Skip to content

AgentMachinist 0.20.0

Choose a tag to compare

@vscarpenter vscarpenter released this 06 Oct 20:41
6496d5c

AgentMachinist 0.20.0 makes the local workflow easier to learn and inspect, and improves recovery before implementation.

Learn the workflow before your first Task

  • machinist rehearse --guided pauses at the saved plan and reviewed candidate so you can practice Approval and acceptance for free. It runs from an installed package without a source checkout.
  • machinist doctor --local --fresh-workshop checks committed code and dependencies in a disposable checkout before a paid run.
  • CLI help and first-run documentation lead with the local plan, review, and integration journey.

Make informed decisions and recover safely

  • machinist revise T1 --feedback "..." revises an initial plan while preserving Task identity, prior Specs, and feedback history. Every revised plan needs fresh Approval.
  • machinist inspect T1 [--json] brings the plan, candidate diff, checks, Review, and next action into one read-only view.
  • machinist config show/set --local makes saved local settings and their effects explicit.
  • Local retries can explicitly select another Harness or model without changing saved preferences.

Reliability and documentation

  • Goose Specs and Reviews retain the final assistant answer from its structured transcript.
  • Failed local Reviews report a concise error and retain raw output for diagnosis.
  • Verification has larger finite time budgets for the full real-Git suite, preserving every check and publication guard.
  • Documentation includes the updated 90-second explainer, clearer recovery guidance, more readable labels and headings, improved contrast in both themes, and simpler layouts. Commands, diagrams, and onboarding controls are preserved.

Install or upgrade:

uv tool install agentmachinist
# Existing uv installation:
uv tool upgrade agentmachinist
machinist --version
machinist rehearse --guided

Python 3.12 or later is required. The free rehearsal makes no model/API calls. Real Tasks use your configured coding assistant and its provider usage.

The release workflow verifies the exact version, builds and smoke-tests the wheel and source distribution, and publishes using PyPI Trusted Publishing. Distributions and SHA-256 checksums are attached after successful publication.