Skip to main content
Change a run while it goes, stop it, and rerun the tasks that didn’t finish. Every command is in the valkyrie run reference. To do the same from Python, see Manage runs with Python. In the commands on this page, replace RUN_ID with the run’s ID, which valkyrie run start prints.

​
Change concurrency

Raising the limit starts more tasks. Lowering it lets running tasks finish and holds back new ones until the run is under the limit. Only a run in progress can be updated.

​
Stop a run

The Valkyrie CLI asks you to confirm. Tasks that haven’t started, and tasks being set up or graded, stop. Agents that are already running finish their task. The run shows STOPPING until they do, then STOPPED. To stop only some tasks and let the rest of the run go on, name them:
--force also stops running agents and deletes the sandboxes:
A forced stop deletes the run’s sandboxes, or the named tasks’ sandboxes, with whatever they hold. You can’t undo it.

​
Shed load when a model hits rate limits

Lowering concurrency waits for running tasks to finish. run shed lowers it and stops the most recently started tasks at once, which have done the least work, until the run is at the new limit:
The CLI lists the tasks to stop and asks you to confirm; --dry-run only lists them. The new limit must be lower than the current one. Tasks being graded count toward the limit but are never stopped. After the rate limits clear, rerun the stopped tasks with valkyrie run retry. They start over.

​
Resume or retry

  • On a finished or stopped run, resume reruns tasks in STOPPED, and retry reruns tasks in STOPPED or ERROR.
  • With --task-ids or --task-ids-file on a finished or stopped run, retry reruns exactly those tasks, whatever their status, and resume reruns them along with the stopped tasks. A dataset task the run never had is added to it.
  • On a run in progress, retry reruns failed and stopped tasks, and resume only applies a new --concurrency.
  • A run that is STOPPING can’t be resumed or retried until it stops.
These options change what the rerun uses: A rerun that restarts the agent gives the task a fresh wall limit.