@olares/lares-core 0.15.11 → 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",
@@ -18,7 +69,22 @@ 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",
@@ -54,11 +120,13 @@ export const ZH = {
   "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.partial": "显卡已切换，但这些应用没能恢复：{apps}",
-  "settings.mode.failed": "切换显卡模式失败：{error}",
+  "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.noGpu": "没有找到 NVIDIA 显卡",
   "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",
@@ -113,6 +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",
@@ -124,7 +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",
@@ -133,7 +253,29 @@ 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",
@@ -169,11 +311,13 @@ export const EN = {
   "settings.mode.streaming": "While writing",
   "settings.mode.switching": "Switching GPU mode…",
   "settings.mode.resumingApp": "Resuming {app}…",
-  "settings.mode.partial": "GPU switched, but these apps did not resume: {apps}",
-  "settings.mode.failed": "GPU switch failed: {error}",
+  "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.noGpu": "No NVIDIA GPU found",
   "settings.mode.retry": "Retry",
+  "settings.list.separator": ", ",
   "settings.gpu.MemorySlice": "memory slicing",
   "settings.gpu.TimeSlice": "time slicing",
   "settings.gpu.Exclusive": "exclusive",

package/larepass/runtime.js

@@ -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"));

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/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@olares/lares-core",
-  "version": "0.15.11",
+  "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": {
@@ -124,6 +124,7 @@
     "./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",
@@ -168,6 +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

@@ -71,7 +71,7 @@ export async function synthesize(request) {
         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";

package/tools/http.js

@@ -1,7 +1,7 @@
 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";
+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) {
@@ -85,6 +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 {
@@ -97,7 +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 },
   });
 }
 

