agent-native-repo-playbook

Run a harness audit that inspects root and nested AGENTS.md files, docs structure, check scripts, and execution surfaces. The audit identifies where a cold agent would stall and returns prioritized gaps with concrete file evidence.

Use the harness readiness rubric, which scores seven dimensions from 0 to 4, including context routing, autonomous execution, mechanical guardrails, and proof of work. The output includes an overall score out of 100, a dimension table, and top blockers.

Root AGENTS.md should be a short router with universal rules and pointers to canonical docs, while nested files only cover local boundary rules that differ. Each line should be universal, operational, enforceable, and durable, otherwise it moves to docs or gets deleted.

It is designed for solo-developer, high-permission workflows where agents write code and docs while humans set intent. It deliberately avoids team-heavy approvals, branch ceremony, and centralized policy layers unless a concrete risk requires them.

Direct-to-main is the default for trusted solo repos with fast recoverable checks, but approval or branch processes are appropriate when they meaningfully reduce real risk, such as secrets handling or irreversible external effects.