ThoughtDAG: a read-only source / editable mirror boundary for agent sessions
Lopsided_Scarcity979 · reddit · 2026-09-02
- The author open-sourced ThoughtDAG, a context-graph project, asking for architecture critique.
- Core design: reads local Codex and Claude Code session files, groups by project, projects a session onto a graph; each user/agent turn becomes a node with paired tool calls, results, truncation markers, and sub-thread identity attached; source files are never modified, and an open mirror appends new turns as sessions grow.
- Key boundary: visibility and model input are separate states — importing three sessions onto a canvas does not put them into the next prompt; only context deliberately connected to a downstream node gets compiled.
- Acknowledged tradeoffs: no automatic relevance selector, no source writeback, only two native session adapters.
- Asks reviewers where it would fail first: format drift, tool-result provenance, incremental sync, or context compilation.
More from coding & agent
- Superpowers creator Jesse Vincent: managing AI agents is like managing junior engineers, 10x productivity — carsonfarmer · 2026-09-03
- Prompt engineering is basically mechanism design for multi-agent systems, argues engineer — carsonfarmer · 2026-09-03
- LukeW: developers now supervise hundreds of agents; Intent open-sources the next abstraction layer — LukeW · 2026-09-03
- Agentic System Trilogy: Design to AWS Deployment with Full Code — kmeanskaran · 2026-09-03
- Agent system trilogy: design, backend, and AWS ECS production deployment — kmeanskaran · 2026-09-03
- MazeBench Author Admits Algorithm-Generated Levels Are Useless for Coding Agents — patience_cave · 2026-09-03