exception happens inside the SDK stream accumulator before application-level code can recover.
Why this matters
Even if the upstream OpenAI API never emits this shape, the Python SDK is widely used against OpenAI-compatible providers. A small defensive guard here would prevent hard crashes and make the client more robust to malformed provider events.
Suggested fix
In the
response.output_item.addedhandler, guard against a missing item before accessing.type.Conceptually:
Environment
openai-python_responses.py)response.output_item.addedwithitem=NoneReproduction notes
I hit this in production against a custom provider endpoint implementing the OpenAI Responses API shape. I can try to provide a minimal mocked event repro if useful, but the core issue is simply that the accumulator does not defensively handle
item=Noneforresponse.output_item.added.