kaize

中文参考

我访问了Jev并立即将其插入Claude Code，看看它是否真的解决了代理工作流程中最大的难题：上下文压缩 >结果：相当狂野 为那些错过它的人提供快速背景： 每个Claude Code会话都有一个上下文窗口，该窗口会快速填充工具调用（文件读取、grep输出、日志等）。 当它填满时，代理运行“压缩”-要求一个单独的LLM总结聊天。这是缓慢的，昂贵的，LLM经常丢弃重要的上下文或用垃圾膨胀摘要： fast-jev-compaction使用Jev代替 链接：https://t.co/WD7kpBfDwL 为什么这么好用： Jev不会生成文本摘要，而是对聊天中的每个工具调用进行评分，并回答“这有多重要？“概率。 >高于阈值→保持它 '下面'删除它或用树桩替换 >总而言之：没有文本生成、没有臃肿、没有丢弃关键上下文 你实际得到的是： 1/更快：Jev并行输出概率，而不是顺序令牌。 2/更便宜：100万个上下文代币=约0.04美元以压缩。 3/确定性：没有“创造性”摘要会删除一半的代码库上下文。 顺便说一句，Typesafe提供5美元的免费积分，足够进行数周的测试。 链接：https://t.co/NzGPArgLvc 诚实的警告： Jev只看到工具Calls，而不是它们的输出，很难说压实质量客观上是否优于默认。 Jev的上下文窗口很小（32 K），因此需要对150 K以上的聊天进行分块。 这不是完全替代品-这是一个智能的预过滤器 而是作为概念的证明。 这正是使代理工作流程大规模可行的外科修复方法。 现在考虑以Jev作为决策层，我的堆栈中还有哪些内容可以提高100倍。

原帖全文 / EN

I got access to Jev and immediately plugged it into Claude Code to see if it actually FIXES the biggest pain of agent workflows: context compaction > result: pretty wild Quick context for those who missed it: Every Claude Code session has a context window that fills up fast with tool calls (file reads, grep outputs, logs, etc). When it fills up, the agent runs "compaction" - asks a separate LLM to summarize the chat. this is slow, expensive, and the LLM often drops important context or bloats the summary with garbage: fast-jev-compaction uses Jev instead Link: https://t.co/WD7kpBfDwL Why this works so well: Instead of generating a text summary, Jev scores every tool call in your chat with "how important is this?" probability. > above threshold → keeps it > below → drops it or replaces with a stub > in summary: no text generation, no bloat, no dropped critical context What you actually get: 1/ way faster: Jev outputs probabilities in parallel, not sequential tokens. 2/ way cheaper: 1M tokens of context = ~$0.04 to compact. 3/ deterministic: no "creative" summary that removes half your codebase context. btw, TypeSafe gives $5 free credits at enough for weeks of testing. Link: https://t.co/NzGPArgLvc Honest caveats: Jev only sees the tool CALLS, not their outputs, hard to say if compaction quality is objectively better than default. Jev's context window is small (32K), so 150K+ chats need to be chunked. This isn't a full replacement - it's a smart pre-filter but as a proof of concept. This is exactly the kind of surgical fix that makes agent workflows viable at scale. Now thinking about what else in my stack could get 100x faster with Jev as the decision layer.