Codex Remote on Android: very large task history causes indefinite cold-load and temporary list disappearance

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

Codex desktop 26.818.61809 (build 7019). The Remote client is ChatGPT for Android 1.2026.230 (versionCode 2623032).

What subscription do you have?

Not included in this public report.

What platform is your computer?

macOS 26.6.2. Remote client: Android 16 on a Samsung device.

What issue are you seeing?

One long-running Codex task with a very large local history consistently cold-opened in Codex Remote on Android to an empty skeleton/loading screen and never completed. Other Remote tasks loaded normally.

The same task also temporarily disappeared from the Android task list and search even though it remained available on desktop. An exact deep open still reached the correct task identity, but the content stayed on the loading skeleton.

While the affected task was already open, newly generated incremental turns could appear. Closing it and cold-opening it again returned to the indefinite skeleton. A native compaction reduced active context usage but did not fix the cold-load failure.

What steps can reproduce the bug?

1. On macOS, use one Codex task long enough for its local rollout/history to become very large. In the observed case it was approximately 244 MB across about 38,000 JSONL records, including several records larger than 1 MiB.
2. Open that task from Codex Remote in the ChatGPT Android app.
3. Observe that the task remains on an empty loading skeleton indefinitely, while smaller Remote tasks open normally.
4. Return to the task list. The affected task may disappear from list/search even though it still exists and opens on desktop.
5. Deep-open the same task identity. It resolves to the task, but the history still does not render.
6. Run native compaction on desktop and retry a cold open. The failure can persist despite the reduced active context.

What is the expected behavior?

Remote should render a bounded recent tail quickly, then page older history on demand. Compaction boundaries should be honored during Remote hydration. A task should remain discoverable independently of whether its full history can be hydrated, and a bounded error/retry state should replace an infinite skeleton if hydration fails.

Additional information

A controlled, backup-backed diagnostic retained the same task identity while reducing the active local rollout to approximately 1.0 MB and 59 records. After that reduction, the task reappeared in the Android list and succeeded on two separate cold opens. This strongly points to unbounded or complete-history hydration rather than authentication, connectivity, or task-identity failure.

The shipped client contains a thread-follower-load-complete-history operation. That name, combined with the behavior above, suggests that Remote cold-open may request the complete legacy display history instead of a bounded or compacted tail. This is an inference, not a confirmed implementation diagnosis.

Suggested regression coverage:

• A task with a rollout larger than 200 MB and individual records larger than 1 MiB.
• Cold-open versus incremental-follow behavior on Android Remote.
• List/search discoverability when history hydration is slow or fails.
• Hydration after native compaction.
• Cursor-based or paginated history loading with explicit size/time limits.

No private task title, account/contact information, task or session identifier, device serial, host name, local path, transcript content, raw log, recording, screenshot, evidence link, or private hash is included here. Detailed evidence has been provided privately to OpenAI Support.