[Codex desktop] Repeated authoritative-data misclassification caused excessive usage consumption

Component

Codex desktop app on Windows, working in a local project task.

What happened

During a long-running local task, the agent repeatedly mishandled an authoritative race-result marker meaning “non-starter/excluded” (출전제외). It assigned a synthetic fallback finishing rank to the excluded entry, then continued defending and reprocessing the derived result after the user had identified the official fact. This caused a prolonged correction loop and many unnecessary tool calls.

The task already contained explicit domain rules requiring official cancellations/non-starters to be excluded and requiring pre-start predictions to remain frozen after the race start.

Expected behavior

• Preserve authoritative cancellation and non-starter markers.
• Never synthesize a finishing rank for a non-starter.
• Once the official start time is reached, preserve the pre-start prediction and update only official facts/results unless the user explicitly authorizes a prediction change.
• When the user supplies an authoritative correction, stop repeating the same failing interpretation and revise the execution path immediately.
• Provide a clear account-support/escalation path when erroneous agent work materially consumes a usage window.

Impact

The repeated correction cycle materially consumed the user’s weekly Codex usage and delayed an active race-day workflow. The local project logic has since been corrected and regression-tested, but the product-level behavior and the resulting usage impact should be reviewed by the Codex team.

Privacy

No project files, private account identifiers, credentials, or private logs are included here. Please contact the issue author privately if account or session identifiers are needed for investigation.

Review the Codex desktop app on Windows against the reported race-result scenario, focusing first on handling the “출전제외” marker and post-start prediction freeze. Confirm done when authoritative non-starters remain excluded, predictions stay unchanged after start, corrections stop the loop, and a support or escalation path is defined; no repository file or test is named.