'Span' object has no attribute '_context_manager_state'

How do you use Sentry?

Sentry Saas (sentry.io)

Version

2.35.0

Steps to Reproduce

Run

import asyncio

import sentry_sdk
from agents import Agent, Runner
from sentry_sdk.integrations.openai_agents import OpenAIAgentsIntegration

sentry_sdk.init(
    dsn="...",
    send_default_pii=True,
    traces_sample_rate=1.0,
    integrations=[
        OpenAIAgentsIntegration(),
    ],
)


main_agent = Agent(
    name="main_agent",
    model="gpt-5",
)


async def run_agent() -> None:
    runner = await Runner.run(
        starting_agent=main_agent,
        input="How are you?",
    )
    print(runner.final_output)


async def main() -> None:
    await asyncio.gather(*[run_agent() for _ in range(2)])


if __name__ == "__main__":
    asyncio.run(main())

Then, the agent that finishes 2nd will fail with:

{
  "error": "'Span' object has no attribute '_context_manager_state'",
  "tool_name": "<agentname>_agent"
}

Expected Result

This error not to happen
That most (if not all) parts of the Sentry SDK are wrapped in try catch blocks, so that if for some reason like now a Span cannot be closed, it doesn't interfere with the normal flow

Actual Result

The agent crashes, and instead of the result that was generated by the nested agent being available to the main agent, the main agent only sees the error message and cannot continue.

@sentrivana

@sentrivana

Played around a bit, and this seems to be quite reliable (100% so far) in repro-ing:

import asyncio

import sentry_sdk
from agents import Agent, Runner
from sentry_sdk.integrations.openai_agents import OpenAIAgentsIntegration

sentry_sdk.init(
    dsn="...",
    send_default_pii=True,
    traces_sample_rate=1.0,
    integrations=[
        OpenAIAgentsIntegration(),
    ],
)


main_agent = Agent(
    name="main_agent",
    model="gpt-5",
)


async def run_agent() -> None:
    runner = await Runner.run(
        starting_agent=main_agent,
        input="How are you?",
    )
    print(runner.final_output)


async def main() -> None:
    await asyncio.gather(*[run_agent() for _ in range(2)])


if __name__ == "__main__":
    asyncio.run(main())

From my tests:

Couldn't repro with just 1 call, needed to be 2 simultaneous (asyncio.gather) Tasks
Doesn't seem to be related to nested agents after all, was able to reproduce without any nested agent and without any function calling
Seems to fail irrespective of the model, tested both with gpt-5 and gpt-4o-mini, so seems to be unrelated to reasoning vs non-reasoning model

(also updated the original Issue so people don't have to scroll :) )

@Afstkla

Awesome, thanks for the repro @Afstkla! We're on it.

Current working theory: scope bleed between the different agent runners. We're missing proper scope isolation. Working on a fix here, should be ready by tomorrow latest.

The fix is basically akin to changing run_agent to

  async def run_agent() -> None:
+     with sentry_sdk.isolation_scope():
          runner = await Runner.run(
              starting_agent=main_agent,
              input="How are you?",
          )
          print(runner.final_output)

@sentrivana

@sentrivana thanks for the update, we've currently had to disable tracing in prod unfortunately, and ideally don't want to depend on a non-release version of the sentry sdk. Can you tell me how it aligns with your release schedule?

Unless anything happens there should be a release with this tomorrow.

@sentrivana

@sentrivana not sure if it helps btw, but I had Claude Code go on an adventure, and it proposed changing

current_span.__exit__(None, None, None)

to

current_span.finish()

In runner.py (

sentry-python/sentry_sdk/integrations/openai_agents/patches/runner.py

Line 38 in 28d0ddd

current_span.__exit__(None, None, None)

) and `invoke_agent.py (

sentry-python/sentry_sdk/integrations/openai_agents/spans/invoke_agent.py

Line 34 in 28d0ddd

current_span.__exit__(None, None, None)

).
Ran a very quick test locally and that indeed seemed to resolve the issue. Ofc I don't know too much about the sdk, so not sure if it actually resolves the underlying bug or if it just makes a lot of other stuff throw up, but sharing in case it helps :)

fix(openai-agents): Isolate agent run (#4720)

not sure if it actually resolves the underlying bug or if it just makes a lot of other stuff throw up, but sharing in case it helps :)

Thanks, but in this case looks like it's the latter. If you look at the transactions produced with the fix, they're all mixed up, so there's still scope bleed going on.

