用 Jev 重建 Claude Code 上下文压缩的五条规则

Freddy @0x_freddy原推发布于 2026.09.21

在 X 上打开原推

Half of AI engineering just got obsolete in five days. The model doing it cannot write a sentence Jev does not generate text. You hand it state and typed questions, it hands back a choice from a set you defined, plus a confidence score. It cannot return something off-schema. That is the whole product. Someone already rebuilt Claude Code's compaction on top of it. Reported 4,575 stars in five days. 200x faster, 400x cheaper per their own tests. Five rules before you wire it in. Separate the deciding from the generating first. Most agent loops burn a full generative call on "which tool" and "should I retry." That is frontier rates for a multiple-choice test. Set your confidence threshold before you ship, not after the first incident. Without one you built a fast oracle and will trust it as blindly as the slow one. Log the confidence next to the real outcome every single time. A drifting calibration is invisible until you have the paired data. "0.9 confident" means nothing until you check what 0.9 actually cashed out to. Treat the benchmarks as marketing until someone independent runs them. The speed numbers are self-reported. The accuracy reference is an average of two frontier models. Vendor grading vendor. Steal the pattern from fast-jev-compaction. Every tool call gets scored in one request, stale ones dropped or truncated, everything kept stays verbatim. No summary step means no summary lies. A model that cannot explain itself, deciding which half of your context lives. Worth asking that question before you wire it in

这条是引用着另一条写的：x.com/yurshevv/status/2100978237547458936