feat: beaver

main/application.cc

@@ -81,6 +81,7 @@ void Application::Initialize() {
         xEventGroupSetBits(event_group_, MAIN_EVENT_WAKE_WORD_DETECTED);
     };
     callbacks.on_vad_change = [this](bool speaking) {
+        vad_speaking_.store(speaking);
         xEventGroupSetBits(event_group_, MAIN_EVENT_VAD_CHANGE);
     };
     audio_service_.SetCallbacks(callbacks);
@@ -233,6 +234,13 @@ void Application::Run() {
             if (GetDeviceState() == kDeviceStateListening) {
                 auto led = Board::GetInstance().GetLed();
                 led->OnStateChanged();
+
+                if (vad_speaking_.load() && vision_text_mode_enabled_.load() &&
+                    !vision_frame_sent_for_current_listen_.exchange(true)) {
+                    if (!SendCurrentVisionFrame()) {
+                        vision_frame_sent_for_current_listen_.store(false);
+                    }
+                }
             }
         }
 
@@ -525,6 +533,9 @@ void Application::InitializeProtocol() {
     protocol_->OnAudioChannelClosed([this, &board]() {
         board.SetPowerSaveLevel(PowerSaveLevel::LOW_POWER);
         Schedule([this]() {
+            if (GetDeviceState() == kDeviceStateConnecting) {
+                return;
+            }
             auto display = Board::GetInstance().GetDisplay();
             display->SetChatMessage("system", "");
             SetDeviceState(kDeviceStateIdle);
@@ -673,17 +684,39 @@ void Application::DismissAlert() {
 }
 
 void Application::ToggleChatState() {
-    vision_text_mode_enabled_.store(false);
-    xEventGroupSetBits(event_group_, MAIN_EVENT_TOGGLE_CHAT);
+    ToggleChatStateForMode(kChatAgentModeNormal, false);
 }
 
 void Application::ToggleChatStateWithVision() {
-    vision_text_mode_enabled_.store(true);
+    ToggleChatStateForMode(kChatAgentModeNormal, true);
+}
+
+void Application::ToggleChatStateForMode(ChatAgentMode agent_mode, bool vision_enabled) {
+    chat_agent_mode_.store(agent_mode);
+    vision_text_mode_enabled_.store(vision_enabled);
+    vision_frame_sent_for_current_listen_.store(false);
     xEventGroupSetBits(event_group_, MAIN_EVENT_TOGGLE_CHAT);
 }
 
+bool Application::IsVisionTextModeEnabled() const {
+    return vision_text_mode_enabled_.load();
+}
+
+const char* Application::GetChatAgentModeName() const {
+    return chat_agent_mode_.load() == kChatAgentModeBeaver ? "beaver" : "normal";
+}
+
+const char* Application::GetChatModeName() const {
+    bool vision_enabled = vision_text_mode_enabled_.load();
+    if (chat_agent_mode_.load() == kChatAgentModeBeaver) {
+        return vision_enabled ? "vision-beaver" : "beaver";
+    }
+    return vision_enabled ? "vision-normal" : "normal";
+}
+
 void Application::StartListening() {
     vision_text_mode_enabled_.store(false);
+    vision_frame_sent_for_current_listen_.store(false);
     xEventGroupSetBits(event_group_, MAIN_EVENT_START_LISTENING);
 }
 
@@ -693,9 +726,6 @@ void Application::StopListening() {
 
 void Application::HandleToggleChatEvent() {
     auto state = GetDeviceState();
-    if (state != kDeviceStateIdle) {
-        vision_text_mode_enabled_.store(false);
-    }
 
     if (state == kDeviceStateActivating) {
         SetDeviceState(kDeviceStateIdle);
@@ -717,7 +747,12 @@ void Application::HandleToggleChatEvent() {
 
     if (state == kDeviceStateIdle) {
         ListeningMode mode = GetDefaultListeningMode();
-        if (!protocol_->IsAudioChannelOpened()) {
+        bool agent_mode_changed = chat_agent_mode_.load() != active_chat_agent_mode_.load();
+        bool vision_mode_changed = vision_text_mode_enabled_.load() != active_vision_text_mode_enabled_.load();
+        if (!protocol_->IsAudioChannelOpened() || agent_mode_changed || vision_mode_changed) {
+            if (protocol_->IsAudioChannelOpened()) {
+                protocol_->CloseAudioChannel();
+            }
             SetDeviceState(kDeviceStateConnecting);
             // Schedule to let the state change be processed first (UI update)
             Schedule([this, mode]() {
@@ -749,6 +784,8 @@ void Application::ContinueOpenAudioChannel(ListeningMode mode) {
         }
     }
 
