feat: async task pipeline for media and llm workflows

2026-03-15 00:12:26 +08:00
--- a/.env.example
+++ b/.env.example
@@ -12,6 +12,7 @@ VITE_OAUTH_PORTAL_URL=
 VITE_FRONTEND_FORGE_API_URL=
 VITE_FRONTEND_FORGE_API_KEY=
 LOCAL_STORAGE_DIR=/data/app/storage
+APP_PUBLIC_BASE_URL=https://te.hao.work/

 # Compose MySQL
 MYSQL_DATABASE=tennis_training_hub
@@ -23,6 +24,9 @@ MYSQL_ROOT_PASSWORD=replace-with-root-password
 LLM_API_URL=https://one.hao.work/v1/chat/completions
 LLM_API_KEY=replace-with-llm-api-key
 LLM_MODEL=qwen3.5-plus
+LLM_VISION_API_URL=https://one.hao.work/v1/chat/completions
+LLM_VISION_API_KEY=replace-with-llm-api-key
+LLM_VISION_MODEL=qwen3-vl-235b-a22b
 LLM_MAX_TOKENS=32768
 LLM_ENABLE_THINKING=0
 LLM_THINKING_BUDGET=128
@@ -32,3 +36,5 @@ VITE_MEDIA_BASE_URL=/media

 # Local app-to-media proxy for development or direct container access
 MEDIA_SERVICE_URL=http://127.0.0.1:8081
+BACKGROUND_TASK_POLL_MS=3000
+BACKGROUND_TASK_STALE_MS=300000
--- a/2
+++ b/2
@@ -21,4 +21,4 @@ COPY patches ./patches
 RUN pnpm install --prod --frozen-lockfile
 COPY --from=build /app/dist ./dist
 EXPOSE 3000
-CMD ["node", "dist/index.js"]
+CMD ["node", "dist/_core/index.js"]
--- a/README.md
+++ b/README.md
@@ -1,12 +1,13 @@
 # Tennis Training Hub

-网球训练管理与分析应用，提供训练计划、姿势分析、实时摄像头分析、在线视频录制与视频库管理。当前版本新增独立 Go 媒体服务，用于处理在线录制、分段上传、实时推流信令和归档回放。
+网球训练管理与分析应用，提供训练计划、姿势分析、实时摄像头分析、在线视频录制与视频库管理。当前版本在媒体服务之外新增数据库驱动的后台任务系统，用于承接训练计划生成、动作纠正、多模态分析和录制归档这类高延迟任务。

 ## Architecture

 - `client/`: React 19 + TypeScript + Tailwind CSS 4 + shadcn/ui
 - `server/`: Express + tRPC + Drizzle + MySQL/TiDB，负责业务 API、登录、训练数据与视频库元数据
 - `media/`: Go 媒体服务，负责录制会话、分段上传、WebRTC 信令、关键片段标记与 FFmpeg 归档
+- `server/worker.ts`: Node 后台 worker，负责执行重任务队列
 - `docker-compose.yml`: 单机部署编排
 - `deploy/nginx.te.hao.work.conf`: `te.hao.work` 的宿主机 nginx 入口配置

@@ -18,7 +19,27 @@
 - 浏览器端 `RTCPeerConnection` 同步建立 WebRTC 低延迟推流链路
 - 客户端运动检测自动写入关键片段 marker，也支持手动标记
 - 摄像头中断后自动重连，保留既有分段与会话
- 服务端 worker 将分段合并归档，并产出 WebM 回放；FFmpeg 可用时额外生成 MP4
+- Go 媒体 worker 将分段合并归档，并产出 WebM 回放；FFmpeg 可用时额外生成 MP4
+- Node app worker 轮询媒体归档状态，归档完成后自动登记到视频库并向任务中心反馈结果
+
+## Background Tasks
+
+统一后台任务覆盖以下路径：
+
+- `training_plan_generate`
+- `training_plan_adjust`
+- `analysis_corrections`
+- `pose_correction_multimodal`
+- `media_finalize`
+
+前端提供全局任务中心，页面本地也会显示任务提交、执行中、完成或失败状态。训练页、分析页和录制页都可以在用户离开页面后继续完成后台任务。
+
+## Multimodal LLM
+
+- 文本类任务使用 `LLM_API_URL` / `LLM_API_KEY` / `LLM_MODEL`
+- 图片类任务可单独指定 `LLM_VISION_API_URL` / `LLM_VISION_API_KEY` / `LLM_VISION_MODEL`
+- 所有图片输入都要求可从公网访问，因此本地相对路径会通过 `APP_PUBLIC_BASE_URL` 规范化为绝对 URL
+- 若视觉模型链路不可用，系统会自动回退到结构化指标驱动的文本纠正，避免任务直接失败

 ## Quick Start

@@ -67,7 +88,7 @@ pnpm exec playwright install chromium
 单机部署推荐：

 1. 宿主机 nginx 处理 `80/443` 和 TLS
-2. `docker compose up -d --build` 启动 `app + media + worker + db`
+2. `docker compose up -d --build` 启动 `app + app-worker + media + media-worker + db`
 3. nginx 将 `/` 转发到宿主机 `127.0.0.1:3002 -> app:3000`，`/media/` 转发到 `127.0.0.1:8081 -> media:8081`
 4. 如需绕过 nginx 直连调试，也可通过公网 4 位端口访问主站：`http://te.hao.work:8302/`

@@ -100,6 +121,10 @@ pnpm exec playwright install chromium
 - `LLM_API_URL`
 - `LLM_API_KEY`
 - `LLM_MODEL`
+- `LLM_VISION_API_URL`
+- `LLM_VISION_API_KEY`
+- `LLM_VISION_MODEL`
+- `APP_PUBLIC_BASE_URL`
 - `LOCAL_STORAGE_DIR`
 - `MEDIA_SERVICE_URL`
 - `VITE_MEDIA_BASE_URL`
--- a/client/src/components/DashboardLayout.tsx
+++ b/client/src/components/DashboardLayout.tsx
@@ -28,6 +28,7 @@ import {
 import { CSSProperties, useEffect, useRef, useState } from "react";
 import { useLocation, Redirect } from "wouter";
 import { DashboardLayoutSkeleton } from './DashboardLayoutSkeleton';
+import { TaskCenter } from "./TaskCenter";

 const menuItems = [
  { icon: LayoutDashboard, label: "仪表盘", path: "/dashboard", group: "main" },
@@ -262,6 +263,9 @@ function DashboardLayoutContent({
          </SidebarContent>

          <SidebarFooter className="p-3">
+            <div className="mb-3">
+              <TaskCenter />
+            </div>
            <DropdownMenu>
              <DropdownMenuTrigger asChild>
                <button className="flex items-center gap-3 rounded-lg px-1 py-1 hover:bg-accent/50 transition-colors w-full text-left group-data-[collapsible=icon]:justify-center focus:outline-none focus-visible:ring-2 focus-visible:ring-ring">
@@ -315,6 +319,7 @@ function DashboardLayoutContent({
                </div>
              </div>
            </div>
+            <TaskCenter compact />
          </div>
        )}
        <main className={`flex-1 p-4 md:p-6 ${isMobile ? "pb-28" : ""}`}>{children}</main>
--- a/client/src/components/TaskCenter.tsx
+++ b/client/src/components/TaskCenter.tsx
@@ -0,0 +1,152 @@
+import { useEffect, useMemo, useRef } from "react";
+import { trpc } from "@/lib/trpc";
+import { Button } from "@/components/ui/button";
+import { Badge } from "@/components/ui/badge";
+import { Progress } from "@/components/ui/progress";
+import { Sheet, SheetContent, SheetHeader, SheetTitle, SheetTrigger } from "@/components/ui/sheet";
+import { ScrollArea } from "@/components/ui/scroll-area";
+import { toast } from "sonner";
+import { AlertTriangle, BellRing, CheckCircle2, Loader2, RefreshCcw } from "lucide-react";
+
+function formatTaskStatus(status: string) {
+  switch (status) {
+    case "running":
+      return "执行中";
+    case "succeeded":
+      return "已完成";
+    case "failed":
+      return "失败";
+    default:
+      return "排队中";
+  }
+}
+
+export function TaskCenter({ compact = false }: { compact?: boolean }) {
+  const utils = trpc.useUtils();
+  const retryMutation = trpc.task.retry.useMutation({
+    onSuccess: () => {
+      utils.task.list.invalidate();
+      toast.success("任务已重新排队");
+    },
+    onError: (error) => {
+      toast.error(`任务重试失败: ${error.message}`);
+    },
+  });
+
+  const taskListQuery = trpc.task.list.useQuery(
+    { limit: 20 },
+    {
+      refetchInterval: (query) => {
+        const hasActiveTask = (query.state.data ?? []).some((task) => task.status === "queued" || task.status === "running");
+        return hasActiveTask ? 3_000 : 8_000;
+      },
+    }
+  );
+
+  const previousStatusesRef = useRef<Record<string, string>>({});
+  useEffect(() => {
+    for (const task of taskListQuery.data ?? []) {
+      const previous = previousStatusesRef.current[task.id];
+      if (previous && previous !== task.status) {
+        if (task.status === "succeeded") {
+          toast.success(`${task.title} 已完成`);
+          if (task.type === "training_plan_generate" || task.type === "training_plan_adjust") {
+            utils.plan.active.invalidate();
+            utils.plan.list.invalidate();
+          }
+          if (task.type === "media_finalize") {
+            utils.video.list.invalidate();
+          }
+        }
+        if (task.status === "failed") {
+          toast.error(`${task.title} 失败${task.error ? `: ${task.error}` : ""}`);
+        }
+      }
+      previousStatusesRef.current[task.id] = task.status;
+    }
+  }, [taskListQuery.data, utils.plan.active, utils.plan.list, utils.video.list]);
+
+  const activeCount = useMemo(
+    () => (taskListQuery.data ?? []).filter((task) => task.status === "queued" || task.status === "running").length,
+    [taskListQuery.data]
+  );
+
+  return (
+    <Sheet>
+      <SheetTrigger asChild>
+        <Button variant={compact ? "ghost" : "outline"} size="sm" className="gap-2">
+          <BellRing className="h-4 w-4" />
+          <span>{compact ? "任务" : "任务中心"}</span>
+          {activeCount > 0 ? <Badge variant="secondary">{activeCount}</Badge> : null}
+        </Button>
+      </SheetTrigger>
+      <SheetContent side="right" className="w-[380px] sm:w-[420px]">
+        <SheetHeader>
+          <SheetTitle>任务中心</SheetTitle>
+        </SheetHeader>
+        <ScrollArea className="h-[calc(100vh-6rem)] pr-4">
+          <div className="mt-6 space-y-3">
+            {(taskListQuery.data ?? []).length === 0 ? (
+              <div className="rounded-xl border border-dashed p-4 text-sm text-muted-foreground">
+                当前没有后台任务。
+              </div>
+            ) : (
+              (taskListQuery.data ?? []).map((task) => (
+                <div key={task.id} className="rounded-2xl border p-4">
+                  <div className="flex items-start justify-between gap-3">
+                    <div className="min-w-0">
+                      <p className="font-medium leading-5">{task.title}</p>
+                      <p className="mt-1 text-xs text-muted-foreground">{task.message || formatTaskStatus(task.status)}</p>
+                    </div>
+                    <Badge variant={task.status === "failed" ? "destructive" : "secondary"}>
+                      {formatTaskStatus(task.status)}
+                    </Badge>
+                  </div>
+
+                  <div className="mt-3">
+                    <Progress value={task.progress} className="h-2" />
+                  </div>
+
+                  {task.error ? (
+                    <div className="mt-3 rounded-xl bg-red-50 px-3 py-2 text-xs text-red-700">
+                      <div className="flex items-start gap-2">
+                        <AlertTriangle className="mt-0.5 h-3.5 w-3.5 shrink-0" />
+                        <span>{task.error}</span>
+                      </div>
+                    </div>
+                  ) : null}
+
+                  <div className="mt-3 flex items-center justify-between text-xs text-muted-foreground">
+                    <span>{new Date(task.createdAt).toLocaleString("zh-CN")}</span>
+                    {task.status === "failed" ? (
+                      <Button
+                        variant="ghost"
+                        size="sm"
+                        className="h-8 gap-1 px-2"
+                        onClick={() => retryMutation.mutate({ taskId: task.id })}
+                        disabled={retryMutation.isPending}
+                      >
+                        {retryMutation.isPending ? <Loader2 className="h-3.5 w-3.5 animate-spin" /> : <RefreshCcw className="h-3.5 w-3.5" />}
+                        重试
+                      </Button>
+                    ) : task.status === "succeeded" ? (
+                      <span className="inline-flex items-center gap-1 text-emerald-600">
+                        <CheckCircle2 className="h-3.5 w-3.5" />
+                        已完成
+                      </span>
+                    ) : (
+                      <span className="inline-flex items-center gap-1 text-primary">
+                        <Loader2 className="h-3.5 w-3.5 animate-spin" />
+                        处理中
+                      </span>
+                    )}
+                  </div>
+                </div>
+              ))
+            )}
+          </div>
+        </ScrollArea>
+      </SheetContent>
+    </Sheet>
+  );
+}
--- a/client/src/hooks/useBackgroundTask.ts
+++ b/client/src/hooks/useBackgroundTask.ts
@@ -0,0 +1,15 @@
+import { trpc } from "@/lib/trpc";
+
+export function useBackgroundTask(taskId: string | null | undefined) {
+  return trpc.task.get.useQuery(
+    { taskId: taskId || "" },
+    {
+      enabled: Boolean(taskId),
+      refetchInterval: (query) => {
+        const task = query.state.data;
+        if (!task) return 3_000;
+        return task.status === "queued" || task.status === "running" ? 3_000 : false;
+      },
+    }
+  );
+}
--- a/client/src/lib/media.ts
+++ b/client/src/lib/media.ts
@@ -53,14 +53,40 @@ export type MediaSession = {
 };

 const MEDIA_BASE = (import.meta.env.VITE_MEDIA_BASE_URL || "/media").replace(/\/$/, "");
+const RETRYABLE_STATUS = new Set([502, 503, 504]);
+
+function sleep(ms: number) {
+  return new Promise((resolve) => setTimeout(resolve, ms));
+}

