Sahek/lean-agent-governance

Ultra-Lean Agent Governance Framework (v4)

A production-ready, drop-in governance layer for AI coding agents (Claude Code, Cursor, Aider).
Copy the files into any repository to establish sandbox enforcement, boundary rules, and CI pipeline checkpoints.

What's Included

File	Purpose
CLAUDE.md	Input firewall for Claude Code
.cursorrules	Input firewall for Cursor Agent
.aiderignore	Vision barrier for Aider
.claude/settings.json	Execution guard — banned/allowed bash operations
scripts/run-agent.sh	Docker sandbox — network, memory, and secret isolation
hooks/pre-commit	Local gatekeeper — diff volume + core architecture protection
.github/workflows/agent-governance.yml	Remote CI validator — authoritative PR audit

Note: hooks/pre-commit should be copied to .git/hooks/pre-commit in your repo and made executable with chmod +x .git/hooks/pre-commit.

Defense Layers

Layer	Mechanism	What It Stops
Input Firewall	.cursorrules / CLAUDE.md	Prompt injection, architectural drift
Execution Guard	.claude/settings.json + Docker	Fork bombs, disk abuse, subshell escapes
Network Guard	--network none in Docker	curl|bash downloads, data exfiltration
Secret Guard	Null bind mounts	.env / SSH key exposure to model context
Local Gatekeeper	Pre-commit hook	Bad agent refactors sneaking into commits
Remote Validator	GitHub Actions	Hook bypass via --no-verify

Quick Setup

# 1. Copy all files into your repo root
# 2. Make scripts executable
chmod +x scripts/run-agent.sh
cp hooks/pre-commit .git/hooks/pre-commit
chmod +x .git/hooks/pre-commit

# 3. Run an agent session inside the sandbox
./scripts/run-agent.sh claude   # or aider

Human Override

The pre-commit hook and CI both support an explicit architect override:

ALLOW_CORE_CHANGES=true git commit -m "Architect manual override: update core types"

---

Published by @sahek.