jev-oxlint

# jev-oxlint

Build an [Oxlint JavaScript plugin](https://oxc.rs/docs/guide/usage/linter/js-plugins) from an agent
skill. Deterministic AST checks handle structural rules, while
[Jev](https://docs.typesafe.ai/introduction) answers narrow, typed questions that require contextual
judgment.

The result is a project-specific linter that can:

- limit analysis to files that import selected packages;
- route relevant skill guidance to each file;
- report broad guidance hints before precise checks exist;
- combine AST prechecks, model answers, and deterministic thresholds; and
- calibrate model-backed checks against a human answer key.

> **Experimental:** Oxlint JavaScript plugins are alpha. Live cache misses make synchronous network
> requests, so this project is best suited to evaluation and targeted workflows. The packages are
> not yet published; use a local clone as described below.

## Quick start

Requirements: Node.js 22 or later, pnpm 12, and a project skill containing `SKILL.md` with optional
`references/*.md` files.

Build the toolkit, then scaffold a linter next to the project that it will inspect:

```bash
git clone <repository-url> /absolute/path/to/jev-oxlint
cd /absolute/path/to/jev-oxlint
pnpm install
pnpm build

node packages/create/dist/cli.js ../my-linter \
  --skills /absolute/path/to/project/.agents/skills \
  --skill my-skill \
  --targets '^@my/(sdk|client)(/.*)?$' \
  --local-repo /absolute/path/to/jev-oxlint

cd ../my-linter
pnpm install
pnpm build
```

The generated linter initially has no precise checks. It uses the skill as a hint tier, identifying
which guidance applies to each in-scope file and reporting likely deviations. Survey a codebase to
find guidance that should become a precise check:

```bash
TYPESAFE_API_KEY=… pnpm survey /absolute/pat