修复直播测试与 TTS 预热

.gitignore

@@ -2,6 +2,7 @@
 __pycache__/
 *.py[cod]
 .venv/
+.venv-tts/
 runtime/
 config/.venv_python_path.txt
 
@@ -21,6 +22,10 @@ data/queue_state.json
 data/song_requests.json
 data/tts_state.json
 data/bilibili_credentials.json
+data/admin_audit.log
+data/admin_auth.json
+data/statistics.sqlite3*
+data/users.json
 web/music_cover.jpg
 
 # IDE
@@ -30,3 +35,4 @@ web/music_cover.jpg
 # Large binaries (kept local only)
 vendor/mpv/mpv.exe
 vendor/mpv/mpv.7z
+vendor/tts-model/

app/danmu_queue.py

@@ -61,6 +61,10 @@ from pathlib import Path
 from typing import Any
 
 MIHOYO_SDK_REGISTRY_SUBKEY = r"Software\miHoYoSDK"
+MANUAL_LIVE_TEST_DURATION_SECONDS = 30 * 60
+BILIBILI_LIVE_STATUS_POLL_TIMEOUT_SECONDS = 45.0
+BILIBILI_LIVE_STATUS_POLL_INTERVAL_SECONDS = 2.0
+BILIBILI_LIVE_STATUS_REFRESH_SECONDS = 60.0
 
 
 def delete_mihoyo_sdk_registry(logger: logging.Logger, subkey: str = MIHOYO_SDK_REGISTRY_SUBKEY) -> dict[str, Any]:
@@ -511,6 +515,7 @@ class Config:
         tts_all["faster-qwen3-tts"].setdefault("chunk_size", 8)
         tts_all["faster-qwen3-tts"].setdefault("append_silence", True)
         tts_all["faster-qwen3-tts"].setdefault("streaming", True)
+        tts_all["faster-qwen3-tts"].setdefault("startup_timeout_seconds", 900)
         self.data.setdefault("frontend", {})
         if legacy_frontend:
             self.data["frontend"].update(legacy_frontend)
@@ -2450,6 +2455,10 @@ class FasterQwen3TTSEngine:
         self.cpu_threads = max(1, min(8, int(self.cfg.get("cpu_threads", 4) or 4)))
         self.cpu_affinity_count = max(0, int(self.cfg.get("cpu_affinity_count", 8) or 0))
         self.process_priority = str(self.cfg.get("process_priority", "below_normal") or "below_normal")
+        self.startup_timeout_seconds = max(
+            60,
+            int(self.cfg.get("startup_timeout_seconds", 900) or 900),
+        )
         self.logger = logger
         self._parent = parent
         self._worker_ready = False
@@ -2473,6 +2482,7 @@ class FasterQwen3TTSEngine:
             },
             logger,
             synthesis_timeout_seconds=120,
+            startup_timeout_seconds=self.startup_timeout_seconds,
         )
         if bool(self.cfg.get("streaming", False)):
             self.logger.warning("[FasterQwenTTS] 独立 worker 仅使用非流式模式，已忽略 streaming=true")
@@ -2720,6 +2730,7 @@ class Broadcaster:
         self._tts_sequence = 0
         self._tts_started = False
         self._tts_warmup_done = asyncio.Event()
+        self._tts_warmup_task: asyncio.Task | None = None
         self._broadcast_channels: dict[str, dict[str, str]] = {}
         self._stop = False
         # 弹幕会话已登录标志: 启动时ping一下 nav API, -101 表示 SESSDATA 失效, 关闭弹幕发送
@@ -2762,9 +2773,14 @@ class Broadcaster:
                     self.logger.warning("[TTS] 配置热更新时无法调度旧 worker 关闭")
         self.logger.info(f"[配置] 播报配置已热更新: danmu={self.enable_danmu}, tts={self.config.tts_provider}")
 
+    async def _warmup_tts_in_background(self, tts: TTSEngine):
+        try:
+            await tts.warmup()
+        finally:
+            self._tts_warmup_done.set()
+
     async def start(self):
         if self._tts_started:
-            await self._tts_warmup_done.wait()
             return
         self._tts_started = True
         if not self.tts.enabled:
@@ -2777,10 +2793,16 @@ class Broadcaster:
         for task in (synth_task, play_task):
             task.add_done_callback(self._tts_worker_tasks.discard)
 
-        try:
+        if bool(self._tts_queue_cfg.get("warmup_on_start", True)):
-            if bool(self._tts_queue_cfg.get("warmup_on_start", True)):
+            warmup_task = asyncio.create_task(
-                await self.tts.warmup()
+                self._warmup_tts_in_background(self.tts),
-        finally:
+                name="TTS后台预热",
+            )
+            self._tts_warmup_task = warmup_task
+            self._tts_worker_tasks.add(warmup_task)
+            warmup_task.add_done_callback(self._tts_worker_tasks.discard)
+            self.logger.info("[TTS队列] 后台预热已启动，不阻塞弹幕监听")
+        else:
             self._tts_warmup_done.set()
         self.logger.info(
             f"[TTS队列] 流水线启动, pending={self._tts_pending.maxsize}, "
@@ -5285,6 +5307,11 @@ class CommandHandler:
         )
 
     def _is_live_time(self) -> bool:
+        if self.system and (
+            self.system.manual_live_override_active()
+            or self.system.bilibili_live_active()
+        ):
+            return True
         return is_within_live_time(self.config.system_cfg)
 
     def _rule_matches(self, rule: dict, text: str) -> bool:
@@ -6867,6 +6894,23 @@ def get_real_room_id(room_id: int, cookie: str = "") -> int:
     return real_room_id
 
