Repository navigation
[macOS Desktop 26.803.41515] Nearly every prompt/goal enters Reconnecting 5/5 and ends with "request timed out" #37660
Description
What version of the Codex App are you using (From “About Codex” dialog)?
ChatGPT/Codex Desktop 26.803.41515 (released Aug 7, 2026)
What subscription do you have?
ChatGPT Pro
What platform is your computer?
- macOS Tahoe
26.5.2 - Apple Silicon / arm64
- MacBook Pro 16-inch
- Apple M5 Max
- 128 GB unified memory
What issue are you seeing?
On the current macOS desktop build, almost every time I send a prompt or start/resume a goal, the request spends several minutes in Working, then enters:
Reconnecting 5/5
and eventually ends with:
request timed out
Representative observed sequence:
Working for 5m 22s
Reconnecting 5/5
and in another occurrence:
Working for 7m 32s
Reconnecting 5/5
request timed out
This is recurrent rather than a one-off failure. In some cases, the task only begins or resumes after the timeout sequence, which makes the error state misleading: the UI appears to report a terminal failure even though work may subsequently continue.
The problem is visible in the ChatGPT macOS desktop application’s Codex surface and affects both ordinary prompts and goal/task starts.
What steps can reproduce the bug?
- Open ChatGPT/Codex Desktop on macOS.
- Open an existing project conversation or goal.
- Send a prompt, or start/resume a goal.
- Wait while the UI remains on
Workingfor several minutes. - Observe
Reconnecting 5/5. - Observe
request timed outafter the reconnect attempts are exhausted. - Retry with another prompt/task; the same sequence occurs frequently.
What is the expected behavior?
- Prompts and goals should begin normally without a multi-minute reconnect loop.
- A transient transport interruption should recover without exhausting all five retries.
- If the underlying task is still alive and later continues, the desktop UI should not present a terminal-looking timeout; it should reconcile and display the authoritative task state.
- If the request truly failed, the app should provide a clear retry/resume action without requiring the user to guess whether work is still running.
Additional information
In-app feedback was submitted with the current ChatGPT session logs and browser tabs/browser logs included.
Feedback ID:
019fda84-a41b-7980-8108-bbb8616dfade
Activity
github-actions commentedon Aug 9, 2026
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
stream disconnected before completionduring long-running tool operations while work continues #37647Powered by Codex Action
qw72tc7fxw-prog commentedon Aug 20, 2026
Adding an independent reproduction from another macOS Codex Desktop user after the recent update.
Observed behavior:
There is also a separate composer regression after the same update: after sending a prompt or using Steer, the submitted text can remain visible in the text box instead of clearing. I am reporting that separately so it can be tracked independently.
qw72tc7fxw-prog commentedon Aug 20, 2026
Additional affected environment from another user reproduction:
26.814.41407Powered by Codex & OWLAug 18, 2026Reconnecting/ timeout behavior after the updateThis suggests the timeout/reconnect regression is still reproducible on the Aug 18 desktop build, not only the earlier Aug 7 build originally reported here.
qw72tc7fxw-prog commentedon Aug 20, 2026
User impact/severity feedback for build
26.814.41407(released Aug 18, 2026):The user describes this as the worst Codex update they have experienced so far because the regression is not isolated to a cosmetic issue. Frequent timeout/reconnecting behavior is disrupting normal work, persists after multiple restarts, and is occurring alongside a separate composer-state regression after Send/Steer. Combined, these issues make the current desktop build materially less usable than the previous version.
qw72tc7fxw-prog commentedon Aug 20, 2026
Additional user impact report for Codex Desktop 26.814.41407 (released Aug 18, 2026):
This is not an occasional timeout. In normal project work, the user is seeing the reconnect/timeout failure roughly every 3–4 commands. When it happens, work stops and the UI does not provide a clear Retry button/action, so the user cannot simply resume the failed turn. Combined with the separate composer-state regression (submitted Send/Steer text sometimes remains in the input box), this makes it unclear what was accepted and materially disrupts ongoing work.
The user has already restarted the app multiple times with no improvement. Their feedback is that this release should have had stronger regression/reliability testing before rollout because the current frequency of failures is directly impacting work progress.
Requested UX/reliability improvements:
soundneedle commentedon Sep 3, 2026
is this the same issue where this temporarily fixes it? This just started happening on my mac and it's so frustrating.
sudo dscacheutil -flushcache
sudo killall -HUP mDNSResponder
johnparkerg commentedon Sep 8, 2026
For me it was my USB ethernet adapter. I worked fine for videocalling so I assumed all was well, but switching to another slower more reliable one proved successful.
hishamfangs commentedon Oct 14, 2026
What the actual F, Open AI.
Got my brand new M5 mac and immediatley hit this reconnection issue. Its been open all year long, how have you not looked at this yet?
It's unsueable. I'm switching to Claude
johnparkerg commentedon Oct 14, 2026
For me it was changing ISPs, I did that and it stopped completely.