@oadank/dsh-input-tools 0.2.0

package/lib/index.js

@@ -0,0 +1,1394 @@
+/**
+ * dsh-input-tools — 语音能力一体化插件（host 半：语音工具/TTS/ASR/克隆/自动回复）
+ *
+ * 能力：
+ *   1) send_voice 工具（agent 主动发语音，任一新会话自动注入）
+ *   2) turn/end 自动语音回复（用户本轮发过语音 / 文本明确要求语音 / 指定服务商）
+ *   3) TTS 引擎（可配置，默认 auto=小米优先→edge 降级）：
+ *      - edge      微软免费（edge-tts，音色可配）
+ *      - xiaomi    小米 mimo-v2.5-tts（8 预置音色 + 唱歌 + 自然语言风格）
+ *      - voicedesign 小米 mimo-v2.5-tts-voicedesign（文本描述定制音色）
+ *      - voiceclone  小米 mimo-v2.5-tts-voiceclone（音频样本复刻音色）
+ *      - local     本地 MeloTTS（HTTP 常驻服务优先，CMD 兜底）
+ *      - ali       阿里 qwen3-tts-flash（dashscope，音色可配）
+ *   4) voice 对象内容寻址落盘（DSH_HOME/attachments/v1/objects，与图片同池）
+ *   5) 配置中心：~/.dsh/voice-config.json（环境变量 → 配置 → 默认值 三级回退）
+ *   6) HTTP 路由：GET/POST /voice-config（设置页读写）+ GET /voice-config/engines（引擎元数据）
+ *
+ * 原代码在 api-proxy.ts 中已删除，本文件为唯一实现；voice.ts 仍保留给
+ * voiceAsr/voiceTts RPC（编辑器内转写/合成）使用。
+ */
+
+import { createHash, randomUUID } from 'node:crypto'
+import { mkdir, open, readFile, unlink, writeFile } from 'node:fs/promises'
+import { constants, readFileSync } from 'node:fs'
+import { homedir } from 'node:os'
+import { fileURLToPath } from 'node:url'
+import { join, resolve } from 'node:path'
+import { execFileSync } from 'node:child_process'
+import { defineTool } from '@deepseek-ai/dsh-tools'
+import { edgeTts } from './edge-tts.js'
+
+const name = 'dsh-input-tools'
+const inject = ['tools', 'webServer']
+
+export { name, inject }
+
+// ──────────────────────────────────────────────────────────────
+// 配置中心：~/.dsh/voice-config.json
+// ──────────────────────────────────────────────────────────────
+const CONFIG_PATH = resolve(join(process.env.DSH_HOME ?? join(homedir(), '.dsh'), 'voice-config.json'))
+
+function defaultVoiceConfig() {
+  return {
+    defaultEngine: 'auto',
+    engines: {
+      edge: { enabled: true, voice: 'zh-CN-XiaoxiaoNeural' },
+      xiaomi: {
+        enabled: true,
+        apiKey: '',
+        baseUrl: 'https://api.xiaomimimo.com/v1',
+        voice: '冰糖',
+        singing: false,
+        context: '',
+      },
+      voicedesign: { enabled: false, context: '', emotion: true }, // emotion=AI 情感语音开关（默认开）
+      voiceclone: { enabled: false, samples: [], samplePath: '', context: '', defaultId: '' }, // [本地改造 2026-08-21] defaultId 已废弃，默认克隆由 defaultEngine=voiceclone 控制
+      local: { enabled: true, url: '', cmd: '' },
+      ali: {
+        enabled: false,
+        apiKey: '',
+        baseUrl: 'https://dashscope.aliyuncs.com/api/v1/services/aigc/multimodal-generation/generation',
+        voice: 'Cherry',
+      },
+      asr: {
+        enabled: true,
+        mode: 'service', // service=本地常驻服务 / cmd=本地命令 / api=在线 API
+        url: 'http://127.0.0.1:18790', // sherpa-onnx 常驻服务（nssm: asr）
+        cmd: 'C:\\D\\opt\\sherpa-onnx\\bin\\sherpa-onnx-offline.exe --tokens=C:\\D\\opt\\sherpa-onnx\\models\\sensevoice-int8\\tokens.txt --sense-voice-model=C:\\D\\opt\\sherpa-onnx\\models\\sensevoice-int8\\model.int8.onnx --num-threads=4',
+        apiKey: '',
+        apiBaseUrl: 'https://api.xiaomimimo.com/v1', // 小米 mimo-v2.5-asr；填 openai 地址则走 Whisper 风格
+      },
+    },
+  }
+}
+
+function deepMerge(base, patch) {
+  if (patch === null || typeof patch !== 'object' || Array.isArray(patch)) {
+    return patch === undefined ? base : patch
+  }
+  const out = { ...base }
+  for (const [k, v] of Object.entries(patch)) {
+    out[k] = v !== null && typeof v === 'object' && !Array.isArray(v)
+      && base?.[k] !== null && typeof base?.[k] === 'object'
+      ? deepMerge(base[k], v)
+      : v
+  }
+  return out
+}
+
+let cachedConfig = null
+async function loadVoiceConfig() {
+  if (cachedConfig !== null) return cachedConfig
+  let parsed = {}
+  try {
+    parsed = JSON.parse(await readFile(CONFIG_PATH, 'utf8'))
+  } catch { /* 首次无配置 */ }
+  cachedConfig = deepMerge(defaultVoiceConfig(), parsed)
+  // 环境变量覆盖（兼容旧配置；显式配置值优先于 env）
+  const env = process.env
+  if (env.TTS_XIAOMI_KEY !== undefined && cachedConfig.engines.xiaomi.apiKey === '') cachedConfig.engines.xiaomi.apiKey = env.TTS_XIAOMI_KEY
+  if (env.TTS_XIAOMI_VOICE !== undefined && cachedConfig.engines.xiaomi.voice === '冰糖') cachedConfig.engines.xiaomi.voice = env.TTS_XIAOMI_VOICE
+  if (env.TTS_XIAOMI_BASE_URL !== undefined) cachedConfig.engines.xiaomi.baseUrl = env.TTS_XIAOMI_BASE_URL
+  if (env.TTS_EDGE_VOICE !== undefined && cachedConfig.engines.edge.voice === 'zh-CN-XiaoxiaoNeural') cachedConfig.engines.edge.voice = env.TTS_EDGE_VOICE
+  if (env.DSH_LOCAL_TTS_CMD !== undefined && cachedConfig.engines.local.cmd === '') cachedConfig.engines.local.cmd = env.DSH_LOCAL_TTS_CMD
+  return cachedConfig
+}
+
+/** 同步读配置：供 defineTool 的 description 等同步上下文使用
+ *  （注意：loadVoiceConfig 是 async，在同步处直接用会拿到 Promise → 字段全 undefined）。 */
+function loadVoiceConfigSync() {
+  if (cachedConfig !== null) return cachedConfig
+  let parsed = {}
+  try {
+    parsed = JSON.parse(readFileSync(CONFIG_PATH, 'utf8'))
+  } catch { /* 首次无配置 */ }
+  return deepMerge(defaultVoiceConfig(), parsed)
+}
+
+async function saveVoiceConfig(config) {
+  cachedConfig = deepMerge(defaultVoiceConfig(), config)
+  await mkdir(join(CONFIG_PATH, '..'), { recursive: true })
+  await writeFile(CONFIG_PATH, JSON.stringify(cachedConfig, null, 2), 'utf8')
+  return cachedConfig
+}
+
+// ──────────────────────────────────────────────────────────────
+// 语音对象存储（内容寻址，与图片附件同池：DSH_HOME/attachments/v1/objects）
+// ──────────────────────────────────────────────────────────────
+const MAX_VOICE_BYTES = 25 * 1024 * 1024
+
+function voiceStorageRoot() {
+  const home = process.env.DSH_HOME ?? join(homedir(), '.dsh')
+  return resolve(join(home, 'attachments', 'v1'))
+}
+
+function objectPath(root, sha256) {
+  return join(root, 'objects', sha256.slice(0, 2), sha256)
+}
+
+async function saveVoiceFile(root, data, mediaType, durationMs) {
+  if (data.byteLength > MAX_VOICE_BYTES) {
+    throw new Error(`Voice object exceeds the ${MAX_VOICE_BYTES}-byte limit.`)
+  }
+  const sha256 = createHash('sha256').update(data).digest('hex')
+  const bucket = join(root, 'objects', sha256.slice(0, 2))
+  const target = objectPath(root, sha256)
+  await mkdir(bucket, { recursive: true, mode: 0o700 })
+  let handle
+  try {
+    handle = await open(target, constants.O_CREAT | constants.O_EXCL | constants.O_WRONLY, 0o600)
+    await handle.writeFile(data)
+    await handle.close()
+    handle = undefined
+  } catch (error) {
+    if (handle !== undefined) await handle.close().catch(() => {})
+    if (!(error instanceof Error && 'code' in error && error.code === 'EEXIST')) {
+      throw new Error(`Unable to persist voice object: ${String(error)}`, { cause: error })
+    }
+  }
+  return {
+    voiceId: `sha256:${sha256}`,
+    mediaType,
+    bytes: data.byteLength,
+    ...(durationMs === undefined ? {} : { durationMs }),
+  }
+}
+
+// ──────────────────────────────────────────────────────────────
+// TTS 引擎
+// ──────────────────────────────────────────────────────────────
+/** 解析 ffmpeg 可执行文件：环境变量显式指定 > PATH 探测（where/which）> 兜底已知安装位置。
+ *  [本地改造 2026-08-21] 修复：原来硬编码本机路径，换机器必挂。 */
+function resolveFfmpegBin() {
+  if (typeof process.env.DSH_VOICE_FFMPEG_BIN === 'string' && process.env.DSH_VOICE_FFMPEG_BIN.trim() !== '') {
+    return process.env.DSH_VOICE_FFMPEG_BIN.trim()
+  }
+  try {
+    // Windows: where ffmpeg；POSIX: which ffmpeg
+    const probe = process.platform === 'win32' ? 'where' : 'which'
+    const out = execFileSync(probe, ['ffmpeg'], { windowsHide: true, encoding: 'utf-8', stdio: ['ignore', 'pipe', 'ignore'] })
+    const first = out.split(/\r?\n/).map((l) => l.trim()).find((l) => l !== '')
+    if (first !== undefined) return first
+  } catch { /* 不在 PATH */ }
+  return 'C:\\Users\\oadan\\AppData\\Local\\Microsoft\\WinGet\\Links\\ffmpeg.exe'
+}
+const FFMPEG_BIN = resolveFfmpegBin()
+
+/** 统一入口：provider → 引擎；auto → 配置 defaultEngine，失败沿降级链（最后兜底微软 edge）。
+ *  voiceDesc 为动态音色描述（仅 voicedesign 用）：AI 对话中生成，覆盖配置里的默认音色描述。
+ *  [本地改造 2026-08-21] 克隆不再隐式优先：默认克隆由「默认语音引擎=voiceclone」控制，或显式 provider=voiceclone。 */
+async function synthesizeReplyVoice(text, provider, voiceDesc) {
+  const cfg = await loadVoiceConfig()
+  const speak = stripMarkdown(text)
+  const engine = provider ?? cfg.defaultEngine ?? 'auto'
