Explicit opt-in
Autopilot runs only when you explicitly request it. The normal Blueprint workflow pauses for human review after the spec and each implementation step.
One bounded pass
Autopilot can select or resume one feature or fix, write and critique the spec,
create the branch, implement small steps, run checks, and create passing
checkpoint commits. When AGENTS.md declares a Verify command, Autopilot uses
that exact shared recipe as its automated gate.
It applies qualityGates.regular from blueprint/config.json. Audit, behavioral
check, and manual try-guide automation all default to manual; conditional and
always modes opt into the matching gate. When an audit runs, Autopilot
validates findings, repairs confirmed high-severity issues that stay within the
approved scope, reruns affected checks, and ends with a review packet. Broader
project cleanup remains a separate audit followed by focused fix work.
See Project Configuration for the gate modes.
Hard boundaries
Autopilot never completes, merges, pushes, deploys, publishes, sends, deletes data, runs irreversible migrations, or performs destructive actions without the separate approval required for that action.
Autopilot does not run rollback specs. Reversing completed work must use the
reviewed /rollback and /implement path so dependency and conflict gates remain
visible.
It stops when product decisions are missing, the working tree is unsafe, a required check or repair fails repeatedly, a confirmed high-severity finding cannot be fixed safely within scope, or continuing would cross one of those boundaries.
Next step
Review the final packet, diff, audit findings, and verification evidence. Run try if you want a manual walkthrough, then complete when you are ready to archive and merge.