Stop a run, and correct a checkpoint instead of ending the run

Saying no at a checkpoint used to end the run, and that is almost never what a no means. The checkpoints in a codified workflow are the ones you put there, at the moments you wanted to look before something went out. So a rejection is usually a correction: wrong recipient, figure from last quarter.

A rejection now goes back to the agent with your reason attached. A rejected gate is revised and asked again, with your words in front of it, and you decide again. You can go back and forth for as long as the run's engagement budget allows. An approval can also carry edited values, and what gets recorded is what you approved rather than what was proposed.

None of that counts as the workflow failing. It raises no drift event, sends no notification and touches no counter, because using a checkpoint the way it was meant to be used is not a defect.

There is also a brake. Cancel stops a live run from the screen you are on, and the confirm dialog offers to stop every live run of that workflow and pause its trigger in the same move. By then you have usually found a flaw in the workflow itself, and cancelling runs one at a time is a losing race against the trigger.

Cancelling says nothing about whether the workflow works: no failure review, no demotion counter, no streak. It has one honest limit, which the dialog states. The brake bounds what a run can still record. A call it already made to the outside world has already happened.