jev-kit for Claude Code

项目摘要

jev-kit packages several Jev decisions around Claude Code: a PreToolUse guard for ambiguous tool calls, a check for oversized sub-agent choices, completion verification, faster file search and a bounded browser agent. Claude remains responsible for writing and planning; Jev handles small typed decisions inside the loop.

来源原文

I've been using TypeSafe's Jev with Claude Code for a couple of days and kept adding more use cases/ideas. So I've cleaned it up and put it in one repo. For anyone who doesn't know Jev; Jev is a small fast model that answers pick-one and yes/no questions (and more) in about 0.3s, which makes it cheap so you can put it inside your agents loop and make calls Claude would otherwise spend a turn on (eating away our already limited weekly usage). What's in it: * A PreToolUse guard. Plain code checks let about 93% of tool calls straight through in ~33ms, the grey ones get one question to Jev. Catches stuff like cat on an env file or a find across the whole disk * Sub-agent right-sizing. It asks Jev what kind of job a dispatch really is and flags or blocks when you send Opus to do a grep. On one day 20 of my 27 dispatches were a tier higher than they needed to be * A check that sends an agent back when it says done without having verified anything * File search pointed at plocate on Linux and Everything on Windows instead of find (like 500 times faster search) * A browse tool for Claude Code where Jev does the clicking * Installer, doctor, a warm daemon so calls stay fast, runs on Linux, WSL and native Windows The browser part is what I spent the last chunk of tokens on. Since Jev only picks one thing on screen at a time and doesn't plan, it's good when you literally spell out the clicks and useless when you don't. I put Sonnet on low effort with thinking off in front of it as a planner, and it's kept warm between calls. 6 Wikipedia tasks run 3 times for cleaner data where every link is pre-planned: |Model|Score|Time per task|Cost| |:-|:-|:-|:-| |Jev alone| 15/18|7s|$0.00001 ?| |Sonnet low + Jev|16/18|16s|$0.05| |Sonnet + Playwright |17/18|26s|$0.1| Tasks where Jev has to find its own way: |Model|Score|Time per task|Cost| |:-|:-|:-|:-| |Jev alone|0/6|DNF|$0.00001 ?| |Sonnet low + Jev|6/6|33s|$0.11| |Sonnet + Playwright |6/6|47s|$0.18| First version of that planner was a full Claude Code session and it sat thinking for nearly a minute per task, so warm + thinking-off matters a lot. About the guard: I ran 30 sessions with and without it and it blocked nothing. Agents mostly behave and it costs almost nothing when nothing's wrong and if Jev is down or times out the call just goes through, so I keep it on as a backstop. Hopefully it helps some of those skill issue posts from happening. MIT license, and needs a TypeSafe key. Browser agent is browser-use's jev-ultrafast vendored in with my changes on top.

原记录的限制

• The Reddit post reports small author-run measurements rather than an independent benchmark.
• The browser workflow works best when the click sequence is explicitly bounded; the author reports that open-ended navigation is not a fit.
• The guard was observed blocking nothing in 30 sessions, so the post does not demonstrate prevented incidents.
• The kit fails open when Jev is unavailable and requires a TypeSafe API key for Jev-backed checks.