How does AI stop being a tool you call, and become something you work alongside? AI 怎样从一个被调用的工具，变成可以一起工作的伙伴？
That is the question my projects keep circling, from different directions. It looks like a question about capability, and the usual answer is a bigger model, more tools, a longer context window. I do not think that is the hard part. 这是我的几个项目从不同方向反复绕着的问题。它看起来像是能力问题，常见的答案是更大的模型、更多工具、更长的上下文。我认为难的不在那里。
A partner is defined by what it remembers, what it is allowed to do without asking, and whether it is still there when the network is not. Those are design decisions, and they get made by whoever builds the thing around the model. 一个伙伴由三件事定义：它记得什么、什么事可以不问就做、断网之后它还在不在。这些是设计决定，由在模型外面搭东西的人做出。
Each of these is a position, not a claim about results. What follows each one is the thing in the code that pays for it. 每一条都是立场，不是对结果的断言。每条后面跟着代码里为它付出的代价。
If a companion holds your messages and your files, the honest place for it is your own machine. Convenience is the argument for the other design, and it is a real argument. I would rather pay the convenience cost. 如果一个伙伴握着你的消息和文件，它诚实的位置是你自己的机器。另一种设计的理由是方便，那是个真实的理由。我宁可付这个方便的代价。
CC is a daemon on your computer. Messages, memory and code stay in a folder on your disk. Your computer has to be awake for it to answer, and that is the cost. CC 是跑在你电脑上的守护进程。消息、记忆和代码留在你磁盘的一个文件夹里。你的电脑必须开着它才能回话，这就是代价。
An agent that can edit files is not a chat window, and should not feel like one at the moment it matters. The pause where it asks is the whole difference between a tool you are using and a tool that is using your machine. 一个能改文件的 Agent 不是聊天窗口，在关键时刻也不该像聊天窗口。它停下来问你的那一刻，正是「你在用工具」和「工具在用你的机器」之间的全部区别。
Every sensitive tool call becomes a permission card in the chat. Nothing runs until you reply. It costs seconds in a demo, and I have not tried to hide that. 每一次敏感的工具调用都会在聊天里变成一张授权卡片，你不回复就不执行。演示时它要多花几秒，我没有试图掩盖这一点。
Generating more text is easy. Staying consistent with what was already generated is the part that breaks, and it breaks quietly. So the state has to live outside the model, in something a person can read. 生成更多文字很容易。和已经生成的东西保持一致才是会坏的那部分，而且坏得很安静。所以状态必须活在模型之外，活在人能读懂的东西里。
Nate & Cynthia keeps a ledger of what the story promised and what contradicts what. hearth proposes changes as reviewable plans over plain markdown you can still read if my software disappears. Nate & Cynthia 维护一张账本，记故事承诺过什么、哪里自相矛盾。hearth 把改动写成可审阅的计划，底下是纯 Markdown，我的软件消失了你依然读得懂。
Most software that reads your health data moves straight to telling you what to do. I want to see whether a product can earn the right to say anything by first showing it understood the day. 大多数读你健康数据的软件会直接跳到告诉你该做什么。我想看看一个产品能不能先证明自己读懂了这一天，再换取说话的资格。
melu is the attempt. It is early, one vertical slice on iOS, and this is the commitment I have the least evidence for. melu 是这个尝试。它还很早，iOS 上只有一个垂直切片，这也是我证据最少的一条。
If people turn the permission step off the first week, then asking before acting is my preference rather than a requirement, and I should stop calling it a principle. 如果人们第一周就把授权关掉，那么「动手前先问」只是我的偏好而不是必需，我就该停止把它叫做原则。
If local models never get good enough for the ordinary cases, then running on your own machine is a hobbyist position, not a product one. 如果本地模型始终不足以应付普通场景，那么「跑在你自己机器上」就是爱好者的立场，不是产品的立场。
And if the companionship turns out to be something people tolerate rather than want, then CC is a good remote control wearing a costume, and I should take the costume off. 还有，如果陪伴这件事最后只是被人忍受而不是被人想要，那 CC 就是一个穿着戏服的好用遥控器，我应该把戏服脱掉。
I write down what these look like in practice as I go. Writing. Or tell me where I am wrong. 实践中的样子我边做边记。Writing。 或者告诉我哪里错了。