openai / openai/codex

Codex workflow regressions are consuming paid usage without delivering usable results

Open
#40,965 2 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

context enhancement model-behavior rate-limits
Dominant language
Rust
Stars
125k
Forks
19.4k
PR merge metrics
PR metrics pending

Description

I am reporting a serious workflow reliability problem with Codex that is causing direct operational loss.

I am using Codex for real production work across repositories, deployment workflows, and technical projects. Repeatedly, the agent has consumed substantial usage while taking retrograde or unnecessary actions, revisiting already-resolved approaches, creating conflicting directions, and failing to deliver a stable, usable end result.

The practical impact is not just inconvenience:

  • Paid Codex/agentic usage is being consumed on work that has to be undone or repeated.
  • I have hit usage limits during business hours while projects are still blocked.
  • Time-sensitive project deadlines are being missed because the workflow is not converging.
  • The lack of a clear boundary between advisory reasoning, execution, Codex, and other ChatGPT surfaces makes it difficult to know what is consuming usage and what environment is actually acting on the project.
  • The product can appear to be progressing while in practice it is redoing work, changing direction, or regressing previously working states.

What I need is a concrete solution, not another workaround. At minimum, Codex should provide:

  1. Clear disclosure before an agentic action consumes Codex/agent usage.
  2. A persistent canonical project/repository context so the agent does not repeatedly rediscover or contradict prior decisions.
  3. Stronger safeguards against modifying or replacing working solutions without explicit evidence that a change is necessary.
  4. Reliable before/after evidence for repo, branch, files changed, tests, deploy state, and resulting URL/output.
  5. Better visibility into usage consumption per task/session, especially when work is abandoned, reverted, or produces no usable outcome.
  6. A reliable escalation path when paid usage is materially consumed by agent regressions.

I have also contacted OpenAI Support separately regarding the account/usage impact and requested review of the consumed usage and possible compensation or restoration. This GitHub issue is specifically to document the Codex product/workflow failure publicly and make it trackable.

The core issue is simple: for production work, "the agent tried" is not a successful outcome. Paid agentic usage should converge toward a verifiable result, not repeatedly consume time and credits while re-opening settled decisions.

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

No files, tests, or entry points are identified in the issue, so begin by collecting a minimal reproducible workflow and the specific repository, branch, files, tests, and deployment state involved. Done should include verifiable before/after evidence, usage consumed, and a stable result that does not regress earlier work.

Written by the indexing model from the issue text.

Assessment

Domain
ai, devtools
Issue type
Bug
Difficulty
5/5
Estimated time
Over a week
Activity status
Active
Clarity
Needs clarification
Newbie friendliness
25/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.