Coding agents break your rules from the very first edit. Abide catches every one and makes your agent fix it

1 in 13 turns break a rule no linter can catch · abide does · 300 ms per check · a tenth of a cent per turn
_{Measured by replaying 93 real Claude Code sessions (1,256 edits, 147 turns) in two repos against their own AGENTS.md, for 22 cents. Jev flagged 39 edits and 15 turns; an independent reviewer confirmed 10 and 11. The turn-level catches (single-use abstractions, oversized files, duplicated logic) held up 11 times in 15. Method, per-rule table and what was wrong: benchmarks/replay.}

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, opencode or pi as usual. That is the whole setup.

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.

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.

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.

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". Scored 0.86 in apps/web/src/pages/api/logout.ts.
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.

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.

Run abide init pi and restart Pi, or use /reload in an existing session. Add --project to install into .pi/extensions/abide.js; Pi must trust the project before loading it. PI_CODING_AGENT_DIR selects a different user-level agent directory.

Tested with Pi 0.99.1. The extension uses the same checker, key lookup, and rubric as the other agents. It checks local edits and complete activity diffs, including shell changes, and can request one final repair continuation. abide replay pi checks recorded Pi edit/write tools on the active persisted branch. Installation, limitations, and live verification.

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.

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

compile and tune use Claude Code by default, or Pi when Claude is unavailable. Choose explicitly with abide compile --agent pi or abide tune --agent pi; add --global to tune global rules. An explicitly selected agent never falls back to another. If neither is available, Abide prints the prompt for you to paste into an agent session. --print prints the prompt without starting an agent.

Pi runs with --print --no-session, using its configured model and authentication. It does not save a session or override project trust: trust the repo in Pi first if compilation needs project settings or extensions. Claude Code uses its saved login without inheriting ANTHROPIC_API_KEY or ANTHROPIC_AUTH_TOKEN from your shell.

.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 your history and switches them off. tune has the agent rewrite them.

• Changed lines go to TypeSafe under your key, and nowhere else. No server of ours is in the path.
• Retention is between you and TypeSafe: their API takes no per-request retention flag, and zero data retention is an account arrangement on their enterprise tier. They state Jev is not trained on customer requests or responses. With a gateway key the call carries zeroDataRetention: true, which asks the Vercel AI Gateway to route only to providers under a zero-retention agreement.
• Key lookup order: the environment, then .env.local and .env at the repo root, then ~/.abide/.env. Never a flag, never logged. Set AI_GATEWAY_API_KEY instead of a TypeSafe key to go through your Vercel AI Gateway.
• To send the direct call somewhere other than typesafe.ai — a self-hosted deployment, a local proxy, a mock — set TYPESAFE_AI_BASE_URL alongside your TypeSafe key. It is read from the same places as the key, never written or logged, and applies only to the direct call, not the gateway.
• A check on this repo's 13 rules is 1,000 to 1,600 input tokens: $0.00004 to $0.00007, about 300 ms for Jev and about 1 s for the whole hook including Node startup. A turn of 15 edits costs a tenth of a cent. Measured 2026-09-18, direct to TypeSafe. abide bench measures yours.
• The hooks cannot break your session. Every path exits 0, has a hard deadline, and prints only what the host expects.
• No key or no network: the edit goes through unchecked and the miss is logged in .abide/events.jsonl, where report counts it.

Four hooks per agent. Session start: hash the instruction files, ask the agent to compile if they changed. Turn start: snapshot the working tree with git. After each edit: run the edit-phase rules on that hunk. End of turn: diff the whole turn against the snapshot and run the turn-phase rules, plus the edit-phase rules for anything a shell command wrote.

abide uninstall            # every agent it was installed into
abide uninstall codex      # one agent; add --project for a project-level install

Removes abide's own entries and nothing else. Rubric files and ~/.abide/.env stay until you delete them.

• packages/schema: the rubric, hook payloads, verdicts and events as zod schemas.
• packages/cli: the abide command, the hook script, OpenCode plugin, and Pi extension.
• skills/abide-compile: the procedure the agent follows to compile a rubric.

MIT