AgentMachinist

AgentMachinist takes a small development Task from intent to a reviewed local change. It coordinates Claude Code, OpenCode, Pi, or Codex, with human Approval of the exact Spec before implementation and human review before integration. GitHub and GitLab are optional sources of Tasks and destinations for publication.

Task → Spec commit → human Approval → implementation → verification → Review
                                                                    │
                                  human review → local integration ◄─┘
                                  optional GitHub PR / GitLab MR

The controller owns commits, Task records, and optional publication. Local integration is an explicit fast-forward operation into your clean base checkout. AgentMachinist never merges remotely or automatically.

Current release: AgentMachinist 0.14.0 on PyPI.

Install

uv tool install agentmachinist

You also need git and one supported Harness executable (claude, opencode, pi, or codex). GitHub operations require authenticated gh; GitLab operations require authenticated glab. The core CLI is tested on macOS and Linux with Python 3.12–3.14. Managed background service commands are macOS-only; on Linux, schedule machinist watch --once with your existing service manager.

machinist update-check compares the installed release against PyPI and prints the upgrade command for how this copy was installed (uv tool, pipx, pip, or a source checkout). machinist doctor reports the same result as a diagnostic row. Set MACHINIST_NO_UPDATE_CHECK=1 to suppress both probes on offline or CI machines.

Upgrading the package is not always the whole upgrade. Managed workflows are projected files, so a workflow change only takes effect once you run machinist sync-workflows. machinist watch reports that drift at startup and machinist update-check reports it alongside the release comparison, so you do not have to run doctor to find out. The advisory never blocks a command and never appears in update-check --json.

Start

Version 0.14.0 includes the guided local workflow and optional GitLab support. If you already installed AgentMachinist with uv, upgrade with uv tool upgrade agentmachinist. The existing GitHub workflow remains available.

cd your-repository
machinist start "Handle an invalid timezone without crashing" --test-cmd "uv run pytest"
# Read the saved Spec and copy the exact Approval command printed by start:
machinist approve --task T1 --spec-sha <full-spec-commit-sha>
# Approval continues implementation, verification, and independent Review.
machinist status T1
# Inspect the diff and report, then integrate explicitly:
machinist integrate T1

First start detects an installed Harness and a verification command. Its local settings and Task records live under .machinist/runs/local/, excluded through Git's local exclude file. It does not require an origin, labels, a daemon, forge authentication, or hosted workflows. Existing machinist.yaml settings remain available. Local orchestration can still use a cloud model; offline inference requires a separately configured local provider.

Verification runs in an isolated committed checkout; dependency folders from your working repository are not copied. Use a self-preparing command such as npm ci && npm test or uv run pytest. A failing baseline stops before the Spec Harness. Correct the Gate in .machinist/runs/local/config.yaml or its dependency setup, then run machinist retry --task T1 --phase spec.

You can publish the same reviewed candidate when collaboration is useful:

machinist publish T1 --provider gitlab
# Or: machinist publish T1 --provider github

Publication requires an origin that matches the selected forge and authenticated CLI. It preserves local work on failure and retries the same branch and change request without repeating Harness work or verification. GitLab support includes nested projects and explicitly bound self-managed hosts; it covers issue intake and merge-request publication, not GitLab-hosted Spec automation.

See the local workflow guide for input files, external issues, amendments, recovery, and use by a solo developer or small team.

GitHub automation

Existing GitHub issue commands, trusted workflow Approval, and watcher operation remain available. Configure that integration separately:

cd your-repository
machinist onboard
# Answer the setup questions, review the generated files, then:
git status --short
git add machinist.yaml .machinist/specs/.gitkeep .gitignore
git add .github/ISSUE_TEMPLATE/agentmachinist-task.yml
git add -p .github/workflows   # review each hunk
git diff --cached              # verify what will be committed
git commit -m "chore: configure AgentMachinist"
git push
machinist doctor --run-gates   # after setup reaches the default branch
machinist watch

machinist doctor --run-gates is the single health check — it already verifies labels, workflow drift, the sealed issue form, and verification gates, and prints the exact fix for any FAIL. Only run the individual machinist sync-labels --check, machinist sync-workflows --check, or machinist task template --check if doctor asks for them.