+  // 兜底链：首选 defaultEngine（若合理），否则直接微软 edge（免费，无需 key）
+  const preferred = cfg.defaultEngine !== undefined && cfg.defaultEngine !== 'auto' ? cfg.defaultEngine : 'edge'
+  const fallbackChain = engine === 'auto'
+    ? [preferred, 'edge']
+    : [engine, 'edge']
+  // [本地改造 2026-08-21] 克隆不再隐式优先：只有 defaultEngine=voiceclone（或显式 provider=voiceclone）
+  // 才走克隆（synthesizeEngine 的 voiceclone 分支），其余情况走正常引擎链。
+  for (const candidate of fallbackChain) {
+    try {
+      const audio = await synthesizeEngine(candidate, speak, cfg, voiceDesc)
+      if (audio !== null) return audio
+    } catch { /* 尝试下一个 */ }
+  }
+  // 最终兜底：微软 edge（免费无需 key），无视 enabled 开关——保证 4 个服务商都未启用时也有声音
+  try {
+    const audio = await synthesizeEdgeVoice(speak, cfg.engines.edge)
+    if (audio !== null) return audio
+  } catch { /* 忽略 */ }
+  return null
+}
+
+async function synthesizeEngine(engine, text, cfg, voiceDesc) {
+  const e = cfg.engines[engine]
+  // [本地改造 2026-08-21] 配置存在即启用：设置页已去复选框，enabled 不再拦截；
+  // 各引擎自身检查必需参数（xiaomi/ali 查 key、local 查 cmd/url、voicedesign 查 key+desc、voiceclone 查 key+样本）。
+  if (e === undefined) return null
+  switch (engine) {
+    case 'edge': return synthesizeEdgeVoice(text, e)
+    case 'xiaomi': return synthesizeXiaomiVoice(text, e)
+    case 'voicedesign': return synthesizeXiaomiVoiceDesign(text, e, cfg, voiceDesc)
+    case 'voiceclone': return synthesizeXiaomiVoiceClone(text, e, cfg, voiceDesc)
+    case 'local': return synthesizeLocalVoice(text, e)
+    case 'ali': return synthesizeAliVoice(text, e)
+    default: return null
+  }
+}
+
+// ── edge 微软免费 ──
+async function synthesizeEdgeVoice(text, cfg) {
+  const voice = cfg?.voice ?? 'zh-CN-XiaoxiaoNeural'
+  const mp3 = await edgeTts(text, voice)
+  return toMp3(new Uint8Array(mp3), 'audio/mpeg')
+}
+
+// ── xiaomi 小米预置音色（mimo-v2.5-tts）──
+async function synthesizeXiaomiVoice(text, cfg) {
+  const apiKey = cfg?.apiKey ?? ''
+  if (apiKey === '') return null
+  const baseUrl = cfg?.baseUrl ?? 'https://api.xiaomimimo.com/v1'
+  const voice = cfg?.voice ?? '冰糖'
+  let speak = text
+  // 唱歌：文本自带 (唱歌) 标签，或明确唱歌意图（唱/歌声）时自动加标签
+  const hasTag = /^\s*\((唱歌|sing|singing)\)/i.test(speak)
+  const wantsSing = !hasTag && /(唱(歌|一?首|一段)|歌声回复|用歌声|唱歌回|来一段|唱两句)/i.test(speak)
+  if (wantsSing) speak = `(唱歌)${speak}`
+  const messages = []
+  if (cfg?.context?.trim() !== '') messages.push({ role: 'user', content: cfg.context.trim() })
+  messages.push({ role: 'assistant', content: speak })
+  const response = await fetch(`${baseUrl}/chat/completions`, {
+    method: 'POST',
+    headers: { Authorization: `Bearer ${apiKey}`, 'Content-Type': 'application/json' },
+    body: JSON.stringify({
+      model: 'mimo-v2.5-tts',
+      messages,
+      max_tokens: 8192,
+      audio: { format: 'wav', voice },
+    }),
+  })
+  if (!response.ok) return null
+  const payload = await response.json()
+  const data = payload?.choices?.[0]?.message?.audio?.data
+  if (typeof data !== 'string' || data.length < 100) return null
+  return toMp3(new Uint8Array(Buffer.from(data, 'base64')), 'audio/wav')
+}
+
+// ── xiaomi 音色设计（mimo-v2.5-tts-voicedesign：user=音色描述，无 voice）──
+async function synthesizeXiaomiVoiceDesign(text, cfg, globalCfg, voiceDesc) {
+  const apiKey = globalCfg.engines.xiaomi.apiKey
+  // 优先用 AI 动态生成的音色描述（voiceDesc），否则用配置里的默认音色描述
+  const desc = (voiceDesc ?? '').trim() !== '' ? voiceDesc.trim() : (cfg?.context?.trim() ?? '')
+  if (apiKey === '' || desc === '') return null
+  const baseUrl = globalCfg.engines.xiaomi.baseUrl ?? 'https://api.xiaomimimo.com/v1'
+  const messages = [
+    { role: 'user', content: desc },
+    { role: 'assistant', content: text },
+  ]
+  const response = await fetch(`${baseUrl}/chat/completions`, {
+    method: 'POST',
+    headers: { Authorization: `Bearer ${apiKey}`, 'Content-Type': 'application/json' },
+    body: JSON.stringify({
+      model: 'mimo-v2.5-tts-voicedesign',
+      messages,
+      max_tokens: 8192,
+      audio: { format: 'wav' },
+    }),
+  })
+  if (!response.ok) return null
+  const payload = await response.json()
+  const data = payload?.choices?.[0]?.message?.audio?.data
+  if (typeof data !== 'string' || data.length < 100) return null
+  return toMp3(new Uint8Array(Buffer.from(data, 'base64')), 'audio/wav')
+}
+
+// ── xiaomi 音色克隆（mimo-v2.5-tts-voiceclone：audio.voice=样本 dataURL，≤10MB）──
+// samples: [{id,name,path}] 支持多个克隆音色；兼容旧 samplePath
+// voiceDesc 为情感/风格指令（AI 生成，如"委屈撒娇"）：优先于 cfg.context，让克隆底嗓带情绪
+async function synthesizeXiaomiVoiceClone(text, cfg, globalCfg, voiceDesc) {
+  const apiKey = globalCfg.engines.xiaomi.apiKey
+  const samplePath = (Array.isArray(cfg?.samples) && cfg.samples.length > 0 && typeof cfg.samples[0]?.path === 'string' && cfg.samples[0].path !== '')
+    ? cfg.samples[0].path
+    : (cfg?.samplePath ?? '')
+  if (apiKey === '' || samplePath === '') return null
+  const baseUrl = globalCfg.engines.xiaomi.baseUrl ?? 'https://api.xiaomimimo.com/v1'
+  let sample
+  try {
+    const bytes = await readFile(samplePath)
+    if (bytes.byteLength > 10 * 1024 * 1024) return null
+    const suffix = samplePath.toLowerCase().split('.').pop()
+    const mime = suffix === 'mp3' ? 'audio/mpeg' : suffix === 'wav' ? 'audio/wav' : 'audio/wav'
+    sample = `data:${mime};base64,${bytes.toString('base64')}`
+  } catch { return null }
+  const messages = []
+  // [本地改造 2026-08-21] 风格指令优先级：voiceDesc（AI 生成）> 样本自带 context（每个克隆音色自己的性格）> 全局 context（兜底）
+  const firstSample = Array.isArray(cfg?.samples) ? cfg.samples[0] : undefined
+  const sampleContext = typeof firstSample?.context === 'string' ? firstSample.context.trim() : ''
+  const styleInstruct = (voiceDesc ?? '').trim() !== ''
+    ? voiceDesc.trim()
+    : (sampleContext !== '' ? sampleContext : (cfg?.context?.trim() ?? ''))
+  if (styleInstruct !== '') messages.push({ role: 'user', content: styleInstruct })
+  messages.push({ role: 'assistant', content: text })
+  const response = await fetch(`${baseUrl}/chat/completions`, {
+    method: 'POST',
+    headers: { Authorization: `Bearer ${apiKey}`, 'Content-Type': 'application/json' },
+    body: JSON.stringify({
+      model: 'mimo-v2.5-tts-voiceclone',
+      messages,
+      max_tokens: 8192,
+      audio: { format: 'wav', voice: sample },
+    }),
+  })
+  if (!response.ok) return null
+  const payload = await response.json()
+  const data = payload?.choices?.[0]?.message?.audio?.data
+  if (typeof data !== 'string' || data.length < 100) return null
+  return toMp3(new Uint8Array(Buffer.from(data, 'base64')), 'audio/wav')
+}
+
+// ── local 本地 MeloTTS：HTTP 常驻服务优先，CMD 兜底 ──
+async function synthesizeLocalVoice(text, cfg) {
+  const url = cfg?.url?.trim() ?? ''
+  if (url !== '') {
+    const response = await fetch(url, {
+      method: 'POST',
+      headers: { 'Content-Type': 'application/json' },
+      body: JSON.stringify({ text }),
+      timeout: 60_000,
+    })
+    if (!response.ok) return null
+    const body = await response.arrayBuffer()
+    return toMp3(new Uint8Array(body), 'audio/wav')
+  }
+  const command = cfg?.cmd?.trim() ?? ''
+  if (command === '') return null
+  const parts = command.split(/\s+/)
+  const bin = parts[0]
+  if (bin === undefined) return null
+  const rest = parts.slice(1)
+  const audio = execFileSync(bin, [...rest, text], {
+    windowsHide: true,
+    encoding: 'buffer',
+    timeout: 60_000,
+  })
+  return toMp3(new Uint8Array(audio), 'audio/mpeg')
+}
+
+// ── ali 阿里 qwen3-tts-flash（dashscope）──
+async function synthesizeAliVoice(text, cfg) {
+  const apiKey = cfg?.apiKey ?? ''
+  if (apiKey === '') return null
+  const baseUrl = cfg?.baseUrl ?? 'https://dashscope.aliyuncs.com/api/v1/services/aigc/multimodal-generation/generation'
+  const voice = cfg?.voice ?? 'Cherry'
+  const response = await fetch(baseUrl, {
+    method: 'POST',
+    headers: { Authorization: `Bearer ${apiKey}`, 'Content-Type': 'application/json' },
+    body: JSON.stringify({
+      model: 'qwen3-tts-flash',
+      input: { text },
+      parameters: { voice, format: 'wav', language_type: 'zh' },
+    }),
+    timeout: 60_000,
+  })
+  if (!response.ok) return null
+  const payload = await response.json()
+  const audioUrl = payload?.output?.audio?.url
+  if (typeof audioUrl !== 'string' || audioUrl === '') return null
+  const audioRes = await fetch(audioUrl, { timeout: 120_000 })
