Add the thought graph to the web UI. Use your harness models, run agent turns, and explore sessions in Atlas. No separate ThoughtDAG app or CLI required.
Install notes macOS builds are signed and notarized by Apple: double-click and go. Windows builds are not signed yet; on the SmartScreen prompt choose "More info", then "Run anyway".
Opening the context black box between agents and models
A model's next answer depends on the context it reads in that turn. The two reports and one collaboration below measure and expose that context; their methods, data and scripts are public.
Oct 2026Research report
Where an agent turn's tokens go
Besides its answer, an agent turn produces tool calls, tool results and thinking, and these can enter the next turn's context. This note measures 3,237 agent turns from one month of the author's own use. On Claude Code, Codex and DeepSeek Harness, the final answer was 1 to 7% of a turn's tokens, tool calls with their results 42 to 95%, and thinking 4 to 53%; Claude Code and DeepSeek Harness kept the previous turn's thinking in the next turn's prompt. ThoughtDAG therefore gives each turn's tool record its own switch, off by default.
Context repair: how one error contaminates the rest of a conversation
In a multi-turn conversation, a wrong claim that enters the context becomes a premise for later answers and contaminates the context; deleting the claim alone can leave its effect in the answers below it. This pilot compares three repairs on 9 model endpoints. Across 162 derailed outcomes (18 cases on 9 endpoints), deleting only the source repaired 152, recomputing the descendants in dependency order repaired 161, and removing the contaminated subgraph repaired all 162. ThoughtDAG supports the last two.
An agent run leaves a long record of steps, from which it is hard to see what the agent did and on what basis. TraceVista draws agent workflows as maps that unfold from the overview down to each action and its original record; answers to questions about a run link back to the records they rest on. The reading layer builds on ThoughtDAG's zoom summaries.
With @Superone77 and contributors · reading layer and map Q&A by Xia Chen
Can I edit context without deleting the conversation?
Yes. Removing a wire removes that route into a later request while keeping the nodes. Check other input paths, attachments, and Recall too. An existing answer does not change until you edit or regenerate it.
How does branching support nonlinear thinking?
Keep alternative questions on separate branches, compare them on the canvas, and connect selected findings to a new downstream question. A branch inherits its upstream conversation path; selecting a passage does not limit context to those words alone.
Will all my past chats appear automatically?
No. The desktop app indexes supported local records you have connected. It cannot restore unsaved history or automatically access chats stored only in a web account. When you use a remote model, the material included in a request is sent to that service.
How does ThoughtDAG decide what the model reads?
It walks the incoming edges of a node, orders the relevant ancestors, and builds the message sequence sent to the selected model. Explicit references and enabled Recall can add material. Preview the request before sending to see each source, its order and the token count.