 
+def get_bilibili_live_status(room_id: int, cookie: str = "") -> int:
+    """返回 B站房间直播状态：0=未开播，1=直播中，其他值按接口原样返回。"""
+    data = _request_bilibili_json(
+        f"https://api.live.bilibili.com/room/v1/Room/get_info?room_id={int(room_id)}",
+        cookie,
+        f"https://live.bilibili.com/{int(room_id)}",
+    )
+    if data.get("code") != 0:
+        raise RuntimeError(
+            f"查询直播状态失败: code={data.get('code')} message={data.get('message', '')}"
+        )
+    try:
+        return int((data.get("data") or {}).get("live_status", -1))
+    except (TypeError, ValueError) as exc:
+        raise RuntimeError("直播状态接口未返回有效 live_status") from exc
+
+
 def get_danmu_server(room_id: int, cookie: str = "") -> dict:
     """使用登录 Cookie 获取弹幕服务器和专用 token；token 不得回退为 SESSDATA。"""
     url = f"https://api.live.bilibili.com/xlive/web-room/v1/index/getDanmuInfo?id={int(room_id)}"
@@ -7241,6 +7285,9 @@ class SystemScheduler:
         self._live_session_started_at: datetime | None = None
         self._startup_compensation_task: asyncio.Task | None = None
         self._startup_compensation_window_key: str | None = None
+        self._last_live_action_error = ""
+        self._last_bilibili_live_status: int | None = None
+        self._last_live_status_refresh_monotonic = 0.0
 
     @property
     def cfg(self) -> dict:
@@ -7452,6 +7499,89 @@ class SystemScheduler:
         user32.EnumWindows(EnumWindowsProc(callback), 0)
         return hwnd_found[0] if hwnd_found else None
 
