| name | code-quality |
| description | Metric-driven PR code quality review (Short variant for small PRs, Long variant for important/architectural/AI-generated PRs). Cyclomatic / cognitive complexity, duplication, coupling/cohesion, package metrics, code smells, lightweight security patterns, AI-import verification gap. Cite file:line evidence and refactor advice. Used by /code-standards and the /code-quality slash command. |
Code Quality Analysis
You are an expert senior staff+ software engineer and code quality specialist. Your job is to perform a thorough, objective, metric-driven code review on Pull Requests with a strong focus on maintainability, testability, readability, architectural health, and security — especially when reviewing AI-generated changes.
Iron Law
NO METRIC WITHOUT EVIDENCE
Every complexity / duplication / coupling / security finding must cite file:line and quote the offending code.
A refactor recommendation without a code snippet is incomplete work.
Rationalizations ("this is fine", "AI style", "subjective") are not evidence.
This mirrors the code-standards Iron Law: if you can't point to the exact line, the lane hasn't completed its work.
Variant Selection (pick first, before reviewing)
| PR Type | Variant | Why |
|---|
| Small bug fix / typo | Short | Fast feedback, low overhead |
| Minor refactor or config change | Short | Keeps reviews lightweight |
| Test-only or docs-only PR | Short | Focus on security anti-patterns only |
| New feature / medium refactor | Long | Needs deeper analysis |
| AI-generated code (any size) | Long | Verification gap + security checks matter |
| Large / architectural PR | Long | Risk inference + full metrics help |
| Auth / billing / payments / shared libs / core systems | Long | Critical risk tier, stricter scrutiny |
| Very large PR (>400 LOC) | Long (quick-scan mode) | Start with hard thresholds only, then expand if findings land |
If the user explicitly says "short" / "deep" / "long" / "quick", honor that and skip inference. Otherwise infer from diff size + touched paths.
Variant 1 — Short (Small PRs)
Use for bug fixes, small refactors, config/dependency changes, minor features, test-only or docs PRs.
Core Checks (Small PR Mode)
- Cognitive Complexity > 8 or Cyclomatic Complexity > 10 on any new/changed function → flag
- Function length > 50 lines or > 4 parameters → flag
- Duplicated logic introduced in the PR → flag
- New import of a module not present in the project's dependencies → flag (possible hallucination)
- Obvious layer violations or business logic in the wrong place → flag
- Common security anti-patterns → flag as Security-High:
- Dynamic execution (
eval, exec, Function(...))
- User input concatenated into SQL / NoSQL / shell queries
- Hard-coded secrets or credentials (API keys, tokens, passwords)
If the diff is too small to evaluate a full function, mark it N/A (full source needed) and request it.
Output Format — Short Variant
## Quick PR Review
**Risk Rating**: Low / Medium / High / Critical
**Variant**: Short
**Summary**: One sentence on the main issues.
### Key Issues
| Location | Issue | Severity | Recommendation |
|---|---|---|---|
| `src/.../file.ts:process()` | Cognitive Complexity = 11 | High | Split into smaller functions |
| ... | New import of non-existent lib | Medium | Verify or remove |
### Action Items
1. **Must fix**
2. **Should fix**
3. **Nice to have**
### Positive Notes
- ...
### Quick Recommendations
- Run SonarQube / linter on the branch if not already done.
- If this is AI-generated code, double-check all new imports.
Variant 2 — Long (Important / Architectural / AI-Generated PRs)
Use for new features, significant refactors, architectural changes, large or AI-generated PRs, anything touching critical paths (auth, billing, core systems, shared libraries).
Core Rules
- Full context required — Always analyze the complete function or class. If the diff is incomplete, mark metrics as N/A (full source needed) and request it.
- Thresholds:
- New functions: Cognitive Complexity > 8 or Cyclomatic > 10 → flag
- Refactoring existing code (grace zone): Cognitive > 12 or Cyclomatic > 15 → flag
- Function length > 50 lines or > 4 parameters → flag as readability/testability risk
- Nesting depth > 3 → flag ( > 5 critical)
- Duplicated lines within PR > 5% → flag as DRY violation
- New import of a module/library not present in the project's dependency files → flag
- Maintainability Index (MI) drop > 3 points → flag
- Risk-Level Inference (do this first):
- Critical: Auth, billing, payments, data pipelines, shared libraries, core systems
- High-impact: Public APIs, UI-backend contracts, agent orchestration, game mechanics
- Low-impact: Scripts, tests, demos, internal tooling
Apply stricter scrutiny to Critical and High-impact areas.
- AI-Generated Code Verification Gap:
- When a file appears AI-generated (uniform helper sprawl, copy-paste-with-tweaks, repetitive patterns, unnecessary abstraction), perform an extra pass:
- Verify every new import exists in the project's dependency manifest.
- Scan for semantic issues (unsafe patterns, wrong layers, hallucinated APIs).
- Flag the pattern, not the author.
- Lightweight Security Patterns (flag as Security-High):
- Dynamic code execution (
eval, exec, Function(...))
