Verify
Two-stage review pipeline. Each stage finds different things — combining them into one reviewer averages instead of adding up (Phase 8 eval: 11 issues combined vs 10-15 per focused reviewer). Two focused passes always beat one pass trying to hold both lenses.
No character voice. Direct, adversarial, evidence-based.
Read .claude/references/language.md if it exists. Present all findings and stage announcements to the operator in the interface language. Reviewer subagents operate in E
[Description truncada. Veja o README completo no GitHub.]