Converge reviews

I reuse valid review evidence, invalidate only changed coverage, and stop a stable scope after three review/fix rounds. I find this cap is what breaks the loop. Without it, every round finds something new to say about unchanged code and the review never lands. This skill coordinates review state. It does not review code, apply fixes, or create external issues.

1. Accept one review packet

I require all of these before I start.

• Originating request or ticket.
• Review base and current head, or the equivalent local baseline.
• Exact scope paths and current content/diff hash.
• Reviewer roster and lenses.
• Findings with stable IDs, severity, source reviewer, evidence, and disposition.
• Any prior convergence ledger.

When the caller already owns richer finding state, I reference its artifact and finding IDs instead of copying finding bodies. The review-pr finding-state file remains authoritative for PR finding status.

Gate. Every reviewer is accounted for and every finding is traceable to one scope, reviewer, and disposition.