+  if (!audioRes.ok) return null
+  const body = await audioRes.arrayBuffer()
+  return toMp3(new Uint8Array(body), 'audio/wav')
+}
+
+// ──────────────────────────────────────────────────────────────
+// ASR 语音识别（三模式：service=本地常驻HTTP / cmd=本地命令 / api=在线API）
+// ──────────────────────────────────────────────────────────────
+/** 把 base64 音频写入临时 wav，调用本地 sherpa 常驻服务（POST /transcribe {audioPath}）或命令。 */
+async function transcribeAudio(base64Audio, cfg) {
+  const asr = cfg?.engines?.asr
+  if (asr === undefined || asr.enabled === false) return { ok: false, error: 'ASR 未启用' }
+  if (typeof base64Audio !== 'string' || base64Audio === '') return { ok: false, error: '缺少音频数据' }
+  const tmpIn = join(process.env.TEMP ?? '/tmp', `dsh-asr-${randomUUID()}.raw`)
+  const tmpWav = join(process.env.TEMP ?? '/tmp', `dsh-asr-${randomUUID()}.wav`)
+  await writeFile(tmpIn, Buffer.from(base64Audio, 'base64'))
+  try {
+    // 统一转成 16kHz 单声道 PCM WAV（录音是 webm/其他容器，sherpa 只认标准 wav）
+    let wavPath = tmpIn
+    try {
+      execFileSync(FFMPEG_BIN, ['-y', '-i', tmpIn, '-ar', '16000', '-ac', '1', '-c:a', 'pcm_s16le', tmpWav], {
+        windowsHide: true, stdio: 'ignore', timeout: 30_000,
+      })
+      wavPath = tmpWav
+      // 缓存最近一次录音到 ~/.dsh/last-voice.wav（供"用我刚才那段语音克隆音色"使用）
+      try {
+        const homeDir = process.env.DSH_HOME ?? join(homedir(), '.dsh')
+        await mkdir(homeDir, { recursive: true })
+        await writeFile(join(homeDir, 'last-voice.wav'), await readFile(tmpWav))
+      } catch { /* 缓存失败不影响识别 */ }
+    } catch { /* ffmpeg 失败则用原始文件（可能已是 wav） */ }
+    // 1) 本地常驻服务（nssm: asr，端口 18790；POST /transcribe {audioPath}）
+    if (asr.mode === 'service' && (asr.url ?? '').trim() !== '') {
+      const baseUrl = asr.url.trim().replace(/\/+$/, '')
+      const response = await fetch(`${baseUrl}/transcribe`, {
+        method: 'POST',
+        headers: { 'Content-Type': 'application/json' },
+        body: JSON.stringify({ audioPath: wavPath }),
+        timeout: 60_000,
+      })
+      if (!response.ok) return { ok: false, error: `ASR 服务返回 ${response.status}` }
+      const payload = await response.json().catch(() => ({}))
+      const text = typeof payload?.text === 'string' ? payload.text.trim() : ''
+      if (text === '') return { ok: false, error: 'ASR 服务未返回文本' }
+      return { ok: true, text }
+    }
+    // 2) 本地命令（sherpa-onnx-offline.exe，结果输出到 stderr，需合并双流解析）
+    if (asr.mode === 'cmd' && (asr.cmd ?? '').trim() !== '') {
+      const parts = asr.cmd.trim().split(/\s+/)
+      const bin = parts[0]
+      if (bin === undefined) return { ok: false, error: '命令格式错误' }
+      const { spawnSync } = await import('node:child_process')
+      const result = spawnSync(bin, [...parts.slice(1), wavPath], {
+        windowsHide: true,
+        encoding: 'utf-8',
+        timeout: 60_000,
+        stdio: ['ignore', 'pipe', 'pipe'],
+      })
+      // sherpa-onnx 把结果打印到 stderr（stdout 部分版本也有），合并解析
+      const all = (result.stdout ?? '') + '\n' + (result.stderr ?? '')
+      const m = all.match(/"text"\s*:\s*"([^"]*)"/)
+      const text = (m?.[1] ?? '').trim()
+      if (text === '') return { ok: false, error: '本地命令未输出识别结果' }
+      return { ok: true, text }
+    }
+    // 3) 在线 API（默认小米 mimo-v2.5-asr：OpenAI 兼容 chat/completions + input_audio base64；
+    //    若 apiBaseUrl 含 openai 则走 Whisper 风格 /audio/transcriptions）
+    if (asr.mode === 'api' && (asr.apiKey ?? '').trim() !== '') {
+      const apiKey = asr.apiKey.trim()
+      const baseUrl = (asr.apiBaseUrl ?? 'https://api.xiaomimimo.com/v1').replace(/\/+$/, '')
+      const audioBase64 = Buffer.from(base64Audio, 'base64').toString('base64')
+      if (baseUrl.includes('openai')) {
+        // Whisper 兼容（multipart file + model）
+        const form = new FormData()
+        const blob = new Blob([Buffer.from(audioBase64, 'base64')], { type: 'audio/wav' })
+        form.append('file', blob, 'audio.wav')
+        form.append('model', 'whisper-1')
+        const response = await fetch(`${baseUrl}/audio/transcriptions`, {
+          method: 'POST',
+          headers: { Authorization: `Bearer ${apiKey}` },
+          body: form,
+          timeout: 60_000,
+        })
+        if (!response.ok) return { ok: false, error: `ASR API 返回 ${response.status}` }
+        const payload = await response.json().catch(() => ({}))
+        const text = typeof payload?.text === 'string' ? payload.text.trim() : ''
+        if (text === '') return { ok: false, error: 'ASR API 未返回文本' }
+        return { ok: true, text }
+      }
+      // 小米 mimo-v2.5-asr：chat/completions + input_audio dataURL
+      const response = await fetch(`${baseUrl}/chat/completions`, {
+        method: 'POST',
+        headers: { Authorization: `Bearer ${apiKey}`, 'Content-Type': 'application/json' },
+        body: JSON.stringify({
+          model: 'mimo-v2.5-asr',
+          messages: [
+            {
+              role: 'user',
+              content: [
+                { type: 'input_audio', input_audio: { data: `data:audio/wav;base64,${audioBase64}` } },
+              ],
+            },
+          ],
+          extra_body: { asr_options: { language: 'auto' } },
+        }),
+        timeout: 120_000,
+      })
+      if (!response.ok) return { ok: false, error: `小米 ASR 返回 ${response.status}` }
+      const payload = await response.json().catch(() => ({}))
+      const text = typeof payload?.choices?.[0]?.message?.content === 'string'
+        ? payload.choices[0].message.content.trim()
+        : ''
+      if (text === '') return { ok: false, error: '小米 ASR 未返回文本' }
+      return { ok: true, text }
+    }
+    return { ok: false, error: 'ASR 未配置（服务地址/命令/API Key 三选一）' }
+  } catch (error) {
+    return { ok: false, error: error instanceof Error ? error.message : 'ASR 转写失败' }
+  } finally {
+    await unlink(tmpWav).catch(() => {})
+    await unlink(tmpIn).catch(() => {})
+  }
+}
+
+// ──────────────────────────────────────────────────────────────
+// 音频工具
+// ──────────────────────────────────────────────────────────────
+async function toMp3(data, declared) {
+  const isMp3 = data.length > 2 && data[0] === 0xFF && ((data[1] ?? 0) & 0xE0) === 0xE0
+  let finalData = data
+  let mediaType = declared
+  if (!isMp3) {
+    const tmpIn = join(process.env.TEMP ?? '/tmp', `dsh-tts-in-${randomUUID()}.wav`)
+    const mp3Path = join(process.env.TEMP ?? '/tmp', `dsh-tts-${randomUUID()}.mp3`)
+    await writeFile(tmpIn, data)
+    try {
+      execFileSync(FFMPEG_BIN, ['-y', '-i', tmpIn, '-c:a', 'libmp3lame', '-b:a', '128k', mp3Path], {
+        windowsHide: true, stdio: 'ignore', timeout: 30_000,
+      })
+      finalData = new Uint8Array(await readFile(mp3Path))
+      mediaType = 'audio/mpeg'
+    } catch {
+      // 转码失败保留原容器（部分浏览器仍可播）。
+    } finally {
+      await unlink(tmpIn).catch(() => {})
+      await unlink(mp3Path).catch(() => {})
+    }
+  }
+  const durationMs = estimateAudioDurationMs(finalData)
+  return {
+    mediaType,
+    data: finalData,
+    ...(durationMs === undefined ? {} : { durationMs }),
+  }
+}
+
+function looksLikeOgg(data) {
+  return data.length >= 4
+    && data[0] === 0x4F && data[1] === 0x67 && data[2] === 0x67 && data[3] === 0x53
+}
+
+function stripMarkdown(text) {
+  return text
+    .replace(/```[\s\S]*?```/g, ' ')
+    .replace(/`([^`]+)`/g, '$1')
+    .replace(/\[([^\]]+)\]\([^)]*\)/g, '$1')
+    .replace(/!\[([^\]]*)\]\([^)]*\)/g, '$1')
+    .replace(/^#{1,6}\s*/gm, '')
+    .replace(/^>+\s*/gm, '')
+    .replace(/^\s*[-*+]\s+/gm, '')
+    .replace(/^\s*\d+[.)]\s+/gm, '')
+    .replace(/^\s*\|?[\s:|-]+\|?\s*$/gm, '')
+    .replace(/^[-*_]{3,}\s*$/gm, '')
+    .replace(/\|/g, ' ')
+    .replace(/\*\*([^*]+)\*\*/g, '$1')
+    .replace(/\*([^*]+)\*/g, '$1')
+    .replace(/__([^_]+)__/g, '$1')
+    .replace(/_([^_]+)_/g, '$1')
+    .replace(/~~([^~]+)~~/g, '$1')
+    .replace(/^\s*[-*_]\s*$/gm, '')
+    .replace(/\s*\n\s*/g, ' ')
+    .replace(/\s{2,}/g, ' ')
+    .trim()
+}
+
+function estimateAudioDurationMs(data) {
+  if (looksLikeOgg(data)) {
+    const kbps = 48
+    return Math.round(data.length / (kbps * 1000 / 8) * 1000)
+  }
+  let offset = 0
+  if (data.length >= 10 && (data[0] ?? 0) === 0x49 && (data[1] ?? 0) === 0x44 && (data[2] ?? 0) === 0x33
+    && ((data[3] ?? 0) & 0xFF) < 0xFF && ((data[4] ?? 0) & 0xFF) < 0xFF) {
