principle-explicit-intentlisted

# Explicit Intent An irreversible act — a merge, a push over published history, a public close, a deletion — fires only on the user's stated intent, never inferred from state. A PR merely being green is not ship intent; a branch merely being behind is not rebase intent; a PR merely being stale is not abandon intent. **Why:** No verification step can undo an irreversible act after the fact. The invocation is the authorization, so it must be deliberate, and its scope must match exactly what it authorizes. **Pattern:** - Granularity matches irreversibility: one yes per irreversible mutation. An approval covers exactly the items it names — a request naming a set covers that set, and a yes to one item never silently extends to an item it did not name. Approving an adjacent class of change never carries the irreversible one. - The grant is scoped and complete: authorization to act is authorization to finish the verified act — and nothing beyond what was stated. - Spend granted authorization; do not re-ask it. Once intent is stated and the gates pass, the run completes without stopping to re-confirm. Confirmation churn erodes the signal a real confirmation carries, so even the churn is bounded and reported. - Guard the entry: a skill whose invocation itself authorizes a side effect MUST state an explicit-intent guard in its description ("Invoke ONLY on … never infer …"); the strictest ones also set `disable-model-invocation` where the host honors it. A skill