Core concepts

DeterministicRouter and SemanticRouter

Author routing as router units instead of hand-writing every edge. Both compile into the FSM; ControlManager still validates and commits.

DeterministicRouter — hard rules

First matching (predicate, target) rule wins. Compiles to deterministic edges at FSM build time. No model call — ever. Predicates receive ctx with ctx.state.

Python SDKpython

auth = DeterministicRouter(
    id="CheckAuth",
    input_schema=CustomerRequest,   # required when this router is the FSM entry
    rules=[
        (lambda ctx: ctx.state.get("token_valid") is True,  intent_router),
        (lambda ctx: ctx.state.get("token_valid") is False, login_node),
    ],
)

SemanticRouter — labeled intent routes

A model picks among labeled targets (routes={label: node_or_group}). The proposal is validated against the graph before anything executes. v1 supports up to 9 labeled routes plus 1 fallback slot.

Python SDKpython

intent = SemanticRouter(
    id="CustomerIntent",
    routes={
        "refund":  refunds_group,       # label → group (scoped to entry)
        "status":  order_status_node,   # label → node
        "billing": billing_group,
    },
    fallback_node=out_of_scope,         # cannot be a group
)

Grounded in neosyntropy-framework/docs/concepts-explained.md §7–8.