+    active_chat_agent_mode_.store(chat_agent_mode_.load());
+    active_vision_text_mode_enabled_.store(vision_text_mode_enabled_.load());
     SetListeningMode(mode);
 }
 
@@ -892,6 +929,7 @@ void Application::HandleStateChangedEvent() {
     switch (new_state) {
         case kDeviceStateUnknown:
         case kDeviceStateIdle:
+            vision_frame_sent_for_current_listen_.store(false);
             display->SetStatus(Lang::Strings::STANDBY);
             display->ClearChatMessages();  // Clear messages first
             display->SetEmotion("neutral"); // Then set emotion (wechat mode checks child count)
@@ -904,6 +942,8 @@ void Application::HandleStateChangedEvent() {
             display->SetChatMessage("system", "");
             break;
         case kDeviceStateListening:
+            vad_speaking_.store(false);
+            vision_frame_sent_for_current_listen_.store(false);
             display->SetStatus(Lang::Strings::LISTENING);
             display->SetEmotion("neutral");
 
@@ -915,9 +955,6 @@ void Application::HandleStateChangedEvent() {
                     audio_service_.WaitForPlaybackQueueEmpty();
                 }
                 
-                if (vision_text_mode_enabled_.load()) {
-                    SendCurrentVisionFrame();
-                }
                 // Send the start listening command
                 protocol_->SendStartListening(listening_mode_);
                 audio_service_.EnableVoiceProcessing(true);
@@ -957,24 +994,25 @@ void Application::HandleStateChangedEvent() {
     }
 }
 
-void Application::SendCurrentVisionFrame() {
+bool Application::SendCurrentVisionFrame() {
     if (!protocol_ || !protocol_->IsAudioChannelOpened()) {
-        return;
+        return false;
     }
 
     auto camera = Board::GetInstance().GetCamera();
     if (camera == nullptr) {
-        return;
+        return false;
     }
 
     std::string jpeg_data;
-    if (!camera->CaptureToJpeg(jpeg_data, false)) {
+    if (!camera->CaptureToJpeg(jpeg_data, true)) {
         ESP_LOGW(TAG, "Failed to capture vision frame");
-        return;
+        return false;
     }
 
     protocol_->SendVisionFrame(jpeg_data);
     ESP_LOGI(TAG, "Sent vision frame, size=%u bytes", static_cast<unsigned>(jpeg_data.size()));
+    return true;
 }
 
 void Application::Schedule(std::function<void()>&& callback) {
@@ -995,6 +1033,8 @@ void Application::AbortSpeaking(AbortReason reason) {
 
 void Application::SetListeningMode(ListeningMode mode) {
     listening_mode_ = mode;
+    vad_speaking_.store(false);
+    vision_frame_sent_for_current_listen_.store(false);
     SetDeviceState(kDeviceStateListening);
 }
 

main/application.h

@@ -41,6 +41,11 @@ enum AecMode {
     kAecOnServerSide,
 };
 
+enum ChatAgentMode {
+    kChatAgentModeNormal,
+    kChatAgentModeBeaver,
+};
+
 class Application {
 public:
     static Application& GetInstance() {
@@ -93,6 +98,11 @@ public:
      */
     void ToggleChatState();
     void ToggleChatStateWithVision();
+    void ToggleChatStateForMode(ChatAgentMode agent_mode, bool vision_enabled);
+    bool IsVisionTextModeEnabled() const;
+    ChatAgentMode GetChatAgentMode() const { return chat_agent_mode_.load(); }
+    const char* GetChatAgentModeName() const;
+    const char* GetChatModeName() const;
 
