Apply the route with modifiers

WWDC26 session 241 shows a brainstorm Profile using PCC plus .reasoningLevel(.deep), and a cheaper analysis Profile on SystemLanguageModel. Your router is the function that picks that branch. The Profile body is the switch.

import FoundationModels
import CoreAI

// Illustrative. Confirm LanguageModel initializers in the installed SDK.
struct RoutedProfile: LanguageModelSession.DynamicProfile {
    let route: String
    let system: SystemLanguageModel
    let pcc: PrivateCloudComputeLanguageModel
    let coreAI: CoreAILanguageModel?

    var body: some LanguageModelSession.DynamicProfile {
        switch route {
        case "pcc":
            Profile {
                Instructions { "Plan the next step. Stay inside the packed facts." }
            }
            .model(pcc)
            .reasoningLevel(.deep)
        case "coreai":
            Profile {
                Instructions { "Answer in the baked voice. Do not invent facts." }
            }
            .model(coreAI!)
        default:
            Profile {
                Instructions { "Answer quickly from the packed facts." }
            }
            .model(system)
        }
    }
}

Force-unwrapping coreAI is only safe after choose proved it is available. Prefer an enum with associated models over a string key in production.

When you hop from PCC to the system model, trim the transcript with historyTransform so the smaller window still fits. Session 242 covers that handoff. Do not send private tool dumps to a larger model if the user bound the turn to the device.

Next, wrap choose and the Profile in one router type.