fix: server error to disconnect

.gitignore

@@ -10,6 +10,7 @@ sdkconfig
 dependencies.lock
 .env
 releases/
+vision_frames/
 main/assets/lang_config.h
 main/mmap_generate_emoji.h
 .DS_Store
@@ -18,3 +19,4 @@ main/mmap_generate_emoji.h
 *.bin
 mmap_generate_*.h
 .clangd
+background_frames/

main/Kconfig.projbuild

@@ -6,6 +6,34 @@ config OTA_URL
     help
         The application will access this URL to check for new firmwares and server address.
 
+config USE_DIRECT_WEBSOCKET
+    bool "Use direct WebSocket without OTA"
+    default n
+    help
+        Skip the OTA server check and use the WebSocket settings below directly.
+
+config WEBSOCKET_URL
+    string "Default WebSocket URL"
+    depends on USE_DIRECT_WEBSOCKET
+    default "ws://172.19.0.240:8080"
+    help
+        The WebSocket server URL used when direct WebSocket mode is enabled.
+
+config WEBSOCKET_TOKEN
+    string "Default WebSocket token"
+    depends on USE_DIRECT_WEBSOCKET
+    default ""
+    help
+        Optional Authorization token for the direct WebSocket server.
+
+config WEBSOCKET_PROTOCOL_VERSION
+    int "Default WebSocket protocol version"
+    depends on USE_DIRECT_WEBSOCKET
+    range 1 3
+    default 1
+    help
+        Protocol-Version header and hello version used by the WebSocket protocol.
+
 choice
     prompt "Flash Assets"
     default FLASH_DEFAULT_ASSETS if !USE_EMOTE_MESSAGE_STYLE

main/application.cc

@@ -1,25 +1,24 @@
 #include "application.h"
+#include "assets.h"
+#include "assets/lang_config.h"
+#include "audio_codec.h"
 #include "board.h"
 #include "display.h"
-#include "system_info.h"
-#include "audio_codec.h"
-#include "mqtt_protocol.h"
-#include "websocket_protocol.h"
-#include "assets/lang_config.h"
 #include "mcp_server.h"
-#include "assets.h"
+#include "mqtt_protocol.h"
 #include "settings.h"
+#include "system_info.h"
+#include "websocket_protocol.h"
 
-#include <cstring>
-#include <esp_log.h>
-#include <cJSON.h>
 #include <driver/gpio.h>
+#include <esp_log.h>
 #include <arpa/inet.h>
+#include <cJSON.h>
 #include <font_awesome.h>
+#include <cstring>
 
 #define TAG "Application"
 
-
 Application::Application() {
     event_group_ = xEventGroupCreate();
 
@@ -33,16 +32,16 @@ Application::Application() {
     aec_mode_ = kAecOff;
 #endif
 
-    esp_timer_create_args_t clock_timer_args = {
+    esp_timer_create_args_t clock_timer_args = {.callback =
-        .callback = [](void* arg) {
+                                                    [](void* arg) {
                                                         Application* app = (Application*)arg;
-            xEventGroupSetBits(app->event_group_, MAIN_EVENT_CLOCK_TICK);
+                                                        xEventGroupSetBits(app->event_group_,
+                                                                           MAIN_EVENT_CLOCK_TICK);
                                                     },
                                                 .arg = this,
                                                 .dispatch_method = ESP_TIMER_TASK,
                                                 .name = "clock_timer",
-        .skip_unhandled_events = true
+                                                .skip_unhandled_events = true};
-    };
     esp_timer_create(&clock_timer_args, &clock_timer_handle_);
 }
 
@@ -54,9 +53,7 @@ Application::~Application() {
     vEventGroupDelete(event_group_);
 }
 
-bool Application::SetDeviceState(DeviceState state) {
+bool Application::SetDeviceState(DeviceState state) { return state_machine_.TransitionTo(state); }
-    return state_machine_.TransitionTo(state);
-}
 
 void Application::Initialize() {
     auto& board = Board::GetInstance();
@@ -81,6 +78,7 @@ 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);
@@ -141,13 +139,16 @@ void Application::Initialize() {
                 display->SetStatus(Lang::Strings::DETECTING_MODULE);
                 break;
             case NetworkEvent::ModemErrorNoSim:
-                Alert(Lang::Strings::ERROR, Lang::Strings::PIN_ERROR, "triangle_exclamation", Lang::Sounds::OGG_ERR_PIN);
+                Alert(Lang::Strings::ERROR, Lang::Strings::PIN_ERROR, "triangle_exclamation",
+                      Lang::Sounds::OGG_ERR_PIN);
                 break;
             case NetworkEvent::ModemErrorRegDenied:
-                Alert(Lang::Strings::ERROR, Lang::Strings::REG_ERROR, "triangle_exclamation", Lang::Sounds::OGG_ERR_REG);
+                Alert(Lang::Strings::ERROR, Lang::Strings::REG_ERROR, "triangle_exclamation",
+                      Lang::Sounds::OGG_ERR_REG);
                 break;
             case NetworkEvent::ModemErrorInitFailed:
-                Alert(Lang::Strings::ERROR, Lang::Strings::MODEM_INIT_ERROR, "triangle_exclamation", Lang::Sounds::OGG_EXCLAMATION);
+                Alert(Lang::Strings::ERROR, Lang::Strings::MODEM_INIT_ERROR, "triangle_exclamation",
+                      Lang::Sounds::OGG_EXCLAMATION);
                 break;
             case NetworkEvent::ModemErrorTimeout:
                 display->SetStatus(Lang::Strings::REGISTERING_NETWORK);
@@ -167,18 +168,10 @@ void Application::Run() {
     vTaskPrioritySet(nullptr, 10);
 
     const EventBits_t ALL_EVENTS =
-        MAIN_EVENT_SCHEDULE |
+        MAIN_EVENT_SCHEDULE | MAIN_EVENT_SEND_AUDIO | MAIN_EVENT_WAKE_WORD_DETECTED |
-        MAIN_EVENT_SEND_AUDIO |
+        MAIN_EVENT_VAD_CHANGE | MAIN_EVENT_CLOCK_TICK | MAIN_EVENT_ERROR |
-        MAIN_EVENT_WAKE_WORD_DETECTED |
+        MAIN_EVENT_NETWORK_CONNECTED | MAIN_EVENT_NETWORK_DISCONNECTED | MAIN_EVENT_TOGGLE_CHAT |
-        MAIN_EVENT_VAD_CHANGE |
+        MAIN_EVENT_START_LISTENING | MAIN_EVENT_STOP_LISTENING | MAIN_EVENT_ACTIVATION_DONE |
-        MAIN_EVENT_CLOCK_TICK |
-        MAIN_EVENT_ERROR |
-        MAIN_EVENT_NETWORK_CONNECTED |
-        MAIN_EVENT_NETWORK_DISCONNECTED |
-        MAIN_EVENT_TOGGLE_CHAT |
-        MAIN_EVENT_START_LISTENING |
-        MAIN_EVENT_STOP_LISTENING |
-        MAIN_EVENT_ACTIVATION_DONE |
         MAIN_EVENT_STATE_CHANGED;
 
     while (true) {
@@ -186,7 +179,8 @@ void Application::Run() {
 
         if (bits & MAIN_EVENT_ERROR) {
             SetDeviceState(kDeviceStateIdle);
-            Alert(Lang::Strings::ERROR, last_error_message_.c_str(), "circle_xmark", Lang::Sounds::OGG_EXCLAMATION);
+            Alert(Lang::Strings::ERROR, last_error_message_.c_str(), "circle_xmark",
+                  Lang::Sounds::OGG_EXCLAMATION);
         }
 
         if (bits & MAIN_EVENT_NETWORK_CONNECTED) {
@@ -233,6 +227,13 @@ 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);
+                    }
+                }
             }
         }
 
@@ -270,12 +271,14 @@ void Application::HandleNetworkConnectedEvent() {
             return;
         }
 
-        xTaskCreate([](void* arg) {
+        xTaskCreate(
+            [](void* arg) {
                 Application* app = static_cast<Application*>(arg);
                 app->ActivationTask();
                 app->activation_task_handle_ = nullptr;
                 vTaskDelete(NULL);
-        }, "activation", 4096 * 2, this, 2, &activation_task_handle_);
+            },
+            "activation", 4096 * 2, this, 2, &activation_task_handle_);
     }
 
     // Update the status bar immediately to show the network state
@@ -286,7 +289,8 @@ void Application::HandleNetworkConnectedEvent() {
 void Application::HandleNetworkDisconnectedEvent() {
     // Close current conversation when network disconnected
     auto state = GetDeviceState();
-    if (state == kDeviceStateConnecting || state == kDeviceStateListening || state == kDeviceStateSpeaking) {
+    if (state == kDeviceStateConnecting || state == kDeviceStateListening ||
+        state == kDeviceStateThinking || state == kDeviceStateSpeaking) {
         ESP_LOGI(TAG, "Closing audio channel due to network disconnection");
         protocol_->CloseAudioChannel();
     }
@@ -302,11 +306,15 @@ void Application::HandleActivationDoneEvent() {
     SystemInfo::PrintHeapStats();
     SetDeviceState(kDeviceStateIdle);
 
