hyhmrright/brooks-lint

AI code reviews grounded in twelve classic engineering books.
Consistent. Traceable. Actionable.

English · 简体中文 · 繁體中文 · 日本語 · 한국어 · Español

Quick Start • The Six Decay Risks • What It Looks Like • Benchmark • Installation

→ Visit the website

"The bearing of a child takes nine months, no matter how many women are assigned." — Frederick Brooks, The Mythical Man-Month (1975)

50 years later, Brooks was still right — and so were McConnell, Fowler, Martin, Hunt & Thomas, Evans, Ousterhout, Winters, Meszaros, Osherove, Feathers, and the Google Testing team.

Most code quality tools count lines and cyclomatic complexity. brooks-lint goes deeper — it diagnoses your code against six decay risk dimensions synthesized from twelve classic engineering books, producing structured findings with book citations, severity labels, and concrete remedies every time.

For the full source-to-skill mapping, including exceptions and false-positive guards, see skills/_shared/source-coverage.md.

# Claude Code
/plugin marketplace add hyhmrright/brooks-lint
/plugin install brooks-lint@brooks-lint-marketplace

# Any other Agent Skills platform — Cursor · Codex · Gemini · Copilot · Windsurf · OpenCode · Kiro · Bob …
curl -fsSL https://raw.githubusercontent.com/hyhmrright/brooks-lint/main/scripts/install.sh | bash -s -- <platform>

Then just ask ("review this PR", "audit the architecture"), or run one of the six commands — /brooks-review, /brooks-audit, /brooks-debt, /brooks-test, /brooks-health, /brooks-sweep (what each one does).

Every finding comes back as Symptom → Source → Consequence → Remedy with a book citation and a 0–100 Health Score. Full install options (10 more platforms) and CI/CD setup are below.

brooks-lint evaluates your code across six production-code decay risks and six test-suite decay risks synthesized from twelve classic engineering books:

Philosophy of SD = A Philosophy of Software Design (Ousterhout) · SE@Google = Software Engineering at Google (Winters et al.)

Given this code:

class UserService:
    def update_profile(self, user_id, name, email, avatar_url):
        user = self.db.query(f"SELECT * FROM users WHERE id = {user_id}")
        user['email'] = email
        ...
        if user['email'] != email:   # always False — silent bug
            self.smtp.send(...)
        points = user['login_count'] * 10 + 500
        self.db.execute(f"UPDATE loyalty SET points={points} WHERE user_id={user_id}")

brooks-lint produces:

Health Score: 28/100

This method concentrates four unrelated business responsibilities into a single function, contains a logic bug that silently suppresses email change notifications, and is wide open to SQL injection.

Symptom: update_profile performs profile field updates, email change notifications, loyalty points recalculation, and cache invalidation all in one method body. Source: Fowler — Refactoring — Divergent Change; Hunt & Thomas — The Pragmatic Programmer — Orthogonality Consequence: Any change to the loyalty formula risks breaking email notifications and vice versa. Every edit carries regression risk across four unrelated domains simultaneously. Remedy: Extract NotificationService, LoyaltyService, and UserCacheInvalidator. UserService.update_profile should orchestrate by calling each — it should hold no implementation logic itself.

Symptom: user['email'] = email overwrites the old value before if user['email'] != email — the condition is always False. The notification is dead code. Source: McConnell — Code Complete — Ch. 17: Unusual Control Structures Consequence: Users are never notified when their email address changes. Silent data integrity failure — the system appears functional while violating a business rule. Remedy: Capture old_email = user['email'] before any mutation. Compare against old_email, not user['email'].

(+ 6 more findings including SQL injection, dependency disorder, magic numbers)

In Mode 2 (Architecture Audit), brooks-lint generates a Mermaid dependency graph at the top of the report. Modules are color-coded by severity: red = Critical findings, yellow = Warning, green = clean.

graph TD
    subgraph src/api
        AuthController
        UserController
    end
    subgraph src/domain
        UserService
        OrderService
    end
    subgraph src/infra
        Database
        EmailClient
    end

    AuthController --> UserService
    UserController --> UserService
    UserController --> OrderService
    OrderService --> UserService
    OrderService --> EmailClient
    UserService --> Database
    EmailClient -.->|circular| OrderService

    classDef critical fill:#ff6b6b,stroke:#c92a2a,color:#fff
    classDef warning fill:#ffd43b,stroke:#e67700
    classDef clean fill:#51cf66,stroke:#2b8a3e,color:#fff

    class OrderService,EmailClient critical
    class AuthController warning
    class UserService,UserController,Database clean

The graph renders natively in GitHub, Notion, and other Markdown environments — no extra tools needed.