 async function request<T>(path: string, init?: RequestInit): Promise<T> {
-  const response = await fetch(`${MEDIA_BASE}${path}`, init);
-  if (!response.ok) {
-    const errorBody = await response.json().catch(() => ({}));
-    throw new Error(errorBody.error || errorBody.message || `Media service error (${response.status})`);
+  let lastError: Error | null = null;
+
+  for (let attempt = 0; attempt < 3; attempt++) {
+    try {
+      const response = await fetch(`${MEDIA_BASE}${path}`, init);
+      if (!response.ok) {
+        const errorBody = await response.json().catch(() => ({}));
+        const error = new Error(errorBody.error || errorBody.message || `Media service error (${response.status})`);
+        if (RETRYABLE_STATUS.has(response.status) && attempt < 2) {
+          lastError = error;
+          await sleep(400 * (attempt + 1));
+          continue;
+        }
+        throw error;
+      }
+      return response.json() as Promise<T>;
+    } catch (error) {
+      lastError = error instanceof Error ? error : new Error("Media request failed");
+      if (attempt < 2) {
+        await sleep(400 * (attempt + 1));
+        continue;
+      }
+      throw lastError;
+    }
  }
-  return response.json() as Promise<T>;
+
+  throw lastError || new Error("Media request failed");
 }

 export async function createMediaSession(payload: {
--- a/client/src/pages/Analysis.tsx
+++ b/client/src/pages/Analysis.tsx
@@ -7,10 +7,12 @@ import { Input } from "@/components/ui/input";
 import { Select, SelectContent, SelectItem, SelectTrigger, SelectValue } from "@/components/ui/select";
 import { Progress } from "@/components/ui/progress";
 import { Badge } from "@/components/ui/badge";
+import { Alert, AlertDescription, AlertTitle } from "@/components/ui/alert";
+import { useBackgroundTask } from "@/hooks/useBackgroundTask";
 import { toast } from "sonner";
 import {
  Upload, Video, Loader2, Play, Pause, RotateCcw,
-  Zap, Target, Activity, TrendingUp, Eye
+  Zap, Target, Activity, TrendingUp, Eye, ListTodo
 } from "lucide-react";
 import { Streamdown } from "streamdown";

@@ -39,6 +41,8 @@ export default function Analysis() {
  const [analysisProgress, setAnalysisProgress] = useState(0);
  const [analysisResult, setAnalysisResult] = useState<AnalysisResult | null>(null);
  const [corrections, setCorrections] = useState<string>("");
+  const [correctionReport, setCorrectionReport] = useState<any>(null);
+  const [correctionTaskId, setCorrectionTaskId] = useState<string | null>(null);
  const [showSkeleton, setShowSkeleton] = useState(false);
  const videoRef = useRef<HTMLVideoElement>(null);
  const canvasRef = useRef<HTMLCanvasElement>(null);
@@ -55,7 +59,16 @@ export default function Analysis() {
      utils.rating.history.invalidate();
    },
  });
-  const correctionMutation = trpc.analysis.getCorrections.useMutation();
+  const correctionMutation = trpc.analysis.getCorrections.useMutation({
+    onSuccess: (data) => {
+      setCorrectionTaskId(data.taskId);
+      toast.success("动作纠正任务已提交");
+    },
+    onError: (error) => {
+      toast.error("动作纠正任务提交失败: " + error.message);
+    },
+  });
+  const correctionTaskQuery = useBackgroundTask(correctionTaskId);

  const handleFileSelect = (e: React.ChangeEvent<HTMLInputElement>) => {
    const file = e.target.files?.[0];
@@ -73,8 +86,22 @@ export default function Analysis() {
    setVideoUrl(URL.createObjectURL(file));
    setAnalysisResult(null);
    setCorrections("");
+    setCorrectionReport(null);
+    setCorrectionTaskId(null);
  };

+  useEffect(() => {
+    if (correctionTaskQuery.data?.status === "succeeded") {
+      const result = correctionTaskQuery.data.result as { corrections?: string; report?: any } | null;
+      setCorrections(result?.corrections || "暂无建议");
+      setCorrectionReport(result?.report || null);
+      setCorrectionTaskId(null);
+    } else if (correctionTaskQuery.data?.status === "failed") {
+      toast.error(`动作纠正失败: ${correctionTaskQuery.data.error || "未知错误"}`);
+      setCorrectionTaskId(null);
+    }
+  }, [correctionTaskQuery.data]);
+
  const analyzeVideo = useCallback(async () => {
    if (!videoRef.current || !canvasRef.current || !videoFile) return;

@@ -267,6 +294,8 @@ export default function Analysis() {
      };

      setAnalysisResult(result);
+      setCorrections("");
+      setCorrectionReport(null);

      // Upload video and save analysis
      const reader = new FileReader();
@@ -293,13 +322,12 @@ export default function Analysis() {
      };
      reader.readAsDataURL(videoFile);

-      // Get AI corrections
+      const snapshots = await extractFrameSnapshots(videoUrl);
      correctionMutation.mutate({
        poseMetrics: result.poseMetrics,
        exerciseType,
        detectedIssues: result.detectedIssues,
-      }, {
-        onSuccess: (data) => setCorrections(data.corrections as string),
+        imageDataUrls: snapshots,
      });

      pose.close();
@@ -318,6 +346,16 @@ export default function Analysis() {
        <p className="text-muted-foreground text-sm mt-1">AI姿势识别与矫正反馈</p>
      </div>

+      {(correctionMutation.isPending || correctionTaskQuery.data?.status === "queued" || correctionTaskQuery.data?.status === "running") ? (
+        <Alert>
+          <ListTodo className="h-4 w-4" />
+          <AlertTitle>后台任务执行中</AlertTitle>
+          <AlertDescription>
+            多模态动作纠正正在后台生成。可以先查看分析结果，完成后任务中心和当前页面都会更新。
+          </AlertDescription>
+        </Alert>
+      ) : null}
+
      {/* Upload section */}
      <Card className="border-0 shadow-sm">
        <CardHeader className="pb-3">
@@ -532,7 +570,12 @@ export default function Analysis() {
              {correctionMutation.isPending ? (
                <div className="flex items-center gap-2 text-muted-foreground">
                  <Loader2 className="h-4 w-4 animate-spin" />
-                  <span className="text-sm">AI正在生成矫正建议...</span>
+                  <span className="text-sm">正在提交动作纠正任务...</span>
+                </div>
+              ) : correctionTaskQuery.data?.status === "queued" || correctionTaskQuery.data?.status === "running" ? (
+                <div className="flex items-center gap-2 text-muted-foreground">
+                  <Loader2 className="h-4 w-4 animate-spin" />
+                  <span className="text-sm">{correctionTaskQuery.data.message || "AI正在后台生成多模态矫正建议..."}</span>
                </div>
              ) : corrections ? (
                <div className="prose prose-sm max-w-none">
@@ -543,6 +586,24 @@ export default function Analysis() {
              )}
            </CardContent>
          </Card>
+
+          {correctionReport?.priorityFixes?.length ? (
+            <Card className="border-0 shadow-sm">
+              <CardHeader className="pb-2">
+                <CardTitle className="text-base">优先修正项</CardTitle>
+              </CardHeader>
+              <CardContent className="space-y-3">
+                {correctionReport.priorityFixes.map((item: any, index: number) => (
+                  <div key={`${item.title}-${index}`} className="rounded-xl border p-3">
+                    <p className="font-medium text-sm">{item.title}</p>
+                    <p className="mt-1 text-sm text-muted-foreground">{item.why}</p>
+                    <p className="mt-2 text-sm"><strong>练习：</strong>{item.howToPractice}</p>
+                    <p className="mt-1 text-xs text-primary"><strong>达标：</strong>{item.successMetric}</p>
+                  </div>
+                ))}
+              </CardContent>
+            </Card>
+          ) : null}
        </>
      )}
    </div>
@@ -667,3 +728,39 @@ function averageAngles(anglesHistory: any[]) {
  }
  return avg;
 }
+
+async function extractFrameSnapshots(sourceUrl: string) {
+  if (!sourceUrl) return [];
+
+  const video = document.createElement("video");
+  video.src = sourceUrl;
+  video.muted = true;
+  video.playsInline = true;
+  video.crossOrigin = "anonymous";
+
+  await new Promise<void>((resolve, reject) => {
+    video.onloadedmetadata = () => resolve();
+    video.onerror = () => reject(new Error("无法读取视频元数据"));
+  });
+
+  const canvas = document.createElement("canvas");
+  canvas.width = video.videoWidth || 1280;
+  canvas.height = video.videoHeight || 720;
+  const ctx = canvas.getContext("2d");
+  if (!ctx) return [];
+
+  const duration = Math.max(video.duration || 0, 1);
+  const checkpoints = [0.15, 0.5, 0.85].map((ratio) => Math.min(duration - 0.05, duration * ratio)).filter((time, index, array) => time >= 0 && array.indexOf(time) === index);
+  const snapshots: string[] = [];
+
+  for (const checkpoint of checkpoints) {
+    await new Promise<void>((resolve) => {
+      video.onseeked = () => resolve();
+      video.currentTime = checkpoint;
+    });
+    ctx.drawImage(video, 0, 0, canvas.width, canvas.height);
+    snapshots.push(canvas.toDataURL("image/jpeg", 0.82));
+  }
+
+  return snapshots;
+}
--- a/client/src/pages/Dashboard.tsx
+++ b/client/src/pages/Dashboard.tsx
@@ -51,11 +51,10 @@ export default function Dashboard() {

  return (
    <div className="space-y-6">
-      {/* Welcome header */}
      <div className="flex flex-col sm:flex-row sm:items-center sm:justify-between gap-4">
        <div>
          <h1 className="text-2xl font-bold tracking-tight" data-testid="dashboard-title">
-            欢迎回来，{user?.name || "球友"}
+            当前用户：{user?.name || "未命名用户"}
          </h1>
          <div className="flex items-center gap-3 mt-2">
            <NTRPBadge rating={stats?.ntrpRating || 1.5} />
--- a/client/src/pages/Login.tsx
+++ b/client/src/pages/Login.tsx
@@ -39,7 +39,7 @@ export default function Login() {
    try {
      const data = await loginMutation.mutateAsync({ username: username.trim() });
      const user = await syncAuthenticatedUser(data.user);
-      toast.success(data.isNew ? `欢迎加入，${user.name}！` : `欢迎回来，${user.name}！`);
+      toast.success(data.isNew ? `已创建用户：${user.name}` : `已登录：${user.name}`);
      setLocation("/dashboard");
    } catch (err) {
      const message = err instanceof Error ? err.message : "未知错误";
--- a/client/src/pages/Recorder.tsx
+++ b/client/src/pages/Recorder.tsx
@@ -18,6 +18,8 @@ import { Button } from "@/components/ui/button";
 import { Card, CardContent, CardHeader, CardTitle } from "@/components/ui/card";
 import { Input } from "@/components/ui/input";
 import { Progress } from "@/components/ui/progress";
+import { Alert, AlertDescription, AlertTitle } from "@/components/ui/alert";
+import { useBackgroundTask } from "@/hooks/useBackgroundTask";
 import { toast } from "sonner";
 import {
  Activity,
@@ -34,6 +36,7 @@ import {
  ShieldAlert,
  Smartphone,
  Sparkles,
+  ListTodo,
  Video,
  VideoOff,
  Wifi,
@@ -126,7 +129,16 @@ function formatFileSize(bytes: number) {

 export default function Recorder() {
  const { user } = useAuth();
-  const registerExternalMutation = trpc.video.registerExternal.useMutation();
+  const utils = trpc.useUtils();
+  const finalizeTaskMutation = trpc.task.createMediaFinalize.useMutation({
+    onSuccess: (data) => {
+      setArchiveTaskId(data.taskId);
+      toast.success("录制归档任务已提交");
+    },
+    onError: (error) => {
+      toast.error(`录制归档任务提交失败: ${error.message}`);
+    },
+  });

  const liveVideoRef = useRef<HTMLVideoElement>(null);
  const playbackVideoRef = useRef<HTMLVideoElement>(null);
@@ -142,11 +154,9 @@ export default function Recorder() {
  const pendingUploadsRef = useRef<PendingSegment[]>([]);
  const uploadInFlightRef = useRef(false);
  const currentSessionRef = useRef<MediaSession | null>(null);
-  const registeredSessionIdRef = useRef<string | null>(null);
  const segmentTickerRef = useRef<ReturnType<typeof setInterval> | null>(null);
  const timerTickerRef = useRef<ReturnType<typeof setInterval> | null>(null);
  const motionTickerRef = useRef<ReturnType<typeof setInterval> | null>(null);
-  const pollTickerRef = useRef<ReturnType<typeof setInterval> | null>(null);
  const reconnectTimeoutRef = useRef<ReturnType<typeof setTimeout> | null>(null);
  const modeRef = useRef<RecorderMode>("idle");
  const reconnectAttemptsRef = useRef(0);
@@ -170,11 +180,13 @@ export default function Recorder() {
  const [markers, setMarkers] = useState<MediaMarker[]>([]);
  const [connectionState, setConnectionState] = useState<RTCPeerConnectionState>("new");
  const [immersivePreview, setImmersivePreview] = useState(false);
+  const [archiveTaskId, setArchiveTaskId] = useState<string | null>(null);

  const mobile = useMemo(() => isMobileDevice(), []);
  const mimeType = useMemo(() => pickRecorderMimeType(), []);
  const currentPlaybackUrl = mediaSession?.playback.mp4Url || mediaSession?.playback.webmUrl || "";
  const archiveProgress = getArchiveProgress(mediaSession);
+  const archiveTaskQuery = useBackgroundTask(archiveTaskId);

  const syncSessionState = useCallback((session: MediaSession | null) => {
    currentSessionRef.current = session;
@@ -196,6 +208,25 @@ export default function Recorder() {
    facingModeRef.current = facingMode;
  }, [facingMode]);

