Thanks — that log is helpful. I’m responding to the ValueError: "process object is closed" that occurred while the agent was attempting to terminate a job process. Summary of what happened • The agent tried to close a supervised subprocess during shutdown/termination. • While awaiting

proc.aclose()

a timeout/cancellation occurred, then the code attempted

self._proc.kill()

. • Python raised ValueError("process object is closed") because the multiprocessing Process had already been closed (exited or otherwise cleaned up) before the kill was attempted. What this means for you • This is a process lifecycle race: the agent attempted to force-kill a process that had already been closed, which produces noisy error logs but does not necessarily indicate the job failed in a way that affects other jobs. • You may still want to eliminate the exception to reduce log noise and ensure clean shutdown behavior. Recommended next steps (actionable) 1. Add a safe-check before calling kill ◦ Update the kill/send-kill logic to skip calling

kill()

when the process is already closed or not alive. Example:

if self._proc is None or not self._proc.is_alive():
       return
   self._proc.kill()

• This prevents ValueError when the process has already terminated. 1. Catch and ignore ValueError at shutdown ◦ Wrap the termination sequence in a try/except so a closed-process error is handled gracefully:

try:
       await self._send_kill_signal()
   except ValueError:
       # process already closed, nothing to do
       pass

1. Make

aclose()

tolerant to already-closed processes ◦ When implementing

aclose()

, check process state and avoid sending signals if closed; also catch

asyncio.TimeoutError

and

ValueError

and proceed to cleanup. 1. Add debug logging around termination ◦ Log process state immediately before attempting termination so you can confirm whether the process was already dead:

logger.debug(f"Terminating job {job_id}; process alive={self._proc.is_alive()}")

1. Check shutdown sequencing and external signals ◦ Verify no external lifecycle (container/platform) signal causes the process to exit prior to agent termination logic. ◦ Confirm

drain_timeout

and graceful-shutdown settings are appropriate for your jobs so forced-kill logic is rarely needed. If you want, I can: • Point to the specific files and lines to change if you share the agent version or repo layout. • Draft a small patch for

SupervisedProc.aclose()

and the

_handle_termination

handler that implements the checks and exception handling above. • Help interpret further logs if this error recurs after changes (please include the agent version and the surrounding logs). Which would you like to do next? Sources: Job lifecycle | LiveKit Documentation | Server options | LiveKit Documentation | Self-hosted deployments | LiveKit Documentation