ThoughtDAG is an editable context graph designed to replace linear chat histories with a structured, inspectable protocol for managing Large Language Model (LLM) conversation context. Unlike traditional interfaces where the specific history entering a prompt remains hidden, ThoughtDAG treats the graph as the context itself; it constructs the message sequence by walking the incoming edges of a node and ordering its relevant ancestors. This allows users to manually prune, branch, or merge specific reasoning paths to control exactly what information the model receives. Technical evaluations demonstrate the efficacy of this approach in context repair: when addressing derailed model outputs, removing the contaminated subgraph successfully repaired 162 out of 162 cases across nine model endpoints. Furthermore, the tool offers a "Session Atlas" feature to unify scattered agent contexts, enabling cross-agent continuity by linking sessions via exact phrases, URLs, or document IDs. By providing a preview of the upcoming request—including node order and token counts—ThoughtDAG enables reproducible context management, allowing users to see how removing a single edge (e.g., reducing context by 47 tokens) directly alters the model's output. It is available via a desktop app and plugins for Claude Code, Codex, and DeepSeek Harness.