Skip to content

Releases

unsupervised v0.1.1

Choose a tag to compare

@github-actions github-actions released this 15 Nov 20:14

unsupervised v0.1.1

Added

  • Software Factory: references/factory.md covers choosing local or Cloud mode
    (the user's own session vs Mutagent Cloud's sandboxes), linking a GitHub
    Project board and repositories during onboarding, what factory.yaml holds
    (board, repositories, label roles, merge preset, autonomy), running work
    locally (/sync, /spec -> /publish -> /dispatch, /execute) and on
    Cloud (architech factory push, enable, runs), following runs, relaying
    gates and escalations, common failures, and the safety rules: the agent
    never adds or removes a person's label, never approves a gate, never merges.
    Its CLI is architech factory, not mutagent factory (the Agent Factory).
    SKILL.md points to it, and the description and trigger cases cover it.
  • The description rules out the Agent Factory (mutagent factory), so a
    request about it no longer loads this skill.

Changed

  • README rewritten around the Software Factory: what you get, how it runs
    locally or on Mutagent Cloud, the install lines, and what stays yours.

Fixed

  • /execute is described as it works: it plans, builds, has the pull request
    reviewed, then merges as your merge rules allow, so in Fast Mode the merge is
    what you approve. Agents tell the user before they start it and relay its
    merge question, and scripts/run-headless.sh refuses an /execute prompt
    as it refuses /ship.
  • The label table: merge:hold stops work whoever adds it. runs approve --request-changes (in --help, not yet in the docs) is listed with the
    other user-only approval commands. The README says onboarding commits two
    files.
  • plugin.json declares "skills": ["./"], which older Claude Code releases
    accept (a bare "." was refused).
  • The Release workflow publishes only after Validate passed on the same
    commit; a manual release needs a passing Validate run on the tag's commit.
  • CI pins the Claude Code version used for claude plugin validate.

unsupervised v0.1.0

Choose a tag to compare

@github-actions github-actions released this 15 Nov 19:02

unsupervised v0.1.0

Added

  • Harness-neutral: works from any agent harness that can read files and run
    shell commands, in three phases (start, setup, operate), with capability
    checks, a short per-host table, and fallbacks without tmux or without a
    shell. AGENTS.md and a README "For agents" section point other harnesses
    to the same steps.
  • First release: install Architech with the public installer, onboard a
    repository, choose a flow (Full-Loop or Fast Mode), and hand work to
    Architech from Claude Code or Codex, headless (scripts/run-headless.sh) or
    in a background tmux session that can ask questions.
  • Background and detached runs start a private tmux server with -f /dev/null,
    so none of the user's tmux config (session restore or autosave plugins) runs
    in it. The runner is always called by its absolute path in the skill folder.
  • A run stopped from outside (its process group got a signal) is graded
    INTERRUPTED, not FAILED; dispatch.md shows how to stop a run that is
    still going by its recorded group.
  • Where to find the skill's folder when a harness shows it only by name or URI,
    without searching the home folder or re-typing the runner.
  • Write or commit questions, and a choice of how to land a change, are relayed
    to the user like plan and merge questions.
  • run-headless.sh writes pgid.txt, so a run whose runner was killed outright
    (no verdict, runner gone) can be stopped exactly; the docs cover that outcome.
  • run-headless.sh stops Architech and writes INTERRUPTED when it is itself
    stopped, refuses a /ship prompt, and is meant to run in the background.
  • Background sessions run on their own tmux server (-L unsupervised) with
    exact targets.
  • Checked against Architech 0.1.49. On macOS: headless runs cannot ask and end
    BLOCKED when they need a decision; a background tmux session shows a
    question as a menu the calling agent can answer; journeys in Claude Code and
    in Codex ended with a headless run that pushed an unmerged branch. Codex's
    default sandbox asks you to approve each tmux -L unsupervised call,
    architech update --check and pgrep, and in one test it answered two
    approval questions itself, which the relay rules above now forbid. On Linux:
    a clean install, doctor --strict and the offline test suites passed on
    Ubuntu 24.04 arm64, with no model run yet. Linux x64 and other harnesses are
    expected to work, not yet tested.