Creed

在评论前先浏览了整个站点。选择 MCP 拉取是正确的交付策略，因为与其将内容粘贴进提示词，不如在回答前拉取文件，这样才能让它在多个工具间起作用，而不仅限于某一个工具内。

Andrew 在上面问到，随着文件增长，代理是否会读取全部内容。我认为更前置的问题是：它是否会缩减。你自己截图中的两个差异对比都只增加行而不删除任何行，一边是 +18 和 +9，另一边是零，而这就是每一款记忆产品最终演变成的样子——因为代理擅长注意到该添加什么，却天生没有感知“什么已经不再成立”的本能。忍受上一年这种只增不减的变化，它会变成每次回答问题前都要拉取的一段冗长前言，其中大部分都已过时，且每次调用都要为所有内容付费。

所以问题是：会不会有任何东西提议删除？一条指出“这行信息已经失效四个月了，以下是原因”的差异记录，才是更难构建的部分，也是能让这份文件始终值得被阅读的关键。

另一个问题是优先级。你自己的示例里写着“周一到周四发布到生产环境”。如果我的 Creed 这么说了，而某个代码仓库自己的上下文文件另有规定，那么在我将它连接到任何真实东西之前，我需要知道代理到底应该相信哪一个。

Went through the site before commenting. The MCP pull is the right delivery choice, since a file fetched before the answer rather than pasted into a prompt is what makes this work across tools instead of inside one.

Andrew asked above whether the agent reads all of the file as it grows. I think the prior question is whether it ever shrinks. Both diffs in your own screenshots add lines and remove none, +18 and +9 with a zero on the other side, and that is the shape every memory product drifts into, because agents are good at noticing what to add and have no instinct for what stopped being true. Accept a year of those and it is a long preamble pulled before every answer, mostly stale, all of it paid for on every call.

So the question: does anything ever propose a deletion? A diff that says this line has not been true for four months, here is why, is the harder thing to build and the thing that keeps the file worth reading.

The other one is precedence. Your own example says ship to production Monday through Thursday. If my Creed says that and a repo's own context file says otherwise, I want to know which one the agent is meant to believe before I connect it to anything real.