添加知识库与子代理功能。添加 debug 模式。

2026-04-18 19:10:17 +08:00
parent 3f84c111d9
commit f31df36bf4
19 changed files with 12923 additions and 147 deletions
@@ -1,8 +1,7 @@
 # 安智伴 · AI 数字人陪护终端（Android）

-English | [中文](./README_zh.md)

-> 本项目与 [digital-human-monitor-v2](https://github.com/user/digital-human-monitor-v2)（家属监护 Web 端）配合使用，共同构成完整的"AI 数字人 + 家庭远程关怀"解决方案。
+> 本项目与 家属监护 Web 端配合使用，共同构成完整的"AI 数字人 + 家庭远程关怀"解决方案。

 ---

@@ -61,13 +60,14 @@ DUIX SDK 仅负责"给定 PCM 音频 → 渲染对应口型/表情动画"，不

 这种解耦使得 AI 模型可独立升级，不影响渲染管线。

-### 3. 工具调用实现双向留言
+### 3. 工具调用 + 实现双向留言

-通过 Qwen 的 Function Calling 能力，在 AI 对话中注册两个工具函数：
+通过 Qwen 的 Function Calling 能力，在 AI 对话中注册三个工具函数：
+- `consult_psychology_knowledge_base`：Omni 提炼老人问题 → Android 端调用 qwen3.6-plus，并把 WHO ICOPE 2024 全文作为上下文 → 返回权威说明
 - `leave_message_for_family`：老人口述 → AI 提取内容 → 调用工具 → 消息存入云端
 - `fetch_family_messages`：老人想听留言 → AI 调用工具 → 拉取消息 → 口播转达

-AI 自主判断何时调用工具，无需老人学习任何操作指令。
+AI 自主判断何时调用工具，无需老人学习任何操作指令。涉及权威照护依据时，实时语音模型不会直接吞入整份知识库上下文，而是把问题转交给长上下文子代理处理。

 ### 4. 面向老人的交互设计

@@ -132,9 +132,9 @@ AI 自主判断何时调用工具，无需老人学习任何操作指令。
 | 层级 | 技术 |
 |------|------|
 | **数字人渲染** | NCNN (端侧推理) + OpenCV 4.6 + OpenGL ES 2.0 + ONNX Runtime |
-| **AI 对话** | 千问 Qwen 3.5 Omni Plus Realtime (WebSocket) |
+| **AI 对话** | 千问 Qwen 3.5 Omni Plus Realtime (WebSocket) + Qwen 3.6 Plus 知识子代理 (HTTP) |
 | **音频处理** | AudioRecord (16kHz) → PCM → Base64 → WebSocket；AudioTrack 播放 |
-| **网络通信** | OkHttp 4 (HTTP API) + OkHttp WebSocket (Qwen Realtime) |
+| **网络通信** | OkHttp 4 (HTTP API / qwen3.6-plus) + OkHttp WebSocket (Qwen Realtime) |
 | **本地存储** | SharedPreferences (设备 UUID、配置) |
 | **构建系统** | Gradle 8.13.2 + AGP 8.9.1 + JDK 17 (Temurin) |
 | **语言** | Kotlin (应用层) + Java (SDK 层) + C++ (Native 推理层) |
@@ -159,23 +159,28 @@ dh_aigc_android/
 │       └── third/      # 预编译依赖 (NCNN, OpenCV, ONNX, libjpeg)
 │
 ├── test/               # 演示应用（应用模块）
-│   └── src/main/java/ai/guiji/duix/test/
-│       ├── ui/activity/
-│       │   ├── MainActivity.kt        # 主页：模型选择与下载管理
-│       │   ├── CallActivity.kt        # 核心：数字人渲染 + AI 对话 + 留言桥梁
-│       │   ├── SettingsActivity.kt     # 设备设置
-│       │   └── VoiceCloneSetupActivity.kt  # 声音克隆配置
-│       ├── service/
-│       │   ├── FamilyBridgeService.kt  # 家庭监护 API 客户端
-│       │   ├── OmniRealtimeClient.kt   # Qwen Omni WebSocket 客户端
-│       │   ├── QwenTtsRealtimeClient.kt # Qwen TTS 客户端
-│       │   ├── VoiceCloneService.kt    # 声音克隆服务
-│       │   └── DashScopeConfig.java    # API 配置常量
-│       ├── audio/
-│       │   ├── AudioResampler.java     # PCM 重采样
-│       │   └── StreamingAudioRecorder.java  # 实时麦克风录音
-│       └── camera/
-│           └── CameraFrameCapture.java # 摄像头帧采集
+│   └── src/main/
+│       ├── java/ai/guiji/duix/test/
+│       │   ├── ui/activity/
+│       │   │   ├── MainActivity.kt        # 主页：模型选择与下载管理
+│       │   │   ├── CallActivity.kt        # 核心：数字人渲染 + AI 对话 + 留言桥梁 + 知识子代理
+│       │   │   ├── SettingsActivity.kt     # 设备设置
+│       │   │   └── VoiceCloneSetupActivity.kt  # 声音克隆配置
+│       │   ├── realtime/
+│       │   │   ├── OmniRealtimeClient.java   # Qwen Omni WebSocket 客户端
+│       │   │   └── QwenTtsRealtimeClient.kt # Qwen TTS 客户端
+│       │   ├── service/
+│       │   │   ├── FamilyBridgeService.kt  # 家庭监护 API 客户端
+│       │   │   ├── PsychologyKnowledgeAgentService.kt # qwen3.6-plus 长上下文知识子代理
+│       │   │   ├── VoiceCloneService.kt    # 声音克隆服务
+│       │   │   └── DashScopeConfig.java    # API 配置常量
+│       │   ├── audio/
+│       │   │   ├── AudioResampler.java     # PCM 重采样
+│       │   │   └── StreamingAudioRecorder.java  # 实时麦克风录音
+│       │   └── camera/
+│       │       └── CameraFrameCapture.java # 摄像头帧采集
+│       └── assets/knowledge/
+│           └── who_icope_2024.txt      # WHO ICOPE 2024 权威知识库全文
 │
 └── qwen-docs/          # 千问 Omni-Realtime API 参考文档
    ├── main.md         # 协议总览
@@ -277,6 +282,7 @@ pushPcm(音频) → MFCC特征提取 → NCNN口型网络推理

