fix(stream): emit queued events before max-turns exception

[OiPunk]

Summary

RunResultStreaming.stream_events() currently breaks as soon as _stored_exception is set, even if there are still queued run-item events waiting in _event_queue.

For max-turn failures this can drop the last tool_call_output_item from the stream, even though the item has already been produced and queued.

This PR changes the loop so we only break on _stored_exception when the event queue is empty.

Why

Issue: #526

This preserves event ordering and guarantees consumers receive all already-queued semantic events before the final exception is raised.

Changes

• In RunResultStreaming.stream_events(), changed:
  • from: break immediately when _stored_exception is set
  • to: break only when _stored_exception is set and _event_queue is empty
• Added regression test:
  • test_streaming_max_turns_emits_pending_tool_output_events
  • Verifies tool_call_item and tool_call_output_item are emitted before MaxTurnsExceeded

Validation

• uv run ruff check src/agents/result.py tests/test_agent_runner_streamed.py
• uv run mypy src/agents/result.py tests/test_agent_runner_streamed.py
• env -u all_proxy -u ALL_PROXY -u http_proxy -u HTTP_PROXY -u https_proxy -u HTTPS_PROXY uv run pytest tests/test_agent_runner_streamed.py -k "max_turns" -q
• env -u all_proxy -u ALL_PROXY -u http_proxy -u HTTP_PROXY -u https_proxy -u HTTPS_PROXY uv run pytest tests/test_agent_runner_streamed.py -q
• make format-check
• env -u all_proxy -u ALL_PROXY -u http_proxy -u HTTP_PROXY -u https_proxy -u HTTPS_PROXY uv run python -m trace --count --coverdir /tmp/openai_agents_526_trace --module pytest tests/test_agent_runner_streamed.py -k "max_turns" -q

Trace output confirms the modified branch in src/agents/result.py is executed.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 1744c15fd8

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Preserve immediate abort for guardrail-triggered exceptions

stream_events() now continues draining _event_queue whenever _stored_exception is set but the queue is non-empty, which changes behavior for all exception types, not just max-turns. _check_errors() sets _stored_exception for input/output guardrail failures and task exceptions, and parallel input guardrails run concurrently with model streaming, so a tripwire can be detected while response events are already queued; with this condition, those queued events are still yielded before raising, potentially leaking assistant output after a guardrail has fired. Please scope the queue-draining behavior to MaxTurnsExceeded (or another explicit allowlist) so guardrail-triggered runs still abort immediately.

Useful? React with 👍 / 👎.