+    if (ota_ != nullptr) {
         has_server_time_ = ota_->HasServerTime();
+    }
 
     auto display = Board::GetInstance().GetDisplay();
+    if (ota_ != nullptr) {
         std::string message = std::string(Lang::Strings::VERSION) + ota_->GetCurrentVersion();
         display->ShowNotification(message.c_str());
+    }
     display->SetChatMessage("system", "");
 
     // Release OTA object after activation is complete
@@ -321,6 +329,10 @@ void Application::HandleActivationDoneEvent() {
 }
 
 void Application::ActivationTask() {
+#if CONFIG_USE_DIRECT_WEBSOCKET
+    CheckAssetsVersion();
+    InitializeProtocol();
+#else
     // Create OTA object for activation process
     ota_ = std::make_unique<Ota>();
 
@@ -332,6 +344,7 @@ void Application::ActivationTask() {
 
     // Initialize the protocol
     InitializeProtocol();
+#endif
 
     // Signal completion to main loop
     xEventGroupSetBits(event_group_, MAIN_EVENT_ACTIVATION_DONE);
@@ -362,7 +375,8 @@ void Application::CheckAssetsVersion() {
 
         char message[256];
         snprintf(message, sizeof(message), Lang::Strings::FOUND_NEW_ASSETS, download_url.c_str());
-        Alert(Lang::Strings::LOADING_ASSETS, message, "cloud_arrow_down", Lang::Sounds::OGG_UPGRADE);
+        Alert(Lang::Strings::LOADING_ASSETS, message, "cloud_arrow_down",
+              Lang::Sounds::OGG_UPGRADE);
 
         // Wait for the audio service to be idle for 3 seconds
         vTaskDelay(pdMS_TO_TICKS(3000));
@@ -370,7 +384,8 @@ void Application::CheckAssetsVersion() {
         board.SetPowerSaveLevel(PowerSaveLevel::PERFORMANCE);
         display->SetChatMessage("system", Lang::Strings::PLEASE_WAIT);
 
-        bool success = assets.Download(download_url, [this, display](int progress, size_t speed) -> void {
+        bool success =
+            assets.Download(download_url, [this, display](int progress, size_t speed) -> void {
                 char buffer[32];
                 snprintf(buffer, sizeof(buffer), "%d%% %uKB/s", progress, speed / 1024);
                 Schedule([display, message = std::string(buffer)]() {
@@ -382,7 +397,8 @@ void Application::CheckAssetsVersion() {
         vTaskDelay(pdMS_TO_TICKS(1000));
 
         if (!success) {
-            Alert(Lang::Strings::ERROR, Lang::Strings::DOWNLOAD_ASSETS_FAILED, "circle_xmark", Lang::Sounds::OGG_EXCLAMATION);
+            Alert(Lang::Strings::ERROR, Lang::Strings::DOWNLOAD_ASSETS_FAILED, "circle_xmark",
+                  Lang::Sounds::OGG_EXCLAMATION);
             vTaskDelay(pdMS_TO_TICKS(2000));
             SetDeviceState(kDeviceStateActivating);
             return;
@@ -414,12 +430,15 @@ void Application::CheckNewVersion() {
             }
 
             char error_message[128];
-            snprintf(error_message, sizeof(error_message), "code=%d, url=%s", err, ota_->GetCheckVersionUrl().c_str());
+            snprintf(error_message, sizeof(error_message), "code=%d, url=%s", err,
+                     ota_->GetCheckVersionUrl().c_str());
             char buffer[256];
-            snprintf(buffer, sizeof(buffer), Lang::Strings::CHECK_NEW_VERSION_FAILED, retry_delay, error_message);
+            snprintf(buffer, sizeof(buffer), Lang::Strings::CHECK_NEW_VERSION_FAILED, retry_delay,
+                     error_message);
             Alert(Lang::Strings::ERROR, buffer, "cloud_slash", Lang::Sounds::OGG_EXCLAMATION);
 
-            ESP_LOGW(TAG, "Check new version failed, retry in %d seconds (%d/%d)", retry_delay, retry_count, MAX_RETRY);
+            ESP_LOGW(TAG, "Check new version failed, retry in %d seconds (%d/%d)", retry_delay,
+                     retry_count, MAX_RETRY);
             for (int i = 0; i < retry_delay; i++) {
                 vTaskDelay(pdMS_TO_TICKS(1000));
                 if (GetDeviceState() == kDeviceStateIdle) {
@@ -477,6 +496,9 @@ void Application::InitializeProtocol() {
 
     display->SetStatus(Lang::Strings::LOADING_PROTOCOL);
 
+#if CONFIG_USE_DIRECT_WEBSOCKET
+    protocol_ = std::make_unique<WebsocketProtocol>();
+#else
     if (ota_->HasMqttConfig()) {
         protocol_ = std::make_unique<MqttProtocol>();
     } else if (ota_->HasWebsocketConfig()) {
@@ -485,10 +507,9 @@ void Application::InitializeProtocol() {
         ESP_LOGW(TAG, "No protocol specified in the OTA config, using MQTT");
         protocol_ = std::make_unique<MqttProtocol>();
     }
+#endif
 
-    protocol_->OnConnected([this]() {
+    protocol_->OnConnected([this]() { DismissAlert(); });
-        DismissAlert();
-    });
 
     protocol_->OnNetworkError([this](const std::string& message) {
         last_error_message_ = message;
@@ -496,7 +517,7 @@ void Application::InitializeProtocol() {
     });
 
     protocol_->OnIncomingAudio([this](std::unique_ptr<AudioStreamPacket> packet) {
-        if (GetDeviceState() == kDeviceStateSpeaking) {
+        if (accepting_tts_audio_.load() || GetDeviceState() == kDeviceStateSpeaking) {
             audio_service_.PushPacketToDecodeQueue(std::move(packet));
         }
     });
@@ -504,14 +525,20 @@ void Application::InitializeProtocol() {
     protocol_->OnAudioChannelOpened([this, codec, &board]() {
         board.SetPowerSaveLevel(PowerSaveLevel::PERFORMANCE);
         if (protocol_->server_sample_rate() != codec->output_sample_rate()) {
-            ESP_LOGW(TAG, "Server sample rate %d does not match device output sample rate %d, resampling may cause distortion",
+            ESP_LOGW(TAG,
+                     "Server sample rate %d does not match device output sample rate %d, "
+                     "resampling may cause distortion",
                      protocol_->server_sample_rate(), codec->output_sample_rate());
         }
     });
 
