@olares/lares-core 0.15.10 → 0.15.12

package/i18n/voice.js

@@ -1,4 +1,55 @@
 export const ZH = {
+  "settings.backend": "语音后端",
+  "settings.backend.pipeline": "STT → Lares → TTS",
+  "settings.realtime.model": "Realtime 模型",
+  "settings.realtime.hint": "Realtime 模型完成实际验收后才可启用。普通语音模式始终可用。",
+  "error.voice_host_upgrade_required": "需要更新 Lares 主机音频助手",
+  "error.voice_device_lost": "音频设备已断开，请重新连接后恢复语音",
+
+  "call.mute": "关闭本次朗读",
+  "call.unmute": "开启本次朗读",
+  "call.history": "语音交流记录",
+  "call.clearHistory": "清除交流记录",
+  "settings.dictationCleanup": "整理听写文字（保留原意）",
+  "settings.preset": "语音工作方式",
+  "settings.chatModel": "交流模型",
+  "settings.preset.dictation": "只做听写",
+  "settings.preset.chat": "纯语音聊天",
+  "settings.preset.assistant": "语音助手委托 Lares",
+  "settings.preset.direct": "每轮直接提交 Lares",
+  "error.voice_idle_timeout": "语音因闲置已结束，点击重新开始",
+  "error.voice_recovery_expired": "语音恢复窗口已结束，点击重新开始",
+  "mic.cancel": "取消录音",
+  "mic.finish": "完成录音",
+  "call.pauseInput": "暂停收听",
+  "call.resumeInput": "恢复收听",
+  "call.inputPaused": "麦克风已暂停",
+  "call.halfDuplex": "播放中，麦克风已暂停",
+  "error.voice_draft_changed": "输入框暂时无法插入，请保留转写文字后重试",
+  "error.voice_background": "语音已暂停，点击恢复",
+  "settings.mobileMode": "移动端默认模式",
+  "settings.webMode": "网页默认模式",
+  "call.start": "语音对话",
+  "call.end": "结束语音",
+  "call.resume": "恢复语音",
+  "call.stopPlayback": "停止朗读",
+  "call.cancelTask": "取消任务",
+  "call.listening": "正在收听",
+  "call.working": "正在执行任务",
+  "call.speaking": "正在播放",
+  "call.waiting": "等待下一轮",
+  "settings.conversation": "语音对话",
+  "settings.mode": "对话方式",
+  "settings.mode.single": "语音输入",
+  "settings.mode.continuous": "连续对话",
+  "settings.device": "音频设备",
+  "settings.device.auto": "默认设备",
+  "settings.device.device": "浏览器麦克风与扬声器",
+  "settings.device.headset": "主机／蓝牙设备",
+  "error.voice_device_busy": "音频设备正在被其他语音连接使用",
+  "error.voice_disconnected": "连接已断开，点击恢复后继续收听",
+  "error.voice_realtime_unavailable": "Realtime 尚未验收，请选择普通语音模式",
+
   "error.voice_permission_denied": "麦克风权限被拒绝",
   "error.voice_audio_too_large": "录音过长",
   "error.voice_audio_unreadable": "音频无法识别，请重录",
@@ -9,7 +60,7 @@ export const ZH = {
   "error.voice_failed": "转写失败，请重试",
   "error.voice_unsupported": "当前浏览器不支持录音",
   "error.voice_bluetooth_not_ready": "蓝牙麦克风还没连上，请重试",
-  "error.voice_bluetooth_unavailable": "这台机器的蓝牙还没就绪",
