Fix stdio_client BrokenResourceError race condition during shutdown

[weiguangli-io]

Summary

Fixes#1960. Supersedes #2219 (which was closed).

During stdio_client shutdown, the finally block closes read_stream (the receiving end of a zero-buffer memory channel) while stdout_reader may still be blocked on send(). When the receiving end closes underneath a pending send(), anyio raises BrokenResourceError — not ClosedResourceError. The existing except clause only caught ClosedResourceError, so BrokenResourceError escaped into an ExceptionGroup.

The fix catches BrokenResourceError alongside ClosedResourceError in both stdout_reader and stdin_writer. This is the minimal, correct change: ClosedResourceError means "you called send on a handle you already closed", while BrokenResourceError means "the other end was closed" — both are expected during shutdown and should be handled identically.

Test plan

• Added regression test test_stdio_client_no_broken_resource_error_on_shutdown that reproduces the exact scenario: server sends a JSON-RPC message, client exits without consuming the read stream
• All 12 existing tests/client/test_stdio.py tests pass

[Bortlesboat]

ACK. Clean minimal fix — BrokenResourceError is the correct counterpart to catch when the receiving end closes under a pending send().

[Bortlesboat]

Noting for future readers: the SSE client (sse.py) avoids this class of race by calling tg.cancel_scope.cancel() in its finally block, which cancels tasks before closing their streams. The stdio client takes a different approach — letting stream closure propagate to tasks and catching the resulting exceptions. Both are valid; this PR correctly handles the full set of exceptions that anyio can raise in this scenario.

[4444J99]

This would also resolve the semaphore leak reported in openai/openai-agents-python#618 — the SDK currently works around it by pre-closing the write stream before exit stack teardown (openai/openai-agents-python#2802), but that reaches into BaseSession._write_stream which is a private attribute. A proper fix here would let the SDK drop that workaround.

Happy to help test if useful.

[weiguangli-io]

Friendly ping — CI is green. Would appreciate a review when you get a chance. Thanks!

[maxisbey]

Closing as a duplicate of #2450, which is already in review for the same fix. Your root-cause analysis on #1960 was a big help in getting that PR right — thanks. If #2450 stalls we can reopen this.

_{AI Disclaimer}