A/B route and roll back

Add an experiment field on the feature: off, or a pair of record ids with a bucket. The Dynamic Profile body reads the bucket and returns one Profile. The transcript stays. Only the instructions and tools change on the next prompt.

struct SupportProfile: LanguageModelSession.DynamicProfile {
    var registry: PromptRegistry
    var experiment: PromptExperiment

    var body: some LanguageModelSession.DynamicProfile {
        switch experiment.choice {
        case .record(let id):
            Profile { SupportCopy(record: registry.record(id: id)) }
        case .control:
            Profile { SupportCopy(record: registry.live!) }
        }
    }
}

mutating func rollback(to id: String) {
    records = records.map { record in
        var copy = record
        if copy.id == id { copy.status = "live" }
        else if copy.status == "live" { copy.status = "retired" }
        return copy
    }
}

Confirm experiment types in your own code. Apple does not ship A/B. You assign the bucket. Keep the assignment stable for a user so the session does not flip every prompt. A flip can invalidate the KV cache and raise TTFT. Measure that in Stream tokens and measure speed.

Do not roll back from a hallway opinion. Run Test AI behavior before you ship on both ids, then switch live. If a live prompt is harmful, roll back first and eval second.

Next, ship the file so the team can edit it.