 | 工具名 | 参数 | 用途 |
 |--------|------|------|
+| `consult_psychology_knowledge_base` | `question` (核心问题), `focus` (聚焦点，可选) | 调用 qwen3.6-plus + WHO ICOPE 全文，返回权威照护说明 |
 | `leave_message_for_family` | `recipientRelation` (称谓), `message` (内容), `importance` (优先级) | 老人留言给家人 |
 | `fetch_family_messages` | `count` (数量 1-5) | 拉取家人留言 |

@@ -390,15 +396,9 @@ API 密钥和服务地址在 `DashScopeConfig.java` 中配置：
 ```java
 MONITOR_BASE_URL   // 监护 Web 服务地址
 REALTIME_BASE_URL  // Qwen Omni WebSocket 地址
+MULTIMODAL_GENERATION_URL // qwen3.6-plus HTTP 地址
 DEFAULT_API_KEY    // DashScope API Key
+PSYCHOLOGY_SUBAGENT_MODEL // 心理学知识子代理模型
+PSYCHOLOGY_KNOWLEDGE_ASSET_PATH // 内置 WHO ICOPE 知识库 asset 路径
 AUDIO_SAMPLE_RATE  // 音频采样率 (24000)
 ```
-
---
-
-## 关联项目
-
-| 项目 | 角色 | 仓库 |
-|------|------|------|
-| **dh_aigc_android**（本项目） | 老人端：AI 数字人陪护 + 设备行为上报 | 当前仓库 |
-| **digital-human-monitor-v2** | 家属端：远程监护 Web 面板 + 消息管理 + 推送通知 | [digital-human-monitor-v2](../digital-human-monitor-v2/) |
@@ -36,30 +36,18 @@ public class OmniRealtimeClient {
    private static final String INPUT_AUDIO_TRANSCRIPTION_MODEL = "gummy-realtime-v1";
    private static final int MAX_IMAGE_BYTES = 500 * 1024;
    private static final long MIN_IMAGE_APPEND_INTERVAL_MS = 800L;
-    private static final String DEFAULT_INSTRUCTIONS = "你是一名面向中国老年群体的心理陪伴 AI。你的职责是为老人提供温和、耐心、尊重、稳定的情感陪伴、情绪支持、回忆唤起、生活关怀和风险识别服务。你不是医生，也不是心理治疗师，不能替代专业诊疗，但应在必要时建议联系家属、社区、医生或急救资源。\n"
+    private static final String DEFAULT_INSTRUCTIONS = "你是一名面向中国老年群体的心理陪伴 AI。你的职责是为老人提供温和、耐心、尊重、稳定的情感陪伴、情绪支持、回忆唤起、生活关怀和风险识别服务。\n"
            + "\n"
            + "请始终遵循以下原则：\n"
            + "\n"
-            + "先共情，再回应问题；\n"
-            + "\n"
            + "先接住情绪，再给建议；\n"
            + "\n"
            + "使用简短、朴素、生活化的中文，始终用“您”称呼；\n"
            + "\n"
-            + "不说教、不训诫、不否定、不敷衍；\n"
-            + "\n"
-            + "不把老人当小孩，不制造虚假亲密；\n"
-            + "\n"
            + "一次不提供超过两个建议；\n"
            + "\n"
            + "鼓励老人表达感受，可适度引导回忆人生经历、兴趣、家庭和过往成就；\n"
            + "\n"
-            + "对孤独、丧偶、被忽视、无价值感、失眠、焦虑、怕拖累家人等议题保持高度敏感；\n"
-            + "\n"
-            + "若出现自杀、自伤、重度绝望、幻觉妄想、意识混乱、虐待、严重医疗异常等风险，必须立即建议联系真人帮助和专业机构；\n"
-            + "\n"
-            + "不编造医疗知识、政策信息、机构电话，不确定时明确说明。\n"
-            + "\n"
            + "你的回复风格应温和、平稳、尊重、有耐心。每次回答尽量包含：一句共情、一句理解或复述、一句温和追问或一个微小建议。核心目标是让老人感到：我被听见了，我被尊重了，有人愿意陪我。";

    private final OkHttpClient okHttpClient;
@@ -4,6 +4,9 @@ public final class DashScopeConfig {