     protocol_->OnAudioChannelClosed([this, &board]() {
         board.SetPowerSaveLevel(PowerSaveLevel::LOW_POWER);
+        accepting_tts_audio_.store(false);
         Schedule([this]() {
+            if (GetDeviceState() == kDeviceStateConnecting) {
+                return;
+            }
             auto display = Board::GetInstance().GetDisplay();
             display->SetChatMessage("system", "");
             SetDeviceState(kDeviceStateIdle);
@@ -523,14 +550,20 @@ void Application::InitializeProtocol() {
         auto type = cJSON_GetObjectItem(root, "type");
         if (strcmp(type->valuestring, "tts") == 0) {
             auto state = cJSON_GetObjectItem(root, "state");
-            if (strcmp(state->valuestring, "start") == 0) {
+            if (strcmp(state->valuestring, "thinking") == 0) {
+                Schedule([this]() { SetDeviceState(kDeviceStateThinking); });
+            } else if (strcmp(state->valuestring, "start") == 0) {
+                audio_service_.ResetDecoder();
+                accepting_tts_audio_.store(true);
                 Schedule([this]() {
                     aborted_ = false;
                     SetDeviceState(kDeviceStateSpeaking);
                 });
             } else if (strcmp(state->valuestring, "stop") == 0) {
+                accepting_tts_audio_.store(false);
                 Schedule([this]() {
-                    if (GetDeviceState() == kDeviceStateSpeaking) {
+                    auto state = GetDeviceState();
+                    if (state == kDeviceStateSpeaking || state == kDeviceStateThinking) {
                         if (listening_mode_ == kListeningModeManualStop) {
                             SetDeviceState(kDeviceStateIdle);
                         } else {
@@ -573,9 +606,7 @@ void Application::InitializeProtocol() {
                 ESP_LOGI(TAG, "System command: %s", command->valuestring);
                 if (strcmp(command->valuestring, "reboot") == 0) {
                     // Do a reboot if user requests a OTA update
-                    Schedule([this]() {
+                    Schedule([this]() { Reboot(); });
-                        Reboot();
-                    });
                 } else {
                     ESP_LOGW(TAG, "Unknown system command: %s", command->valuestring);
                 }
@@ -585,7 +616,8 @@ void Application::InitializeProtocol() {
             auto message = cJSON_GetObjectItem(root, "message");
             auto emotion = cJSON_GetObjectItem(root, "emotion");
             if (cJSON_IsString(status) && cJSON_IsString(message) && cJSON_IsString(emotion)) {
-                Alert(status->valuestring, message->valuestring, emotion->valuestring, Lang::Sounds::OGG_VIBRATION);
+                Alert(status->valuestring, message->valuestring, emotion->valuestring,
+                      Lang::Sounds::OGG_VIBRATION);
             } else {
                 ESP_LOGW(TAG, "Alert command requires status, message and emotion");
             }
@@ -594,7 +626,8 @@ void Application::InitializeProtocol() {
             auto payload = cJSON_GetObjectItem(root, "payload");
             ESP_LOGI(TAG, "Received custom message: %s", cJSON_PrintUnformatted(root));
             if (cJSON_IsObject(payload)) {
-                Schedule([this, display, payload_str = std::string(cJSON_PrintUnformatted(payload))]() {
+                Schedule(
+                    [this, display, payload_str = std::string(cJSON_PrintUnformatted(payload))]() {
                         display->SetChatMessage("system", payload_str.c_str());
                     });
             } else {
@@ -614,18 +647,12 @@ void Application::ShowActivationCode(const std::string& code, const std::string&
         char digit;
         const std::string_view& sound;
     };
-    static const std::array<digit_sound, 10> digit_sounds{{
+    static const std::array<digit_sound, 10> digit_sounds{
-        digit_sound{'0', Lang::Sounds::OGG_0},
+        {digit_sound{'0', Lang::Sounds::OGG_0}, digit_sound{'1', Lang::Sounds::OGG_1},
-        digit_sound{'1', Lang::Sounds::OGG_1}, 
+         digit_sound{'2', Lang::Sounds::OGG_2}, digit_sound{'3', Lang::Sounds::OGG_3},
-        digit_sound{'2', Lang::Sounds::OGG_2},
+         digit_sound{'4', Lang::Sounds::OGG_4}, digit_sound{'5', Lang::Sounds::OGG_5},
-        digit_sound{'3', Lang::Sounds::OGG_3},
+         digit_sound{'6', Lang::Sounds::OGG_6}, digit_sound{'7', Lang::Sounds::OGG_7},
-        digit_sound{'4', Lang::Sounds::OGG_4},
+         digit_sound{'8', Lang::Sounds::OGG_8}, digit_sound{'9', Lang::Sounds::OGG_9}}};
-        digit_sound{'5', Lang::Sounds::OGG_5},
-        digit_sound{'6', Lang::Sounds::OGG_6},
-        digit_sound{'7', Lang::Sounds::OGG_7},
-        digit_sound{'8', Lang::Sounds::OGG_8},
-        digit_sound{'9', Lang::Sounds::OGG_9}
-    }};
 
     // This sentence uses 9KB of SRAM, so we need to wait for it to finish
     Alert(Lang::Strings::ACTIVATION, message.c_str(), "link", Lang::Sounds::OGG_ACTIVATION);
@@ -639,7 +666,8 @@ void Application::ShowActivationCode(const std::string& code, const std::string&
     }
 }
 
-void Application::Alert(const char* status, const char* message, const char* emotion, const std::string_view& sound) {
+void Application::Alert(const char* status, const char* message, const char* emotion,
+                        const std::string_view& sound) {
     ESP_LOGW(TAG, "Alert [%s] %s: %s", emotion, status, message);
     auto display = Board::GetInstance().GetDisplay();
     display->SetStatus(status);
@@ -659,17 +687,40 @@ void Application::DismissAlert() {
     }
 }
 
-void Application::ToggleChatState() {
+void Application::ToggleChatState() { ToggleChatStateForMode(kChatAgentModeNormal, false); }
+
+void Application::ToggleChatStateWithVision() {
+    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);
 }
 
-void Application::StopListening() {
+void Application::StopListening() { xEventGroupSetBits(event_group_, MAIN_EVENT_STOP_LISTENING); }
-    xEventGroupSetBits(event_group_, MAIN_EVENT_STOP_LISTENING);
-}
 
 void Application::HandleToggleChatEvent() {
     auto state = GetDeviceState();
@@ -694,17 +745,22 @@ 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]() {
+            Schedule([this, mode]() { ContinueOpenAudioChannel(mode); });
-                ContinueOpenAudioChannel(mode);
-            });
             return;
         }
         SetListeningMode(mode);
-    } else if (state == kDeviceStateSpeaking) {
+    } else if (state == kDeviceStateSpeaking || state == kDeviceStateThinking) {
         AbortSpeaking(kAbortReasonNone);
+        SetListeningMode(GetDefaultListeningMode());
     } else if (state == kDeviceStateListening) {
         protocol_->CloseAudioChannel();
     }
@@ -726,6 +782,8 @@ 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);
 }
 
@@ -750,13 +808,11 @@ void Application::HandleStartListeningEvent() {
         if (!protocol_->IsAudioChannelOpened()) {
             SetDeviceState(kDeviceStateConnecting);
             // Schedule to let the state change be processed first (UI update)
-            Schedule([this]() {
+            Schedule([this]() { ContinueOpenAudioChannel(kListeningModeManualStop); });
-                ContinueOpenAudioChannel(kListeningModeManualStop);
-            });
             return;
         }
         SetListeningMode(kListeningModeManualStop);
-    } else if (state == kDeviceStateSpeaking) {
+    } else if (state == kDeviceStateSpeaking || state == kDeviceStateThinking) {
         AbortSpeaking(kAbortReasonNone);
         SetListeningMode(kListeningModeManualStop);
     }
@@ -794,17 +850,14 @@ void Application::HandleWakeWordDetectedEvent() {
             SetDeviceState(kDeviceStateConnecting);
             // Schedule to let the state change be processed first (UI update),
             // then continue with OpenAudioChannel which may block for ~1 second
-            Schedule([this, wake_word]() {
+            Schedule([this, wake_word]() { ContinueWakeWordInvoke(wake_word); });
-                ContinueWakeWordInvoke(wake_word);
-            });
             return;
         }
         // Channel already opened, continue directly
         ContinueWakeWordInvoke(wake_word);
-    } else if (state == kDeviceStateSpeaking || state == kDeviceStateListening) {
+    } else if (state == kDeviceStateSpeaking || state == kDeviceStateThinking ||
+               state == kDeviceStateListening) {
         AbortSpeaking(kAbortReasonWakeWordDetected);
-        // Clear send queue to avoid sending residues to server
-        while (audio_service_.PopPacketFromSendQueue());
 
         if (state == kDeviceStateListening) {
             protocol_->SendStartListening(GetDefaultListeningMode());
@@ -869,6 +922,7 @@ 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)
@@ -881,21 +935,19 @@ 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");
 
-            // Make sure the audio processor is running
+            // Re-entering listening after an interrupt must restart the capture path even if the
-            if (play_popup_on_listening_ || !audio_service_.IsAudioProcessorRunning()) {
+            // processor task is still marked running, otherwise realtime mode can show Listening
-                // For auto mode, wait for playback queue to be empty before enabling voice processing
+            // while no fresh mic frames are sent.
-                // This prevents audio truncation when STOP arrives late due to network jitter
             if (listening_mode_ == kListeningModeAutoStop) {
                 audio_service_.WaitForPlaybackQueueEmpty();
             }
-                
-                // Send the start listening command
             protocol_->SendStartListening(listening_mode_);
             audio_service_.EnableVoiceProcessing(true);
-            }
 
 #ifdef CONFIG_WAKE_WORD_DETECTION_IN_LISTENING
             // Enable wake word detection in listening mode (configured via Kconfig)
@@ -911,6 +963,16 @@ void Application::HandleStateChangedEvent() {
                 audio_service_.PlaySound(Lang::Sounds::OGG_POPUP);
             }
             break;
+        case kDeviceStateThinking:
+            vad_speaking_.store(false);
+            display->SetStatus(Lang::Strings::THINKING);
+            display->SetEmotion("thinking");
+
+            if (listening_mode_ != kListeningModeRealtime) {
+                audio_service_.EnableVoiceProcessing(false);
+                audio_service_.EnableWakeWordDetection(audio_service_.IsAfeWakeWord());
+            }
+            break;
         case kDeviceStateSpeaking:
             display->SetStatus(Lang::Strings::SPEAKING);
 
@@ -919,7 +981,9 @@ void Application::HandleStateChangedEvent() {
                 // Only AFE wake word can be detected in speaking mode
                 audio_service_.EnableWakeWordDetection(audio_service_.IsAfeWakeWord());
             }
+            if (!accepting_tts_audio_.load()) {
                 audio_service_.ResetDecoder();
+            }
             break;
         case kDeviceStateWifiConfiguring:
             audio_service_.EnableVoiceProcessing(false);
@@ -931,6 +995,27 @@ void Application::HandleStateChangedEvent() {
     }
 }
 
