Automatic context compaction hangs for 20+ minutes and immediately restarts after stop on GPT-6 Astra Ultra

What version of the Codex App are you using (From “About Codex” dialog)?

26.901.5280.0

What subscription do you have?

Pro

What platform is your computer?

Microsoft Windows NT 10.0.26200.0 x64

What issue are you seeing?

In a long-running Codex App project thread using GPT-6 Astra Ultra, automatic context compaction became stuck for more than 20 minutes and the task did not resume.

The UI remained on Context automatically compacting, then showed Reconnecting 2/2 and stream disconnected before completion: idle timeout on socket. The thread had made a file change immediately beforehand, but produced no completion or recoverable result after the compaction began.

I manually stopped the stalled turn after more than 20 minutes and sent a new message. Instead of recovering or responding normally, the app immediately started Context automatically compacting again. This leaves the thread trapped in a repeat-compaction loop even after a manual stop.

At the time, the context window showed 83% full: 215k / 258k tokens used.

Feedback ID: 01a07197-ed40-7ce0-933f-ab0c19da0224

What steps can reproduce the bug?

1. Open a long-running project thread in the Codex App on Windows.
2. Use GPT-6 Astra Ultra for an extended coding/integration task with substantial accumulated context.
3. Continue the conversation until automatic context compaction starts.
4. Observe that the UI remains on Context automatically compacting for more than 20 minutes.
5. The app attempts to reconnect (Reconnecting 2/2) and reports an idle socket timeout, but the turn does not complete or recover.
6. Manually stop the stalled turn.
7. Send a new, ordinary message in the same thread.
8. Observe that the app immediately starts automatic context compaction again instead of answering the new message.

Model: GPT-6 Astra Ultra
Context-window usage: 83% full, 215k / 258k tokens used
Feedback ID: 01a07197-ed40-7ce0-933f-ab0c19da0224

What is the expected behavior?

Automatic compaction should finish within a reasonable time and resume the same task without losing its work. If compaction or the response stream fails, the app should fail promptly with an actionable retry/recovery option rather than remaining indefinitely stuck. Manually stopping the stalled turn should allow the next message to run normally, rather than immediately entering compaction again.

Additional information

Observed on 5 September 2026 in a project thread titled Astra - Lead Developer. The screen showed one file changed (+11 -1) immediately before or during the stalled turn. The first compaction indicator remained active for over 20 minutes. After I stopped it and sent the message Stopped as was compacting for over 20 mins, a second turn displayed Context automatically compacting again after about one minute, without providing a normal response. The context indicator showed 83% full (215k / 258k tokens used). Screenshots of the stalled turn, repeated compaction and context usage are available.

This appears related to open issues #14342 and #31834, but this report provides a current Windows/GPT-6 Astra Ultra reproduction, the failed-stop recovery sequence, exact context usage and a feedback ID for server-side correlation.

Start by reproducing the Windows Codex App sequence in a long-running GPT-6 Astra Ultra thread at roughly 215k/258k tokens, then inspect the compaction and reconnect behavior described in the report. Use feedback ID 01a07197-ed40-7ce0-933f-ab0c19da0224 and compare with issues #14342 and #31834. Done means compaction completes or fails promptly, and stopping it allows the next message to run normally.