/review

name: review
description: "Validate plans, execution, or PRs against wish criteria — returns SHIP / FIX-FIRST / BLOCKED with severity-tagged gaps."

review — Universal Review Gate

**Runtime syntax:** in Codex, invoke the plugin copy with the owner-qualified `$genie:<skill>` selector; use bare `$<skill>` only when intentionally selecting a user-tier copy (a separately installed personal copy; Genie no longer seeds this tier). Claude Code and Hermes use `/<skill>`. Cross-skill prose below uses bare names as portable semantic routes; the orchestrator resolves the selector for the active tier.

Validate a design, wish plan, completed execution, or PR against its governing criteria. Dispatch as a subagent — never review your own work inline. The deliverable is findings plus a verdict: report and stop — never implement fixes, however small.

Context Injection

When spawned as a reviewer subagent, your dispatch prompt carries the curated scope: the target (DESIGN.md, wish draft, completed work, or PR diff), its exact path, and the extracted design or acceptance criteria. A design review does not require a wish path; later pipelines include `.genie/wishes/<slug>/WISH.md`. Use the supplied context directly — do not re-parse information already provided.

When to Use

• After `brainstorm` — validate DESIGN.md before converting it into a wish
• Before `work` — validate a wish plan is ready for execution
• After `work` — verify implementation meets acceptance criteria
• Before merge — check a PR diff against wish scope

Flow

1. **Detect target** — DESIGN.md, wish draft, completed work, or PR diff. 2. **Select pipeline** — the matching checklist below. 3. **Run checklist** — evaluate each criterion, collecting evidence. 4. **Run validations** — execute validation commands; capture pass/fail output and confirm their scope is proportional to the diff's risk and reach. Documentation-only changes, including deterministic generated documentation or plugin skill mirrors, use relevant format, link, example, generator, parity, or content-contract checks; runtime changes use focused behavior tests and add type, lint, or build checks for boundaries reached. Shared runtime/core behavior, dependency or lockfile, generated executable or runtime artifact, configuration or schema, CI or release, broad-refactor, or uncertain-impact changes require the repository full gate plus affected build or end-to-end checks. Reject zero validation and under-validation. A passing full-suite run is always valid evidence; a missing scope rationale is at most a MEDIUM gap, never by itself grounds for FIX-FIRST. A repository-documented gate is by itself sufficient justification for its scope. Preserve required aggregate integration and release gates. 5. **Tag gaps** — classify every unmet criterion by severity. 6. **Return verdict** — SHIP, FIX-FIRST, or BLOCKED, with exact fixes (files, commands, what to change) for each gap.

Escalation Diagnosis

Use this policy before any model or effort change; keep this contract identical in `fix`, `review`, and `work`.

| Cause | Diagnostic evidence | Corrective route | |-------|---------------------|------------------| | `model-capacity` | The supplied context is complete, the spec is decidable, the environment works, and attempt output shows the assigned model or effort still cannot perform the reasoning. | May raise model or effort one step, but only with new evidence and available caps. | | `missing-context` | The attempt identifies absent files, history, criteria, logs, or other inputs needed to decide. | Supply the missing context and retry at the same model and effort; MUST NOT escalate model or effort. | | `ambiguous-spec` | Two or more materially different behaviors remain consistent with the stated criteria. | Request a human decision or wish clarification; MUST NOT escalate model or effort. | | `env-tool-failure` | A reproducible environment, dependency, permission, timeout, or tool error prevents valid execution. | Repair or retry the environment/tool, or report blocked with the error; MUST NOT escalate model or effort. | | `overdesigned-plan` | Gaps cluster in optional machinery that lacks a current criterion or measurement, while a simpler design satisfies the user stories with fewer durable states or recovery paths. | Stop the fix loop and return to `brainstorm`/`wish` to remove or defer the mechanism. Re-review the amended design/plan; MUST NOT spend retries or model escalation defending it. |

Escalation eligibility requires **new evidence** produced since the previous attempt: attach the new failing output or diagnostic result, the correction already tried, and why it rules out the other four causes. A repeated verdict or unchanged failure is not new evidence and cannot authorize a model or effort change.

Model and reasoning effort belong in the active runtime's session or named-agent configuration, never in skill frontmatter. Inherit the active model by default. Only an evidenced `model-capacity` diagnosis may justify one higher-effort fresh agent, with at most two escalation attempts per group. The runtime's highest supported effort is appropriate only for a final gate or similarly demanding review when the user requested it or the evidence warrants it. Further escalation requires an explicit human decision recorded with the wish/group, old and new settings, reason, approver, and timestamp.

If an ordinary reviewer and the `final-gate` disagree, log an appeal with the wish/group, both verdicts and evidence, the contested criterion, and the human resolution. Neither verdict silently overrides the other, and the group remains `in_progress` until the appeal is resolved.

Pipelines

Design Review (after `brainstorm`)

• [ ] Problem is explicit, consequential, and readable one way
• [ ] Scope has concrete IN and OUT boundaries that fit one wish
• [ ] Chosen approach names its rationale and rejected alternatives
• [ ] Simplici