Review-loop-autopilot

Drive the gated loop to APPROVE on your own, but stop dead at outward actions. The unattended mode skips the mandatory audit — treat it as the risk-accepting exception.

Contract

Run the loop autonomously: prompt → cross-model review → orchestrator audit → fix → re-review, repeating until APPROVE, with a hard stop at every outward action (push / PR / merge / publish). The audit step is mandatory each iteration. The fully-unattended mode (do <stage> --auto, and --auto --flip) forwards a failed review straight to fix/revise with no audit step — so it is NOT protocol-compliant and is used only when that risk is explicitly accepted.

When to trigger

• You have an APPROVE target and want to converge without a human between every micro-step.
• You are choosing between driving stages one at a time and an unattended --auto run.
• You hit a push / PR / merge / publish — stop and get explicit sign-off.
• You want to set reviewer effort or an iteration cap for the run.

Protocol

1. Drive single stages: do plan, do run, etc. — one stage, with an eyeball on the result.
2. After each review, run the orchestrator audit (see orchestrator-verify) before any fix — every iteration, not just at a flip. Local rule if that skill is unavailable: for each Critical/Important finding, open the cited file:line, classify it confirmed / theoretical / false-positive, record the classification in audit-iter<n>.md, and forward only the confirmed scope to the fix.
3. Loop prompt → review → audit → fix → re-review until APPROVE; cap iterations (default 3) so a non-converging loop forces a human decision.
4. Calibrate reviewer effort to budget: high effort for a single deep review, medium for wide batches; detect usage caps by exit code, not by grepping output.
5. Switch to alternating author/reviewer only late (polish-class findings), never during architectural churn.
6. STOP at every outward action. Push / PR / merge / publish each require explicit OK, re-authorized per session for destructive forms.

Anti-patterns

• Auto-pushing or auto-merging at the end of a loop — outward actions are always gated.
• Treating --auto / --auto --flip output as if it were audited — those modes skip the mandatory verification.
• Running an uncapped loop with no iteration ceiling.
• Flipping author/reviewer during an architectural pivot instead of after convergence on small findings.

Worked micro-example

The compliant gated loop, with the hard stop:

do run <task-id>            # author implements
→ cross-model review        # reviewer returns NEEDS WORK + citations
→ orchestrator audit        # re-read each finding, classify, write audit-iter<n>.md
→ do fix <task-id>          # apply only the verified scope
→ re-review                 # APPROVE
STOP → push / open PR        # requires explicit human OK

The --auto form collapses the middle into review→fix with no audit — convenient, but it is the risk-accepting exception, not this loop.