Codex repeatedly wastes tokens and user time on preventable setup mistakes
What version of the Codex App are you using (From “About Codex” dialog)?
01a00ddb-00b4-7092-9a3a-1e8a225a3a70
What subscription do you have?
01a00ddb-00b4-7092-9a3a-1e8a225a3a70
What platform is your computer?
_No response_
What issue are you seeing?
While working on a local project, Codex repeatedly provided incorrect instructions for running verification. These failures were not caused by the project code. They occurred because the agent did not inspect the actual repository and environment state before taking action.
Specifically, Codex:
modified a handoff file before committing it, even though the verification gate required it to match HEAD exactly;
failed to account for the current workflow test and an outdated branch name;
instructed me to use a .venv without first checking whether it existed;
attempted Git operations without confirming that its sandbox could write to .git;
repeatedly asked me to rerun a lengthy full verification suite after failures caused by its own preparation mistakes rather than code failures.
As a result, a substantial number of tokens were spent diagnosing and correcting problems created by the agent itself. I also wasted time running unnecessary verification attempts.
The most frustrating issue is that this happens repeatedly: Codex acts on assumptions, encounters a predictable failure, consumes more tokens analyzing it, and only then checks the facts it should have verified beforehand.
Expected behavior:
Before issuing commands, the agent should verify file existence, the available Python interpreter, the current branch and HEAD, Git status, and sandbox restrictions.
Before asking the user to run a full verification suite, it should complete every available fast preflight check.
It should not ask the user to repeat a lengthy verification run when the previous failure was caused by the agent’s own setup mistake.
If Git operations may be blocked, it should determine this before modifying files or proposing a workflow that depends on committing.
Codex should account for token cost and avoid cycles of “assumption → failure → lengthy diagnosis → another assumption.”
Please improve Codex’s preflight reliability and its ability to validate instructions against the actual environment before acting. At present, it regularly wastes tokens and user time on preventable mistakes of its own.
What steps can reproduce the bug?
Feedback ID: 01a00ddb-00b4-7092-9a3a-1e8a225a3a70
What is the expected behavior?
_No response_
Additional information
_No response_