Agents do not search like humans. They start with a hunch, inspect a result, refine the query, and keep following evidence. cxs indexes the work process itself: requests, decisions, commands, errors, and fixes from local Codex and Claude Code sessions. Agent 不像人一样搜索。它会先拿一个大致判断去查，读一段结果，再改查询，继续顺着证据往下找。cxs 索引的是工作过程本身：请求、决策、命令、报错和修复，都来自本机 Codex 与 Claude Code 会话。
No daemon, no cloud, no full-transcript dump. The index stays local, and the agent reads only the slice it needs. 无守护进程，无云端，不倾倒完整 transcript。索引留在本机，Agent 只读取当前需要的片段。
The npm package installs only the CLI. Install the agent skill separately, then restart Codex or open a new session. npm 包只安装 CLI。Agent skill 需要单独安装，装完后重启 Codex 或开启新会话。
Most knowledge tools assume somebody will stop and write a clean note. Agent work rarely waits for that. The useful memory is already present in the messy trail: what was asked, what was tried, what failed, and what finally held. 大多数知识工具假设总会有人停下来写一篇干净的笔记。Agent 工作通常不会等这一步。真正有用的记忆已经在过程里：当时问了什么、试了什么、哪里失败、最后哪个修法站住了。
cxs gives that trail back to the agent. It does not need to guess the whole answer through one embedding lookup. The agent can search, inspect, refine, and read progressively, just like it already works. cxs 把这条轨迹还给 Agent。 它不需要靠一次 embedding 查询猜完整答案。Agent 可以搜索、检查、改写查询、逐步读取，这正是 Agent 本来就擅长的工作方式。