Leash — Development skill for Claude Code

Leash

Keep a coding agent's output quality on a short leash. Leash judges every turn against your project's un-lintable rules using [Jev](https://typesafe.ai) (TypeSafe's System One decision model), and tells the agent exactly which rule it broke so it fixes it before moving on. About 300 ms and a fraction of a cent per turn.

Jev-powered. Early (v0.1): the offline core (rubric, engine, ratchet, CLI). Agent hooks land next. Inspired by [Abide](https://www.npmjs.com/package/@coldtea/abide); the difference is the ratchet (below).

Why

Your `CLAUDE.md` / `AGENTS.md` is full of rules no linter can check: "no premature abstractions", "never let a raw error reach a user", "small single-purpose functions". Nothing enforces them, so the agent breaks them from the first edit. A frontier LLM could judge each rule, but at a cent and a few seconds per check it never pays off. Jev answers one typed question per rule with a calibrated probability in ~300 ms for a fraction of a cent, which is what makes checking every turn viable.

What makes it different: the ratchet

Judging every file as if freshly written buries you in findings on a real repo. Leash borrows Foundry's accepted-debt baseline: existing violations are baselined once, and Leash only ever flags what a turn **newly** introduces. The baseline is one-way, it can only shrink. That is what makes it adoptable on an existing codebase from day one.

Turn-check is the default (once per turn, on the whole diff, where the un-lintable questions actually have an answer). A per-edit mode is planned but off by default: at Jev's edit-level precision, per-edit auto-repair risks the agent chasing phantoms.

Install

npm install -g leash   # or: npx leash 

Set a key: `export JEV_API_KEY=...` (or `TYPESAFE_AI_BASE_URL` for a self-host / proxy / mock). No key means Leash no-ops and lets the edit through, always.

Use

leash report              # list the rules in .leash/rubric.json
leash audit