verify-gate

name	verify-gate
description	Dispatches the independent verifier agent, enforces Reflexion x1 (one retry-with-critique per stage) and bounded-retry-then-escalate at 3 total verify loops. This is what /verify invokes and what the Stop hook ultimately points at. Auto-invocable.
model	sonnet

verify-gate — independent verification with bounded retry

This skill is the operational home of CONSTITUTION.md N3 (Reflexion ×1, bounded retry-then-escalate) and N4 (fail-closed, three verdicts only). It dispatches the verifier agent — never a generic/untyped stand-in (N7) — and manages the retry-counter bookkeeping that N3 depends on.

Procedure

1. Read the stage state. Load .fable/<run_id>/stages/<stage>.json. If it doesn't exist yet, create it with {"stage": "<stage>", "verify_loops": 0, "reflexion_used": false, "last_verdict": null, "history": []}.
2. Dispatch verifier via Task (typed agent, N7) — never let a general-purpose Task stand in. The verifier runs with no shared context with the builder, re-derives claims from disk rather than trusting self-report, and returns exactly one of pass, pass-with-notes, reject (N4). If the verifier itself errors or cannot complete, treat that as reject — "a validator that cannot validate has already failed."
3. On pass or pass-with-notes: record the verdict in stages/<stage>.json, append to history, and stop — the stage clears.
4. On reject: a. Increment verify_loops in stages/<stage>.json. b. If verify_loops has now reached 3 total verify loops, do not retry again — escalate to the human instead of looping. Report the verifier's reject reasons and the full retry history, and wait for explicit human direction. Do not silently re-attempt a 4th time under any circumstance. c. If verify_loops is still under 3 and reflexion_used is false: this is the single Reflexion retry (N3 — at most one retry-with-critique per stage). Set reflexion_used: true, bundle the verifier's critique with the original generator prompt (this is what reflexion-coach does — dispatch it rather than hand-rolling the retry prompt here), re-run the stage's producer, then loop back to step 2. d. If reflexion_used is already true and a further loop is needed before hitting the loop-3 ceiling: this can happen when a stage is re-entered later (e.g. after the human resolves an earlier escalation) — still count toward the 3-loop ceiling; do not reset reflexion_used to allow indefinite additional reflexion retries. Reflexion ×1 means one critique-guided retry per stage, full stop; the loop-3 ceiling is a separate, additional cap on total verify attempts.

Relationship to the Stop hook

The Stop hook (wired in a later phase, per plan.md's build sequence) points at this skill via a /verify invocation. It must guard with stop_hook_active to avoid recursive triggering, and it inherits this skill's same 3-loop escalation ceiling rather than defining its own — do not let the hook and this skill drift out of sync on what "3 loops" means.

Constitution citations

Refused per N3: Reflexion ×1 already used for this stage — additional failures escalate to human, not silent retry. / Refused per N4: verifier returned reject (or could not complete) — treating as hard stage failure.