Jev Oxlint

It is a toolkit that builds an Oxlint JavaScript plugin from an agent skill, aimed at teams who want a project-specific linter. Deterministic AST checks handle structural rules while Jev answers the narrow, typed questions that require contextual judgment.

Code decides everything it can: import gates, AST facts, triggers, prechecks and probability thresholds. Jev answers atomic typed questions, and policy stays in the request state because checks cite complete guidance files instead of parsing Markdown sections.

Clone the toolkit, run pnpm install and pnpm build, then run node packages/create/dist/cli.js ../my-linter with --skills pointing at your skills folder, --skill naming the skill, --targets as an import regex and --local-repo pointing at the clone until the packages are published. Then build the generated linter.

Run jev-lint propose with --plugin, --guidance and --calibrate to create a redacted context packet; Anthropic is the default provider and --provider codex uses the Codex SDK. It returns a draft check with fixtures and answer-key entries, and --dry-run writes only the packet.

TYPESAFE_API_KEY for Jev-backed linting and surveys, plus ANTHROPIC_API_KEY when using the default proposal provider, or an existing Codex login or CODEX_API_KEY with the Codex provider. The survey and oxlint commands in the README are shown with TYPESAFE_API_KEY set.

Oxlint JavaScript plugins are alpha, live cache misses make synchronous network requests, and the @jev-oxlint packages are not yet published, so the README recommends a local clone and says the toolkit is best suited to evaluation and targeted workflows.

Node.js 22 or later, pnpm 12, and a project skill containing SKILL.md with optional references files. The default targets pattern limits analysis to files importing selected packages, and generated proposals are drafts that should be reviewed with their questions, thresholds and fixtures before use.