Agent-Skills.md

Agent Skills: Subagent Review

|

UncategorizedID: i9wa4/dotfiles/subagent-review

Repository

i9wa4/dotfiles
i9wa4License: MIT
6

Install this agent skill to your local

pnpm dlx add-skill https://github.com/i9wa4/dotfiles/tree/HEAD/skills/subagent-review

Skill Files

Browse the full folder contents for subagent-review.

Download Skill

Loading file tree…

skills/subagent-review/SKILL.md

Skill Metadata

Name
subagent-review
Description
"USE FOR: Native five-perspective guardian/critic review. DO NOT USE FOR: implementation, dispatcher fan-out, model/tier choice, or public comment posting."

Subagent Review

UTILITY SKILL: Use the current runtime's five native reviewer perspectives for guardian/critic review. The active role owns synthesis and approval. Do not implement, delegate approval, dispatch fan-out, choose models/tiers, own implementation, or build cross-engine reviewer pools.

Normal lane: orchestrator -> guardian -> critic -> guardian -> orchestrator.

Guardian is the Codex review owner. Critic is the Claude subordinate reviewer. Guardian packages evidence for critic, waits for the recommendation, then returns the verdict.

Default set: security, architecture, historian, code, and QA. Guardian uses Codex-native perspectives; critic uses Claude-native perspectives.

Boundary

This skill owns reviewer perspectives, guardian/critic handoff, synthesis, severity, and output expectations. postman.md owns only the live route, reply flow, verdicts, no-bypass gates, and public-posting approval gates. github owns GitHub mechanics; create-review-comment owns user-facing PR comment drafts.

Workflow

  1. Run the cheapest relevant verifier first.
  2. For substantive reviews, use the five defaults before synthesizing. Direct review is acceptable for trivial follow-ups if stated.
  3. Add data or research reviewers only for specialized questions; do not replace the defaults.
  4. Use native subagents only. No dispatcher, model/tier selection, or cross-engine pool.
  5. Give each subagent bounded read-only paths, context, and output shape.
  6. Suppress unverified findings, deduplicate, and produce the verdict or recommendation.

Role Rules

  • Guardian uses Codex reviewers, owns the final verdict, and treats the critic recommendation as evidence.
  • Critic uses Claude native reviewers and returns a recommendation.
  • Critic sends the review recommendation to guardian, not orchestrator.
  • Reviewer subagents must not edit, commit, push, or approve.
  • If evidence is incomplete, continue direct review.

Output

Preserve perspective, paths, checked evidence, severity, confidence, and gaps. Final verdicts and recommendations use the guardian-mediated surface, not raw subagent output.