Show HN: ThoughtDAG – An editable context graph for LLM conversations

Chat history is long but context remains invisible: the interface shows what was said, not which parts actually enter the next request. ThoughtDAG solves this by treating the graph as the context, replacing linear chat history with an explicit, editable structure. Users can ask from a source passage, clip relevant parts into source-linked nodes, and preview the exact tokens the model will receive, including ordering and token count.

The core workflow is externalize, inspect, and edit. Users delete an edge (e.g., an unrelated 'dinner detour') and the removed branch truly leaves the request — no hidden memory selector. In the example, deleting one edge reduced the outgoing context by 47 tokens (from 1,284), and regenerating with the identical prompt produced a clean answer free of the dinner suggestion. ThoughtDAG also records provenance for every source and shows context diffs so changes are visible and reproducible.

ThoughtDAG frames itself as a 'context protocol before generation' rather than a generic canvas. By making context visible, editable, and inspectable, it aims to eliminate the pollution that occurs when unrelated history silently alters model outputs, giving users deterministic control over what the LLM sees.