The Full Gallery has real brooks-lint output across Python, TypeScript, Go, and Java — including PR reviews, architecture audits with Mermaid dependency graphs, tech debt assessments, and test quality reviews.

New to the decay risks? The Decay Risk Field Guide explains all six — diagnostic question, signature symptoms, source books, and remedy for each.

Tested across 3 real-world scenarios (PR review, architecture audit, tech debt assessment):

The gap isn't what Claude can find — it's what it consistently finds, with traceable evidence and actionable remedies every time.

The table above is illustrative. These numbers are deterministic and you can reproduce them locally:

Parser fidelity — SARIF export and the CI gates depend on parsing the model's Markdown report correctly. Against a frozen corpus of 30 real, model-generated reports spanning all six modes (evals/benchmark-corpus.json), each paired with an independently graded finding inventory (a separate model pass, spot-checked by hand), the shipped parser scores — run npm run benchmark:

Because the parser is deterministic and the corpus is frozen, npm run benchmark gives everyone the same result, and npm test guards it as a regression. The corpus deliberately includes 9 false-positive / tradeoff reports (e.g. a ports-and-adapters design that looks like a dependency cycle) that must stay clean.

Scoring determinism — for a fixed finding set (2 Critical / 3 Warning / 1 Suggestion), the strictness presets produce exactly the scores their common.md table predicts: strict 34, balanced 54, legacy-friendly 74 — and only legacy-friendly leads with the top-three fixes.

Model quality — whether the model finds the right risks on real code is measured by the 58-scenario eval suite (evals/evals.json): npm run evals (structural) and npm run evals:live (live, needs ANTHROPIC_API_KEY).

Scope & honesty: the parser numbers are deterministic and exactly reproducible. The strictness and eval-suite figures are single-run live measurements against the model and vary slightly run to run. The parser benchmark measures report-parsing fidelity (does the tooling read every finding the report states?), not whether a given finding is "correct." The severity-count match is the fully independent signal; risk-code agreement also reflects the shared canonical name→code legend.

~ = occasionally / inconsistently

brooks-lint doesn't replace your linter. It catches what linters can't: architectural drift, knowledge silos, and domain model distortion — the problems that slow teams down for months before anyone notices.

/plugin marketplace add hyhmrright/brooks-lint
/plugin install brooks-lint@brooks-lint-marketplace

Short-form commands (/brooks-review) are auto-installed on first session start — or run bash hooks/session-start yourself. To skip the marketplace: mkdir -p ~/.claude/skills/brooks-lint && cp -r skills/* ~/.claude/skills/brooks-lint/.

/extensions install https://github.com/hyhmrright/brooks-lint   # Gemini CLI

Install the brooks-lint skill from hyhmrright/brooks-lint       # ask inside a Codex session

Or use the installer below: ./scripts/install.sh gemini / ./scripts/install.sh codex.

brooks-lint ships as standard Agent Skills. Any agent that loads Agent Skills runs all six modes with no conversion — one command installs them:

# pick your platform; --project installs into the current repo instead of your global config
curl -fsSL https://raw.githubusercontent.com/hyhmrright/brooks-lint/main/scripts/install.sh | bash -s -- <platform>
#   <platform> = opencode · cursor · windsurf · antigravity · pi · kiro · copilot · droid · dsh · gemini · codex · claude · bob · agents

The installer copies the skills flat into the right folder, so the shared framework (../_shared/) always resolves — you can't get the layout wrong. Then just ask ("review this PR", "audit the architecture") and the matching skill auto-triggers from its description.

OpenCode v2, Kiro, Factory Droid, and DeepSeek Harness also auto-register /brooks-review. New to skills, or using an agent not listed? See docs/getting-started.md.

Maintainer testing policy (2026-10-16). My Claude account has been banned. Going forward, I will only test and verify Codex-related capabilities. Earlier verification of Claude Code, Gemini CLI, or other platforms is historical and does not imply ongoing testing. Existing integrations remain available, but their current behavior is not verified by the maintainer; community test reports and contributions are welcome.

🧪 Historical verification. Claude Code, Gemini CLI, and Codex CLI were previously verified by the maintainer. The ten platforms above are documented from each tool's official skill spec and verified at the file-layout level (the installer is tested), but not yet run end-to-end by the maintainer on every platform. Tried one — working or broken? Open an issue with the platform, version, and what you saw. Another Agent-Skills agent? It almost certainly works the same way — tell us and we'll add it.