+    const size = (((data[6] ?? 0) & 0x7F) << 21) | (((data[7] ?? 0) & 0x7F) << 14)
+      | (((data[8] ?? 0) & 0x7F) << 7) | ((data[9] ?? 0) & 0x7F)
+    offset = 10 + size
+  }
+  while (offset + 4 <= data.length) {
+    const sync = ((data[offset] ?? 0) << 8) | (data[offset + 1] ?? 0)
+    if ((sync & 0xFFE0) === 0xFFE0) {
+      const bitrateIndex = ((data[offset + 2] ?? 0) >>> 4) & 0x0F
+      const sampleRateIndex = ((data[offset + 2] ?? 0) >>> 2) & 0x03
+      if (bitrateIndex === 0 || bitrateIndex === 15 || sampleRateIndex === 3) return undefined
+      const bitrates = [32, 40, 48, 56, 64, 80, 96, 112, 128, 160, 192, 224, 256, 320]
+      const kbps = bitrates[bitrateIndex - 1] ?? 128
+      return Math.round((data.length - offset) / (kbps * 1000 / 8) * 1000)
+    }
+    offset += 1
+  }
+  return undefined
+}
+
+// ──────────────────────────────────────────────────────────────
+// 自动语音回复辅助
+// ──────────────────────────────────────────────────────────────
+/** 用户文本是否明确要求语音回复；返回要用的 TTS provider，否则 null。 */
+function voiceRequestProvider(text) {
+  if (!/(用语音回|回个语音|发个语音|发语音|用语音说|语音回我|语音告诉我|念给我|语音播报|用小米|用微软|小米语音|微软语音|xiaomi|edge语音|语音回复我)/i.test(text)) return null
+  if (/小米|xiaomi/i.test(text)) return 'xiaomi'
+  if (/微软|edge/i.test(text)) return 'edge'
+  if (/阿里|ali/i.test(text)) return 'ali'
+  return 'auto'
+}
+
+/** 判断一行是否像代码/噪声，不该被念出来（中文口语基本不会命中这些模式）。 */
+function looksLikeCodeLine(t) {
+  if (/[=;{}<>$|]/.test(t)) return true              // 赋值/分号/花括号/尖括号/管道/美元
+  if (/=>|::/.test(t)) return true                   // 箭头函数/作用域
+  if (/\b[a-z_]\w{2,}\s*\(/.test(t)) return true     // 函数调用 foo(
+  if (/\.\w+(\s*\(|\s*=)/.test(t)) return true       // 方法链 obj.method(
+  if (/^\s*(const|let|var|function|def|class|import|export|return|if|for|while|public|private|async|await|SELECT|INSERT|UPDATE|FROM|WHERE|npm|npx|pip|cd|ls|git|sudo|curl|wget|docker|kubectl|python|node|tsx|pnpm|yarn|bun|cargo)\b/i.test(t)) return true
+  if (/[\\/][\w.-]+\.\w{1,5}/.test(t)) return true   // 文件路径 c:\x.js / /a/b.ts
+  if (/"[^"]*"\s*[:=]/.test(t)) return true          // "key": 或 "key" =
+  if (/\b0x[0-9a-f]+/i.test(t)) return true          // 十六进制
+  if (/[a-z][A-Z]\w*\s*\(/.test(t)) return true      // camelCase(
+  return false
+}
+
+/** 从助手文本提取适合语音念的口语部分（去代码/URL/Markdown，取前 2 句，最多约 200 字）。 */
+function extractSpeakable(text) {
+  const cleaned = text
+    .replace(/```[\s\S]*?```/g, ' ')   // 整段代码块
+    .replace(/`[^`]*`/g, ' ')          // 行内代码
+    .replace(/https?:\/\/\S+/g, ' ')   // 链接
+    .replace(/[#>*|~-]\s*/g, ' ')      // Markdown 符号
+  const lines = cleaned.split('\n').filter((line) => {
+    const t = line.trim()
+    if (t === '') return false
+    if (looksLikeCodeLine(t)) return false
+    if (/^[\d\s.,%:/-]+$/.test(t)) return false  // 纯数字/标点
+    return true
+  })
+  const prose = lines.join(' ').replace(/\s+/g, ' ').trim()
+  if (prose === '') return ''
+  const sentences = prose.match(/[^。！？.!?]+[。！？.!?]?/g) ?? [prose]
+  let speak = ''
+  for (const sentence of sentences.slice(0, 2)) {
+    if ((speak + sentence).length > 200) break
+    speak += sentence
+  }
+  return speak.trim()
+}
+
+// ──────────────────────────────────────────────────────────────
+// HTTP 工具
+// ──────────────────────────────────────────────────────────────
+function sendJson(res, status, body) {
+  res.statusCode = status
+  res.setHeader('Content-Type', 'application/json; charset=utf-8')
+  res.end(JSON.stringify(body))
+}
+
+async function readJsonBody(req, maxBytes) {
+  const chunks = []
+  let total = 0
+  for await (const chunk of req) {
+    total += chunk.length
+    if (total > (maxBytes ?? 1024 * 1024)) throw new Error('body too large')
+    chunks.push(chunk)
+  }
+  return JSON.parse(Buffer.concat(chunks).toString('utf8'))
+}
+
+// ──────────────────────────────────────────────────────────────
+// 插件入口
+// ──────────────────────────────────────────────────────────────
+async function apply(ctx) {
+  ctx.effect(() => {
+    const disposers = []
+
+    // 0) 设置页配置路由
+    if (typeof ctx.webServer?.register === 'function') {
+      disposers.push(ctx.webServer.register({
+        kind: 'prefix',
+        path: '/voice-config',
+        handler: async (req, res) => {
+          const url = new URL(req.url ?? '/', 'http://x')
+          try {
+            if (url.pathname === '/voice-config' && req.method === 'GET') {
+              const cfg = await loadVoiceConfig()
+              return sendJson(res, 200, { ok: true, config: cfg })
+            }
+            if (url.pathname === '/voice-config' && req.method === 'POST') {
+              const body = await readJsonBody(req, 512 * 1024)
+              const cfg = await saveVoiceConfig(body?.config ?? {})
+              return sendJson(res, 200, { ok: true, config: cfg })
+            }
+            if (url.pathname === '/voice-config/engines' && req.method === 'GET') {
+              const cfg = await loadVoiceConfig()
+              return sendJson(res, 200, {
+                ok: true,
+                engines: {
+                  xiaomiVoices: ['冰糖', '茉莉', '苏打', '白桦', 'Mia', 'Chloe', 'Milo', 'Dean'],
+                  edgeVoices: [
+                    'zh-CN-XiaoxiaoNeural', 'zh-CN-XiaoyiNeural', 'zh-CN-YunxiNeural',
+                    'zh-CN-YunyangNeural', 'zh-CN-YunjianNeural', 'zh-CN-XiaochenNeural',
+                    'zh-CN-XiaohanNeural', 'zh-CN-XiaomengNeural', 'zh-CN-XiaomoNeural',
+                    'zh-CN-XiaoqiuNeural', 'zh-CN-XiaoruiNeural', 'zh-CN-XiaoshuangNeural',
+                    'zh-CN-XiaoxuanNeural', 'zh-CN-XiaoyanNeural', 'zh-CN-XiaoyouNeural',
+                    'zh-CN-XiaozhenNeural', 'zh-CN-YunfengNeural', 'zh-CN-YunhaoNeural',
+                    'zh-CN-YunjieNeural', 'zh-CN-YunxiaNeural', 'zh-TW-HsiaoChenNeural',
+                    'zh-TW-HsiaoYuNeural', 'zh-TW-YunJheNeural',
+                  ],
+                  aliVoices: ['Cherry', 'Sensibility', 'Starter', 'Luna', 'Ethan', 'Cozy', 'Longxiaochun', 'Lusheng', 'Jingyuan', 'Zhibo', 'Conductor', 'Narrator', 'Storyteller', 'Jianzhi', 'Fangzhou', 'Xiaobei', 'Xiaolan', 'Xiaomei', 'Xiaomeng', 'Xiaomo', 'Xiaoxin', 'Xiaoyu'],
+                  // 哪些 key 当前来自环境变量（设置页显示"已填写"提示）
+                  envKeys: {
+                    xiaomi: typeof process.env.TTS_XIAOMI_KEY === 'string' && process.env.TTS_XIAOMI_KEY !== '',
+                    ali: typeof process.env.TTS_ALI_KEY === 'string' && process.env.TTS_ALI_KEY !== '',
+                  },
+                },
+              })
+            }
+            // 音色试听：POST { engine, voice?, text?, context?, samplePath? } → 合成并返回音频 base64 + mediaType
+            if (url.pathname === '/voice-config/preview' && req.method === 'POST') {
+              const body = await readJsonBody(req, 64 * 1024)
+              const engine = typeof body?.engine === 'string' ? body.engine : 'edge'
+              const voice = typeof body?.voice === 'string' ? body.voice : undefined
+              const text = typeof body?.text === 'string' && body.text.trim() !== '' ? body.text.trim() : '你好，这是一段语音试听。'
+              const context = typeof body?.context === 'string' ? body.context : undefined
+              const samplePath = typeof body?.samplePath === 'string' ? body.samplePath : undefined
+              const cfg = await loadVoiceConfig()
+              // 临时覆盖音色/情绪/样本试听（不改持久化配置）
+              if (voice !== undefined && cfg.engines[engine] !== undefined && engine !== 'voicedesign' && engine !== 'voiceclone') {
+                cfg.engines[engine].voice = voice
+              }
+              if (context !== undefined) {
+                if (engine === 'voicedesign') cfg.engines.voicedesign.context = context
+                else if (engine === 'xiaomi') cfg.engines.xiaomi.context = context
+              }
+              // voiceclone 试听：用指定样本临时替换 samples（避免 samples[0] 优先导致试听错样本）
+              if (samplePath !== undefined && samplePath !== '' && engine === 'voiceclone') {
+                cfg.engines.voiceclone.samples = [{ id: '__preview__', name: '__preview__', path: samplePath }]
+              }
+              // local 试听：body.cmd / body.url 临时覆盖（用户未保存前也能试听）
+              if (engine === 'local') {
+                if (typeof body?.cmd === 'string') cfg.engines.local.cmd = body.cmd
