Question 1

Why not just use a linter for all of this?

Accepted Answer

A linter owns syntax and dataflow such as prefer-const; Jev owns natural-language rule intent. The repo is explicit that exact characters and cross-file state like unseen package.json dependencies score poorly and belong elsewhere.

Question 2

How accurate is it really at the default threshold?

Accepted Answer

59 labeled examples over 19 style, architecture, and security rules: 93.3% precision and recall at the 0.7 default, 95.3% and 91.1% at 0.8, with 2 of 28 clean texts flagged at default. Reproduce from bench results with npm run bench.

Question 3

When do the checks run and can they loop forever?

Accepted Answer

A Stop hook checks the final reply and a PostToolUse hook checks Edit, Write, and MultiEdit, reading the same rule files with Claude Code precedence. End mode records edits and checks everything once per turn; broken rules return decision block exactly once so the fix cannot loop.

Question 4

What happens if I install it but never set a key?

Accepted Answer

Install the marketplace plugin and set TYPESAFE_API_KEY in Claude settings; with no key there are no checks and sessions keep working. The same CLI reads the key from the environment, so it doubles as a changed-files rule check in CI.

Question 5

How do I try it before installing anything into Claude Code?

Accepted Answer

Nothing to install if you have Node 18 and npx: run npx jev-enforce check --as code src/orders.ts, or --as reply with a text file. It exits 1 when a rule is broken. Three other commands: jev-enforce rules prints your rules and how each was sorted, and the same CLI reading TYPESAFE_API_KEY from the environment doubles as a rule check on changed files in CI. Tests run offline against a fake Jev implementation.

Question 6

What exactly gets sent to the API on each check?

Accepted Answer

Your reply text when a reply is checked, the edited file contents when an edit is checked, and the rules themselves, one question per rule, in a single request at about 350ms. The README is plain about the cost of that: about three cents per 1,000 checks, and the whole 59-example benchmark cost $0.0027 at a 348ms median. Set JEV_ENFORCE_OFF=1 to disable, or move JEV_ENFORCE_THRESHOLD, which defaults to 0.7.

Question 7

Is this production-ready, and what license is it?

Accepted Answer

It is a check, not a guarantee. At the default threshold about one flag in fifteen is wrong and about one broken rule in fifteen slips through, it misreads exact characters, and it cannot see state outside the text, so a never-add-a-dependency rule scored 0.56 on a fresh axios import. TypeScript with no runtime dependencies, MIT licensed, and dist/src/ is committed because Claude Code runs plugin hooks straight from the repo.

Jev Enforce

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev community