jev-enforce

Claude Code plugin that checks every reply and every file edit against your CLAUDE.md (or AGENTS.md), using TypeSafe's Jev. Break a rule and Claude gets it quoted back and fixes it, in the same turn.

jev-enforce catching six broken rules in one check

_{The handler it is checking is demo/src/orders.ts.}

Why

CLAUDE.md is context, not a constraint. It holds for a few turns, then a handler shows up with raw SQL in it and a test quietly becomes it.skip. Nothing checks the output against the rules, so this does.

Jev makes it cheap enough to run every turn: it answers typed yes/no questions instead of writing text, all of them in one request. One question per rule, about 350ms, about 3 cents per 1,000 checks.

Try it

Nothing to install, Node 18+ and npx are enough:

npx jev-enforce check --as code src/orders.ts

Install

claude plugin marketplace add erkamyaman/jev-enforce
claude plugin install jev-enforce@jev-enforce

Add your key (from console.typesafe.ai) to ~/.claude/settings.json:

{ "env": { "TYPESAFE_API_KEY": "..." } }

No key, no checks. It cannot break a session.

How it works

Reads the same rule files Claude Code does: ~/.claude/CLAUDE.md, CLAUDE.md, .claude/CLAUDE.md, CLAUDE.local.md, .claude/rules/**, and AGENTS.md when the project has no CLAUDE.md, matching Claude Code's own precedence. Each bullet is one rule.
Sorts each rule once, cached: does it constrain replies, code, both, or neither. Git and workflow rules land in neither and are skipped.
A Stop hook checks the final reply. A PostToolUse hook checks Edit, Write and MultiEdit.
With JEV_ENFORCE_MODE=end, edits are only recorded while Claude works. The Stop hook checks every edited file and the reply together, in parallel, so a turn pays for one check no matter how many edits it made.
Broken rules come back as decision: "block" with the rules quoted, so Claude fixes it before you see it. Once per turn, so it cannot loop.

CLI

Installed as a plugin, or through npx jev-enforce:

jev-enforce rules                          # your rules and how each was sorted
jev-enforce check --as code src/orders.ts  # exits 1 if a rule is broken
jev-enforce check --as reply message.txt

The CLI reads TYPESAFE_API_KEY from the environment, so it works in CI as a rule check on changed files.

Settings

Env	Default
`TYPESAFE_API_KEY`	none	Required
`JEV_ENFORCE_THRESHOLD`	`0.7`	How sure Jev must be before Claude is told
`JEV_ENFORCE_MODE`	`turn`	`end` checks all edits once when the turn ends, instead of after each edit
`JEV_ENFORCE_OFF`	unset	`1` disables it

Benchmark

59 labeled examples against 19 rules, half style and half architecture and security. Full output in bench/results.md, reproduce with npm run bench.

Threshold	Precision	Recall	Clean texts wrongly flagged
0.6	87.8%	95.6%	3 of 28
0.7 (default)	93.3%	93.3%	2 of 28
0.8	95.3%	91.1%	2 of 28
0.9	97.4%	84.4%	1 of 28

348ms p50, $0.0027 for the whole run.

What it is bad at

Exact characters. It read 2019–2024 as an em dash.
Dataflow. prefer-const rules need a linter, not a model.
Rules that depend on state outside the text. "Never add an npm dependency" scored 0.56 on a new axios import, because it cannot see package.json. Editing an old migration scored 0.06.
Your reply text, edits and rules go to TypeSafe's API.

Development

npm install
npm test

TypeScript, no runtime dependencies, tests run offline against a fake Jev. dist/src/ is committed because Claude Code runs plugin hooks straight from the repo, so run npm run build before committing (npm run check-dist verifies).

Sponsor

If this saves you from one more "I told it not to do that" moment, you can sponsor the work.

License

MIT