+bool Application::SendCurrentVisionFrame() {
+    if (!protocol_ || !protocol_->IsAudioChannelOpened()) {
+        return false;
+    }
+
+    auto camera = Board::GetInstance().GetCamera();
+    if (camera == nullptr) {
+        return false;
+    }
+
+    std::string jpeg_data;
+    if (!camera->CaptureToJpeg(jpeg_data, true)) {
+        ESP_LOGW(TAG, "Failed to capture vision frame");
+        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) {
     {
         std::lock_guard<std::mutex> lock(mutex_);
@@ -942,6 +1027,8 @@ void Application::Schedule(std::function<void()>&& callback) {
 void Application::AbortSpeaking(AbortReason reason) {
     ESP_LOGI(TAG, "Abort speaking");
     aborted_ = true;
+    accepting_tts_audio_.store(false);
+    audio_service_.ResetDecoder();
     if (protocol_) {
         protocol_->SendAbortSpeaking(reason);
     }
@@ -949,6 +1036,8 @@ 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);
 }
 
@@ -983,7 +1072,8 @@ bool Application::UpgradeFirmware(const std::string& url, const std::string& ver
     }
     ESP_LOGI(TAG, "Starting firmware upgrade from URL: %s", upgrade_url.c_str());
 
-    Alert(Lang::Strings::OTA_UPGRADE, Lang::Strings::UPGRADING, "download", Lang::Sounds::OGG_UPGRADE);
+    Alert(Lang::Strings::OTA_UPGRADE, Lang::Strings::UPGRADING, "download",
+          Lang::Sounds::OGG_UPGRADE);
     vTaskDelay(pdMS_TO_TICKS(3000));
 
     SetDeviceState(kDeviceStateUpgrading);
@@ -1005,10 +1095,12 @@ bool Application::UpgradeFirmware(const std::string& url, const std::string& ver
 
     if (!upgrade_success) {
         // Upgrade failed, restart audio service and continue running
-        ESP_LOGE(TAG, "Firmware upgrade failed, restarting audio service and continuing operation...");
+        ESP_LOGE(TAG,
+                 "Firmware upgrade failed, restarting audio service and continuing operation...");
         audio_service_.Start();                              // Restart audio service
         board.SetPowerSaveLevel(PowerSaveLevel::LOW_POWER);  // Restore power save level
-        Alert(Lang::Strings::ERROR, Lang::Strings::UPGRADE_FAILED, "circle_xmark", Lang::Sounds::OGG_EXCLAMATION);
+        Alert(Lang::Strings::ERROR, Lang::Strings::UPGRADE_FAILED, "circle_xmark",
+              Lang::Sounds::OGG_EXCLAMATION);
         vTaskDelay(pdMS_TO_TICKS(3000));
         return false;
     } else {
@@ -1034,17 +1126,13 @@ void Application::WakeWordInvoke(const std::string& wake_word) {
         if (!protocol_->IsAudioChannelOpened()) {
             SetDeviceState(kDeviceStateConnecting);
             // Schedule to let the state change be processed first (UI update)
-            Schedule([this, wake_word]() {
+            Schedule([this, wake_word]() { ContinueWakeWordInvoke(wake_word); });
-                ContinueWakeWordInvoke(wake_word);
-            });
             return;
         }
         // Channel already opened, continue directly
         ContinueWakeWordInvoke(wake_word);
-    } else if (state == kDeviceStateSpeaking) {
+    } else if (state == kDeviceStateSpeaking || state == kDeviceStateThinking) {
-        Schedule([this]() {
+        Schedule([this]() { AbortSpeaking(kAbortReasonNone); });
-            AbortSpeaking(kAbortReasonNone);
-        });
     } else if (state == kDeviceStateListening) {
         Schedule([this]() {
             if (protocol_) {
@@ -1077,7 +1165,7 @@ void Application::RegisterMcpBroadcastCallback(std::function<void(const std::str
 
 void Application::SendMcpMessage(const std::string& payload) {
     // Always schedule to run in main task for thread safety
-    Schedule([this, payload](){ 
+    Schedule([this, payload]() {
         if (protocol_) {
             protocol_->SendMcpMessage(payload);
         }
@@ -1114,9 +1202,7 @@ void Application::SetAecMode(AecMode mode) {
     });
 }
 
-void Application::PlaySound(const std::string_view& sound) {
+void Application::PlaySound(const std::string_view& sound) { audio_service_.PlaySound(sound); }
-    audio_service_.PlaySound(sound);
-}
 
 void Application::ResetProtocol() {
     Schedule([this]() {
@@ -1128,4 +1214,3 @@ void Application::ResetProtocol() {
         protocol_.reset();
     });
 }
-

main/application.h

@@ -11,6 +11,7 @@
 #include <deque>
 #include <memory>
 #include <functional>
+#include <atomic>
 
 #include "protocol.h"
 #include "ota.h"
@@ -40,6 +41,11 @@ enum AecMode {
     kAecOnServerSide,
 };
 
+enum ChatAgentMode {
+    kChatAgentModeNormal,
+    kChatAgentModeBeaver,
+};
+
 class Application {
 public:
     static Application& GetInstance() {
@@ -91,6 +97,12 @@ public:
      * Sends MAIN_EVENT_TOGGLE_CHAT to be handled in Run()
      */
     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)
@@ -144,6 +156,13 @@ 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;
+    std::atomic<bool> accepting_tts_audio_ = false;
     int clock_ticks_ = 0;
     TaskHandle_t activation_task_handle_ = nullptr;
 
@@ -159,6 +178,7 @@ private:
     void HandleWakeWordDetectedEvent();
     void ContinueOpenAudioChannel(ListeningMode mode);
     void ContinueWakeWordInvoke(const std::string& wake_word);
+    bool SendCurrentVisionFrame();
 
     // Activation task (runs in background)
     void ActivationTask();

main/assets/locales/en-US/language.json

@@ -26,6 +26,7 @@
         "CONNECTION_SUCCESSFUL": "Connection Successful",
         "CONNECTED_TO": "Connected to ",
         "LISTENING": "Listening...",
+        "THINKING": "Thinking...",
         "SPEAKING": "Speaking...",
         "SERVER_NOT_FOUND": "Looking for available service",
         "SERVER_NOT_CONNECTED": "Unable to connect to service, please try again later",

main/assets/locales/zh-CN/language.json

@@ -23,6 +23,7 @@
         "CONNECTING": "连接中...",
         "CONNECTED_TO": "已连接 ",
         "LISTENING": "聆听中...",
+        "THINKING": "思考中...",
         "SPEAKING": "说话中...",
         "SERVER_NOT_FOUND": "正在寻找可用服务",
         "SERVER_NOT_CONNECTED": "无法连接服务，请稍后再试",

main/audio/audio_service.cc

@@ -579,6 +579,7 @@ void AudioService::EnableWakeWordDetection(bool enable) {
 void AudioService::EnableVoiceProcessing(bool enable) {
     ESP_LOGD(TAG, "%s voice processing", enable ? "Enabling" : "Disabling");
     if (enable) {
+        bool was_running = IsAudioProcessorRunning();
         if (!audio_processor_initialized_) {
             audio_processor_->Initialize(codec_, OPUS_FRAME_DURATION_MS, models_list_);
             audio_processor_initialized_ = true;
@@ -586,7 +587,7 @@ void AudioService::EnableVoiceProcessing(bool enable) {
 