+  "error.voice_bluetooth_unavailable": "这台机器的蓝牙还没就绪。请确认机器有蓝牙并已打开，然后在 Olares 市场重启 Lares，重启时会重新安装蓝牙组件",
   "error.voice_bluetooth_scan_failed": "扫描附近设备失败",
   "error.voice_bluetooth_connect_failed": "连接失败，请确认设备仍在等待连接",
   "error.voice_bluetooth_invalid": "蓝牙地址无效",
@@ -18,28 +69,90 @@ export const ZH = {
   "error.voice_speak_failed": "语音播报失败，请重试",
   "error.voice_speak_timeout": "语音播报超时，请重试",
   "error.voice_speak_unready": "当前不能保存语音播报开关",
-  "mic.preparing": "正在连接蓝牙麦克风…",
+  "error.voice_gpu_failed": "切换显卡模式失败（{detail}）。可以点重试；仍然失败，请到 Olares「设置 → GPU」查看显卡状态",
+  "error.voice_gpu_busy": "显卡正在切换，请等它完成后再操作",
+  "error.voice_gpu_no_gpu": "这台机器没有可用的 NVIDIA 显卡，不能边写边读，请使用「写完再读」",
+  "error.voice_gpu_unsupported": "{model} 不支持显存分片，不能边写边读，请使用「写完再读」",
+  "error.voice_gpu_plan_empty": "还没有应用分到显存，显卡模式没有改变。请先在下方「显存分片」里开启要用的应用并填好显存，保存后点重试",
+  "error.voice_gpu_plan_over": "这些应用至少需要 {need} GiB 显存，超过了显卡的 {total} GiB，显卡模式没有改变。请在下方「显存分片」里关掉一些应用或调小显存，保存后点重试",
+  "error.voice_gpu_plan_floor": "{app}的显存不能低于 {size} GiB",
+  "error.voice_gpu_unknown_app": "{app}不在这张显卡上，请刷新后重试",
+  "error.voice_gpu_no_slice": "{app}还没有分到显存。请先在「显存分片」里开启它并保存",
+  "error.voice_gpu_forbidden": "当前 Olares 账号没有管理显卡的权限，显卡模式没有改变。请用管理员账号安装 Lares，或请管理员到 Olares「设置 → GPU」里切换",
+  "error.voice_gpu_login": "Lares 访问 Olares 的登录凭据失效了。请在 Olares 市场重启 Lares 后再试",
+  "error.voice_gpu_blocked": "有应用停不下来，显卡模式没有改变（{detail}）。请先在 Olares 市场手动停止它，再点重试",
+  "error.voice_gpu_stop_failed": "{app}停不下来，显卡模式没有改变（{detail}）。请在 Olares 市场查看它的状态，再点重试",
+  "error.voice_gpu_memory": "{app}显存不够，没能启动。请在「显存分片」里调大它的显存，或关掉其他应用后重试",
+  "error.voice_gpu_resume_failed": "{app}没能启动（{detail}）。请在 Olares 市场查看它的状态，再点重试",
+  "mic.preparing": "正在准备麦克风…",
   "mic.stop": "点击结束录音",
   "mic.transcribing": "转写中…",
   "mic.idle": "语音输入",
   "mic.blocked": "当前无法输入",
   "lang.auto": "自动检测",
-  "status.auto": "自动",
   "settings.title": "语音设置",
-  "settings.intro": "点击麦克风开始说话，说完自动转成文字。语音播报开启后，只朗读这一轮的最终回答。",
   "settings.refresh": "刷新",
   "settings.refreshing": "刷新中…",
   "settings.router": "Router 控制台",
   "settings.loading": "正在读取语音设置…",
   "settings.loadFailed": "读取语音设置失败",
   "settings.refreshFailed": "刷新语音模型列表失败",
-  "settings.status.ready": "识别就绪 · {model}",
-  "settings.status.notReady": "还没有语音识别模型",
   "settings.stt.title": "语音识别",
   "settings.stt.hint": "把麦克风里的话转成文字",
   "settings.tts.title": "语音播报",
