用 Jev 接管 Claude Code 的 yes/no 路由决策
把「重试还是停」「读下一个文件吗」这类路由问题交给 Jev 做 0.1 秒的类型化决策，替代 Opus 5.5 每次花 0.4 美元、1.4 秒的昂贵调用。
99% of Claude Code users are running a $20/hour brain on $2/hour tasks and they'll keep doing it because nobody surfaces the routing layer in the docs every session has 5-6 decisions per loop - "which file", "retry or stop", "call this tool", "is this done" those aren't generative questions. they're routing questions Opus 5.5 burning 800 tokens to return "yes, read the next file" isn't intelligence - it's overhead you don't see on the surface three settings that fix it: > effortLevel: high in ~/.claude/settings.json for the main session only > explorer, worker, researcher subagents in .claude/agents/ on medium > Jev handling yes/no forks before an expensive call ever fires Jev turns "retry or stop" into a 0.1s typed decision for almost nothing Opus was spending $0.40 and 1.4 seconds on that same call most people hit "o5.5 is too expensive" and upgrade their plan. the actual fix was never the plan model didn't get dumber. you just stopped asking a surgeon to schedule his own appointments
这条是引用着另一条写的：x.com/l1vsun/status/2105630781578125651