[Windows] Desktop resumes live goal output without backfilling a missing transcript interval

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

Windows Codex App 26.818.3698.0; app-server 0.149.0

What subscription do you have?

ChatGPT Pro

What platform is your computer?

Microsoft Windows NT 10.0.26200.0 x64

What issue are you seeing?

I started a long-running goal in Windows Codex Desktop. Around automatic context compaction, Desktop stopped showing new output even though the task continued and its history was still being saved.

I opened the task on iOS only to confirm that it was still progressing. After I sent a follow-up as a recovery probe, Desktop began showing new activity again but never restored the missed section. The transcript now has an earlier prefix and a later suffix with a permanent gap between them.

This is urgent because Desktop looks current after updates resume. Nothing warns me that part of the completed work is missing, so I cannot reliably tell what Codex has already done to my files from the Desktop transcript alone.

What steps can reproduce the bug?

1. Start an event-heavy goal in Windows Codex Desktop and keep the task view open.
2. Let it run through automatic context compaction.
3. Observe that Desktop may stop appending output while execution continues.
4. Confirm elsewhere that later events were persisted, then send a follow-up from another client.
5. Observe that Desktop displays new events without backfilling the earlier gap.

The failure is intermittent.

What is the expected behavior?

If Desktop misses live events, it should rebuild the transcript from stored history before showing newer ones. If that fails, it should report a sync error instead of presenting an incomplete transcript as current.

Additional information

The incident occurred on August 22, 2026, around 15:30–15:50 CDT (UTC-05:00). I submitted in-product feedback, but no Feedback ID was displayed.

The canonical rollout contains the missing activity and normal turn completion. Its JSONL is valid and chronological, the SQLite backups pass PRAGMA quick_check, and the app-server logs contain no writer conflict, parse failure, or database corruption. The logs also show running thread_resume responses with an active turn after Desktop fell behind, but the visible history remained incomplete.

The 0.149.0 resume path explains how the gap becomes permanent. It rebuilds persisted turns only when the client requests turns or an initial turns page. A metadata-only resume can reattach the live listener without rebuilding history. That allows later events to reach an already incomplete Desktop transcript. The logs do not expose the request’s includeTurns value, so I cannot confirm that flag or determine whether the initial stall began in transport sequencing or the renderer.

Related issues:

• #32466 also covers missed Desktop updates, but Desktop later catches up. Here, the gap remains.
• #35727 reports a partial transcript after a sequence gap. Here, the local Windows view fails first; the second client is only a recovery probe.
• #35955 concerns stale running state, not missing history between later visible events.
• #35746 concerns paginated projection and ordinal reuse. This thread uses legacy history and has no projection ordinals.

I can provide side-by-side screenshots, a redacted event timeline, and integrity-check results privately.