+  useEffect(() => {
+    if (archiveTaskQuery.data?.status === "succeeded") {
+      void (async () => {
+        if (currentSessionRef.current?.id) {
+          const response = await getMediaSession(currentSessionRef.current.id);
+          syncSessionState(response.session);
+        }
+        setMode("archived");
+        utils.video.list.invalidate();
+        toast.success("回放文件已归档完成");
+        setArchiveTaskId(null);
+      })();
+    } else if (archiveTaskQuery.data?.status === "failed") {
+      toast.error(`录制归档失败: ${archiveTaskQuery.data.error || "未知错误"}`);
+      setMode("idle");
+      setArchiveTaskId(null);
+    }
+  }, [archiveTaskQuery.data, syncSessionState, utils.video.list]);
+
  const stopTickers = useCallback(() => {
    if (segmentTickerRef.current) clearInterval(segmentTickerRef.current);
    if (timerTickerRef.current) clearInterval(timerTickerRef.current);
@@ -556,10 +587,10 @@ export default function Recorder() {
      setUploadBytes(0);
      setQueuedSegments(0);
      setReconnectAttempts(0);
+      setArchiveTaskId(null);
      segmentSequenceRef.current = 0;
      motionFrameRef.current = null;
      pendingUploadsRef.current = [];
-      registeredSessionIdRef.current = null;

      const stream = await ensurePreviewStream();
      const sessionResponse = await createMediaSession({
@@ -602,62 +633,19 @@ export default function Recorder() {
        durationMs: Date.now() - recordingStartedAtRef.current,
      });
      syncSessionState(response.session);
-      toast.success("录制已提交，正在生成回放文件");
+      await finalizeTaskMutation.mutateAsync({
+        sessionId: session.id,
+        title: title.trim() || session.title,
+        exerciseType: "recording",
+      });
+      toast.success("录制已提交，后台正在整理回放文件");
    } catch (error: any) {
      toast.error(`结束录制失败: ${error?.message || "未知错误"}`);
      setMode("recording");
-      return;
    }
-
-    if (pollTickerRef.current) {
-      clearInterval(pollTickerRef.current);
-    }
-
-    pollTickerRef.current = setInterval(async () => {
-      const current = currentSessionRef.current;
-      if (!current?.id) {
-        return;
-      }
-      try {
-        const response = await getMediaSession(current.id);
-        syncSessionState(response.session);
-
-        if (response.session.archiveStatus === "completed") {
-          if (pollTickerRef.current) clearInterval(pollTickerRef.current);
-          setMode("archived");
-          toast.success("回放文件已归档完成");
-
-          if (registeredSessionIdRef.current !== response.session.id) {
-            const playbackUrl = response.session.playback.webmUrl || response.session.playback.mp4Url;
-            const playbackFormat = response.session.playback.webmUrl ? "webm" : response.session.playback.mp4Url ? "mp4" : "";
-            if (!playbackUrl || !playbackFormat) {
-              return;
-            }
-            registeredSessionIdRef.current = response.session.id;
-            await registerExternalMutation.mutateAsync({
-              title: title.trim() || response.session.title,
-              url: playbackUrl,
-              fileKey: `media/sessions/${response.session.id}/recording.${playbackFormat}`,
-              format: playbackFormat,
-              fileSize: response.session.playback.webmSize || response.session.playback.mp4Size,
-              duration: response.session.durationMs / 1000,
-              exerciseType: "recording",
-            });
-          }
-        }
-
-        if (response.session.archiveStatus === "failed") {
-          if (pollTickerRef.current) clearInterval(pollTickerRef.current);
-          toast.error(response.session.lastError || "归档失败");
-        }
-      } catch {
-        // keep polling
-      }
-    }, 3_000);
-  }, [closePeer, flushPendingSegments, registerExternalMutation, stopCamera, stopRecorder, syncSessionState, title]);
+  }, [closePeer, finalizeTaskMutation, flushPendingSegments, stopCamera, stopRecorder, syncSessionState, title]);

  const resetRecorder = useCallback(async () => {
-    if (pollTickerRef.current) clearInterval(pollTickerRef.current);
    if (reconnectTimeoutRef.current) clearTimeout(reconnectTimeoutRef.current);
    stopTickers();
    await stopRecorder().catch(() => {});
@@ -667,7 +655,7 @@ export default function Recorder() {
    uploadInFlightRef.current = false;
    motionFrameRef.current = null;
    currentSessionRef.current = null;
-    registeredSessionIdRef.current = null;
+    setArchiveTaskId(null);
    setMediaSession(null);
    setMarkers([]);
    setDurationMs(0);
@@ -755,7 +743,6 @@ export default function Recorder() {

  useEffect(() => {
    return () => {
-      if (pollTickerRef.current) clearInterval(pollTickerRef.current);
      if (reconnectTimeoutRef.current) clearTimeout(reconnectTimeoutRef.current);
      stopTickers();
      if (recorderRef.current && recorderRef.current.state !== "inactive") {
@@ -988,6 +975,17 @@ export default function Recorder() {
        </div>
      </section>

+      {(finalizeTaskMutation.isPending || archiveTaskQuery.data?.status === "queued" || archiveTaskQuery.data?.status === "running") ? (
+        <Alert>
+          <ListTodo className="h-4 w-4" />
+          <AlertTitle>后台归档处理中</AlertTitle>
+          <AlertDescription>
+            {archiveTaskQuery.data?.message || "录制文件正在后台整理、转码并登记到视频库。"}
+            你可以离开当前页面，完成后任务中心会提示结果。
+          </AlertDescription>
+        </Alert>
+      ) : null}
+
      <div className="grid gap-4 xl:grid-cols-[minmax(0,1.7fr)_minmax(340px,0.9fr)]">
        <section className="space-y-4">
          <Card className="overflow-hidden border-0 shadow-lg">
--- a/client/src/pages/Training.tsx
+++ b/client/src/pages/Training.tsx
@@ -1,4 +1,4 @@
-import { useState, useMemo } from "react";
+import { useEffect, useMemo, useState } from "react";
 import { useAuth } from "@/_core/hooks/useAuth";
 import { trpc } from "@/lib/trpc";
 import { Card, CardContent, CardHeader, CardTitle, CardDescription } from "@/components/ui/card";
@@ -6,10 +6,12 @@ import { Button } from "@/components/ui/button";
 import { Badge } from "@/components/ui/badge";
 import { Select, SelectContent, SelectItem, SelectTrigger, SelectValue } from "@/components/ui/select";
 import { Skeleton } from "@/components/ui/skeleton";
+import { Alert, AlertDescription, AlertTitle } from "@/components/ui/alert";
+import { useBackgroundTask } from "@/hooks/useBackgroundTask";
 import { toast } from "sonner";
 import {
  Target, Loader2, CheckCircle2, Circle, Clock, Dumbbell,
-  RefreshCw, Footprints, Hand, ArrowRight, Sparkles
+  RefreshCw, Footprints, Hand, ArrowRight, Sparkles, ListTodo
 } from "lucide-react";

 const categoryIcons: Record<string, React.ReactNode> = {
@@ -42,24 +44,26 @@ export default function Training() {
  const [skillLevel, setSkillLevel] = useState<"beginner" | "intermediate" | "advanced">("beginner");
  const [durationDays, setDurationDays] = useState(7);
  const [selectedDay, setSelectedDay] = useState(1);
+  const [generateTaskId, setGenerateTaskId] = useState<string | null>(null);
+  const [adjustTaskId, setAdjustTaskId] = useState<string | null>(null);

  const utils = trpc.useUtils();
  const { data: activePlan, isLoading: planLoading } = trpc.plan.active.useQuery();
+  const generateTaskQuery = useBackgroundTask(generateTaskId);
+  const adjustTaskQuery = useBackgroundTask(adjustTaskId);

  const generateMutation = trpc.plan.generate.useMutation({
-    onSuccess: () => {
-      toast.success("训练计划已生成！");
-      utils.plan.active.invalidate();
-      utils.plan.list.invalidate();
+    onSuccess: (data) => {
+      setGenerateTaskId(data.taskId);
+      toast.success("训练计划任务已提交");
    },
    onError: (err) => toast.error("生成失败: " + err.message),
  });

  const adjustMutation = trpc.plan.adjust.useMutation({
    onSuccess: (data) => {
-      toast.success("训练计划已调整！");
-      utils.plan.active.invalidate();
-      if (data.adjustmentNotes) toast.info("调整说明: " + data.adjustmentNotes);
+      setAdjustTaskId(data.taskId);
+      toast.success("训练计划调整任务已提交");
    },
    onError: (err) => toast.error("调整失败: " + err.message),
  });
@@ -81,6 +85,36 @@ export default function Training() {
  }, [activePlan, selectedDay]);

  const totalDays = activePlan?.durationDays || 7;
+  const generating = generateMutation.isPending || generateTaskQuery.data?.status === "queued" || generateTaskQuery.data?.status === "running";
+  const adjusting = adjustMutation.isPending || adjustTaskQuery.data?.status === "queued" || adjustTaskQuery.data?.status === "running";
+
+  useEffect(() => {
+    if (generateTaskQuery.data?.status === "succeeded") {
+      toast.success("训练计划已生成");
+      utils.plan.active.invalidate();
+      utils.plan.list.invalidate();
+      setGenerateTaskId(null);
+    } else if (generateTaskQuery.data?.status === "failed") {
+      toast.error(`训练计划生成失败: ${generateTaskQuery.data.error || "未知错误"}`);
+      setGenerateTaskId(null);
+    }
+  }, [generateTaskQuery.data, utils.plan.active, utils.plan.list]);
+
+  useEffect(() => {
+    if (adjustTaskQuery.data?.status === "succeeded") {
+      toast.success("训练计划已调整");
+      utils.plan.active.invalidate();
+      utils.plan.list.invalidate();
+      const adjustmentNotes = (adjustTaskQuery.data.result as { adjustmentNotes?: string } | null)?.adjustmentNotes;
+      if (adjustmentNotes) {
+        toast.info(`调整说明: ${adjustmentNotes}`);
+      }
+      setAdjustTaskId(null);
+    } else if (adjustTaskQuery.data?.status === "failed") {
+      toast.error(`训练计划调整失败: ${adjustTaskQuery.data.error || "未知错误"}`);
+      setAdjustTaskId(null);
+    }
+  }, [adjustTaskQuery.data, utils.plan.active, utils.plan.list]);

  if (planLoading) {
    return (
@@ -100,6 +134,17 @@ export default function Training() {
        </div>
      </div>

+      {generating || adjusting ? (
+        <Alert>
+          <ListTodo className="h-4 w-4" />
+          <AlertTitle>后台任务执行中</AlertTitle>
+          <AlertDescription>
+            {generating ? "训练计划正在后台生成。" : "训练计划正在根据最近分析结果调整。"}
+            你可以切换到其他页面，完成后会在任务中心显示结果。
+          </AlertDescription>
+        </Alert>
+      ) : null}
+
      {!activePlan ? (
        /* Generate new plan */
        <Card className="border-0 shadow-sm">
@@ -145,11 +190,11 @@ export default function Training() {
            <Button
              data-testid="training-generate-button"
              onClick={() => generateMutation.mutate({ skillLevel, durationDays })}
-              disabled={generateMutation.isPending}
+              disabled={generating}
              className="w-full sm:w-auto gap-2"
            >
-              {generateMutation.isPending ? (
-                <><Loader2 className="h-4 w-4 animate-spin" />生成中...</>
+              {generating ? (
+                <><Loader2 className="h-4 w-4 animate-spin" />后台生成中...</>
              ) : (
                <><Sparkles className="h-4 w-4" />生成训练计划</>
              )}
@@ -178,10 +223,10 @@ export default function Training() {
                  variant="outline"
                  size="sm"
                  onClick={() => adjustMutation.mutate({ planId: activePlan.id })}
-                  disabled={adjustMutation.isPending}
+                  disabled={adjusting}
                  className="gap-1"
                >
-                  {adjustMutation.isPending ? (
+                  {adjusting ? (
                    <Loader2 className="h-3 w-3 animate-spin" />
                  ) : (
                    <RefreshCw className="h-3 w-3" />
@@ -274,11 +319,11 @@ export default function Training() {
              onClick={() => {
                generateMutation.mutate({ skillLevel, durationDays });
              }}
-              disabled={generateMutation.isPending}
+              disabled={generating}
              className="gap-2"
            >
-              <Sparkles className="h-4 w-4" />
-              重新生成计划
+              {generating ? <Loader2 className="h-4 w-4 animate-spin" /> : <Sparkles className="h-4 w-4" />}
+              {generating ? "后台生成中..." : "重新生成计划"}
            </Button>
          </div>
        </>
--- a/docker-compose.yml
+++ b/docker-compose.yml
@@ -75,7 +75,7 @@ services:
      - media-data:/data/media
    restart: unless-stopped

-  worker:
+  media-worker:
    build:
      context: ./media
      dockerfile: Dockerfile
@@ -89,6 +89,29 @@ services:
      - media
    restart: unless-stopped

+  app-worker:
+    build:
+      context: .
+      dockerfile: Dockerfile
+    command: ["node", "dist/worker.js"]
+    env_file:
+      - .env
+    environment:
+      DATABASE_URL: mysql://${MYSQL_USER:-tennis}:${MYSQL_PASSWORD:-tennis_password}@db:3306/${MYSQL_DATABASE:-tennis_training_hub}
+      MEDIA_SERVICE_URL: http://media:8081
+      LOCAL_STORAGE_DIR: /data/app/storage
+      NODE_ENV: production
+    volumes:
+      - app-data:/data/app
+    depends_on:
+      db:
+        condition: service_healthy
+      migrate:
+        condition: service_completed_successfully
+      media:
+        condition: service_started
+    restart: unless-stopped
+
 volumes:
  app-data:
  db-data:
--- a/docs/API.md
+++ b/docs/API.md
@@ -75,7 +75,7 @@
 | 类型 | Mutation |
 | 认证 | **需认证** |
 | 输入 | `{ skillLevel: enum, durationDays: number, focusAreas?: string[] }` |
-| 输出 | `{ planId: number, plan: TrainingPlanData }` |
+| 输出 | `{ taskId: string, task: BackgroundTask }` |