         /* We should make sure no audio is playing */
         ResetDecoder();
-        audio_input_need_warmup_ = true;
+        audio_input_need_warmup_ = !was_running;
         // Reset input resampler to clear cached data from previous mode (e.g. WakeWord)
         // This prevents buffer overflow when switching between different feed sizes
         {

main/background_capture_service.cc

@@ -0,0 +1,177 @@
+#include "background_capture_service.h"
+
+#include "board.h"
+#include "camera.h"
+
+#include <algorithm>
+#include <esp_heap_caps.h>
+#include <esp_log.h>
+
+#define TAG "BgCapture"
+
+BackgroundCaptureService::BackgroundCaptureService() = default;
+
+BackgroundCaptureService::~BackgroundCaptureService() {
+    Stop();
+}
+
+void BackgroundCaptureService::Start() {
+#if CONFIG_BACKGROUND_CAPTURE_ENABLE
+    if (running_.exchange(true)) {
+        return;
+    }
+
+    auto result = xTaskCreate(
+        &BackgroundCaptureService::TaskEntry,
+        "bg_capture",
+        CONFIG_BACKGROUND_CAPTURE_TASK_STACK_SIZE,
+        this,
+        CONFIG_BACKGROUND_CAPTURE_TASK_PRIORITY,
+        &task_handle_);
+    if (result != pdPASS) {
+        running_.store(false);
+        task_handle_ = nullptr;
+        ESP_LOGE(TAG, "Failed to create background capture task");
+    }
+#endif
+}
+
+void BackgroundCaptureService::Stop() {
+#if CONFIG_BACKGROUND_CAPTURE_ENABLE
+    if (!running_.exchange(false)) {
+        return;
+    }
+
+    while (task_handle_ != nullptr) {
+        vTaskDelay(pdMS_TO_TICKS(20));
+    }
+#endif
+}
+
+void BackgroundCaptureService::TaskEntry(void* arg) {
+#if CONFIG_BACKGROUND_CAPTURE_ENABLE
+    auto* service = static_cast<BackgroundCaptureService*>(arg);
+    service->Run();
+    service->task_handle_ = nullptr;
+#else
+    (void)arg;
+#endif
+    vTaskDelete(nullptr);
+}
+
+void BackgroundCaptureService::Run() {
+#if CONFIG_BACKGROUND_CAPTURE_ENABLE
+    ESP_LOGI(TAG, "Background capture task started");
+
+    while (running_.load()) {
+        if (!CaptureAndSendFrame()) {
+            consecutive_failures_++;
+            auto delay_ms = GetFailureDelayMs();
+            ESP_LOGW(TAG, "Background capture retry in %u ms, failures=%u",
+                     delay_ms, consecutive_failures_);
+            vTaskDelay(pdMS_TO_TICKS(delay_ms));
+            continue;
+        }
+
+        consecutive_failures_ = 0;
+        vTaskDelay(pdMS_TO_TICKS(CONFIG_BACKGROUND_CAPTURE_FRAME_INTERVAL_MS));
+    }
+
+    ESP_LOGI(TAG, "Background capture task stopped");
+#endif
+}
+
+bool BackgroundCaptureService::CaptureAndSendFrame() {
+#if CONFIG_BACKGROUND_CAPTURE_ENABLE
+    const size_t free_internal_heap = heap_caps_get_free_size(MALLOC_CAP_INTERNAL);
+    if (free_internal_heap < CONFIG_BACKGROUND_CAPTURE_MIN_FREE_INTERNAL_HEAP) {
+        ESP_LOGW(TAG, "Skip background capture, low internal heap: free=%u threshold=%u",
+                 static_cast<unsigned>(free_internal_heap),
+                 static_cast<unsigned>(CONFIG_BACKGROUND_CAPTURE_MIN_FREE_INTERNAL_HEAP));
+        return false;
+    }
+
+    auto camera = Board::GetInstance().GetCamera();
+    if (camera == nullptr) {
+        ESP_LOGW(TAG, "No camera available for background capture");
+        return false;
+    }
+
+    std::string jpeg_data;
+    if (!camera->CaptureToJpeg(jpeg_data, false)) {
+        ESP_LOGW(TAG, "Failed to capture background frame");
+        return false;
+    }
+
+    if (jpeg_data.empty()) {
+        ESP_LOGW(TAG, "Captured empty background frame");
+        return false;
+    }
+
+    return UploadJpegFrame(jpeg_data);
+#else
+    return false;
+#endif
+}
+
+uint32_t BackgroundCaptureService::GetFailureDelayMs() const {
+#if CONFIG_BACKGROUND_CAPTURE_ENABLE
+    const uint32_t base_delay_ms = CONFIG_BACKGROUND_CAPTURE_RETRY_INTERVAL_MS;
+    const uint32_t max_delay_ms = CONFIG_BACKGROUND_CAPTURE_MAX_BACKOFF_MS;
+    const uint32_t shift = std::min<uint32_t>(consecutive_failures_ - 1, 4);
+    return std::min<uint32_t>(base_delay_ms << shift, max_delay_ms);
+#else
+    return 0;
+#endif
+}
+
+bool BackgroundCaptureService::UploadJpegFrame(const std::string& jpeg_data) {
+#if CONFIG_BACKGROUND_CAPTURE_ENABLE
+    const std::string url = CONFIG_BACKGROUND_CAPTURE_UPLOAD_URL;
+    if (url.empty()) {
+        ESP_LOGI(TAG, "Captured background frame: %u bytes", static_cast<unsigned>(jpeg_data.size()));
+        return true;
+    }
+
+    auto network = Board::GetInstance().GetNetwork();
+    if (network == nullptr) {
+        ESP_LOGW(TAG, "No network available for background upload");
+        return false;
+    }
+
+    const std::string boundary = "----XIAOZHI_BACKGROUND_CAPTURE_BOUNDARY";
+    auto http = network->CreateHttp(3);
+    http->SetHeader("Content-Type", "multipart/form-data; boundary=" + boundary);
+
+    if (!http->Open("POST", url)) {
+        ESP_LOGW(TAG, "Failed to open background upload URL: %s", url.c_str());
+        return false;
+    }
+
+    std::string file_header;
+    file_header += "--" + boundary + "\r\n";
+    file_header += "Content-Disposition: form-data; name=\"file\"; filename=\"frame.jpg\"\r\n";
+    file_header += "Content-Type: image/jpeg\r\n\r\n";
+    http->Write(file_header.c_str(), file_header.size());
+    http->Write(jpeg_data.data(), jpeg_data.size());
+
+    std::string footer;
+    footer += "\r\n--" + boundary + "--\r\n";
+    http->Write(footer.c_str(), footer.size());
+    http->Write("", 0);
+
+    const int status_code = http->GetStatusCode();
+    http->Close();
+
+    if (status_code < 200 || status_code >= 300) {
+        ESP_LOGW(TAG, "Background upload failed, status=%d", status_code);
+        return false;
+    }
+
+    ESP_LOGI(TAG, "Uploaded background frame: %u bytes", static_cast<unsigned>(jpeg_data.size()));
+    return true;
+#else
+    (void)jpeg_data;
+    return false;
+#endif
+}

main/background_capture_service.h

@@ -0,0 +1,32 @@
+#ifndef BACKGROUND_CAPTURE_SERVICE_H
+#define BACKGROUND_CAPTURE_SERVICE_H
+
+#include <freertos/FreeRTOS.h>
+#include <freertos/task.h>
+
+#include <atomic>
+#include <cstdint>
+#include <string>
+
+class BackgroundCaptureService {
+public:
+    BackgroundCaptureService();
+    ~BackgroundCaptureService();
+
+    void Start();
+    void Stop();
+    bool IsRunning() const { return running_.load(); }
+
+private:
+    TaskHandle_t task_handle_ = nullptr;
+    std::atomic<bool> running_ = false;
+    uint32_t consecutive_failures_ = 0;
+
+    static void TaskEntry(void* arg);
+    void Run();
+    bool CaptureAndSendFrame();
+    bool UploadJpegFrame(const std::string& jpeg_data);
+    uint32_t GetFailureDelayMs() const;
+};
+
+#endif // BACKGROUND_CAPTURE_SERVICE_H

main/boards/atoms3r-echo-pyramid/atoms3r_echo_pyramid.cc

@@ -214,6 +214,9 @@ public:
             case kDeviceStateSpeaking:
                 ctrl_->SetStatusColor(64, 0, 0);   // red
                 break;
+            case kDeviceStateThinking:
+                ctrl_->SetStatusColor(0, 0, 64);   // blue
+                break;
             default:
                 ctrl_->SetStatusColor(0, 0, 64);   // blue
                 break;

main/boards/common/camera.h

@@ -7,6 +7,8 @@ class Camera {
 public:
     virtual void SetExplainUrl(const std::string& url, const std::string& token) = 0;
     virtual bool Capture() = 0;
+    virtual bool CaptureBackground() { return Capture(); }
+    virtual bool CaptureToJpeg(std::string& jpeg_data, bool show_preview = false) { return false; }
     virtual bool SetHMirror(bool enabled) = 0;
     virtual bool SetVFlip(bool enabled) = 0;
     virtual bool SetSwapBytes(bool enabled) { return false; }  // Optional, default no-op

main/boards/common/esp_video.cc

@@ -24,6 +24,7 @@
 #include "lvgl_display.h"
 #include "mcp_server.h"
 #include "system_info.h"
+#include "esp_timer.h"
 
 #ifdef CONFIG_XIAOZHI_ENABLE_CAMERA_DEBUG_MODE
 #undef LOG_LOCAL_LEVEL
@@ -55,6 +56,7 @@
 
 
 #define TAG "EspVideo"
+#define FOREGROUND_CAPTURE_PROTECTION_US (10 * 1000 * 1000)
 
