jev-kit：把 Claude Code 拆成写手、决策者、执行者

Deniss @learnodd原推发布于 2026.10.16

在 X 上打开原推

It is becoming f*cking dangerous: Claude Code agents stopped thinking about every yes/no and started handing it to a second brain TypeSafe built Jev, a model that never writes a single word. You give it a state and a typed question, and it returns a choice, a score or a yes/no with a confidence attached Someone wired it into Claude Code, measured everything and open-sourced the whole setup as jev-kit The split is one rule: if it creates → LLM. if it picks, scores or gates → Jev. if it does the thing → code Each link in the team got a clear scope of work: > Claude writes the code, the drafts, the briefings and the text in form fields > Jev picks the next worker, scores the draft and gates every tool call > plain code saves the files, hits the APIs and clicks the browser > a tier guard checks every sub-agent dispatch and blocks the ones 2+ tiers too expensive The danger is that a frontier model will happily burn 30 cents of tokens "thinking about" a fork it already knows how to name So every decision has a bar to clear: a block only fires at 0.80+ confidence with a 0.40 margin over the runner-up, and any error or timeout just lets the call through For every call, the log shows: the question that was asked the choice Jev made the confidence behind it whether the call was allowed, warned or blocked The results: on a browser agent, with Jev picking every click, it reached 9/9 goals for $0.0008 of Claude per run. With Sonnet deciding: the same 9/9 for $0.1868. The tier guard was right on 56 of 57 dispatches with zero false blocks If you want to run something similar, open Claude Code in your agent's repo and give it this command: Split this agent into a writer, a decider and an actor: 1. Find every place where the LLM answers yes/no, picks from a list or scores something. List each one with its file and line 2. Turn each one into a typed Jev question with explicit options and a confidence threshold 3. Keep the writing on Claude, keep side effects in plain code, and log every decision with its confidence 4. Run it in shadow mode first: log what Jev would decide, change nothing, then show me the diff It is wiser to start with one decision, the one your agent makes most often, and scale up from there You will need a TypeSafe API key for the Jev calls. jev-kit's installer also has a --check-only flag that prints the plan before it touches anything All 12 use cases and how to wire them are in the article below

这条是引用着另一条写的：x.com/thegreatest_sv/status/2102802378378621355