Skip to content

v0.1.0 — first public release

Pre-release
Pre-release

Choose a tag to compare

View all tags
@waspa waspa released this 26 Oct 17:38
· 6 commits to main since this release
v0.1.0
3670c1c

The first public version of ARS, an engine that learns how you write code from one of your projects, turns it into Claude Code skills, and measures on unseen tasks whether a coding agent follows them.

This is an early release. Expect rough edges, and please open an issue when a rule ARS learns from your code is wrong.

Highlights

  • Signature in six layers: formatting, syntax idioms, naming, component anatomy, architecture and voice. Every rule links to the files it was learned from, with counterexamples, and is self-tested against a planted violation.
  • Review inbox: approve, reject or reword each rule. Only approved rules reach the skills and the scores.
  • Versioned skill packages: SKILL.md files, de-domained exemplars, a formatter config, and an edit hook that formats each file and blocks off-style edits while the agent works. Install them from the UI or with ars install.
  • Lab: reconstruction (T1), transfer to new apps (T2) and indistinguishability (T3) tests, each run under three conditions (no skills, a length-matched generic guide, your skills), with repeats and bootstrap confidence intervals.
  • Improvement loop: diagnoses why a rule failed, proposes a fix, and promotes it only if it passes a regression gate on held-out tasks.
  • CLI: ars analyze, ars check, ars install and the Claude Code hook.

Requirements

  • Node 22.13 or newer, git and npm.
  • Analysis, review and skill generation need nothing else.
  • Deep-read and judges need an Anthropic API key. Agent runs need an API key or a Claude Code login.

Getting started

git clone https://github.com/ars-engine/ars.git && cd ars
npm install
npm run dev   # UI on http://localhost:5317

See Getting started and Costs before running Lab experiments, and Security for what leaves your machine.

Known limitations

  • Deep analysis covers TypeScript, JavaScript and React only. Other languages get a basic inventory.
  • There is no automated test suite yet.
  • The T3 "spot the real file" game can pick runs from any condition, not only the ARS one.
  • The Lab verdict compares means, not confidence intervals, so small runs can pass or fail by noise.
  • The baseline probe is only available after a skill version exists.

What's next

  • Codex and other coding agents, compared side by side in the same Lab.
  • Deep analysis for more languages.
  • Automated tests and CI.
Assets 2
Loading