abide-compile

---

npx @coldtea/abide login    # pick a key type, paste it once
npx @coldtea/abide init     # hooks into every agent on this machine

Then start claude, codex or opencode as usual. That is the whole setup.

What it does

Your AGENTS.md, CLAUDE.md and the rest of your project instructions are full of rules no linter can check. "Never let a raw error reach a user." "Don't create premature abstractions." Nothing can script those, so nothing enforces them. In 93 real sessions, the agent broke one on 1 turn in 13, from the first edit on.

https://github.com/user-attachments/assets/2d45f6b0-c889-474c-ab4a-8d019fdc7140

Abide enforces exactly those rules. On every edit (or turn) it asks Jev, TypeSafe's decision model, one question per rule and gets a probability back. Jev sees the rule and the diff, never the conversation, so edit 200 is checked like edit 1. Break a rule and the agent is told which one and fixes it in the same turn.

• One call per edit, about 300 ms, a few thousandths of a cent.
• Rules a linter could check are handed to your linter instead.
• No built-in rules. No instruction files, nothing to enforce.
• Your key, your data. Nothing here talks to a server of ours.

Not previously possible

Checking every edit or turn against every rule was never worth doing (economically and latency-wise) with an ordinary LLM. A check is about 2,500 tokens. At typical model prices that is a cent or more, and a few seconds, per edit, and the answer comes back as prose you then have to parse and cannot fully trust. Two hundred edits a day made it a non-starter.

Jev changes the arithmetic. It is a decision model, so it answers a typed question with a calibrated probability and nothing else. There is no free text, so there is nothing to make up. It is up to 100x cheaper than a typical LLM and answers in about 300 ms. That is what makes it reasonable to check every edit, every time.

Three minutes to the first catch

1. Get a TypeSafe API key at typesafe.ai, or use a Vercel AI Gateway key you already have.
2. Run npx @coldtea/abide login, pick which kind of key it is and where it lives, and paste it. It goes to ~/.abide/.env for every repo on the machine, or to .env.local in this repo, owner-only either way. A .env you already have at the repo root works too.
3. Run npx @coldtea/abide init in your repo.
4. Start your agent. Its first turn compiles your rules into .abide/rubric.json and tells you what it found.
5. Ask for something your rules forbid. An AGENTS.md that says "use Yup, never validate by hand" produces this the moment the agent writes a manual guard:

Abide: This edit appears to break a rule from this repository's instructions.
- Rule "api-validation-uses-yup" from ~/.codex/AGENTS.md line 65: "When writing API endpoints, do NOT write input validations manually. Use Yup (with clear validation messages) + early return in the API handler". (0.86)
Repair apps/web/src/pages/api/logout.ts now, then continue with the task.

The agent repairs it before moving on. No human in the loop.

Agents

init with no name installs into every agent it finds. Add --project to install into the repo instead, so teammates get it with the checkout.

Codex only: start codex, type /hooks, and accept the four abide entries. Codex asks this once for any new hook. Codex edits through apply_patch; abide reads the patch and judges every file in it.

OpenCode only: there are no hook processes, so abide runs as a plugin. Same checks, same messages: an edit that breaks a rule gets the repair request appended to its tool result, and a turn that ends with one gets a single follow-up message.

See what your codebase already breaks

abide audit src/

Every file is judged as if it had just been written. You get a table by rule and a list by file. On 33 API routes of a real Next.js app: 12 seconds, about a cent.

Commands

report, check, audit, bench and calibrate take --json.

The rubric is yours

.abide/rubric.json is a committed, readable file. Every verdict names a rule in it, and every rule quotes the line of your instruction file it came from. A wrong verdict is a rule you can rewrite.

• Each rule runs at one of two moments: edit after each edit, turn once at the end against the whole diff. "Did this add more than was asked" has no answer after edit 1 of 12.
• Each rule can carry a scope of globs, so an API route and a stylesheet get different questions.
• Verdicts are banded. 0.8 and above: the agent is told to repair. 0.5 to 0.8: you see a note, the agent does not. Below 0.5: nothing.
• A badly worded rule scores 0.4 on everything and never fires. calibrate finds those against twenty real hunks from