package/voice/agent-session.js

@@ -0,0 +1,133 @@
+import { mintRpcId } from "../larepass/rpc.js";
+import { createSessionFollow } from "../larepass/mux.js";
+import { reduceUtterance } from "./speech.js";
+
+export const SPOKEN_PROMPT = "This is a voice conversation. Keep user-facing answers concise and natural for speech. Do not read reasoning or tool payloads aloud. Preserve the task's factual results and use the user's language.";
+
+/** dsh opens a turn before claiming input; rpcId associates input with that open turn. */
+export function createVoiceAgentSession(api, { sessionId = null, preset, workspaceId, restoredRequests = [], metric = () => {} } = {}) {
+  let creating;
+  let following;
+  let closed = false;
+  let utterance = null;
+  let error = null;
+  let version = 0;
+  let cursor = -1;
+  let openTurn = null;
+  let latestRequest = null;
+  let followTask = null;
+  let ready = null;
+  let readyTimer;
+  let observedRunning = false;
+  const requests = new Map();
+  for (const restored of restoredRequests.slice(-256)) {
+    requests.set(restored.requestId, { requestId: restored.requestId, turn: null, status: "unknown", utterance: null, error: null, at: Date.now(), firstTextAt: null });
+  }
+  const decoder = createSessionFollow();
+  const unfinished = (task) => ["accepted", "running", "waiting", "unknown"].includes(task.status);
+  const snapshot = () => ({ sessionId, running: observedRunning || [...requests.values()].some(unfinished), utterance, error, version,
+    requestId: latestRequest,
+    tasks: [...requests.values()].map(({ requestId, turn, status, utterance: reply, error: taskError }) => ({ requestId, turn, status, text: reply?.silent ? "" : reply?.text || "", error: taskError })),
+  });
+  const ensure = async () => {
+    if (sessionId) return sessionId;
+    creating ??= Promise.resolve(api.create({ agentPreset: preset, workspaceId })).then((value) => {
+      if (!value?.sessionId) throw new Error("voice_session_failed");
+      sessionId = value.sessionId;
+      return sessionId;
+    });
+    return creating;
+  };
+  const publish = (event, replay = false) => {
+    if (!event || typeof event !== "object") return;
+    if (typeof event.seq === "number") {
+      if (event.seq <= cursor) return;
+      cursor = event.seq;
+    }
+    if (event.type === "turn/start") { openTurn = event.data.turn; observedRunning = true; }
+    if (event.type === "user/message") {
+      const id = event.data?.requestId ?? event.data?.source?.rpcId;
+      const task = requests.get(id);
+      if (task && openTurn !== null) { task.turn = openTurn; task.status = "running"; }
+    }
+    const turn = event.data?.turn ?? openTurn;
+    const owned = [...requests.values()].filter((task) => task.turn != null && task.turn === turn && unfinished(task));
+    for (const task of owned) {
+      task.utterance = reduceUtterance(task.utterance, event);
+      if (event.type === "approval/asked") task.status = "waiting";
+      if (event.type === "approval/decided") task.status = "running";
+      if (event.type === "turn/end") {
+        const reason = event.data?.reason?.kind;
+        task.status = reason === "error" ? "failed" : reason === "cancelled" ? "cancelled" : "completed";
+        task.error = reason === "error" ? "voice_agent_failed" : reason === "cancelled" ? "voice_cancelled" : null;
+      }
+      if (!replay && task.utterance?.text && !task.firstTextAt) {
+        task.firstTextAt = Date.now();
+        metric({ stage: "agent_first_text", requestId: task.requestId, durationMs: task.firstTextAt - task.at });
+      }
+      if (!replay && task.requestId === latestRequest) { utterance = task.utterance; error = task.error; }
+    }
+    if (event.type === "turn/end" && openTurn === event.data?.turn) { openTurn = null; observedRunning = false; }
+    version++;
+  };
+  const follow = () => {
+    if (following && !following.signal.aborted) return ready;
+    const abort = new AbortController();
+    following = abort;
+    let resolveReady;
+    let rejectReady;
+    ready = new Promise((resolve, reject) => { resolveReady = resolve; rejectReady = reject; });
+    readyTimer = setTimeout(() => { rejectReady(new Error("voice_agent_timeout")); abort.abort(); }, 15_000);
+    ready.catch(() => {});
+    abort.signal.addEventListener("abort", () => rejectReady(new Error("voice_session_closed")), { once: true });
+    followTask = (async () => {
+      try {
+        for await (const frame of api.follow({ address: { kind: "session", sessionId }, assistantStream: true }, abort.signal)) {
+          const decoded = decoder.push(frame);
+          if (frame.type === "snapshot") {
+            for (const event of decoded?.events ?? []) publish(event, true);
+            clearTimeout(readyTimer); resolveReady();
+          } else for (const event of decoded?.events ?? []) publish(event);
+        }
+        if (!abort.signal.aborted) throw new Error("voice_agent_disconnected");
+      } catch (cause) {
+        clearTimeout(readyTimer); rejectReady(cause);
+        if (!abort.signal.aborted) { error = "voice_agent_disconnected"; version++; abort.abort(); }
+      }
+    })();
+    followTask.catch(() => {});
+    return ready;
+  };
+  return {
+    snapshot,
+    async open() { if (sessionId) await follow(); },
+    async resume() {
+      following?.abort(); await followTask;
+      closed = false; error = null; utterance = null; cursor = -1; openTurn = null; observedRunning = false; decoder.reset();
+      if (sessionId) await follow();
+    },
+    async submit(text, requestId = mintRpcId(), mode = "queue") {
+      if (closed) throw new Error("voice_session_closed");
+      if (!String(text).trim()) return snapshot();
+      if (requests.has(requestId)) return requests.get(requestId).operation || snapshot();
+      if (requests.size >= 256) throw new Error("voice_session_limit");
+      const task = { requestId, turn: null, status: "accepted", utterance: null, error: null, at: Date.now(), firstTextAt: null };
+      requests.set(requestId, task);
+      latestRequest = requestId; utterance = null; error = null; version++;
+      task.operation = (async () => {
+        try {
+          await ensure();
+          await follow();
+          await api.prompt({ sessionId, requestId, mode: mode === "steer" ? "steer" : "queue", content: [{ type: "text", text }] }, following.signal);
+          return snapshot();
+        } catch (cause) {
+          task.status = "unknown"; task.error = "submission_outcome_unknown"; version++;
+          throw cause;
+        }
+      })();
+      return task.operation;
+    },
+    async cancel() { if (sessionId) await api.cancel({ sessionId }); },
+    async close() { closed = true; clearTimeout(readyTimer); following?.abort(); await followTask; },
+  };
+}

package/voice/agent.js

@@ -29,7 +29,15 @@ export async function liveAgent(options = {}) {
   return agentIsLive(info, now()) ? info : null;
 }
 
-export async function askAgent(op, fields, options) {
+const requests = new Map();
+export function askAgent(op, fields, options = {}) {
+  const key = voiceDir(options);
+  const result = (requests.get(key) ?? Promise.resolve()).catch(() => {}).then(() => performAgentRequest(op, fields, options));
+  requests.set(key, result);
+  void result.finally(() => { if (requests.get(key) === result) requests.delete(key); }).catch(() => {});
+  return result;
+}
+async function performAgentRequest(op, fields, options) {
   const dir = voiceDir(options);
   const now = options.now ?? Date.now;
   const sleep = options.sleep ?? ((ms) => new Promise((resolve) => setTimeout(resolve, ms)));