Goals and non-goals
What the first version covers and what it deliberately leaves out.
Goals
- Preserve signal. Give every agent session a server-side brain whose life is bound to the session, so compaction and resume of the same named session do not lose working state.
- Make updates global. Answer "what happened in project X" as a first-class, time-ordered query for both agents and the human.
- Give the human a home. One calm surface for finished work, questions, approvals, artifacts, and pruning, reachable from a phone.
- Stay lean. One binary, one engine, and the smallest set of moving parts that delivers the above.
- Stay local. All data on the operator's node; no vendor cloud and no brain off the node.
Non-goals for v1
- No vendor or remote cloud, and no synchronising a brain off the node.
- No multi-tenant organisations. The v1 shape is a single human with a fleet of agents.
- No real-time chat with agents. Interaction is asynchronous, under mailbox semantics.
- No automatic retention or expiry. The human prunes.
- No CRDTs, no custom distributed consensus, and no second search service.
- No reimplementation of AgentFS. The hub wraps the real thing, and FUSE is not used because agents are remote MCP clients.
Milestone shape
Work proceeds in independently shippable milestones: a skeleton with the engine and the AgentFS wrapper, then the feed, the brain, the inbox and Q&A, artifacts, the PWA, search, security, and finally polish. The completion bar is a single binary that runs on a node, where an agent can signal, keep a brain across compaction, publish an artifact, and ask the human a question, and the human can see all of it and prune.
The architecture describes the intended design those milestones build toward.