prevc-workflow

Use /prevc run <goal-id> only after PREVC has recorded explicit operator approval for the current plan. The run executes only the approved scope, files, permissions, and budgets, then automates Review, Execute, Validate, Judge, and Refine before stopping at awaiting_confirmation.

Risk is classified as low, medium, high, or untrusted during Prepare, informed by the harness-risk-router but not replaced by it. The router deterministically produces low, medium, or high from path globs and diff signals, while untrusted is declared only by the operator for untrusted provenance.

Auto tier reads no diff and relies on the green full-mode gate report plus Judge verdict. Sampling runs code review over sampled files only. Full tier requires complete code review, a security analyst pass, a human diff read, and a recorded review citing the gate report's sourceHash.

No. Agents and skills are bounded capabilities that cannot transition lifecycle state, approve evidence, or self-confirm completion. Only the operator may confirm, request revision, or abort from awaiting_confirmation, and this gate is mandatory in every tier.

The run stops immediately as blocked or needs_input, naming the failing metric, its value, and its threshold. A gate exit code of 2 is reported as a harness blocker rather than a code failure, and no open-ended repair loop is started.

Use it when the operator instructs execution of a named plan or spec end to end; that instruction itself is the run authorization. PREVC automates every task's cycle and transitions to awaiting_confirmation once for the whole spec, halting only for real blockers like scope changes or unrecoverable validation failures.