Run the gate

Run the Gate

The in-session author cannot review the in-session author. A spec written this session carries this session's blind spots straight into the implementation. The only fix is a reviewer with zero session history — it starts from the code, not from your reasoning, so it structurally catches what your own checklist cannot.

This skill replaces /codex (retired) with a fresh-context Opus subagent as that reviewer.

When this is mandatory

Run BEFORE writing code when the change touches ANY of:

• BalanceCalculator / money math / MoneySerializer
• security/firestore.rules or Cloud Functions auth/validation
• routing — app_router.dart, route tree, deep links, back guards
• a schema / field-name change with BOTH a read-path and a write-path

Outside these (a one-sentence diff, no money/route/schema/rules surface): skip the Gate, just do it. Don't burn a review round on a typo.

The loop

1. Write the spec to a file — docs/plans/<YYYY-MM-DD>-<topic>.md. While writing, run the 7 verification principles yourself and report results out loud. The embedded check is the reviewer's rubric, not a substitute — but a first pass narrows what the reviewer hunts for.

2. Spawn TWO reviewers in parallel — both fresh Opus subagents with zero history, launched in a single message so they run concurrently:

   • Rubric reviewer — prompt: the full contents of reviewer-prompt.md (this skill's dir) + the absolute path to the spec file. Verifies the spec's own claims against live code.
   • Orthogonal-axis adversary — prompt: the full contents of adversary-prompt.md + the same spec path. Hunts what an on-the-map review structurally misses: regressions on coupled axes the spec doesn't cover, and derived surfaces (share cards, exports, activity rows, l10n pairs) the change silently reaches. Why two: #857 — a typography spec passed TWO single-reviewer rounds while a share-card cover-band caption needing AR translation slipped through; the builder caught it, not the Gate. A same-rubric reviewer re-walks the author's map; the adversary walks off it.
   • Both: Agent, subagent_type: general-purpose, model: opus. Each reads ONLY the spec + live code — neither sees this conversation NOR the other's output. That double isolation is the point; do not paste your reasoning or one verdict into the other's prompt.

3. Apply the union of [P1]s, re-spawn. Rewrite the spec to resolve every [P1] from BOTH reviewers. Then spawn a new pair — fresh Agent calls, NEVER SendMessage to continue an old one. Continuation re-imports the context you're trying to escape; each round must start from zero.

4. Stop when a round's union has no [P1]s — both the rubric reviewer AND the adversary must come back clean in the SAME round. ~2 rounds is typical. A 3rd round means the spec was over-scoped — split it and gate the pieces.

Non-negotiables

• One reviewer = one fresh Agent call. Round 2's pair must not know round 1 existed, and neither reviewer in a round sees the other's verdict.
• A round is clean only when BOTH verdicts have zero [P1]s — a clean rubric verdict alone does not end the loop.
• Pass the spec's file path and make the agent read it — review the spec against code, not your summary of the spec.
• A green test suite is NOT the Gate (worked example #185 in docs/SPEC-VERIFICATION.md: rules suite was green while the client write violated the rule). "Tests pass" never ends a round.
• The reviewer's findings are merge/spec claims, not gospel — but [P1]s block implementation until resolved or explicitly refuted against code.
• Convergence pressure — momentum, prior approval, sycophancy — is exactly what this interrupts. Treat each round as v1.

Full reasoning + scar-tissue worked examples: docs/SPEC-VERIFICATION.md. Trigger rationale: The Gate in CLAUDE.md.