 **输入验证：**
 - `skillLevel`：`"beginner"` / `"intermediate"` / `"advanced"`
@@ -105,7 +105,7 @@
 | 类型 | Mutation |
 | 认证 | **需认证** |
 | 输入 | `{ planId: number }` |
-| 输出 | `{ success: true, adjustmentNotes: string }` |
+| 输出 | `{ taskId: string, task: BackgroundTask }` |

 ---

@@ -187,8 +187,10 @@
 |------|-----|
 | 类型 | Mutation |
 | 认证 | **需认证** |
-| 输入 | `{ poseMetrics: object, exerciseType: string, detectedIssues: array }` |
-| 输出 | `{ corrections: string }` |
+| 输入 | `{ poseMetrics: object, exerciseType: string, detectedIssues: array, imageUrls?: string[], imageDataUrls?: string[] }` |
+| 输出 | `{ taskId: string, task: BackgroundTask }` |
+
+该接口始终走后台任务。若提供 `imageUrls` 或 `imageDataUrls`，服务端会优先走多模态纠正链路，并把相对地址规范化为可公网访问的绝对 URL。

 #### `analysis.list` - 获取用户所有分析记录

@@ -211,6 +213,48 @@

 ### 6. 训练记录模块 (`record`)

+### 5.1 后台任务模块 (`task`)
+
+#### `task.list` - 获取当前用户后台任务
+
+| 属性 | 值 |
+|------|-----|
+| 类型 | Query |
+| 认证 | **需认证** |
+| 输入 | `{ limit?: number }` |
+| 输出 | `BackgroundTask[]` |
+
+#### `task.get` - 获取单个后台任务
+
+| 属性 | 值 |
+|------|-----|
+| 类型 | Query |
+| 认证 | **需认证** |
+| 输入 | `{ taskId: string }` |
+| 输出 | `BackgroundTask | null` |
+
+#### `task.retry` - 重试失败任务
+
+| 属性 | 值 |
+|------|-----|
+| 类型 | Mutation |
+| 认证 | **需认证** |
+| 输入 | `{ taskId: string }` |
+| 输出 | `{ task: BackgroundTask }` |
+
+#### `task.createMediaFinalize` - 提交录制归档后台任务
+
+| 属性 | 值 |
+|------|-----|
+| 类型 | Mutation |
+| 认证 | **需认证** |
+| 输入 | `{ sessionId: string, title: string, exerciseType?: string }` |
+| 输出 | `{ taskId: string, task: BackgroundTask }` |
+
+该接口会校验媒体会话所属用户，并由后台 worker 轮询 Go 媒体服务状态，归档完成后自动登记到视频库。
+
+### 6. 训练记录模块 (`record`)
+
 #### `record.create` - 创建训练记录

 | 属性 | 值 |
--- a/docs/FEATURES.md
+++ b/docs/FEATURES.md
@@ -10,7 +10,7 @@
 ### 用户与训练

 - 用户名登录：无需注册，输入用户名即可进入训练工作台
- 训练计划：按技能等级和训练周期生成训练计划
+- 训练计划：按技能等级和训练周期生成训练计划，改为后台异步生成
 - 训练进度：展示训练次数、时长、评分趋势、最近分析结果
 - 每日打卡与提醒：支持训练打卡、提醒、通知记录

@@ -18,18 +18,22 @@

 - 视频上传分析：上传 `webm/mp4` 视频进入视频库并触发分析流程
 - 实时摄像头分析：浏览器端调用 MediaPipe，进行姿势识别和反馈展示
+- 动作纠正：支持文本纠正和多模态纠正两条链路，统一通过后台任务执行
+- 多模态图片输入：上传关键帧后会转换为公网可访问的绝对 URL，再提交给视觉模型
 - 视频库：集中展示录制结果、上传结果和分析摘要

 ### 在线录制与媒体链路

 - Go 媒体服务：独立处理录制会话、分段上传、marker、归档和回放资源
+- Node app worker：统一处理训练计划、动作纠正和录制归档结果登记
 - WebRTC 推流：录制时并行建立低延迟实时推流链路
 - MediaRecorder 分段：浏览器本地压缩录制并每 60 秒自动分段上传
 - 自动标记：客户端通过轻量运动检测创建关键片段 marker
 - 手动标记：录制中支持手动插入剪辑点
 - 自动重连：摄像头 track 断开时自动尝试恢复
 - 归档回放：worker 合并片段并生成 WebM，FFmpeg 可用时额外生成 MP4
- 视频库登记：归档完成后自动写回现有视频库
+- 视频库登记：归档完成后由 app worker 自动写回现有视频库
+- 上传稳定性：媒体分段上传遇到 `502/503/504` 会自动重试

 ## 前端能力

@@ -46,12 +50,14 @@

 - 统一工作台导航
 - 仪表盘、训练、视频、录制、分析等模块一致的布局结构
+- 全局任务中心：桌面侧边栏和移动端头部都可查看后台任务
 - 为后续 PC 粗剪时间线预留媒体域与文档规范

 ## 架构能力

 - Node 应用负责业务 API、登录、训练数据与视频库元数据
 - Go 服务负责媒体链路与归档
+- 后台任务表 `background_tasks` 统一承接重任务
 - `Docker Compose + 宿主机 nginx` 作为标准单机部署方式
 - 统一的本地验证命令：
  - `pnpm check`
@@ -67,10 +73,27 @@
 - 当前 WebRTC 重点是浏览器到服务端的实时上行，不是多观众直播分发
 - 当前 PC 剪辑仍处于基础媒体域准备阶段，未交付完整多轨编辑器
 - 当前存储策略为本地卷优先，未接入对象存储归档
+- 当前 `.env` 配置的视觉网关若忽略 `LLM_VISION_MODEL`，系统会回退到文本纠正；代码已支持独立视觉模型配置，但上游网关能力仍需单独确认

 ## 后续增强方向

- PC 时间线粗剪与 clip plan 持久化
- 更细粒度的设备能力自适应
- 更强的媒体回放和片段导出能力
- 更深入的前端域拆分和懒加载优化
+### 移动端个性化增强
+
+- 根据网络、机型和电量状态动态切换录制档位、分段大小与上传节流策略
+- 将录制焦点视图扩展为单手操作布局，支持拇指热区、自定义主按钮顺序和横竖屏独立面板
+- 为不同训练项目提供场景化预设，例如发球、正手、反手、步伐训练各自保存摄像头方向、裁切比例和提示文案
+- 增加弱网回传面板，向用户展示排队片段、预计上传耗时和失败重试建议
+
+### PC 轻剪与训练回放
+
+- 交付单轨时间线粗剪：入点、出点、片段删除、关键帧封面和 marker 跳转
+- 增加“剪辑计划”实体，允许把自动 marker、手动 marker 和 AI 建议片段一起保存
+- 提供双栏回放模式：左侧原视频，右侧姿态轨迹、节奏评分和文字纠正同步滚动
+- 支持从视频库直接发起导出任务，在后台生成训练集锦或问题片段合集
+
+### 高性能前端重构
+
+- 将训练、分析、录制、视频库拆分为按域加载的路由包，继续降低首屏主包体积
+- 把共享媒体状态、任务状态和用户状态从页面本地逻辑收拢为稳定的数据域层
+- 统一上传、任务轮询、错误提示和绝对 URL 规范化逻辑，减少当前多处重复实现
+- 为重计算页面增加惰性加载、按需图表加载和更严格的移动端资源预算
--- a/docs/deploy.md
+++ b/docs/deploy.md
@@ -6,9 +6,10 @@
 - `db` 容器：MySQL 8，数据持久化到 `db-data`
 - `migrate` 容器：一次性执行 Drizzle 迁移，成功后退出
 - `app` 容器：Node 应用，端口 `3000`
+- `app-worker` 容器：Node 后台任务 worker，共享应用卷与数据库
 - 宿主机公开调试端口：`8302 -> app:3000`
 - `media` 容器：Go 媒体服务，端口 `8081`
- `worker` 容器：Go 媒体归档 worker，共享媒体卷
+- `media-worker` 容器：Go 媒体归档 worker，共享媒体卷
 - `app-data` 卷：上传视频等本地文件存储
 - `db-data` 卷：MySQL 数据目录
 - `media-data` 卷：录制片段、会话状态、归档成片
@@ -32,6 +33,13 @@ docker compose up -d --build
 - `MYSQL_PASSWORD`
 - `MYSQL_ROOT_PASSWORD`
 - `LLM_API_KEY`
+- `APP_PUBLIC_BASE_URL`
+- `LLM_VISION_MODEL`
+
+如需启用独立视觉模型端点，再补：
+
+- `LLM_VISION_API_URL`
+- `LLM_VISION_API_KEY`

 ## nginx

@@ -54,6 +62,7 @@ systemctl reload nginx
 - `curl http://127.0.0.1:3002/api/trpc/auth.me`
 - `curl http://te.hao.work:8302/`
 - `curl http://127.0.0.1:8081/media/health`
+- `docker compose exec app-worker node dist/worker.js --help` 不适用；应通过 `docker compose ps app-worker` 确认 worker 常驻

 ## External access links

@@ -77,4 +86,4 @@ systemctl reload nginx
 2. 回退 Git 版本
 3. 重新执行 `docker compose up -d --build`

-如果只需停止录制链路，可单独关闭 `media` 与 `worker`，主站业务仍可继续运行。
+如果只需停止录制链路，可单独关闭 `media` 与 `media-worker`，主站业务仍可继续运行；如需暂停训练计划/动作纠正等后台任务，再额外停止 `app-worker`。
--- a/docs/testing.md
+++ b/docs/testing.md
@@ -20,6 +20,7 @@

 - Node/tRPC 路由输入校验与权限检查
 - LLM 模块请求配置与环境变量回退逻辑
+- 视觉模型 per-request model override 能力
 - 媒体工具函数，例如录制时长格式化与码率选择

 ### 3. Go 媒体服务测试
@@ -43,6 +44,7 @@
 - 注入假媒体设备、假 `MediaRecorder` 和假 `RTCPeerConnection`

 这样可以自动验证前端主流程，而不依赖数据库、真实摄像头权限和真实 WebRTC 网络环境。
+当前 E2E 已覆盖新的后台任务流和任务中心依赖的接口 mock。

 ## Unified verification

@@ -75,6 +77,14 @@ pnpm test:llm -- "你好，做个自我介绍"
 - 适合验证 `LLM_API_KEY`、`LLM_MODEL` 和网关连通性
 - 不建议纳入 `pnpm verify`，因为它依赖外部网络和真实密钥

+多模态链路建议额外执行一次手工 smoke test：
+
+```bash
+pnpm exec tsx -e 'import "dotenv/config"; import { invokeLLM } from "./server/_core/llm"; const result = await invokeLLM({ model: process.env.LLM_VISION_MODEL, apiUrl: process.env.LLM_VISION_API_URL, apiKey: process.env.LLM_VISION_API_KEY, messages: [{ role: "user", content: [{ type: "text", text: "请用中文一句话描述图片" }, { type: "image_url", image_url: { url: "https://..." } }] }] }); console.log(result.model, result.choices[0]?.message?.content);'
+```
+
+如果返回模型与 `LLM_VISION_MODEL` 不一致，说明上游网关忽略了视觉模型选择，业务任务会自动回退到文本纠正结果。
+
 ## Production smoke checks

 部署到宿主机后，建议至少补以下联测：
--- a/docs/verified-features.md
+++ b/docs/verified-features.md
@@ -1,12 +1,12 @@
 # Verified Features

-本文档记录当前已经通过自动化验证或构建验证的项目。更新时间：2026-03-14 22:24 CST。
+本文档记录当前已经通过自动化验证或构建验证的项目。更新时间：2026-03-15 00:11 CST。

 ## 最新完整验证记录

 - 通过命令：`pnpm verify`
- 验证时间：2026-03-14 22:23 CST
- 结果摘要：`pnpm check` 通过，`pnpm test` 通过（74/74），`pnpm test:go` 通过，`pnpm build` 通过，`pnpm test:e2e` 通过（5/5）
+- 验证时间：2026-03-15 00:10 CST
+- 结果摘要：`pnpm check` 通过，`pnpm test` 通过（80/80），`pnpm test:go` 通过，`pnpm build` 通过，`pnpm test:e2e` 通过（6/6），`pnpm test:llm` 通过

 ## 生产部署联测

@@ -15,10 +15,13 @@
 | `https://te.hao.work/` HTTPS 访问 | `curl -I https://te.hao.work/` | 通过 |
 | `http://te.hao.work:8302/` 4 位端口访问 | `curl -I http://te.hao.work:8302/` | 通过 |
 | 站点 TLS 证书 | Let’s Encrypt ECDSA 证书已签发并由宿主机 nginx 加载 | 通过 |
-| 生产首页、登录页、录制页浏览器打开 | Playwright 访问 `https://te.hao.work/`、`/login`、`/recorder` | 通过 |
+| 生产登录与首次进入工作台 | Playwright 登录真实站点并跳转 `/dashboard` | 通过 |
+| 生产训练 / 实时分析 / 录制 / 视频库页面加载 | Playwright 访问 `/training`、`/live-camera`、`/recorder`、`/videos` | 通过 |
+| 生产训练计划后台任务提交 | Playwright 点击训练计划生成按钮并收到后台任务反馈 | 通过 |
+| 生产移动端录制焦点视图 | Playwright 移动端视口打开 `/recorder` 并验证焦点入口与操作壳层 | 通过 |
 | 生产前端运行时异常检查 | Playwright `pageerror` / `console.error` 检查 | 通过 |
 | 媒体健康检查 | `curl http://127.0.0.1:8081/media/health` | 通过 |
-| compose 自包含服务 | `docker compose ps` 中 `app` / `db` / `media` / `worker` 正常运行，`migrate` 成功退出 | 通过 |
+| compose 自包含服务 | `docker compose ps -a` 中 `app` / `app-worker` / `db` / `media` / `media-worker` 正常运行，`migrate` 成功退出 | 通过 |

 ## 构建与编译通过

@@ -43,6 +46,7 @@
 | badge | `pnpm test` | 通过 |
 | leaderboard | `pnpm test` | 通过 |
 | tutorial / reminder / notification 路由校验 | `pnpm test` | 通过 |
+| task 后台任务路由 | `pnpm test` / `pnpm test:e2e` | 通过 |
 | media 工具函数 | `pnpm test` | 通过 |
 | 登录 URL 回退逻辑 | `pnpm test` | 通过 |