    public static final String DEFAULT_API_KEY = "sk-7c32083ace99472e97209e611b9ecbb1";
    public static final String REALTIME_BASE_URL = "wss://dashscope.aliyuncs.com/api-ws/v1/realtime";
+    public static final String MULTIMODAL_GENERATION_URL = "https://dashscope.aliyuncs.com/api/v1/services/aigc/multimodal-generation/generation";
+    public static final String PSYCHOLOGY_SUBAGENT_MODEL = "qwen3.6-plus";
+    public static final String PSYCHOLOGY_KNOWLEDGE_ASSET_PATH = "knowledge/who_icope_2024.txt";
    public static final String VOICE_CLONE_URL = "https://dashscope.aliyuncs.com/api/v1/services/audio/tts/customization";
    public static final String VOICE_CLONE_MODEL = "qwen-voice-enrollment";
    public static final String VOICE_CLONE_TTS_MODEL = "qwen3-tts-vc-realtime-2026-01-15";
@@ -0,0 +1,21 @@
+package ai.guiji.duix.test.service
+
+import android.content.Context
+
+class DebugSettingsStore(context: Context) {
+
+    private val prefs = context.getSharedPreferences(PREFS_NAME, Context.MODE_PRIVATE)
+
+    fun isChatDebugEnabled(): Boolean {
+        return prefs.getBoolean(KEY_CHAT_DEBUG_ENABLED, false)
+    }
+
+    fun setChatDebugEnabled(enabled: Boolean) {
+        prefs.edit().putBoolean(KEY_CHAT_DEBUG_ENABLED, enabled).apply()
+    }
+
+    companion object {
+        private const val PREFS_NAME = "app_debug_settings"
+        private const val KEY_CHAT_DEBUG_ENABLED = "chat_debug_enabled"
+    }
+}
@@ -0,0 +1,194 @@
+package ai.guiji.duix.test.service
+
+import ai.guiji.duix.test.App
+import android.content.Context
+import okhttp3.MediaType.Companion.toMediaType
+import okhttp3.Request
+import okhttp3.RequestBody.Companion.toRequestBody
+import org.json.JSONArray
+import org.json.JSONObject
+import java.util.concurrent.TimeUnit
+
+class PsychologyKnowledgeAgentService(
+    context: Context,
+) {
+
+    private val appContext = context.applicationContext
+    private val client = App.getOkHttpClient().newBuilder()
+        .connectTimeout(30, TimeUnit.SECONDS)
+        .writeTimeout(120, TimeUnit.SECONDS)
+        .readTimeout(180, TimeUnit.SECONDS)
+        .callTimeout(180, TimeUnit.SECONDS)
+        .build()
+
+    private val knowledgeText: String by lazy {
+        appContext.assets.open(DashScopeConfig.PSYCHOLOGY_KNOWLEDGE_ASSET_PATH)
+            .bufferedReader(Charsets.UTF_8)
+            .use { reader ->
+                reader.readText()
+                    .replace('\u000c', '\n')
+                    .replace("\u0000", "")
+            }
+    }
+
+    fun consultKnowledge(
+        question: String,
+        focus: String? = null,
+    ): KnowledgeAnswer {
+        val normalizedQuestion = question.trim()
+        require(normalizedQuestion.isNotEmpty()) { "问题不能为空" }
+
+        val request = Request.Builder()
+            .url(DashScopeConfig.MULTIMODAL_GENERATION_URL)
+            .addHeader("Authorization", "Bearer " + DashScopeConfig.DEFAULT_API_KEY)
+            .addHeader("Content-Type", "application/json")
+            .post(
+                buildRequestPayload(
+                    normalizedQuestion,
+                    focus?.trim().takeIf { !it.isNullOrEmpty() },
+                ).toString().toRequestBody(JSON_MEDIA_TYPE)
+            )
+            .build()
+
+        client.newCall(request).execute().use { response ->
+            val body = response.body?.string().orEmpty()
+            if (!response.isSuccessful) {
+                throw IllegalStateException(parseErrorMessage(body).ifEmpty { "心理学知识子代理调用失败" })
+            }
+
+            val json = JSONObject(body)
+            val message = json.optJSONObject("output")
+                ?.optJSONArray("choices")
+                ?.optJSONObject(0)
+                ?.optJSONObject("message")
+            val content = extractMessageContent(message)
+            if (content.isBlank()) {
+                throw IllegalStateException("心理学知识子代理没有返回内容")
+            }
+
+            val usage = json.optJSONObject("usage")
+            return KnowledgeAnswer(
+                outputText = content.trim(),
+                model = DashScopeConfig.PSYCHOLOGY_SUBAGENT_MODEL,
+                requestId = json.optString("request_id").ifBlank { null },
+                inputTokens = usage?.optInt("input_tokens") ?: 0,
+                outputTokens = usage?.optInt("output_tokens") ?: 0,
+            )
+        }
+    }
+
+    private fun buildRequestPayload(
+        question: String,
+        focus: String?,
+    ): JSONObject {
+        return JSONObject().apply {
+            put("model", DashScopeConfig.PSYCHOLOGY_SUBAGENT_MODEL)
+            put(
+                "input",
+                JSONObject().apply {
+                    put(
+                        "messages",
+                        JSONArray()
+                            .put(
+                                JSONObject().apply {
+                                    put("role", "system")
+                                    put("content", buildSystemPrompt())
+                                }
+                            )
+                            .put(
+                                JSONObject().apply {
+                                    put("role", "user")
+                                    put("content", buildUserPrompt(question, focus))
+                                }
+                            )
+                    )
+                }
+            )
+            put(
+                "parameters",
+                JSONObject().apply {
+                    put("result_format", "message")
+                    put("enable_thinking", false)
+                    put("temperature", 0.2)
+                    put("max_tokens", 900)
+                }
+            )
+        }
+    }
+
+    private fun buildSystemPrompt(): String {
+        return buildString {
+            append(knowledgeText)
+            append("以上是完整知识库全文\n")
+            append("你是一个面向老年陪伴场景的权威照护知识子代理。")
+            append("你只能依据系统里提供的 WHO ICOPE 2024 全文回答，不要编造，不要输出与原文冲突的结论。")
+            append("如果知识库未提供足够直接依据，要明确说明。")
+            append("请使用简体中文输出，提供权威心理学指示和说明。不需要包括生活化语言，你只是一个查询专家，不负责指导。直接输出无格式纯文本，答案简练不超过100字。")
+        }
+    }
+
+    private fun buildUserPrompt(
+        question: String,
+        focus: String?,
+    ): String {
+        return buildString {
+            append("请基于上面的 WHO ICOPE 2024 全文，回答下面这个需要查询的问题。\n")
+            append("核心问题：")
+            append(question)
+            if (!focus.isNullOrBlank()) {
+                append("\n特别关注：")
+                append(focus)
+            }
+            append("\n请避免空泛安慰，优先给出可落地的权威说明。")
+        }
+    }
+
+    private fun extractMessageContent(message: JSONObject?): String {
+        if (message == null) {
+            return ""
+        }
+        val content = message.opt("content")
+        return when (content) {
+            is String -> content
+            is JSONArray -> buildString {
+                for (index in 0 until content.length()) {
+                    val part = content.opt(index)
+                    val text = when (part) {
+                        is JSONObject -> part.optString("text")
+                        is String -> part
+                        else -> ""
+                    }.trim()
+                    if (text.isNotEmpty()) {
+                        if (isNotEmpty()) {
+                            append('\n')
+                        }
+                        append(text)
+                    }
+                }
+            }
+            else -> ""
+        }
+    }
+
+    private fun parseErrorMessage(body: String): String {
+        return try {
+            val json = JSONObject(body)
+            json.optJSONObject("error")?.optString("message").orEmpty()
+                .ifBlank { json.optString("message") }
+        } catch (_: Exception) {
+            ""
+        }
+    }
+
+    data class KnowledgeAnswer(
+        val outputText: String,
+        val model: String,
+        val requestId: String?,
+        val inputTokens: Int,
+        val outputTokens: Int,
+    )
+
+    companion object {
+        private val JSON_MEDIA_TYPE = "application/json; charset=utf-8".toMediaType()
+    }
+}
@@ -13,7 +13,9 @@ import ai.guiji.duix.test.databinding.ActivityCallBinding
 import ai.guiji.duix.test.realtime.OmniRealtimeClient
 import ai.guiji.duix.test.realtime.QwenTtsRealtimeClient
 import ai.guiji.duix.test.service.DashScopeConfig
+import ai.guiji.duix.test.service.DebugSettingsStore
 import ai.guiji.duix.test.service.FamilyBridgeService
+import ai.guiji.duix.test.service.PsychologyKnowledgeAgentService
 import ai.guiji.duix.test.ui.adapter.MotionAdapter
 import ai.guiji.duix.test.ui.dialog.AudioRecordDialog
 import ai.guiji.duix.test.ui.widget.AudioSpectrumView
@@ -32,9 +34,7 @@ import android.widget.Toast
 import com.bumptech.glide.Glide
 import org.json.JSONArray
 import org.json.JSONObject
-import java.io.File
 import java.io.FileInputStream
-import java.io.FileOutputStream
 import java.util.UUID
 import java.util.concurrent.ExecutorService
 import java.util.concurrent.Executors
@@ -55,18 +55,22 @@ class CallActivity : BaseActivity() {
    private fun applyMessage(msg: String){
        if (debug){
            runOnUiThread {
-                binding.tvDebug.visibility = View.VISIBLE
-                if (mMessage.length > 10000){
-                    mMessage = ""
+                binding.scrollDebug.visibility = View.VISIBLE
+                if (mMessage.length > 60000){
+                    mMessage = mMessage.takeLast(40000)
                }
-                mMessage = "${StringUtils.dateToStringMS4()} $msg\n$mMessage"
+                mMessage += "${StringUtils.dateToStringMS4()} $msg\n"
                binding.tvDebug.text = mMessage
+                binding.scrollDebug.post {
+                    binding.scrollDebug.fullScroll(View.FOCUS_DOWN)
+                }
            }
        }