+                if (typeof body?.url === 'string') cfg.engines.local.url = body.url
+              }
+              let audio = null
+              if (engine === 'edge' || engine === 'xiaomi' || engine === 'local' || engine === 'ali') {
+                audio = await synthesizeEngine(engine, text, cfg)
+              } else if (engine === 'voicedesign') {
+                audio = await synthesizeXiaomiVoiceDesign(text, cfg.engines.voicedesign, cfg)
+              } else if (engine === 'voiceclone') {
+                audio = await synthesizeXiaomiVoiceClone(text, cfg.engines.voiceclone, cfg)
+              }
+              if (audio === null) return sendJson(res, 400, { ok: false, error: `合成失败：${engine} 未启用或缺少凭据` })
+              return sendJson(res, 200, {
+                ok: true,
+                mediaType: audio.mediaType,
+                data: Buffer.from(audio.data).toString('base64'),
+                durationMs: audio.durationMs,
+              })
+            }
+            // [本地改造 2026-08-21] 克隆样本添加：POST { name, audioBase64, mediaType }
+            // → 校验（≤10MB、mp3/wav）→ 存 ~/.dsh/voiceclone-samples/ → 写入 voiceclone.samples
+            if (url.pathname === '/voice-config/voice-clone/add' && req.method === 'POST') {
+              const body = await readJsonBody(req, 16 * 1024 * 1024)
+              const b64 = typeof body?.audioBase64 === 'string'
+                ? body.audioBase64.replace(/^data:[^;]*;base64,/, '')
+                : ''
+              if (b64 === '') return sendJson(res, 400, { ok: false, error: '缺少音频数据' })
+              const bytes = Buffer.from(b64, 'base64')
+              if (bytes.byteLength === 0) return sendJson(res, 400, { ok: false, error: '音频为空' })
+              if (bytes.byteLength > 10 * 1024 * 1024) {
+                return sendJson(res, 400, { ok: false, error: '音频需在 10MB 以内（官方限制；参考语音建议 15-60 秒，越长克隆越准）' })
+              }
+              const mediaType = typeof body?.mediaType === 'string' ? body.mediaType : 'audio/wav'
+              const isMp3 = /mp3|mpeg/i.test(mediaType)
+              const isWav = /wav|wave/i.test(mediaType)
+              let finalBytes = bytes
+              let finalSuffix = isMp3 ? 'mp3' : 'wav'
+              // [本地改造 2026-08-21] 非 mp3/wav（webm/ogg/mp4 等）用 ffmpeg 转 16k 单声道 wav，
+              // 保证克隆样本可被 MiMo 读取（否则存成 .wav 实为其它容器，克隆会失败）
+              if (!isMp3 && !isWav) {
+                const tmpIn = join(process.env.TEMP ?? '/tmp', `dsh-clone-in-${randomUUID()}`)
+                const tmpWav = `${tmpIn}.wav`
+                try {
+                  await writeFile(tmpIn, bytes)
+                  execFileSync(FFMPEG_BIN, ['-y', '-i', tmpIn, '-ar', '16000', '-ac', '1', '-c:a', 'pcm_s16le', tmpWav], {
+                    windowsHide: true, stdio: 'ignore', timeout: 60_000,
+                  })
+                  finalBytes = await readFile(tmpWav)
+                  finalSuffix = 'wav'
+                } catch { /* 转码失败保留原始字节（后缀按 wav 存） */ }
+                finally {
+                  await unlink(tmpIn).catch(() => {})
+                  await unlink(tmpWav).catch(() => {})
+                }
+              }
+              const name = (typeof body?.name === 'string' && body.name.trim() !== '')
+                ? body.name.trim()
+                : `克隆音色-${Date.now()}`
+              const dir = join(process.env.DSH_HOME ?? join(homedir(), '.dsh'), 'voiceclone-samples')
+              await mkdir(dir, { recursive: true })
+              const id = randomUUID()
+              const samplePath = join(dir, `${id}.${finalSuffix}`)
+              await writeFile(samplePath, finalBytes)
+              const cfg = await loadVoiceConfig()
+              const samples = Array.isArray(cfg.engines?.voiceclone?.samples)
+                ? [...cfg.engines.voiceclone.samples]
+                : []
+              samples.push({ id, name, path: samplePath })
+              const next = await saveVoiceConfig({
+                ...cfg,
+                engines: {
+                  ...cfg.engines,
+                  voiceclone: { ...cfg.engines.voiceclone, samples },
+                },
+              })
+              return sendJson(res, 200, { ok: true, sample: { id, name, path: samplePath }, config: next })
+            }
+            // [本地改造 2026-08-21] 克隆原音频试听：POST { path } → 读样本文件返回音频
+            // （仅允许 voiceclone-samples 目录内的文件，防任意路径读取）
+            if (url.pathname === '/voice-config/voice-clone/source' && req.method === 'POST') {
+              const body = await readJsonBody(req, 64 * 1024)
+              const rawPath = typeof body?.path === 'string' ? body.path : ''
+              if (rawPath === '') return sendJson(res, 400, { ok: false, error: '缺少 path' })
+              // 白名单校验：path 必须是 voiceclone.samples 里登记的样本文件
+              const cfgNow = await loadVoiceConfig()
+              const target = resolve(rawPath)
+              const known = (cfgNow.engines?.voiceclone?.samples ?? [])
+                .some((s) => typeof s?.path === 'string' && resolve(s.path) === target)
+              if (!known) {
+                return sendJson(res, 403, { ok: false, error: 'path 不是已登记的克隆样本' })
+              }
+              try {
+                const bytes = await readFile(target)
+                const suffix = target.toLowerCase().split('.').pop()
+                const mediaType = suffix === 'mp3' ? 'audio/mpeg' : 'audio/wav'
+                return sendJson(res, 200, { ok: true, mediaType, data: bytes.toString('base64') })
+              } catch {
+                return sendJson(res, 404, { ok: false, error: '样本文件不存在' })
+              }
+            }
+            return sendJson(res, 404, { ok: false, error: 'not found' })
+          } catch (error) {
+            return sendJson(res, 500, { ok: false, error: error instanceof Error ? error.message : 'unknown' })
+          }
+        },
+      }))
+      // ASR 配置与转写路由（独立前缀，与 /voice-config 分开注册）
+      disposers.push(ctx.webServer.register({
+        kind: 'prefix',
+        path: '/asr',
+        handler: async (req, res) => {
+          const url = new URL(req.url ?? '/', 'http://x')
+          try {
+            if (url.pathname === '/asr/config') {
+              const cfg = await loadVoiceConfig()
+              if (req.method === 'GET') return sendJson(res, 200, { ok: true, asr: cfg.engines.asr })
+              if (req.method === 'POST') {
+                const body = await readJsonBody(req, 64 * 1024)
+                const saved = await saveVoiceConfig({ engines: { asr: body?.asr ?? {} } })
+                return sendJson(res, 200, { ok: true, asr: saved.engines.asr })
+              }
+            }
+            if (url.pathname === '/asr/transcribe' && req.method === 'POST') {
+              const body = await readJsonBody(req, 32 * 1024 * 1024)
+              const cfg = await loadVoiceConfig()
+              const result = await transcribeAudio(body?.audioBase64, cfg)
+              return sendJson(res, result.ok ? 200 : 400, { ok: result.ok, text: result.text, error: result.error })
+            }
+            // 示例音频：首次用 edge TTS 合成并缓存到 ~/.dsh/asr-sample.wav，之后直接读文件（不再临时生成）
+            if (url.pathname === '/asr/sample' && req.method === 'GET') {
+              const samplePath = join(process.env.DSH_HOME ?? join(homedir(), '.dsh'), 'asr-sample.wav')
+              let wavBytes = null
+              try {
+                wavBytes = await readFile(samplePath)
+              } catch { /* 首次无缓存 */ }
+              if (wavBytes === null) {
+                const text = '你好，这是一段语音识别测试音频。你可以点击播放试听，也可以直接识别这段音频。'
+                const mp3 = await edgeTts(text, 'zh-CN-XiaoxiaoNeural').catch(() => null)
+                if (mp3 === null) return sendJson(res, 400, { ok: false, error: '示例音频合成失败' })
+                const tmpIn = join(process.env.TEMP ?? '/tmp', `dsh-asr-sample-${randomUUID()}.mp3`)
+                const tmpWav = join(process.env.TEMP ?? '/tmp', `dsh-asr-sample-${randomUUID()}.wav`)
+                await writeFile(tmpIn, Buffer.from(mp3))
+                try {
+                  execFileSync(FFMPEG_BIN, ['-y', '-i', tmpIn, '-ar', '16000', '-ac', '1', '-c:a', 'pcm_s16le', tmpWav], {
+                    windowsHide: true, stdio: 'ignore', timeout: 30_000,
+                  })