@@ -63,7 +67,9 @@
 | 训练计划 | 训练计划页加载与生成入口可见 | 通过 |
 | 视频库 | 视频卡片渲染 | 通过 |
 | 实时分析 | 摄像头启动入口渲染 | 通过 |
+| 实时分析打分 | 启动分析后出现实时评分结果 | 通过 |
 | 在线录制 | 启动摄像头、开始录制、手动标记、结束归档 | 通过 |
+| 录制焦点视图 | 移动端最大化焦点视图与主操作按钮渲染 | 通过 |
 | 录制结果入库 | 归档完成后视频库可见录制结果 | 通过 |

 ## LLM 模块验证
@@ -72,12 +78,14 @@
 |------|----------|------|
 | `.env` 中的 `LLM_API_URL` / `LLM_API_KEY` / `LLM_MODEL` | `pnpm test:llm` | 通过 |
 | `https://one.hao.work/v1/chat/completions` 联通性 | `pnpm test:llm` 实际返回文本 | 通过 |
+| 视觉模型独立配置路径 | `server/_core/llm.test.ts` + 手工 smoke 检查 | 通过 |

 ## 已知非阻断警告

 - 测试与开发日志中会出现 `OAUTH_SERVER_URL` 未配置提示；当前 mocked auth 和本地验证链路不依赖真实 OAuth 服务，因此不会导致失败
 - `pnpm build` 仍有 Vite 大 chunk 警告；当前属于性能优化待办，不影响本次产物生成
 - Playwright 运行依赖 mocked media/network，不等价于真机摄像头、真实弱网和真实 WebRTC 质量验收
+- 当前上游视觉网关可能忽略 `LLM_VISION_MODEL` 并回退为文本模型；服务端已实现自动降级，任务不会因此直接失败

 ## 当前未纳入自动验证的内容

--- a/drizzle/0005_lively_taskmaster.sql
+++ b/drizzle/0005_lively_taskmaster.sql
@@ -0,0 +1,22 @@
+CREATE TABLE `background_tasks` (
+	`id` varchar(36) NOT NULL,
+	`userId` int NOT NULL,
+	`type` enum('media_finalize','training_plan_generate','training_plan_adjust','analysis_corrections','pose_correction_multimodal') NOT NULL,
+	`status` enum('queued','running','succeeded','failed') NOT NULL DEFAULT 'queued',
+	`title` varchar(256) NOT NULL,
+	`message` text,
+	`progress` int NOT NULL DEFAULT 0,
+	`payload` json NOT NULL,
+	`result` json,
+	`error` text,
+	`attempts` int NOT NULL DEFAULT 0,
+	`maxAttempts` int NOT NULL DEFAULT 3,
+	`workerId` varchar(96),
+	`runAfter` timestamp NOT NULL DEFAULT (now()),
+	`lockedAt` timestamp,
+	`startedAt` timestamp,
+	`completedAt` timestamp,
+	`createdAt` timestamp NOT NULL DEFAULT (now()),
+	`updatedAt` timestamp NOT NULL DEFAULT (now()) ON UPDATE CURRENT_TIMESTAMP,
+	CONSTRAINT `background_tasks_id` PRIMARY KEY(`id`)
+);
--- a/drizzle/meta/_journal.json
+++ b/drizzle/meta/_journal.json
@@ -36,6 +36,13 @@
      "when": 1773490358606,
      "tag": "0004_exotic_randall",
      "breakpoints": true
+    },
+    {
+      "idx": 5,
+      "version": "5",
+      "when": 1773504000000,
+      "tag": "0005_lively_taskmaster",
+      "breakpoints": true
    }
  ]
-}
+}
--- a/drizzle/schema.ts
+++ b/drizzle/schema.ts
@@ -301,3 +301,36 @@ export const notificationLog = mysqlTable("notification_log", {
 export type NotificationLogEntry = typeof notificationLog.$inferSelect;
 export type InsertNotificationLog = typeof notificationLog.$inferInsert;

+/**
+ * Background task queue for long-running or retryable work.
+ */
+export const backgroundTasks = mysqlTable("background_tasks", {
+  id: varchar("id", { length: 36 }).primaryKey(),
+  userId: int("userId").notNull(),
+  type: mysqlEnum("type", [
+    "media_finalize",
+    "training_plan_generate",
+    "training_plan_adjust",
+    "analysis_corrections",
+    "pose_correction_multimodal",
+  ]).notNull(),
+  status: mysqlEnum("status", ["queued", "running", "succeeded", "failed"]).notNull().default("queued"),
+  title: varchar("title", { length: 256 }).notNull(),
+  message: text("message"),
+  progress: int("progress").notNull().default(0),
+  payload: json("payload").notNull(),
+  result: json("result"),
+  error: text("error"),
+  attempts: int("attempts").notNull().default(0),
+  maxAttempts: int("maxAttempts").notNull().default(3),
+  workerId: varchar("workerId", { length: 96 }),
+  runAfter: timestamp("runAfter").defaultNow().notNull(),
+  lockedAt: timestamp("lockedAt"),
+  startedAt: timestamp("startedAt"),
+  completedAt: timestamp("completedAt"),
+  createdAt: timestamp("createdAt").defaultNow().notNull(),
+  updatedAt: timestamp("updatedAt").defaultNow().onUpdateNow().notNull(),
+});
+
+export type BackgroundTask = typeof backgroundTasks.$inferSelect;
+export type InsertBackgroundTask = typeof backgroundTasks.$inferInsert;
--- a/package.json
+++ b/package.json
@@ -6,8 +6,9 @@
  "scripts": {
    "dev": "NODE_ENV=development tsx watch server/_core/index.ts",
    "dev:test": "PORT=41731 STRICT_PORT=1 VITE_APP_ID=test-app VITE_OAUTH_PORTAL_URL=http://127.0.0.1:41731 NODE_ENV=development tsx server/_core/index.ts",
-    "build": "vite build && esbuild server/_core/index.ts --platform=node --packages=external --bundle --format=esm --outdir=dist",
-    "start": "NODE_ENV=production node dist/index.js",
+    "build": "vite build && esbuild server/_core/index.ts server/worker.ts --platform=node --packages=external --bundle --format=esm --outdir=dist",
+    "start": "NODE_ENV=production node dist/_core/index.js",
+    "start:worker": "NODE_ENV=production node dist/worker.js",
    "check": "tsc --noEmit",
    "format": "prettier --write .",
    "test": "vitest run",
--- a/server/_core/env.ts
+++ b/server/_core/env.ts
@@ -11,6 +11,7 @@ const parseBoolean = (value: string | undefined, fallback: boolean) => {

 export const ENV = {
  appId: process.env.VITE_APP_ID ?? "",
+  appPublicBaseUrl: process.env.APP_PUBLIC_BASE_URL ?? "",
  cookieSecret: process.env.JWT_SECRET ?? "",
  databaseUrl: process.env.DATABASE_URL ?? "",
  oAuthServerUrl: process.env.OAUTH_SERVER_URL ?? "",
@@ -27,7 +28,22 @@ export const ENV = {
  llmApiKey:
    process.env.LLM_API_KEY ?? process.env.BUILT_IN_FORGE_API_KEY ?? "",
  llmModel: process.env.LLM_MODEL ?? "gemini-2.5-flash",
+  llmVisionApiUrl:
+    process.env.LLM_VISION_API_URL ??
+    process.env.LLM_API_URL ??
+    (process.env.BUILT_IN_FORGE_API_URL
+      ? `${process.env.BUILT_IN_FORGE_API_URL.replace(/\/$/, "")}/v1/chat/completions`
+      : ""),
+  llmVisionApiKey:
+    process.env.LLM_VISION_API_KEY ??
+    process.env.LLM_API_KEY ??
+    process.env.BUILT_IN_FORGE_API_KEY ??
+    "",
+  llmVisionModel: process.env.LLM_VISION_MODEL ?? process.env.LLM_MODEL ?? "gemini-2.5-flash",
  llmMaxTokens: parseInteger(process.env.LLM_MAX_TOKENS, 32768),
  llmEnableThinking: parseBoolean(process.env.LLM_ENABLE_THINKING, false),
  llmThinkingBudget: parseInteger(process.env.LLM_THINKING_BUDGET, 128),
+  mediaServiceUrl: process.env.MEDIA_SERVICE_URL ?? "",
+  backgroundTaskPollMs: parseInteger(process.env.BACKGROUND_TASK_POLL_MS, 3000),
+  backgroundTaskStaleMs: parseInteger(process.env.BACKGROUND_TASK_STALE_MS, 300000),
 };
--- a/server/_core/llm.test.ts
+++ b/server/_core/llm.test.ts
@@ -68,6 +68,29 @@ describe("invokeLLM", () => {
    expect(JSON.parse(request.body)).not.toHaveProperty("thinking");
  });

+  it("allows overriding the model per request", async () => {
+    process.env.LLM_API_URL = "https://one.hao.work/v1/chat/completions";
+    process.env.LLM_API_KEY = "test-key";
+    process.env.LLM_MODEL = "qwen3.5-plus";
+
+    const fetchMock = vi.fn().mockResolvedValue({
+      ok: true,
+      json: async () => mockSuccessResponse,
+    });
+    vi.stubGlobal("fetch", fetchMock);
+
+    const { invokeLLM } = await import("./llm");
+    await invokeLLM({
+      model: "qwen3-vl-235b-a22b",
+      messages: [{ role: "user", content: "describe image" }],
+    });
+
+    const [, request] = fetchMock.mock.calls[0] as [string, { body: string }];
+    expect(JSON.parse(request.body)).toMatchObject({
+      model: "qwen3-vl-235b-a22b",
+    });
+  });
+
  it("falls back to legacy forge variables when LLM_* values are absent", async () => {
    delete process.env.LLM_API_URL;
    delete process.env.LLM_API_KEY;
--- a/server/_core/llm.ts
+++ b/server/_core/llm.ts
@@ -57,6 +57,9 @@ export type ToolChoice =

 export type InvokeParams = {
  messages: Message[];
+  model?: string;
+  apiUrl?: string;
+  apiKey?: string;
  tools?: Tool[];
  toolChoice?: ToolChoice;
  tool_choice?: ToolChoice;
@@ -209,13 +212,15 @@ const normalizeToolChoice = (
  return toolChoice;
 };

-const resolveApiUrl = () =>
-  ENV.llmApiUrl && ENV.llmApiUrl.trim().length > 0
+const resolveApiUrl = (apiUrl?: string) =>
+  apiUrl && apiUrl.trim().length > 0
+    ? apiUrl
+    : ENV.llmApiUrl && ENV.llmApiUrl.trim().length > 0
    ? ENV.llmApiUrl
    : "https://forge.manus.im/v1/chat/completions";

-const assertApiKey = () => {
-  if (!ENV.llmApiKey) {
+const assertApiKey = (apiKey?: string) => {
+  if (!(apiKey || ENV.llmApiKey)) {
    throw new Error("LLM_API_KEY is not configured");
  }
 };
@@ -266,10 +271,13 @@ const normalizeResponseFormat = ({
 };

 export async function invokeLLM(params: InvokeParams): Promise<InvokeResult> {
-  assertApiKey();
+  assertApiKey(params.apiKey);

  const {
    messages,
+    model,
+    apiUrl,
+    apiKey,
    tools,
    toolChoice,
    tool_choice,
@@ -280,7 +288,7 @@ export async function invokeLLM(params: InvokeParams): Promise<InvokeResult> {
  } = params;

  const payload: Record<string, unknown> = {
-    model: ENV.llmModel,
+    model: model || ENV.llmModel,
    messages: messages.map(normalizeMessage),
  };

@@ -315,11 +323,11 @@ export async function invokeLLM(params: InvokeParams): Promise<InvokeResult> {
    payload.response_format = normalizedResponseFormat;
  }

-  const response = await fetch(resolveApiUrl(), {
+  const response = await fetch(resolveApiUrl(apiUrl), {
    method: "POST",
    headers: {
      "content-type": "application/json",
-      authorization: `Bearer ${ENV.llmApiKey}`,
+      authorization: `Bearer ${apiKey || ENV.llmApiKey}`,
    },
    body: JSON.stringify(payload),
  });
--- a/server/_core/static.ts
+++ b/server/_core/static.ts
@@ -6,7 +6,7 @@ export function serveStatic(app: Express) {
  const distPath =
    process.env.NODE_ENV === "development"
      ? path.resolve(import.meta.dirname, "../..", "dist", "public")
-      : path.resolve(import.meta.dirname, "public");
+      : path.resolve(import.meta.dirname, "..", "public");
  if (!fs.existsSync(distPath)) {
    console.error(
      `Could not find the build directory: ${distPath}, make sure to build the client first`
--- a/server/db.ts
+++ b/server/db.ts
@@ -1,4 +1,4 @@
-import { eq, desc, and, sql } from "drizzle-orm";
+import { eq, desc, and, asc, lte, sql } from "drizzle-orm";
 import { drizzle } from "drizzle-orm/mysql2";
 import {
  InsertUser, users,
@@ -14,6 +14,7 @@ import {
  tutorialProgress, InsertTutorialProgress,
  trainingReminders, InsertTrainingReminder,
  notificationLog, InsertNotificationLog,
+  backgroundTasks, InsertBackgroundTask,
 } from "../drizzle/schema";
 import { ENV } from './_core/env';

@@ -179,6 +180,15 @@ export async function getVideoById(videoId: number) {
  return result.length > 0 ? result[0] : undefined;
 }

+export async function getVideoByFileKey(userId: number, fileKey: string) {
+  const db = await getDb();
+  if (!db) return undefined;
+  const result = await db.select().from(trainingVideos)
+    .where(and(eq(trainingVideos.userId, userId), eq(trainingVideos.fileKey, fileKey)))
+    .limit(1);
+  return result.length > 0 ? result[0] : undefined;
+}
+
 export async function updateVideoStatus(videoId: number, status: "pending" | "analyzing" | "completed" | "failed") {
  const db = await getDb();
  if (!db) return;
@@ -660,6 +670,162 @@ export async function getUnreadNotificationCount(userId: number) {
  return result[0]?.count || 0;
 }

