code-quality-skills-kit
servas-ai/code-quality-skills-kit
Self-bootstrapping multi-agent code-quality audit kit for Claude Code. Universal — works on any codebase. AI-friendly compact output, 15 audit dimensions, parallel sub-agents.
Instalação
npx convez-ai servas-ai/code-quality-skills-kit1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
README
A tradução para o português deste README ainda está na fila.
code-quality-skills-kit
Self-bootstrapping multi-agent code-quality audit kit for Claude Code. Universal — works on any codebase.
One prompt. Any language (TS / Python / Rust / Go / Java / Ruby / Elixir / polyglot). 15 audit dimensions running in parallel via sub-agents. Token-optimised compact output. Single source of truth.
One-line install
curl -sSL https://raw.githubusercontent.com/servas-ai/code-quality-skills-kit/main/install.sh | sh
Or manual:
git clone --depth 1 https://github.com/servas-ai/code-quality-skills-kit /tmp/cqsk
cp /tmp/cqsk/code-quality-checker.md ./
cp -r /tmp/cqsk/code-quality-skills ./
echo "audit-reports/" >> .gitignore
Then in Claude Code:
/code-quality-checker # whole repo
/code-quality-checker src/foo # scope to a path
/code-quality-checker '#42' # PR audit
/code-quality-checker --no-tools # greps only, skip typecheck/test/build
v4: External Logic Offloading (MCO + Static Pre-Filter)
v4 removes ~80% of bash from the kit (≈3000 → 602 LOC) by delegating each layer to a single, dedicated tool. Four bins, each ~150 lines, replace the v3 monolith:
| Bin | Layer | Delegates to |
|---|---|---|
cqc-prefilter | L0 — static pre-filter (parallel) | Knip + jscpd by default; Semgrep + ast-grep optional |
cqc-orchestrate | L1 — multi-CLI review across providers | mco review (Multi-CLI Orchestrator) |
cqc-score | L3 — KPI snapshot / before-after diff | cloc, jscpd, ts-complex, knip |
cqc-autofix | L3 — closed-loop dup-removal PRs | mco run in a git worktree |
Pre-filter runs all 4 tools in parallel and writes clean-files.txt; cqc-orchestrate uses it to skip files with zero static findings, so the LLM layer only sees dirty paths.
When cqc-orchestrate starts (anything but --plan-only), cqc-watch boots cqc-ui on :4020 if not already up and opens it in the default browser (or prints the URL when no DISPLAY). The dashboard shows a per-run files_done / files_total progress bar plus each CLI's daily and weekly plan limits pulled from subscription.limits.
Example:
cqc-orchestrate --scope . --max-parallel 10
What you get
- 15 dimensions — D1 Correctness · D2 Types · D3 Tests · D4 Security · D5 Performance · D6 Architecture · D7 A11y · D8 Dead Code · D9 Docs · D10 CI Health · D11 i18n · D12 Deps · D13 Cache · D14 CSS Tokens · D15 Feature Flags
- Parallel sub-agents in 5 ordered groups (quality first, security last)
- Token-efficient compact format —
@file:line excerpt(≤72 chars per citation), JSONL findings (~200 bytes/finding) - Critical-Cap scoring — 1 critical → score ≤40 (no security-by-bargaining)
- Single source of truth —
_run.json(state) +_findings.jsonl(append-only) — parallel-safe, resumable - AI-feed output —
REPORT.compact.txt≤4 KB for 50 findings, paste-able into any other agent - Coverage gate — ≥99 % file coverage enforced; inline sweeper catches unowned files
- Universal stack adapter — auto-detects toolchain (8+ stacks);
cqc.config.yamlfor explicit override
Run output
audit-reports/<YYYY-MM-DD>__<short-sha>/
├── _run.json ← state · agents · files · stats · hotspots
├── _findings.jsonl ← append-only NDJSON (one finding per line)
├── _profile.yaml ← detected stack + invariants
├── _status.txt ← live status (rewritten every 5 s)
├── d1.md … d15.md ← per-dimension reports (≤200 lines each)
├── REPORT.md ← human markdown
├── REPORT.compact.txt ← AI-feed format (~2.5 KB / 50 findings)
├── dashboard.html ← one-page HTML, no JS, ≤2 KB minimal
└── fix-prompts.md ← tag-clustered paste-able fix prompts
How it stays simple
- One file is the single source of truth —
code-quality-checker.mdat the repo root contains everything: orchestrator workflow, shared sub-skill contract, all 15 skill templates, output formats. The on-diskcode-quality-skills/folder is a derivable cache regenerated on every run. - Two coordination files only —
_run.json(atomic read/modify/write with 3-retry guard) +_findings.jsonl(append-only, parallel-safe). - 15 sub-skills, one shared contract — every sub-agent does CLAIM → AUDIT → REPORT → CHECK OUT against the same two files.
- No external dependencies — works with bash + jq + rg + git.
Dispatch order (intentional)
GROUP 1 — Quality (parallel): d1 d6 d8 d9 d13 d14 d15 ← FIRST
GROUP 2 — Types & Tests (sequential): d2 d3
GROUP 3 — Performance (sequential): d5 d10 d12
GROUP 4 — UX & i18n (parallel): d7 d11
GROUP 5 — Security (sequential): d4 ← LAST
Code-quality runs first to give context. Security runs last so its findings headline the report.
Universal stack adapter (auto-detect)
| Detected | Typecheck | Test | Build | Lint |
|---|---|---|---|---|
package.json + pnpm | pnpm typecheck | pnpm test --run | pnpm build | pnpm lint |
package.json (npm/yarn) | npx tsc --noEmit | npm test | npm run build | npx eslint . |
pyproject.toml | mypy . | pytest | python -m build | ruff check . |
Cargo.toml | cargo check | cargo test | cargo build --release | cargo clippy |
go.mod | go vet ./... | go test ./... | go build ./... | golangci-lint run |
pom.xml / build.gradle | mvn compile / gradle compileJava | mvn test / gradle test | mvn package | mvn checkstyle:check |
Gemfile | bundle exec srb tc | bundle exec rspec | — | bundle exec rubocop |
mix.exs | mix compile --warnings-as-errors | mix test | mix release | mix credo |
Polyglot repos: all detected stacks run; results aggregate.
Customise per project (optional)
Drop cqc.config.yaml in repo root. See cqc.config.example.yaml for the schema.
language_policy:
ui_strings: de # preserve German UI strings
invariants:
- name: All API calls go through src/lib/api.ts
verify: rg "fetch\\(" src/ | grep -v "src/lib/api.ts"
scope_blocklist: [billing, admin]
Inspirations (skills.sh + community)
- skills.sh — open agent skills marketplace (Vercel, 2026)
- Anthropic skills — frontmatter convention
- Trail of Bits skills — security-domain skill folder layout
- codeprobe — multi-agent orchestrator with per-skill reports
- code-health-check — universal SKILL.md format
- Code Auditor (Grade A) — 6-dimension rubric
- FlowX audit-completeness + auditable agentic AI (ACM 2025) — 99 %-coverage / hash-linked entry patterns
Live demo
servas-ai.github.io/code-quality-skills-kit — interactive landing page with live demo, install tabs, configure form.
Status
v3.2 · End-to-end tested against a real Next.js codebase (211 files, 3 parallel sub-agents, all findings traced to real file:line citations, Critical-Cap scoring verified).
License
MIT — see LICENSE.