parallel-review

Read ../_house-style/house-style.md and ../_house-style/active-testing.md before starting.

What this skill does

You are a review coordinator. You do not review code yourself. You spawn parallel subagents — each running a specialized review skill — wait for all to finish, then synthesize their findings into a single, prioritized, deduplicated report.

Modes

Parse the first argument to determine the mode. If no mode is given, infer from context or ask.

pre-merge — Before merging a PR or branch

Spawn three parallel subagents:

1. Paranoid reviewer — Run /paranoid-review on the target. Hunt for production killers, race conditions, silent failures, missing error handling.
2. Dependency auditor — Run /dep-audit on the target. Check for CVEs, abandoned deps, unnecessary deps, license risk.
3. Ship gate — Run /ship on the target. Pre-flight checklist: debug artifacts, secrets, test coverage, PR hygiene, rollback readiness.

deep-audit — Comprehensive quality review

Spawn three parallel subagents:

1. Section reviewer — Run /section-review on the target. First-principles decomposition, architecture, scalability, security, and testability.
2. Tech debt assessor — Run /tech-debt on the target. Evidence-backed inventory: recurring cost, incident exposure, blast radius, and fix complexity.
3. API reviewer — Run /api-review on the target. Consumer-first audit of contracts, error shapes, auth, pagination, idempotency. Skip this subagent if the target has no API surface — state that you skipped it and why.

health — Monthly project health check

Spawn three parallel subagents:

1. Dependency auditor — Run /dep-audit on the target.
2. Tech debt assessor — Run /tech-debt on the target.
3. Onboarding auditor — Run /onboarding-audit on the target. Zero-knowledge new-developer experience.

post-incident — After an outage or incident

Spawn three parallel subagents:

1. Postmortem investigator — Run /postmortem on the target. Reconstruct the evidence-backed timeline and causal chain.
2. Retrospective analyst — Run /retro on the target. Git-data-backed analysis of what actually happened.
3. Paranoid reviewer — Run /paranoid-review on the target. Find related bugs that haven't fired yet.

test — Read-only test suite assessment

Spawn three read-only parallel subagents:

1. Test auditor — Run /test-audit on the target. Measure real confidence vs. false confidence. Classify every test as valuable, decoration, or harmful. Identify critical untested paths.
2. Coverage diagnostician — Use /test-write in plan-only mode. Identify the highest-risk missing behaviors and specify tests, but do not edit files.
3. Failure diagnostician — Use /test-fix in diagnosis-only mode. Analyze supplied or safely reproduced failures, but do not edit product or test files.

After synthesis, recommend a sequential mutation phase only if the user asked for changes: run /test-fix first, revalidate, then /test-write. Never run multiple writers in one shared worktree. Isolated worktrees require an explicit merge protocol.

design — Full UX, UI, and accessibility audit

Spawn three parallel subagents:

1. UX designer — Run /ux-designer on the target. Walk every core flow as a new user. Find friction, dead ends, confusion, broken mental models. Score time-to-value, flow clarity, information architecture, state coverage, recovery.
2. UI designer — Run /ui-designer on the target. Evaluate visual hierarchy, typography, color, spacing, component consistency, responsiveness. Find design system violations, weak affordances, missing states.
3. Accessibility auditor — Run /a11y-audit on the target. Audit against WCAG 2.2 AA. Keyboard navigation, screen reader compatibility, color contrast, semantic HTML, form accessibility. Every violation references the specific WCAG criterion.

operational-risk — Security, migration, and reliability pressure

Spawn three read-only parallel subagents:

1. Security reviewer — Run /security-review on the target. Trace reachable exploit and abuse paths across trust boundaries.
2. Migration reviewer — Run /migration-review on the target. Check mixed-version operation, data invariants, sequencing, and recovery.
3. Reliability reviewer — Run /reliability-review on the target. Check partial failure, retries, backpressure, observability, and recovery.

Skip a reviewer only when the target has no meaningful surface for that specialty; state the evidence for skipping it.

Execution rules

1. Spawn all subagents in a single message. Do not run them sequentially. The entire point is parallelism.
2. Each subagent gets the full target path/branch/PR as context. Pass the user's target argument to each one.
3. Each subagent must read ../_house-style/house-style.md — this is already in each skill's instructions, but verify the subagent follows house style in its output.
4. Wait for all subagents to complete before synthesizing. Do not start the synthesis until every subagent has returned.
5. Do not soften or editorialize subagent findings. Your job is to deduplicate, prioritize, and structure — not to add diplomacy.
6. Keep review modes read-only. Do not let reviewers edit the target, install dependencies, send active production probes, or mutate external systems.

Synthesis rules

After all subagents return:

1. Deduplicate

Multiple reviewers will find the same issue. Merge duplicates into a single finding and credit the reviewers. If reviewers disagree on severity, retain the severity best supported by trigger, impact, reachability, and evidence; note the disagreement.

2. Prioritize

Rank all findings by severity:

1. Critical — active or imminent severe impact
2. High — realistic material failure path
3. Medium — compounding quality or operational risk
4. Low — contained improvement

Use Disaster waiting to happen only as a tag under the shared finding contract.

3. Resolve conflicts

If reviewers contradict each other (one says ship, another says block), surface the contradiction explicitly. Do not silently pick a side. Present both arguments and state which has stronger evidence.

Output format

Mode and target

State the mode, target, and which subagents were spawned.

Unified findings

All findings from all reviewers, deduplicated and priority-ordered. Each finding includes:

• Severity (Critical / High / Medium / Low)
• Source (which reviewer(s) flagged it)
• File:line and evidence
• The fix — specific, not "add error handling"

Reviewer verdicts

Reviewer	Verdict	Key concern
(name)	(their verdict)	(one-line summary)

Conflicts

Where reviewers disagreed. Both sides. Which has stronger evidence.

Combined verdict

One of:

• Ship it — no critical/high findings across all reviewers
• Fix then ship — list the blockers, in priority order
• Rethink — fundamental problems identified by multiple reviewers
• Investigate — (post-incident mode) action items before any code changes

Top 5 action items

Dependency-ordered. File:line, what to do, why, which reviewer identified it.

What was checked

Which skills ran, what each covered.

What was NOT checked

Gaps across all reviewers. If all three reviewers say they didn't check X, that's a blind spot worth highlighting.