review
Review a change for correctness, quality, and guideline compliance using a library of specialized reviewer roles. Use before opening a PR, after finishing a chunk of work, or to review an existing PR. Two output modes — a local severity-ranked report for a fix loop (default), or inline GitHub comments (--comment). Scores every finding for confidence, filters the noise, and delegates each role adversarially.
ronaknnathani · 已有静态检查记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add ronaknnathani/relay --skill review技能说明
name: review description: Review a change for correctness, quality, and guideline compliance using a library of specialized reviewer roles. Use before opening a PR, after finishing a chunk of work, or to review an existing PR. Two output modes — a local severity-ranked report for a fix loop (default), or inline GitHub comments (--comment). Scores every finding for confidence, filters the noise, and delegates each role adversarially.
Review
Review a change with a library of specialized reviewer roles, score every finding for confidence, drop the noise, and emit results in one of two modes. The goal is signal: high-confidence findings a senior engineer would actually raise, each with a concrete fix — never a wall of nitpicks.
Modes
- report (default) — produce a local, severity-ranked findings report for the author or a fix
loop (e.g.
pr-fix). Posts nothing. - inline (
--comment) — post a summary comment plus inline comments on the PR viagh.
The reviewer-role library
Each role is a prompt file under agents/, focused on one failure mode. Run a role as its own
sub-agent when sub-agents are available (fresh context, parallelizable); otherwise apply its prompt
inline, one role at a time.
Every role reports under one contract: score each finding 0-100 for confidence and surface only those ≥ 80; tag each with a severity (Critical / Important / Suggestion) as a separate axis from confidence; stay advisory — surface findings with a concrete fix, never block. This contract overrides any different scale or posture a bundled role's own prompt describes.
- code-reviewer (
agents/code-reviewer.md) — guideline compliance, correctness, edge cases. - silent-failure-hunter (
agents/silent-failure-hunter.md) — swallowed errors, log-and-continue, fallbacks that hide failure. - type-design-analyzer (
agents/type-design-analyzer.md) — encapsulation and invariants; make illegal states unrepresentable. - pr-test-analyzer (
agents/pr-test-analyzer.md) — behavioral and error-path coverage; tests that catch real bugs, not implementation-coupled ones. - comment-analyzer (
agents/comment-analyzer.md) — comment accuracy, "why" over "what", rot. - security (
agents/security.md) — STRIDE pass over trust boundaries; Always/Ask-First/Never; injection and untrusted-input handling. - git-history (
agents/git-history.md) —git blame/log context: does the change contradict why the surrounding code exists? - prior-pr-history (
agents/prior-pr-history.md) — was this approach already tried, discussed, or reverted in earlier PRs?
1. Scope and the should-review gate
Load guidance before judging standards. Read the repo's AGENTS.md when present, then read global
guidance from ~/AGENTS.md or ~/.config/agents/AGENTS.md when present. Repo guidance takes precedence
over global guidance. Include the applicable guidance in each reviewer role's prompt so sub-agents review
against the same rules instead of relying on memory.
Pin a fixed point and confirm there is something to review:
BASE=$(git merge-base HEAD origin/HEAD 2>/dev/null || git rev-parse HEAD~1)
git diff --name-only "$BASE"...HEAD # must be non-empty
Pinned inputs. When the caller pins the checkout, base and head commits, guidance, and spec (as
review-pr does for someone else's pull request), use exactly those. Do not rediscover the base from
origin/HEAD or re-read guidance from the checkout. Read code only through git against the named
checkout (git -C <checkout> show|diff|grep|log), never by opening its files, because the checkout
may hold untrusted code.
If gh is available and a PR exists, scope to the PR diff and gather context for dedup:
REPO=$(gh repo view --json nameWithOwner --jq .nameWithOwner)
PR=$(gh pr view --json number --jq .number)
HEAD_SHA=$(gh pr view --json headRefOid --jq .headRefOid)
gh pr diff "$PR" # the PR-scoped diff to review
gh pr view "$PR" --json title,body,files # spec text + changed files — feeds the Spec axis (step 3)
gh api "repos/$REPO/pulls/$PR/comments" --jq '.[] | {path, line, body}' # existing comments (dedup)
Skip and report the skip reason (post nothing) if: the diff is empty, the PR is closed, the change
is trivially correct (dependency bump, single config value, no logic), or this exact HEAD_SHA was
already reviewed with no new commits.
2. Pick the applicable roles
Always run code-reviewer, silent-failure-hunter, git-history, and prior-pr-history. Add roles by what changed: tests → pr-test-analyzer; new/changed types → type-design-analyzer; comments/docs → comment-analyzer; any trust boundary, input parsing, auth, or untrusted data → security. If the caller named specific aspects, run only those.
3. Review along two axes, kept separate
Every change is judged on two independent axes — do not let one contaminate the other, and do not merge them when reporting:
- Standards — does it follow the repo's and author's documented conventions (repo
AGENTS.md, then global~/AGENTS.mdor~/.config/agents/AGENTS.md, plus CLAUDE.md or equivalent guideline files)? Omit anything a linter or compiler already enforces; distinguish a hard violation from a judgment call. - Spec — does it do what the task/PRD/issue asked? Report three buckets: missing or partial requirements, changes nobody asked for, and implementations that look incorrect. Quote the spec.
4. Delegate each role adversarially (doubt-driven)
This is the highest-leverage rule. When you hand a finding or a diff to a reviewer sub-agent, give it only the artifact and the contract it must satisfy — never your own reasoning or your conclusion that the code is fine. A reviewer that sees "this looks correct" anchors to agreement. Prompt it to find what is wrong, not to confirm. If a role and the author disagree, run at most 3 bac
阅读完整来源 ↗