   "settings.tts.hint": "只朗读最终回答，思考和工具调用不播报",
   "settings.speak.label": "开启语音播报",
+  "settings.group.stt": "语音识别",
+  "settings.group.tts": "语音播报",
+  "settings.group.bluetooth": "输出设备",
+  "settings.stt.model": "识别模型",
+  "settings.tts.model": "播报模型",
+  "settings.speak.title": "自动朗读回答",
+  "settings.phase.running": "运行中",
+  "settings.phase.stopped": "已停止",
+  "settings.phase.busy": "处理中",
+  "settings.phase.failed": "异常",
+  "settings.phase.absent": "未安装",
+  "settings.mode.title": "播报方式",
+  "settings.mode.hint": "边写边读用显存分片，写完再读用时间分片。切换约 2–3 分钟，期间模型不可用",
+  "settings.mode.gpu": "显卡：{type}",
+  "settings.mode.turns": "写完再读",
+  "settings.mode.streaming": "边写边读",
+  "settings.mode.switching": "正在切换显卡模式…",
+  "settings.mode.resumingApp": "正在恢复 {app}…",
+  "settings.mode.restoring": "切换没成功，正在恢复原来的应用…",
+  "settings.mode.restored": "原来运行的应用已恢复。",
+  "settings.mode.notRestored": "这些应用没能恢复，请在 Olares 市场手动启动：{apps}",
+  "settings.mode.partial": "显卡已切换，但有应用没能启动。",
+  "settings.mode.mismatch": "显卡当前是{type}，和播报方式不一致",
+  "settings.mode.retry": "重试",
+  "settings.list.separator": "、",
+  "settings.gpu.MemorySlice": "显存分片",
+  "settings.gpu.TimeSlice": "时间分片",
+  "settings.gpu.Exclusive": "独占",
+  "settings.apps.title": "显卡应用",
+  "settings.apps.hint": "时间分片下这些应用轮流使用整张显卡；同时运行太多，显存可能不够",
+  "settings.apps.required": "至少 {size} GiB 显存",
+  "settings.apps.start": "启动",
+  "settings.apps.stop": "停止",
+  "settings.apps.starting": "启动中…",
+  "settings.apps.stopping": "停止中…",
+  "settings.apps.empty": "没有需要显卡的应用",
+  "settings.slices.title": "显存分片",
+  "settings.slices.gpu": "显卡",
+  "settings.slices.used": "已分配 {used} / {total} GiB",
+  "settings.slices.hint": "选择哪些应用常驻显卡、各分多少显存。保存后只重启有改动的应用",
+  "settings.slices.hintPending": "切到边写边读时，按这里的配置分配显存",
+  "settings.slices.size": "{app}的显存",
+  "settings.slices.enable": "{app}参与显存分片",
+  "settings.slices.empty": "还没有应用绑定到这张显卡",
+  "settings.slices.over": "超出显卡显存 {over} GiB",
+  "settings.slices.invalid": "开启的应用显存不能低于它的最低要求",
+  "settings.slices.restart": "有改动的应用会重启",
+  "settings.slices.reset": "还原",
+  "settings.slices.save": "保存",
+  "settings.slices.apply": "保存并应用",
+  "settings.slices.saving": "保存中…",
   "settings.install": "安装",
   "settings.installing": "安装中…",
   "settings.installStarted": "已开始安装，模型下载完成后点刷新",
@@ -68,6 +181,57 @@ export const ZH = {
 };
 
