Count tokens and set a budget

xcode-select --install
xcodebuild -version
# Add FoundationModels. Target iOS 27 or macOS 27.

import FoundationModels

// Illustrative. Confirm names in the installed SDK.
struct TokenBudget {
    let window: Int
    let replyReserve: Int
    let question: Int
    let instructions: Int

    var leftover: Int {
        window - replyReserve - question - instructions
    }
}

func makeBudget(model: SystemLanguageModel, question: String, instructions: String) async throws -> TokenBudget {
    let window = model.contextSize
    let q = try await model.tokenCount(for: question)
    let i = try await model.tokenCount(for: instructions)
    return TokenBudget(
        window: window,
        replyReserve: max(256, window / 8),
        question: q,
        instructions: i
    )
}

Count with the same model that will answer. A SystemLanguageModel count is wrong for a PrivateCloudComputeLanguageModel pack, and the reverse is also wrong. CoreAILanguageModel has its own window. Read that model's contextSize if the SDK exposes it.

If token counting is missing in your drop, say so and use a conservative character estimate only as a stand-in. Write that fallback in the README. Do not pretend it is exact.

Next, rank the slices that compete for leftover tokens.