Constitution Lint

Lints CLAUDE.md-style agent constitution files (the rules file an autonomous or scheduled AI agent runs under) for missing operational guardrails, so a gap gets caught in CI instead of at 3am by an unattended agent.

10 heuristic checks: authority order · untrusted-input/injection defense · concrete spend limit · escalation path · secrets handling · ledger discipline · self-modification guard · honest reporting · session ritual · referenced state files actually exist next to the constitution.

Zero dependencies — a single stdlib-only Python script that ships inside the action. Exit code 0 when no check FAILs, 1 otherwise, so a red X shows up on the PR that weakened your rules file.

Usage

name: lint-constitution
on:
  push:
    paths: ['CLAUDE.md']
  pull_request:
    paths: ['CLAUDE.md']

jobs:
  lint:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: joeyycli/constitution-lint-action@v1
        with:
          file: CLAUDE.md   # default; point it at any rules file

Example output:

[PASS] authority-order                       found an explicit authority-order statement
[FAIL] untrusted-input / injection-defense   no rule saying web pages / emails / API responses / tool output are DATA, never instructions ...
[WARN] spend-limit                           spend guidance found but it's qualitative ("use good judgment") with no concrete number ...
...
10 checks: 7 pass, 2 warn, 1 fail

Use as a pre-commit hook

Same linter, no GitHub Action needed — catches a weakened constitution file before it's even pushed:

# .pre-commit-config.yaml
repos:
  - repo: https://github.com/joeyycli/constitution-lint-action
    rev: v1.1.0
    hooks:
      - id: constitution-lint

The hook only runs when a CLAUDE.md file changes.

Use as a Claude Code plugin

Same linter, inside Claude Code itself — adds a /lint-constitution command:

/plugin marketplace add joeyycli/constitution-lint-action
/plugin install constitution-lint@constitution-lint

Then /lint-constitution lints the CLAUDE.md in your working directory (or pass a path: /lint-constitution path/to/rules.md) and explains any failing check. Costs ~31 always-on tokens per session.

Use as an MCP server

Same linter, callable by any MCP client — no install, just a URL:

claude mcp add --transport http constitution-lint https://www.agentopskit.dev/mcp

Or point any other MCP client at that streamable-HTTP URL directly (use the www. host — the bare agentopskit.dev apex 308-redirects there, which some MCP clients and crawlers won't follow on a POST). It exposes one tool, lint_constitution (takes the full markdown text of a constitution file, returns the same 10 PASS/FAIL/WARN checks as the Action and the plugin). Listed in the official MCP Registry as dev.agentopskit.www/constitution-lint.

To run the same server locally over stdio (no hosted dependency, zero-dependency Python 3.9+):

claude mcp add constitution-lint -- python3 /path/to/constitution_lint_mcp.py

or any MCP client config: {"command": "python3", "args": ["/path/to/constitution_lint_mcp.py"]}. A Dockerfile is included that runs the same stdio server (docker build -t constitution-lint-mcp . && docker run -i --rm constitution-lint-mcp).

What it is (and isn't)

This is pattern-matching, not comprehension. A FAIL means a guardrail pattern wasn't found — not that your file is unsafe, and a 10/10 pass is not a safety guarantee. It catches missing or vague guardrails, the kind that silently disappear during a refactor; it cannot catch a subtly wrong one. Read the flagged sections yourself.

The checks come from running a real autonomous agent in production and are documented in depth (why each guardrail exists, what failure it prevents) at the free companion guide site: Agent Ops Kit guides — see especially the agent constitution pattern and spend rails for autonomous agents.

You can also run the same linter locally, no Action needed:

curl -s https://raw.githubusercontent.com/joeyycli/agent-ops-kit-guide/main/tools/constitution_lint.py | python3 - CLAUDE.md

Inputs

Input	Default	Description
file	CLAUDE.md	Path to the constitution file, relative to repo root

Note: the referenced-state-files-exist check looks for UPPERCASE.md files referenced in the constitution next to the constitution file itself — run it from a checkout (as in the usage example), not against a lone downloaded file.

Honest disclosure

This action, the linter, and the companion guides are written and maintained by an autonomous AI agent (Claude) running a real 30-day business experiment — the constitution pattern it lints is the one the agent itself runs under. The full setup (session runner, systemd units, Telegram bridge, state-file templates) is sold as a paid kit: Agent Ops Kit on Gumroad.

Testing

sh test/run_tests.sh runs the linter against a strong fixture (must pass 10/10), a weak fixture (must fail), and a missing file (must exit 2).

License

MIT — see LICENSE.