 export const EN = {
+  "settings.backend": "Voice backend",
+  "settings.backend.pipeline": "STT → Lares → TTS",
+  "settings.realtime.model": "Realtime model",
+  "settings.realtime.hint": "Realtime is available only after model validation. Pipeline mode remains available.",
+  "error.voice_host_upgrade_required": "Update the Lares host audio helper",
+  "error.voice_device_lost": "Audio device disconnected. Reconnect it and resume voice.",
+
+  "call.mute": "Mute replies",
+  "call.unmute": "Enable spoken replies",
+  "call.history": "Voice history",
+  "call.clearHistory": "Clear voice history",
+  "settings.dictationCleanup": "Clean up dictation without rewriting",
+  "settings.preset": "Voice behavior",
+  "settings.chatModel": "Conversation model",
+  "settings.preset.dictation": "Dictation only",
+  "settings.preset.chat": "Voice chat",
+  "settings.preset.assistant": "Voice assistant delegates to Lares",
+  "settings.preset.direct": "Send every turn to Lares",
+  "error.voice_idle_timeout": "Voice ended after inactivity. Start again to continue.",
+  "error.voice_recovery_expired": "Voice recovery expired. Start again to continue.",
+  "mic.cancel": "Cancel recording",
+  "mic.finish": "Finish recording",
+  "call.pauseInput": "Pause microphone",
+  "call.resumeInput": "Resume microphone",
+  "call.inputPaused": "Microphone paused",
+  "call.halfDuplex": "Microphone paused during playback",
+  "error.voice_draft_changed": "The composer cannot accept text yet. Keep the transcript and retry.",
+  "error.voice_background": "Voice paused. Resume explicitly.",
+  "settings.mobileMode": "Mobile default mode",
+  "settings.webMode": "Web default mode",
+  "call.start": "Voice conversation",
+  "call.end": "End voice",
+  "call.resume": "Resume voice",
+  "call.stopPlayback": "Stop speaking",
+  "call.cancelTask": "Cancel task",
+  "call.listening": "Listening",
+  "call.working": "Working",
+  "call.speaking": "Speaking",
+  "call.waiting": "Waiting",
+  "settings.conversation": "Voice conversation",
+  "settings.mode": "Conversation mode",
+  "settings.mode.single": "Dictation",
+  "settings.mode.continuous": "Continuous conversation",
+  "settings.device": "Audio device",
+  "settings.device.auto": "Default device",
+  "settings.device.device": "Browser microphone and speaker",
+  "settings.device.headset": "Host / Bluetooth device",
+  "error.voice_device_busy": "Audio device is in use by another connection",
+  "error.voice_disconnected": "Connection lost. Resume to listen again",
+  "error.voice_realtime_unavailable": "Realtime has not been validated. Select pipeline mode",
+
   "error.voice_permission_denied": "Microphone access denied",
   "error.voice_audio_too_large": "Recording too long",
   "error.voice_audio_unreadable": "Audio unreadable, please re-record",
@@ -79,7 +243,8 @@ export const EN = {
   "error.voice_failed": "Transcription failed, please retry",
   "error.voice_unsupported": "This browser does not support recording",
   "error.voice_bluetooth_not_ready": "Bluetooth microphone is not ready, please retry",
-  "error.voice_bluetooth_unavailable": "Bluetooth is not ready on this machine",
+  "error.voice_bluetooth_unavailable":
+    "Bluetooth is not ready on this machine. Make sure it has Bluetooth turned on, then restart Lares in the Olares Market; the restart reinstalls the Bluetooth pieces",
   "error.voice_bluetooth_scan_failed": "Failed to scan for nearby devices",
   "error.voice_bluetooth_connect_failed": "Connection failed. Confirm the device is still waiting to connect",
   "error.voice_bluetooth_invalid": "Invalid Bluetooth address",
@@ -88,29 +253,97 @@ export const EN = {
   "error.voice_speak_failed": "Couldn't speak the answer, please retry",
   "error.voice_speak_timeout": "Speech timed out, please retry",
   "error.voice_speak_unready": "Speech playback can't be saved right now",
-  "mic.preparing": "Connecting the Bluetooth microphone…",
+  "error.voice_gpu_failed":
+    "Couldn't switch the GPU mode ({detail}). Retry; if it keeps failing, check the GPU in Olares Settings → GPU",
+  "error.voice_gpu_busy": "The GPU is switching. Wait for it to finish",
+  "error.voice_gpu_no_gpu": "This machine has no usable NVIDIA GPU, so reading while writing is unavailable. Use \"When done\"",
+  "error.voice_gpu_unsupported": "{model} does not support memory slicing, so reading while writing is unavailable. Use \"When done\"",
+  "error.voice_gpu_plan_empty":
+    "No app has a memory slice yet, so the GPU mode was left unchanged. Turn on the apps you need under Memory slices below, set their memory, save, then retry",
+  "error.voice_gpu_plan_over":
+    "These apps need at least {need} GiB, more than the GPU's {total} GiB, so the GPU mode was left unchanged. Turn some off or lower their memory under Memory slices below, save, then retry",
+  "error.voice_gpu_plan_floor": "{app} needs at least {size} GiB",
+  "error.voice_gpu_unknown_app": "{app} is not on this GPU. Refresh and try again",
+  "error.voice_gpu_no_slice": "{app} has no memory slice yet. Turn it on under Memory slices and save",
