pattern8 0.2.0

Pattern 8 (P8) — AI Agent Governance Framework

Pattern 8 (P8)

🎱 AI Agent Governance Framework

P8 constrains how AI Agents (Claude, Cursor, Gemini, etc.) behave in your project. Law (SKILL files) + Police (code engine) + Zero-Trust (Hook + Rules) = Agents can't jailbreak.

Architecture

Developer-editable (Law)              Read-only Engine (Police)
┌──────────────────────┐          ┌──────────────────────────┐
│ SKILL.md             │          │ SecurityGuard            │
│ checklist.yaml       │  read →  │  ↳ regex blacklist       │
│ template.yaml        │          │  ↳ path restrictions     │
│ guidelines.yaml      │          │ Reviewer                 │
│ security.yaml        │          │  ↳ static rule engine    │
│                      │          │  ↳ P8AuditError rollback │
│ "Constitution"       │          │ "Police"                 │
└──────────────────────┘          └──────────────────────────┘
                ↕ Agent calls via MCP ↕

Three Defense Layers

Layer	Method	Can Agent bypass?
Layer 1	AGENTS.md + Cursor Rules prompt injection	⚠️ Theoretically yes
Layer 2	MCP Tools (submit_review / execute_tool)	❌ Code-enforced
Layer 3	Git pre-commit hook	❌ Impossible to bypass

Progressive Disclosure (aligned with Google ADK)

Agent loads lightweight Resources on startup; audit rules are loaded internally on demand:

Type	Interface	Visible to Agent?
Resource	skill://index — SKILL list	✅
Resource	skill://{name}/checklist — Entry checklist	✅
Resource	skill://{name}/template — Output template	✅
Tool	submit_review — Submit for audit	✅ Interface visible
Tool	execute_tool — Request execution	✅ Interface visible
🔒	guidelines.yaml — Audit rules	❌ Loaded internally
🔒	security.yaml — Security red lines	❌ Loaded internally

Installation

# Basic install (CLI + SKILL management)
pip install pattern8

# Full install (with MCP enforcement server)
pip install 'pattern8[enforcement]'

Quick Start

# 1. Initialize P8 (auto-installs 5 components)
p8 init my-project
#   ✅ AGENTS.md              — Global agent instructions
#   ✅ skills/ (5 SKILLs)     — Built-in governance rules
#   ✅ .gitignore
#   ✅ .git/hooks/pre-commit   — Commit-time audit
#   ✅ .cursor/rules/          — Forces agent to use MCP

# 2. Manage SKILLs
p8 list                         # List all SKILLs
p8 new my_custom_skill          # Create custom SKILL
p8 validate skills/example      # Validate SKILL integrity

# 3. Start enforcement server
p8 serve                        # Start MCP enforcement server
p8 mcp-config --client cursor   # Generate Cursor MCP config

Connect to Cursor

p8 mcp-config --client cursor

Paste the output into .cursor/mcp.json:

{
  "mcpServers": {
    "pattern8": {
      "command": "p8",
      "args": ["serve"]
    }
  }
}

5 Patterns × 5 Built-in SKILLs

Pattern	Controls	Config File
Pipeline	Step execution order	SKILL.md
Inversion	Blocks if info is missing, asks user	checklist.yaml
Generator	Output format constraints	template.yaml
Tool Wrapper	Dangerous command interception	security.yaml
Reviewer	Output quality audit	guidelines.yaml

SKILL	Use Case
example	PRD document generation
code_review	Code review (correctness / security / performance / maintainability)
bug_fix	Bug fixing (locate → root cause → fix → regression)
refactor	Code refactoring (functional equivalence guarantee)
feature_dev	Feature development (requirements → design → implement → verify)

SKILL Directory Structure

skills/code_review/
├── SKILL.md                   # Pipeline + <HARD-GATE> blockers
├── assets/
│   ├── checklist.yaml         # Inversion entry checklist
│   └── template.yaml          # Generator output template
└── references/
    ├── guidelines.yaml        # Reviewer audit rules (🔒 invisible to Agent)
    └── security.yaml          # Tool Wrapper security red lines (🔒 invisible to Agent)

License

MIT