Contents

x0ne OS / Handbook / Project

Maintained by agents

Works today (measured 2026-10-09): the repository ships AGENTS.md, machine-readable results and a deny-by-default harness policy.

Planned — later: the policy loaded by a real x0ne harness binary, and the maintainer agent pipeline.

AI agents will maintain x0ne OS continuously: x0ne harness agents first, Claude as a fallback. Humans will sign every release.

The pipeline

  1. Watch (agent): Debian security advisories, upstream versions, CVEs.
  2. Rebuild (agent): in a disposable container.
  3. Test (agent): in a VM.
  4. Propose (agent): a change with logs, SBOM and build comparison.
  5. Sign off (human): review the diff and the hash.
  6. Publish (human): sign offline, publish.

AGENTS.md

The repository tells agents how to work on it. Its hard rules:

  • Never read, copy or print a signing key.
  • Everything runs in containers. No sudo, no loop device.
  • Tests first: show the test fail, then make it pass.
  • A test for something not possible yet is an expected failure with a reason. It is never deleted.
  • No network except Debian mirrors and pinned registries.
  • Agents never push. A human publishes.

The harness policy

The agent runs under a policy that denies by default. It may run git, make and podman, in the repository only. Key files are denied even there. Every change goes through a human gate.

Machine-readable state

what an agent reads
$ make status
$ ls out/results/
acceptance.json  auto.json  autoi.json  byoi.json  iso.json  public.json  security.json

Exit codes: 0 success, 1 a test failed, 2 a build error.

What an agent never does

  • sign anything;
  • add a package, an upstream source or a trusted key;
  • change the kernel configuration or the apt policy;
  • touch the infrastructure.

See also