+    @property
+    def last_live_action_error(self) -> str:
+        return self._last_live_action_error
+
+    @staticmethod
+    def _force_foreground_window(hwnd: int) -> bool:
+        """在同一登录会话中临时合并输入线程，可靠地前置目标窗口。"""
+        user32 = ctypes.windll.user32
+        kernel32 = ctypes.windll.kernel32
+        current_thread = int(kernel32.GetCurrentThreadId())
+        foreground = int(user32.GetForegroundWindow() or 0)
+        target_pid = ctypes.wintypes.DWORD()
+        target_thread = int(user32.GetWindowThreadProcessId(hwnd, ctypes.byref(target_pid)) or 0)
+        foreground_thread = 0
+        if foreground:
+            foreground_pid = ctypes.wintypes.DWORD()
+            foreground_thread = int(
+                user32.GetWindowThreadProcessId(foreground, ctypes.byref(foreground_pid)) or 0
+            )
+
+        attached: list[int] = []
+        try:
+            for thread_id in {target_thread, foreground_thread}:
+                if thread_id and thread_id != current_thread:
+                    if user32.AttachThreadInput(current_thread, thread_id, True):
+                        attached.append(thread_id)
+            if user32.IsIconic(hwnd):
+                user32.ShowWindow(hwnd, 9)  # SW_RESTORE
+            else:
+                user32.ShowWindow(hwnd, 5)  # SW_SHOW
+            user32.BringWindowToTop(hwnd)
+            user32.SetForegroundWindow(hwnd)
+            user32.SetActiveWindow(hwnd)
+            user32.SetFocus(hwnd)
+        finally:
+            for thread_id in reversed(attached):
+                user32.AttachThreadInput(current_thread, thread_id, False)
+        return int(user32.GetForegroundWindow() or 0) == int(hwnd)
+
+    async def _get_bilibili_live_status(self, *, record_error: bool = True) -> int | None:
+        try:
+            status = await asyncio.to_thread(
+                get_bilibili_live_status,
+                int(self.config.room_id),
+                self.config.bilibili_cookie,
+            )
+            self._last_bilibili_live_status = status
+            if self.system:
+                self.system.set_bilibili_live_status(status)
+            return status
+        except Exception as exc:
+            message = f"查询 B站直播状态失败: {exc}"
+            if record_error:
+                self._last_live_action_error = message
+                self.logger.warning(f"[系统定时] {message}")
+            else:
+                self.logger.debug(f"[系统定时] {message}")
+            return None
+
+    async def _wait_for_bilibili_live_status(
+        self,
+        expected_status: int,
+        *,
+        timeout_seconds: float = BILIBILI_LIVE_STATUS_POLL_TIMEOUT_SECONDS,
+    ) -> bool:
+        effective_timeout = max(1.0, float(timeout_seconds))
+        deadline = time.monotonic() + effective_timeout
+        while True:
+            status = await self._get_bilibili_live_status()
+            if status == expected_status:
+                self._last_live_action_error = ""
+                return True
+            remaining = deadline - time.monotonic()
+            if remaining <= 0:
+                state_name = "开播" if expected_status == 1 else "关播"
+                self._last_live_action_error = (
+                    f"已点击直播姬，但 B站在 {int(effective_timeout)} 秒内未确认{state_name}"
+                    f"（live_status={status}）"
+                )
+                self.logger.warning(f"[系统定时] {self._last_live_action_error}")
+                return False
+            await asyncio.sleep(min(BILIBILI_LIVE_STATUS_POLL_INTERVAL_SECONDS, remaining))
+
     async def _click_bilibili_live(
         self,
         *,
@@ -7460,57 +7590,137 @@ class SystemScheduler:
         action_name: str,
         confirm_enter: bool = False,
     ) -> bool:
-        """把直播姬窗口前置，在指定比例坐标点击；可选按 Enter 确认弹窗。"""
+        """确认直播姬获得前台后，在恢复后的窗口坐标内点击。"""
+        self._last_live_action_error = ""
         if os.name != "nt":
-            self.logger.warning(f"[系统定时] 当前系统不支持自动{action_name}")
+            self._last_live_action_error = f"当前系统不支持自动{action_name}"
+            self.logger.warning(f"[系统定时] {self._last_live_action_error}")
             return False
         keyword = self.cfg.get("bilibili_push_window_keyword", "直播姬")
         hwnd = self._find_window_by_keyword(keyword)
         if not hwnd:
-            self.logger.warning(f"[系统定时] 未找到直播姬窗口，关键字: {keyword}")
+            self._last_live_action_error = f"未找到直播姬窗口，关键字: {keyword}"
+            self.logger.warning(f"[系统定时] {self._last_live_action_error}")
             return False
         user32 = ctypes.windll.user32
-        rect = ctypes.wintypes.RECT()
+        previous_dpi_context = None
-        if not user32.GetWindowRect(hwnd, ctypes.byref(rect)):
+        set_thread_dpi_context = getattr(user32, "SetThreadDpiAwarenessContext", None)
-            self.logger.warning("[系统定时] 获取直播姬窗口位置失败")
-            return False
-        x_ratio = float(self.cfg.get(x_ratio_key, 0.741))
-        y_ratio = float(self.cfg.get(y_ratio_key, 0.907))
-        x = int(rect.left + max(0.0, min(1.0, x_ratio)) * (rect.right - rect.left))
-        y = int(rect.top + max(0.0, min(1.0, y_ratio)) * (rect.bottom - rect.top))
         try:
-            user32.ShowWindow(hwnd, 9)  # SW_RESTORE
+            if set_thread_dpi_context:
-            user32.SetForegroundWindow(hwnd)
+                set_thread_dpi_context.restype = ctypes.c_void_p
+                previous_dpi_context = set_thread_dpi_context(ctypes.c_void_p(-4))
+
+            if user32.IsIconic(hwnd):
+                user32.ShowWindow(hwnd, 9)  # SW_RESTORE
+            else:
+                user32.ShowWindow(hwnd, 5)  # SW_SHOW
             await asyncio.sleep(0.5)
-            user32.SetCursorPos(x, y)
+            if not self._force_foreground_window(hwnd):
+                await asyncio.sleep(0.3)
+            if int(user32.GetForegroundWindow() or 0) != int(hwnd):
+                self._last_live_action_error = "直播姬窗口无法获得前台，已取消点击以避免误操作其他窗口"
+                self.logger.warning(f"[系统定时] {self._last_live_action_error}")
+                return False
+
+            rect = ctypes.wintypes.RECT()
+            if not user32.GetWindowRect(hwnd, ctypes.byref(rect)):
+                self._last_live_action_error = "获取直播姬窗口位置失败"
+                self.logger.warning(f"[系统定时] {self._last_live_action_error}")
+                return False
+            width = int(rect.right - rect.left)
+            height = int(rect.bottom - rect.top)
+            if width < 200 or height < 150:
+                self._last_live_action_error = (
+                    f"直播姬窗口尺寸异常: rect=({rect.left},{rect.top},{rect.right},{rect.bottom})"
+                )
+                self.logger.warning(f"[系统定时] {self._last_live_action_error}")
+                return False
+
+            x_ratio = float(self.cfg.get(x_ratio_key, 0.741))
+            y_ratio = float(self.cfg.get(y_ratio_key, 0.907))
+            x = int(rect.left + max(0.0, min(1.0, x_ratio)) * width)
+            y = int(rect.top + max(0.0, min(1.0, y_ratio)) * height)
+            virtual_left = int(user32.GetSystemMetrics(76))
+            virtual_top = int(user32.GetSystemMetrics(77))
+            virtual_right = virtual_left + int(user32.GetSystemMetrics(78))
+            virtual_bottom = virtual_top + int(user32.GetSystemMetrics(79))
+            if not (virtual_left <= x < virtual_right and virtual_top <= y < virtual_bottom):
+                self._last_live_action_error = (
+                    f"直播姬点击坐标超出屏幕: x={x}, y={y}, "
+                    f"screen=({virtual_left},{virtual_top},{virtual_right},{virtual_bottom})"
+                )
+                self.logger.warning(f"[系统定时] {self._last_live_action_error}")
+                return False
+
+            previous_cursor = ctypes.wintypes.POINT()
+            has_previous_cursor = bool(user32.GetCursorPos(ctypes.byref(previous_cursor)))
+            if not user32.SetCursorPos(x, y):
+                self._last_live_action_error = (
+                    f"无法移动鼠标到直播姬按钮位置: x={x}, y={y}。"
+                    "请通过桌面上可见的 run.bat 控制台启动服务后重试"
+                )
+                self.logger.warning(f"[系统定时] {self._last_live_action_error}")
+                return False
             await asyncio.sleep(0.1)
+            if int(user32.GetForegroundWindow() or 0) != int(hwnd):
+                self._last_live_action_error = "点击前直播姬失去前台，已取消操作"
+                self.logger.warning(f"[系统定时] {self._last_live_action_error}")
+                if has_previous_cursor:
+                    user32.SetCursorPos(previous_cursor.x, previous_cursor.y)
+                return False
             user32.mouse_event(0x0002, 0, 0, 0, 0)  # LEFTDOWN
             await asyncio.sleep(0.05)
             user32.mouse_event(0x0004, 0, 0, 0, 0)  # LEFTUP
+            if has_previous_cursor:
+                user32.SetCursorPos(previous_cursor.x, previous_cursor.y)
             if confirm_enter:
                 await asyncio.sleep(0.8)
                 user32.keybd_event(0x0D, 0, 0, 0)  # VK_RETURN key down
                 user32.keybd_event(0x0D, 0, 0x0002, 0)  # key up
-            self.logger.info(f"[系统定时] 已点击直播姬{action_name}: hwnd={hwnd}, x={x}, y={y}")
+            self.logger.info(
+                f"[系统定时] 已点击直播姬{action_name}: hwnd={hwnd}, x={x}, y={y}, "
+                f"rect=({rect.left},{rect.top},{rect.right},{rect.bottom})"
+            )
             return True
         except Exception as e:
-            self.logger.warning(f"[系统定时] 点击直播姬{action_name}失败: {e}")
+            self._last_live_action_error = f"点击直播姬{action_name}失败: {e}"
+            self.logger.warning(f"[系统定时] {self._last_live_action_error}")
             return False
+        finally:
+            if set_thread_dpi_context and previous_dpi_context:
+                try:
+                    set_thread_dpi_context(ctypes.c_void_p(previous_dpi_context))
+                except Exception:
+                    pass
 
     async def push_bilibili_live(self) -> bool:
-        return await self._click_bilibili_live(
+        self._last_live_action_error = ""
+        if await self._get_bilibili_live_status() == 1:
+            self.logger.info("[系统定时] B站已处于开播状态，跳过重复点击")
+            return True
+        clicked = await self._click_bilibili_live(
             x_ratio_key="bilibili_push_click_x_ratio",
             y_ratio_key="bilibili_push_click_y_ratio",
             action_name="开启推流按钮",
         )
+        if not clicked:
+            return False
+        return await self._wait_for_bilibili_live_status(1)
 
