示例：Neural Imprint 生命周期

这个页面记录 settings/lifecycle integration contract，不是 standalone runnable bridge。This page documents the settings/lifecycle integration contract, not a standalone runnable bridge.

它做四件事：

记录本地用户信号。
展示 profile readiness。
校验 Neural Imprint capsule。
仅在兼容性检查通过后恢复 capsule。

Profile job 和产物生成应放在你的本地数据层后面，或直接参考 Edge Scaffold 流程。

SwiftUI 设置页流程

import EdgeHalo
import Foundation
import SwiftUI

struct PersonalizationView: View {
    @StateObject private var model = PersonalizationViewModel()

    var body: some View {
        Form {
            Section("State") {
                LabeledContent("Evolution", value: model.evolutionText)
                LabeledContent("Capsule", value: model.capsuleText)
            }

            if let profile = model.profile {
                Section("Local Profile") {
                    Text(profile.narrative.isEmpty ? "Profile ready" : profile.narrative)
                    LabeledContent("Samples", value: "\(profile.sampleCount)")
                    LabeledContent("Stability", value: profile.stabilityText)
                }
            }

            Section("Signals") {
                Button("Record positive feedback") {
                    model.recordFeedback(accepted: true)
                }
                Button("Record correction") {
                    model.recordCorrection()
                }
            }

            Section("Neural Imprint") {
                Button("Refresh state") {
                    Task { await model.refreshState() }
                }
                Button("Restore local capsule") {
                    Task { await model.restoreLocalCapsule() }
                }
                Button("Reset personalization") {
                    Task { await model.resetPersonalization() }
                }
            }

            if !model.status.isEmpty {
                Text(model.status)
                    .font(.caption)
                    .foregroundStyle(.secondary)
            }
        }
        .navigationTitle("Personalization")
        .task { await model.refreshState() }
    }
}

@MainActor
final class PersonalizationViewModel: ObservableObject {
    @Published var evolutionText = "Idle"
    @Published var capsuleText = "Base model"
    @Published var status = ""
    @Published var profile: UserProfile?

    private let halo: EdgeHalo
    private let events: AsyncStream<HaloDataEvent>.Continuation
    private let capsuleStore = LocalCapsuleStore()

    init() {
        let stream = AsyncStream.makeStream(of: HaloDataEvent.self)
        events = stream.continuation

        halo = EdgeHalo(
            engine: AppEngineSession(),
            generator: AppTextGenerator(),
            dataStream: stream.stream
        )
    }

    func recordFeedback(accepted: Bool) {
        events.yield(.feedback(
            accepted: accepted,
            conversationID: UUID().uuidString
        ))
        status = "Feedback recorded locally"
    }

    func recordCorrection() {
        events.yield(.correction(
            original: "Original assistant reply",
            corrected: "User-corrected reply",
            conversationID: UUID().uuidString
        ))
        status = "Correction recorded locally"
    }

    func refreshState() async {
        profile = await halo.currentProfile

        switch await halo.evolutionState {
        case .idle:
            evolutionText = "Idle"
        case .collecting(let progress):
            evolutionText = "Collecting \(progress.collected)/\(progress.threshold)"
        case .readyToBuildCapsule:
            evolutionText = "Ready to build"
        case .buildingCapsule:
            evolutionText = "Building"
        case .validating(let capsuleID):
            evolutionText = "Validating \(capsuleID)"
        case .evolved(let capsuleID):
            evolutionText = "Personalized \(capsuleID)"
        }

        switch await halo.haloState {
        case .idle:
            capsuleText = "Base model"
        case .collecting(let factsCount, let threshold):
            capsuleText = "Collecting \(factsCount)/\(threshold)"
        case .profiling:
            capsuleText = "Preparing profile"
        case .buildingCapsule:
            capsuleText = "Building capsule"
        case .validating(let capsuleID):
            capsuleText = "Validating \(capsuleID)"
        case .active(let capsuleID):
            capsuleText = "Neural Imprint active \(capsuleID)"
        case .incompatible(_, let reason):
            capsuleText = "Incompatible: \(reason)"
        case .failed(let reason):
            capsuleText = "Failed: \(reason)"
        }
    }

    func restoreLocalCapsule() async {
        do {
            let capsule = try capsuleStore.loadLatestCapsule()
            let current = try capsuleStore.currentRuntimeRequirements()
            let compatibility = await halo.validateCapsule(
                capsule.manifest,
                currentRequirements: current
            )

            guard compatibility.isCompatible else {
                status = "Capsule does not match the loaded model. Regenerate it."
                await refreshState()
                return
            }

            try await halo.activateCapsule(
                capsule,
                currentRequirements: current
            )
            status = "Neural Imprint restored"
            await refreshState()
        } catch {
            status = "Restore failed: \(error.localizedDescription)"
        }
    }

    func resetPersonalization() async {
        capsuleStore.removeLocalArtifacts()
        status = "Personalization artifacts removed. Reload the base model."
        await refreshState()
    }
}

private extension UserProfile {
    var stabilityText: String {
        stabilityScore.formatted(.number.precision(.fractionLength(2)))
    }
}

Runtime bridge reference

Use the Edge Scaffold 运行时 bridge as the runnable reference. 可运行参考见 edge-scaffold 仓库中的 EdgeScaffold/AI/ScaffoldHaloRuntimeAdapter.swift。

Scaffold adapter 属于 app-layer code，不属于基础框架：

ScaffoldHaloRuntimeAdapter 实现 HaloTextGenerator 和 HaloEngineSession。
它通过 chat 使用的已加载 LLMEngine 或 VLMEngine session 做 tokenize 和 generate。
它通过 LLMEngine.captureHiddenStates(...) 或 VLMEngine.captureHiddenStates(...) 捕获 profile activations。
它通过 AIManager.restoreNeuralImprintCacheForHalo(from:) 恢复兼容的 Neural Imprint 产物。

不要在 app code 中 stub generation output 或 hidden states。如果已加载 runtime 不可用、不兼容，或缺少需要的 capture/restore path，app 应失败即关闭，并保持 base model active。

Capsule storage 与 runtime requirements

生产 App 通常从 Edge Studio export、Edge Scaffold flow、App 拥有 local storage 或可信 EdgeMesh transfer 加载 capsule manifest 与本地产物 URL。调用 validateCapsule 和 activateCapsule 前，runtime requirements 应来自当前已加载模型、tokenizer、工具 schema 和 runtime identity。

Storage 和 reset policy 留在 App 层。Edge Kit 与 Edge Halo 提供可复用 runtime 和 lifecycle infrastructure；它们不应拥有 App 专属 records 或产品策略。

集成注意事项

原始用户文本留在本地 app storage。
诊断使用 hashes 和状态回执s。
只有 validateCapsule 成功后才 restore。
个性化缺失或不兼容时，保持 base model 可用。
不要把 profile text 加进每次 system prompt；应恢复 artifact。
提供用户可见的 reset 路径。

下一步

阅读模型进化能力指南。
如需了解 Neural Imprint 产物的跨 App 兼容性检查，请阅读 Artifact 复用。
当产物需要在受信任的用户自有设备间移动时，使用设备 Mesh。
使用 Edge Scaffold 作为参考 iOS 实现。

SwiftUI 设置页流程​

Runtime bridge reference​

Capsule storage 与 runtime requirements​

集成注意事项​

下一步​

SwiftUI 设置页流程

Runtime bridge reference

Capsule storage 与 runtime requirements

集成注意事项

下一步