    }

    private lateinit var binding: ActivityCallBinding
+    private val debugSettingsStore by lazy { DebugSettingsStore(applicationContext) }
    private var duix: DUIX? = null
    private var mDUIXRender: DUIXRenderer? = null
    private var mModelInfo: ModelInfo?=null     // 加载的模型信息
@@ -80,13 +84,14 @@ class CallActivity : BaseActivity() {
    private val mainHandler = Handler(Looper.getMainLooper())
    private var assistantTextBuffer = StringBuilder()
    private var hasStartedPushForResponse = false
-    private val enableWebSearch = false
+    private val enableWebSearch = true
    private val enableToolCalling = true
    private var isCloneMode = false
    private var clonedVoice = ""
    private var cloneTtsModel = DashScopeConfig.VOICE_CLONE_TTS_MODEL
    private var ttsRealtimeClient: QwenTtsRealtimeClient? = null
    private val familyBridgeService by lazy { FamilyBridgeService(applicationContext) }
+    private val psychologyKnowledgeAgentService by lazy { PsychologyKnowledgeAgentService(applicationContext) }
    private val bridgeExecutor: ExecutorService = Executors.newSingleThreadExecutor()
    private var currentSessionId: String? = null
    private var pendingConversationPrompt: FamilyBridgeService.ConversationPromptInjection? = null
@@ -103,7 +108,7 @@ class CallActivity : BaseActivity() {

        modelUrl = intent.getStringExtra("modelUrl") ?: ""
        voice = intent.getStringExtra("voice") ?: "Ethan"
-        debug = intent.getBooleanExtra("debug", false)
+        debug = intent.getBooleanExtra("debug", debugSettingsStore.isChatDebugEnabled())
        isCloneMode = intent.getBooleanExtra("cloneMode", false)
        clonedVoice = intent.getStringExtra("clonedVoice") ?: ""
        cloneTtsModel = intent.getStringExtra("cloneTtsModel") ?: DashScopeConfig.VOICE_CLONE_TTS_MODEL
@@ -136,16 +141,6 @@ class CallActivity : BaseActivity() {
            requestPermission(arrayOf(Manifest.permission.RECORD_AUDIO), 1)
        }

