GitHub - redreamality/jev-skill-selection: Pre-message hook: use TypeSafe Jev to keep/drop skills and shrink agent context
Pre-message hook: use TypeSafe Jev to keep/drop skills and shrink agent context - redreamality/jev-s
最近算了笔账：agent 技能一多，真正贵的是把无关 SKILL.md 整段塞进上下文。用 Jev 在首条消息前 keep/drop，再硬过滤——仓库小 demo 是 1431→393，大约砍掉七成 skill 正文；目录到几十上百个时，skill 块往往能砍九成量级。有 prompt cache 也别只靠「全量塞进去再打折」：基数大，命中仍在付；过滤后再缓存，写入和命中都会瘦一圈。