把 Jev 放在 Claude Code 前面做成本路由

Wifter @0xWifter原推发布于 2026.09.27

在 X 上打开原推

your coding agent is burning money before it writes a single f**king line of code so i put jev in front of claude code and made it decide what actually deserves expensive reasoning here’s how the setup works: step 1 → give jev a small state. the task, changed files, failing tests, and whether the action can break anything step 2 → use CHOICE to route the work. simple sh*t stays local, real coding goes to claude, risky stuff gets kicked back to a human step 3 → use NOUL before anything sensitive. if jev isn’t confident the action is safe, the agent doesn’t just yolo the command step 4 → SCORE every change by risk. a reversible one-file edit shouldn’t burn the same reasoning as touching auth, infra, or production step 5 → keep claude for the hard stuff. debugging, architecture, planning, ambiguous problems, and actually writing the code step 6 → run jev in shadow mode first. let it make decisions without controlling anything and watch where your agent is wasting tokens step 7 → once the routes make sense, wire them into the real loop. jev decides, code routes, claude does the expensive work jev handles the decisions. claude stays the brain and only steps in when the work actually needs expensive reasoning. copy the setup below and try it on your own coding agent ↓