AOS 1.0 prerelease

Staged gates for coding agents, closed only on evidence

AOS runs inside Claude Code as a plugin. It splits work into stages and closes each one only on evidence: test runs it observed itself, reviews the agent records, and approvals only you can give.

  1. 1. In Claude Code, install the plugin

    /plugin install aos --marketplace aos-engineer/harness
  2. 2. In your terminal, install the CLI you approve gates with

    npm install -g aos-harness@next
  3. 3. Back in Claude Code, start a run

    /aos:run add a --shout flag to the greet command

Needs Claude Code 2.1.295 or later, Bun 1.3.11 or later on your PATH, and git. Full walkthrough.

A guardrail, not a security boundary

AOS keeps a cooperating agent on the process and makes skipping a step visible. It does not contain a determined one: an agent with shell access can reach the run log through processes the hooks never see, and the shell guards read command text. Every run is an append-only, hash-chained log under .aos/runs/, so an edit that breaks the chain is detected, but the chain is not signed. Read the security model.

This release supports Claude Code, and spike and bounded runs; briefs the reference pack classifies as architectural are refused. The release notes list what is not in it yet.