+  "error.voice_gpu_forbidden":
+    "This Olares account cannot manage the GPU, so the GPU mode was left unchanged. Install Lares from an admin account, or ask an admin to switch it in Olares Settings → GPU",
+  "error.voice_gpu_login": "Lares lost its Olares sign-in. Restart Lares in the Olares Market and try again",
+  "error.voice_gpu_blocked":
+    "An app would not stop, so the GPU mode was left unchanged ({detail}). Stop it in the Olares Market, then retry",
+  "error.voice_gpu_stop_failed":
+    "{app} would not stop, so the GPU mode was left unchanged ({detail}). Check it in the Olares Market, then retry",
+  "error.voice_gpu_memory":
+    "{app} ran out of GPU memory and did not start. Give it more memory under Memory slices, or turn other apps off, then retry",
+  "error.voice_gpu_resume_failed": "{app} did not start ({detail}). Check it in the Olares Market, then retry",
+  "mic.preparing": "Preparing microphone…",
   "mic.stop": "Click to stop recording",
   "mic.transcribing": "Transcribing…",
   "mic.idle": "Voice input",
   "mic.blocked": "Input unavailable",
   "lang.auto": "Auto-detect",
-  "status.auto": "auto",
   "settings.title": "Voice settings",
-  "settings.intro":
-    "Click the microphone to start speaking; the take is transcribed into the composer. With playback on, only the turn's final answer is spoken.",
   "settings.refresh": "Refresh",
   "settings.refreshing": "Refreshing…",
   "settings.router": "Router console",
   "settings.loading": "Loading voice settings…",
   "settings.loadFailed": "Failed to load voice settings",
   "settings.refreshFailed": "Failed to refresh voice models",
-  "settings.status.ready": "Recognition ready · {model}",
-  "settings.status.notReady": "No speech-recognition model yet",
   "settings.stt.title": "Speech recognition",
   "settings.stt.hint": "Turn the microphone into text",
   "settings.tts.title": "Speech playback",
   "settings.tts.hint": "Read only the final answer, not reasoning or tool calls",
   "settings.speak.label": "Speak the final answer",
+  "settings.group.stt": "Speech recognition",
+  "settings.group.tts": "Speech playback",
+  "settings.group.bluetooth": "Output device",
+  "settings.stt.model": "Recognition model",
+  "settings.tts.model": "Speech model",
+  "settings.speak.title": "Read answers aloud",
+  "settings.phase.running": "Running",
+  "settings.phase.stopped": "Stopped",
+  "settings.phase.busy": "Working",
+  "settings.phase.failed": "Failed",
+  "settings.phase.absent": "Not installed",
+  "settings.mode.title": "Reading mode",
+  "settings.mode.hint": "While writing uses memory slicing; when done uses time slicing. Switching takes 2–3 minutes; models are unavailable meanwhile",
+  "settings.mode.gpu": "GPU: {type}",
+  "settings.mode.turns": "When done",
+  "settings.mode.streaming": "While writing",
+  "settings.mode.switching": "Switching GPU mode…",
+  "settings.mode.resumingApp": "Resuming {app}…",
+  "settings.mode.restoring": "The switch failed. Restoring the apps that were running…",
+  "settings.mode.restored": "The apps that were running are back.",
+  "settings.mode.notRestored": "These apps did not come back; start them in the Olares Market: {apps}",
+  "settings.mode.partial": "The GPU switched, but some apps did not start.",
+  "settings.mode.mismatch": "The GPU is in {type}, which does not match the reading mode",
+  "settings.mode.retry": "Retry",
+  "settings.list.separator": ", ",
+  "settings.gpu.MemorySlice": "memory slicing",
+  "settings.gpu.TimeSlice": "time slicing",
+  "settings.gpu.Exclusive": "exclusive",
+  "settings.apps.title": "GPU apps",
+  "settings.apps.hint": "With time slicing these apps take turns on the whole GPU; running too many at once may run out of memory",
+  "settings.apps.required": "needs {size} GiB",
+  "settings.apps.start": "Start",
+  "settings.apps.stop": "Stop",
+  "settings.apps.starting": "Starting…",
+  "settings.apps.stopping": "Stopping…",
+  "settings.apps.empty": "No apps need the GPU",
+  "settings.slices.title": "Memory slices",
+  "settings.slices.gpu": "GPU",
+  "settings.slices.used": "{used} / {total} GiB allocated",
+  "settings.slices.hint": "Choose which apps stay on the GPU and how much memory each gets. Saving restarts only the apps that changed",
+  "settings.slices.hintPending": "Applied when switching to reading while writing",
+  "settings.slices.size": "Memory for {app}",
+  "settings.slices.enable": "{app} uses a memory slice",
+  "settings.slices.empty": "No apps are bound to this GPU yet",
+  "settings.slices.over": "{over} GiB over the GPU memory",
+  "settings.slices.invalid": "An enabled app cannot go below its minimum memory",
+  "settings.slices.restart": "Apps that changed will restart",
+  "settings.slices.reset": "Reset",
+  "settings.slices.save": "Save",
+  "settings.slices.apply": "Save and apply",
+  "settings.slices.saving": "Saving…",
   "settings.install": "Install",
   "settings.installing": "Installing…",
   "settings.installStarted": "Install started. Refresh after the model finishes downloading",

