fix(sdk): keep the completion a Responses stream yielded

[alanhuangyoo]

HUMAN:

Ran the streaming path and confirmed the completion the Responses stream yielded is the one kept, on both the sync and async call sites.

---

AGENT:

Why

responses() and aresponses() capture the ResponseCompletedEvent while draining the stream, then read it back off the wrapper:

completed_response = getattr(ret, "completed_response", completed_response)

getattr's default only applies when the attribute is absent. A wrapper that exposes completed_response and leaves it None therefore overwrites the event iteration just found, and the call fails with

LLMNoResponseError: Responses stream finished without a completed response

even though the stream completed normally. Both the sync and async paths carried the same line.

Summary

• Take the wrapper's value only when it has one. A wrapper that sets the attribute late can still supply the completion; a stale None no longer clobbers what iteration already found.
• Two tests, sync and async, driving a wrapper that yields the completion and also exposes completed_response = None.

Issue Number

Closes #4769

How to Test

uv run pytest tests/sdk/llm/ -q
# 981 passed

uv run ruff check && uv run ruff format --check
# clean

test_responses_streaming_keeps_the_yielded_completion and test_aresponses_streaming_keeps_the_yielded_completion both fail on main with exactly the reported error:

openhands.sdk.llm.exceptions.types.LLMNoResponseError: Responses stream finished without a completed response

and pass here, with response.raw_response is completed_response.

The test double is the smallest thing that reproduces it:

class _StaleCompletedResponseStream:
    def __init__(self, events):
        self._events = events
        self.completed_response = None      # present, and still None after iteration

    def __iter__(self):
        return iter(self._events)

Video/Screenshots

Not applicable — no GUI surface; the test output above is the evidence.

Design Doc

Not added; the change is one conditional, applied to the two paths that shared the line.

[all-hands-bot]

👋 This PR needs a couple of things fixed before OpenHands can review it:

• the PR description's HUMAN: section needs at least 20 characters describing what you tested, not just the template placeholder

Push an update once this is addressed and this check re-runs automatically.

This is an automated check - no AI was used to generate this comment.

[all-hands-bot]

🚦 CI is currently failing on this PR's latest commit.

Please fix the failing checks before OpenHands reviews it - this is re-checked automatically once you push a new commit. (A maintainer can also request @all-hands-bot as a reviewer to have it reviewed regardless of CI status.)

This is an automated check - no AI was used to generate this comment.

[alanhuangyoo]

The only red here is Validate PR description, and it is not something I can clear from the PR side:

Linked issue(s) (#4769) carry neither `ready-for-dev` nor a pre-rollout creation date.

#4769 was filed by @yifanxiong272 on 2026-08-30, after the ready-for-dev rollout date the checker grandfathers against (READY_FOR_DEV_ROLLOUT_ISO = 2026-08-13), and it carries bug, llm, priority:medium, sdk but not ready-for-dev. Dropping the link is not a way out either — validate_linked_issue_ready errors on a PR with no linked issue at all.

So this needs a maintainer to add ready-for-dev to #4769, after which the check re-runs on the next event. Everything else on this PR is green.

Content-wise nothing has changed since the last push: the regression test fails on main with the stale wrapper state overwriting the yielded completion event, and passes here.

[alanhuangyoo]

@neubig #4976 names this PR as the fix for the yielded-versus-wrapper completion bug, so here is where it stands.

I re-checked it against main at c370074, which includes #4968's llm.py changes:

• The branch still merges into main with no conflicts.
• With main's llm.py, both regression tests fail with LLMNoResponseError: Responses stream finished without a completed response.
• With this change, the whole file passes (20/20).

So the bug is still on main, and the fix still applies.

The only red check is Validate PR description. It fails because #4769 has no ready-for-dev label, and I can't add that from here. Could you add it to #4769, or tell me who triages it?