+// ===== BACKGROUND TASK OPERATIONS =====
+
+export async function createBackgroundTask(task: InsertBackgroundTask) {
+  const db = await getDb();
+  if (!db) throw new Error("Database not available");
+  await db.insert(backgroundTasks).values(task);
+  return task.id;
+}
+
+export async function listUserBackgroundTasks(userId: number, limit = 20) {
+  const db = await getDb();
+  if (!db) return [];
+  return db.select().from(backgroundTasks)
+    .where(eq(backgroundTasks.userId, userId))
+    .orderBy(desc(backgroundTasks.createdAt))
+    .limit(limit);
+}
+
+export async function getBackgroundTaskById(taskId: string) {
+  const db = await getDb();
+  if (!db) return undefined;
+  const result = await db.select().from(backgroundTasks)
+    .where(eq(backgroundTasks.id, taskId))
+    .limit(1);
+  return result[0];
+}
+
+export async function getUserBackgroundTaskById(userId: number, taskId: string) {
+  const db = await getDb();
+  if (!db) return undefined;
+  const result = await db.select().from(backgroundTasks)
+    .where(and(eq(backgroundTasks.id, taskId), eq(backgroundTasks.userId, userId)))
+    .limit(1);
+  return result[0];
+}
+
+export async function claimNextBackgroundTask(workerId: string) {
+  const db = await getDb();
+  if (!db) return null;
+
+  const now = new Date();
+  const [nextTask] = await db.select().from(backgroundTasks)
+    .where(and(eq(backgroundTasks.status, "queued"), lte(backgroundTasks.runAfter, now)))
+    .orderBy(asc(backgroundTasks.runAfter), asc(backgroundTasks.createdAt))
+    .limit(1);
+
+  if (!nextTask) {
+    return null;
+  }
+
+  await db.update(backgroundTasks).set({
+    status: "running",
+    workerId,
+    attempts: sql`${backgroundTasks.attempts} + 1`,
+    lockedAt: now,
+    startedAt: now,
+    updatedAt: now,
+  }).where(eq(backgroundTasks.id, nextTask.id));
+
+  return getBackgroundTaskById(nextTask.id);
+}
+
+export async function heartbeatBackgroundTask(taskId: string, workerId: string) {
+  const db = await getDb();
+  if (!db) return;
+  await db.update(backgroundTasks).set({
+    workerId,
+    lockedAt: new Date(),
+  }).where(eq(backgroundTasks.id, taskId));
+}
+
+export async function updateBackgroundTask(taskId: string, data: Partial<InsertBackgroundTask>) {
+  const db = await getDb();
+  if (!db) return;
+  await db.update(backgroundTasks).set(data).where(eq(backgroundTasks.id, taskId));
+}
+
+export async function completeBackgroundTask(taskId: string, result: unknown, message?: string) {
+  const db = await getDb();
+  if (!db) return;
+  await db.update(backgroundTasks).set({
+    status: "succeeded",
+    progress: 100,
+    message: message ?? "已完成",
+    result,
+    error: null,
+    workerId: null,
+    lockedAt: null,
+    completedAt: new Date(),
+  }).where(eq(backgroundTasks.id, taskId));
+}
+
+export async function failBackgroundTask(taskId: string, error: string) {
+  const db = await getDb();
+  if (!db) return;
+  await db.update(backgroundTasks).set({
+    status: "failed",
+    error,
+    workerId: null,
+    lockedAt: null,
+    completedAt: new Date(),
+  }).where(eq(backgroundTasks.id, taskId));
+}
+
+export async function rescheduleBackgroundTask(taskId: string, params: {
+  progress?: number;
+  message?: string;
+  error?: string | null;
+  delayMs?: number;
+}) {
+  const db = await getDb();
+  if (!db) return;
+  await db.update(backgroundTasks).set({
+    status: "queued",
+    progress: params.progress,
+    message: params.message,
+    error: params.error ?? null,
+    workerId: null,
+    lockedAt: null,
+    runAfter: new Date(Date.now() + (params.delayMs ?? 0)),
+  }).where(eq(backgroundTasks.id, taskId));
+}
+
+export async function retryBackgroundTask(userId: number, taskId: string) {
+  const db = await getDb();
+  if (!db) throw new Error("Database not available");
+  const task = await getUserBackgroundTaskById(userId, taskId);
+  if (!task) {
+    throw new Error("Task not found");
+  }
+  await db.update(backgroundTasks).set({
+    status: "queued",
+    progress: 0,
+    message: "任务已重新排队",
+    error: null,
+    result: null,
+    workerId: null,
+    lockedAt: null,
+    completedAt: null,
+    runAfter: new Date(),
+  }).where(eq(backgroundTasks.id, taskId));
+  return getBackgroundTaskById(taskId);
+}
+
+export async function requeueStaleBackgroundTasks(staleBefore: Date) {
+  const db = await getDb();
+  if (!db) return;
+  await db.update(backgroundTasks).set({
+    status: "queued",
+    message: "检测到任务中断，已重新排队",
+    workerId: null,
+    lockedAt: null,
+    runAfter: new Date(),
+  }).where(and(eq(backgroundTasks.status, "running"), lte(backgroundTasks.lockedAt, staleBefore)));
+}
+
 // ===== STATS HELPERS =====

 export async function getUserStats(userId: number) {
--- a/server/mediaService.ts
+++ b/server/mediaService.ts
@@ -0,0 +1,34 @@
+import { ENV } from "./_core/env";
+
+export type RemoteMediaSession = {
+  id: string;
+  userId: string;
+  title: string;
+  archiveStatus: "idle" | "queued" | "processing" | "completed" | "failed";
+  playback: {
+    webmUrl?: string;
+    mp4Url?: string;
+    webmSize?: number;
+    mp4Size?: number;
+    ready: boolean;
+    previewUrl?: string;
+  };
+  lastError?: string;
+};
+
+function getMediaBaseUrl() {
+  if (!ENV.mediaServiceUrl) {
+    throw new Error("MEDIA_SERVICE_URL is not configured");
+  }
+  return ENV.mediaServiceUrl.replace(/\/+$/, "");
+}
+
+export async function getRemoteMediaSession(sessionId: string) {
+  const response = await fetch(`${getMediaBaseUrl()}/sessions/${sessionId}`);
+  if (!response.ok) {
+    const message = await response.text().catch(() => response.statusText);
+    throw new Error(`Media service request failed (${response.status}): ${message}`);
+  }
+  const payload = await response.json() as { session: RemoteMediaSession };
+  return payload.session;
+}
--- a/server/prompts.ts
+++ b/server/prompts.ts
@@ -0,0 +1,255 @@
+type RecentScore = {
+  score: number | null;
+  issues: unknown;
+  exerciseType: string | null;
+  shotCount: number | null;
+  strokeConsistency: number | null;
+  footworkScore: number | null;
+};
+
+type RecentAnalysis = {
+  score: number | null;
+  issues: unknown;
+  corrections: unknown;
+  shotCount: number | null;
+  strokeConsistency: number | null;
+  footworkScore: number | null;
+  fluidityScore: number | null;
+};
+
+function skillLevelLabel(skillLevel: "beginner" | "intermediate" | "advanced") {
+  switch (skillLevel) {
+    case "intermediate":
+      return "中级";
+    case "advanced":
+      return "高级";
+    default:
+      return "初级";
+  }
+}
+
+export function buildTrainingPlanPrompt(input: {
+  skillLevel: "beginner" | "intermediate" | "advanced";
+  durationDays: number;
+  focusAreas?: string[];
+  recentScores: RecentScore[];
+}) {
+  return [
+    `你是一位专业网球教练。请为一位${skillLevelLabel(input.skillLevel)}水平的网球学员生成 ${input.durationDays} 天训练计划。`,
+    "训练条件与要求：",
+    "- 训练以个人可执行为主，可使用球拍、弹力带、标志盘、墙面等常见器材。",
+    "- 每天训练 30-60 分钟，结构要清晰：热身、专项、脚步、力量/稳定、放松。",
+    "- 输出内容要适合直接执行，不写空话，不写营销语，不写额外说明。",
+    input.focusAreas?.length ? `- 重点关注：${input.focusAreas.join("、")}` : "- 如未指定重点，请自动平衡技术、脚步和体能。",
+    input.recentScores.length > 0
+      ? `- 用户最近分析摘要：${JSON.stringify(input.recentScores)}`
+      : "- 暂无历史分析数据，请基于该水平的常见薄弱项设计。",
+    "每个训练项都要给出目标、动作描述、组次/次数、关键提示，避免重复堆砌。",
+  ].join("\n");
+}
+
+export function buildAdjustedTrainingPlanPrompt(input: {
+  currentExercises: unknown;
+  recentAnalyses: RecentAnalysis[];
+}) {
+  return [
+    "你是一位专业网球教练，需要根据最近训练分析结果调整现有训练计划。",
+    `当前计划：${JSON.stringify(input.currentExercises)}`,
+    `最近分析结果：${JSON.stringify(input.recentAnalyses)}`,
+    "请优先修复最近最频繁、最影响击球质量的问题。",
+    "要求：",
+    "- 保留原计划中仍然有效的训练项，不要全部推倒重来。",
+    "- 增加动作纠正、脚步节奏、稳定性和专项力量训练。",
+    "- adjustmentNotes 需要说明为什么这样调整，以及下一阶段重点。",
+    "- 输出仅返回结构化 JSON。",
+  ].join("\n");
+}
+
+export function buildTextCorrectionPrompt(input: {
+  exerciseType: string;
+  poseMetrics: unknown;
+  detectedIssues: unknown;
+}) {
+  return [
+    "你是一位网球技术教练与动作纠正分析师。",
+    `动作类型：${input.exerciseType}`,
+    `姿态指标：${JSON.stringify(input.poseMetrics)}`,
+    `已检测问题：${JSON.stringify(input.detectedIssues)}`,
+    "请用中文输出专业、直接、可执行的纠正建议，使用 Markdown。",
+    "内容结构必须包括：",
+    "1. 动作概览",
+    "2. 最高优先级的 3 个修正点",
+    "3. 每个修正点对应的练习方法、感受提示、完成标准",
+    "4. 下一次拍摄或训练时的注意事项",
+  ].join("\n");
+}
+
+export const multimodalCorrectionSchema = {
+  type: "object",
+  properties: {
+    summary: { type: "string" },
+    overallScore: { type: "number" },
+    confidence: { type: "number" },
+    phaseFindings: {
+      type: "array",
+      items: {
+        type: "object",
+        properties: {
+          phase: { type: "string" },
+          score: { type: "number" },
+          observation: { type: "string" },
+          impact: { type: "string" },
+        },
+        required: ["phase", "score", "observation", "impact"],
+        additionalProperties: false,
+      },
+    },
+    bodyPartFindings: {
+      type: "array",
+      items: {
+        type: "object",
+        properties: {
+          bodyPart: { type: "string" },
+          issue: { type: "string" },
+          recommendation: { type: "string" },
+        },
+        required: ["bodyPart", "issue", "recommendation"],
+        additionalProperties: false,
+      },
+    },
+    priorityFixes: {
+      type: "array",
+      items: {
+        type: "object",
+        properties: {
+          title: { type: "string" },
+          why: { type: "string" },
+          howToPractice: { type: "string" },
+          successMetric: { type: "string" },
+        },
+        required: ["title", "why", "howToPractice", "successMetric"],
+        additionalProperties: false,
+      },
+    },
+    drills: {
+      type: "array",
+      items: {
+        type: "object",
+        properties: {
+          name: { type: "string" },
+          purpose: { type: "string" },
+          durationMinutes: { type: "number" },
+          steps: {
+            type: "array",
+            items: { type: "string" },
+          },
+          coachingCues: {
+            type: "array",
+            items: { type: "string" },
+          },
+        },
+        required: ["name", "purpose", "durationMinutes", "steps", "coachingCues"],
+        additionalProperties: false,
+      },
+    },
+    safetyRisks: {
+      type: "array",
+      items: { type: "string" },
+    },
+    nextSessionFocus: {
+      type: "array",
+      items: { type: "string" },
+    },
+    recommendedCaptureTips: {
+      type: "array",
+      items: { type: "string" },
+    },
+  },
+  required: [
+    "summary",
+    "overallScore",
+    "confidence",
+    "phaseFindings",
+    "bodyPartFindings",
+    "priorityFixes",
+    "drills",
+    "safetyRisks",
+    "nextSessionFocus",
+    "recommendedCaptureTips",
+  ],
+  additionalProperties: false,
+};
+
+export function buildMultimodalCorrectionPrompt(input: {
+  exerciseType: string;
+  poseMetrics: unknown;
+  detectedIssues: unknown;
+  imageCount: number;
+}) {
+  return [
+    "你是一位专业网球技术教练，正在审阅学员的动作截图。",
+    `动作类型：${input.exerciseType}`,
+    `结构化姿态指标：${JSON.stringify(input.poseMetrics)}`,
+    `已有问题标签：${JSON.stringify(input.detectedIssues)}`,
+    `本次共提供 ${input.imageCount} 张关键帧图片。`,
+    "请严格依据图片和结构化指标交叉判断，不要编造看不到的动作细节。",
+    "分析要求：",
+    "- 识别准备、引拍、击球/发力、收拍几个阶段的质量。",
+    "- 指出躯干、肩髋、击球臂、非持拍手、重心和脚步的主要问题。",
+    "- priorityFixes 只保留最重要、最值得优先修正的项目。",
+    "- drills 要足够具体，适合下一次训练直接执行。",
+    "- recommendedCaptureTips 说明下次如何补拍，以便提高判断准确度。",
+    "输出仅返回 JSON，不要附加解释。",
+  ].join("\n");
+}
+
+export function renderMultimodalCorrectionMarkdown(report: {
+  summary: string;
+  overallScore: number;
+  confidence: number;
+  priorityFixes: Array<{ title: string; why: string; howToPractice: string; successMetric: string }>;
+  drills: Array<{ name: string; purpose: string; durationMinutes: number; coachingCues: string[] }>;
+  safetyRisks: string[];
+  nextSessionFocus: string[];
+  recommendedCaptureTips: string[];
+}) {
+  const priorityFixes = report.priorityFixes
+    .map((item, index) => [
+      `${index + 1}. ${item.title}`,
+      `- 原因：${item.why}`,
+      `- 练习：${item.howToPractice}`,
+      `- 达标：${item.successMetric}`,
+    ].join("\n"))
+    .join("\n");
+
+  const drills = report.drills
+    .map((item) => [
+      `- ${item.name}（${item.durationMinutes} 分钟）`,
+      `  目的：${item.purpose}`,
+      `  口令：${item.coachingCues.join("；")}`,
+    ].join("\n"))
+    .join("\n");
+
+  return [
+    `## 动作概览`,
+    report.summary,
+    "",
+    `- 综合评分：${Math.round(report.overallScore)}/100`,
+    `- 置信度：${Math.round(report.confidence)}%`,
+    "",
+    "## 优先修正",
+    priorityFixes || "- 暂无",
+    "",
+    "## 推荐练习",
+    drills || "- 暂无",
+    "",
+    "## 风险提醒",
+    report.safetyRisks.length > 0 ? report.safetyRisks.map(item => `- ${item}`).join("\n") : "- 暂无明显风险",
+    "",
+    "## 下次训练重点",
+    report.nextSessionFocus.length > 0 ? report.nextSessionFocus.map(item => `- ${item}`).join("\n") : "- 保持当前节奏",
+    "",
+    "## 下次拍摄建议",
+    report.recommendedCaptureTips.length > 0 ? report.recommendedCaptureTips.map(item => `- ${item}`).join("\n") : "- 保持当前拍摄方式",
+  ].join("\n");
+}
--- a/server/publicUrl.ts
+++ b/server/publicUrl.ts
@@ -0,0 +1,22 @@
+import { ENV } from "./_core/env";
+
+function hasProtocol(value: string) {
+  return /^[a-z][a-z0-9+.-]*:\/\//i.test(value);
+}
+
+export function toPublicUrl(pathOrUrl: string) {
+  const value = pathOrUrl.trim();
+  if (!value) {
+    throw new Error("Public URL value is empty");
+  }
+
+  if (hasProtocol(value)) {
+    return value;
+  }
+
+  if (!ENV.appPublicBaseUrl) {
+    throw new Error("APP_PUBLIC_BASE_URL is required for externally accessible asset URLs");
+  }
+
+  return new URL(value.startsWith("/") ? value : `/${value}`, ENV.appPublicBaseUrl).toString();
+}
--- a/server/routers.ts
+++ b/server/routers.ts
@@ -4,53 +4,34 @@ import { systemRouter } from "./_core/systemRouter";
 import { publicProcedure, protectedProcedure, router } from "./_core/trpc";
 import { z } from "zod";
 import { sdk } from "./_core/sdk";
-import { invokeLLM } from "./_core/llm";
 import { storagePut } from "./storage";
 import * as db from "./db";
 import { nanoid } from "nanoid";
-import {
-  normalizeAdjustedPlanResponse,
-  normalizeTrainingPlanResponse,
-} from "./trainingPlan";
+import { getRemoteMediaSession } from "./mediaService";
+import { prepareCorrectionImageUrls } from "./taskWorker";
+import { toPublicUrl } from "./publicUrl";