     /**
      * Start listening (event-based, thread-safe)
@@ -146,7 +156,12 @@ private:
     bool aborted_ = false;
     bool assets_version_checked_ = false;
     bool play_popup_on_listening_ = false;  // Flag to play popup sound after state changes to listening
+    std::atomic<ChatAgentMode> chat_agent_mode_ = kChatAgentModeNormal;
+    std::atomic<ChatAgentMode> active_chat_agent_mode_ = kChatAgentModeNormal;
     std::atomic<bool> vision_text_mode_enabled_ = false;
+    std::atomic<bool> active_vision_text_mode_enabled_ = false;
+    std::atomic<bool> vad_speaking_ = false;
+    std::atomic<bool> vision_frame_sent_for_current_listen_ = false;
     int clock_ticks_ = 0;
     TaskHandle_t activation_task_handle_ = nullptr;
 
@@ -162,7 +177,7 @@ private:
     void HandleWakeWordDetectedEvent();
     void ContinueOpenAudioChannel(ListeningMode mode);
     void ContinueWakeWordInvoke(const std::string& wake_word);
-    void SendCurrentVisionFrame();
+    bool SendCurrentVisionFrame();
 
     // Activation task (runs in background)
     void ActivationTask();

main/boards/m5stack-core-s3/m5stack_core_s3.cc

@@ -192,6 +192,7 @@ private:
     void PollTouchpad() {
         static bool was_touched = false;
         static int64_t touch_start_time = 0;
+        static int touch_start_x = -1;
         const int64_t TOUCH_THRESHOLD_MS = 500;  // 触摸时长阈值，超过500ms视为长按
 
         ft6336_->UpdateTouchPoint();
@@ -201,11 +202,14 @@ private:
         if (touch_point.num > 0 && !was_touched) {
             was_touched = true;
             touch_start_time = esp_timer_get_time() / 1000;  // 转换为毫秒
+            touch_start_x = touch_point.x;
         }
         // 检测触摸释放
         else if (touch_point.num == 0 && was_touched) {
             was_touched = false;
             int64_t touch_duration = (esp_timer_get_time() / 1000) - touch_start_time;
+            bool beaver_mode = touch_start_x >= DISPLAY_WIDTH / 2;
+            auto agent_mode = beaver_mode ? kChatAgentModeBeaver : kChatAgentModeNormal;
 
             if (touch_duration < TOUCH_THRESHOLD_MS) {
                 auto& app = Application::GetInstance();
@@ -213,12 +217,12 @@ private:
                     EnterWifiConfigMode();
                     return;
                 }
-                ESP_LOGI(TAG, "Touch short: text-only mode");
-                app.ToggleChatState();
+                ESP_LOGI(TAG, "Touch short: %s text-only mode", beaver_mode ? "beaver" : "normal");
+                app.ToggleChatStateForMode(agent_mode, false);
             } else {
                 auto& app = Application::GetInstance();
-                ESP_LOGI(TAG, "Touch long: vision+text mode");
-                app.ToggleChatStateWithVision();
+                ESP_LOGI(TAG, "Touch long: %s vision+text mode", beaver_mode ? "beaver" : "normal");
+                app.ToggleChatStateForMode(agent_mode, true);
             }
         }
     }
@@ -344,18 +348,23 @@ private:
                 vTaskDelay(pdMS_TO_TICKS(BACKGROUND_VISION_INITIAL_DELAY_MS));
 
                 while (true) {
+                    if (!Application::GetInstance().IsVisionTextModeEnabled()) {
+                        vTaskDelay(pdMS_TO_TICKS(BACKGROUND_VISION_SAMPLE_INTERVAL_MS));
+                        continue;
+                    }
+
                     if (board->camera_ == nullptr) {
                         vTaskDelay(pdMS_TO_TICKS(BACKGROUND_VISION_SAMPLE_INTERVAL_MS));
                         continue;
                     }
 
