The kanban-as-task-queue framing for AI agents is smart — it forces you to decompose work before handing it off, which tends to produce much better outputs than free-form chat. One challenge we ran into with multi-repo agent workflows is context window management when files across repos have conflicting conventions. Does Hive do anything to normalize context before feeding it to the agent?
The idea of maintaining a persistent markdown wiki instead of stuffing everything into a vector DB resonates with me. We built AFFiNE (an open-source knowledge base) and one of the biggest UX pain points was that users couldn't easily inspect or trust what the system "knew." Human-readable flat files solve the transparency problem elegantly. Curious how you handle conflicting updates when two sources say different things about the same entity?