用 Jev 打分压缩 agent 上下文
演示用 Jev 给每行内容打丢弃概率，保留报错原文、删掉无关行，token 从 15.6 万降到 6.2 万且不重写内容。
YOUR AGENT FORGETS THE ERROR MESSAGE THE SECOND IT RUNS OUT OF CONTEXT. JEV KEEPS IT WORD FOR WORD the fix everyone ships is to ask a model to summarize the session. you wait, you pay for the tokens, and you get back a paraphrase of your own work. the exact failing line is gone. score every tool call → drop what's irrelevant → keep the rest verbatim jev scores each line with a drop probability instead. "read 212 lines" gets 0.97 and goes. the failing assertion, expected 1200 and received 120, gets 0.09 and stays exactly as it was written. 156,000 tokens down to 62,000. context from 78% to 31%. 10 lines kept verbatim, 6 removed, nothing rewritten in between. the takeaway is simple: summaries lose the detail you'll need in ten minutes. deletion doesn't. you keep the real line or you don't keep it at all.
这条是引用着另一条写的：「从零安装 Jev 的十步指南：把 2028 工程师栈搬到今天」