macOS Desktop: completed recurring automations retain CUA node_repl workers

Start by reproducing the recurring-automation process count with the supplied ps command and inspect the private CUA spawn site in the ChatGPT Desktop bundle. Consult the public code-mode host runner and app-server integration for lifecycle context. Done means job-owned workers exit and are reaped on every terminal path, or a shared runtime proves its instance count remains bounded and reused.

Summary

Completed recurring automations leave
/Applications/ChatGPT.app/Contents/Resources/cua_node/bin/node_repl
workers alive for hours. These are not zombies: they remain sleeping (S) and
are direct children of one codex app-server.

Environment and evidence

• macOS, ChatGPT Desktop bundle
• Parent command:
  /Applications/ChatGPT.app/Contents/Resources/codex -c features.code_mode_host=true app-server --analytics-default-enabled
• Observed during investigation: 284 node_repl workers, all S, all direct
  children of the same app-server PID
• Worker elapsed times ranged from about one to eight hours
• Earlier observation: 268 workers and 806 automation session records that day
• Worker creation times correlated with recurring-automation executions
• The executable is a private Desktop resource:
  /Applications/ChatGPT.app/Contents/Resources/cua_node/bin/node_repl

Public Codex repository investigation

The public openai/codex checkout has no cua_node or node_repl source
reference. Its generic code-mode host runner can accept an external host path,
but it is designed to reuse one live host connection per ThreadManager and
already has child kill + wait/reap handling when that shared connection is
dropped.

That model does not account for hundreds of direct node_repl children under
one app-server. The concrete CUA executable selection and per-automation
lifecycle appear to be in the Desktop bundle/private CUA runtime.

This is adjacent to #12491, but it should not be assumed to be fixed by #19753:
#19753 explicitly addresses stdio MCP server shutdown, whereas the observed
workers are cua_node/bin/node_repl direct app-server children.

Reproduction

1. Restart Desktop and record the baseline node_repl count.
2. Run a short recurring automation repeatedly and wait for each run to report
   completion.
3. After each completion, record:

   ps -axo pid,ppid,state,etime,command |
     rg '/Applications/ChatGPT\.app/Contents/Resources/cua_node/bin/node_repl'
   

4. Wait at least one hour and compare with the baseline.
5. Expected: no job-owned worker remains after its run reaches a terminal state.
   Actual: the count rises and completed-run workers remain sleeping.

Requested owner investigation

Please route to ChatGPT Desktop / CUA runtime owners, with Codex app-server
code-mode maintainers consulted for integration.

Instrument the private CUA spawn site with:

• automation job/run ID, spawned PID, parent PID, and spawn call site;
• terminal reason: complete, cancel, error, channel close, or app-server stop;
• terminate result and wait/reap result;
• a bounded per-job registry count.

Do not use global pkill. If workers are job-owned, add a regression test that
runs several short jobs and asserts each owned process exits and is reaped on
all terminal paths. If the runtime is intentionally shared, assert a bounded
single-instance count and prove it is reused rather than recreated per run.

Related

• #12491
• #19753
• #24510