ammar-abagumbul/tokencake

TokenCake is an agent-aware KV-cache scheduler for multi-agent LLM serving. In the documented current milestone, an agent workflow follows a dependency DAG, calls an LLM and external tools, and brackets tool stalls with frontend events. The frontend delegates those events to the temporal controller, which forecasts stall duration, gates offload decisions, and plans cache uploads; the connector integrates KV migration with vLLM. The README also describes a spatial scheduler, but the stated milestone targets temporal scheduling and no spatial implementation path appears in the tree. Unsampled policy wiring is shown only where documented.