Goal auto-continuation can enter an unbounded no-progress loop and generate thousands of duplicate turns

Summary

Codex Goal auto-continuation entered a self-sustaining loop after the active task began waiting on an external durable process. No new user messages were sent, but each assistant task_complete was followed by a new task_started 5–8 ms later and another <codex_internal_context source="goal"> injection.

Impact

• 7,480 Goal continuation turns over 10:37:57.
• 4,857 identical visible no-op replies over 6:01:55.
• Mean interval during the duplicate phase: 4.472 seconds.
• Local Goal usage counter increased by 26,321,578 during the full window.
• A separately configured heartbeat ran every 10 minutes and was not the tight-loop trigger.

Environment

• Codex Windows App 26.715.3651.0
• Session CLI 0.144.2
• Windows 11 10.0.26100
• GPT-5.6 Sol, xhigh

Expected behavior

Waiting/no-progress turns should pause or back off. Goal should detect duplicate output, cap consecutive automatic continuations, and trigger a usage anomaly circuit breaker.

Actual behavior

The scheduler immediately requeued the same Goal after every no-progress completion, producing thousands of visible duplicate replies and substantial usage impact.

Requested fix

Add no-progress detection, deduplication, exponential backoff, a continuation/time/usage cap, and explicit confirmation before ordinary continuation language is converted into persistent Goal mode.

The full private session and account-impact evidence have been submitted through OpenAI Support. The private evidence is intentionally not included in this public issue.