 #if defined(CONFIG_CAMERA_SENSOR_SWAP_PIXEL_BYTE_ORDER) || defined(CONFIG_XIAOZHI_ENABLE_CAMERA_ENDIANNESS_SWAP)
 #warning \
@@ -381,11 +383,47 @@ EspVideo::~EspVideo() {
 }
 
 void EspVideo::SetExplainUrl(const std::string& url, const std::string& token) {
+    std::lock_guard<std::mutex> lock(frame_mutex_);
     explain_url_ = url;
     explain_token_ = token;
 }
 
 bool EspVideo::Capture() {
+    return CaptureFrame(true);
+}
+
+bool EspVideo::CaptureBackground() {
+    return CaptureFrame(false);
+}
+
+bool EspVideo::CaptureToJpeg(std::string& jpeg_data, bool show_preview) {
+    jpeg_data.clear();
+    if (!CaptureFrame(show_preview)) {
+        return false;
+    }
+
+    std::lock_guard<std::mutex> lock(frame_mutex_);
+    if (frame_.data == nullptr || frame_.len == 0) {
+        return false;
+    }
+
+    uint16_t w = frame_.width ? frame_.width : 320;
+    uint16_t h = frame_.height ? frame_.height : 240;
+    return image_to_jpeg_cb(
+        frame_.data, frame_.len, w, h, frame_.format, 60,
+        [](void* arg, size_t index, const void* data, size_t len) -> size_t {
+            auto jpeg_data = static_cast<std::string*>(arg);
+            if (data != nullptr && len > 0) {
+                jpeg_data->append(static_cast<const char*>(data), len);
+            }
+            return len;
+        },
+        &jpeg_data);
+}
+
+bool EspVideo::CaptureFrame(bool show_preview) {
+    std::lock_guard<std::mutex> lock(frame_mutex_);
+
     if (encoder_thread_.joinable()) {
         encoder_thread_.join();
     }
@@ -394,6 +432,10 @@ bool EspVideo::Capture() {
         return false;
     }
 
+    if (!show_preview && esp_timer_get_time() < foreground_capture_protected_until_us_) {
+        return true;
+    }
+
     for (int i = 0; i < 3; i++) {
         struct v4l2_buffer buf = {};
         buf.type = V4L2_BUF_TYPE_VIDEO_CAPTURE;
@@ -729,6 +771,11 @@ bool EspVideo::Capture() {
         }
     }
 
+    if (show_preview) {
+        foreground_capture_protected_until_us_ = esp_timer_get_time() + FOREGROUND_CAPTURE_PROTECTION_US;
+    }
+
+    if (show_preview) {
         // 显示预览图片
         auto display = dynamic_cast<LvglDisplay*>(Board::GetInstance().GetDisplay());
         if (display != nullptr) {
@@ -837,6 +884,7 @@ bool EspVideo::Capture() {
         auto image = std::make_unique<LvglAllocatedImage>(data, lvgl_image_size, w, h, stride, color_format);
         display->SetPreviewImage(std::move(image));
         }
+    }
     return true;
 }
 
@@ -898,10 +946,16 @@ bool EspVideo::SetVFlip(bool enabled) {
  * @warning 如果摄像头缓冲区为空或网络连接失败，将返回错误信息
  */
 std::string EspVideo::Explain(const std::string& question) {
+    std::lock_guard<std::mutex> lock(frame_mutex_);
+
     if (explain_url_.empty()) {
         throw std::runtime_error("Image explain URL or token is not set");
     }
 
+    if (frame_.data == nullptr || frame_.len == 0) {
+        throw std::runtime_error("No camera frame captured");
+    }
+
     // 创建局部的 JPEG 队列, 40 entries is about to store 512 * 40 = 20480 bytes of JPEG data
     QueueHandle_t jpeg_queue = xQueueCreate(40, sizeof(JpegChunk));
     if (jpeg_queue == nullptr) {

main/boards/common/esp_video.h

@@ -5,6 +5,8 @@
 #include <thread>
 #include <memory>
 #include <vector>
+#include <mutex>
+#include <cstdint>
 
 #include <freertos/FreeRTOS.h>
 #include <freertos/queue.h>
@@ -39,6 +41,10 @@ private:
     std::string explain_url_;
     std::string explain_token_;
     std::thread encoder_thread_;
+    std::mutex frame_mutex_;
+    int64_t foreground_capture_protected_until_us_ = 0;
+
+    bool CaptureFrame(bool show_preview);
 
 public:
     EspVideo(const esp_video_init_config_t& config);
@@ -46,6 +52,8 @@ public:
 
     virtual void SetExplainUrl(const std::string& url, const std::string& token);
     virtual bool Capture();
+    virtual bool CaptureBackground() override;
+    virtual bool CaptureToJpeg(std::string& jpeg_data, bool show_preview = false) override;
     // 翻转控制函数
     virtual bool SetHMirror(bool enabled) override;
     virtual bool SetVFlip(bool enabled) override;

main/boards/common/wifi_board.cc

@@ -203,7 +203,10 @@ void WifiBoard::EnterWifiConfigMode() {
     auto& app = Application::GetInstance();
     auto state = app.GetDeviceState();
 
-    if (state == kDeviceStateSpeaking || state == kDeviceStateListening || state == kDeviceStateIdle) {
+    if (state == kDeviceStateSpeaking ||
+            state == kDeviceStateThinking ||
+            state == kDeviceStateListening ||
+            state == kDeviceStateIdle) {
         // Reset protocol (close audio channel, reset protocol)
         Application::GetInstance().ResetProtocol();
 

main/boards/electron-bot/electron_emoji_display.cc

@@ -85,6 +85,13 @@ void ElectronEmojiDisplay::SetStatus(const char* status) {
         lv_obj_add_flag(network_label_, LV_OBJ_FLAG_HIDDEN);
         lv_obj_add_flag(battery_label_, LV_OBJ_FLAG_HIDDEN);
         return;
+    } else if (strcmp(status, Lang::Strings::THINKING) == 0) {
+        lv_obj_set_style_text_font(status_label_, text_font, 0);
+        lv_label_set_text(status_label_, status);
+        lv_obj_clear_flag(status_label_, LV_OBJ_FLAG_HIDDEN);
+        lv_obj_add_flag(network_label_, LV_OBJ_FLAG_HIDDEN);
+        lv_obj_add_flag(battery_label_, LV_OBJ_FLAG_HIDDEN);
+        return;
     } else if (strcmp(status, Lang::Strings::CONNECTING) == 0) {
         lv_obj_set_style_text_font(status_label_, &OTTO_ICON_FONT, 0);
         lv_label_set_text(status_label_, "\xEF\x83\x81");  // U+F0c1 连接图标

main/boards/esp-hi/emoji_display.cc

@@ -155,6 +155,8 @@ void EmojiWidget::SetStatus(const char* status)
     if (player_) {
         if (strcmp(status, Lang::Strings::LISTENING) == 0) {
             player_->StartPlayer("asking", true, 15);
+        } else if (strcmp(status, Lang::Strings::THINKING) == 0) {
+            player_->StartPlayer("thinking", true, 15);
         } else if (strcmp(status, Lang::Strings::STANDBY) == 0) {
             player_->StartPlayer("wake", true, 15);
         }

main/boards/jiuchuan-s3/jiuchuan_dev_board.cc

@@ -231,9 +231,9 @@ private:
                 // 如果当前是聆听状态，切换到待命状态
                 ESP_LOGI(TAG, "从聆听状态切换到待命状态");
                 app.ToggleChatState(); // 切换到待命状态
-            } else if (current_state == kDeviceStateSpeaking) {
+            } else if (current_state == kDeviceStateSpeaking || current_state == kDeviceStateThinking) {
-                // 如果当前是说话状态，终止说话并切换到待命状态
+                // 如果当前是说话或思考状态，终止并切换到待命状态
-                ESP_LOGI(TAG, "从说话状态切换到待命状态");
+                ESP_LOGI(TAG, "从说话/思考状态切换到待命状态");
                 app.ToggleChatState(); // 终止说话
             } else {
                 // 其他状态下只唤醒设备

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

@@ -1,21 +1,23 @@
-#include "wifi_board.h"
+#include "application.h"
+#include "axp2101.h"
+#include "config.h"
 #include "cores3_audio_codec.h"
 #include "display/lcd_display.h"
-#include "application.h"
-#include "config.h"
-#include "power_save_timer.h"
 #include "i2c_device.h"
-#include "axp2101.h"
+#include "power_save_timer.h"
+#include "wifi_board.h"
 
-#include <esp_log.h>
 #include <driver/i2c_master.h>
+#include <esp_lcd_ili9341.h>
 #include <esp_lcd_panel_io.h>
 #include <esp_lcd_panel_ops.h>
-#include <esp_lcd_ili9341.h>
+#include <esp_log.h>
 #include <esp_timer.h>
 #include "esp_video.h"
 
