A shared OpenAI SDK client plus the model to call it with, routed by the same verdict the
chat-model path reads. Cached per transport, not as one singleton: the verdict is per-request, so
a single client would pin whichever routing built it first onto everyone after.
A shared OpenAI SDK client plus the model to call it with, routed by the same verdict the chat-model path reads. Cached per transport, not as one singleton: the verdict is per-request, so a single client would pin whichever routing built it first onto everyone after.