     async def stop_bilibili_live(self) -> bool:
-        return await self._click_bilibili_live(
+        self._last_live_action_error = ""
+        if await self._get_bilibili_live_status() == 0:
+            self.logger.info("[系统定时] B站已处于未开播状态，跳过重复关播点击")
+            return True
+        clicked = await self._click_bilibili_live(
             x_ratio_key="bilibili_stop_push_click_x_ratio",
             y_ratio_key="bilibili_stop_push_click_y_ratio",
             action_name="关闭推流按钮",
             confirm_enter=bool(self.cfg.get("bilibili_stop_push_confirm_enter", True)),
         )
+        if not clicked:
+            return False
+        return await self._wait_for_bilibili_live_status(0)
 
     def _schedule_reboot_after_stop(self, stopped_at: datetime) -> bool:
         if not self._is_enabled("reboot_after_stop_enabled"):
@@ -7677,6 +7887,13 @@ class SystemScheduler:
             now = datetime.now().replace(second=0, microsecond=0)
             today = date.today().isoformat()
             try:
+                monotonic_now = time.monotonic()
+                if (
+                    monotonic_now - self._last_live_status_refresh_monotonic
+                    >= BILIBILI_LIVE_STATUS_REFRESH_SECONDS
+                ):
+                    self._last_live_status_refresh_monotonic = monotonic_now
+                    await self._get_bilibili_live_status(record_error=False)
                 for event, when in self._live_occurrences(now):
                     key = self._event_key(event, when)
                     if key in self._triggered_events or not self._event_due(now, when):
@@ -7744,6 +7961,11 @@ class QueueSystem:
         self.stats_store = stats_store
         self.health = ServiceRegistry(stats_store)
         self.live_client = None
+        self._manual_live_override_until_monotonic = 0.0
+        self._manual_live_override_until_wall = 0.0
+        self._bilibili_live_status: int | None = None
+        self._bilibili_live_status_updated_monotonic = 0.0
+        self._bilibili_live_status_updated_wall = 0.0
         self.health.set("主程序", ServiceRegistry.STARTING, "初始化")
         self.health.set("配置", ServiceRegistry.RUNNING, f"revision={config.revision}")
         self.user_mgr = UserManager(
@@ -7792,6 +8014,54 @@ class QueueSystem:
         self.handler.login_monitor = self.login_monitor
         self.health.set("主程序", ServiceRegistry.RUNNING, "初始化完成")
 
+    def enable_manual_live_override(
+        self,
+        duration_seconds: int = MANUAL_LIVE_TEST_DURATION_SECONDS,
+    ) -> float:
+        duration = max(60, min(int(duration_seconds), 6 * 60 * 60))
+        self._manual_live_override_until_monotonic = time.monotonic() + duration
+        self._manual_live_override_until_wall = time.time() + duration
+        self.logger.info(f"[直播测试] 已临时开放弹幕指令 {duration // 60} 分钟")
+        return self._manual_live_override_until_wall
+
+    def clear_manual_live_override(self) -> None:
+        was_active = self.manual_live_override_active()
+        self._manual_live_override_until_monotonic = 0.0
+        self._manual_live_override_until_wall = 0.0
+        if was_active:
+            self.logger.info("[直播测试] 已关闭手动直播指令时段")
+
+    def manual_live_override_active(self) -> bool:
+        if self._manual_live_override_until_monotonic <= time.monotonic():
+            self._manual_live_override_until_monotonic = 0.0
+            self._manual_live_override_until_wall = 0.0
+            return False
+        return True
+
+    def manual_live_override_snapshot(self) -> dict[str, Any]:
+        active = self.manual_live_override_active()
+        return {
+            "active": active,
+            "until": self._manual_live_override_until_wall if active else 0.0,
+        }
+
+    def set_bilibili_live_status(self, status: int) -> None:
+        self._bilibili_live_status = int(status)
+        self._bilibili_live_status_updated_monotonic = time.monotonic()
+        self._bilibili_live_status_updated_wall = time.time()
+
+    def bilibili_live_active(self, max_age_seconds: float = 120.0) -> bool:
+        status = getattr(self, "_bilibili_live_status", None)
+        updated = float(getattr(self, "_bilibili_live_status_updated_monotonic", 0.0) or 0.0)
+        return status == 1 and time.monotonic() - updated <= max(1.0, max_age_seconds)
+
+    def bilibili_live_status_snapshot(self) -> dict[str, Any]:
+        return {
+            "status": getattr(self, "_bilibili_live_status", None),
+            "updated_at": float(getattr(self, "_bilibili_live_status_updated_wall", 0.0) or 0.0),
+            "active": self.bilibili_live_active(),
+        }
+
     def apply_config(self):
         self.config.apply_runtime_settings()
         self.user_mgr.config = self.config
@@ -9223,6 +9493,8 @@ class WebServer:
             "active_song_request": song_state.get("active"),
             "recent_danmu": list(getattr(self.system.handler, "recent_danmu", [])[-25:]),
             "recent_gifts": list(getattr(self.system.handler, "recent_gifts", [])[-25:]),
+            "manual_live_test": self.system.manual_live_override_snapshot(),
+            "bilibili_live": self.system.bilibili_live_status_snapshot(),
         })
         return state
 
@@ -10284,13 +10556,19 @@ class WebServer:
         elif action == "test_bilibili_push":
             ok = await self.scheduler.push_bilibili_live()
             if not ok:
-                return web.json_response({"success": False, "error": "未找到直播姬窗口或点击失败"}, status=400)
+                error = self.scheduler.last_live_action_error or "直播姬开播测试失败"
+                return web.json_response({"success": False, "error": error}, status=400)
+            manual_until = self.system.enable_manual_live_override()
             audit_target = "bilibili_push"
+            audit_detail = f"live confirmed manual_until={manual_until:.0f}"
         elif action == "test_bilibili_stop_push":
             ok = await self.scheduler.stop_bilibili_live()
             if not ok:
-                return web.json_response({"success": False, "error": "未找到直播姬窗口或关闭推流点击失败"}, status=400)
+                error = self.scheduler.last_live_action_error or "直播姬关播测试失败"
+                return web.json_response({"success": False, "error": error}, status=400)
+            self.system.clear_manual_live_override()
             audit_target = "bilibili_stop_push"
+            audit_detail = "offline confirmed"
         elif action == "save_system_schedule_config":
             data = await request.json()
             self.config.reload()

app/faster_qwen_worker.py

@@ -3,6 +3,7 @@ from __future__ import annotations
 import io
 import multiprocessing
 import os
+import sys
 import threading
 import time
 import traceback
@@ -13,13 +14,14 @@ from typing import Any, Callable
 
