AGENT SKILL

adversarial-review

Adversarial code review of a target package. Explore the package, run the real gates (typecheck, lint, build, tests), triage and refute optimistic claims, then save a prioritized findings table to a report file rather than dumping it inline. Use when asked to review, audit, or adversarially check a package's quality. Invoke with /adversarial-review (renamed from /review to avoid colliding with gstack's pre-landing PR review).

使用场景代码审查
License 许可
MIT
最近更新
2026-11-01
Stars
★ 2
Forks
1

01 前言 Front Matter

name
adversarial-review
description
Adversarial code review of a target package. Explore the package, run the real gates (typecheck, lint, build, tests), triage and refute optimistic claims, then save a prioritized findings table to a report file rather than dumping it inline. Use when asked to review, audit, or adversarially check a package's quality. Invoke with /adversarial-review (renamed from /review to avoid colliding with gstack's pre-landing PR review).

03 SKILL.md 正文

Adversarial Code Review

A skeptical, evidence-first review pass. Every claim must be grounded in real command output — no findings asserted from reading alone, no "looks fine" without running the gate.

Procedure

  1. Explore the target package. Map its entry points, public surface, tests, and config. Identify what the package promises and where it could break.

  2. Run the gates and record real output. Execute pnpm typecheck, pnpm lint, pnpm build, and pnpm test (scoped to the package). Capture actual stdout/stderr — exit codes, error lines, counts. Never paraphrase a gate you did not run.

  3. Triage findings; refute optimistic claims. For each candidate finding, try to disprove it against source and gate output before accepting it. Separate real regressions from pre-existing/infra noise (see the repo's known-issue memories). Drop anything you cannot reproduce.

  4. Save a prioritized findings table to a report file. Write the results to a report file (e.g. docs/reviews/<package>-review.md or $CLAUDE_JOB_DIR/tmp/<package>-review.md) — a P0/P1/P2 table with file:line, evidence, and proposed fix. Do not dump the full table inline; summarize and point to the file.

Notes

  • Tabs, not spaces; Biome over Prettier (see Formatting & Conventions).
  • Treat unverified claims as refuted by default.
  • Keep the inline reply concise; the report file holds the detail.

04 同仓库更多 Skills

ancyloce
ancyloce/anvilkit-studio

react-doctor

Use when finishing a feature, fixing a bug, before committing React code, or when the user types `/doctor`, asks to scan, triage, or clean up React diagnostics. Covers lint, accessibility, bundle size, architecture. Includes a regression check and a full local-triage workflow that fetches the canonical playbook.

前端开发
★ 2
ancyloce
ancyloce/anvilkit-studio

review-remediate

Use for the review-then-fix-everything loop on a package or diff — deliver a grounded review report immediately (no scoping questions, no plan-approval), then remediate every finding as tracked numbered items, gating each batch on typecheck/lint/test/build plus an independent Codex pass until clean. Trigger on "review and fix <package>", "remediate the review", "review-fix-verify", or when the user wants a deep review followed by autonomous remediation. Distinct from gstack `review` (pre-landing PR review) and `code-review` (diff only) — this reviews from scratch AND fixes to green.

★ 2
ancyloce
ancyloce/anvilkit-studio

review-fixes

Close out an EXISTING code-review findings report — one subagent per finding, each with a regression test — then regenerate EN/zh/ja/ko docs, run all gates, and verify no stale or unexported symbols remain. Trigger on "/review-fixes", "fix all N review findings", "close the review findings", or when the user hands over a findings report to remediate. Distinct from review-remediate, which produces the review itself first — this skill starts from a report that already exists.

★ 2
ancyloce
ancyloce/anvilkit-studio

release-prep

Pre-release verification + go/no-go checklist for publishing @anvilkit packages. Detects the release surface (main workspace fixed-group vs components submodule), runs changeset status and the check:all release gates, enforces this repo's publish gotchas (version.ts drift, ui peer ranges, never-publish canvas-templates, bundle/api-snapshot), then hands the exact publish command to the user. Use when asked to "prep a release", "check release readiness", "/release-prep", or before cutting a version. Does NOT commit or publish.

★ 2
ancyloce
ancyloce/anvilkit-studio

react-doctor

Use when finishing a feature, fixing a bug, before committing React code, or when the user types `/doctor`, asks to scan, triage, or clean up React diagnostics. Covers lint, accessibility, bundle size, architecture. Includes a regression check and a full local-triage workflow that fetches the canonical playbook.

前端开发
★ 2
ancyloce
ancyloce/anvilkit-studio

pre-refactor

暂无简介

代码审查
★ 2