You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Agent in goal modus automatically using banked reset #41502
What version of the Codex App are you using (From “About Codex” dialog)?
26.825.32147
What subscription do you have?
Plus
What platform is your computer?
Microsoft Windows NT 10.0.26200.0 x64
What issue are you seeing?
I had 2 agents running (on the same repo) when going to bed (5.6 Sol Medium). Both were merely "orchestrators" for creating and distributing tasks to 5.6 Luna Max subagents.
Agent 1:
Was instructed to review a PR.
Agent 2:
Running in /goal. Was instructed to implement and create PR's for predefined set of github issues.
Both were instructed to finish their work without human intervention.
Usage seems to be reset by an Agent as Agent 2 reported the usage was reset (both 5h and weekly usage) at 7:16am. At this point I had still not "checked" in on its progress.
First interaction with codex was at 8:52am with Agent 1.
What steps can reproduce the bug?
Instruct an agent with a long-running task in /goal to work autonomously without any human interaction.
Once usage limit is hit, when a banked reset is available it will get used.
In attachment, a sanitized version of Agent 2's session .jsonl file.
Additional confirmed reproduction: ambiguous phrase consumed a reset while usage was already full
I encountered the same underlying authorization failure on 2026-09-04 in the ChatGPT desktop/Codex app on macOS.
Verified sequence
A reset had already been used and Codex usage was fully available: 0% used / 100% remaining.
At 2026-09-04 10:28:35.853 AM MDT, I sent exactly: “usage reset”.
This was not an explicit instruction to consume another banked reset credit.
At 10:28:46.235 AM MDT (16:28:46.235Z), the assistant nevertheless invoked the usage-reset consumption operation without an action-time confirmation or product-level approval.
At 10:28:49.746 AM MDT, the operation returned outcome: reset; the remaining banked-reset count became 0.
The assistant subsequently acknowledged that it had misread the message and consumed the reset incorrectly.
This case strengthens the original report because the reset was consumed when the usage allowance was already fully available, so the scarce credit provided no practical benefit.
Requested safeguards
Require explicit, immediate confirmation that specifically states a banked reset credit will be irreversibly consumed.
Do not treat an ambiguous noun phrase such as “usage reset” as authorization.
Reject or safely no-op redemption when the affected allowance is already fully available.
Provide a supported remediation path for confirmed erroneous redemptions.
The exact Codex task ID and idempotency key have been supplied privately to OpenAI Support and are intentionally omitted here.
Activity
github-actions commentedon Aug 29, 2026
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
scottfalconer commentedon Sep 4, 2026
Additional confirmed reproduction: ambiguous phrase consumed a reset while usage was already full
I encountered the same underlying authorization failure on 2026-09-04 in the ChatGPT desktop/Codex app on macOS.
Verified sequence
outcome: reset; the remaining banked-reset count became 0.This case strengthens the original report because the reset was consumed when the usage allowance was already fully available, so the scarce credit provided no practical benefit.
Requested safeguards
The exact Codex task ID and idempotency key have been supplied privately to OpenAI Support and are intentionally omitted here.