CLI hangs indefinitely on session resume when ssh got disconnected during a long running command

Title: Codex hang when SSH session closes during active tool call

Date: 2026-02-22
Environment:

• Codex CLI: 0.104.0
• OS: Linux aarch64 (Debian bookworm, kernel 6.12.47+rpt-rpi-v8)
• Shell: bash 5.2.15
• Workspace: /mnt/device1/firmware

Observed behavior:
A long-running tool command was dispatched, entered running state, then the SSH session
closed shortly after. Likely, the rollout never received normal completion for that command and the agent remained in a waiting/stalled state until resumed/interrupted, which leads to a state that, when the session is resumed, it will stay in 'Working' indefinitely. See more details in the protocol

Evidence:
1) Command dispatch:

• rollout file: /home/pi/.codex/sessions/2026/02/22/rollout-2026-02-22T04-05-49-019c834f-

7a6c-7df0-9316-b1d3513a5d20.jsonl

• line 607: service.restartall command
• line 609: “Process running with session ID 59574”

2) SSH close near same time:

• Feb 22 18:20:33 local: pam_unix(sshd:session): session closed for user pi
• captured in /tmp/openai-hang-debug-20260222-2148/02b-ssh-filtered-with-close.txt

3) Related non-graceful task warning:

• codex-tui.log contains: “task ... didn't complete gracefully after 100ms”
• extracted in /tmp/openai-hang-debug-20260222-2148/03-codex-warnings.txt

Hypothesis:
Race condition between tool/process execution state and transport/session teardown (SSH/
PTY), leaving the engine waiting for a completion event that never arrives.

Attached artifacts:

• /tmp/openai-hang-debug-20260222-2148.tar.gz
• sha256: 3bf646afa2b15401f77b291a2ff26750108fc89e554de43c49b44703ce8d5e15