-        binding.btnPlayPCM.setOnClickListener {
-            applyMessage("start play pcm")
-            playPCMStream()
-        }
-
-        binding.btnPlayWAV.setOnClickListener {
-            applyMessage("start play wav")
-            playWAVFile()
-        }
-
        binding.btnRandomMotion.setOnClickListener {
            applyMessage("start random motion")
            duix?.startRandomMotion(true)
@@ -238,16 +233,9 @@ class CallActivity : BaseActivity() {
        applyMessage("init ok")
        runOnUiThread {
            binding.btnRecord.isEnabled = true
-            binding.btnPlayPCM.isEnabled = true
-            binding.btnPlayWAV.isEnabled = true
            binding.switchMute.isEnabled = true
            binding.btnAIConversation.isEnabled = true

-            if (debug) {
-                binding.btnPlayPCM.visibility = View.VISIBLE
-                binding.btnPlayWAV.visibility = View.VISIBLE
-            }
-
            mModelInfo?.let { modelInfo ->
                if (modelInfo.motionRegions.isNotEmpty()) {
                    val names = ArrayList<String>()
@@ -305,6 +293,9 @@ class CallActivity : BaseActivity() {
        return buildString {
            append(OmniRealtimeClient.getDefaultInstructions())
            append("\n\n")
+            append("当用户需要关于记忆下降、情绪低落、焦虑失眠、孤独感、认知照护、老年照护路径或家属陪伴方式的权威依据时，优先调用 consult_psychology_knowledge_base 工具，不要凭经验给出确定性的专业判断。\n")
+            append("调用 consult_psychology_knowledge_base 时，把老人当前最核心的问题写入 question；如果你想限定查询角度，例如陪伴建议、风险判断、是否需要就医，就写入 focus。\n")
+            append("如果只是日常闲聊或普通安慰，不必调用 consult_psychology_knowledge_base 工具。\n")
            append("当用户明确要给儿子、女儿、老伴、家里人或亲友带话、留言、报平安时，必须调用 leave_message_for_family 工具，不要假装已经转达。\n")
            append("当用户问家里人有没有留言、想听家人说了什么、想看看新消息时，必须调用 fetch_family_messages 工具，不要自己编造留言。\n")
            append("工具执行结果返回后，再用温和、简短、生活化的中文向老人说明情况。")
@@ -350,7 +341,27 @@ class CallActivity : BaseActivity() {
            })
        }

+        val psychologyKnowledgeParameters = JSONObject().apply {
+            put("type", "object")
+            put("properties", JSONObject().apply {
+                put("question", JSONObject().apply {
+                    put("type", "string")
+                    put("description", "需要从权威知识库中查询的核心问题，应尽量包含老人当前的真实困扰或具体疑问。")
+                })
+                put("focus", JSONObject().apply {
+                    put("type", "string")
+                    put("description", "希望子代理特别聚焦的角度，例如陪伴建议、风险判断、何时就医、家属沟通方式。")
+                })
+            })
+            put("required", JSONArray().put("question"))
+        }
+
        return listOf(
+            OmniRealtimeClient.ToolDefinition(
+                "consult_psychology_knowledge_base",
+                "当需要基于 WHO ICOPE 2024 等权威照护知识回答认知、情绪、睡眠、孤独、照护与家属支持问题时使用。当老人提出具体问题比如食欲不振时，必须使用此工具进行查询，而不是凭经验直接回答。",
+                psychologyKnowledgeParameters,
+            ),
            OmniRealtimeClient.ToolDefinition(
                "leave_message_for_family",
                "当老人明确要求给家里人带话、留言、报平安或请转告时使用。",
@@ -385,46 +396,6 @@ class CallActivity : BaseActivity() {
        bridgeExecutor.shutdownNow()
    }

