Scheduled Codex automation metadata calls stall while substantive run succeeds

Open 💬 0 comments Opened Aug 26, 2026 by adaighy

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

0.149.0-alpha.4.3

What subscription do you have?

Not disclosed

What platform is your computer?

Microsoft Windows NT 10.0.26200.0 x64

What issue are you seeing?

Summary

A standalone Codex Desktop scheduled task can complete its substantive work successfully, while supported task-metadata operations issued from the scheduled context remain running for several minutes without returning success or an error. The affected operations were the current-task title update and the task-message callback.

Environment

  • Codex Desktop on Windows
  • Codex app/CLI version from session metadata: 0.149.0-alpha.4.3
  • Windows: Microsoft Windows NT 10.0.26200.0 x64
  • Local project-scoped cron automation; Europe/Amsterdam timezone
  • No raw session log uploaded

Observed evidence

On 26 Aug 2026, the X Overnight Intelligence Briefing scheduled run began around 07:30 Amsterdam (05:30 UTC). The substantive briefing published successfully and its saved outputs passed normal verification. In the same run:

  • The opening current-task title call remained running for more than four minutes and was terminated.
  • The final title call remained running for about three minutes and was terminated.
  • The task-message callback also stalled and did not return.
  • The run completed with the metadata operations unconfirmed.

Sanitized diagnostic excerpts:

  • source: vscode; thread_source: automation; cli_version: 0.149.0-alpha.4.3
  • Script running with cell ID ...
  • Script terminated
  • task_complete was recorded after substantive work, while metadata calls had not returned.

Similar scheduled-context title/callback failures are recorded in available runs from 23, 24, and 25 Aug 2026, with exact duration varying by run. The same title operation from a normal foreground Chief task, using the exact task ID, returned in approximately 0.2 seconds.

Impact

The briefing can succeed, but the scheduled run cannot reliably update its dated title or notify its parent task. This leaves stale task metadata and requires a separate recovery workflow. The current workaround is for a later task to identify and rename the exact completed run; the scheduled task still lacks a prompt completion or clear failure for supported metadata calls.

What steps can reproduce the bug?

  1. In Codex Desktop on Windows, create a standalone local cron automation in a project with a prompt that performs a small substantive action, calls the current-task title operation at the start or end, and sends a completion message to a known parent task.
  2. Let it fire on schedule with the app open and machine awake.
  3. Observe the scheduled run execute its substantive work.
  4. Observe the title operation and/or task-message callback remain running for several minutes without success or an error.
  5. Compare by issuing the same title operation from a normal foreground task, targeting the exact task ID; it returns promptly (about 0.2 seconds here).
  6. Repeat across daily scheduled runs; the scheduled-context behavior recurs.

A minimal reproduction does not require X, publishing, financial data, credentials, or any external connector; the issue is the scheduled-context metadata bridge.

What is the expected behavior?

A scheduled task should be able to use its documented/current callable task-metadata operations. Each metadata call should either complete promptly and report success, or return a clear actionable error within a bounded time. A successful substantive run should not be left with silently unconfirmed title or callback operations.

Additional information

This report is intentionally sanitized. The affected workflow is a standalone X overnight briefing, but the failure is at the Codex task-metadata layer and does not depend on its content. No full task identifiers, account details, email address, private project contents, credentials, or raw session log are included.

Closest public reports reviewed: #33503 (scheduler-vs-Run-now tool-manifest difference), #33544 (automation run identity/status), #22475 (dynamic automation/session titles), and #19969/#21344/#26086 (empty scheduled sessions). These are adjacent but not duplicates.

View original on GitHub ↗