用 Jev 做 Claude 命令安全预检，一晚仅 0.6 美分
作者用 PreToolUse 钩子把每条 bash 命令先交给 Jev 做风险判断，300 次调用整晚只花 $0.0063。
Your Claude loop spends more on deciding than on building. Jev does the deciding in 100 ms every "is this command safe" goes to the same model that writes your code. on Fable 5.1 a yes or no like that costs about 4 cents and takes seconds. the fix is a PreToolUse hook that sorts every bash command before Claude is allowed to run it. hard rules → Jev risk choice → allow, deny or ask code blocks rm -rf /, force pushes and piped shells first, because a command can be written to mislead a classifier. then Jev picks one of 5 risk labels, each with a confidence. read only and local edits at 0.9 or higher run. destructive at 0.9 is denied. anything external lands on you, whatever the score. a heavy night is 300 bash calls at ~500 tokens each. on Jev the whole night costs $0.0063. the takeaway: let Claude build, not judge. every yes or no belongs to a model that is cheaper and calibrated.
这条是引用着另一条写的：x.com/polydao/status/2103689373774483815