-                    if (board->camera_->CaptureBackground()) {
+                    if (board->camera_->Capture()) {
                         if (!has_logged_success) {
-                            ESP_LOGI(TAG, "Background vision sampler started");
+                            ESP_LOGI(TAG, "Vision preview sampler started");
                             has_logged_success = true;
                         }
                     } else if (!has_logged_failure) {
-                        ESP_LOGW(TAG, "Background vision sampler is waiting for camera");
+                        ESP_LOGW(TAG, "Vision preview sampler is waiting for camera");
                         has_logged_failure = true;
                     }
 

main/bridge_server.py

@@ -2,6 +2,7 @@ import asyncio
 import base64
 import json
 import os
+import re
 import shutil
 import struct
 import sys
@@ -28,7 +29,18 @@ TOKEN_URL = "http://172.19.0.240:8000/getToken"
 LIVEKIT_WS_URL = "ws://172.19.0.240:7880"
 ROOM_PREFIX = "test-livekit"
 IDENTITY_PREFIX = "uv-livekit"
-AGENT_NAME = "my-agent"
+LEGACY_AGENT_NAME = os.getenv("LIVEKIT_AGENT_NAME", "normal-agent")
+DEFAULT_AGENT_MODE = os.getenv("LIVEKIT_DEFAULT_AGENT_MODE", "normal").strip().lower()
+AGENT_NAMES = {
+    "normal": os.getenv("LIVEKIT_NORMAL_AGENT_NAME", LEGACY_AGENT_NAME),
+    "beaver": os.getenv("LIVEKIT_BEAVER_AGENT_NAME", "beaver-agent"),
+}
+CHAT_MODE_AGENT_NAMES = {
+    "normal": AGENT_NAMES["normal"],
+    "beaver": AGENT_NAMES["beaver"],
+    "vision-normal": os.getenv("LIVEKIT_VISION_NORMAL_AGENT_NAME", "vision-normal-agent"),
+    "vision-beaver": os.getenv("LIVEKIT_VISION_BEAVER_AGENT_NAME", "vision-beaver-agent"),
+}
 CONNECT_TIMEOUT_SECONDS = float(os.getenv("LIVEKIT_CONNECT_TIMEOUT_SECONDS", "20.0"))
 AGENT_READY_TIMEOUT_SECONDS = float(os.getenv("LIVEKIT_AGENT_READY_TIMEOUT_SECONDS", "10.0"))
 WS_PORT = 8080
@@ -55,6 +67,16 @@ TTS_DISPLAY_SCROLL_WIDTH = int(os.getenv("TTS_DISPLAY_SCROLL_WIDTH", "18"))
 TTS_DISPLAY_SCROLL_INTERVAL_SECONDS = float(os.getenv("TTS_DISPLAY_SCROLL_INTERVAL_SECONDS", "0.18"))
 TTS_DISPLAY_SCROLL_GAP = "   "
 TTS_INTERRUPT_SUPPRESS_SECONDS = float(os.getenv("TTS_INTERRUPT_SUPPRESS_SECONDS", "0.8"))
+EMOTION_TEXT_PATTERN = re.compile(
+    r"^\s*<?\s*emotion\s*=\s*([^\s>,，;；]+)\s*>?[\s,，;；]*(.*)$",
+    re.DOTALL,
+)
+EMOTION_TEST_SEQUENCE = [
+    emotion.strip()
+    for emotion in os.getenv("BRIDGE_EMOTION_TEST_SEQUENCE", "").split(",")
+    if emotion.strip()
+]
+EMOTION_TEST_INTERVAL_SECONDS = float(os.getenv("BRIDGE_EMOTION_TEST_INTERVAL_SECONDS", "2.0"))
 
 
 @dataclass
@@ -64,6 +86,10 @@ class DeviceSession:
     protocol_version: int
     room_name: str
     identity: str
+    chat_mode: str
+    agent_mode: str
+    agent_name: str
+    vision_enabled: bool
     room: rtc.Room
     mic_source: AudioSource
     agent_ready: asyncio.Event
@@ -75,6 +101,7 @@ class DeviceSession:
     tts_transcript_text: str = ""
     tts_display_text: str = ""
     tts_display_final: bool = False
+    tts_emotion: str = ""
     tts_suppressed_until: float = 0.0
     agent_dispatch_task: Optional[asyncio.Task] = None
     closed: bool = False
@@ -82,10 +109,10 @@ class DeviceSession:
     first_capture_log_time: float = 0.0
 