I've merged the PR, will cut a release shortly.

@Afstkla

@Afstkla Out now in 2.35.1. Let us know if the fix works for you :)

@sentrivana

@sentrivana I thought it was resolved, but it seems to not have been fully resolved unfortunately...
(unfortunately I can't re-open the issue, so if you could (or create a new one, whichever fits your process better) that'd be great)

From our traces, I can't see anything obvious unfortunately, adding a screenshot in case it helps from one of the traces that has errored. This was a nested agent running (agent.as_tool(), which calls Runner.run, nested inside a Runner.run_streamed)

So unfortunately we're again reverting the integration from our prod, and I'll again vote to at least try the calls so sth like this doesn't negatively interfere with our agents that are running...

Tagging @seynadio for vis as well.

@Afstkla

Hey @Afstkla, yesterday we fixed yet another bug that was presenting itself with the same error in b1a8b63. We'll cut a release with this today/tomorrow -- maybe with the try..except on top for good measure. I'll ping you when it's out.

Wrap span restoration in `__exit__` in `capture_internal_exceptions` (#…

Hey @Afstkla, yesterday we fixed yet another bug that was presenting itself with the same error in b1a8b63. We'll cut a release with this today/tomorrow -- maybe with the try..except on top for good measure. I'll ping you when it's out.

Thx for the update. I cannot have this same issue break production again, so I’m realistically not going to re-enable it on prod until there’s error handling around it (and I hope it’s also something you see as well, that enabling tracing in any way should never be responsible for breaking our main flow in prod…).

So please keep me posted, and sharing the above just so you know where we’re coming from :)

@Afstkla

Hey @Afstkla, of course, it goes without saying that the SDK should never break a user's app. That being said, the OpenAI agents integration is still in beta so there might be (and apparently are) things that need ironing out. We're on them with high prio -- also thanks to you raising them and following up. 🙏🏻

Re: release, v2.36.0 of the SDK is out. It includes the fix #4736 I mentioned above as well as some error silencing around span/profiler restoration (#4698, #4719). Do you happen to have a non-prod environment where you could give this a spin?

Hey @Afstkla, of course, it goes without saying that the SDK should never break a user's app. That being said, the OpenAI agents integration is still in beta so there might be (and apparently are) things that need ironing out. We're on them with high prio -- also thanks to you raising them and following up. 🙏🏻

Re: release, v2.36.0 of the SDK is out. It includes the fix #4736 I mentioned above as well as some error silencing around span/profiler restoration (#4698, #4719). Do you happen to have a non-prod environment where you could give this a spin?

Thx for the response! We'll test it out (tho not in prod) but as it seems mostly a multithreading issue, it might be quite difficult to reproduce in an env where we have not the traffic of prod.. Will keep you posted when we test & encounter anything

@sentrivana

@sentrivana haven't seen any occurrences of this happening anymore, so deployed to prod and so far so good. Thanks for the support!

This issue was fixed by the patch mentioned above. I will close this issue.

Navigation Menu

Repository navigation

'Span' object has no attribute '_context_manager_state' #4718

How do you use Sentry?

Version

Steps to Reproduce

Expected Result

Actual Result

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

'Span' object has no attribute '_context_manager_state' #4718

Description

How do you use Sentry?

Version

Steps to Reproduce

Expected Result

Actual Result

Activity

Afstkla commented on Aug 25, 2025

sentrivana commented on Aug 25, 2025

sentrivana commented on Aug 25, 2025

Afstkla commented on Aug 25, 2025

sentrivana commented on Aug 25, 2025

Afstkla commented on Aug 25, 2025

sentrivana commented on Aug 26, 2025

sentrivana commented on Aug 26, 2025

Afstkla commented on Sep 2, 2025

sentrivana commented on Sep 3, 2025

Afstkla commented on Sep 4, 2025

sentrivana commented on Sep 4, 2025

Afstkla commented on Sep 4, 2025

Afstkla commented on Sep 10, 2025

antonpirker commented on Sep 11, 2025

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions