pr-code-review

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-discipline Part 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

1. Eligibility — Open, non-draft, needs review; not already reviewed this session on this tip. Else stop.
2. Resolve depth — full (default) or light from the caller.
3. 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).
4. 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).
5. 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).
6. Summarize — Short change summary (title, intent, diff shape).
7. 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.
8. 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.
9. Strengths — If any, list briefly before issues (accurate praise builds trust in the rest).
10. Re-check eligibility — Still open / same tip before publish.
11. 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 (full or light).
• 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-discipline Part 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).