package/larepass/host.js

@@ -3,6 +3,12 @@ export const LARES_API_PREFIX = "/api/lares";
 export const MODELS_PATH = `${LARES_API_PREFIX}/models`;
 export const PC_TEST_PROXY = "/laresHost";
 
+/**
+ * Packaged LarePass page origins (Capacitor Android `androidScheme: http`, iOS).
+ * Binary bodies leave those pages as a cross-origin webview `fetch`.
+ */
+export const LAREPASS_PAGE_ORIGINS = ["http://localhost", "https://localhost", "capacitor://localhost"];
+
 /** Host plugin routes (`file-preview`, models, voice, …). Same path on PC and LarePass. */
 export function isLaresPluginPath(path) {
   const suffix = String(path ?? "").startsWith("/") ? path : `/${path}`;

package/larepass/mux.js

@@ -77,11 +77,12 @@ export function classifyRemoteEvent(value) {
   return { kind: "pass", eventId: value.eventId };
 }
 
+// A tool call starting is what turns the step being read aloud silent.
+const LIVE_CHUNKS = new Set(["text-delta", "reasoning-delta", "tool-call-delta", "block-end"]);
+
 function liveChunkEvent(seq, turn, step, chunk) {
   if (!chunk || typeof chunk !== "object") return null;
-  if (chunk.type !== "text-delta" && chunk.type !== "reasoning-delta" && chunk.type !== "block-end") {
-    return null;
-  }
+  if (!LIVE_CHUNKS.has(chunk.type)) return null;
   return {
     type: "assistant/chunk",
     seq,

package/larepass/runtime.js

@@ -47,7 +47,7 @@ function viewOf(state, cache, pendingQuestions, pendingApprovals, inbox) {
     messages: snap.messages,
     running: snap.running,
     runningSince: snap.runningSince,
-    reply: snap.reply,
+    utterance: snap.utterance,
     error: state.error || snap.error,
     failed: state.failed,
     phase: state.phase,
@@ -862,6 +862,8 @@ export function createChatRuntime(client) {
       }
       return client.upload(id, file, options);
     },
+    voiceSocket(ticket) { return client.voiceSocket(ticket); },
+    voiceRequest(path, options) { return client.voiceRequest(path, options); },
     transcribe(blob, language, signal) {
       if (typeof client.transcribe !== "function") {
         return Promise.reject(new Error("voice_failed"));
@@ -872,9 +874,9 @@ export function createChatRuntime(client) {
       if (typeof client.reportLog !== "function") return Promise.resolve(false);
       return client.reportLog(scope, event, fields);
     },
-    speak(text) {
+    speak(request) {
       if (typeof client.speak !== "function") return Promise.resolve(null);
-      return client.speak(text);
+      return client.speak(request);
     },
     interruptSpeech() {
       if (typeof client.interruptSpeech !== "function") return Promise.resolve(false);

package/larepass/settings.js

@@ -101,6 +101,9 @@ export function createHostSettings(request) {
           status: payload(status, "/api/lares/voice/status"),
           sttModels: modelIds(listed, "stt"),
           ttsModels: modelIds(listed, "tts"),
+          chatModels: modelIds(listed, "chat"),
+          realtimeModels: modelIds(listed, "realtime"),
+          realtimeReady: modelIds(listed, "realtimeReady"),
         };
       }, { force: Boolean(force) });
     },

package/larepass/transcript.js

@@ -1,6 +1,6 @@
 import { producedPathsFromView, turnMediaPathsFromEvent } from "../files/deliverables.js";
 import { durablePathFromToolCall, parseToolArguments } from "../files/published-tools.js";
-import { concludingReply } from "../voice/speech.js";
+import { reduceUtterance } from "../voice/speech.js";
 import { contextForm, contextProvenance, isHumanUserSource } from "./context-provenance.js";
 import { toolRowModel } from "./tool-row.js";
 import { createTurnMetrics } from "./turn-metrics.js";
@@ -111,7 +111,7 @@ export function foldTranscript(entries) {
   let running = false;
   let runningSince = 0;
   let error = "";
-  let reply = null;
+  let utterance = null;
   let streamCursor = { loc: "", kind: "", index: 0 };
   let activeTurn = 0;
   const ended = new Map();
@@ -238,6 +238,7 @@ export function foldTranscript(entries) {
     const data = event.data ?? {};
     const view = entryView(entry, event);
     metrics.observe(event);
+    utterance = reduceUtterance(utterance, event);
 
     if (event.type === "turn/start") {
       closeTurn();
@@ -248,7 +249,6 @@ export function foldTranscript(entries) {
       running = true;
       runningSince = Number(event.time) || 0;
       error = "";
-      reply = null;
       continue;
     }
 
