Heard：让 Claude Code 和 Codex 开口说话

@galdayan 是的，3 个 Agent 正是那道明确的门槛。窗口平铺到这里就开始失效了，最后你不得不频繁按 Alt-Tab 去查看那些其实并不需要你介入的事情。

这确实就是你猜到的两部分。有些事件属于硬信号——权限请求、工具调用失败、一次运行结束——无论如何都会被语音播报。其他所有信息都会先经过一层处理，这一层会在整个会话期间保留上下文，因此不会孤立地逐行判断。它会掌握此前发生过什么，了解你真正关心的内容，并决定何时存在值得提醒的信息。

你提到的两种失败模式正是整个问题所在，大部分调优也都集中于此。第一种是关键内容埋在常规日志中却没有被提醒：硬信号列表会捕捉这类情况，即使权限请求出现在一大段输出的中间，也会被提取出来。第二种是在一段很长、看似例行的流程中保持沉默，但实际上该流程需要你做决定：这更难处理，而保留上下文正是能够捕捉到它的关键——即使周围输出看起来很无聊，系统也知道一个决策节点即将到来。

很好奇，你通常会同时运行多少个 Agent？对你来说，从多少个开始会出现问题？

@galdayan yeah agent 3 is the exact wall. tiling breaks down and you end up alt-tabbing to check on things that didn't need you.

it's both of the things you guessed. some events are hard signals - permission prompts, tool-call failures, a run exiting - those get spoken no matter what. everything else goes through a pass that keeps context across the session, so it's not judging each line in isolation. it holds what's happened so far, learns what you actually care about, and decides when there's something worth saying.

the two failure modes you named are the whole problem and most of the tuning went there. talking over something buried in routine logs: the hard-signal list catches that, a permission prompt gets pulled even if it's mid-wall. going quiet during a long routine stretch that actually needed a decision: that's the harder one, and holding context is what makes it possible to catch - it knows a decision point is coming even when the surrounding output looks boring.

curious how many agents are you running, and where does it start breaking for you?