-async function invokeStructuredPlan<T>(params: {
-  baseMessages: Array<{ role: "system" | "user"; content: string }>;
-  responseFormat: {
-    type: "json_schema";
-    json_schema: {
-      name: string;
-      strict: true;
-      schema: Record<string, unknown>;
-    };
-  };
-  parse: (content: unknown) => T;
+async function enqueueTask(params: {
+  userId: number;
+  type: "media_finalize" | "training_plan_generate" | "training_plan_adjust" | "analysis_corrections" | "pose_correction_multimodal";
+  title: string;
+  payload: Record<string, unknown>;
+  message: string;
 }) {
-  let lastError: unknown;
+  const taskId = nanoid();
+  await db.createBackgroundTask({
+    id: taskId,
+    userId: params.userId,
+    type: params.type,
+    title: params.title,
+    message: params.message,
+    payload: params.payload,
+    progress: 0,
+    maxAttempts: params.type === "media_finalize" ? 90 : 3,
+  });

-  for (let attempt = 0; attempt < 3; attempt++) {
-    const retryHint =
-      attempt === 0 || !(lastError instanceof Error)
-        ? []
-        : [{
-            role: "user" as const,
-            content:
-              `上一次输出无法被系统解析，错误是：${lastError.message}。` +
-              "请只返回一个合法、完整、可解析的 JSON 对象，不要包含额外说明、注释或 Markdown 代码块。",
-          }];
-
-    const response = await invokeLLM({
-      messages: [...params.baseMessages, ...retryHint],
-      response_format: params.responseFormat,
-    });
-
-    try {
-      return params.parse(response.choices[0]?.message?.content);
-    } catch (error) {
-      lastError = error;
-    }
-  }
-
-  throw lastError instanceof Error ? lastError : new Error("Failed to parse structured LLM response");
+  const task = await db.getBackgroundTaskById(taskId);
+  return { taskId, task };
 }

 export const appRouter = router({
@@ -104,86 +85,13 @@ export const appRouter = router({
        focusAreas: z.array(z.string()).optional(),
      }))
      .mutation(async ({ ctx, input }) => {
-        const user = ctx.user;
-        // Get user's recent analyses for personalization
-        const analyses = await db.getUserAnalyses(user.id);
-        const recentScores = analyses.slice(0, 5).map(a => ({
-          score: a.overallScore,
-          issues: a.detectedIssues,
-          exerciseType: a.exerciseType,
-          shotCount: a.shotCount,
-          strokeConsistency: a.strokeConsistency,
-          footworkScore: a.footworkScore,
-        }));
-
-        const prompt = `你是一位网球教练。请为一位${
-          input.skillLevel === "beginner" ? "初级" : input.skillLevel === "intermediate" ? "中级" : "高级"
-        }水平的网球学员生成一个${input.durationDays}天的训练计划。
-
-要求：
- 只需要球拍，不需要球场和球网
- 包含影子挥拍、墙壁练习、脚步移动、体能训练等
- 每天训练30-60分钟
-${input.focusAreas?.length ? `- 重点关注: ${input.focusAreas.join(", ")}` : ""}
-${recentScores.length > 0 ? `- 用户最近的分析数据: ${JSON.stringify(recentScores)}` : ""}
-
-请返回JSON格式，包含每天的训练内容。`;
-
-        const parsed = await invokeStructuredPlan({
-          baseMessages: [
-            { role: "system", content: "你是网球训练计划生成器。返回严格的JSON格式。" },
-            { role: "user", content: prompt },
-          ],
-          responseFormat: {
-            type: "json_schema",
-            json_schema: {
-              name: "training_plan",
-              strict: true,
-              schema: {
-                type: "object",
-                properties: {
-                  title: { type: "string", description: "训练计划标题" },
-                  exercises: {
-                    type: "array",
-                    items: {
-                      type: "object",
-                      properties: {
-                        day: { type: "number" },
-                        name: { type: "string" },
-                        category: { type: "string" },
-                        duration: { type: "number", description: "分钟" },
-                        description: { type: "string" },
-                        tips: { type: "string" },
-                        sets: { type: "number" },
-                        reps: { type: "number" },
-                      },
-                      required: ["day", "name", "category", "duration", "description", "tips", "sets", "reps"],
-                      additionalProperties: false,
-                    },
-                  },
-                },
-                required: ["title", "exercises"],
-                additionalProperties: false,
-              },
-            },
-          },
-          parse: (content) => normalizeTrainingPlanResponse({
-            content,
-            fallbackTitle: `${input.durationDays}天训练计划`,
-          }),
+        return enqueueTask({
+          userId: ctx.user.id,
+          type: "training_plan_generate",
+          title: `${input.durationDays}天训练计划生成`,
+          message: "训练计划已加入后台队列",
+          payload: input,
        });
-
-        const planId = await db.createTrainingPlan({
-          userId: user.id,
-          title: parsed.title,
-          skillLevel: input.skillLevel,
-          durationDays: input.durationDays,
-          exercises: parsed.exercises,
-          isActive: 1,
-          version: 1,
-        });
-
-        return { planId, plan: parsed };
      }),

    list: protectedProcedure.query(async ({ ctx }) => {
@@ -197,78 +105,15 @@ ${recentScores.length > 0 ? `- 用户最近的分析数据: ${JSON.stringify(rec
    adjust: protectedProcedure
      .input(z.object({ planId: z.number() }))
      .mutation(async ({ ctx, input }) => {
-        const analyses = await db.getUserAnalyses(ctx.user.id);
-        const recentAnalyses = analyses.slice(0, 5);
        const currentPlan = (await db.getUserTrainingPlans(ctx.user.id)).find(p => p.id === input.planId);
        if (!currentPlan) throw new Error("Plan not found");
-
-        const prompt = `基于以下用户的姿势分析结果，调整训练计划：
-
-当前计划: ${JSON.stringify(currentPlan.exercises)}
-最近分析结果: ${JSON.stringify(recentAnalyses.map(a => ({
-          score: a.overallScore,
-          issues: a.detectedIssues,
-          corrections: a.corrections,
-          shotCount: a.shotCount,
-          strokeConsistency: a.strokeConsistency,
-          footworkScore: a.footworkScore,
-          fluidityScore: a.fluidityScore,
-        })))}
-
-请根据分析结果调整训练计划，增加针对薄弱环节的训练，返回与原计划相同格式的JSON。`;
-
-        const parsed = await invokeStructuredPlan({
-          baseMessages: [
-            { role: "system", content: "你是网球训练计划调整器。返回严格的JSON格式。" },
-            { role: "user", content: prompt },
-          ],
-          responseFormat: {
-            type: "json_schema",
-            json_schema: {
-              name: "adjusted_plan",
-              strict: true,
-              schema: {
-                type: "object",
-                properties: {
-                  title: { type: "string" },
-                  adjustmentNotes: { type: "string", description: "调整说明" },
-                  exercises: {
-                    type: "array",
-                    items: {
-                      type: "object",
-                      properties: {
-                        day: { type: "number" },
-                        name: { type: "string" },
-                        category: { type: "string" },
-                        duration: { type: "number" },
-                        description: { type: "string" },
-                        tips: { type: "string" },
-                        sets: { type: "number" },
-                        reps: { type: "number" },
-                      },
-                      required: ["day", "name", "category", "duration", "description", "tips", "sets", "reps"],
-                      additionalProperties: false,
-                    },
-                  },
-                },
-                required: ["title", "adjustmentNotes", "exercises"],
-                additionalProperties: false,
-              },
-            },
-          },
-          parse: (content) => normalizeAdjustedPlanResponse({
-            content,
-            fallbackTitle: currentPlan.title,
-          }),
+        return enqueueTask({
+          userId: ctx.user.id,
+          type: "training_plan_adjust",
+          title: `${currentPlan.title} 调整`,
+          message: "训练计划调整任务已加入后台队列",
+          payload: input,
        });
-
-        await db.updateTrainingPlan(input.planId, {
-          exercises: parsed.exercises,
-          adjustmentNotes: parsed.adjustmentNotes,
-          version: (currentPlan.version || 1) + 1,
-        });
-
-        return { success: true, adjustmentNotes: parsed.adjustmentNotes };
      }),
  }),

@@ -287,19 +132,20 @@ ${recentScores.length > 0 ? `- 用户最近的分析数据: ${JSON.stringify(rec
        const fileKey = `videos/${ctx.user.id}/${nanoid()}.${input.format}`;
        const contentType = input.format === "webm" ? "video/webm" : "video/mp4";
        const { url } = await storagePut(fileKey, fileBuffer, contentType);
+        const publicUrl = toPublicUrl(url);

        const videoId = await db.createVideo({
          userId: ctx.user.id,
          title: input.title,
          fileKey,
-          url,
+          url: publicUrl,
          format: input.format,
          fileSize: input.fileSize,
          exerciseType: input.exerciseType || null,
          analysisStatus: "pending",
        });

-        return { videoId, url };
+        return { videoId, url: publicUrl };
      }),

    registerExternal: protectedProcedure
@@ -313,11 +159,12 @@ ${recentScores.length > 0 ? `- 用户最近的分析数据: ${JSON.stringify(rec
        exerciseType: z.string().optional(),
      }))
      .mutation(async ({ ctx, input }) => {
+        const publicUrl = toPublicUrl(input.url);
        const videoId = await db.createVideo({
          userId: ctx.user.id,
          title: input.title,
          fileKey: input.fileKey,
-          url: input.url,
+          url: publicUrl,
          format: input.format,
          fileSize: input.fileSize ?? null,
          duration: input.duration ?? null,
@@ -325,7 +172,7 @@ ${recentScores.length > 0 ? `- 用户最近的分析数据: ${JSON.stringify(rec
          analysisStatus: "completed",
        });

-        return { videoId, url: input.url };
+        return { videoId, url: publicUrl };
      }),

    list: protectedProcedure.query(async ({ ctx }) => {
@@ -399,32 +246,70 @@ ${recentScores.length > 0 ? `- 用户最近的分析数据: ${JSON.stringify(rec
        poseMetrics: z.any(),
        exerciseType: z.string(),
        detectedIssues: z.any(),
+        imageUrls: z.array(z.string()).optional(),
+        imageDataUrls: z.array(z.string()).max(4).optional(),
      }))
-      .mutation(async ({ input }) => {
-        const response = await invokeLLM({
-          messages: [
-            {
-              role: "system",
-              content: "你是一位网球动作分析员。根据MediaPipe姿势分析数据，给出具体的姿势矫正建议。用中文回答。",
-            },
-            {
-              role: "user",
-              content: `分析以下网球动作数据并给出矫正建议：
-动作类型: ${input.exerciseType}
-姿势指标: ${JSON.stringify(input.poseMetrics)}
-检测到的问题: ${JSON.stringify(input.detectedIssues)}
-
-请给出：
-1. 每个问题的具体矫正方法
-2. 推荐的练习动作
-3. 需要注意的关键点`,
-            },
-          ],
+      .mutation(async ({ ctx, input }) => {
+        const imageUrls = await prepareCorrectionImageUrls({
+          userId: ctx.user.id,
+          imageUrls: input.imageUrls,
+          imageDataUrls: input.imageDataUrls,
        });