 #define TAG "M5StackCoreS3Board"
+#define BACKGROUND_VISION_INITIAL_DELAY_MS 8000
+#define BACKGROUND_VISION_SAMPLE_INTERVAL_MS 100
 
 class Pmic : public Axp2101 {
 public:
@@ -41,7 +43,7 @@ public:
 
 class CustomBacklight : public Backlight {
 public:
-    CustomBacklight(Pmic *pmic) : pmic_(pmic) {}
+    CustomBacklight(Pmic* pmic) : pmic_(pmic) {}
 
     void SetBrightnessImpl(uint8_t brightness) override {
         pmic_->SetBrightness(target_brightness_);
@@ -49,7 +51,7 @@ public:
     }
 
 private:
-    Pmic *pmic_;
+    Pmic* pmic_;
 };
 
 class Aw9523 : public I2cDevice {
@@ -96,9 +98,7 @@ public:
         read_buffer_ = new uint8_t[6];
     }
 
-    ~Ft6336() {
+    ~Ft6336() { delete[] read_buffer_; }
-        delete[] read_buffer_;
-    }
 
     void UpdateTouchPoint() {
         ReadRegs(0x02, read_buffer_, 6);
@@ -107,9 +107,7 @@ public:
         tp_.y = ((read_buffer_[3] & 0x0F) << 8) | read_buffer_[4];
     }
 
-    inline const TouchPoint_t& GetTouchPoint() {
+    inline const TouchPoint_t& GetTouchPoint() { return tp_; }
-        return tp_;
-    }
 
 private:
     uint8_t* read_buffer_ = nullptr;
@@ -137,9 +135,7 @@ private:
             GetDisplay()->SetPowerSaveMode(false);
             GetBacklight()->RestoreBrightness();
         });
-        power_save_timer_->OnShutdownRequest([this]() {
+        power_save_timer_->OnShutdownRequest([this]() { pmic_->PowerOff(); });
-            pmic_->PowerOff();
-        });
         power_save_timer_->SetEnabled(true);
     }
 
@@ -153,7 +149,8 @@ private:
             .glitch_ignore_cnt = 7,
             .intr_priority = 0,
             .trans_queue_depth = 0,
-            .flags = {
+            .flags =
+                {
                     .enable_internal_pullup = 1,
                 },
         };
@@ -195,6 +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();
@@ -204,20 +202,27 @@ 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;
 
-            // 只有短触才触发
             if (touch_duration < TOUCH_THRESHOLD_MS) {
                 auto& app = Application::GetInstance();
                 if (app.GetDeviceState() == kDeviceStateStarting) {
                     EnterWifiConfigMode();
                     return;
                 }
-                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: %s vision+text mode", beaver_mode ? "beaver" : "normal");
+                app.ToggleChatStateForMode(agent_mode, true);
             }
         }
     }
@@ -228,7 +233,8 @@ private:
 
         // 创建定时器，20ms 间隔
         esp_timer_create_args_t timer_args = {
-            .callback = [](void* arg) {
+            .callback =
+                [](void* arg) {
                     M5StackCoreS3Board* board = (M5StackCoreS3Board*)arg;
                     board->PollTouchpad();
                 },
@@ -285,14 +291,16 @@ private:
         esp_lcd_panel_swap_xy(panel, DISPLAY_SWAP_XY);
         esp_lcd_panel_mirror(panel, DISPLAY_MIRROR_X, DISPLAY_MIRROR_Y);
 
-        display_ = new SpiLcdDisplay(panel_io, panel,
+        display_ = new SpiLcdDisplay(panel_io, panel, DISPLAY_WIDTH, DISPLAY_HEIGHT,
-                                    DISPLAY_WIDTH, DISPLAY_HEIGHT, DISPLAY_OFFSET_X, DISPLAY_OFFSET_Y, DISPLAY_MIRROR_X, DISPLAY_MIRROR_Y, DISPLAY_SWAP_XY);
+                                     DISPLAY_OFFSET_X, DISPLAY_OFFSET_Y, DISPLAY_MIRROR_X,
+                                     DISPLAY_MIRROR_Y, DISPLAY_SWAP_XY);
     }
 
     void InitializeCamera() {
         static esp_cam_ctlr_dvp_pin_config_t dvp_pin_config = {
             .data_width = CAM_CTLR_DATA_WIDTH_8,
-            .data_io = {
+            .data_io =
+                {
                     [0] = CAMERA_PIN_D0,
                     [1] = CAMERA_PIN_D1,
                     [2] = CAMERA_PIN_D2,
@@ -330,6 +338,42 @@ private:
         camera_->SetHMirror(false);
     }
 
+    void InitializeBackgroundVisionSampler() {
+        xTaskCreate(
+            [](void* arg) {
+                auto board = static_cast<M5StackCoreS3Board*>(arg);
+                bool has_logged_success = false;
+                bool has_logged_failure = false;
+
+                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_->Capture()) {
+                        if (!has_logged_success) {
+                            ESP_LOGI(TAG, "Vision preview sampler started");
+                            has_logged_success = true;
+                        }
+                    } else if (!has_logged_failure) {
+                        ESP_LOGW(TAG, "Vision preview sampler is waiting for camera");
+                        has_logged_failure = true;
+                    }
+
+                    vTaskDelay(pdMS_TO_TICKS(BACKGROUND_VISION_SAMPLE_INTERVAL_MS));
+                }
+            },
+            "BgVisionSampler", 4096, this, 1, nullptr);
+    }
+
 public:
     M5StackCoreS3Board() {
         InitializePowerSaveTimer();
@@ -340,34 +384,24 @@ public:
         InitializeSpi();
         InitializeIli9342Display();
         InitializeCamera();
+        InitializeBackgroundVisionSampler();
         InitializeFt6336TouchPad();
         GetBacklight()->RestoreBrightness();
     }
 
     virtual AudioCodec* GetAudioCodec() override {
-        static CoreS3AudioCodec audio_codec(i2c_bus_,
+        static CoreS3AudioCodec audio_codec(
-            AUDIO_INPUT_SAMPLE_RATE,
+            i2c_bus_, AUDIO_INPUT_SAMPLE_RATE, AUDIO_OUTPUT_SAMPLE_RATE, AUDIO_I2S_GPIO_MCLK,
-            AUDIO_OUTPUT_SAMPLE_RATE,
+            AUDIO_I2S_GPIO_BCLK, AUDIO_I2S_GPIO_WS, AUDIO_I2S_GPIO_DOUT, AUDIO_I2S_GPIO_DIN,
-            AUDIO_I2S_GPIO_MCLK,
+            AUDIO_CODEC_AW88298_ADDR, AUDIO_CODEC_ES7210_ADDR, AUDIO_INPUT_REFERENCE);
-            AUDIO_I2S_GPIO_BCLK,
-            AUDIO_I2S_GPIO_WS,
-            AUDIO_I2S_GPIO_DOUT,
-            AUDIO_I2S_GPIO_DIN,
-            AUDIO_CODEC_AW88298_ADDR,
-            AUDIO_CODEC_ES7210_ADDR,
-            AUDIO_INPUT_REFERENCE);
         return &audio_codec;
     }
 
-    virtual Display* GetDisplay() override {
+    virtual Display* GetDisplay() override { return display_; }
-        return display_;
-    }
 
-    virtual Camera* GetCamera() override {
+    virtual Camera* GetCamera() override { return camera_; }
-        return camera_;
-    }
 
-    virtual bool GetBatteryLevel(int &level, bool& charging, bool& discharging) override {
+    virtual bool GetBatteryLevel(int& level, bool& charging, bool& discharging) override {
         static bool last_discharging = false;
         charging = pmic_->IsCharging();
         discharging = pmic_->IsDischarging();
@@ -387,7 +421,7 @@ public:
         WifiBoard::SetPowerSaveLevel(level);
     }
 
-    virtual Backlight *GetBacklight() override {
+    virtual Backlight* GetBacklight() override {
         static CustomBacklight backlight(pmic_);
         return &backlight;
     }

main/boards/otto-robot/otto_emoji_display.cc

@@ -77,6 +77,13 @@ void OttoEmojiDisplay::SetStatus(const char* status) {
         lv_obj_add_flag(network_label_, LV_OBJ_FLAG_HIDDEN);
         lv_obj_add_flag(battery_label_, LV_OBJ_FLAG_HIDDEN);
         return;
+    } else if (strcmp(status, Lang::Strings::THINKING) == 0) {
+        lv_obj_set_style_text_font(status_label_, text_font, 0);
+        lv_label_set_text(status_label_, status);
+        lv_obj_clear_flag(status_label_, LV_OBJ_FLAG_HIDDEN);
+        lv_obj_add_flag(network_label_, LV_OBJ_FLAG_HIDDEN);
+        lv_obj_add_flag(battery_label_, LV_OBJ_FLAG_HIDDEN);
+        return;
     } else if (strcmp(status, Lang::Strings::CONNECTING) == 0) {
         lv_obj_set_style_text_font(status_label_, &OTTO_ICON_FONT, 0);
         lv_label_set_text(status_label_, "\xEF\x83\x81");  // U+F0c1 连接图标

