Retry, Re-run, Re-verify, and Cancel

Choose the operation that preserves the right identity and history.

On this page

Choose the correct operation

Operation	Scope	What it creates	Use when
Retry	One failed rollout	The same rollout reopens; its next execution creates a new Attempt	Execution should try that exact Task slot again.
Re-run	A prior selection	A new Experiment and new Runs	You want a fresh evaluation plan or changed settings.
Re-verify	One Verifier on one sealed Attempt	A separate verification run for that Verifier	Grade the same sealed evidence again without rerunning the agent.
Cancel	A Run or Experiment	Cancellation intent; terminal status after in-flight work settles	No new rollout work should be claimed.

Automatic retry happens before manual recovery

Runtime automatically retries failures classified as retryable transient or infrastructure failures while the rollout's attempt budget remains. The default permits two total Attempts — one automatic retry — with exponential backoff starting at two seconds and capped at sixty seconds. Every Attempt remains in the rollout lineage. If an eligible rollout is still failed afterward, use the manual operator command below.

Retry one rollout

Terminal

fleet-runtime runs retry \
  run-... "task-key#0" --target fleet-alpha

Manual retry is currently a CLI operation. It is allowed only for a persisted failed rollout whose recorded failure is retryable. It reopens that rollout and may reopen a finalized or failed Run; the next execution appends a new Attempt. Earlier Attempt IDs, evidence, and Verifier outcomes remain immutable history. A cancelled Run cannot be reopened by retry.

Re-run from Registry

Re-run opens a new Experiment form prefilled from the prior Run: its immutable TaskSet, Task selection, rollout count, available results project, and agent identity or model. Review every field—especially agent options and credentials—before creating the new Experiment. The prior Run and Attempt history remain unchanged.

Re-verify sealed evidence

Terminal

fleet-runtime verify attempt-... \
  --verifier-id verifier-id --target fleet-alpha

Re-verification creates a separate immutable verification run. It never changes the original verdict or lets later live Environment state enter an already sealed Attempt. With --target, Fleet materializes remote sealed evidence and grades and stores the new verification locally; it does not update the remote Run.

In Registry, open the rollout's Attempt, open Details on mobile, select Outcome, and choose Re-verify on the specific Verifier. Each action reruns only that Verifier, not every Verifier on the Attempt. Re-verify is unavailable when Registry cannot load the exact Verifier document or the Attempt has reached its 100-rerun limit.

Cancel pending work

Terminal

fleet-runtime runs cancel run-... --target fleet-alpha

Cancellation stops new claims and signals active holders, but it does not interrupt an agent task midway through its active Attempt. A result that settles during cancellation remains recorded; cancellation never erases sealed history. Cancelling an Experiment fans out to its member Runs and reports a separate outcome for each member.