 MAX_NEW_TOKENS = 384
 DEFAULT_SYNTHESIS_TIMEOUT_SECONDS = 120.0
-DEFAULT_STARTUP_TIMEOUT_SECONDS = 480.0
+DEFAULT_STARTUP_TIMEOUT_SECONDS = 900.0
 # 启动失败后再次拉起 worker 的最小间隔：避免"启动超时→立即重启→再超时"的死循环
 # 在系统高负载时持续加载 torch/CUDA，进一步加剧卡顿。
 STARTUP_FAILURE_BACKOFF_SECONDS = 60.0
 DEFAULT_CPU_THREADS = 4
 DEFAULT_CPU_AFFINITY_COUNT = 8
 DEFAULT_PROCESS_PRIORITY = "below_normal"
+STARTUP_PROGRESS_POLL_SECONDS = 1.0
 
 
 class FasterQwenWorkerError(RuntimeError):
@@ -122,15 +124,26 @@ def _apply_worker_process_limits(settings: dict[str, Any]) -> dict[str, Any]:
     }
 
 
-def _load_runtime(settings: dict[str, Any]) -> dict[str, Any]:
+def _load_runtime(
+    settings: dict[str, Any],
+    progress: Callable[[str], None] | None = None,
+) -> dict[str, Any]:
+    def report(stage: str) -> None:
+        if progress:
+            progress(stage)
+
     device = str(settings.get("device") or "cuda")
     if device == "cpu" and "CUDA_VISIBLE_DEVICES" not in os.environ:
         os.environ["CUDA_VISIBLE_DEVICES"] = ""
 
     cpu_threads = _configure_worker_environment(settings)
+    report("torch_import_started")
     import torch
+    report("torch_import_complete")
     _configure_torch_threads(torch, cpu_threads)
+    report("tts_library_import_started")
     from faster_qwen3_tts import FasterQwen3TTS
+    report("tts_library_import_complete")
 
     if device == "cpu":
         torch.cuda.is_available = lambda: False
@@ -138,10 +151,12 @@ def _load_runtime(settings: dict[str, Any]) -> dict[str, Any]:
     load_kwargs: dict[str, Any] = {}
     if device == "cpu":
         load_kwargs["device"] = "cpu"
+    report("from_pretrained_started")
     model = FasterQwen3TTS.from_pretrained(
         str(settings.get("model_name_or_path") or "Qwen/Qwen3-TTS-12Hz-0.6B-Base"),
         **load_kwargs,
     )
+    report("from_pretrained_complete")
 
     # voice_clone_prompt 不再预计算：当前 faster_qwen3_tts 的 FasterQwen3TTS 没有
     # create_voice_clone_prompt 方法，预计算只会失败。改为每次合成时在
@@ -193,21 +208,63 @@ def _synthesize_wav(runtime: dict[str, Any], text: str) -> bytes:
 
 
 def faster_qwen_worker_main(connection: Connection, settings: dict[str, Any]) -> None:
+    def report(stage: str, **details: Any) -> None:
+        try:
+            connection.send({
+                "type": "progress",
+                "stage": stage,
+                "pid": os.getpid(),
+                **details,
+            })
+        except Exception:
+            pass
+
     try:
+        report("process_started", executable=sys.executable, prefix=sys.prefix)
         cpu_threads = _configure_worker_environment(settings)
-        process_limits = _apply_worker_process_limits(settings)
+        report("environment_ready", cpu_threads=cpu_threads)
         load_started = time.monotonic()
-        runtime = _load_runtime(settings)
+        report("model_loading", model_name=str(settings.get("model_name_or_path") or ""))
+        runtime = _load_runtime(settings, progress=report)
         load_ms = int((time.monotonic() - load_started) * 1000)
+        report("model_loaded", load_ms=load_ms)
 
         warmup_started = time.monotonic()
+        graph_warmup_ms = 0
+        report("warmup_started")
+        warmup = getattr(runtime["model"], "warmup", None)
+        if callable(warmup):
+            graph_warmup_started = time.monotonic()
+            warmup()
+            graph_warmup_ms = int((time.monotonic() - graph_warmup_started) * 1000)
+            report("graph_warmup_complete", graph_warmup_ms=graph_warmup_ms)
+
+        # model.warmup() 只捕获模型内部 CUDA Graph，不会初始化参考音频和
+        # voice-clone 生成路径。若把这一步留到首条播报，后台会在近一分钟内
+        # 看起来毫无响应，因此在降低进程优先级之前完成一次真实短句合成。
+        voice_clone_warmup_started = time.monotonic()
+        report("voice_clone_warmup_started")
         _synthesize_wav(runtime, "系统启动")
+        voice_clone_warmup_ms = int((time.monotonic() - voice_clone_warmup_started) * 1000)
+        report("voice_clone_warmup_complete", voice_clone_warmup_ms=voice_clone_warmup_ms)
         warmup_ms = int((time.monotonic() - warmup_started) * 1000)
+        report(
+            "warmup_complete",
+            warmup_ms=warmup_ms,
+            graph_warmup_ms=graph_warmup_ms,
+            voice_clone_warmup_ms=voice_clone_warmup_ms,
+        )
+
+        # 加载和 CUDA Graph 捕获需要完整 CPU 调度能力。仅在预热完成后降低
+        # worker 优先级和亲和性，避免繁忙直播环境下启动时间被放大到超时。
+        process_limits = _apply_worker_process_limits(settings)
         connection.send({
             "type": "ready",
             "pid": os.getpid(),
             "load_ms": load_ms,
             "warmup_ms": warmup_ms,
+            "graph_warmup_ms": graph_warmup_ms,
+            "voice_clone_warmup_ms": voice_clone_warmup_ms,
             "max_new_tokens": MAX_NEW_TOKENS,
             "cpu_threads": cpu_threads,
             **process_limits,
@@ -278,7 +335,7 @@ class FasterQwenWorkerClient:
         self.settings = dict(settings)
         self.logger = logger
         self.synthesis_timeout_seconds = max(1.0, float(synthesis_timeout_seconds))
-        self.startup_timeout_seconds = max(10.0, float(startup_timeout_seconds))
+        self.startup_timeout_seconds = max(0.1, float(startup_timeout_seconds))
         self._context = context or multiprocessing.get_context("spawn")
         self._process_target = process_target or faster_qwen_worker_main
         self._lock = threading.RLock()
@@ -330,21 +387,55 @@ class FasterQwenWorkerClient:
         self._worker_pid = int(getattr(process, "pid", 0) or 0)
         self._log("info", "[FasterQwenTTS] worker 已启动, pid=%s，正在加载和预热", self.worker_pid)
 