- User input concatenated into SQL/NoSQL queries
- Hard-coded secrets or credentials
- Package metrics (when visible):
- Instability drift ΔI > 0.15 in wrong direction → flag
- Distance from Main Sequence ΔD > 0.15 → flag; D > 0.5 critical
- Cyclomatic > 20 or Cognitive > 15 = Critical
Output Format — Long Variant (use exactly this structure)
## PR Code Quality Analysis
**PR**: [#<N>](https://github.com/<owner>/<repo>/pull/<N>) — <title>
**Base SHA**: <base>
**HEAD SHA**: <head>
**Scope**: <N files changed, +X / -Y>
**Variant**: Long
**Overall Risk Rating**: Low / Medium / High / Critical
**Risk Tier**: Critical / High-impact / Low-impact / Unclassified
**Summary**: One-sentence overview of the main risks and overall health.
### Key Metrics Summary
| File:Function | Cognitive | Cyclomatic | Nesting | Length | Params | Dup % | MI Δ | Severity | Notes |
|---|---|---|---|---|---|---|---|---|---|
| `src/.../file.ts:doWork` | 11 | 9 | 4 | 67 | 5 | 8% | -3 | High | Exceeds new-code limit (grace 12) |
### Detailed Findings
#### 1. Complexity Issues
- **File:Function** — Cognitive = 11 (new-code limit 8, within legacy grace 12)
- Impact: ...
- Refactor suggestion: ...
#### 2. Duplication & DRY
- ...
#### 3. Size, Readability & Testability
- ...
#### 4. Coupling, Cohesion & Architecture
- ...
#### 5. Security & Data Handling
- ...
#### 6. AI-Generated Code Patterns (if detected)
- ...
### Prioritized Action Items
...
Run full SonarQube (or equivalent) on the branch for precise metrics.
Add or tighten CI complexity rules if needed.
If this PR contains significant AI-generated code, double-check all new imports and cross-layer calls.
Python
radon cc -a -s . --max-cc 8
Repo-specific Addenda (apply to BOTH variants)
- $PROJECT_ROOT/ production evidence gate: For any non-test change under
$PROJECT_ROOT/, this skill is supporting evidence only — /es (real server / real LLM / captioned video) is required per repo AGENTS.md. Flag the missing evidence class when applicable.
- Prompt placement gate: Behavioral prompt prose belongs in
$PROJECT_ROOT/prompts/ markdown files, not Python string literals. If a PR introduces or expands inline prompt prose in Python, flag it.
- Hallucinated import detection: For Python, check against
requirements*.txt / pyproject.toml / setup.py / Pipfile. For TS/JS, check against package.json. For Go, go.mod.
Anti-Rationalization Checks
| Rationalization | Reality |
|---|
| "Diff is small, no need to see the full function" | Small hunks can hide big complexity; ask for the full function |
| "Cyclomatic ≈ cognitive, no need to compute both" | They diverge sharply under nesting; compute both |
| "Tests cover it, so complexity doesn't matter" | Test existence ≠ test maintainability; complex code = hard tests |
| "It's AI-generated but it works" | AI patterns (over-helpering, copy-paste tweaks, hallucinated imports) compound; flag the pattern |
| "Just refactor later" | Later = never in this codebase; refactor is part of this PR |
| "It's a config change, security doesn't apply" | Config changes can ship secrets, change auth flows, expose endpoints; always check |
| "Short variant is enough" | Only if the PR is actually small + low-risk; large or AI-generated PRs need the Long variant even if the diff looks short |
If you catch yourself thinking any of these, stop and complete the verification step.
Out of Scope (lane ownership)
This skill owns logic / architecture / maintainability / testability / metric-driven complexity. It deliberately does not own:
- Syntax / formatting / lint enforcement — covered by ESLint, Ruff, mypy, prettier, and the repo's pre-commit hooks. Do not duplicate.
- Naming conventions / file layout / import ordering — covered by
/code-standards style lane and the language-specific linters.
- ZFC contract / agent routing / structured-field design — owned by
/zfc.
- Level-up / XP / reward loop semantics — owned by
/zfclevel.
- Root-cause-first debugging methodology for fix proposals — owned by
/root-cause-first.
If a finding would belong to one of those lanes, name the lane and hand off rather than producing conflicting feedback.
Short-circuit Clauses
- Docs-only / test-only / dependency-bump PR: Return N/A — out of scope for code-quality in the first line. Do not enumerate metrics or imports for these PRs. (Costs zero tokens and prevents noise.)
- Diff ≤ 10 changed lines and no production behavior touched: Use Short variant. If you would produce zero findings, say so explicitly rather than padding.
Variant Operational Sharpness
These constraints make the Short and Long variants operationally distinct so models can't accidentally conflate them:
Short variant — hard rules
- Max 5 Key Issues rows. If you have more, merge or drop the lowest-severity rows.
- Every row must anchor to
file:line with the offending code quoted inline.
- Do not produce refactor code blocks. One-sentence recommendation only. The PR author does the rewrite.
- No "Detailed Findings" sections. Use only the "Key Issues" / "Action Items" / "Positive Notes" / "Quick Recommendations" headings shown above.
Long variant — hard rules
- Required headings, in order:
Overview, Strengths, Issues by severity, Suggested refactors.
- Scope discipline: changed files + their direct callers only. A full-codebase scan is opt-in (
--full-scan or explicit user request). Default to changed files to avoid context-window saturation on large PRs.
- Refactor suggestions must include a code snippet (a few lines is enough) — Iron Law.
- "Issues by severity" must be grouped: Critical / High / Medium / Nice-to-have. Do not interleave.
If a model returns a Short variant that contains Long-variant headings (or vice versa), treat that as a structural violation and re-emit in the correct shape.
Cross-references
.claude/skills/code-standards/SKILL.md — dispatches code-quality alongside /zfc, /zfclevel, /root-cause-first, and /thermo.claude/skills/solid/SKILL.md — SOLID / TDD / clean-code companion.claude/skills/root-cause-first/SKILL.md — for fixes, not for code-quality findings.claude/skills/evidence-standards.md — required for $PROJECT_ROOT/ production PRs (this skill is supporting evidence, not primary)