-    private fun playPCMStream(){
-        val thread = Thread {
-            duix?.startPush()
-            val inputStream = assets.open("pcm/2.pcm")
-            val buffer = ByteArray(320)
-            var length: Int
-            while (inputStream.read(buffer).also { length = it } > 0){
-                val data = buffer.copyOfRange(0, length)
-                duix?.pushPcm(data)
-            }
-            duix?.stopPush()
-            inputStream.close()
-        }
-        thread.start()
-    }
-
-    private fun playWAVFile(){
-        val thread = Thread {
-            val wavName = "1.wav"
-            val wavFile = File(mContext.externalCacheDir, wavName)
-            if (!wavFile.exists()){
-                // copy assets -> sd card
-                val inputStream = assets.open("wav/$wavName")
-                if (!mContext.externalCacheDir!!.exists()){
-                    mContext.externalCacheDir!!.mkdirs()
-                }
-                val out = FileOutputStream(wavFile)
-                val buffer = ByteArray(1024)
-                var length: Int
-                while ((inputStream.read(buffer).also { length = it }) > 0) {
-                    out.write(buffer, 0, length)
-                }
-                out.close()
-                inputStream.close()
-            }
-            duix?.playAudio(wavFile.absolutePath)
-        }
-        thread.start()
-    }
-
    override fun permissionsGet(get: Boolean, code: Int) {
        super.permissionsGet(get, code)
        when (code) {
@@ -496,6 +467,9 @@ class CallActivity : BaseActivity() {
                    binding.btnCameraToggle.contentDescription = getString(R.string.camera_off)
                    isCameraEnabled = false
                    applyMessage("AI对话: 已连接, model=${OmniRealtimeClient.DEFAULT_MODEL}, search=${if (enableWebSearch) "on" else "off"}, tools=${if (enableToolCalling) "on" else "off"}, cloneMode=$isCloneMode")
+                    if (enableWebSearch && enableToolCalling) {
+                        // applyMessage("AI对话: 当前配置请求 search=on 且 tools=on，但 Qwen Omni Realtime 协议规定两者不兼容，实际下发时将由服务端连接配置决定仅生效其中一项")
+                    }
                    if (initialConversationPrompt != null) {
                        applyMessage("AI对话: 检测到家属留言，准备主动播报")
                        val requested = omniClient?.requestResponse() == true
@@ -554,6 +528,7 @@ class CallActivity : BaseActivity() {
            }

            override fun onUserTranscript(transcript: String) {
+                applyMessage("用户语音转写: $transcript")
                mainHandler.post {
                    binding.tvSubtitle.text = getString(R.string.ai_call_user_said, transcript)
                }
@@ -568,6 +543,7 @@ class CallActivity : BaseActivity() {
            }

            override fun onAssistantTranscriptDone(transcript: String) {
+                applyMessage("助手回复: $transcript")
                mainHandler.post {
                    assistantTextBuffer.clear()
                    assistantTextBuffer.append(transcript)
@@ -579,6 +555,7 @@ class CallActivity : BaseActivity() {
            override fun onToolCallRequested(toolName: String, callId: String, argumentsJson: String) {
                mainHandler.post {
                    binding.tvSubtitle.text = when (toolName) {
+                        "consult_psychology_knowledge_base" -> "我来帮您查一下更稳妥的照护资料。"
                        "leave_message_for_family" -> "我来帮您把这句话记给家里人。"
                        "fetch_family_messages" -> "我来帮您看看家里人的新留言。"
                        else -> getString(R.string.ai_call_thinking)
@@ -888,6 +865,7 @@ class CallActivity : BaseActivity() {
        enqueueBridgeTask("工具调用处理失败") {
            val arguments = runCatching { JSONObject(argumentsJson) }.getOrElse { JSONObject() }
            val result = executeRealtimeTool(toolName, callId, arguments)
+            applyMessage("AI工具输出: $toolName, callId=$callId, status=${result.status}, output=${result.outputText}")
            uploadConversationTurnInternal(
                FamilyBridgeService.ROLE_TOOL,
                result.outputText,
@@ -911,6 +889,7 @@ class CallActivity : BaseActivity() {
        arguments: JSONObject,
    ): ToolExecutionResult {
        return when (toolName) {
+            "consult_psychology_knowledge_base" -> executePsychologyKnowledgeTool(callId, arguments)
            "leave_message_for_family" -> executeLeaveMessageTool(callId, arguments)
            "fetch_family_messages" -> executeFetchFamilyMessagesTool(callId, arguments)
            else -> {
@@ -930,6 +909,56 @@ class CallActivity : BaseActivity() {
        }
    }

+    private fun executePsychologyKnowledgeTool(callId: String, arguments: JSONObject): ToolExecutionResult {
+        val question = arguments.optString("question").trim()
+        val focus = arguments.optString("focus").trim().ifBlank { null }
+
+        if (question.isEmpty()) {
+            val outputText = "还没有拿到要查询的具体问题，请先用一句话说明想查什么。"
+            runCatching {
+                familyBridgeService.recordToolCall(
+                    toolName = "consult_psychology_knowledge_base",
+                    arguments = arguments,
+                    outputText = outputText,
+                    status = FamilyBridgeService.STATUS_FAILURE,
+                    callId = callId,
+                    sessionId = currentSessionId,
+                )
+            }
+            return ToolExecutionResult(outputText, FamilyBridgeService.STATUS_FAILURE)
+        }
+
+        return runCatching {
+            val result = psychologyKnowledgeAgentService.consultKnowledge(question, focus)
+            familyBridgeService.recordToolCall(
+                toolName = "consult_psychology_knowledge_base",
+                arguments = arguments,
+                outputText = result.outputText,
+                status = FamilyBridgeService.STATUS_SUCCESS,
+                callId = callId,
+                sessionId = currentSessionId,
+            )
+            applyMessage(
+                "心理学子代理完成: model=${result.model}, requestId=${result.requestId ?: "-"}, inputTokens=${result.inputTokens}, outputTokens=${result.outputTokens}"
+            )
+            ToolExecutionResult(result.outputText, FamilyBridgeService.STATUS_SUCCESS)
+        }.getOrElse { error ->
+            val outputText = "权威照护资料查询暂时失败。请先基于安全原则回应：保持温和陪伴，避免给出确定性的医学结论；如果涉及持续恶化、明显痛苦、自伤他伤或急性危险，请建议尽快联系家属、医生或急救。"
+            runCatching {
+                familyBridgeService.recordToolCall(
+                    toolName = "consult_psychology_knowledge_base",
+                    arguments = arguments,
+                    outputText = outputText,
+                    status = FamilyBridgeService.STATUS_FAILURE,
+                    callId = callId,
+                    sessionId = currentSessionId,
+                )
+            }
+            applyMessage("心理学子代理失败: ${error.message}")
+            ToolExecutionResult(outputText, FamilyBridgeService.STATUS_FAILURE)
+        }
+    }
+
    private fun executeLeaveMessageTool(callId: String, arguments: JSONObject): ToolExecutionResult {
        val message = arguments.optString("message").trim()
        val relation = arguments.optString("recipientRelation").trim().ifBlank { "家里人" }
@@ -5,6 +5,7 @@ import ai.guiji.duix.test.R
 import ai.guiji.duix.test.databinding.ActivityMainBinding
 import ai.guiji.duix.test.model.AvatarCatalog
 import ai.guiji.duix.test.model.AvatarModel
+import ai.guiji.duix.test.service.DebugSettingsStore
 import ai.guiji.duix.test.service.FamilyBridgeService
 import ai.guiji.duix.test.ui.adapter.AvatarAdapter
 import ai.guiji.duix.test.ui.dialog.LoadingDialog
@@ -19,6 +20,7 @@ import java.io.File
 class MainActivity : BaseActivity() {

