Codex Desktop: deleted 15-minute automation reappears, creates repeated threads, and consumes tokens

Open 💬 0 comments Opened Aug 19, 2026 by luckymgoon-dotcom

Summary

On Codex Desktop for Windows, a 15-minute scheduled automation was deleted multiple times from the Scheduled UI, but it reappeared and continued creating new threads on the same 15-minute cadence.

This is not only a stale-card problem: repeated automation-created threads accumulated, and 15 of them contain recorded model token usage. The delete action should durably cancel the next trigger and prevent scheduler/cache rehydration.

Environment

  • Product: Codex Desktop
  • Platform: Windows 11, OS build 10.0.26200.0
  • Automation type: local cron automation
  • Schedule: every 15 minutes
  • Automation name/id, local paths, account identity, and private prompt contents are intentionally redacted

Steps to reproduce

  1. Create a local scheduled automation that runs every 15 minutes.
  2. Let it create one or more scheduled-run threads.
  3. Open Scheduled in Codex Desktop and delete the automation.
  4. Observe that the automation appears deleted.
  5. Wait until the next 15-minute boundary.
  6. The same automation reappears and creates another thread.
  7. Repeat deletion; the behavior repeats.

Actual behavior

  • The user deleted the same scheduled automation several times.
  • The automation returned on the same 15-minute cadence.
  • 27 automation-created threads accumulated.
  • The Scheduled UI, local automation state, and sidebar thread state did not remain consistent.
  • A later programmatic delete reported success and the local automation definition became absent, while the UI still displayed the scheduled item and generated threads until separate cleanup.
  • This caused repeated unwanted model runs, thread clutter, troubleshooting work, and possible rate-limit or credit consumption.

Expected behavior

Deleting an automation should:

  1. durably remove the backend/local automation definition;
  2. cancel the already-planned next trigger;
  3. write a deletion tombstone or equivalent guard so stale scheduler/cache state cannot recreate it;
  4. update the Scheduled UI only after durable deletion succeeds;
  5. return an explicit error instead of optimistic success if any deletion layer fails;
  6. prevent any model usage from a run that starts after deletion was acknowledged.

Client-observed token impact

Token usage was read from the final cumulative token_count event in each affected local thread. These are client-observed counters, not an invoice, and cached input may be accounted differently for rate limits or billing.

Unwanted automation-created threads

  • Threads created: 27
  • Threads with token events: 15
  • Zero-token threads: 12
  • Input tokens: 3,659,779
  • Cached input tokens: 2,852,352
  • Uncached input tokens: 807,427
  • Output tokens: 31,780
  • Reasoning output tokens: 5,210 (subset of output)
  • Total tokens: 3,691,559

This 3,691,559-token total is directly associated with the unwanted scheduled-run threads.

Troubleshooting/remediation thread

  • Input tokens: 4,446,096
  • Cached input tokens: 4,274,176
  • Uncached input tokens: 171,920
  • Output tokens: 26,534
  • Reasoning output tokens: 11,000 (subset of output)
  • Total tokens: 4,472,630

This troubleshooting total includes legitimate user messages and remediation work, so it should not all be classified as waste. It is included to show the additional cost imposed by diagnosing and cleaning up the deletion failure.

Combined observed usage

  • Input tokens: 8,105,875
  • Cached input tokens: 7,126,528
  • Uncached input tokens: 979,347
  • Output tokens: 58,314
  • Reasoning output tokens: 16,210 (subset of output)
  • Total tokens: 8,164,189

Requested remediation

  1. Fix automation deletion so UI, scheduler, local store, and queued next-run state are changed as one durable operation.
  2. Add a final deletion check immediately before every scheduled model run.
  3. Do not charge usage or rate-limit allowance for runs started after deletion was acknowledged.
  4. Provide a user-visible usage ledger identifying which automation/run consumed tokens.
  5. Audit affected post-delete runs and restore any credits or quota consumed by them.
  6. Add regression tests for repeated delete/relaunch/next-trigger scenarios on Windows.

Related issues

  • #35378 — viewing a deleted automation can render a phantom scheduled-task card
  • #37445 — background activity can consume Codex usage without clear user intent

No transcript, screenshot, username, local path, company information, credentials, or private automation prompt is included.

View original on GitHub ↗