Repeated Python child-process SIGSEGVs during concurrent Codex desktop agent execution on macOS

Environment

• Codex desktop: 26.908.70816, build 9275
• Bundle identifier: com.openai.codex
• macOS: 26.6.2 (25G83)

Summary

During one local multi-agent Codex desktop session, six distinct Python child processes crashed within approximately 105 seconds. Codex remained running, but macOS displayed a separate Python crash dialog for each event.

Observed behavior

All six events reported:

• Parent process: codex
• Responsible application: Codex desktop
• Exception: EXC_BAD_ACCESS
• Signal: SIGSEGV / signal 11
• Crashing thread: Python main thread
• Process lifetime: approximately 0.16–0.32 seconds

Two crashes used Apple Python 3.9.6 and four used Homebrew Python 3.14.7. Frames varied across CPython allocation, garbage collection, type handling, _ctypes, and _json. This appears consistent with native memory corruption, although the initiating command is unknown because the reports omit child argv and working directory.

Impact

• Repeated macOS crash dialogs during agent work
• Python-backed commands may fail before producing trustworthy output
• Concurrent-agent results cannot safely be assumed complete after a child crash

Observed trigger conditions

1. Open a local project in Codex desktop.
2. Run multiple concurrent local agents.
3. Have those agents invoke short Python-backed commands within a narrow interval.
4. Multiple Python children may terminate with SIGSEGV.

This is not yet a deterministic reproduction.

Post-restart control

After restarting Codex desktop, this controlled probe completed successfully:

PYTHONFAULTHANDLER=1 PYTHONMALLOC=debug python3 -c 'import json; print(json.dumps({"probe":"ok"}))'

Result: Python 3.14.7, exit status 0, no stderr or crash dialog.

This suggests the incident may depend on concurrency, accumulated session state, inherited environment, or a particular Python-backed command rather than a bare Python launch.

Expected behavior

Python child processes launched by Codex should complete normally or return an ordinary command failure without crashing the interpreter.

Diagnostic request

Which Codex desktop logs capture the exact argv and inherited environment for local child processes? That would allow a controlled reproduction to isolate Codex supervision, inherited environment, or a native Python helper.

Start by identifying which Codex desktop logs capture argv and the inherited environment for local child processes. Reproduce concurrent Python-backed commands if possible and compare those logs with the successful PYTHONFAULTHANDLER/PYTHONMALLOC probe. Done means the failing child invocation and whether the fault lies in supervision, inherited state, or a native Python helper are isolated.