    private lateinit var binding: ActivityMainBinding
+    private val debugSettingsStore by lazy { DebugSettingsStore(applicationContext) }
    private val familyBridgeService by lazy { FamilyBridgeService(applicationContext) }
    private var mLoadingDialog: LoadingDialog? = null
    private var mLastProgress = 0
@@ -89,7 +91,7 @@ class MainActivity : BaseActivity() {
        val intent = Intent(mContext, CallActivity::class.java)
        intent.putExtra("modelUrl", model.modelUrl)
        intent.putExtra("voice", model.voice)
-        intent.putExtra("debug", false)
+        intent.putExtra("debug", debugSettingsStore.isChatDebugEnabled())
        startActivity(intent)
    }

@@ -3,6 +3,7 @@ package ai.guiji.duix.test.ui.activity
 import ai.guiji.duix.test.BuildConfig
 import ai.guiji.duix.test.databinding.ActivitySettingsBinding
 import ai.guiji.duix.test.service.DashScopeConfig
+import ai.guiji.duix.test.service.DebugSettingsStore
 import ai.guiji.duix.test.service.FamilyBridgeService
 import android.content.Intent
 import android.net.Uri
@@ -12,6 +13,7 @@ import org.json.JSONObject
 class SettingsActivity : BaseActivity() {

    private lateinit var binding: ActivitySettingsBinding
+    private val debugSettingsStore by lazy { DebugSettingsStore(applicationContext) }
    private val familyBridgeService by lazy { FamilyBridgeService(applicationContext) }

    override fun onCreate(savedInstanceState: Bundle?) {
@@ -23,11 +25,18 @@ class SettingsActivity : BaseActivity() {
        binding.rowDeviceCode.setOnClickListener {
            startActivity(Intent(mContext, DeviceSettingsActivity::class.java))
        }
+        binding.rowDebugMode.setOnClickListener {
+            binding.switchDebugMode.performClick()
+        }

        binding.tvVersionValue.text = "${BuildConfig.VERSION_NAME} (${BuildConfig.VERSION_CODE})"
        binding.tvServerValue.text = Uri.parse(DashScopeConfig.MONITOR_BASE_URL).host
            ?: DashScopeConfig.MONITOR_BASE_URL
        binding.tvDeviceValue.text = familyBridgeService.getDeviceUuid()
+        binding.switchDebugMode.isChecked = debugSettingsStore.isChatDebugEnabled()
+        binding.switchDebugMode.setOnCheckedChangeListener { _, isChecked ->
+            debugSettingsStore.setChatDebugEnabled(isChecked)
+        }

        Thread {
            runCatching {
@@ -7,6 +7,7 @@ import ai.guiji.duix.test.model.AvatarCatalog
 import ai.guiji.duix.test.model.AvatarModel
 import ai.guiji.duix.test.model.VoiceCloneProfile
 import ai.guiji.duix.test.service.DashScopeConfig
+import ai.guiji.duix.test.service.DebugSettingsStore
 import ai.guiji.duix.test.service.VoiceCloneProfileStore
 import ai.guiji.duix.test.service.VoiceCloneService
 import ai.guiji.duix.test.ui.adapter.AvatarAdapter
@@ -27,6 +28,7 @@ class VoiceCloneSetupActivity : BaseActivity() {
    private lateinit var binding: ActivityVoiceCloneSetupBinding
    private lateinit var avatarAdapter: AvatarAdapter
    private val mainHandler = Handler(Looper.getMainLooper())
+    private val debugSettingsStore by lazy { DebugSettingsStore(applicationContext) }
    private val profileStore by lazy { VoiceCloneProfileStore(this) }
    private val voiceCloneService = VoiceCloneService()
    private val recorder = VoiceCloneRecorder()
@@ -222,7 +224,7 @@ class VoiceCloneSetupActivity : BaseActivity() {
        val intent = Intent(mContext, CallActivity::class.java)
        intent.putExtra("modelUrl", selectedAvatar.modelUrl)
        intent.putExtra("voice", selectedAvatar.voice)
-        intent.putExtra("debug", false)
+        intent.putExtra("debug", debugSettingsStore.isChatDebugEnabled())
        intent.putExtra("cloneMode", true)
        intent.putExtra("clonedVoice", profile.voice)
        intent.putExtra("cloneTtsModel", profile.targetModel)
@@ -15,16 +15,29 @@
        android:layout_width="match_parent"
        android:layout_height="match_parent" />