 
-async def fetch_token(room_name: str, identity: str) -> str:
+async def fetch_token(room_name: str, identity: str, agent_name: str) -> str:
     params = {"room": room_name, "identity": identity}
     if AGENT_DISPATCH_MODE == "token":
-        params["agent_name"] = AGENT_NAME
+        params["agent_name"] = agent_name
 
     async with httpx.AsyncClient(timeout=15.0, follow_redirects=True) as client:
         response = await client.get(TOKEN_URL, params=params)
@@ -125,15 +152,61 @@ class ESP32LiveKitBridge:
         print(f"[session] device={device_id} room={room_name} identity={identity}")
         return room_name, identity
 
-    def _is_agent_participant(self, participant: rtc.RemoteParticipant) -> bool:
+    def _resolve_agent_selection(self, headers: Any) -> tuple[str, str, str, bool]:
+        requested_chat_mode = (
+            headers.get("Chat-Mode")
+            or headers.get("X-Chat-Mode")
+            or ""
+        ).strip().lower()
+        chat_mode_to_agent = {
+            "normal": ("normal", False),
+            "beaver": ("beaver", False),
+            "vision-normal": ("normal", True),
+            "vision-beaver": ("beaver", True),
+        }
+
+        if requested_chat_mode in chat_mode_to_agent:
+            requested_mode, vision_enabled = chat_mode_to_agent[requested_chat_mode]
+        else:
+            if requested_chat_mode:
+                print(f"未知 Chat-Mode={requested_chat_mode!r}，回退到 Agent-Mode")
+            requested_mode = (
+                headers.get("Agent-Mode")
+                or headers.get("X-Agent-Mode")
+                or DEFAULT_AGENT_MODE
+                or "normal"
+            ).strip().lower()
+            vision_enabled = False
+            requested_chat_mode = requested_mode
+
+        requested_name = headers.get("Agent-Name") or headers.get("X-Agent-Name")
+        if requested_name:
+            return requested_chat_mode, "custom", requested_name, vision_enabled
+
+        if requested_mode not in AGENT_NAMES:
+            print(f"未知 Agent-Mode={requested_mode!r}，回退到 normal")
+            requested_mode = "normal"
+            requested_chat_mode = "vision-normal" if vision_enabled else "normal"
+
+        if requested_chat_mode in CHAT_MODE_AGENT_NAMES:
+            return (
+                requested_chat_mode,
+                requested_mode,
+                CHAT_MODE_AGENT_NAMES[requested_chat_mode],
+                vision_enabled,
+            )
+
+        return requested_chat_mode, requested_mode, AGENT_NAMES[requested_mode], vision_enabled
+
+    def _is_agent_participant(self, participant: rtc.RemoteParticipant, agent_name: str) -> bool:
         identity = getattr(participant, "identity", "") or ""
-        return identity.startswith("agent-") or AGENT_NAME in identity
+        return identity.startswith("agent-") or agent_name in identity
 
     def _get_agent_identities(self, session: DeviceSession) -> list[str]:
         return [
             participant.identity
             for participant in session.room.remote_participants.values()
-            if self._is_agent_participant(participant)
+            if self._is_agent_participant(participant, session.agent_name)
         ]
 
     def _log_agent_participants(self, session: DeviceSession, source: str) -> None:
@@ -175,7 +248,7 @@ class ESP32LiveKitBridge:
                 dispatch_agent_name = getattr(dispatch, "agent_name", None)
                 dispatch_room = getattr(dispatch, "room", None)
                 if dispatch_room == session.room_name and (
-                    dispatch_agent_name == AGENT_NAME or dispatch_agent_name is None
+                    dispatch_agent_name == session.agent_name or dispatch_agent_name is None
                 ):
                     print(
                         f"检测到已有 dispatch: room={session.room_name} "
@@ -198,7 +271,7 @@ class ESP32LiveKitBridge:
             return
 
         for participant in session.room.remote_participants.values():
-            if self._is_agent_participant(participant):
+            if self._is_agent_participant(participant, session.agent_name):
                 # print(f"Agent 已在房间中，跳过 dispatch: {participant.identity}")
                 return
 
