Controlling a run
A run does not have to see itself through. From the Runs board or a run’s detail page you can pause it, stop it, retry a failed phase, or restart the whole pass on a fresh branch. Knowing which to reach for saves time and tokens.
#Where the controls live
Every run carries its controls in two places · the run’s card on the Runs board, and its full detail page. The controls available depend on the run’s current status, so you only ever see the actions that make sense right now.
#Pause and resume
Pause holds a running run where it is · nothing is lost and the branch stays put. Resume picks it back up from the phase it was on. Use pause when you need the agent to wait, not to abandon the work.
#Stop
Stop ends a run immediately. The run moves to Stopped · a deliberate terminal state, not a failure, that stays on the board so you can still read what it did. The branch and any pull request it opened stay exactly where they were.
Stopping is something you chose · it reads as Stopped, separate from a run that Failed on its own. A stopped run can still be restarted later.
#Retry
When a run has Failed, Retry re-runs the phase that failed · it does not redo the phases that already succeeded. The re-run is recorded as a fresh attempt of that phase, so the earlier failed attempt stays in the history. Reach for retry when one phase hit a transient problem and the brief and plan are still good.
#Restart
Restart re-runs the entire pipeline from the start · the design’s entry phase, Refine in the Starter SWE design. It discards the agent’s branch and pull request and cuts a fresh branch, so tokens already spent are not refunded. Use restart when the brief itself was wrong and you want a clean pass.
Restart is blocked while a run is live · stop it first, then restart. This stops you from cutting a new branch out from under a run that is still working.
Retry is cheap and surgical · it repeats the failed phase on the same branch. Restart is a full do-over on a new branch. When only one phase misbehaved, retry it rather than restarting.
#When a run needs you
A run in Needs input is waiting on a decision it could not make from the repository alone · answer or approve in the run panel and it resumes from where it paused. This is how a review gate or a mid-run question hands control back to you without losing the work so far.