@@ -280,8 +280,6 @@ export function foldTranscript(entries) {
 
     if (event.type === "assistant/message") {
       applyAssistantMessage(data);
-      const text = concludingReply(event);
-      if (text) reply = { seq: event.seq, text };
       continue;
     }
 
@@ -376,6 +374,6 @@ export function foldTranscript(entries) {
     running,
     runningSince: running ? runningSince : 0,
     error,
-    reply,
+    utterance,
   };
 }

package/olares/trusted-host.js

@@ -2,9 +2,19 @@ export function loopbackAuthority(port) {
   return `127.0.0.1:${port}`;
 }
 
+/** The browser's Origin, kept for CORS after the loopback rewrite replaces it. */
+const pageOrigins = new WeakMap();
+
+export function pageOrigin(headers) {
+  return pageOrigins.get(headers) ?? headers?.origin;
+}
+
 export function applyLoopbackHeaders(headers, authority) {
   headers.host = authority;
-  if (headers.origin !== undefined) headers.origin = `http://${authority}`;
+  if (headers.origin !== undefined) {
+    if (!pageOrigins.has(headers)) pageOrigins.set(headers, headers.origin);
+    headers.origin = `http://${authority}`;
+  }
   // Fetch Metadata outlives the origin rewrite. The browser decides this one and
   // we cannot argue with it: a packaged page calling the entrance is `cross-site`
   // however the host and origin now read, and dsh refuses that outright — the

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@olares/lares-core",
-  "version": "0.15.10",
+  "version": "0.15.12",
   "description": "Lares 共享业务核心：Router、workspace、文件/媒体、Olares 身份（无 UI、无平台壳）",
   "type": "module",
   "repository": {
@@ -123,6 +123,8 @@
     "./voice/problem": "./voice/problem.js",
     "./voice/capture": "./voice/capture.js",
     "./voice/install": "./voice/install.js",
+    "./voice/gpu-mode": "./voice/gpu-mode.js",
+    "./voice/gpu-problem": "./voice/gpu-problem.js",
     "./search/config": "./search/config.js",
     "./search/menu": "./search/menu.js",
     "./search/client": "./search/client.js",
@@ -167,6 +169,29 @@
       "types": "./skills/state.d.ts",
       "default": "./skills/state.js"
     },
-    "./skills/client": "./skills/client.js"
+    "./skills/client": "./skills/client.js",
+    "./voice/controller": "./voice/controller.js",
+    "./voice/lease": "./voice/lease.js",
+    "./voice/browser-capture": "./voice/browser-capture.js",
+    "./voice/agent-session": "./voice/agent-session.js",
+    "./voice/device-input": "./voice/device-input.js",
+    "./voice/host-input": "./voice/host-input.js",
+    "./voice/conversation": "./voice/conversation.js",
+    "./voice/browser-playback": "./voice/browser-playback.js",
+    "./voice/realtime": "./voice/realtime.js",
+    "./voice/realtime-gateway": "./voice/realtime-gateway.js",
+    "./voice/realtime-audio": "./voice/realtime-audio.js",
+    "./voice/realtime-conversation": "./voice/realtime-conversation.js",
+    "./voice/host-service": "./voice/host-service.js",
+    "./voice/options": "./voice/options.js",
+    "./voice/records": "./voice/records.js",
+    "./voice/dialogue": "./voice/dialogue.js",
+    "./voice/external": "./voice/external.js",
+    "./voice/composite-gateway": "./voice/composite-gateway.js",
+    "./voice/router-audio": "./voice/router-audio.js",
+    "./voice/protocol-client": "./voice/protocol-client.js"
+  },
+  "dependencies": {
+    "ws": "^8.22.0"
   }
 }

package/router/tts.js

@@ -5,6 +5,28 @@ const REQUEST_TIMEOUT_MS = 180_000;
 
 const PCM_RATE = 24000;
 
