Scan the Skill, Govern the Action: Composing Registry Verdicts with Runtime Consequence Control

RESUMO

Agent skill registries screen what they publish, but scanners overlap on at most 10.4% of combined positives. We report three measurements over 66,192 public ClawHub skill versions: 705 skills across 135 publishers rated clean by every scanner nonetheless instruct a prohibited action; of 144 live-executed commands, 34.7% carried a consequence class absent from documentation; and a deterministic gate stopped all 23 cases where an agent reached for an undocumented action among 53 cleared skills.