Use machinist onboard --setup-pr when you want AgentMachinist to put only its managed setup files on a pushed chore/agentmachinist-setup branch and open a draft PR. Fresh setup requires a clean default branch; a recognized partial setup resumes without replacing your preferences. Setup checks local readiness before publication. Merge setup, then run full doctor to verify the deployed workflows. machinist rehearse exercises the production local Phases, real Git, verification, and integration with a fake Harness; --harness explicitly opts into configured providers in the disposable repository.

In a terminal, machinist onboard (the GitHub setup entry point) asks a short set of setup questions — dispatch mode, managed workflows, harness, test gate, and notifications — each with a one-line explanation and a safe default. Flags such as --harness, --test-cmd, --spec-source, and --notifications pre-answer their questions; --yes accepts all safe defaults and auto-enables the detected test command for hands-free quickstart; --no-input (or a non-interactive shell) also skips questions but does not auto-enable the test command unless you pass --test-cmd. Errors outside a configured repo now point you to machinist onboard and the first-run guide. machinist init is the same setup step without the guided receipt — prefer onboard for new repositories. machinist --help groups commands as Setup, Tasks, Build, and Operate — daily vs Operate — advanced.

Review the staged diff before committing. The managed Task form is installed even when Actions workflows are externally managed. Managed workflows must be pushed before GitHub comment or label approval can record SHA-bound evidence. machinist init also adds /.machinist/runs/ to .gitignore. If you manage workflows yourself, machinist init --no-workflows records github.manage_workflows: false; doctor then reports that its drift check was intentionally skipped.

The default github.spec_source: local makes watch own spec generation. Choose github-actions and run machinist sync-workflows if CI should own that phase instead. Exactly one source is active, preventing duplicate spec runs. Managed CI installs the selected Spec adapter and reads its declared secret name; built-ins support Claude Code, Codex, OpenCode, and Pi.

Approval is bound to the exact PR head commit. Use either:

machinist approve --issue 57
# or: machinist approve --pr 18
# or post the SHA-bound comment shown in the Spec PR body:
# /machinist-execute <full-spec-commit-sha>

GitHub Approval takes exactly one of --issue or --pr; local Approval instead uses --task T1 --spec-sha <sha>. These selectors keep local Tasks, issues, and pull requests distinct.

Editing the spec after approval makes that approval stale and blocks execution until the new head is approved.

With the starter configuration, a successful Execute run leaves the PR draft for an independent read-only Review Task Run. Review compares the approved Spec, diff, and verification evidence, posts a structured advisory report, and alone marks the exact implementation head ready. Findings never trigger an automatic repair or merge. Use machinist review <issue> manually or machinist retry <issue> --phase review after a failed Review.

The CLI approval command submits the SHA-bound comment; the managed GitHub workflow independently verifies the current head and approver's write access, then records the marker and label. Until that workflow completes, status remains awaiting approval rather than pretending execution is authorized. approval pending specifically means the label is visible but trusted SHA Evidence has not arrived yet.

Revise or explicitly abandon a successful Spec by issue number:

machinist spec 42 --dry-run
machinist spec 42 --revise
machinist spec 42 --abandon --reason "requirements changed"

The dry run prints the proposed Spec without commits, pushes, or a PR. Revision regenerates the Spec on its existing branch and draft PR. Abandonment records the reason, removes the trigger and approval labels, and closes the open draft PR.

Commands

Documentation

• TL;DR
• Getting started
• Local workflow and optional publication
• GitHub visual first-run field guide
• Architecture and lifecycle
• Operator runbook
• Trust model
• Harness support
• Architecture decisions
• Contributing
• Changelog

The trust model is deliberately narrower than “the agent cannot use git.” Harness flags, credential reduction, repository postconditions, and push leases reduce risk, but local harnesses still execute with the operating-system access of the user who launched them. Read the trust model before unattended use.

Releasing

Releases use PyPI Trusted Publishing. Bump pyproject.toml, update the changelog, and publish a GitHub Release tagged v<version>. The release job first checks tag/version equality, runs tests and workflow checks, builds both distributions, smoke-tests the installed wheel and sdist through a generated first-run project, and records SHA-256 hashes. A minimal job then publishes those verified artifacts. Only after publication do separate jobs attach the distributions and checksum file to the GitHub Release and verify that the exact version is visible and installable from PyPI.

License

MIT — see LICENSE.