Anvil

Execution Intelligence for AI Agents

AI agents can generate code. But they can't explain why your build just broke. Anvil can.

What It Does

• Constitution — Immutable trust anchor. Rules only humans can change
• Gate Chain — Script-based quality gates. Exit code = verdict
• Heritage — ADR + FMEA. Your project remembers decisions and failures
• Trace — Capture what the agent actually did
• Insight — Explain why things broke
• Pattern — Compress recurring failures into rules
• Guard — Learned rules applied at runtime
• Sandbox — Isolated execution environments
• Replay — Reproduce and compare past executions

The Core Loop

run → observe → understand → improve → run

Install

claude plugin add anvil-claude

Quick Start

/anvil-claude:init

This creates the .anvil/ directory with governance files, then guides you through setting up your project's constitution.

Skills

| Skill | Command | Purpose | |-------|---------|---------| | Init | /anvil-claude:init | First-time project setup | | Constitution | /anvil-claude:constitution | View/verify trust anchor | | Self-Model | /anvil-claude:self-model | Project identity and state | | ADR | /anvil-claude:adr | Record architecture decisions | | FMEA | /anvil-claude:fmea | Record failure patterns | | Gate Chain | /anvil-claude:gate-chain | Run quality gates | | Sprint | /anvil-claude:sprint <quest> | Build cycle | | Sprint | /anvil-claude:sprint | Heal/maintain cycle | | Structure | /anvil-claude:structure | Code tier classification | | Commit | /anvil-claude:commit | Commit with conventions | | Research | /anvil-claude:research | Evidence-based research | | Health | /anvil-claude:health | Quick project scan | | Trace | /anvil-claude:trace | Execution trace capture | | Insight | /anvil-claude:insight | Failure analysis | | Pattern | /anvil-claude:pattern | Failure pattern detection | | Guard | /anvil-claude:guard | Adaptive runtime guardrails | | Sandbox | /anvil-claude:sandbox | Isolated execution | | Replay | /anvil-claude:replay | Execution replay |

How It Works

User input → skill-detector.sh → matches trigger → loads SKILL.md as context
                                                  → agent follows skill instructions

1. Hooks (hooks/hooks.json) — run on session start, user prompt, and tool use
2. Scripts (scripts/*.sh) — detect triggers, inject context, guard constraints
3. Skills (skills/*/SKILL.md) — instructions the AI follows for each capability
4. Gate scripts (skills/*/scripts/*.sh) — executable checks with exit 0/1

Philosophy

1. Environment is intelligence — the quality of the substrate determines the capability of the agent
2. Design physics, not control — design environments, not commands
3. Success is becoming unnecessary — as trust accumulates, intervention decreases

License

MIT