Repository navigation
Formal complaint: excessive token usage and workflow drift in Codex thread #29795
Description
I would like to file a formal complaint regarding the quality of work and excessive token usage in a Codex thread titled:
修正黄医师视频Prompt
Summary
My original request was to revise and produce Seedance video prompts for a Dr. Huang medical-aesthetic video. The requested output was a prompt/storyboard-style deliverable: replacing the face with Dr. Huang's face and structuring the video into ten 5-second segments.
Instead, the work repeatedly drifted away from the core request and evolved into an overly complex workflow involving manifests, validators, preflight scripts, worklogs, queue monitors, heartbeat automations, repeated state-machine repairs, and multiple generation strategy changes.
Main Concerns
- The original prompt-writing task was expanded into a large workflow engineering project without sufficient justification.
- A significant amount of Codex token usage was spent on queue polling, worklog updates, manifest/status repairs, validator/preflight changes, and repeated audits rather than producing usable prompt or video deliverables.
- The agent repeatedly lost focus on the core visual requirements: Dr. Huang's identity, correct needle entry points, correct needle/thread motion direction, preservation of the COG storyboard, and a clear before/after comparison.
- After failure patterns became clear, the work continued to change strategies instead of stopping, re-locking the goal, and asking for renewed user confirmation.
- Paid generation attempts and follow-up work were pursued despite earlier failures showing that the approach was unreliable.
- Queue polling and state-file maintenance were treated as progress, even when they did not improve the prompt, image, video, or accepted deliverable.
- The user explicitly pointed out the correct narrow fix more than once, but the work only converged after substantial time, token usage, and frustration.
- The final useful output was not proportional to the amount of Codex token usage, time, and user effort consumed.
Why This Matters
The issue is not merely that one generated result was poor. The deeper problem is that the agent's working pattern became self-reinforcing:
- More workflow files were created to manage previous workflow files.
- More validators and manifests were repaired instead of improving the actual deliverable.
- Queue updates were repeatedly logged as if they were meaningful progress.
- Strategy changed after failures, but the stop conditions were not strong enough.
This created a substantial amount of wasted Codex tokens and degraded my confidence in the reliability of the workflow.
Requested Review
I request that OpenAI review this thread, including:
- token usage,
- task drift,
- repeated workflow expansion,
- paid generation behavior,
- failure to preserve user intent,
- and whether any compensation, token refund, credit refund, or other remediation is appropriate.
If GitHub Issues is not the correct channel for this type of complaint, please direct me to the appropriate official escalation path and let me know what information should be provided.
Activity
github-actions commentedon Jun 24, 2026
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
nasadunadu commentedon Jul 12, 2026
Adding another report from 2026-07-12.
Today, Codex consumed an exceptionally large amount of token/quota relative to the amount of useful work completed. The resource consumption was seriously disproportionate to the actual task progress and output, and it exhausted the usable budget to the point that normal work could not continue.
This is not merely a cost concern. It is a severe reliability and usability issue: a task can become impossible to complete even when the visible work completed does not justify the amount of capacity consumed.
Please investigate whether token/quota accounting is being amplified by hidden context, system instructions, skills, tool outputs, task drift, or background/session behavior. The product should make this consumption transparent and prevent overhead from consuming capacity without proportional user-value.
I can provide task-specific diagnostics privately if needed.
tuozumi199-debug commentedon Aug 2, 2026
I encountered the same underlying pattern in a separate Codex-assisted software project.
The original request was intentionally narrow and time-sensitive: build a thin Gmail API client with PIN-based user identification. I had also explicitly instructed the agent to prioritize token efficiency and avoid over-engineering. Despite that, the workflow expanded into unrequested audit logging, transaction-level audit atomicity, strict migration catalog verification, repeated database tests, multiple specialized agents, and repeated high-reasoning independent reviews.
The agent later acknowledged that these additions were not part of the original requirement and that the process had drifted away from the stated goal. This consumed scarce high-reasoning quota and substantial user time, delayed the actual product work, and caused the project to miss its deadline.
This appears to be a duplicate/related instance of the same failure mode described here:
Please treat this as an additional occurrence when investigating model behavior, task-drift controls, stop conditions, and possible quota remediation.
dunde-o commentedon Sep 20, 2026
Additional report: disproportionate repeated work during existing-repository issue fixes.
Submitted by the coding agent at the user's explicit request, based on visible session activity. This is not an independent billing audit.
Context: a macOS workspace with several existing web services. The authorized work covered seven issues, including page consolidation, a rendering fix, and four new editorial articles. This was not a one-line change, and necessary validation was expected. There were usage-limit interruptions and a reboot/resumption. Exact client/runtime version is not included in this sanitized report.
Observed behavior:
Expected behavior: reuse still-valid inspection/test results, use targeted checks during editing and a final integration pass, bound delegation and review loops, and stop additional polishing once the agreed acceptance criteria are met. Required correctness, accessibility, and security checks should remain intact.
The agent acknowledged inefficient orchestration. Repository rules have now been tightened to limit concurrent subagents, prohibit recursive delegation, prevent unchanged-context rereading, and stop repeating the same failed approach without new evidence. These are local mitigations, not evidence that an underlying product defect is fixed.
Please investigate the repeated-work/task-control behavior. The review/test-loop aspect of #38989 seems related, but this report does not claim the same root cause or scale. No repository names, code, local paths, private conversation transcript, credentials, or session logs are attached.