Nuke Verify

An implementation just landed — produced by a cheaper agent, another tool, or a colleague — and must be confirmed against its intent and the quality bar, then repaired until clean. Family positioning (map: references/family-map.md): audit finds everything (read-only) · review judges a change (read-only) · verify confirms an implementation matches intent and repairs it (writes) · exec executes a spec or delegated task (writes).

Core principle — asymmetric verification: the implementer is presumed cheap; validators sit one tier above it, and every defect they catch returns to a cheap fixer as an exact finding. Detection plus the fix loop is what turns a weak implementation into a strong one.

Arguments

[mode] — light (default) | full | plan. plan runs preflight, prints the plan block, and STOPS — nothing is written. [intent] — the source of truth to verify against: a spec/plan/task file path, or free text (intent: add retry with backoff to the fetch layer). [scope] — changed (vs HEAD, default) | staged | branch (vs main/master) | <path> --ask — pause at the preflight plan for confirmation; default is no gate — the plan prints and the run starts (references/preflight.md).

Mode tokens match exactly; anything else in first position is the intent.

Modes

Tier vocabulary (clerk/worker/session/top) and per-platform mechanics: references/model-tiers.md. Platforms without per-agent overrides → every role inherits the session model.