Review

review — Universal Review Gate

Runtime syntax: invoke the plugin copy through the active runtime's owner-qualified skill selector; use a bare selector only when intentionally selecting a user-tier copy (a separately installed personal copy; Genie no longer seeds this tier). Cross-skill prose below uses bare names as portable semantic routes; the orchestrator resolves the selector for the active runtime.

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.

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
• Simplicity Case states the simplest complete design; every added mechanism has a present requirement or measurement, and future complexity has a concrete adoption trigger
• Current state is bounded and history is separated before deltas, sharding, caches, or distributed synchronization are considered
• Decisions are consistent with the approach and repository constraints
• Risks and assumptions name mitigations or explicit acceptance
• Success criteria are testable without requiring execution-group details
• Next step is wish; DESIGN.md contains no TODO/TBD placeholders

Plan Review (before work)

• Problem statement is one sentence and testable
• Scope IN has concrete deliverables; Scope OUT is explicit
• Every task has testable acceptance criteria
• Tasks are bite-sized and independently shippable
• Dependencies tagged (depends-on / blocks)
• Every group has non-zero validation proportional to its planned risk and reach, with escalation and scope rationale
• Aggregate integration and release gates remain present where the repository requires them
• Simplicity Case is executable: no group builds machinery marked deferred, and every stateful mechanism maps to a current success criterion

Execution Review (after work)

• All acceptance criteria met with evidence
• Validation commands run and passing; their recorded scope matches the actual diff's risk and reach
• No under-validation or zero validation; full-suite runs carry a scope rationale (a missing rationale is at most MEDIUM)
• No scope creep — only wish-scoped changes
• Work is auditable — commands and outcomes captured
• Quality pass: security, maintainability, correctness
• No regressions introduced
• Implementation did not introduce caches, synchronization states, configuration, or abstractions absent from the approved Simplicity Case

PR Review (before merge)

• Diff matches wish scope — no unrelated changes
• File list matches wish's "Files to Create/Modify"
• No secrets, credentials, or hardcoded tokens in diff
• Risk-proportional validation passes, and required aggregate gates remain intact
• Commit messages reference wish slug

Severity & Verdicts

In design and plan review, unjustified stateful machinery—such as speculative caches, deltas, sharding, background coordination, or retry state machines—is a HIGH gap because it creates permanent correctness and maintenance obligations without delivering a current criterion. If removing it changes the governing approach, return BLOCKED with overdesigned-plan and replan instead of asking a fixer to preserve it.

Persistence handoff

The reviewer is read-only. Return a timestampable evidence block containing the review context, target SHA/path, commands and outcomes, verdict, and gaps. For a design review, return the exact reviewed-content SHA-256 defined by the bounded evidence block in DESIGN.md as reviewed-sha256; the invoking orchestrator passes that value unchanged to the stamp command as --reviewed-sha256. Stamping rejects a current design that differs from the reviewed content, verification rejects any later edit, and the reviewer never recomputes a digest for content it did not review. For plan, execution, and PR review, the orchestrator appends the block under the wish's ## Review Results and owns every durable transition:

• plan SHIP → APPROVED; plan FIX-FIRST → FIX-FIRST; plan BLOCKED → BLOCKED;
• execution and PR verdicts are appended while the wish remains IN_PROGRESS;
• only an authorized merge plus required QA changes the wish to SHIPPED.

Do not claim the next stage is active until the orchestrator confirms the write. Never edit WISH.md, the brainstorm jar, or task state as the reviewer.

SHIP next-steps

FIX-FIRST loop

1. Diagnose first. For overdesigned-plan, return to brainstorm/wish without consuming a fix attempt.
2. Otherwise auto-invoke fix with the severity-tagged gap list.
3. After fix completes, re-run review (max 2 fix loops).
4. Still FIX-FIRST after 2 loops → return BLOCKED with an Escalation Diagnosis; never raise model or effort automatically.

