zombie processes & `autoRestart` failures

[lukeed]

We've seen some long(er)-running sessions (~20m) that fail unexpectedly and become fully unresponsive, despite having autoRestart enabled.

I was able to get one of these to happen in a Docker container, so I had Opus dig around and leave some notes. Not sure how helpful this will be to y'all but leaving it here just in case:

FWIW: I think its correct. Node's .on("exit") is generally for healthy / intentional exits

---

Root Cause: The Copilot CLI subprocess (PID 181, MainThread) crashed and became a zombie process. The SDK's autoRestart feature should have reconnected, but there are issues:

1. Zombie processes: The container has 15+ zombie processes (bash, git, MainThread) — all children of the bun process (PID 13). This indicates Node.js isn't properly reaping child processes.

2. Timeline of failure: - Last logged event: session.truncation at 18:21:20.501Z

   • Followed by assistant.turn_start for turn 14
   • Then silence — no more events logged
   • The VM restarted twice (seen in log with duplicate NODE_ENV production lines)
   • But the old zombie processes remain from before the restarts

3. Why autoRestart didn't work: The SDK's reconnect logic fires on exit event:

   this.cliProcess.on("exit", (code) => {
     if (this.options.autoRestart && this.state === "connected") {
       void this.reconnect();
     }
   });

   But if the stdio pipes get corrupted or the process crashes hard (SIGSEGV), the exit handler might not fire correctly.

4. Most likely culprit: The Copilot CLI uses native prebuilds (keytar.node, pty.node). A crash in native code (segfault) would explain:

   • Abrupt stop of events after turn 14 started
   • Zombie state (parent didn't get proper exit notification)
   • No error logged

[github-actions]

Thanks for the detailed report @lukeed!

I've labeled this as:

• bug — process crash and autoRestart failure
• sdk/nodejs — affects the Node.js SDK implementation
• priority/high — impacts stability of long-running sessions

The analysis of zombie processes and the autoRestart logic looks thorough. The hypothesis about native module crashes (segfault) not triggering the exit handler makes sense and would explain why autoRestart isn't firing.

The team will take a look at improving the process lifecycle handling.

AI generated by Issue Triage Agent

[friggeri]

@lukeed I've been unable to reproduce this bug, is there anything more you can share about what made the CLI crash (can you try passing {logLevel: "debug"} to Client and getting the logs for a session)? So far everything I've tried to kill the CLI properly triggers either the exit or close event handler.