@@ -210,7 +283,7 @@ class ESP32LiveKitBridge:
         await session.agent_dispatch_task
 
     async def _dispatch_agent(self, session: DeviceSession) -> None:
-        print(f"准备 dispatch agent: room={session.room_name}, agent={AGENT_NAME}")
+        print(f"准备 dispatch agent: room={session.room_name}, agent={session.agent_name}")
 
         try:
             if await self._dispatch_agent_with_sdk(session):
@@ -240,13 +313,17 @@ class ESP32LiveKitBridge:
         try:
             dispatch = await lkapi.agent_dispatch.create_dispatch(
                 livekit_api.CreateAgentDispatchRequest(
-                    agent_name=AGENT_NAME,
+                    agent_name=session.agent_name,
                     room=session.room_name,
                     metadata=json.dumps(
                         {
                             "source": "bridge_server",
                             "identity": session.identity,
                             "device_id": session.device_id,
+                            "chat_mode": session.chat_mode,
+                            "agent_mode": session.agent_mode,
+                            "agent_name": session.agent_name,
+                            "vision_enabled": session.vision_enabled,
                         }
                     ),
                 )
@@ -269,13 +346,17 @@ class ESP32LiveKitBridge:
             "--room",
             session.room_name,
             "--agent-name",
-            AGENT_NAME,
+            session.agent_name,
             "--metadata",
             json.dumps(
                 {
                     "source": "bridge_server",
                     "identity": session.identity,
                     "device_id": session.device_id,
+                    "chat_mode": session.chat_mode,
+                    "agent_mode": session.agent_mode,
+                    "agent_name": session.agent_name,
+                    "vision_enabled": session.vision_enabled,
                 }
             ),
             stdout=asyncio.subprocess.PIPE,
@@ -294,7 +375,7 @@ class ESP32LiveKitBridge:
             print(f"lk dispatch create 失败，退出码: {process.returncode}")
             return False
 
-        print(f"Agent dispatch 已通过 lk CLI 创建: room={session.room_name}, agent={AGENT_NAME}")
+        print(f"Agent dispatch 已通过 lk CLI 创建: room={session.room_name}, agent={session.agent_name}")
         return True
 
     async def _publish_agent_event(self, session: DeviceSession, payload: dict[str, Any]) -> bool:
@@ -415,9 +496,40 @@ class ESP32LiveKitBridge:
         await session.websocket.send(json.dumps({"type": "tts", "state": state}))
         print(f"已发送 tts {state}: device={session.device_id}")
 
+    async def _send_emotion(self, session: DeviceSession, emotion: str) -> None:
+        if session.websocket is None:
+            print(f"跳过 emotion {emotion}，ESP32 尚未连接")
+            return
+        await session.websocket.send(json.dumps({"type": "llm", "emotion": emotion}))
+        print(f"已发送 emotion: device={session.device_id} emotion={emotion}")
+
+    def _parse_emotion_text(self, text: str) -> tuple[Optional[str], str]:
+        match = EMOTION_TEXT_PATTERN.match(text)
+        if match is None:
+            return None, text.strip()
+        emotion, tts_text = match.groups()
+        return emotion.strip(), tts_text.strip()
+
+    async def _run_emotion_test_sequence(self, session: DeviceSession) -> None:
+        if not EMOTION_TEST_SEQUENCE:
+            return
+
+        for index, emotion in enumerate(EMOTION_TEST_SEQUENCE):
+            if session.websocket is None or session.closed:
+                return
+            if index > 0:
+                await asyncio.sleep(EMOTION_TEST_INTERVAL_SECONDS)
+            await self._send_emotion(session, emotion)
+
     async def _send_tts_text(self, session: DeviceSession, text: str, final: bool) -> None:
         if session.websocket is None:
             return
+        raw_text = text
+        _emotion, text = self._parse_emotion_text(text)
+        if not text:
+            print(f"[tts->esp32] skip empty text: raw={raw_text!r} final={final}")
+            return
+        print(f"[tts->esp32] text={text!r} final={final}")
         await session.websocket.send(
             json.dumps(
                 {
@@ -493,6 +605,7 @@ class ESP32LiveKitBridge:
         if not session.tts_display_text:
             session.tts_transcript_text = ""
             session.tts_display_final = False
+            session.tts_emotion = ""
             self._cancel_tts_display_task(session)
         await self._send_tts_state(session, "start")
         session.tts_active = True
@@ -507,6 +620,7 @@ class ESP32LiveKitBridge:
         session.tts_transcript_text = ""
         session.tts_display_text = ""
         session.tts_display_final = False
+        session.tts_emotion = ""
 