main/boards/waveshare/esp32-s3-touch-lcd-4.3c/sdkconfig.4_3c

@@ -598,6 +598,10 @@ CONFIG_PARTITION_TABLE_MD5=y
 # Xiaozhi Assistant
 #
 CONFIG_OTA_URL="https://api.tenclass.net/xiaozhi/ota/"
+CONFIG_USE_DIRECT_WEBSOCKET=y
+CONFIG_WEBSOCKET_URL="ws://172.19.0.240:8080"
+CONFIG_WEBSOCKET_TOKEN=""
+CONFIG_WEBSOCKET_PROTOCOL_VERSION=1
 # CONFIG_FLASH_NONE_ASSETS is not set
 CONFIG_FLASH_DEFAULT_ASSETS=y
 # CONFIG_FLASH_CUSTOM_ASSETS is not set

main/device_state.h

@@ -8,6 +8,7 @@ enum DeviceState {
     kDeviceStateIdle,
     kDeviceStateConnecting,
     kDeviceStateListening,
+    kDeviceStateThinking,
     kDeviceStateSpeaking,
     kDeviceStateUpgrading,
     kDeviceStateActivating,

main/device_state_machine.cc

@@ -13,6 +13,7 @@ static const char* const STATE_STRINGS[] = {
     "idle",
     "connecting",
     "listening",
+    "thinking",
     "speaking",
     "upgrading",
     "activating",
@@ -69,9 +70,10 @@ bool DeviceStateMachine::IsValidTransition(DeviceState from, DeviceState to) con
                    to == kDeviceStateActivating;
 
         case kDeviceStateIdle:
-            // Can go to connecting, listening (manual mode), speaking, activating, upgrading, or wifi configuring
+            // Can go to connecting, listening (manual mode), thinking, speaking, activating, upgrading, or wifi configuring
             return to == kDeviceStateConnecting ||
                    to == kDeviceStateListening ||
+                   to == kDeviceStateThinking ||
                    to == kDeviceStateSpeaking ||
                    to == kDeviceStateActivating ||
                    to == kDeviceStateUpgrading ||
@@ -83,8 +85,15 @@ bool DeviceStateMachine::IsValidTransition(DeviceState from, DeviceState to) con
                    to == kDeviceStateListening;
 
         case kDeviceStateListening:
-            // Can go to speaking or idle
+            // Can go to thinking, speaking, or idle
+            return to == kDeviceStateThinking ||
+                   to == kDeviceStateSpeaking ||
+                   to == kDeviceStateIdle;
+
+        case kDeviceStateThinking:
+            // Can go to speaking, listening, or idle
             return to == kDeviceStateSpeaking ||
+                   to == kDeviceStateListening ||
                    to == kDeviceStateIdle;
 
         case kDeviceStateSpeaking:

main/display/emote_display.cc

@@ -167,6 +167,8 @@ void EmoteDisplay::SetStatus(const char* const status)
             emote_set_event_msg(emote_handle_, EMOTE_MGR_EVT_LISTEN, NULL);
         } else if (std::strcmp(status, Lang::Strings::STANDBY) == 0) {
             emote_set_event_msg(emote_handle_, EMOTE_MGR_EVT_IDLE, NULL);
+        } else if (std::strcmp(status, Lang::Strings::THINKING) == 0) {
+            emote_set_event_msg(emote_handle_, EMOTE_MGR_EVT_LISTEN, NULL);
         } else if (std::strcmp(status, Lang::Strings::SPEAKING) == 0) {
             emote_set_event_msg(emote_handle_, EMOTE_MGR_EVT_SPEAK, NULL);
         } else if (std::strcmp(status, Lang::Strings::ERROR) == 0) {

main/display/lvgl_display/lvgl_display.cc

@@ -203,6 +203,7 @@ void LvglDisplay::UpdateStatusBar(bool update_all) {
             kDeviceStateStarting,
             kDeviceStateWifiConfiguring,
             kDeviceStateListening,
+            kDeviceStateThinking,
             kDeviceStateActivating,
         };
         if (std::find(allowed_states.begin(), allowed_states.end(), device_state) != allowed_states.end()) {

main/led/circular_strip.cc

@@ -228,6 +228,11 @@ void CircularStrip::OnStateChanged() {
             SetAllColor(color);
             break;
         }
+        case kDeviceStateThinking: {
+            StripColor color = { low_brightness_, low_brightness_, default_brightness_ };
+            Blink(color, 300);
+            break;
+        }
         case kDeviceStateUpgrading: {
             StripColor color = { low_brightness_, default_brightness_, low_brightness_ };
             Blink(color, 100);

main/led/gpio_led.cc

@@ -235,6 +235,10 @@ void GpioLed::OnStateChanged() {
             // TurnOn();
             StartFadeTask();
             break;
+        case kDeviceStateThinking:
+            SetBrightness(DEFAULT_BRIGHTNESS);
+            StartContinuousBlink(300);
+            break;
         case kDeviceStateSpeaking:
             SetBrightness(SPEAKING_BRIGHTNESS);
             TurnOn();

main/led/single_led.cc

@@ -152,6 +152,10 @@ void SingleLed::OnStateChanged() {
             SetColor(0, DEFAULT_BRIGHTNESS, 0);
             TurnOn();
             break;
+        case kDeviceStateThinking:
+            SetColor(0, 0, DEFAULT_BRIGHTNESS);
+            StartContinuousBlink(300);
+            break;
         case kDeviceStateUpgrading:
             SetColor(0, DEFAULT_BRIGHTNESS, 0);
             StartContinuousBlink(100);

main/protocols/protocol.cc

@@ -1,9 +1,22 @@
 #include "protocol.h"
 
 #include <esp_log.h>
+#include <mbedtls/base64.h>
 
 #define TAG "Protocol"
 
+static std::string Base64Encode(const std::string& data) {
+    size_t encoded_length = 0;
+    size_t output_length = 0;
+    mbedtls_base64_encode(nullptr, 0, &encoded_length,
+                          reinterpret_cast<const unsigned char*>(data.data()), data.size());
+    std::string result(encoded_length, 0);
+    mbedtls_base64_encode(reinterpret_cast<unsigned char*>(result.data()), result.size(), &output_length,
+                          reinterpret_cast<const unsigned char*>(data.data()), data.size());
+    result.resize(output_length);
+    return result;
+}
+
 void Protocol::OnIncomingJson(std::function<void(const cJSON* root)> callback) {
     on_incoming_json_ = callback;
 }
@@ -78,6 +91,27 @@ void Protocol::SendMcpMessage(const std::string& payload) {
     SendText(message);
 }
 
+void Protocol::SendVisionFrame(const std::string& jpeg_data) {
+    if (jpeg_data.empty()) {
+        return;
+    }
+
+    cJSON* root = cJSON_CreateObject();
+    cJSON_AddStringToObject(root, "session_id", session_id_.c_str());
+    cJSON_AddStringToObject(root, "type", "vision");
+    cJSON_AddStringToObject(root, "state", "frame");
+    cJSON_AddStringToObject(root, "mime_type", "image/jpeg");
+    auto encoded = Base64Encode(jpeg_data);
+    cJSON_AddStringToObject(root, "image", encoded.c_str());
+
+    char* json_str = cJSON_PrintUnformatted(root);
+    if (json_str != nullptr) {
+        SendText(json_str);
+        cJSON_free(json_str);
+    }
+    cJSON_Delete(root);
+}
+
 bool Protocol::IsTimeout() const {
     const int kTimeoutSeconds = 120;
     auto now = std::chrono::steady_clock::now();

main/protocols/protocol.h

@@ -73,6 +73,7 @@ public:
     virtual void SendStopListening();
     virtual void SendAbortSpeaking(AbortReason reason);
     virtual void SendMcpMessage(const std::string& message);
+    virtual void SendVisionFrame(const std::string& jpeg_data);
 
 protected:
     std::function<void(const cJSON* root)> on_incoming_json_;
@@ -95,4 +96,3 @@ protected:
 };
 
 #endif // PROTOCOL_H
-

main/protocols/websocket_protocol.cc

@@ -85,10 +85,21 @@ bool WebsocketProtocol::OpenAudioChannel() {
     std::string url = settings.GetString("url");
     std::string token = settings.GetString("token");
     int version = settings.GetInt("version");
+#if CONFIG_USE_DIRECT_WEBSOCKET
+    url = CONFIG_WEBSOCKET_URL;
+    token = CONFIG_WEBSOCKET_TOKEN;
+    version = CONFIG_WEBSOCKET_PROTOCOL_VERSION;
+#endif
     if (version != 0) {
         version_ = version;
     }
 
+    if (url.empty()) {
+        ESP_LOGE(TAG, "Websocket URL is not set");
+        SetError(Lang::Strings::SERVER_NOT_CONNECTED);
+        return false;
+    }
+
     error_occurred_ = false;
 
     auto network = Board::GetInstance().GetNetwork();
@@ -108,6 +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) {