[Quota compensation requested] Codex Goals continued for 5 hours after acknowledged pause (2,466 turns) #46904

Summary

I explicitly asked Codex to pause an ongoing task. Codex acknowledged that work and its running adviser task had stopped and said it would wait for my explicit instruction to resume. Nevertheless, the active Goal kept automatically starting new model turns for approximately five more hours. Most responses were empty, so the continuing usage was not represented by useful visible output.

This report concerns continued model invocation and quota consumption after a natural-language pause request. It does not claim that I clicked the app's Stop/Pause control at the initial pause, or that the agent kept editing project files throughout the loop.

Compensation requested

I am explicitly requesting compensation for the paid allowance consumed after I told Codex to pause. Please restore all quota attributable to the unintended post-pause Goal continuations. If restoring quota is not possible, provide equivalent usage credits or a refund for that wasted usage.

I should not lose paid allowance to thousands of automatic turns after Codex has acknowledged my instruction to stop. Please audit the session, quantify the attributable loss, and provide a concrete compensation resolution. If account verification or billing review requires a private support channel, please route this complaint there and provide the escalation reference.

Environment

• Component: Codex Desktop, connected to a remote Linux workspace over SSH.
• Session metadata: originator: Codex Desktop, bundled runtime cli_version: 0.154.0, model_provider: openai.
• Model and reasoning effort recorded in the session: gpt-6-astra, max.
• Goal mode was active; an adviser subagent had been running before the pause.
• Subscription: user-reported 20x plan. The exact plan label has not been independently verified.
• Desktop UI build and client OS: not yet collected. Runtime version above is not asserted to be the desktop UI version.
• Session ID: 01a0bed6-e081-7970-a59e-110f74e3eabc.

Observed timeline

Local time below is Asia/Hong_Kong (UTC+08:00), 2026-09-21.

No further human message appears in the parent-session transcript after the pause request. There are 2,466 internally generated Goal continuation inputs and 2,466 new turn starts after that request. Including the turn already in progress when the user paused, there are 2,467 final responses afterward, of which 2,461 are empty.

Elapsed time from the pause request to the last observed completion: approximately 5 hours 1 minute. The initial acknowledgment turn completes at 18:09:51.595Z and the next turn starts only 9 milliseconds later.

Usage and user impact

The account-wide weekly usage readings in the transcript rose from 58% immediately before the pause to 98% at the end, with the same 10,080-minute window and reset timestamp. This is a 40-percentage-point increase during the post-pause period. Account-wide readings cannot prove that this single task caused the entire increase.

I estimate that approximately 50% of my 20x allowance was consumed overnight. The exact attributable loss must be checked against account-side records. My request is to restore the allowance wasted by this incident or provide equivalent compensation, not merely to investigate the bug.

Deltas in the parent transcript's event_msg.token_count.info.total_token_usage counters, from 18:09:20.055Z to 23:10:20.028Z:

These are local cumulative-counter deltas, not a billing statement or a charge estimate. Cached input must not be added again to input. Compaction/accounting counters and billing may use different definitions; the figures should be reconciled against server-side records. Child-agent usage has not been independently audited.

Expected behavior

• An explicit pause request should suspend automatic Goal-driven model invocation until the user explicitly resumes it.
• A paused task should not consume inference quota merely to repeatedly acknowledge that it is paused or produce empty responses.
• If conversational pause cannot currently suspend the Goal scheduler, the app should make that limitation unmistakable and provide an immediate way to stop continued inference.
• Repeated empty/no-progress completions should not create an unbounded rapid continuation loop.

Actual behavior

The assistant stopped useful work but the Goal mechanism kept invoking the model thousands of times. An acknowledgment that work was paused did not correspond to a paused Goal state. The log first records that state change about five hours later.

The evidence points to a mismatch between conversational pause handling and Goal continuation scheduling. The exact implementation fault is not yet established.

Reproduction context

Observed sequence, not a newly executed reproduction:

1. Have an active long-running Goal in Codex Desktop.
2. While work is active, send a natural-language request to pause.
3. Receive an acknowledgment that work is paused and explicit resumption is required.
4. Do not send a resume message.
5. Observe new Goal continuation turns, mostly empty final responses, and increasing usage.

I have not deliberately rerun this scenario because it could waste additional quota. This is not a claim that the UI Stop/Pause button itself is broken.

Evidence and requested investigation

The original session is preserved locally and was restored from accidental archival without sending any instruction to resume it. A sanitized extract records timestamps, line numbers, turn counts, pause acknowledgments, usage checkpoints, and the final paused-state event. Full project contents, local account paths, credentials, and model reasoning are excluded from that extract.

Please fix the Goal continuation behavior after user stop/pause intent, add protection against repeated empty/no-progress turns, and compensate the allowance wasted by this incident. I am requesting both a technical resolution and a concrete quota/credit/refund resolution. Full session sharing and any necessary account verification can be handled privately instead of publishing the raw transcript.