     async def _force_stop_tts(self, session: DeviceSession, reason: str) -> None:
         self._cancel_tts_display_task(session)
@@ -517,6 +631,7 @@ class ESP32LiveKitBridge:
         session.tts_transcript_text = ""
         session.tts_display_text = ""
         session.tts_display_final = False
+        session.tts_emotion = ""
         await self._send_tts_state(session, "stop")
         print(f"已强制停止本地 TTS: device={session.device_id} reason={reason}")
 
@@ -681,16 +796,16 @@ class ESP32LiveKitBridge:
             print(f"✅ 成功连接到 LiveKit 房间: room={session.room_name}")
             self._log_agent_participants(session, "connected")
             for participant in session.room.remote_participants.values():
-                if self._is_agent_participant(participant):
+                if self._is_agent_participant(participant, session.agent_name):
                     session.agent_ready.set()
                     self._scan_participant_audio_tracks(session, participant, "connected_scan")
 
         @session.room.on("participant_connected")
         def on_participant_connected(participant: rtc.RemoteParticipant) -> None:
-            role = "Agent" if self._is_agent_participant(participant) else "Remote participant"
+            role = "Agent" if self._is_agent_participant(participant, session.agent_name) else "Remote participant"
             print(f"👋 {role} ({participant.identity}) 已加入房间: room={session.room_name}")
             self._log_agent_participants(session, "participant_connected")
-            if self._is_agent_participant(participant):
+            if self._is_agent_participant(participant, session.agent_name):
                 session.agent_ready.set()
                 self._scan_participant_audio_tracks(
                     session, participant, "participant_connected_scan"
@@ -723,14 +838,26 @@ class ESP32LiveKitBridge:
             track_pub: rtc.TrackPublication,
         ) -> None:
             identity = participant.identity if participant else "未知"
-            is_agent = isinstance(participant, rtc.RemoteParticipant) and self._is_agent_participant(participant)
+            is_agent = isinstance(participant, rtc.RemoteParticipant) and self._is_agent_participant(
+                participant, session.agent_name
+            )
             for segment in segments:
                 status = "✅ 最终结果" if segment.final else "⏳ 正在思考/中间结果"
                 print(f"🗣️  [{status} | room={session.room_name} | {identity}]: {segment.text}")
                 if is_agent:
                     if time.monotonic() < session.tts_suppressed_until:
                         continue
-                    display_text = self._current_tts_display_text(segment.text)
+                    print(f"[livekit-llm] raw={segment.text!r} final={segment.final}")
+                    emotion, tts_text = self._parse_emotion_text(segment.text)
+                    print(
+                        f"[livekit-llm] parsed emotion={emotion!r} "
+                        f"tts_text={tts_text!r} final={segment.final}"
+                    )
+                    if emotion and emotion != session.tts_emotion:
+                        session.tts_emotion = emotion
+                        asyncio.create_task(self._send_emotion(session, emotion))
+                    display_text = self._current_tts_display_text(tts_text)
+                    print(f"[livekit-llm] display_text={display_text!r} final={segment.final}")
                     if not display_text or display_text == session.tts_transcript_text:
                         continue
                     session.tts_transcript_text = display_text
@@ -785,7 +912,7 @@ class ESP32LiveKitBridge:
         # print(f"[config] token_url={TOKEN_URL}")
         # print(f"[config] room={session.room_name} identity={session.identity}")
         # print(f"[config] livekit_connect_timeout={CONNECT_TIMEOUT_SECONDS}")
-        token = await fetch_token(session.room_name, session.identity)
+        token = await fetch_token(session.room_name, session.identity, session.agent_name)
 
