Home

agentic-harness

A small, opinionated harness for doing production-quality engineering with AI coding agents — Claude Code, Antigravity, Codex, Gemini CLI, and any tool that reads AGENTS.md. Six phase-gated slash commands, two sub-agents, deterministic verification, on-disk state. This wiki is the project's own documentation, maintained by the documenter sub-agent at phase boundaries.

⚡ Quick Reference

Question	Answer
What is this repo?	Product-Intent
How do I install it into a project?	Getting-Started
How do I cut a release?	Runbook — "Cutting a release"
How does install.sh --update behave?	Runbook — "Updating an installed harness"
Where does the code live?	Overview
Why phase-gated?	ADR 0001
Why phase-boundary docs (not inline)?	ADR 0002

🛠 Development

Build, install, and contribute to the harness itself.

• Getting-Started — install the harness into a target project; run the local test suite.

📟 Operational

Run, release, and maintain the harness.

• Runbook — install.sh --update, cutting a release with ship-release, CI gates, the dogfood-freshness check.

🎨 Design

What the harness is for, who it's for, and why it's shaped the way it is.

• Product-Intent — problem statement, target user, non-goals.

🏗 Architecture

How the pieces fit together.

• Overview — repo layout: harness/ (canonical specs), adapters/ (per-tool shims), templates/ (what install.sh drops into a project), scripts/ (test infra, never propagated).
• ADR 0001: Phase-gated workflow — why setup → plan → work → review → release with hard boundaries and one task per /work session.
• ADR 0002: Documentation convention — why docs live in wiki/, are written by a sub-agent at phase boundaries, and why the installer boundary keeps this repo's own wiki/ out of target projects.