Stay for the gates. Then you can walk away.
Install, /setup, one real goal, sit through every human stop. After that the committed policy is what a teammate inherits.
First hour on the mill, then the settings that let a team walk away without a silent merge.
Guide — first hour on the mill
Install Loomwright
Use the app you actually type in. Open the install steps. Do not paste /plugin into zsh — that is a chat command, not a file path.
See what is wired
/setup is a dashboard, not a wizard that overwrites you. Check observability, notifications, Twin, Beads, then turn on only what you want.
/setup
Pick a lane
One new goal: /autonomous. An existing PR: /review-pr. A pile of stories: /automate. Do not start all three at once.
/autonomous "add a daily challenge that resets at midnight UTC"
Stay for the gates
This is foreground-assisted, not fire-and-forget. Save or refine the brief. Answer rubric gates. Desktop banners fire when the mill needs you.
You merge
A healed PR is still an open PR. Read the diff. Merge it. Then run /dreaming after a week so the mill keeps the lessons you actually want.
/review-pr https://github.com/you/repo/pull/42 # bounded review → fix → re-review. Never merges.
Run it across a team
One person can hold a mill in their head. Four cannot. So the settings that matter are committed, the dangerous ones are off, and the ones that cost money are a flag you passed on purpose.
Write your policy down
Flags live in one person’s shell history. .supervisor/config.json is committed, reviewed, and applies to every run in the repo, including the unattended ones. Toggle the knobs to build yours.
Commit this · .supervisor/config.json
{}
Pass this · per run
/autonomous "…"
The one place a merge can happen
Every other surface — /supervisor, /review-pr, the heal loop — terminates with the PR still open. Auto-merge exists in exactly one place, is off unless you pass --auto-merge, and needs all five of these to hold. Anything it cannot read counts as a failure.
What leaves the building
The usual first question from anyone who has to sign off on this. The mill reads your repo, your history, and your logs from where they already are. Three things can go outward, and all three are off until someone turns them on.
A rollout that survives contact
Do not hand a team /automate --auto-merge in week one. Earn each rung.
Prove it on your own branch
One goal, watched end to end
Run /setup, then one /autonomous on real but unglamorous work. Sit through every gate. You are calibrating how good the brief has to be, not measuring throughput.
Commit the policy
Rules and config, in the repo
Seed /rules so a worker reads your conventions while writing, not after review catches them. Commit .supervisor/config.json. Set the real --base-branch. Now a second person gets your setup by cloning.
Make the trunk defensible
Protection, review, and a scoreboard
Turn on branch protection with a required check. Turn on red_team_high_risk if you touch auth, money, or migrations. Run /insights weekly and /pr-postmortem on the PRs that took four review rounds.
And only where it is boring
Queues, and maybe a trusted merge
/automate on a backlog, parked at the gate for a human. Add --auto-merge only on a path where a wrong merge is cheap to revert: dependency bumps, copy changes, generated files. Never on the surface you would not let a new hire merge alone.
Managing it once it is running
A mill nobody inspects turns into a mill nobody trusts. These four are the operating loop.
Weekly · is it getting better
/insights
Local scoreboard: heal rate, review rounds, where sessions burn time. Rendered from your own logs, on your own machine. If the trend is flat, the brief is usually the problem, not the model.
When a PR churns
/pr-postmortem
Sorts every review round into a root cause and attributes it to a stage. Four rounds of the same nit means a missing house rule, not a bad reviewer.
When someone new arrives
/handoff
Decision, why, what was tried and rejected, current state, provenance, each with its own freshness. Two minutes to inherit context that would otherwise live in one person’s head.
Monthly · keep what worked
/dreaming + /rules
Distil sessions into proposed lessons and harvested conventions, delivered as a pull request your team reviews like any other. Accept item by item. Nothing writes itself in.
Roadmap — trust is earned, not scheduled
Shipped
The mill is live
Plan-first Launch Pad, parallel Supervisor, review-and-heal, stacked autonomous loops, automate queues, advisory Twin, dreaming, insights, house rules. You still merge.
In the warp
Ground-truth verification
Run the software in the heal pass. Playwright for web apps. Headless loops against a corpus. A hard pass/fail — not a vibes score — before anything is allowed to gate.
Evidence-gated
Twin starts enforcing
Flip advisory conformance to a real gate only after the benchmark has caught real regressions without a false-positive habit. No calendar flip.
Four guardrails first
A proactive mill
A watcher that can open PRs unprompted — cadence expiry, per-run budget, circuit-breaker, heartbeat. Not started. Highest-risk rung. Intentionally last.
Director model
Point at the repo. Own this.
You set intent and approve outcomes. Diff-review becomes optional because the mill already proved the behavior. Everyone else rents a coder who starts cold every morning.
