Abide

What it does

• Four hooks per agent (session start, turn start, after each edit, end of turn) — npx @coldtea/abide init wires every agent it finds, or name one: claude / codex / opencode
• Rules live in .abide/rubric.json — every verdict names a rule and every rule quotes the instruction line it came from, so a wrong verdict is a rule you can rewrite
• Two rule moments: edit-phase after each hunk, turn-phase against the whole diff (“did this add more than was asked” has no answer after edit 1 of 12); scope globs let API routes and stylesheets get different questions
• calibrate scores every rule against twenty real hunks from your git history and switches off the ones that never fire; tune has the agent rewrite them; audit judges existing files — 33 API routes of a real Next.js app took 12 seconds and about a cent

When to use it

Repos whose AGENTS.md or CLAUDE.md encodes rules no linter can check — “never let a raw error reach a user”, “don't create premature abstractions” — and teams tired of finding the violations in review. Also works as a batch auditor for code the agent has already written.

Getting started

npx @coldtea/abide login    # paste a TypeSafe or Vercel AI Gateway key once
npx @coldtea/abide init     # hooks into every agent on this machine
# or one agent: npx @coldtea/abide init claude
# then start claude / codex / opencode as usual

Things to know

Every check is a cloud call to TypeSafe under your own key — about 300 ms and $0.00004–0.00007 per the project's 2026-09-18 measurement. Changed lines go to TypeSafe and nowhere else, but retention is an arrangement between you and TypeSafe: zero retention means the enterprise tier or a Vercel AI Gateway key sending zeroDataRetention: true. No key or no network means the edit passes unchecked and the miss is logged to .abide/events.jsonl for report to count. Cost and accuracy figures are the project's own benchmark. MIT.