-        if not parent_connection.poll(self.startup_timeout_seconds):
+        started = time.monotonic()
-            self._next_start_after = time.monotonic() + STARTUP_FAILURE_BACKOFF_SECONDS
+        deadline = started + self.startup_timeout_seconds
-            self._terminate_worker_locked("startup_timeout", graceful=False)
+        last_stage = "process_spawned"
-            raise FasterQwenWorkerError(
+        message: Any = None
-                f"Faster-Qwen3-TTS worker startup exceeded {self.startup_timeout_seconds:.0f}s"
+        while True:
-            )
+            remaining = deadline - time.monotonic()
-        try:
+            if remaining <= 0:
-            message = parent_connection.recv()
+                self._next_start_after = time.monotonic() + STARTUP_FAILURE_BACKOFF_SECONDS
-        except (EOFError, OSError) as exc:
+                self._terminate_worker_locked("startup_timeout", graceful=False)
-            self._next_start_after = time.monotonic() + STARTUP_FAILURE_BACKOFF_SECONDS
+                raise FasterQwenWorkerError(
-            self._terminate_worker_locked("startup_connection_closed", graceful=False)
+                    f"Faster-Qwen3-TTS worker startup exceeded {self.startup_timeout_seconds:.0f}s "
-            raise FasterQwenWorkerError("Faster-Qwen3-TTS worker exited during startup") from exc
+                    f"(last_stage={last_stage})"
+                )
+            if not parent_connection.poll(min(STARTUP_PROGRESS_POLL_SECONDS, remaining)):
+                if not process.is_alive():
+                    exit_code = getattr(process, "exitcode", None)
+                    self._next_start_after = time.monotonic() + STARTUP_FAILURE_BACKOFF_SECONDS
+                    self._terminate_worker_locked("startup_process_exited", graceful=False)
+                    raise FasterQwenWorkerError(
+                        f"Faster-Qwen3-TTS worker exited during startup "
+                        f"(exit_code={exit_code}, last_stage={last_stage})"
+                    )
+                continue
+            try:
+                message = parent_connection.recv()
+            except (EOFError, OSError) as exc:
+                self._next_start_after = time.monotonic() + STARTUP_FAILURE_BACKOFF_SECONDS
+                self._terminate_worker_locked("startup_connection_closed", graceful=False)
+                raise FasterQwenWorkerError(
+                    f"Faster-Qwen3-TTS worker connection closed during startup "
+                    f"(last_stage={last_stage})"
+                ) from exc
+            if isinstance(message, dict) and message.get("type") == "progress":
+                last_stage = str(message.get("stage") or "unknown")
+                elapsed_ms = int((time.monotonic() - started) * 1000)
+                self._log(
+                    "info",
+                    "[FasterQwenTTS] worker 启动进度: %s, elapsed=%sms",
+                    last_stage,
+                    elapsed_ms,
+                )
+                continue
+            break
         if not isinstance(message, dict) or message.get("type") != "ready":
             if isinstance(message, dict):
                 error = str(message.get("error") or message.get("error_type") or "unknown startup error")
+                worker_traceback = str(message.get("traceback") or "").strip()
+                if worker_traceback:
+                    self._log("error", "[FasterQwenTTS] worker 启动异常:\n%s", worker_traceback)
             else:
                 error = "invalid startup response"
             self._next_start_after = time.monotonic() + STARTUP_FAILURE_BACKOFF_SECONDS
@@ -353,10 +444,13 @@ class FasterQwenWorkerClient:
         self._next_start_after = 0.0
         self._log(
             "info",
-            "[FasterQwenTTS] worker 预热完成, pid=%s load=%sms warmup=%sms max_new_tokens=%s",
+            "[FasterQwenTTS] worker 预热完成, pid=%s load=%sms warmup=%sms "
+            "graph=%sms voice_clone=%sms max_new_tokens=%s",
             self.worker_pid,
             message.get("load_ms"),
             message.get("warmup_ms"),
+            message.get("graph_warmup_ms"),
+            message.get("voice_clone_warmup_ms"),
             message.get("max_new_tokens"),
         )
         self._log(

config/config.json

@@ -59,7 +59,8 @@
         "streaming": false,
         "cpu_threads": 3,
         "cpu_affinity_count": 6,
-        "process_priority": "below_normal"
+        "process_priority": "below_normal",
+        "startup_timeout_seconds": 900
       }
     },
     "enable_danmu_reply": true,

tests/test_faster_qwen_worker.py

@@ -9,6 +9,7 @@ from app.faster_qwen_worker import (
     MAX_NEW_TOKENS,
     FasterQwenWorkerClient,
     FasterQwenWorkerTimeout,
+    faster_qwen_worker_main,
     _generation_kwargs,
     _configure_worker_environment,
 )
@@ -91,6 +92,14 @@ def _ready(pid):
     }
 
 
+def _progress(stage):
+    return {
+        "type": "progress",
+        "stage": stage,
+        "pid": 4000,
+    }
+
+
 def _probe_worker_main(connection, _settings):
     connection.send({
         "type": "ready",
@@ -153,6 +162,8 @@ class FasterQwenWorkerTests(unittest.TestCase):
 
     def test_client_returns_worker_audio(self):
         connection = _FakeConnection([
+            _progress("model_loading"),
+            _progress("warmup_started"),
             _ready(4100),
             {
                 "type": "result",
@@ -181,6 +192,81 @@ class FasterQwenWorkerTests(unittest.TestCase):
         self.assertEqual(len(context.processes), 1)
         client.close()
 
