Windows 26.901.4073.0 / CLI 0.153.1: duplicate rollout ordinal stalls history; later conversations disappear after restart #42999

Summary

In a long local paginated conversation, canonical JSONL kept saving user/assistant messages, tool activity and generated-image events, but history projection stopped at a duplicate ordinal. Ongoing work was initially visible; after restarting the desktop app the later conversation was no longer displayed. The screen jumped from an earlier installation-test preparation to the new recovery message.

This is separate from Canvas report #42985. That earlier report followed an in-app feedback submission (feedback ID: 01a05eb4-6d9d-7cd2-8bae-207d9839f69b). No new in-app feedback upload for this history bug is being claimed.

Environment and chronology

• Windows x64 desktop package 26.901.4073.0; installed app-server codex-cli 0.153.1.
• September 5, 2026, UTC+8; one logical local conversation with four linked rollout segments.
• No original rollout or projection was manually modified before diagnosis.
• Last history visible after reopening: installation preparation turn started 14:39, records through ~14:48.
• Canonical records still contained 17:25 first-screen delivery, 17:29–17:32 installation/activation design, 17:39 follow-up requests, 17:46 later mockups and evening feedback discussion.
• Restarting to reload history instead made the later conversation disappear from the display. Comparison found all four backed-up pre-restart file prefixes still byte-for-byte present; the active segment had grown.
• App kept appending canonical records while repeatedly logging the same projection error, without an actionable history-integrity warning to the user.

Direct evidence

Only one non-consecutive ordinal boundary was found in the inspected active file. Every line parsed as JSON. Both SQLite databases passed integrity checks.

Affected segment entry in thread_history_projection_state:

next_rollout_byte_offset = 29202943
next_rollout_ordinal = 8961

The offset pointed exactly to the second record with ordinal 8960. Repeated codex_thread_store::local::live_writer warnings ended with:

failed to project durable rollout ...
thread-store internal error: thread history projection ... expected ordinal 8961, got 8960

Diagnostic detail: linked conversation projection rows are keyed by segment ID. Looking up only the root task ID returns the oldest segment cursor and can mislead diagnosis. Valid JSON/SQLite does not establish readable paginated history.

Was compaction responsible?

The user remembered a recent compaction and suspected it. Relevant later compactions were at 15:34 and 19:28, after the duplicate at 15:13. These do not establish compaction as the cause. The direct failure is duplicate ordinal plus strict projection rejection; the exact writer/resume trigger has not been deterministically reproduced.

Isolated recovery experiment and live workaround

1. Back up all four rollouts with SHA-256 manifests and state/history databases using SQLite online Backup API, including committed WAL data. Keep originals untouched and backups local.
2. Run the same installed 0.153.1 app-server against isolated copies, with copied rollout paths redirected, no credentials and no model turns. Baseline native history stopped at the old installation turn; all six target markers (five screen deliveries plus feedback issue) were absent.
3. In the copy only, advance the affected segment's projection byte offset exactly 649 bytes, 29202943 -> 29203592, keeping next_rollout_ordinal=8961. This bypasses only the duplicate settings event for UI projection; it does not alter/delete the canonical event.
4. Load the copied thread with thread/resume without turn/start and read paginated history. All six markers returned. A second cold app-server read also passed with no ordinal errors. All original JSONL hashes remained unchanged; SQLite integrity was OK.
5. After a fresh backup, apply the same guarded transactional single-row cursor adjustment live at 20:46:51. The running app automatically projected the suffix and kept advancing with new activity. No subsequent matching ordinal errors were observed in the checked interval.
6. All 91 inspected user/assistant completion-event IDs after the boundary existed in repaired projection. This includes recovery messages, not a claim exactly 91 messages disappeared. Pre-repair canonical file prefixes remained unchanged.
7. Projection continued advancing after another compaction during the investigation; latest checked database integrity remained OK.

Desktop-visible refresh confirmation remains pending. Native data reading is recovered, but full UI and subsequent completed-turn validation are not being claimed. This is a local workaround, not a product fix or general advice to edit cursors. A future full reindex with the same strict parser could encounter the preserved duplicate again.

Expected behavior / requested fix

• Derive next writer ordinal reliably from durable state across interrupted turns, restarts and linked segments.
• Safely reconcile duplicate metadata ordinals without hiding all later valid history.
• Preserve canonical records and show an actionable warning when saved and displayed history diverge.
• Provide supported task-scoped diagnosis/rebuild with backup and validation, without hand-editing SQLite cursors.

Related reports: #41657, #41566 and #38792. This report adds evidence from desktop 26.901.4073.0 / CLI 0.153.1 and a tested projection-only recovery. Current main-branch materializer includes duplicate/regressed ordinal handling; please clarify which release contains it and whether it recovers already-stalled projections.

Full transcripts, credentials, local user paths, private conversation screenshots and raw logs are intentionally not attached.