Jev-guard：在 Claude Code 工具调用前打分拦截
给 Claude Code 的每次状态变更工具调用加一道 0–3 风险评分，按分数放行、暂停或拒绝，单次检查约 0.00004 美元。
How Jev scores 1,000 actions for 4 cents Jev-guard plants a scoring model between Claude Code and every tool call that changes state. Reads walk past. Cost per check: roughly $0.00004. Three questions per call. How risky, on a 0 to 3 scale. Would an engineer want to approve. Did the user's recent messages actually ask for it. > INTERCEPT - tool call hits the guard before execution > SCORE - one risk number between 0 and 3 > DENY - 2.5 or higher, or injection flagged past 0.7 > ASK - 1.5 or higher, call pauses for a human yes > ALLOW - user explicitly requested, 0.85+, call runs 580 ms at p50 through the gateway. 100 gated actions in a morning batch add about a minute. The guard repo sits at 121 stars against 133.9k on the graph tool it protects. 1,100x less attention on the piece that decides what gets through. Without a scoring gate, agent permissions are binary: everything runs or nothing does. This one adds a ruler. Each call lands on it, and the score picks the lane. Through, paused or killed. The guard fails open by default. For a desk with live mail and a real browser, set it to fail-closed before the first scheduled run. Terminal below: items flow through the graph, hit the ruler, split into three lanes ↓
这条是引用着另一条写的：x.com/0xkarmi/status/2114102976788808056