+    def test_client_default_startup_timeout_allows_slow_model_load(self):
+        client = FasterQwenWorkerClient({})
+
+        self.assertEqual(client.startup_timeout_seconds, 900.0)
+        client.close()
+
+    def test_worker_applies_process_limits_after_model_warmup(self):
+        order = []
+
+        class FakeModel:
+            def warmup(self):
+                order.append("warmup")
+
+        class WorkerConnection:
+            def __init__(self):
+                self.sent = []
+                self.closed = False
+
+            def send(self, message):
+                self.sent.append(message)
+
+            def recv(self):
+                return {"command": "stop"}
+
+            def close(self):
+                self.closed = True
+
+        connection = WorkerConnection()
+        with (
+            patch("app.faster_qwen_worker._configure_worker_environment", return_value=3),
+            patch(
+                "app.faster_qwen_worker._load_runtime",
+                side_effect=lambda _settings, progress=None: (
+                    progress("from_pretrained_started") if progress else None,
+                    order.append("load"),
+                    {"model": FakeModel()},
+                )[-1],
+            ),
+            patch(
+                "app.faster_qwen_worker._apply_worker_process_limits",
+                side_effect=lambda _settings: order.append("limits") or {
+                    "cpu_affinity_count": 6,
+                    "process_priority": "below_normal",
+                },
+            ),
+            patch(
+                "app.faster_qwen_worker._synthesize_wav",
+                side_effect=lambda _runtime, _text: order.append("voice_clone") or b"wav",
+            ),
+        ):
+            faster_qwen_worker_main(connection, {})
+
+        self.assertEqual(order, ["load", "warmup", "voice_clone", "limits"])
+        progress_stages = [
+            item.get("stage")
+            for item in connection.sent
+            if item.get("type") == "progress"
+        ]
+        self.assertEqual(
+            progress_stages,
+            [
+                "process_started",
+                "environment_ready",
+                "model_loading",
+                "from_pretrained_started",
+                "model_loaded",
+                "warmup_started",
+                "graph_warmup_complete",
+                "voice_clone_warmup_started",
+                "voice_clone_warmup_complete",
+                "warmup_complete",
+            ],
+        )
+        self.assertEqual(connection.sent[-1]["type"], "ready")
+
     def test_timeout_terminates_worker_and_starts_a_prewarmed_replacement(self):
         first_connection = _FakeConnection([_ready(4200)])
         replacement_connection = _FakeConnection([_ready(4300)])

tests/test_system_scheduler.py

@@ -1,10 +1,12 @@
 import asyncio
+import ctypes
 import logging
 import unittest
 from datetime import datetime, timedelta
+from types import SimpleNamespace
 from unittest.mock import AsyncMock, MagicMock, patch
 
-from app.danmu_queue import SystemScheduler
+from app.danmu_queue import CommandHandler, QueueSystem, SystemScheduler
 
 
 class _FakeConfig:
@@ -155,5 +157,167 @@ class StartupLiveCompensationTests(unittest.IsolatedAsyncioTestCase):
         )
 
