Hanesu

Hanesu installs into a repository via `npx create-hanesu` and adds a control layer that AI coding agents execute against. Rather than receiving one long prompt and running to completion, an agent working through Hanesu follows a sequence of task files organized into phases: search and discovery, artifact writing, phased coding, quality gate checkpoints, and a final verification step. The agent doesn’t decide when to move forward — the gate structure does, with human review required at configured checkpoints before the next phase unlocks.

The differentiating idea is the artifact trail. Each phase produces written outputs — decision records, findings, plans — that persist in the repo. For audit-sensitive or security-related changes, this means the reasoning behind a refactor isn’t buried in a chat window; it’s a committed file. Teams doing compliance work or post-incident reviews can read exactly what the agent considered before it touched anything.

Hanesu is explicitly scoped to ambiguous, multi-step, or risky work. The docs describe it as sitting inside your repo alongside agent runtimes rather than replacing them, which means it adds surface area to maintain. For teams already running Claude Code or Codex with their own prompt conventions, adopting Hanesu means learning a second layer of abstractions and keeping both in sync. Early-stage repositories with evolving structures will feel that friction most sharply.

The project is MIT-licensed with public source on GitHub, self-hosted by definition since it runs locally inside your repo, and carries no API dependency of its own — behavior is determined by whichever agent runtime you point at it.