Builder
Spec-driven development for coding agents, where the host is the only judge of done.
What it is
I built Builder for my own projects after watching agents report success on work that had never run. A change moves through six phases: specify, design, review, plan, implement and verify. The first four stop for my approval; implement and verify run on their own.
How it works
- The host decides
- Verify runs the declared commands on the host and gates on their exit code. What the agent reports is a claim. The repo has a demo of an agent saying SUCCEEDED and being rejected.
- Where it runs
- Prompts and a small CLI for GitHub Copilot Chat, Claude Code and Codex. An optional validator, Python standard library only, checks the task structure and the red/green evidence.
- What it does not prove
- The red half of test-first is still reported by the agent, so it is trusted rather than proven. The README says the same.
What is true today
Public on GitHub under Apache-2.0, at version 0.3.1. Most of what I build goes through it. The walkthrough covers the phases and the gates in detail.