A training-free orchestration framework for building interactive omni-modal assistants by composing off-the-shelf modality experts, explicit LLM routing, text-centric cross-modal memory, and interruption-aware streaming interaction.