Repository navigation
Description
The ChatGPT desktop app's Codex command execution channel repeatedly becomes unresponsive on macOS. The chat and other app functionality remain responsive, but even a trivial command invoked from /tmp produces no output and does not exit through the tool channel. Fully restarting the app temporarily restores execution, but the failure recurs.
This is not a command that is legitimately running for a long time and is not waiting for an approval.
Environment
- ChatGPT desktop app with Codex:
26.1117.31558 - Bundle build:
29513 - macOS:
27.0.1(26A434) - Architecture: Apple Silicon (
arm64) - Session ID:
019fd2ea-cd5c-7920-a6f3-bb8aa51aeb77
Minimal reproduction / health check
After the runner enters the bad state, execute:
command: /bin/echo runner-ok; /bin/date; /usr/bin/uptime
working directory: /tmp
login shell: false
TTY: false/default pipes
initial yield: 3 seconds
Actual behavior
- No output is returned, including from
/bin/echo. - The tool remains running and must be terminated.
- One captured call returned only a running-cell identifier after approximately 31 seconds, still with no command output.
- Computer-use automation and ordinary chat remain responsive.
- Restarting the desktop app restores command execution temporarily, but the failure has recurred several times after restart.
Expected behavior
The command should return immediately. If the command worker or output transport is unhealthy, the app should fail quickly with an actionable error or automatically restart the worker.
Isolation evidence
- Reproduces from
/tmp; it is not dependent on the workspace directory. - Uses absolute-path system binaries and disables login-shell startup.
- Filesystem searches complete immediately whenever the runner is healthy.
- Spotlight was effectively idle during one occurrence (
mdsabout 0.2% CPU,mds_stores0.1%/0.0%,mdworker0.0%). - Other Codex facilities remain responsive during the command-channel failure.
Sanitized desktop-log evidence
During the affected window, the desktop log repeatedly contains:
[IpcClient] Received broadcast but no handler is configured method=thread-queued-followups-changed
At 2026-11-21T14:33:42Z, around restart/recovery, it contains repeated:
[IpcClient] Received broadcast but no handler is configured method=ipc-connection-reset
followed by:
[AppServerConnection] app_server_connection.state_changed cause=stop_process ... previous=connected next=disconnected ... transport=stdio
The local transcript metadata can report a trivial command as completed with zero duration while the calling tool waits roughly 31 seconds and receives no output. That suggests disagreement between execution state and delivery of the tool result.
Privacy note
The complete session transcript contains confidential legal/client information and cannot be posted publicly. I can provide additional sanitized excerpts or use a private upload route if an OpenAI maintainer requests them.
Related issues reviewed
- Desktop app hangs on Thinking indefinitely on macOS; forcing external codex fixes it #13709 covers the macOS desktop remaining in
Thinking; this report differs because the conversation remains responsive while individual command results fail to return. - VS Code Codex queues prompts after completed turn; thread remains markedStreaming=true #50118 contains the same
thread-queued-followups-changedwarning, but concerns follow-up queue state in VS Code rather than repeated local command-runner output loss in the macOS desktop app.
Activity
github-actions commentedon Nov 21, 2026
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action