Robot production supply.

A persistent multi-agent simulation separates the warehouse engine from its robot controllers. Controllers can be model-backed or scripted; models and fleet size are configurable. The engine owns routing, collision resolution, batteries and parcel ownership, while models choose from the actions available to their robot and can communicate to resolve obstructions. Inference runs asynchronously, so other robots keep moving while a decision is being computed.

Production jobs span component collection, shared kit assembly, packing and truck loading. SQLite transactions save simulation checkpoints alongside recorded changes; an independent gateway streams live activity and serves historical playback without rerunning model calls. Maintenance is part of the simulation: robots secure their cargo, exit through repair and return to the saved world. The viewer exposes decision reasons, inference timing and transient update bubbles, while its timeline reconstructs recorded states across restarts.