OptScale Kompressor:开源中间层压缩构建日志,宣称节省约两成账单
开源项目 OptScale Kompressor 是坐在编程代理与模型之间的单二进制,只压缩重复的构建日志等机器输出并保留可取回句柄;作者称基准上真实账单约降两成二。
编程代理会反复把同一大段构建日志、编译输出和进度条送回模型,每一次重读都计费。OptScale Kompressor 试图砍的是这类「机器噪音」,而不是改你的提示词。
据 @rise_raise_ai 介绍,该工具为免费开源的单一二进制,架在代理与模型之间,确定性压缩嘈杂的机器输出,并声称因此能保留提示缓存折扣;被去掉的细节仍可通过句柄取回。作者称在终端基准上真实供应商账单下降约两成二,成功率无明显下降,约为此前最佳开源压缩器的两倍。安装后以包装命令包裹代理运行;公开仓库名为 hystax/optscale-kompressor。
原帖配图带醒目折扣风标识,本轮不引用。数字来自作者基准叙述,未独立复测;是否与各云厂商的缓存计费规则完全兼容,需按实际账单核对。对日志极多的持续集成与调试型会话,这类中间层值得小范围对照试验,先看压缩后是否丢掉关键报错行。
以上整理仅依据原帖可见事实与公开表述,未额外编造未出现的版本号、客户名单或未披露指标。
OptScale Kompressor for coding agent logs
Slash your AI coding agent bills ~23% without touching a single prompt. Coding agents constantly re-send the same giant build logs, compiler output, and progress bars. You pay for every re-read. OptScale Kompressor is a free, open-source (MIT) single Rust binary that sits between your agent and the model. It deterministically compresses only that noisy machine output (so prompt-cache discounts stay intact) while keeping every removed detail retrievable via a simple handle. On Terminal-Bench it cut real provider bills 22.7% with no meaningful drop in success rate — roughly twice as good as the previous best open compressor. Install the binary, then just run kompressor wrap codex (or the equivalent for your agent). GitHub: hystax/optscale-kompressor