wrok Bot manages assistants and model sources separately. An assistant has a role in the task; a model connection supplies an inference service. A model choice includes both its source and the capabilities that connection actually provides.
Custom connection management, a persistent catalog, and acceptance, snapshots, context and startup for current-database custom v2 have bounded implementations. Shared inference also has an implementation foundation. Saving configuration alone does not prove reachability, valid credentials or a complete task. Tool use, cancellation and recovery require their own validation.
SDK authorization, registration, dispatch and token-owner integration have bounded deliveries. These records do not establish complete readiness across gateways, account bridges and custom models. Model selection, real conversations, tools, cancellation and recovery still require combined acceptance. A passing authorization component is not proof that every connection works.
Calls to external models or tools send the relevant requests to the configured service. Support, account conditions and charges follow that service’s documentation and actual verification. Local model management and additional platforms follow later phases; this is not a promise to support every model.