Codex 5.3: Parallel Long-Running Tasks Can Cause Session Crashes / Truncated Output

When running multiple long-running audit or validation tasks in parallel under Codex 5.3, the execution environment may become unstable, leading to session termination, truncated output, or incomplete task results.

This issue is reproducible when two resource-intensive jobs are launched concurrently. Running the same tasks sequentially (one at a time) resolves the problem consistently.

This appears to be an environment/runtime stability issue rather than a project-specific problem.

Thread ID for reference: 019c324c-bee8-7b71-a5cb-1e0d46615aaa

---

Environment

• Codex Version: 5.3
• Execution Mode: Interactive shell / background terminal
• Host Context: Long-running build/audit workflows (Nix-based, but not project-specific)
• Platform: Codex-hosted environment

---

Background / Context

While performing validation work involving long-running audit pipelines, I observed repeated failures when tasks were executed in parallel.

Typical symptoms included:

• Session termination before completion
• Missing final output lines
• Silent failures with no explicit error
• Background terminals closing unexpectedly

The same workflows consistently completed successfully when executed sequentially.

---

Steps to Reproduce

1. Start two long-running, CPU- and I/O-intensive validation tasks in parallel in Codex.

Example pattern:

```bash
# Terminal 1
long_running_task_A &

# Terminal 2
long_running_task_B &
```

1. Allow both jobs to run concurrently for several minutes.

1. Observe behavior:

• One or both sessions may terminate
• Output may be truncated
• Final success/failure status may never be reported

1. Re-run the same tasks sequentially:

``bash
long_running_task_A
long_running_task_B
``

1. Observe that both complete successfully.

---

Expected Behavior

• Parallel long-running tasks should complete reliably
• Output streams should not be truncated
• Sessions should remain stable until all jobs finish
• Exit statuses should be preserved

---

Actual Behavior

• Sessions may close unexpectedly
• Output is sometimes cut off
• Final status is missing
• Jobs may appear to “hang” or silently terminate

---

Impact

This affects workflows that rely on:

• Parallel validation
• Concurrent builds
• Multi-worktree audits
• CI-like testing inside Codex

It reduces reliability for users performing heavy or long-running verification work and makes parallel execution unsafe.

---

Workaround

Run long-running tasks sequentially instead of in parallel.

Example:

long_running_task_A && long_running_task_B

This consistently avoids crashes.

---

Additional Notes

• The issue does not appear related to the specific tooling used.
• It reproduces across multiple runs.
• It only manifests under parallel execution.
• Sequential execution with identical commands is stable.

This suggests a resource management, session lifecycle, or output buffering issue in Codex 5.3.

---

Request

Please investigate possible issues in Codex 5.3 related to:

• Resource limits under parallel workloads
• PTY/session lifecycle handling
• Output buffering
• Background job supervision
• Long-running process management

Improving stability here would significantly help users running complex verification pipelines.