Context compaction happens in the dark. I made it happen on a map.

AI对话工具有个隐藏问题：聊太久后，系统会悄悄把之前的对话“压缩”掉，而用户完全不知情，重要的决策细节可能就这样被抹掉了。一位开发者为此打造了可视化工具ThoughtDAG，让这个过程变得透明可控。简单来说，AI对话就像一张不断扩张的地图，太长时系统会自动“简写”前面的内容来腾出空间，但用户看不到删了什么。ThoughtDAG把对话变成一张节点图，每个问题和回答都是看得见的节点，按压缩键时，系统会把废话连篇的常规问答折叠起来，同时保留你标记过的重要决策。最贴心的是，压缩不会改动原始对话，而是在旁边生成一份新版本，你可以对照着看，亲手勾选哪些部分要保留。这意味着以后和AI长谈时，我们不再担心关键信息被悄悄“灭口”了。无论是写代码、做研究还是规划工作，用户都能掌控AI的记忆边界，让对话更可靠、更值得信赖。这项设计理念也给AI工具开发提了个醒：技术决策应当让用户看得见、可参与，而非躲在幕后擅自做主。

Every long AI session ends the same way. The context window fills up, and something has to give. Most tools handle this moment for you. A summarizer wakes up, rewrites your history into a paragraph you never see, and the conversation carries on as if nothing happened. Coding agents call it auto-compact. Agent frameworks call it a condenser. The names differ; the shape is the same: the most consequential edit of your session is made by a machine, in the dark, and you were not in the room. What did it keep? What did it drop? Was the decision you spent twenty minutes reaching preserved, or was it flattened into "the user explored several options"? You cannot answer any of these questions, because there is nothing to look at. I have been building <a href="https://github.com/chenxiachan/thoughtdag" rel="noopener noreferrer">ThoughtDAG</a> around one rule that makes this unacceptable: a wire is context. Your conversation lives as a graph on an infinite canvas. Every answer is a node; the wires between nodes decide exactly what the model sees on the next question. Delete one edge, get a different answer. <a href="https://dev.to{{ARTICLE_1_URL}}">I wrote about that rule here</a>. This post is about what that rule forces you to do when the map itself grows too heavy. Compaction has to happen somewhere. The question is whether it happens where you can see it. <h2> Condensing, where you can watch </h2> In ThoughtDAG, condensing is not something that happens to your conversation. It is something you watch happen beside it. Press condense, and the app scans the graph for what I call runs: straight corridors of Q&A where the conversation just marched forward. Three or more turns, one way in, one way out, nothing decisive in the middle. These are the acres of "okay, now try it this way" that every real working session accumulates. The scan is a pure graph walk. No model gets to vote on what counts as collapsible; the structure decides. Each candidate run shows up in a list, checked by default, with the token saving printed next to it. Click a run and the canvas flies there, so you can look at what you are about to fold before you fold it. Uncheck anything you want to keep whole. Then the build starts, and this is the part I care most about: your original tree is never touched. Condensing builds a second tree beside the first one. Checked runs collapse into distill nodes; every other node is copied as it stands; every wire is rebuilt. The two trees sit on the same canvas, side by side. Nothing was edited. Something new was made. <h2> What never gets compressed </h2> The distiller works under fidelity rules that are enforced by structure, not by prompt: <ul> <li> Decisions and pivots never collapse. Any turn you marked as a ruling-out, a decision, or a change of direction is a landmark. Landmarks survive as full nodes in the condensed tree, always.</li> <li> Your highlights survive verbatim. Machine prose gets summarized. Text a human marked as important is carried through word for word. The principle in one line: what the machine said can be compressed; what the human marked must survive as written.</li> <li> The working frontier is off limits. The last three steps above every leaf are where you are thinking right now. The scanner never proposes them, no matter what a model might consider "old".</li> <li> Every distill node knows where it came from. A provenance chip on each distilled node jumps back to the exact run it replaced, in the original tree. The audit trail is one click, not an act of faith.</li> </ul> <h2> The part where it becomes an experiment </h2> Here is what a visible copy buys you that no black box can offer: you can test the compression.