砍掉重复历史省八成
Strip repeat context, save 86%
要点
- Codex 每次请求会重发整段历史，含已读过的大文件和命令输出
- 路由器把旧输出换成短备注，模型需要时再重新读取
- 同请求对比：22071 对 2991 输入 token，约省百分之八十六
- 仓库是 codex-router，不合适可在托盘关掉
- 有人警告这会打坏缓存、可能更贵，先小流量试
原帖开头
Just added a feature we can reduce token usage in Codex. Here's how it works: 1. Codex resends your whole chat history on every request. That includes big file dumps and command outputs the model already read once. You pay for them again every time. 2. My router now cuts those …
