gpt-5.5 standard streams disconnect after Codex Desktop 26.609.30741 update; high-speed profile works
GitHub Issue Draft: gpt-5.5 standard streams disconnect after Codex Desktop update; high-speed profile works
Title
gpt-5.5 standard streams disconnect after Codex Desktop 26.609.30741 update; high-speed profile works
Body
What happened?
After updating Codex Desktop to 26.609.30741, many existing sessions fail before producing a response when using gpt-5.5 with the standard/default speed profile.
The app repeatedly shows reconnect attempts such as:
Reconnecting... 4/5
stream disconnected before completion
One error path also shows:
Falling back from WebSockets to HTTPS transport.
stream disconnected before completion
So this is not only a WebSocket transport issue; the HTTPS fallback also fails.
Environment
- Codex Desktop app:
26.609.30741 - Bundle version:
3808 - Bundled CLI:
codex-cli 0.140.0-alpha.2 - OS: macOS
- Update time: 2026-06-11 around 15:16 CDT
Behavior
Failing path:
gpt-5.5standard/default speed profile- Existing sessions with substantial context
- Repeated streaming disconnects / reconnect attempts
- Some sessions fail during pre-sampling compaction
Working path:
gpt-5.5with speed set to high works in the same local environment- Other lighter profiles have also worked
Local log snippets
Sanitized local logs show:
remote compaction v2 stream failed; retrying request after delay
compact_error=stream disconnected before completion
Failed to run pre-sampling compact
stream disconnected - retrying sampling request (4/5 ...)
stream disconnected - retrying sampling request (5/5 ...)
Why this looks profile-specific
The same local machine and project can continue if the model/speed profile is changed. The failure is concentrated on gpt-5.5 standard/default, while gpt-5.5 high-speed works.
Expected behavior
Existing sessions should compact and continue, or fail gracefully without becoming unusable in the UI.
Actual behavior
The client reconnects repeatedly and fails. Some old sessions cannot be reopened normally through the UI even though local JSONL transcripts are still present.
Notes
Request IDs and detailed logs are available privately to OpenAI support. I am not including request IDs or account-specific logs in this public issue.
Local session data was backed up before further troubleshooting.
This issue has 1 comment on GitHub. Read the full discussion on GitHub ↗