Design specs¶
Per-service design specifications for the binabik brain's components — one file per MCP server / component. They sit behind the Architecture overview and the Coding guidelines.
The specs¶
One spec per MCP server / shared component of the current stack. Every service below is shipped;
what differs is when it runs — robot-mcp-memory comes up with every brain, while
binabik-r1-vision is opt-in on the robot (each spec says so at the top). Deployment
itself is not in these specs: brain-side is Running a brain, robot-side
is R1 robot stack.
| Spec | What it covers |
|---|---|
| Grasp skill | L3 grasp strategy engine — side / top / two_hand FSMs over the robot's stateless primitives. |
| World state | The one "what is the world like" snapshot — body, cameras, scene — read by the planner, the L3 skills and the watchdog. Sits in front of the abstraction and is ROS-free. |
| Memory (MCP) | Central, group-scoped fleet/process memory — teaching, recall, Fleet admin, group learning. |
| MCP kit | Shared fleet library — ToolClient, EventPoster, ParamStore/lessons, RecoveryAgent, server helpers. |
| Voice chat agent | The brain app — planner, chat / Visual-Command / observer UI, mission executor. |
Four services are gone, and so are their specs
robot-mcp-perception (the v2 standing vision node) and mcp-skill-control went in
BIN-95 — superseded by galaxea's on-demand perception plus binabik-r1-vision, and by the
brain calling MCP tools directly. mcp-perception-buffer (the ambient "what do you see"
buffer, :9202) and ros2-memory (episodic recall, :9203) were retired in BIN-306:
nothing launches either — no r1ctl opt-in, no brainctl flag, no manifest entry. "What
do you see" is now the world state, answered on
demand with no VLM poll behind it, and nothing replaces temporal recall.
Their specs have been deleted rather than shelved. Read them in git history if you
want the designs: git log --diff-filter=D -- docs/binabik-software/archive/.