pr-code-review
Dual-axis (Standards∥Spec) multi-perspective PR review with confidence ≥80 filtering, severity calibration, plan alignment, and optional GitHub/GitLab review comment. Best-of: Claude /code-review pipeline + mattpocock dual-axis + Superpowers plan/severity habits. Triggers — 「PR 代码审查」「审查这个 PR」「code-review」「/code-review」「审 PR」「pull request review」「双轴审查 PR」 / pr code review, review this PR. Do NOT use as a name alias for mattpocock code-review or Superpowers requesting-code-review.
How do I install this agent skill?
npx skills add https://github.com/fudesign2008/open-skills --skill pr-code-reviewIs this agent skill safe to install?
- Gen Agent Trust Hubpass
This skill automates pull request reviews by evaluating code against specifications and coding standards. It utilizes official GitHub and GitLab CLI tools to fetch pull request data and publish review comments. A minor security risk is present due to potential indirect prompt injection, as the skill processes untrusted text from pull request descriptions and issue bodies which could be crafted to influence the agent's review behavior. The skill is otherwise well-structured and uses official tools for its repository interactions.
- Socketpass
No alerts
- Snykwarn
Risk: MEDIUM · 1 issue
What does this agent skill do?
PR Code Review
Portable PR review for open-skills. Combines Anthropic Claude Code code-review plugin flow (eligibility, multi-perspective, ≥80 confidence, permalink comments), mattpocock Standards∥Spec separation + fixed-point diff, and Superpowers plan alignment / severity / strengths-first habits. Platform-agnostic agent dispatch (no host model-tier hardcoding).
Install:
npx skills add FuDesign2008/open-skills -g --skill pr-code-review --yes
When this applies
- User asks to review an open PR/MR, or
merge-disciplinePart R requires pre-merge review. - Skip (state reason, do not comment): PR closed, draft, trivial/automated, or already reviewed by this skill this session/tip.
Depth (full | light)
Callers (especially merge-discipline Part R) MAY pass depth=full or depth=light. Default: full.
| Depth | Behavior |
|---|---|
| full | Dual-axis Standards∥Spec; multi-perspective review preferred in parallel when the host supports it (current default process below) |
| light | Dual-axis Standards∥Spec on the pinned tip is still mandatory; parallel multi-perspective swarm is not required. Confidence ≥80 Critical/Important clearance and publish rules are unchanged |
Standalone user triggers without a depth parameter MUST use full.
Process
- Eligibility — Open, non-draft, needs review; not already reviewed this session on this tip. Else stop.
- Resolve depth —
full(default) orlightfrom the caller. - Pin fixed point — PR/MR base ref, or user-supplied commit/branch, or
origin/<default-branch>for merge candidates. Confirm ref resolves and three-dot diff vs tip is non-empty. Empty diff → abort (do not spawn perspectives). - Resolve Spec source (first hit wins): PR/issue body + linked tickets → user path → OpenSpec change delta / related
openspec/specs→ ask user → else Spec axis = skipped (no spec available). - Resolve Standards sources —
AGENTS.md/CLAUDE.md(root + dirs the PR touches), plus coding-standards docs if present. Optional smell baseline: reference.md (repo docs override; smells are judgement calls, never sole hard violations). - Summarize — Short change summary (title, intent, diff shape).
- Dual-axis review (do not merge-rank across axes):
- Standards axis — Documented guidance breaches (cite file + rule); optional smell heuristics.
- Spec axis — Missing/partial planned behavior; unjustified scope creep; wrong implementation of a stated requirement (quote spec/plan line). Skip entire axis if no Spec source.
- full only: also prefer parallel perspectives (bugs-in-diff-only, blame/history, prior PR comments on same files, in-file comment guidance).
- light: skip the mandatory parallel swarm; single-pass dual-axis is enough.
- Calibrate — Each surviving candidate: severity Critical / Important / Minor, then confidence 0–100 (rubric below). Drop scores below 80. Map: Critical/Important usually land ≥75–100 if verified; Minor usually drops unless guidance-hard.
- Strengths — If any, list briefly before issues (accurate praise builds trust in the rest).
- Re-check eligibility — Still open / same tip before publish.
- Publish — Comment via
gh/glab(or report in-session if host forbids comment). Dual-axis sections in the comment; full-SHA permalinks. Templates: reference.md.
Confidence rubric
| Score | Meaning |
|---|---|
| 0 | False positive / pre-existing / fails scrutiny |
| 25 | Unverified / stylistic / not in guidance |
| 50 | Real but nit or rare |
| 75 | Likely real and important; or explicitly required by guidance/plan |
| 100 | Definitely real; evidence confirms |
False positives to discard
Pre-existing; lookalike non-bugs; pedantic nits; linter/typechecker/compiler catchable; generic “more tests/docs/security” unless guidance/plan requires it; unchanged lines; intentional PR-scoped behavior.
Host contract (merge-discipline Part R)
- Run against the open PR/MR about to merge, at the depth Part R selected (
fullorlight). - Pass → neither axis retains ≥80 Critical or Important issues (Minor-only or all scores below 80 = pass). Light depth MUST NOT weaken this gate.
- Fail → report fail to the caller.
merge-disciplinePart R owns fix → re-review until clean (or user explicit Part R skip / accept-residual 留痕) before Part D. Implementing findings alone is not a Part R pass. - CI green / coverage skip is not a substitute.
- Do not require a full receiving-code-review loop to pass Part R.
After feedback (optional pointer)
When acting on review comments (human or this skill): verify against the codebase before changing anything; no performative agreement; reasoned technical pushback when wrong; clarify unclear items before implementing. Full reception discipline is optional and out of Part R.
Integration guide
- Strong-dep from
merge-discipline; missing → abort with
npx skills add FuDesign2008/open-skills -g --skill pr-code-review --yes - Do not rename to
code-review(external name collision).
How can the creator link this skill?
Add the canonical catalog link to the repository README so users can inspect current installs and available audits. The publishing guide covers the complete discovery path.
<a href="https://skillzs.dev/skills/fudesign2008/open-skills/pr-code-review">View pr-code-review on skillZs</a>