+/**
+ * Engines pick a speaker per request when none is named, so the readout voice
+ * changed between answers. First listed voice the model offers wins.
+ */
+const PREFERRED_VOICES = ["serena", "vivian", "uncle_fu"];
+
+export function pickTtsVoice(voices) {
+  const ids = voices.filter((id) => typeof id === "string" && id);
+  return PREFERRED_VOICES.find((id) => ids.includes(id)) ?? [...ids].sort()[0] ?? null;
+}
+
+/** @returns {Promise<string[]>} voice ids the model serves; empty when it lists none. */
+export async function listTtsVoices(model) {
+  const res = await fetch(`${routerShimBaseUrl()}/audio/voices?model=${encodeURIComponent(model)}`, {
+    signal: AbortSignal.timeout(10_000),
+  });
+  if (!res.ok) return [];
+  const body = await res.json().catch(() => null);
+  const voices = Array.isArray(body?.voices) ? body.voices : [];
+  return voices.map((voice) => (typeof voice === "string" ? voice : voice?.id));
+}
+
 /** Kokoro concatenates chunks with no gap. This is the silence we insert between sentences. */
 export function silencePcm(ms, sampleRate = PCM_RATE) {
   const samples = Math.max(0, Math.round((sampleRate * ms) / 1000));
@@ -47,8 +69,9 @@ export async function synthesize(request) {
         model: request.model,
         input: text,
         response_format: "wav",
+        ...(request.voice ? { voice: request.voice } : {}),
       }),
-      signal: AbortSignal.timeout(request.timeoutMs ?? REQUEST_TIMEOUT_MS),
+      signal: request.signal ? AbortSignal.any([request.signal, AbortSignal.timeout(request.timeoutMs ?? REQUEST_TIMEOUT_MS)]) : AbortSignal.timeout(request.timeoutMs ?? REQUEST_TIMEOUT_MS),
     });
   } catch (err) {
     const timedOut = err?.name === "TimeoutError" || err?.name === "AbortError";
@@ -89,6 +112,7 @@ export async function synthesizeStream(request, onChunk) {
         input: text,
         response_format: "pcm",
         stream: true,
+        ...(request.voice ? { voice: request.voice } : {}),
       }),
       signal,
     });

package/tools/http.js

@@ -1,3 +1,18 @@
+import { LAREPASS_PAGE_ORIGINS } from "../larepass/host.js";
+import { pageOrigin } from "../olares/trusted-host.js";
+
+const CORS_HEADERS = "content-type, x-authorization, x-unauth-error, x-lares-session-id, x-lares-file-name, x-lares-upload-request-id, x-lares-voice-id, x-lares-voice-device";
+
+/** @returns {boolean} true when the request came from a LarePass page. */
+function allowLarepassOrigin(req, res) {
+  const origin = pageOrigin(req.headers);
+  if (!origin || !LAREPASS_PAGE_ORIGINS.includes(origin)) return false;
+  res.setHeader("access-control-allow-origin", origin);
+  res.setHeader("access-control-allow-credentials", "true");
+  res.setHeader("vary", "Origin");
+  return true;
+}
+
 export class HttpError extends Error {
   constructor(code, status, message) {
     super(message);
@@ -70,6 +85,7 @@ function errorResponse(err, fallbackCode) {
       status: err.status >= 400 ? err.status : 500,
       code: err.code,
       message: err instanceof Error ? err.message : String(err),
+      params: err.params && typeof err.params === "object" ? err.params : undefined,
     };
   }
   return {
@@ -82,7 +98,7 @@ function errorResponse(err, fallbackCode) {
 export function sendError(res, err, fallbackCode) {
   const failure = errorResponse(err, fallbackCode);
   sendJson(res, failure.status, {
-    error: { code: failure.code, message: failure.message },
+    error: { code: failure.code, message: failure.message, params: failure.params },
   });
 }
 
@@ -99,12 +115,22 @@ export function createRouteHandler(options) {
     const suffix = new URL(req.url ?? "/", "http://x").pathname.slice(options.prefix.length) || "/";
     const path = suffix.replace(/\/+$/, "") || "/";
     const handlers = options.routes[path];
+    const cors = allowLarepassOrigin(req, res);
     if (!handlers) {
       sendJson(res, 404, {
         error: { code: "not_found", message: `no route ${path}` },
       });
       return;
     }
+    if (cors && method === "OPTIONS") {
+      res.writeHead(204, {
+        "access-control-allow-methods": Object.keys(handlers).join(", "),
+        "access-control-allow-headers": CORS_HEADERS,
+        "access-control-max-age": "600",
+      });
+      res.end();
+      return;
+    }
     const route = handlers[method];
     if (!route) {
       sendJson(res, 405, {