WSL tab close during running turn leaves resumed session stuck in Working (missing task_complete) #12843

Summary

When an interactive Codex session is still running and the WSL Ubuntu terminal tab is closed via the tab-bar x button, reopening with codex resume can leave the session stuck in Working (...) forever.

In the affected session file, the last turn has task_started but no matching task_complete (and no turn_aborted).

Environment

• OS: WSL Ubuntu (user environment)
• Codex CLI: 0.104.0
• Date observed: 2026-02-25 to 2026-02-26 (UTC timestamps in logs)

Reproduction (observed flow)

1. Start a normal interactive Codex session.
2. While a turn is still running (Working shown), close the terminal tab using the tab bar x button (abrupt terminal close).
3. Reopen terminal and run codex resume for that session.
4. UI remains in Working (..s • esc to interrupt) and does not recover.

Expected

Session should recover gracefully on resume, or auto-mark the interrupted turn as aborted/completed so Working does not stay forever.

Actual

Working remains indefinitely. The underlying session JSONL has an unmatched task_started turn.

Evidence

1) Missing terminal turn-end event in session logs

Session A:

• file: ~/.codex/sessions/2026/02/25/rollout-2026-02-25T23-20-34-019c952c-4e6d-7dc3-bdfd-b070a914b7aa.jsonl
• has: task_started for turn_id=019c9578-4a8a-7e33-9d1e-d65708ffa679
• originally missing: task_complete / turn_aborted for same turn

Session B:

• file: ~/.codex/sessions/2026/02/15/rollout-2026-02-15T18-21-15-019c609a-ad65-7761-a862-3af2e958375a.jsonl
• has: task_started for turn_id=019c9517-3f2d-7b50-8057-32c3df1d850d
• originally missing: task_complete / turn_aborted for same turn

2) TUI error around the same period

From ~/.codex/log/codex-tui.log:

• 305225: 2026-02-26T00:02:57.055045Z ERROR codex_tui::chatwidget::agent: failed to submit op: internal error; agent loop died unexpectedly
• 305226: 2026-02-26T00:02:57.055061Z ERROR codex_tui::chatwidget::agent: failed to submit op: internal error; agent loop died unexpectedly

Temporary workaround (worked)

Manual log repair:

1. Backup the affected session JSONL.
2. Append a synthetic event:

{"timestamp":"<utc>","type":"event_msg","payload":{"type":"task_complete","turn_id":"<stuck_turn_id>","last_agent_message":null}}

3. Resume/restart Codex.

After this, stuck Working was cleared in both reproduced sessions.

Notes

This looks like resume recovery does not reconcile dangling task_started turns after abrupt terminal/tab termination.