         try:
             await session.room.connect(
@@ -830,8 +957,12 @@ class ESP32LiveKitBridge:
         # print(f"等待 agent 加入: room={session.room_name}")
         try:
             await asyncio.wait_for(session.agent_ready.wait(), timeout=AGENT_READY_TIMEOUT_SECONDS)
+            if session.closed:
+                return
             # print(f"✅ agent 已就绪: room={session.room_name}")
         except asyncio.TimeoutError:
+            if session.closed:
+                return
             print(f"⚠️ agent 等待超时: room={session.room_name}")
 
     async def start(self) -> None:
@@ -839,6 +970,20 @@ class ESP32LiveKitBridge:
         print(f"[config] livekit_ws_url={LIVEKIT_WS_URL}")
         print(f"[config] token_url={TOKEN_URL}")
         print(f"[config] agent_dispatch_mode={AGENT_DISPATCH_MODE}")
+        print(
+            "[config] agents="
+            f"normal:{CHAT_MODE_AGENT_NAMES['normal']} "
+            f"beaver:{CHAT_MODE_AGENT_NAMES['beaver']} "
+            f"vision-normal:{CHAT_MODE_AGENT_NAMES['vision-normal']} "
+            f"vision-beaver:{CHAT_MODE_AGENT_NAMES['vision-beaver']} "
+            f"default_mode:{DEFAULT_AGENT_MODE}"
+        )
+        if EMOTION_TEST_SEQUENCE:
+            print(
+                "[config] emotion_test_sequence="
+                f"{','.join(EMOTION_TEST_SEQUENCE)} "
+                f"interval={EMOTION_TEST_INTERVAL_SECONDS}s"
+            )
 
     async def close(self) -> None:
         for session in list(self.device_sessions.values()):
@@ -849,6 +994,7 @@ class ESP32LiveKitBridge:
             return
         session.closed = True
         session.websocket = None
+        session.agent_ready.set()
         session.tts_active = False
         session.tts_stream_id += 1
         if session.tts_idle_task is not None:
@@ -998,6 +1144,7 @@ class ESP32LiveKitBridge:
             await self._close_session(existing_session)
             self.device_sessions.pop(device_id, None)
 
+        chat_mode, agent_mode, agent_name, vision_enabled = self._resolve_agent_selection(websocket.request.headers)
         room_name, identity = self._build_session_names(device_id)
         session = DeviceSession(
             device_id=device_id,
@@ -1005,6 +1152,10 @@ class ESP32LiveKitBridge:
             protocol_version=protocol_version,
             room_name=room_name,
             identity=identity,
+            chat_mode=chat_mode,
+            agent_mode=agent_mode,
+            agent_name=agent_name,
+            vision_enabled=vision_enabled,
             room=rtc.Room(),
             mic_source=AudioSource(sample_rate=INPUT_SAMPLE_RATE, num_channels=1),
             agent_ready=asyncio.Event(),
@@ -1013,6 +1164,11 @@ class ESP32LiveKitBridge:
 
         print(f"ESP32 已连接: device={device_id}")
         print(f"ESP32 协议版本: {session.protocol_version}")
+        print(
+            f"ESP32 mode: chat={session.chat_mode} "
+            f"agent={session.agent_mode}/{session.agent_name} "
+            f"vision={session.vision_enabled}"
+        )
         session.tts_stream_id += 1
         opus_decoder = None
 
@@ -1033,6 +1189,7 @@ class ESP32LiveKitBridge:
             }
             await websocket.send(json.dumps(hello_msg))
             print(f"已发送 server hello: device={device_id} room={session.room_name}")
+            asyncio.create_task(self._run_emotion_test_sequence(session))
 
             await self._connect_session_room(session)
 

main/protocols/websocket_protocol.cc

@@ -119,6 +119,8 @@ bool WebsocketProtocol::OpenAudioChannel() {
     websocket_->SetHeader("Protocol-Version", std::to_string(version_).c_str());
     websocket_->SetHeader("Device-Id", SystemInfo::GetMacAddress().c_str());
     websocket_->SetHeader("Client-Id", Board::GetInstance().GetUuid().c_str());
+    websocket_->SetHeader("Agent-Mode", Application::GetInstance().GetChatAgentModeName());
+    websocket_->SetHeader("Chat-Mode", Application::GetInstance().GetChatModeName());
 
     websocket_->OnData([this](const char* data, size_t len, bool binary) {
         if (binary) {