Desktop goal run continued for hours and consumed millions of tokens without a progress cutoff

[rickgude]

Summary

A long-running Codex Desktop goal continued autonomously for several hours after progress had clearly diminished. The user observed roughly seven hours of wall-clock work. The goal telemetry at completion reported 8,719,248 tokens and 16,622 seconds (about 4h37m) inside the tracked goal alone.

What happened

• The task involved debugging and testing a native C++ plugin with local build and GUI-runtime probes.
• The agent repeatedly iterated through diagnosis, edits, builds, and runtime tests inside one broad goal.
• It did not establish or enforce a time/token checkpoint.
• It continued across context compactions and revisited hypotheses instead of stopping to report diminishing progress.
• The long run did not produce a proportional improvement in quality; the user experienced it as slower and less reliable than prior behavior.
• The final response exposed the unexpectedly large token count, confirming that the run had become excessive.

Expected behavior

For long local-agent goals, Codex Desktop should:

1. Detect excessive token/time use or repeated low-progress loops.
2. Pause at a bounded checkpoint and provide a concrete status report.
3. Ask the user before continuing a materially longer investigation.
4. Prefer a narrow reproducible probe over repeated broad rebuild/runtime cycles.
5. Avoid silently continuing through multiple compactions when coherence is degrading.

Suggested safeguards

• User-visible elapsed-time and token-budget warnings.
• Automatic checkpointing after a configurable duration or repeated failed iterations.
• A non-progress detector for repeated build/test loops.
• A hard or soft goal budget that requires explicit user confirmation to extend.
• A concise continuation summary that preserves the current hypothesis and prevents repeated work after compaction.

Environment

• Codex Desktop on Windows
• Local repository workflow with shell commands, native builds, and GUI-runtime tests

No proprietary project paths or source details are included here.

[github-actions]

Potential duplicates detected. Please review them and close your issue if it is a duplicate.

• Long-running Codex agent lacked a no-progress watchdog after false derived holds (5.345B token-events) #36056
• Long-running tasks keep expanding work instead of converging or stopping #35892
• [Windows][Regression] Context compaction loses task state, repeats completed work, and exhausts weekly usage #35935
• Context auto-compaction loop repeatedly rereads files, loses progress, and consumes paid Codex credits #35226
• Codex 5.6 Sol doing simple task before I left to work came home 8hrs later and huge chunk of my usage was wasted on a error loop #35165

Powered by Codex Action