fix: Exit Cleanly When Native Executor Shuts Down Concurrently

[danny-avila]

Summary

A native BYOM worker running under systemd with KillMode=control-group exited with status 1 on an ordinary systemctl stop or restart, even while idle, and the journal recorded Native executor is unavailable. The worker restarted and reconnected fine, so every deploy looked like a service failure without being one. Closes #190.

What raced. systemd delivers SIGTERM to the parent CLI and the forked SRT executor at the same moment. The child marks itself shutting down, begins closing its SRT sandbox, and ignores further IPC. The parent aborts the worker loop and reaches nativeCommandSandbox.close() in the CLI finally almost immediately. If the child is still connected at that point, stop() sends the close RPC, the child drops it, and when the child finally exits the lost handler rejects the pending RPC with Native executor is unavailable. That rejection propagated through close(), the CLI finally, and main().catch() into exit status 1. The same sequence applies to Ctrl-C and any supervisor that signals the whole process group.

• Tolerate a lost, refused, or stalled close handshake in NativeProcessWorkspaceCommandSandbox.stop(). The child is terminated in finally whether or not it replies, and stop() has already drained any active command, so those cases carry no mutation ambiguity. A negative close reply from the executor, such as a failed SRT reset or scratch cleanup, is a real failure and still rejects so NativeWorkspaceCommandPool.close() can aggregate it. Quarantine decisions live on the execute path and are unchanged.
• Report exit status 0 from the executor child when its own SRT teardown succeeds on a signal-driven shutdown, and keep status 1 for a failed or timed-out teardown. The parent never inspected this code, so this only makes the journal honest.

Tests

• New: close() resolves when the child exits during the close handshake without replying. This fails on main with the same unavailable rejection seen in the issue.
• New: close() skips the handshake entirely once the child is already lost, and later execute() calls remain fenced.
• New: close() still rejects with the executor's own message when the child replies with an explicit cleanup failure, and the child is terminated.
• Updated: the child signal-routing test now expects exit status 0 for SIGINT, SIGHUP, and SIGTERM.
• Existing coverage for draining an active command before closing IPC, shutdown-receipt fencing, and pre-dispatch atomicity is unchanged and passes.

npm test in packages/code passes 382 of 386 locally. The 4 failures are storage tests that create fixtures under /mnt/c, a WSL DrvFS mount with mode 777, and are unrelated to this change.

Note

The commits on this branch are unsigned because GPG could not prompt for a passphrase in the environment it was authored from.

[danny-avila]

@codex review

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-09-13T08:04:34.855714Z	2ff14c8	Manual request

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" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Nice work!

Reviewed commit: 2ff14c8273

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".