Syntax by platform. Claude Code also accepts the namespaced form /brooks-lint:brooks-review — short forms are auto-installed on first session start by the session-start hook. Codex CLI uses $brooks-review. Gemini CLI and OpenCode v2 use the table as written. Cursor, Antigravity, pi, and DeepSeek Harness invoke Agent Skills from each skill's description, so just ask ("review this PR", "where's our worst tech debt?"); for explicit invocation use the platform's own syntax (pi registers each skill as /skill:brooks-review; dsh takes the table as written, from its / menu or typed inline). On every platform the skills also trigger automatically when you discuss code quality, architecture, or test health.

PR reviews include a lightweight Step 7 Quick Test Check automatically (skipped for docs-only diffs). For a full test audit, run /brooks-test; for a deep dive on any single dimension, use that dimension's own skill rather than /brooks-health.

Place a .brooks-lint.yaml in your project root to customize review behavior:

version: 1

strictness: balanced   # strict | balanced (default) | legacy-friendly — softer scoring for legacy code

disable:
  - T5   # skip coverage metrics check — we don't enforce coverage

severity:
  R1: suggestion   # downgrade Cognitive Overload findings for this domain

ignore:
  - "**/*.generated.*"
  - "**/vendor/**"

# custom_risks:   # define project-specific Cx codes — see skills/_shared/custom-risks-guide.md
# suppress:       # downgrade specific findings by risk + path (e.g. accepted legacy debt)

Copy .brooks-lint.example.yaml as a starting point. All settings are optional — omit the file entirely for default behavior.

"The complexity of software is an essential property, not an accidental one." — Frederick Brooks

AI can help you write code faster, but it can't tell you whether you're building a cathedral or a tar pit — and the decay risks these authors identified only get sharper as generation gets cheaper. Adding an AI assistant doesn't fix cognitive overload or domain model distortion; generating more code increases change propagation and knowledge duplication; moving faster makes accidental complexity and dependency disorder more dangerous.

Every skill is one SKILL.md (trigger + process skeleton) plus its own guide:

brooks-lint/
├── .claude-plugin/ · .codex-plugin/  # plugin metadata per platform
├── skills/
│   ├── _shared/          # common.md (Iron Law, config, report template, Health Score)
│   │                     # source-coverage.md · decay-risks.md (R1–R6)
│   │                     # test-decay-risks.md (T1–T6) · remedy-guide.md · custom-risks-guide.md
│   ├── brooks-review/    # Mode 1: PR Review          → pr-review-guide.md
│   ├── brooks-audit/     # Mode 2: Architecture Audit → architecture-guide.md, onboarding-guide.md
│   ├── brooks-debt/      # Mode 3: Tech Debt          → debt-guide.md
│   ├── brooks-test/      # Mode 4: Test Quality       → test-guide.md
│   ├── brooks-health/    # Mode 5: Health Dashboard   → health-guide.md
│   └── brooks-sweep/     # Mode 6: Full Sweep         → sweep-guide.md
├── hooks/                # SessionStart hook
├── commands/             # short-form command wrappers (auto-installed by the hook)
├── evals/                # 58-scenario eval suite + frozen parser-fidelity corpus
└── assets/               # logo, banner, demo

Automate brooks-lint on every PR using the GitHub Action:

# .github/workflows/brooks-lint.yml
name: Brooks-Lint PR Review
on:
  pull_request:
    types: [opened, synchronize, reopened]

jobs:
  brooks-lint:
    runs-on: ubuntu-latest
    permissions:
      pull-requests: write
    steps:
      - uses: actions/checkout@v7
        with:
          fetch-depth: 0
      - uses: hyhmrright/brooks-lint/.github/actions/brooks-lint@v1.8.0
        with:
          mode: review
          anthropic-api-key: ${{ secrets.ANTHROPIC_API_KEY }}
          fail-below: 70

See docs/github-action-example.yml for the full template.

The action posts the review as a PR comment and optionally fails the check if the Health Score drops below a threshold. If .brooks-lint-history.json is committed to your repo, the comment also includes a trend delta (e.g., "85 → 82 (−3) over last 3 runs").

mode: sweep in CI reports but never applies fixes. The action makes a single tool-less API call, so Full Sweep's consent gate (Step 0), its in-place file edits, and its iteration loop cannot run — the Fix Log and Iteration History sections come back empty, and sweep.max_iterations from .brooks-lint.yaml is inert here. What you get is a description of a sweep, not a sweep. For fixes that actually land, run /brooks-sweep interactively in a harness that has file tools and can answer the Step 0 consent prompt. The five read-only modes — review, audit, debt, test, health — behave in CI exactly as they do locally.