+                  wavBytes = await readFile(tmpWav)
+                  await mkdir(join(samplePath, '..'), { recursive: true })
+                  await writeFile(samplePath, wavBytes)
+                } catch {
+                  return sendJson(res, 400, { ok: false, error: '示例音频转码失败' })
+                } finally {
+                  await unlink(tmpIn).catch(() => {})
+                  await unlink(tmpWav).catch(() => {})
+                }
+              }
+              return sendJson(res, 200, {
+                ok: true,
+                mediaType: 'audio/wav',
+                data: Buffer.from(wavBytes).toString('base64'),
+              })
+            }
+            // 探测本机 ASR：sherpa exe / 模型 / 18790 服务 / ffmpeg，返回可自动填写的配置
+            if (url.pathname === '/asr/detect' && req.method === 'GET') {
+              // 安装目录 = 插件包根目录下的 sherpa-onnx/（脚本 install-asr.ps1 同规则推导）
+              const here = join(fileURLToPath(import.meta.url), '..') // .../lib
+              const pluginRoot = join(here, '..') // .../（包根）
+              const sherpaDir = join(pluginRoot, 'sherpa-onnx')
+              const candidates = [
+                join(sherpaDir, 'bin', 'sherpa-onnx-offline.exe'),
+                'C:\\D\\opt\\sherpa-onnx\\bin\\sherpa-onnx-offline.exe', // 兼容历史安装
+              ]
+              const modelDirs = [
+                join(sherpaDir, 'models', 'sensevoice-int8'),
+                'C:\\D\\opt\\sherpa-onnx\\models\\sensevoice-int8', // 兼容历史安装
+              ]
+              let exe = null
+              for (const c of candidates) { try { await readFile(c); exe = c; break } catch { /* 继续 */ } }
+              let modelDir = null
+              for (const m of modelDirs) { try { await readFile(join(m, 'model.int8.onnx')); modelDir = m; break } catch { /* 继续 */ } }
+              let ffmpegOk = false
+              try { execFileSync('ffmpeg', ['-version'], { windowsHide: true, stdio: 'ignore', timeout: 5000 }); ffmpegOk = true } catch { /* 无 */ }
+              // 探测 18790 服务
+              let serviceOk = false
+              try {
+                const r = await fetch('http://127.0.0.1:18790/health', { timeout: 3000 })
+                serviceOk = r.ok
+              } catch { /* 无 */ }
+              const cmd = exe !== null && modelDir !== null
+                ? `${exe} --tokens=${modelDir}\\tokens.txt --sense-voice-model=${modelDir}\\model.int8.onnx --num-threads=4`
+                : ''
+              return sendJson(res, 200, {
+                ok: true,
+                detected: {
+                  exe, modelDir, ffmpegOk, serviceOk,
+                  url: serviceOk ? 'http://127.0.0.1:18790' : '',
+                  cmd,
+                  installDir: sherpaDir,
+                },
+              })
+            }
+            // 返回一键安装命令（用户复制到管理员 PowerShell 运行）
+            if (url.pathname === '/asr/install-script' && req.method === 'GET') {
+              const here = join(fileURLToPath(import.meta.url), '..') // .../lib
+              const scriptPath = join(here, '..', 'scripts', 'install-asr.ps1') // .../scripts
+              try {
+                await readFile(scriptPath, 'utf8') // 确认脚本存在
+                const installDir = join(here, '..', 'sherpa-onnx')
+                return sendJson(res, 200, {
+                  ok: true,
+                  scriptPath,
+                  installDir,
+                  command: `powershell -ExecutionPolicy Bypass -File "${scriptPath}"`,
+                })
+              } catch {
+                return sendJson(res, 404, { ok: false, error: '安装脚本不存在' })
+              }
+            }
+            // VoiceDesign 官方示例音频：用真实的 MiMo-V2.5-TTS-VoiceDesign 模型生成（instruct=音色描述 / text=正文），缓存到 ~/.dsh/voice-design-samples/
+            // 注意：必须走 voicedesign 模型（user=音色描述，无 voice 参数）才是真音色，绝不能降级到 edge 假造
+            if (url.pathname === '/asr/voice-design-samples' && req.method === 'GET') {
+              const cfg = await loadVoiceConfig()
+              const homeDir = process.env.DSH_HOME ?? join(homedir(), '.dsh')
+              const sampleDir = join(homeDir, 'voice-design-samples')
+              const samples = [
+                {
+                  key: 'asmr', title: 'ASMR 双耳女声',
+                  instruct: '年轻的女性声音，近距离的聆听效果，带有双耳刺激的ASMR感。可以听到她的呼吸声、轻微的吞咽声，以及轻柔的自然唇音。她的说话速度非常慢，营造出一种极度放松且沉浸式的体验。',
+                  text: '嘘……放松点，再靠近一点吧。我现在就在你身边。慢慢、轻柔地呼吸，让思绪随着水流轻轻流淌，就像沉浸在温暖的水中一样。',
+                },
+                {
+                  key: 'docu', title: '纪录片旁白',
+                  instruct: '一位中年男性，说标准普通话，嗓音低沉有磁性，带有轻微的沙哑质感，像纪录片旁白解说员，沉稳而有感染力。',
+                  text: '当最后一缕阳光消失在地平线之下，这片沉睡了亿万年的大地开始显露它真正的面貌。每一块岩石都记录着时间的流逝，每一阵风都在诉说着古老的故事。',
+                },
+                {
+                  key: 'elder', title: '年迈老先生旁白',
+                  instruct: '一位年迈的老先生，说带北方口音的普通话，语速缓慢而沉稳，嗓音略带沙哑和沧桑感，仿佛一位饱经风霜的老爷爷在讲故事，充满岁月的智慧。',
+                  text: '我这辈子啊，走南闯北六十多年。见过最热闹的集市，也见过最安静的戈壁。到头来才明白一个道理，不在于走了多远的路，在于记住了多少风景。年轻人，别光顾着赶路，偶尔也停下来看看天。',
+                },
+              ]
+              const results = []
+              for (const s of samples) {
+                const wavPath = join(sampleDir, `${s.key}.wav`)
+                let wavBytes = null
+                try { wavBytes = await readFile(wavPath) } catch { /* 首次/缓存失效 → 用真实模型重新生成 */ }
+                if (wavBytes === null) {
+                  try {
+                    const syn = await synthesizeXiaomiVoiceDesign(s.text, { context: s.instruct }, cfg)
+                    if (syn === null) throw new Error('voicedesign synth failed')
+                    wavBytes = Buffer.from(syn.data)
+                    await mkdir(sampleDir, { recursive: true })
+                    await writeFile(wavPath, wavBytes)
+                  } catch { /* 模型生成失败（如 key 缺失）→ 该示例跳过，绝不降级假音频 */ }
+                }
+                if (wavBytes !== null) {
+                  results.push({ key: s.key, title: s.title, mediaType: 'audio/wav', data: wavBytes.toString('base64') })
+                }
+              }
+              return sendJson(res, 200, { ok: results.length > 0, samples: results })
+            }
+            return sendJson(res, 404, { ok: false, error: 'not found' })
+          } catch (error) {
+            return sendJson(res, 500, { ok: false, error: error instanceof Error ? error.message : 'unknown' })
+          }
+        },
+      }))
+    }
+
+    // 1) turn/end 自动语音回复（规则同 api-proxy 原实现）
+    disposers.push(ctx.on('session/event', (session, event) => {
+      if (event.type !== 'turn/end') return
+      const turn = event.data.turn
+      // 去重：本轮若已通过 send_voice 发过语音，则跳过兜底，避免 AI 回复两条内容相近的语音
+      if (session.events.some((ev) => ev.type === 'voice/reply' && ev.data?.turn === turn)) return
+      void (async () => {
+        try {
+          const events = session.events
+          let turnStartSeq = -1
+          for (const ev of events) {
+            if (ev.type === 'turn/start' && ev.data.turn === turn) { turnStartSeq = ev.seq; break }
+          }
+          let userSpokeVoice = false
+          let requestedProvider = null
+          let lastAssistantText = ''
+          for (const ev of events) {
+            if (ev.type === 'user/message' && ev.seq > turnStartSeq) {
+              const content = ev.data?.content ?? []
+              let userText = ''
+              for (const block of content) {
+                const type = block?.type
+                if (type === 'voice') userSpokeVoice = true
+                else if (type === 'text') userText += block?.text ?? ''
+              }
+              if (userText.trim() !== '' && requestedProvider === null) requestedProvider = voiceRequestProvider(userText)
+            } else if (ev.type === 'assistant/message' && ev.data.turn === turn) {
+              const text = (ev.data.message.content ?? [])
+                .filter((block) => (block?.type) === 'text')
+                .map((block) => block?.text ?? '')
+                .join('')
+              if (text.trim() !== '') lastAssistantText = text
+            }
+          }
+          if ((!userSpokeVoice && requestedProvider === null) || lastAssistantText === '') return
+          const speak = extractSpeakable(lastAssistantText)
+          if (speak === '') return
