Codex Desktop: completed threads become inaccessible after update

Summary

After updating Codex Desktop and signing in again, some completed chat threads become inaccessible from the sidebar. The UI shows a generic error:

Oops, something went wrong

This is not limited to old pre-update conversations. A new thread created after the update also became inaccessible immediately after a normal completed turn.

In an earlier affected thread, the UI exposed a more specific resume error showing a path-normalization mismatch between a normal Windows path and an extended-length Windows path:

cannot resume running thread ... with stale path: requested <normal Windows Codex session path>, active <extended-length Windows Codex session path>

The visible mismatch was essentially:

requested: C:\...
active:    \\?\C:\...

Environment

• Product: Codex Desktop
• OS: Windows
• Codex Desktop version observed in logs: 26.513.31313
• Codex CLI/runtime version for the new affected thread: 0.131.0-alpha.9
• Auth mode: ChatGPT account
• Local timezone: UTC+05:00

Affected Threads

New post-update thread

• Thread id: 019e2fa8-bd2c-7911-bd70-ec40e63c98ad
• Thread was created after the Codex Desktop update.
• Rollout JSONL file exists in the local Codex sessions folder.
• Rollout file size: about 5.4 MB.
• Created: 2026-05-16 07:20:48 UTC
• Last updated: 2026-05-16 09:10:53 UTC
• Thread state:
  • parent rollout JSONL exists
  • parent rollout tail contains the final assistant answer
  • parent rollout tail contains task_complete
  • all 14 child subagent threads are marked closed
  • the Git commit created during the turn exists
  • project working tree was clean after the commit

Last commit created by the affected thread:

a9d3c42 Fix profile-scoped settings and sources

Checks reported by the completed thread before it became inaccessible:

ruff check autobrief tests -- passed
pytest -- 353 passed
git diff --check -- no errors, only standard CRLF warnings
git status --short --ignored=no -- clean after commit

Older affected thread with explicit stale-path error

• Thread id: 019e27a1-3c0e-7d61-ab57-9b3360b3c3ad
• Earlier CLI/runtime version: 0.130.0-alpha.5
• The resume error reported a path-normalization mismatch between normal Windows paths and extended-length Windows paths:

requested: C:\...
active:    \\?\C:\...

Reproduction Pattern

1. Update Codex Desktop.
2. Fully exit Codex Desktop and sign in again.
3. Open or create a thread in a Windows project workspace.
4. Work normally in the thread. In the new post-update case, the thread used subagents/Critic, completed code changes, created a Git commit, and returned a final answer.
5. The rollout JSONL records a normal final answer and task_complete.
6. Return to the thread from the sidebar or try to resume it.
7. The Codex UI fails to open the conversation and displays a generic error page.

Expected Result

A completed thread with a valid rollout JSONL and task_complete should remain readable and resumable from the sidebar.

Actual Result

The thread becomes inaccessible from the sidebar. The UI shows a generic failure screen. In at least one older affected thread, the underlying error was a stale-path mismatch involving:

C:\...
\\?\C:\...

Relevant Log Evidence

Around the completion time of the new affected thread, the logs show normal completion:

2026-05-16 09:10:53 UTC
codex_core::stream_events_utils: final answer emitted
codex_core::session::turn: token_limit_reached=false model_needs_follow_up=false has_pending_input=false needs_follow_up=false
rollout JSONL: task_complete

Afterwards, during startup/resume attempts, the logs show repeated backend/auth-related failures:

remote plugin sync request to https://chatgpt.com/backend-api/plugins/featured failed with status 403 Forbidden
remote plugin sync request to https://chatgpt.com/backend-api/plugins/list failed with status 403 Forbidden
startup websocket prewarm setup failed: unexpected status 403 Forbidden, url: wss://chatgpt.com/backend-api/codex/responses
failed to connect to websocket: HTTP error: 403 Forbidden, url: wss://chatgpt.com/backend-api/codex/responses
codex_core::client: falling back to HTTP
stream disconnected - retrying sampling request

There are also local skill-loading warnings during session startup:

failed to load skill ... SKILL.md: missing YAML frontmatter delimited by ---

These skill warnings may be unrelated, but they occur during the same resume/startup phase. Even if local custom skills are malformed, a completed historical thread should remain readable or degrade gracefully.

Local Checks Already Performed

• Verified that the new affected rollout JSONL exists.
• Verified that the rollout tail contains a normal final answer and task_complete.
• Verified that all child subagent thread edges for the new affected thread are closed.
• Verified that the Git commit mentioned by the final answer exists.
• Verified that the project working tree was clean after the commit.
• For older affected threads, tried closing stale open child edges in local SQLite state; this did not fix resume.
• For the older stale-path thread, tried changing the stored thread rollout path from normal Windows form to extended-length \\?\ form; this also did not fix resume.

These checks suggest the problem is likely in Codex Desktop's resume/render/state layer or backend-auth integration after the update, not simply an unfinished local task or a corrupt missing rollout file.

Impact

This blocks access to active project history in Codex Desktop. The affected user loses the ability to reopen important working threads after they complete, including a new thread created after the update.

Privacy Note

Raw local logs should not be attached publicly without redaction. They may contain:

• local filesystem paths
• account identifiers or email addresses
• environment variables
• token-like secrets

Sanitized log excerpts and screenshots can be provided through a secure OpenAI support channel.

Attachments To Provide

• Screenshots of the sidebar thread and the generic error screen.
• Sanitized log excerpts around:
  • 2026-05-16 09:10:53 UTC
  • subsequent resume/startup attempts
• A sanitized tail excerpt of the affected rollout JSONL showing the final answer and task_complete.
• Sanitized state rows for the affected thread ids from local Codex SQLite state.

Hypothesis

There may be a regression in Codex Desktop after the recent update involving thread resume state, path normalization, or backend-auth/WebSocket/plugin startup behavior. The older explicit error shows a normal Windows path versus an extended-length Windows path. The newer post-update thread no longer exposes the same detailed error in the UI, but it fails in the same user-facing way despite a clean completed local thread state.