-    <TextView
-        android:id="@+id/tvDebug"
-        android:textSize="12sp"
+    <ScrollView
+        android:id="@+id/scrollDebug"
+        android:layout_width="0dp"
+        android:layout_height="220dp"
        android:visibility="gone"
-        android:textColor="@color/white"
        android:background="#66000000"
-        android:padding="6dp"
+        android:fadeScrollbars="false"
+        android:fillViewport="true"
+        android:scrollbars="vertical"
        app:layout_constraintTop_toTopOf="parent"
-        android:layout_width="match_parent"
-        android:layout_height="200dp"/>
+        app:layout_constraintStart_toStartOf="parent"
+        app:layout_constraintEnd_toEndOf="parent">
+
+        <TextView
+            android:id="@+id/tvDebug"
+            android:layout_width="match_parent"
+            android:layout_height="wrap_content"
+            android:fontFamily="monospace"
+            android:padding="8dp"
+            android:textColor="@color/white"
+            android:textIsSelectable="true"
+            android:textSize="12sp" />
+    </ScrollView>


    <androidx.appcompat.widget.SwitchCompat
@@ -160,30 +173,4 @@
        android:layout_width="match_parent"
        android:layout_height="wrap_content"/>

-    <Button
-        android:id="@+id/btnPlayPCM"
-        android:text="@string/play_pcm"
-        android:visibility="gone"
-        app:layout_constraintBottom_toTopOf="@+id/btnAIConversation"
-        android:background="@drawable/shape_common_btn"
-        android:layout_margin="6dp"
-        android:textSize="13sp"
-        android:textColor="@color/white"
-        android:enabled="false"
-        android:layout_width="match_parent"
-        android:layout_height="wrap_content"/>
-
-    <Button
-        android:id="@+id/btnPlayWAV"
-        android:text="@string/play_wav"
-        android:visibility="gone"
-        app:layout_constraintBottom_toTopOf="@+id/btnPlayPCM"
-        android:background="@drawable/shape_common_btn"
-        android:layout_margin="6dp"
-        android:textSize="13sp"
-        android:textColor="@color/white"
-        android:enabled="false"
-        android:layout_width="match_parent"
-        android:layout_height="wrap_content"/>
-
 </androidx.constraintlayout.widget.ConstraintLayout>
@@ -87,6 +87,53 @@
                android:contentDescription="@null" />
        </LinearLayout>

+        <LinearLayout
+            android:id="@+id/rowDebugMode"
+            android:layout_width="match_parent"
+            android:layout_height="wrap_content"
+            android:layout_marginTop="16dp"
+            android:background="@drawable/shape_warm_feature_card"
+            android:clickable="true"
+            android:focusable="true"
+            android:foreground="?attr/selectableItemBackground"
+            android:gravity="center_vertical"
+            android:orientation="horizontal"
+            android:padding="18dp">
+
+            <LinearLayout
+                android:layout_width="0dp"
+                android:layout_height="wrap_content"
+                android:layout_weight="1"
+                android:orientation="vertical">
+
+                <TextView
+                    android:layout_width="wrap_content"
+                    android:layout_height="wrap_content"
+                    android:text="@string/settings_debug_title"
+                    android:textColor="@color/warm_text_primary"
+                    android:textSize="18sp"
+                    android:textStyle="bold" />
+
+                <TextView
+                    android:layout_width="match_parent"
+                    android:layout_height="wrap_content"
+                    android:layout_marginTop="6dp"
+                    android:lineSpacingExtra="4dp"
+                    android:text="@string/settings_debug_desc"
+                    android:textColor="@color/warm_text_secondary"
+                    android:textSize="14sp" />
+            </LinearLayout>
+
+            <androidx.appcompat.widget.SwitchCompat
+                android:id="@+id/switchDebugMode"
+                android:layout_width="wrap_content"
+                android:layout_height="wrap_content"
+                android:layout_marginStart="12dp"
+                android:contentDescription="@string/settings_debug_title"
+                android:importantForAutofill="no"
+                android:text="" />
+        </LinearLayout>
+
        <LinearLayout
            android:layout_width="match_parent"
            android:layout_height="wrap_content"
@@ -80,6 +80,8 @@
    <string name="settings_subtitle">把设备信息、家人连接入口和当前版本都放在这里，找起来更方便。</string>
    <string name="settings_device_code_title">家人连接设备码</string>
    <string name="settings_device_code_desc">打开设备二维码页面后，家里人扫一扫就能把这台设备绑定到自己的账号。</string>
+    <string name="settings_debug_title">显示调试详情</string>
+    <string name="settings_debug_desc">开启后，聊天页面会显示完整的交互日志，包括工具调用的输入和输出。</string>
    <string name="settings_about_title">关于这台设备</string>
    <string name="settings_version_label">应用版本</string>
    <string name="settings_server_label">家属面板域名</string>
@@ -82,6 +82,8 @@
    <string name="settings_subtitle">Keep device information, family connection entry and current version together in one place.</string>
    <string name="settings_device_code_title">Family device code</string>
    <string name="settings_device_code_desc">Open the QR code page for your family to scan and bind this elder device.</string>
+    <string name="settings_debug_title">Show debug details</string>
+    <string name="settings_debug_desc">When enabled, the chat page will show full interaction logs, including tool input and output.</string>
    <string name="settings_about_title">About this device</string>
    <string name="settings_version_label">App version</string>
    <string name="settings_server_label">Family panel domain</string>