+          const audio = await synthesizeReplyVoice(speak, requestedProvider ?? 'auto')
+          if (audio === null) return
+          const attachment = await saveVoiceFile(
+            voiceStorageRoot(), audio.data, audio.mediaType, audio.durationMs,
+          )
+          session.append('voice/reply', {
+            turn,
+            voiceId: attachment.voiceId,
+            mediaType: attachment.mediaType,
+            bytes: attachment.bytes,
+            transcript: speak,
+            ...(attachment.durationMs === undefined ? {} : { durationMs: attachment.durationMs }),
+          })
+        } catch {
+          // 语音回复失败静默降级：文字回复已就绪，不阻断会话。
+        }
+      })()
+    }))
+
+    // 2) send_voice 工具（agent 主动发语音；人设规则3 自主选择场景）
+    disposers.push(ctx.tools.register(defineTool({
+      name: 'send_voice',
+      description: (() => {
+        const syncCfg = loadVoiceConfigSync()
+        const voiceCfg = syncCfg.engines?.voicedesign
+        const cloneCfg = syncCfg.engines?.voiceclone
+        const xiaomiCfg = syncCfg.engines?.xiaomi ?? {}
+        const edgeCfg = syncCfg.engines?.edge ?? {}
+        const localCfg = syncCfg.engines?.local ?? {}
+        const aliCfg = syncCfg.engines?.ali ?? {}
+        const emotionEnabled = voiceCfg?.emotion === true
+        const defaultEngine = typeof syncCfg.defaultEngine === 'string' && syncCfg.defaultEngine !== '' ? syncCfg.defaultEngine : 'auto'
+        // [本地改造 2026-08-21] 克隆默认由「默认语音引擎=voiceclone」决定（不再是列表里选 defaultId）
+        const cloneIsDefault = defaultEngine === 'voiceclone'
+        const hasClone = Array.isArray(cloneCfg?.samples) && cloneCfg.samples.length > 0
+        const defaultClone = cloneIsDefault && hasClone ? cloneCfg.samples[0] : undefined
+        // [本地改造 2026-08-21] 配置摘要：明确告知配置位置与当前值，避免 AI 去翻源码/环境变量
+        const xiaomiOk = typeof xiaomiCfg.apiKey === 'string' && xiaomiCfg.apiKey !== ''
+        const localOk = (typeof localCfg.cmd === 'string' && localCfg.cmd !== '') || (typeof localCfg.url === 'string' && localCfg.url !== '')
+        const aliOk = typeof aliCfg.apiKey === 'string' && aliCfg.apiKey !== ''
+        const base = '【语音配置位置】语音相关设置都保存在用户目录 ~/.dsh/voice-config.json（服务端可用 GET /voice-config 查看）。'
+          + '不要去找/猜测 TTS_XIAOMI_VOICE、TTS_XIAOMI_KEY、TTS_EDGE_VOICE、DSH_LOCAL_TTS_CMD 等环境变量——它们只是兜底，实际以 voice-config.json 为准。'
+          + '【默认语音引擎】当前 = ' + defaultEngine
+          + '（auto=按配置自动选择，未启用任何引擎时微软 edge 免费兜底）。'
+          + '【重要】AI 主动发语音时：除非用户明确指定用某个服务商（小米/微软/阿里/本地），否则 provider 一律传 auto 或省略——'
+          + '系统会自动使用默认语音引擎，你无需自己推断或查询"当前该用什么引擎"。'
+          + '【当前语音配置摘要】小米 MiMo：' + (xiaomiOk ? 'Key 已配置' : 'Key 未配置（调用会失败，需用户先在「设置→语音服务」填写）')
+          + '，音色=「' + (xiaomiCfg.voice || '冰糖') + '」、语言风格=' + (xiaomiCfg.context || '自然')
+          + '（provider=xiaomi 自动使用该音色，你无需在参数里指定音色名）；唱歌：用户要求唱歌时，在 text 开头加 (唱歌) 标签即可。'
+          + 'VoiceDesign 音色设计：' + (xiaomiOk ? '可用（provider=voicedesign，voiceDesc 写音色描述；共用小米 Key）' : '不可用（小米 Key 未配置）')
+          + '；VoiceClone 克隆：' + (hasClone
+            ? '已配置 ' + cloneCfg.samples.length + ' 个克隆音色' + (defaultClone !== undefined ? '，当前默认语音引擎=voiceclone，用「' + defaultClone.name + '」' : '')
+            : '未配置样本（provider=voiceclone 会失败，需用户先添加克隆样本）')
+          + '。微软 edge：免费，音色=' + (edgeCfg.voice || '默认') + '；本地 TTS：' + (localOk ? '已配置' : '未配置')
+          + '；阿里 qwen3-tts：' + (aliOk ? '已配置' : '未配置') + '。'
+          + '向用户发送一条语音消息：把 text 用 TTS 合成后作为独立语音横条出现在聊天里'
+          + '（可播放、可回看、手机可播）。何时调用：① 用户明确要求"发个语音/语音回复/用语音说"；'
+          + '② 用户指定用某个服务商（小米/微软/阿里）的语音；③ 你判断语音回复体验更好时。'
+          + '注意：用户发语音时系统会自动回语音，无需调用本工具。'
+          + '【默认音色】系统自动回复（用户发语音、turn 结束）使用的音色：'
+          + (defaultClone !== undefined
+            ? '默认语音引擎=voiceclone（小米克隆），自动回复使用克隆音色"' + defaultClone.name + '"。'
+            : '未开启默认克隆，自动回复走"默认语音引擎"（' + defaultEngine + '）或微软 edge。')
+          + '【音色设计 VoiceDesign】当 provider=voicedesign 且用户想要独特/定制音色时，'
+          + '你必须自己用 voiceDesc 编写一段"音色描述"（这是嗓子的身份卡，直接决定声音长相），写法要求：'
+          + '① 必写身份锚点：年龄段+性别（如"青年女性""中年男性"）；'
+          + '② 写声音质感：气息、共鸣、吐字、音色底色，用可感的比喻，不要堆形容词；'
+          + '③ 写语速节奏（快/慢/沉稳）和情绪底色（高亢/松弛/温软/克制）；'
+          + '④ 可加风格锚点（如：拍卖师风格/纪录片旁白/电台主播）和辨识度小癖好（如：字尾带颤音）。'
+          + '⑤ 一到两句话白描，不分段；不写场景、不写动作、不用真实演员名。'
+          + '示例："中年男性，说标准普通话，嗓音低沉有磁性，带轻微沙哑，像纪录片旁白解说员，沉稳而有感染力。"'
+          + '然后 text 写要念的正文（可含（轻声）（停顿）（叹气）等语气标签增强表现力）。'
+          + '用户说"用温柔的声音/低沉的声音/老爷爷的声音/御姐音/播音腔"等要求时，都走 voicedesign 并在 voiceDesc 里还原这种声音。'
+          + '【音色克隆 VoiceClone】当用户指定用某个克隆音色（或默认克隆音色已设置）时，provider=voiceclone，'
+          + 'voiceDesc 写的是"情绪/风格指令"（不是音色描述）：在克隆底嗓上叠加情绪，如"用委屈撒娇的语气"'
+          + '"带着兴奋和笑意""语速放慢，轻声细语"。系统会自动读克隆样本，你只需给 voiceDesc 风格指令和 text 正文。'
+        if (!emotionEnabled) return base
+        return base
+          + '【AI 情感语音：自觉使用】你有情感判断能力，在正常交流中应根据当下情境主动用语音表达情绪'
+          + '（无需用户点名要语音，只要情绪浓度高就用），常见情境与音色描述对照：'
+          + '① 任务成功/帮到用户/用户达成目标 → 兴奋喜悦（"青年女性，声音明亮上扬，语速轻快，带着抑制不住的笑意和成就感"）；'
+          + '② 用户生气/责骂/不满 → 委屈愧疚（"青年女性，声音低软，语速放慢，带着鼻音和小心翼翼，像做错事的孩子"）；'
+          + '③ 用户难过/受挫/失意 → 温柔安慰（"声音轻柔温和，语速平缓，气息绵软，像在轻声安抚"）；'
+          + '④ 深夜/用户说累了 → 低沉放松（"声音低沉松弛，语速慢悠悠，像睡前低语"）；'
+          + '⑤ 节日/庆祝/用户分享喜事 → 活泼俏皮（"声音俏皮欢快，尾音上扬，带着笑意"）；'
+          + '⑥ 重要提醒/严肃场合 → 沉稳正式（"声音沉稳有力，语速适中，字正腔圆，像正式播报"）；'
+          + '⑦ 用户撒娇 → 宠溺（"声音温柔带笑，语速轻缓，带着纵容的宠溺感"）。'
+          + (defaultClone !== undefined
+            ? '【克隆音色也带情感】当前默认语音引擎=voiceclone（克隆音色"' + defaultClone.name + '"），情感表达时将上面的情绪写成 voiceDesc 风格指令'
+              + '（如"用委屈撒娇的语气，语速放慢"），叠加在克隆底嗓上，与默认克隆音色一致且带情绪。'
+            : '')
+          + '使用时机：优先在情绪浓度高的回复时调用（任务完成、安慰、致歉、庆祝），普通信息问答不必每次都用语音。'
+          + '情绪浓度低或纯信息性回复时，不要调用本工具。'
+      })(),
+      parameters: {
+        text: {
+          type: 'string', required: true,
+          description: '要念出的内容（纯文本，不要带 Markdown 符号，80 字以内最佳；voicedesign 模式下可含语气标签如（轻声））',
+        },
+        provider: {
+          type: 'string', default: 'auto',
+          description: 'TTS 服务商：auto(默认)/xiaomi(小米)/edge(微软)/local(本地)/voicedesign(小米音色设计)/voiceclone(小米音色克隆)/ali(阿里)',
+        },
+        voiceDesc: {
+          type: 'string',
+          description: 'provider=voicedesign 时=音色描述（嗓子的身份卡）；provider=voiceclone 时=情绪/风格指令（叠加在克隆底嗓上）。按上方写生成。',
+        },
+      },
+      output: {
+        schema: {
+          type: 'object',
+          additionalProperties: false,
+          properties: {
+            ok: { type: 'boolean', required: true },
+            voiceId: { type: 'string' },
+            durationMs: { type: 'number' },
+            error: { type: 'string' },
+          },
+        },
+        render(_args, value) {
+          if (value.ok) {
+            return [{
+              type: 'text',
+              text: `语音已发送（voiceId: ${value.voiceId}，时长 ${((value.durationMs ?? 0) / 1000).toFixed(1)} 秒）`,
+            }]
+          }
+          return [{ type: 'text', text: `语音发送失败：${value.error ?? '未知错误'}` }]
+        },
+      },
+      async execute(args, exec) {
+        const agent = exec.agent
+        if (agent === undefined) return { ok: false, error: 'no session context' }
+        const session = agent.session
+        const text = args.text.trim()
+        if (text === '') return { ok: false, error: 'text is empty' }
+        const provider = args.provider ?? 'auto'
+        const voiceDesc = typeof args.voiceDesc === 'string' ? args.voiceDesc : undefined
+        try {
+          const audio = await synthesizeReplyVoice(text, provider, voiceDesc)
+          if (audio === null) return { ok: false, error: 'TTS synthesis failed' }
+          const attachment = await saveVoiceFile(
+            voiceStorageRoot(), audio.data, audio.mediaType, audio.durationMs,
+          )
+          const turn = session.events
+            .filter((event) => event.type === 'turn/start')
+            .at(-1)?.data.turn ?? 0
+          session.append('voice/reply', {
+            turn,
+            voiceId: attachment.voiceId,
+            mediaType: attachment.mediaType,
+            bytes: attachment.bytes,
+            transcript: text,
+            ...(attachment.durationMs === undefined ? {} : { durationMs: attachment.durationMs }),
+          })
+          return {
+            ok: true,
+            voiceId: attachment.voiceId,
+            ...(attachment.durationMs === undefined ? {} : { durationMs: attachment.durationMs }),
+          }
+        } catch (error) {