 
+class LivehimePushVerificationTests(unittest.IsolatedAsyncioTestCase):
+    def setUp(self):
+        self.logger = logging.getLogger("test-livehime-push-verification")
+        self.scheduler = SystemScheduler(
+            _FakeConfig({
+                "bilibili_push_window_keyword": "直播姬",
+                "bilibili_push_click_x_ratio": 0.787,
+                "bilibili_push_click_y_ratio": 0.927,
+                "bilibili_stop_push_click_x_ratio": 0.787,
+                "bilibili_stop_push_click_y_ratio": 0.927,
+            }),
+            self.logger,
+        )
+
+    async def test_push_skips_click_when_bilibili_is_already_live(self):
+        self.scheduler._get_bilibili_live_status = AsyncMock(return_value=1)
+        self.scheduler._click_bilibili_live = AsyncMock()
+
+        self.assertTrue(await self.scheduler.push_bilibili_live())
+
+        self.scheduler._click_bilibili_live.assert_not_awaited()
+
+    async def test_push_clicks_then_waits_for_confirmed_live_status(self):
+        self.scheduler._get_bilibili_live_status = AsyncMock(side_effect=[0, 0, 1])
+        self.scheduler._click_bilibili_live = AsyncMock(return_value=True)
+
+        with patch("app.danmu_queue.asyncio.sleep", new_callable=AsyncMock):
+            result = await self.scheduler.push_bilibili_live()
+
+        self.assertTrue(result)
+        self.scheduler._click_bilibili_live.assert_awaited_once()
+
+    async def test_stop_skips_click_when_bilibili_is_already_offline(self):
+        self.scheduler._get_bilibili_live_status = AsyncMock(return_value=0)
+        self.scheduler._click_bilibili_live = AsyncMock()
+
+        self.assertTrue(await self.scheduler.stop_bilibili_live())
+
+        self.scheduler._click_bilibili_live.assert_not_awaited()
+
+    async def test_live_status_timeout_returns_specific_error(self):
+        self.scheduler._get_bilibili_live_status = AsyncMock(return_value=0)
+
+        with patch("app.danmu_queue.asyncio.sleep", new_callable=AsyncMock):
+            result = await self.scheduler._wait_for_bilibili_live_status(
+                1,
+                timeout_seconds=0.01,
+            )
+
+        self.assertFalse(result)
+        self.assertIn("未确认开播", self.scheduler.last_live_action_error)
+
+    @patch("app.danmu_queue.os.name", "nt")
+    async def test_click_restores_and_focuses_before_reading_physical_window_rect(self):
+        hwnd = 12345
+        user32 = MagicMock()
+        call_order = []
+        user32.IsIconic.return_value = True
+        user32.GetForegroundWindow.return_value = hwnd
+        user32.GetSystemMetrics.side_effect = lambda index: {
+            76: 0,
+            77: 0,
+            78: 2560,
+            79: 1440,
+        }[index]
+        user32.SetCursorPos.return_value = True
+        user32.GetCursorPos.return_value = False
+        user32.SetThreadDpiAwarenessContext.return_value = None
+
+        def show_window(_hwnd, _command):
+            call_order.append("restore")
+            return True
+
+        def get_window_rect(_hwnd, pointer):
+            call_order.append("rect")
+            rect = ctypes.cast(pointer, ctypes.POINTER(ctypes.wintypes.RECT)).contents
+            rect.left = -11
+            rect.top = -11
+            rect.right = 2571
+            rect.bottom = 1379
+            return True
+
+        user32.ShowWindow.side_effect = show_window
+        user32.GetWindowRect.side_effect = get_window_rect
+        self.scheduler._find_window_by_keyword = MagicMock(return_value=hwnd)
+        self.scheduler._force_foreground_window = MagicMock(return_value=True)
+
+        with (
+            patch("app.danmu_queue.ctypes.windll.user32", user32),
+            patch("app.danmu_queue.asyncio.sleep", new_callable=AsyncMock),
+        ):
+            result = await self.scheduler._click_bilibili_live(
+                x_ratio_key="bilibili_push_click_x_ratio",
+                y_ratio_key="bilibili_push_click_y_ratio",
+                action_name="开启推流按钮",
+            )
+
+        self.assertTrue(result)
+        self.assertLess(call_order.index("restore"), call_order.index("rect"))
+        user32.SetCursorPos.assert_called_once_with(2021, 1277)
+        self.assertEqual(user32.mouse_event.call_count, 2)
+
+
+class ManualLiveOverrideTests(unittest.TestCase):
+    def setUp(self):
+        self.system = QueueSystem.__new__(QueueSystem)
+        self.system.logger = logging.getLogger("test-manual-live-override")
+        self.system._manual_live_override_until_monotonic = 0.0
+        self.system._manual_live_override_until_wall = 0.0
+        self.system._bilibili_live_status = None
+        self.system._bilibili_live_status_updated_monotonic = 0.0
+        self.system._bilibili_live_status_updated_wall = 0.0
+
+    def test_manual_live_override_allows_commands_outside_schedule(self):
+        handler = CommandHandler.__new__(CommandHandler)
+        handler.system = self.system
+        handler.config = SimpleNamespace(
+            system_cfg={"live_start_time": "08:30", "live_end_time": "23:30"}
+        )
+
+        with patch("app.danmu_queue.time.monotonic", return_value=100.0):
+            self.system.enable_manual_live_override(300)
+            self.assertTrue(handler._is_live_time())
+            snapshot = self.system.manual_live_override_snapshot()
+
+        self.assertTrue(snapshot["active"])
+        self.assertGreater(snapshot["until"], 0)
+
+    def test_clear_manual_live_override_restores_schedule_gate(self):
+        handler = CommandHandler.__new__(CommandHandler)
+        handler.system = self.system
+        handler.config = SimpleNamespace(
+            system_cfg={"live_start_time": "08:30", "live_end_time": "23:30"}
+        )
+
+        with (
+            patch("app.danmu_queue.time.monotonic", return_value=100.0),
+            patch("app.danmu_queue.datetime") as mocked_datetime,
+        ):
+            mocked_datetime.now.return_value = datetime(2026, 8, 16, 3, 30)
+            mocked_datetime.strptime = datetime.strptime
+            self.system.enable_manual_live_override(300)
+            self.system.clear_manual_live_override()
+            self.assertFalse(handler._is_live_time())
+
+    def test_confirmed_bilibili_live_status_allows_commands_outside_schedule(self):
+        handler = CommandHandler.__new__(CommandHandler)
+        handler.system = self.system
+        handler.config = SimpleNamespace(
+            system_cfg={"live_start_time": "08:30", "live_end_time": "23:30"}
+        )
+
+        with (
+            patch("app.danmu_queue.time.monotonic", return_value=100.0),
+            patch("app.danmu_queue.datetime") as mocked_datetime,
+        ):
+            mocked_datetime.now.return_value = datetime(2026, 8, 16, 3, 30)
+            mocked_datetime.strptime = datetime.strptime
+            self.system.set_bilibili_live_status(1)
+            self.assertTrue(handler._is_live_time())
+
+
 if __name__ == "__main__":
     unittest.main()

tests/test_tts_pipeline.py

@@ -2,6 +2,7 @@ import asyncio
 import logging
 import time
 import unittest
+from types import SimpleNamespace
 from unittest.mock import patch
 
 import numpy as np
@@ -144,6 +145,49 @@ class PipelineOverlapTests(unittest.IsolatedAsyncioTestCase):
         await asyncio.gather(synth_worker, play_worker, return_exceptions=True)
 
 
+class BackgroundWarmupTests(unittest.IsolatedAsyncioTestCase):
+    async def test_start_returns_before_tts_warmup_finishes(self):
+        warmup_started = asyncio.Event()
+        release_warmup = asyncio.Event()
+
+        class FakeTTS:
+            enabled = True
+
+            async def warmup(self):
+                warmup_started.set()
+                await release_warmup.wait()
+
+        async def idle_worker():
+            await asyncio.Event().wait()
+
+        broadcaster = Broadcaster.__new__(Broadcaster)
+        broadcaster._tts_started = False
+        broadcaster._tts_warmup_done = asyncio.Event()
+        broadcaster._tts_warmup_task = None
+        broadcaster._tts_worker_tasks = set()
+        broadcaster._tts_queue_cfg = {"warmup_on_start": True}
+        broadcaster._tts_pending = SimpleNamespace(maxsize=8)
+        broadcaster._tts_playback_queue = SimpleNamespace(maxsize=2)
+        broadcaster._tts_synthesis_loop = idle_worker
+        broadcaster._tts_playback_loop = idle_worker
+        broadcaster.tts = FakeTTS()
+        broadcaster.logger = logging.getLogger("tts-background-warmup-test")
+
+        await asyncio.wait_for(broadcaster.start(), timeout=0.2)
+        await asyncio.wait_for(warmup_started.wait(), timeout=0.2)
+
+        self.assertFalse(broadcaster._tts_warmup_done.is_set())
+        self.assertIsNotNone(broadcaster._tts_warmup_task)
+
+        release_warmup.set()
+        await asyncio.wait_for(broadcaster._tts_warmup_done.wait(), timeout=0.2)
+
+        tasks = list(broadcaster._tts_worker_tasks)
+        for task in tasks:
+            task.cancel()
+        await asyncio.gather(*tasks, return_exceptions=True)
+
+
 class TTSExpiryPolicyTests(unittest.TestCase):
     def test_default_expiry_windows_match_configured_policy(self):
         broadcaster = Broadcaster.__new__(Broadcaster)