Codex Desktop: new Worktree chat shows “Conversation not found” during thread/resume race

Codex Desktop version

• ChatGPT/Codex Desktop: 26.810.41047 (build 6570)
• Bundled codex-cli: 0.148.0-alpha.9

Platform

macOS

Summary

Starting a new chat directly in Worktree mode can show a red “Conversation not found” toast even though the managed worktree was created successfully and the chat may continue working afterward.

Sanitized diagnostic events indicate a timing-sensitive race between managed-worktree ownership setup, rollout materialization, renderer registration, and thread/resume.

Steps to reproduce

1. Open a Git repository in Codex Desktop.
2. Start a new task.
3. Select Worktree mode and a starting branch.
4. Submit a prompt.
5. Wait for managed-worktree setup to finish.
6. Observe a red “Conversation not found” toast shortly after the task starts.

This has reproduced repeatedly when creating fresh Worktree-mode tasks.

Expected behavior

The new task should open without an error toast. Desktop should wait or retry until the rollout is available, or use the live thread state returned by thread/start, before attempting to resume/hydrate the conversation.

Conversation events should not be treated as belonging to an unknown conversation while the task is being registered.

Actual behavior

In one sanitized reproduction, the relative sequence was:

T+0 ms    [worktree-create] ready
T+194 ms  thread/start succeeded
T+479 ms  thread/read succeeded
T+598 ms  thread/resume failed with -32600:
           "no rollout found for thread id <redacted>"
T+605 ms  [worktree-create] owner-thread-set
afterward  "Received turn/started for unknown conversation"
           appeared in both the primary renderer and the hidden avatarOverlay renderer

In another reproduction, thread/resume succeeded quickly, but the renderer still logged events for an unknown conversation. This suggests the issue is timing-dependent rather than a failed Git worktree creation.

The resulting worktree is valid, and the task can remain usable after the transient toast.

Additional context

This is distinct from the earlier fixed-delay owner-discovery symptom discussed in #37398:

• The previous maybe_resume_owner_discovery_failed timeout is no longer present.
• The related helper process is no longer connected.
• The toast still occurs on the newer Desktop build above.

Potentially related reports, but none appears to cover this exact managed-worktree startup flow:

• #32728 — no rollout found during new task creation, but not Worktree-specific
• #25621 — empty rollout followed by unknown-conversation errors, with a different trigger
• #34504 — valid worktree followed by conversation-start failure, but without a successful thread/start
• #21167 — unknown-conversation routing, not Worktree-specific
• #37398 — similar visible toast caused by a different owner-discovery timeout

Privacy

No raw logs, screenshots, prompts, account details, repository names, local paths, or full thread/worktree identifiers are included.