openai / openai/codex

Windows app hangs and missing history: duplicate rollout ordinal persists on 26.903.9818.0 / CLI 0.153.4

Open
#44,511 1 comment 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

app bug performance session windows-os
Dominant language
Rust
Stars
125k
Forks
19.4k
PR merge metrics
PR metrics pending

Description

Incident

Codex Desktop on Windows intermittently becomes unresponsive, and a long-running local task loses recent visible conversation history after reopening. Investigation on 2026-09-10 confirmed a duplicate rollout ordinal blocking history projection. The raw conversation remains on disk. This is a new incident consistent with #41566; related reports include #42241 and #42197.

Environment
  • Current installed Windows x64 package: OpenAI.Codex 26.903.9818.0
  • Running bundled CLI: 0.153.4
  • Windows 11 Pro 10.0.26200
  • Windows recorded an Application Hang for earlier package 26.903.8094.0 (ChatGPT.exe file version 152.0.7977.83).
  • Subscription omitted.
Confirmed history defect

The affected task uses paginated history and has an earlier rollout plus a continuation segment. In the continuation, two consecutive records have ordinal 11985:

Physical line (1-based) Ordinal Event UTC timestamp
5686 11984 custom_tool_call_output 2026-09-09 23:29:47.753
5687 11985 token_count 2026-09-09 23:29:47.777
5688 11985 thread_settings_applied 2026-09-09 23:31:01.247
5689 11986 task_started 2026-09-09 23:31:01.273

The projection cursor is exactly at byte 185793634, immediately after physical line 5687, with next_rollout_ordinal=11986.

Sanitized error:

failed to project durable rollout:
thread-store internal error: thread history projection
expected ordinal 11986, got 11985

The same mismatch rejects thread/revert. Desktop logs show it by 2026-09-09 23:46 UTC and again on 2026-09-10 09:29 UTC. The currently retained SQLite logs contain 554 matching warnings between 2026-09-10 11:15:46 and 11:57:20 UTC. The problem persists across app launches and the newer installed package.

Validation
  • Continuation rollout: 416218331 bytes, 11048 JSONL records, zero malformed JSON lines.
  • Earlier rollout: 255463052 bytes, 6308 records, contiguous ordinals.
  • Continuation contains one duplicate ordinal at the boundary above.
  • Read-only PRAGMA quick_check returns ok for state_5.sqlite and thread_history_1.sqlite. This is semantic event/index inconsistency, not detected SQLite structural corruption.
  • migrate-rollouts --thread --json (without --apply) reports already_paginated and bytes_processed=0 for both segments.
  • Both raw rollouts and SQLite online-backup snapshots were preserved locally. A separate readable recovery export was produced. Live rollout and database contents were not edited.
App unresponsiveness: confirmed symptom, cause unresolved

Windows Application event 1002 at 2026-09-09 23:28:32.185 UTC says ChatGPT.exe stopped interacting with Windows and was closed; HangType is "Top level window is idle".

Desktop logs also contain 571 "ResizeObserver loop completed with undelivered notifications" errors on 2026-09-10, across multiple app launches, during approximately 00:04–11:57 UTC. Several recent process logs end near these errors, but this does NOT establish them as the crash cause. No Crashpad dump was found in the inspected app locations.

The user suspects computer use. That relationship is unconfirmed. There was ample free RAM and disk at inspection, which does not rule out earlier transient resource pressure. Please investigate hangs separately from the proven history defect.

Reproduction status and expected behavior

No destructive crash reproduction was attempted. Observed sequence: lengthy task with tools, app interruption/reopen, then thread_settings_applied repeating the preceding token_count ordinal. The history projector should recover safely from an interrupted turn and continue showing subsequent saved messages, or offer a supported repair/reindex operation.

Please provide a safe recovery path for already-paginated continuation histories and fix duplicate ordinal emission. Please also investigate the Windows UI hangs/layout errors.

Privacy: raw transcripts, screenshots, private project names, account identifiers, machine/user names, filesystem paths, and session IDs are intentionally excluded.

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start by separating the confirmed duplicate ordinal in the paginated JSONL continuation from the unconfirmed Windows UI hang. Review the preserved rollout data and the migrate-rollouts --thread dry-run results, then trace the history projection and recovery path. Done means safe recovery for affected histories, corrected duplicate ordinal emission, and a separately supported investigation of the Windows hang.

Written by the indexing model from the issue text.

Assessment

Tech stack
rust, sqlite
Domain
databases, desktop, operating-systems
Issue type
Bug
Difficulty
5/5
Estimated time
Over a week
Activity status
Active
Clarity
Mostly clear
Newbie friendliness
28/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.