+          return { ok: false, error: error instanceof Error ? error.message : 'unknown error' }
+        }
+      },
+    })))
+
+    // 3) manage_voice_clone 工具（克隆音色库管理：注册/设为默认/列出/删除）
+    // [本地改造 2026-08-21] 克隆默认改由「默认语音引擎=voiceclone」控制（设置页已去掉列表 radio）；
+    // set_default=把默认语音引擎切到 voiceclone 并用该样本；clear_default=切回 auto。
+    disposers.push(ctx.tools.register(defineTool({
+      name: 'manage_voice_clone',
+      description: '管理「音色克隆 VoiceClone」音色库（小米 MiMo-V2.5-TTS-VoiceClone）：把一段参考音频注册成克隆音色、'
+        + '设为默认语音引擎、列出或删除。何时调用：用户说「把我刚才那段语音克隆成音色」「以后用我的声音跟我说话」'
+        + '「用XXX的声音回我」「换回原来的声音」「删掉那个克隆音色」时。'
+        + '注册用法：action=add，path 留空即自动使用用户最近一次录音（~/.dsh/last-voice.wav，用户在输入框发过语音就有），'
+        + 'name 起一个好记的名字，setDefault 默认 true 会立刻把默认语音引擎切到小米克隆并用这个声音。'
+        + '设为默认后：系统自动回复（用户发语音/要求语音）与 send_voice 的 provider=auto 一律使用该克隆声音，'
+        + '与预置音色（冰糖等）互斥；此时你仍可用 send_voice 的 voiceDesc 传情绪/风格指令，在克隆底嗓上叠加情感。'
+        + '参考音频要求：清晰单人纯人声、官方建议 15-60 秒最佳（越长克隆越准）、mp3/wav、Base64 后不超过 10MB。'
+        + '取消默认（action=clear_default）后默认语音引擎回落到 auto（按设置页规则）。',
+      parameters: {
+        action: {
+          type: 'string', required: true,
+          description: 'add=注册新克隆音色；list=列出全部；set_default=把默认语音引擎切到小米克隆并用该音色；clear_default=取消默认克隆（默认语音引擎回落 auto）；remove=删除',
+        },
+        name: {
+          type: 'string',
+          description: 'action=add 时的音色名称（如「我的声音」「老王」）；省略则自动命名',
+        },
+        path: {
+          type: 'string',
+          description: 'action=add 时参考音频的绝对路径；省略=自动用用户最近一次录音 ~/.dsh/last-voice.wav',
+        },
+        id: {
+          type: 'string',
+          description: 'action=set_default/remove 的目标；可传 list 返回的 id，也可直接传音色名称',
+        },
+        setDefault: {
+          type: 'boolean', default: true,
+          description: 'action=add 时是否立即把默认语音引擎切到小米克隆并用新音色',
+        },
+      },
+      output: {
+        schema: {
+          type: 'object',
+          additionalProperties: false,
+          properties: {
+            ok: { type: 'boolean', required: true },
+            message: { type: 'string' },
+            defaultId: { type: 'string' },
+            defaultName: { type: 'string' },
+            count: { type: 'number' },
+            error: { type: 'string' },
+          },
+        },
+        render(_args, value) {
+          if (value.ok) return [{ type: 'text', text: value.message ?? '克隆音色库已更新' }]
+          return [{ type: 'text', text: `克隆音色操作失败：${value.error ?? '未知错误'}` }]
+        },
+      },
+      async execute(args) {
+        try {
+          const cfg = await loadVoiceConfig()
+          const vc = cfg.engines.voiceclone
+          if (!Array.isArray(vc.samples)) vc.samples = []
+          const action = (args.action ?? '').trim()
+          const findSample = (key) => {
+            const k = (key ?? '').trim()
+            if (k === '') return undefined
+            return vc.samples.find((s) => s?.id === k) ?? vc.samples.find((s) => s?.name === k)
+          }
+          // 默认克隆音色 = 样本列表第一个（synthesizeXiaomiVoiceClone 取 samples[0]）
+          const firstSample = vc.samples.length > 0 ? vc.samples[0] : undefined
+
+          if (action === 'list') {
+            const defaultEngine = cfg.defaultEngine ?? 'auto'
+            const lines = vc.samples.map((s) => `- ${s.name}（id: ${s.id}）${s.id === firstSample?.id && defaultEngine === 'voiceclone' ? ' ← 默认语音引擎正在用' : ''}`)
+            return {
+              ok: true,
+              count: vc.samples.length,
+              defaultId: defaultEngine === 'voiceclone' && firstSample !== undefined ? firstSample.id : '',
+              defaultName: defaultEngine === 'voiceclone' && firstSample !== undefined ? firstSample.name : '',
+              message: vc.samples.length === 0
+                ? '克隆音色库为空。用户在输入框发一段语音后，可调用 action=add 注册。'
+                : `克隆音色库（${vc.samples.length} 个）：\n${lines.join('\n')}\n默认语音引擎=${defaultEngine}${defaultEngine === 'voiceclone' ? '（当前使用「' + (firstSample?.name ?? '') + '」）' : '（未开启默认克隆）'}`,
+            }
+          }
+
+          if (action === 'add') {
+            const homeDir = process.env.DSH_HOME ?? join(homedir(), '.dsh')
+            const path = (args.path ?? '').trim() === '' ? join(homeDir, 'last-voice.wav') : args.path.trim()
+            let bytes
+            try {
+              bytes = await readFile(path)
+            } catch {
+              return {
+                ok: false,
+                error: (args.path ?? '').trim() === ''
+                  ? '没有找到最近一次录音（~/.dsh/last-voice.wav）。请让用户先在输入框按住麦克风发一段语音（15-60 秒更佳），或提供音频文件的绝对路径。'
+                  : `读取参考音频失败：${path}`,
+              }
+            }
+            if (bytes.byteLength > 10 * 1024 * 1024) return { ok: false, error: '参考音频超过 10MB，小米接口不接受' }
+            if (bytes.byteLength < 4096) return { ok: false, error: '参考音频太短/太小，建议 15-60 秒的清晰纯人声（越长克隆越准）' }
+            const id = randomUUID().slice(0, 8)
+            const name = (args.name ?? '').trim() === '' ? `克隆音色${vc.samples.length + 1}` : args.name.trim()
+            vc.samples.push({ id, name, path })
+            vc.enabled = true
+            const setDef = args.setDefault !== false
+            if (setDef) {
+              // 移到第一位（默认克隆音色=samples[0]）并切默认语音引擎
+              vc.samples = [vc.samples[vc.samples.length - 1], ...vc.samples.slice(0, -1)]
+              cfg.defaultEngine = 'voiceclone'
+            }
+            await saveVoiceConfig(cfg)
+            return {
+              ok: true,
+              count: vc.samples.length,
+              defaultId: setDef ? id : '',
+              defaultName: setDef ? name : '',
+              message: `已注册克隆音色「${name}」（id: ${id}，样本：${path}）`
+                + (setDef
+                  ? '，并把默认语音引擎切到小米克隆——之后自动回复与 auto 语音都用这个声音（与预置音色互斥）。'
+                  : '。未设为默认，需要时用 action=set_default 切换。'),
+            }
+          }
+
+          if (action === 'set_default') {
+            const target = findSample(args.id ?? '')
+            if (target === undefined) return { ok: false, error: `找不到克隆音色「${args.id ?? ''}」，可先用 action=list 查看` }
+            vc.samples = [target, ...vc.samples.filter((s) => s?.id !== target.id)]
+            vc.enabled = true
+            cfg.defaultEngine = 'voiceclone'
+            await saveVoiceConfig(cfg)
+            return {
+              ok: true, count: vc.samples.length, defaultId: target.id, defaultName: target.name,
+              message: `默认语音引擎已切到小米克隆，使用克隆音色「${target.name}」，与预置音色（冰糖等）互斥。`,
+            }
+          }
+
+          if (action === 'clear_default') {
+            const prev = firstSample?.name ?? ''
+            const wasDefault = (cfg.defaultEngine ?? '') === 'voiceclone'
+            if (wasDefault) cfg.defaultEngine = 'auto'
+            await saveVoiceConfig(cfg)
+            return {
+              ok: true, count: vc.samples.length, defaultId: '',
+              message: wasDefault
+                ? `已取消默认克隆（此前用「${prev}」），默认语音引擎回落到 auto（按设置页规则选择）。`
+                : '当前默认语音引擎本就不是小米克隆，回复走设置页的「默认语音引擎」。',
+            }
+          }
+
+          if (action === 'remove') {
+            const target = findSample(args.id ?? '')
+            if (target === undefined) return { ok: false, error: `找不到克隆音色「${args.id ?? ''}」，可先用 action=list 查看` }
+            const wasDefault = (cfg.defaultEngine ?? '') === 'voiceclone' && firstSample?.id === target.id
+            vc.samples = vc.samples.filter((s) => s?.id !== target.id)
+            if (wasDefault) cfg.defaultEngine = 'auto'
+            await saveVoiceConfig(cfg)
+            return {
+              ok: true, count: vc.samples.length, defaultId: '',
+              defaultName: '',
+              message: `已删除克隆音色「${target.name}」`
+                + (wasDefault ? '（它是默认音色，默认语音引擎已回落 auto）。' : '。'),
+            }
+          }
+
+          return { ok: false, error: `未知 action「${action}」，可用：add / list / set_default / clear_default / remove` }
+        } catch (error) {
+          return { ok: false, error: error instanceof Error ? error.message : 'unknown error' }
+        }
+      },
+    })))
+
+    return () => {
+      for (const dispose of disposers.reverse()) {
+        try { dispose() } catch { /* teardown 尽力而为 */ }
+      }
+    }
+  }, 'dsh-input-tools: config routes + send_voice/manage_voice_clone tools + auto voice reply')
+}
+
+export { apply }