code-review

## R — Reference (≤100 words) Upstream: `skills/engineering/code-review/SKILL.md` (87 lines). Process: (1) Pin fixed point via `git diff <fixed>...HEAD` (three-dot merge-base). (2) Identify spec source: commit refs → user path → docs/specs match → ask user. (3) Identify standards: repo docs (CODING_STANDARDS.md) + **smell baseline** (13 Fowler smells: Mysterious Name, Duplicated Code, Feature Envy, Data Clumps, Primitive Obsession, Repeated Switches, Shotgun Surgery, Divergent Change, Speculative Generality, Message Chains, Middle Man, Refused Bequest). Repo overrides baseline; baseline always judgement calls. (4) Spawn parallel sub-agents with diff+standards/spec. (5) Aggregate under ## Standards / ## Spec, no cross-axis rerank. Summary: findings per axis + worst per axis. ## I — Methodology (in own words) **Core principle**: Separate *how code is written* (Standards) from *what code does* (Spec) to prevent one axis masking the other. **Two-axis independence**: - Standards = conformance to documented conventions + universal smell heuristics - Spec = traceability to originating requirements (issue, spec doc, PR description) - Parallel sub-agents ensure no context pollution; each sees only its axis inputs **Smell baseline as floor**: Even repos with zero docs get 13 structural heuristics. Each smell = labelled heuristic ("possible Feature Envy"), never hard violation. Tooling-enforced rules (lint, format) are skipped. **Spec resolution hierarchy**: Commit references (strongest) → user-provided path → filesystem search → user clarification → "no spec" (Spec axis skipped). **Output discipline**: Two separate sections, verbatim sub-agent reports lightly cleaned. No merging, no global ranking. One-line summary per axis only. ## A1 — Upstream Typical Applications - PR review automation: "review since main" → Standards catches naming/duplication; Spec catches missing acceptance criteria - Pre-merge gate: CI runs code-review on feature branch vs merge-base - Post-hoc audit: "review since v1.0.0" on a released branch - Standards drift detection: baseline smells reveal creeping complexity - Spec scope creep detection: diff contains behaviour not in issue ## A2 — Trigger Scenarios (3 contexts + language signals + sibling distinction) | Context | User Language Signals (≥3) | vs Sibling Skills | |---------|---------------------------|-------------------| | **PR/branch review request** | "review this PR", "review since main", "check my branch", "code review please" | vs diagnosing-bugs: code-review assumes code *exists*; diagnosing-bugs starts from *symptom* | | **Standards compliance check** | "does this follow our conventions", "check for code smells", "lint this diff" | vs codebase-design: code-review evaluates *existing* diff; codebase-design shapes *future* modules | | **Spec traceability audit** | "verify this implements the issue", "any scope creep", "missing requirements" | vs grill-with-docs: grill-with-docs sharpens *ideas*; code-review validates *delivered* code | ## E — Execution Steps (each with "完成标准:") 1. **Resolve fixed point** — Parse user input (commit SHA, branch, tag, `HEAD~n`). Run `git rev-parse` to validate; run `git diff <fixed>...HEAD` to capture diff; confirm non-empty. 完成标准: Fixed point resolves, diff command captured, diff non-empty, commit list noted. 2. **Locate spec source** — Search commit messages for issue refs; check user-provided path; search docs/specs/.scratch for matching names; prompt user if none found. 完成标准: Spec source identified (path or "none") with resolution method recorded. 3. **Collect standards sources** — Gather repo standards files (CODING_STANDARDS.md, CONTRIBUTING.md, .eslintrc, etc.). Append full 13-item smell baseline. 完成标准: Standards package = repo files + complete smell baseline, ready for sub-agent. 4. **Spawn parallel sub-agents** — Launch Standards agent (diff + standards package) and Spec agent (diff + spec source). Enforce 400-word limit each. 完成标准: Both sub-agents return reports; no cross-talk between them. 5. **Aggregate & present** — Render ## Standards and ## Spec sections verbatim. Add one-line summary: findings count per axis + worst issue per axis. 完成标准: Two distinct sections, no merged findings, summary line present. ## B — Boundaries & Limits **不适用**: - No git history / not a git repo - Reviewing uncommitted working tree changes without a fixed point (use `git diff HEAD` explicitly) - User wants architectural redesign (use codebase-design + grill-with-docs) - User wants bug root cause (use diagnosing-bugs) **局限**: Requires issue tracker config (docs/agents/issue-tracker.md) for commit-ref fetching. Smell baseline is heuristic—false positives possible. Spec axis degrades to "no spec" if traceability broken. ## Related Skills - **diagnosing-bugs** (bug root cause), **codebase-design** (module design vocabulary), **implement** (builds code that gets reviewed), **tdd** (drives implementation), **to-spec/to-tickets** (produces specs that Spec axis checks), **grill-with-docs** (sharpens specs before code), **setup-matt-pocock-skills** (precondition for issue tracker) ## Audit V1✓ V2✓ V3✓ | 2026-08-26 | Source: D:\outcome\self\upstream\mattpocock__skills\skills\engineering\code-review\SKILL.md