fix(sandbox): reap killed subprocess on UnixLocal exec timeout#3208
fix(sandbox): reap killed subprocess on UnixLocal exec timeout#3208adityasingh2400 wants to merge 3 commits into
Conversation
|
Fixed CI - typecheck and 3.10 test pass locally now. Thanks for the heads up. |
|
This PR is stale because it has been open for 10 days with no activity. |
When an exec timeout fires, the killed subprocess is never awaited, leaving the asyncio SubprocessTransport (pipes, child-watcher entry) attached to a returncode-still-None Process object until GC eventually collects it. The unconditional os.killpg also signals a pid that may already have exited and been reused, hitting an unrelated process group. Skip the kill when proc.returncode is already set, narrow the swallowed exception to ProcessLookupError, and await proc.wait() (bounded) so the transport tears down inside the failing call.
- raise asyncio.TimeoutError (not bare TimeoutError) in the wait_for fixture: on Python 3.10 these are distinct classes, so the source's except asyncio.TimeoutError clause was missing the bare one and the exception was upgraded to ExecTransportError. - import asyncio/os directly instead of reaching through unix_local_module.asyncio / .os, which mypy strict export rejects with attr-defined errors. - type captured[] as list[asyncio.subprocess.Process] so proc.returncode is statically known and the type: ignore comments are no longer needed.
69bc8f7 to
d01e5d1
Compare
|
CI is green now. I rebased onto current main and re-ran the checks locally: ruff format, ruff lint, and mypy all pass, and the 10 tests in tests/sandbox/test_unix_local.py pass. The earlier failures were the py3.10 typing compat and a strict mypy issue in the test, both fixed. The fix itself reaps the killed subprocess on a UnixLocal exec timeout so it does not leave a zombie. Ready for review when you have time. |
|
Re-verified today: the branch merges cleanly against main and the CI errors that prompted the change-request are fixed. The earlier failures were a py3.10 typing-compat issue and a strict mypy issue in the test, both resolved. ruff format, ruff lint, and mypy are clean locally, and the 10 tests in tests/sandbox/test_unix_local.py pass. The fix in _exec_internal in src/agents/sandbox/sandboxes/unix_local.py does two things on an exec timeout: it only sends os.killpg when proc.returncode is None and proc.pid is not None, suppressing ProcessLookupError, so it cannot signal an already-exited and possibly reused pid, and it then awaits proc.wait with a bounded timeout to reap the killed subprocess so asyncio releases the transport instead of leaving a delayed-GC warning. Ready for another look when you have time. |
2 participants
Summary
UnixLocalSandboxSession._exec_internalhandlesasyncio.TimeoutErrorby SIGKILL-ing the subprocess group, but never awaits the killedasyncio.subprocess.Process. Two issues fall out of that:proc.wait()is never awaited,proc.returncodestaysNoneand the underlyingSubprocessTransport(pipes + child-watcher entry) hangs around until GC eventually finds it. Repeated timeouts queue up live transports for the lifetime of the event loop.os.killpg(proc.pid, SIGKILL)runs unconditionally. If the process happened to exit betweenwait_forcancellation and the kill, the pid may already have been reused, and we'd sendSIGKILLto an unrelated process group. Theexcept Exception: passmasks this.The fix mirrors what
_terminate_pty_entryalready does for PTY entries (if process.returncode is None ... ProcessLookupError suppressed):proc.returncode is None(andproc.pid is not None).ProcessLookupErrorinstead ofException.await asyncio.wait_for(proc.wait(), timeout=5.0)so asyncio actually closes the transport beforeExecTimeoutErroris raised.Two regression tests cover both behaviors:
test_timeout_reaps_subprocess_so_returncode_is_setcaptures the asyncio Process used by_exec_internaland assertsreturncodeis populated after the timeout fires (fails on main:returncode is None).test_timeout_skips_killpg_for_already_exited_pidsimulates the "process exited just before the timeout handler" race and asserts nokillpgis issued (fails on main: a stalekillpgis recorded).Test plan
pytest tests/sandbox/test_unix_local.py -v(all 10 pass)pytest tests/sandbox/(765 passed, 19 skipped)ruff check+ruff format --checkpyright src/agents/sandbox/sandboxes/unix_local.py tests/sandbox/test_unix_local.pymain(bug repros), pass with the fix