Large language model calls are routed through an in-namespace LiteLLM proxy that fronts Azure OpenAI gpt-5.4 so that model access is mediated by the local proxy layer.