Retry a failed run

Start another attempt in an existing failed run after fixing its cause.

Prerequisites

A failed run and project operator access or a higher role. Inspect its failure and worker preflight before retrying.

Procedure

1. Open the failed run from Needs attention or its task’s run list.
2. Read the attempt outcome. Fix infrastructure, harness, or credential failures before continuing. A question or blocker needs an operator answer.
3. Select Retry to retain the effective task and agent. To change either, select Edit and retry, edit the task or choose a rostered agent, and select Save and retry. Selecting the same agent uses its current definition.
4. If Cycle budget exhausted prevents new work on the item, inspect the admitted rounds in Needs attention before selecting Reset cycle budget.

Check the result

A new attempt appears in the same run. The event stays unchanged. Edited task and agent definitions apply to all later attempts until another edit. A salvage branch, when available, supplies the prior attempt’s work.

The engine automatically retries infrastructure failures up to three attempts per run. A CI wait that times out or ends with partly run CI also records an infrastructure failure, but isn’t retried automatically: rerun or fix the CI, then retry the run. When CI didn’t finish and a rerun on the same head passes, select Apply CI labels in Needs attention to apply the wait’s completion labels without another attempt; see Review and fix loops. Agent failures, internal encoding defects, runaway limits, child-session budget failures, questions, and blockers require operator action. A retry uses the existing run’s cycle-budget reservation.

If the parent task cannot be confirmed stopped, its checkout and worker slot remain reserved. Wait for reconciliation or restore access to the worker; starting duplicate work does not release that slot.