There is now a stronger official datapoint for this exact question.
In a separate thread about tasks stopping at the limit, @OpenAI_Support posted that the current guide says an active turn can continue after a limit, subject to fair use, and acknowledged that the user’s work had still been interrupted:
That does not prove every mid-task stop is a bug. A separate fair-use or execution boundary can still exist. But it does mean “Codex always stops immediately when the meter hits zero” is not simply the documented default behavior.
So the next natural specimen should pin the boundary down instead of debating the policy in the abstract:
- task start time;
- 5-hour and weekly meter immediately before the turn;
- exact time one of them reaches the limit;
- whether the turn was already active at that moment;
- exact terminal message / status;
- model + reasoning effort + Codex surface;
- session/thread ID;
- last known-good date when an active task continued through the boundary.
If an already-running turn is cut exactly at the allowance boundary while the official guidance still says active turns can continue, that gives Support/maintainers a very specific behavior change to classify.