Quality gates and Code Scanning. Beyond fail-below, the action exposes:

        with:
          mode: review
          anthropic-api-key: ${{ secrets.ANTHROPIC_API_KEY }}
          fail-on: critical            # fail on any Critical finding (none | warning | critical)
          fail-on-regression: true     # fail if the Health Score dropped vs the last run
          sarif-file: brooks-lint.sarif  # also upload findings to GitHub Code Scanning

fail-on-regression reads .brooks-lint-history.json, so commit that file to enforce "no new regressions". Setting sarif-file makes findings appear inline on the PR's Files changed tab and requires security-events: write permission on the job.

Project config in CI. The action applies your .brooks-lint.yaml — strictness, disable, focus, severity, ignore, custom_risks — and reads it from the base branch, not from the PR's checkout. A pull request therefore cannot weaken the policy that reviews it: a diff that disables every risk would otherwise score 100 and sail through fail-below before anyone approved the config change. A config a PR adds for the first time takes effect once merged, for the same reason. The report's Config: line names the provenance it used (e.g. applied from base (main)). This needs the base object in the clone, which is why the template checks out with fetch-depth: 0; in a PR, missing base refs stop the action with an error instructing you to use fetch-depth: 0. It never falls back to the PR's policy. Outside PRs (including local runs), unavailable base refs fall back to the working-tree config with an explicit provenance note. Set config-ref: head to use the PR's own copy instead — appropriate when .brooks-lint.yaml is CODEOWNERS-protected and config changes should take effect in the PR that makes them:

        with:
          mode: review
          anthropic-api-key: ${{ secrets.ANTHROPIC_API_KEY }}
          config-ref: base   # base (default) | head

Custom API endpoint. api-base-url points the action at any Anthropic-compatible /v1/messages endpoint — a self-hosted proxy, an LLM gateway, a regional mirror — instead of api.anthropic.com. Pass that endpoint's key as anthropic-api-key and the model id it expects as model:

        with:
          mode: review
          api-base-url: https://your-gateway.example.com
          anthropic-api-key: ${{ secrets.GATEWAY_API_KEY }}
          model: gateway-model-id

brooks-lint sends your diff to whatever host you name here, so only point it at one you trust with your source. Running scripts/ci-review.mjs yourself needs no flag at all — the Anthropic SDK reads ANTHROPIC_BASE_URL directly.

Claude on Amazon Bedrock. Bedrock serves the Anthropic Messages API natively, so no new client is needed: point api-base-url (action v1.6.0+) at https://bedrock-runtime.{region}.amazonaws.com/anthropic and pass a short-term Bedrock API key as anthropic-api-key — minted inside the job from an OIDC-assumed role, so no long-lived key sits in a secret. docs/github-action-example.yml carries the full wiring as commented steps: role assumption, token minting, model ids, the IAM permissions needed, and the bedrock-mantle alternative. The trust note above still applies — with Bedrock the diff goes to your own AWS account, in the Region you name.

This path cannot review pull requests from forks. GitHub grants a forked pull_request run neither repository secrets nor id-token: write, so there is no OIDC token to exchange for the role and no stored key to fall back on. Fork coverage needs a different trigger (pull_request_target, with all the care that running against untrusted branches demands) or a different key source.

Output ceiling. max-tokens: 8192 raises the report's output ceiling (default 4096). A long sweep or health report can hit that ceiling, and truncation fails in the wrong direction: the findings that were cut off are never counted, so the Health Score comes back higher and a fail-below gate can pass a run it should have failed. You do not have to notice that yourself: a run whose report was cut off carries a ::warning:: annotation, repeats the warning in the PR comment, and sets "truncated": true in brooks-lint-report.json — raise max-tokens and re-run. The run is not failed over it.

Cost: ~$0.05–0.15 per PR run depending on diff size and model. Recommend running on pull_request events only.

Current state (v1.4): 12-book foundation, 6 production decay risks (R1–R6) + 6 test decay risks (T1–T6), 6 skills, CI quality gates, SARIF output for GitHub Code Scanning, strictness presets, and a reproducible parser-fidelity benchmark.

See CONTRIBUTING.md. The most valuable contributions right now are new eval test cases and improved decay-risk symptom patterns. Run /brooks-review on your own PR — we review contributions with the tool we're building.

MIT License — see LICENSE for details.

This project stands on the shoulders of twelve giants — see The Twelve Books above for the full list with editions. The decay risks encoded in this tool are our synthesis of their ideas, applied to modern code quality assessment.

⭐ If this tool helped you see your codebase differently, give it a star!