Problem to Fix
Review outcomes today all stop and ask, regardless of whether the review found anything actionable. Reviews with no findings or only informational notes don't need human judgment to merge.
Why it Matters
The human checkpoint should fire when judgment is actually required. Stopping for clean reviews dilutes attention from real must-address findings.
General Approach
Classify review output into three tiers: none (no findings) → auto-proceed; informational (notes, no blockers) → auto-proceed but capture notes for the follow-up-spawning child; must-address (blockers or required changes) → stop and ask as today. Make the tier classification explicit in the review output so the operator can see why the skill proceeded.
Complexity
Verification / QA effort: significant
Misclassifying a must-address as informational would be a real failure. Need a clear rubric for what counts as "must-address" and explicit operator visibility into the tier decision.
Acceptance Criteria
- Review output reports its tier classification before any auto-proceed.
- Clean reviews merge without prompting.
- Informational reviews merge but surface the notes (handed off to the follow-up-spawning child).
- Must-address reviews behave exactly as today (stop, ask, iterate).
- Anything touching auth/payments/secrets/prod/destructive is forced to must-address regardless of finding count.
Parent story
Part of #167.
Problem to Fix
Review outcomes today all stop and ask, regardless of whether the review found anything actionable. Reviews with no findings or only informational notes don't need human judgment to merge.
Why it Matters
The human checkpoint should fire when judgment is actually required. Stopping for clean reviews dilutes attention from real must-address findings.
General Approach
Classify review output into three tiers: none (no findings) → auto-proceed; informational (notes, no blockers) → auto-proceed but capture notes for the follow-up-spawning child; must-address (blockers or required changes) → stop and ask as today. Make the tier classification explicit in the review output so the operator can see why the skill proceeded.
Complexity
Verification / QA effort: significant
Misclassifying a must-address as informational would be a real failure. Need a clear rubric for what counts as "must-address" and explicit operator visibility into the tier decision.
Acceptance Criteria
Parent story
Part of #167.