When a failure's root cause is unclear, invoke trace before dispatching fix — fix then applies the cause-specific correction from the trace report. An unclear cause is not evidence of model-capacity.

Dispatch

Reviewer ≠ engineer. The orchestrator dispatches review as a separate subagent via the native delegation surface — an agent never reviews its own work. Follow-ups to a running reviewer go through native follow-up messaging. For change-types that warrant deeper scrutiny, the orchestrator also convenes a Lens Panel (below); those lenses advise, but the checklist still owns the verdict.

Reviewer snapshot

When dispatching a review of committed work, the orchestrator pins the review to an immutable tree — a detached worktree at the exact commit under review — so that writers continuing in the primary checkout cannot move the ground under a verdict in progress. Pin by default when other groups are still writing in the primary checkout; skip it when reviewing an uncommitted working tree, which cannot be pinned:

git worktree add --detach <worktreesBase>/<repo>-review-<shortsha>-<unique> <commit>   # provision
git worktree remove <path>                                                             # teardown, after the verdict

(<worktreesBase> is $GENIE_WORKTREES_DIR, else <GENIE_HOME>/worktrees — the base the doctor launch-residue check also scans. <unique> is a per-review disambiguator — the reviewer/group name or a mktemp-style random suffix — because a bare <repo>-review-<shortsha> collides across concurrent reviewers, repeated reviews of one commit, and repos sharing a basename.)

No helper wraps these, because the commands are already fail-safe: worktree add --detach creates without touching any branch, and worktree remove refuses a dirty tree by default. Teardown is the orchestrator's job and nothing else does it — genie doctor's launch-worktree check only classifies worktrees on a wish/<slug>-<group> branch, so a detached snapshot is foreign to it: surfaced only inside the aggregate "other checkouts" count and never removed by --fix. If a review crashes before teardown, remove the leftover explicitly with git worktree remove <path>. The reviewer works READ-ONLY in the snapshot path (no install, no build, no writes) and the verdict cites the pinned commit. Provisioning and teardown are orchestrator-side plumbing: the git-state freeze in AGENTS.md permits git worktree add/remove/prune on snapshot paths, while reviewers themselves never mutate git state.

Lens Panels

When the change-type warrants it, the orchestrator dispatches lens reviewers alongside the standard reviewer — each a separate subagent whose prompt carries its lens file (path + content) and the curated review scope. Convene a lens only when the change actually touches its surface; lenses advise, but the verdict still comes from the checklist above — never from a lens.

Resolve references/lenses/questioner.md from the directory containing this loaded SKILL.md. Resolve a sibling lane from that skill directory's parent (for example ../supply-chain/SKILL.md). If a separately installed skill is missing a sibling, mark that advisory lane unavailable instead of guessing a source-checkout or global plugin path.

Verdict Reporting

The verdict plus severity-tagged gaps ARE the review output — deliver them in your final message. For a plan or PR, the invoking orchestrator persists the returned block in git. The reviewer never mutates files or task state:

genie task done belongs to the orchestrator, after a clean verdict — never to the reviewer.

Rules

• Never mark PASS without evidence from this session — verify, don't assume.
• Never ship with CRITICAL or HIGH gaps.
• Report findings and stop — no unrequested fixes; corrections belong to fix.
• Every FAIL includes an actionable fix (file, command, what to change).
• Keep output concise, severity-ordered, and executable.

Session close (required)

When spawned as a native subagent, your final message IS the completion signal — the orchestrator is notified when you finish; do not poll or emit a separate contract call. State the verdict, then end with exactly one terminal outcome as the last word:

• done — review completed; verdict (SHIP / FIX-FIRST / BLOCKED) and severity-tagged gaps stated.
• blocked — could not complete the review (missing artifact, unrunnable validation). State exactly what you need.
• failed — aborted or irrecoverable. State why.

blocked / failed must include a one-line reason.