-        return {
-          corrections: response.choices[0]?.message?.content || "暂无建议",
-        };
+        return enqueueTask({
+          userId: ctx.user.id,
+          type: imageUrls.length > 0 ? "pose_correction_multimodal" : "analysis_corrections",
+          title: `${input.exerciseType} 动作纠正`,
+          message: imageUrls.length > 0 ? "多模态动作纠正任务已加入后台队列" : "动作纠正任务已加入后台队列",
+          payload: {
+            poseMetrics: input.poseMetrics,
+            exerciseType: input.exerciseType,
+            detectedIssues: input.detectedIssues,
+            imageUrls,
+          },
+        });
+      }),
+  }),
+
+  task: router({
+    list: protectedProcedure
+      .input(z.object({ limit: z.number().min(1).max(50).default(20) }).optional())
+      .query(async ({ ctx, input }) => {
+        return db.listUserBackgroundTasks(ctx.user.id, input?.limit ?? 20);
+      }),
+
+    get: protectedProcedure
+      .input(z.object({ taskId: z.string().min(1) }))
+      .query(async ({ ctx, input }) => {
+        return db.getUserBackgroundTaskById(ctx.user.id, input.taskId);
+      }),
+
+    retry: protectedProcedure
+      .input(z.object({ taskId: z.string().min(1) }))
+      .mutation(async ({ ctx, input }) => {
+        const task = await db.retryBackgroundTask(ctx.user.id, input.taskId);
+        return { task };
+      }),
+
+    createMediaFinalize: protectedProcedure
+      .input(z.object({
+        sessionId: z.string().min(1),
+        title: z.string().min(1).max(256),
+        exerciseType: z.string().optional(),
+      }))
+      .mutation(async ({ ctx, input }) => {
+        const session = await getRemoteMediaSession(input.sessionId);
+        if (session.userId !== String(ctx.user.id)) {
+          throw new Error("Media session not found");
+        }
+
+        return enqueueTask({
+          userId: ctx.user.id,
+          type: "media_finalize",
+          title: `${input.title} 归档`,
+          message: "录制文件归档任务已加入后台队列",
+          payload: input,
+        });
      }),
  }),

--- a/server/storage.test.ts
+++ b/server/storage.test.ts
@@ -37,4 +37,16 @@ describe("storage fallback", () => {
      url: "/uploads/videos/test/sample.webm",
    });
  });
+
+  it("builds externally accessible URLs for local assets", async () => {
+    process.env.APP_PUBLIC_BASE_URL = "https://te.hao.work/";
+    const { toExternalAssetUrl } = await import("./storage");
+
+    expect(toExternalAssetUrl("/uploads/videos/test/sample.webm")).toBe(
+      "https://te.hao.work/uploads/videos/test/sample.webm"
+    );
+    expect(toExternalAssetUrl("https://cdn.example.com/demo.jpg")).toBe(
+      "https://cdn.example.com/demo.jpg"
+    );
+  });
 });
--- a/server/storage.ts
+++ b/server/storage.ts
@@ -4,6 +4,7 @@
 import { mkdir, readFile, writeFile } from "node:fs/promises";
 import path from "node:path";
 import { ENV } from './_core/env';
+import { toPublicUrl } from "./publicUrl";

 type StorageConfig = { baseUrl: string; apiKey: string };

@@ -141,3 +142,7 @@ export async function storageGet(relKey: string): Promise<{ key: string; url: st
    url: await buildDownloadUrl(baseUrl, key, apiKey),
  };
 }
+
+export function toExternalAssetUrl(pathOrUrl: string) {
+  return toPublicUrl(pathOrUrl);
+}
--- a/server/taskWorker.ts
+++ b/server/taskWorker.ts
@@ -0,0 +1,470 @@
+import { nanoid } from "nanoid";
+import { ENV } from "./_core/env";
+import { invokeLLM, type Message } from "./_core/llm";
+import * as db from "./db";
+import { getRemoteMediaSession } from "./mediaService";
+import {
+  buildAdjustedTrainingPlanPrompt,
+  buildMultimodalCorrectionPrompt,
+  buildTextCorrectionPrompt,
+  buildTrainingPlanPrompt,
+  multimodalCorrectionSchema,
+  renderMultimodalCorrectionMarkdown,
+} from "./prompts";
+import { toPublicUrl } from "./publicUrl";
+import { storagePut } from "./storage";
+import {
+  normalizeAdjustedPlanResponse,
+  normalizeTrainingPlanResponse,
+} from "./trainingPlan";
+
+type TaskRow = Awaited<ReturnType<typeof db.getBackgroundTaskById>>;
+
+type StructuredParams<T> = {
+  model?: string;
+  baseMessages: Array<{ role: "system" | "user"; content: string | Message["content"] }>;
+  responseFormat: {
+    type: "json_schema";
+    json_schema: {
+      name: string;
+      strict: true;
+      schema: Record<string, unknown>;
+    };
+  };
+  parse: (content: unknown) => T;
+};
+
+async function invokeStructured<T>(params: StructuredParams<T>) {
+  let lastError: unknown;
+
+  for (let attempt = 0; attempt < 3; attempt++) {
+    const retryHint =
+      attempt === 0 || !(lastError instanceof Error)
+        ? []
+        : [{
+            role: "user" as const,
+            content:
+              `上一次输出无法被系统解析，错误是：${lastError.message}。` +
+              "请只返回合法完整的 JSON 对象，不要附加 Markdown 或说明。",
+          }];
+
+    const response = await invokeLLM({
+      apiUrl: params.model === ENV.llmVisionModel ? ENV.llmVisionApiUrl : undefined,
+      apiKey: params.model === ENV.llmVisionModel ? ENV.llmVisionApiKey : undefined,
+      model: params.model,
+      messages: [...params.baseMessages, ...retryHint],
+      response_format: params.responseFormat,
+    });
+
+    try {
+      return params.parse(response.choices[0]?.message?.content);
+    } catch (error) {
+      lastError = error;
+    }
+  }
+
+  throw lastError instanceof Error ? lastError : new Error("Failed to parse structured LLM response");
+}
+
+function parseDataUrl(input: string) {
+  const match = input.match(/^data:(.+?);base64,(.+)$/);
+  if (!match) {
+    throw new Error("Invalid image data URL");
+  }
+  return {
+    contentType: match[1],
+    buffer: Buffer.from(match[2], "base64"),
+  };
+}
+
+async function persistInlineImages(userId: number, imageDataUrls: string[]) {
+  const persistedUrls: string[] = [];
+  for (let index = 0; index < imageDataUrls.length; index++) {
+    const { contentType, buffer } = parseDataUrl(imageDataUrls[index]);
+    const extension = contentType.includes("png") ? "png" : "jpg";
+    const key = `analysis-images/${userId}/${nanoid()}.${extension}`;
+    const uploaded = await storagePut(key, buffer, contentType);
+    persistedUrls.push(toPublicUrl(uploaded.url));
+  }
+  return persistedUrls;
+}
+
+export async function prepareCorrectionImageUrls(input: {
+  userId: number;
+  imageUrls?: string[];
+  imageDataUrls?: string[];
+}) {
+  const directUrls = (input.imageUrls ?? []).map((item) => toPublicUrl(item));
+  const uploadedUrls = input.imageDataUrls?.length
+    ? await persistInlineImages(input.userId, input.imageDataUrls)
+    : [];
+  return [...directUrls, ...uploadedUrls];
+}
+
+async function runTrainingPlanGenerateTask(task: NonNullable<TaskRow>) {
+  const payload = task.payload as {
+    skillLevel: "beginner" | "intermediate" | "advanced";
+    durationDays: number;
+    focusAreas?: string[];
+  };
+  const analyses = await db.getUserAnalyses(task.userId);
+  const recentScores = analyses.slice(0, 5).map((analysis) => ({
+    score: analysis.overallScore ?? null,
+    issues: analysis.detectedIssues,
+    exerciseType: analysis.exerciseType ?? null,
+    shotCount: analysis.shotCount ?? null,
+    strokeConsistency: analysis.strokeConsistency ?? null,
+    footworkScore: analysis.footworkScore ?? null,
+  }));
+
+  const parsed = await invokeStructured({
+    baseMessages: [
+      { role: "system", content: "你是网球训练计划生成器。返回严格的 JSON 格式。" },
+      {
+        role: "user",
+        content: buildTrainingPlanPrompt({
+          ...payload,
+          recentScores,
+        }),
+      },
+    ],
+    responseFormat: {
+      type: "json_schema",
+      json_schema: {
+        name: "training_plan",
+        strict: true,
+        schema: {
+          type: "object",
+          properties: {
+            title: { type: "string" },
+            exercises: {
+              type: "array",
+              items: {
+                type: "object",
+                properties: {
+                  day: { type: "number" },
+                  name: { type: "string" },
+                  category: { type: "string" },
+                  duration: { type: "number" },
+                  description: { type: "string" },
+                  tips: { type: "string" },
+                  sets: { type: "number" },
+                  reps: { type: "number" },
+                },
+                required: ["day", "name", "category", "duration", "description", "tips", "sets", "reps"],
+                additionalProperties: false,
+              },
+            },
+          },
+          required: ["title", "exercises"],
+          additionalProperties: false,
+        },
+      },
+    },
+    parse: (content) => normalizeTrainingPlanResponse({
+      content,
+      fallbackTitle: `${payload.durationDays}天训练计划`,
+    }),
+  });
+
+  const planId = await db.createTrainingPlan({
+    userId: task.userId,
+    title: parsed.title,
+    skillLevel: payload.skillLevel,
+    durationDays: payload.durationDays,
+    exercises: parsed.exercises,
+    isActive: 1,
+    version: 1,
+  });
+
+  return {
+    kind: "training_plan_generate" as const,
+    planId,
+    plan: parsed,
+  };
+}
+
+async function runTrainingPlanAdjustTask(task: NonNullable<TaskRow>) {
+  const payload = task.payload as { planId: number };
+  const analyses = await db.getUserAnalyses(task.userId);
+  const recentAnalyses = analyses.slice(0, 5);
+  const currentPlan = (await db.getUserTrainingPlans(task.userId)).find((plan) => plan.id === payload.planId);
+
+  if (!currentPlan) {
+    throw new Error("Plan not found");
+  }
+
+  const parsed = await invokeStructured({
+    baseMessages: [
+      { role: "system", content: "你是网球训练计划调整器。返回严格的 JSON 格式。" },
+      {
+        role: "user",
+        content: buildAdjustedTrainingPlanPrompt({
+          currentExercises: currentPlan.exercises,
+          recentAnalyses: recentAnalyses.map((analysis) => ({
+            score: analysis.overallScore ?? null,
+            issues: analysis.detectedIssues,
+            corrections: analysis.corrections,
+            shotCount: analysis.shotCount ?? null,
+            strokeConsistency: analysis.strokeConsistency ?? null,
+            footworkScore: analysis.footworkScore ?? null,
+            fluidityScore: analysis.fluidityScore ?? null,
+          })),
+        }),
+      },
+    ],
+    responseFormat: {
+      type: "json_schema",
+      json_schema: {
+        name: "adjusted_plan",
+        strict: true,
+        schema: {
+          type: "object",
+          properties: {
+            title: { type: "string" },
+            adjustmentNotes: { type: "string" },
+            exercises: {
+              type: "array",
+              items: {
+                type: "object",
+                properties: {
+                  day: { type: "number" },
+                  name: { type: "string" },
+                  category: { type: "string" },
+                  duration: { type: "number" },
+                  description: { type: "string" },
+                  tips: { type: "string" },
+                  sets: { type: "number" },
+                  reps: { type: "number" },
+                },
+                required: ["day", "name", "category", "duration", "description", "tips", "sets", "reps"],
+                additionalProperties: false,
+              },
+            },
+          },
+          required: ["title", "adjustmentNotes", "exercises"],
+          additionalProperties: false,
+        },
+      },
+    },
+    parse: (content) => normalizeAdjustedPlanResponse({
+      content,
+      fallbackTitle: currentPlan.title,
+    }),
+  });
+
+  await db.updateTrainingPlan(payload.planId, {
+    exercises: parsed.exercises,
+    adjustmentNotes: parsed.adjustmentNotes,
+    version: (currentPlan.version || 1) + 1,
+  });
+
+  return {
+    kind: "training_plan_adjust" as const,
+    planId: payload.planId,
+    plan: parsed,
+    adjustmentNotes: parsed.adjustmentNotes,
+  };
+}
+
+async function runTextCorrectionTask(task: NonNullable<TaskRow>) {
+  const payload = task.payload as {
+    exerciseType: string;
+    poseMetrics: unknown;
+    detectedIssues: unknown;
+  };
+  return createTextCorrectionResult(payload);
+}
+
+async function createTextCorrectionResult(payload: {
+  exerciseType: string;
+  poseMetrics: unknown;
+  detectedIssues: unknown;
+}) {
+  const response = await invokeLLM({
+    messages: [
+      {
+        role: "system",
+        content: "你是一位专业网球技术教练。输出中文 Markdown，内容具体、克制、可执行。",
+      },
+      {
+        role: "user",
+        content: buildTextCorrectionPrompt(payload),
+      },
+    ],
+  });
+
+  return {
+    kind: "analysis_corrections" as const,
+    corrections: response.choices[0]?.message?.content || "暂无建议",
+  };
+}
+
+async function runMultimodalCorrectionTask(task: NonNullable<TaskRow>) {
+  const payload = task.payload as {
+    exerciseType: string;
+    poseMetrics: unknown;
+    detectedIssues: unknown;
+    imageUrls: string[];
+  };
+  try {
+    const report = await invokeStructured({
+      model: ENV.llmVisionModel,
+      baseMessages: [
+        { role: "system", content: "你是专业网球教练。请基于图片和结构化姿态指标输出严格 JSON。" },
+        {
+          role: "user",
+          content: [
+            { type: "text", text: buildMultimodalCorrectionPrompt({
+              exerciseType: payload.exerciseType,
+              poseMetrics: payload.poseMetrics,
+              detectedIssues: payload.detectedIssues,
+              imageCount: payload.imageUrls.length,
+            }) },
+            ...payload.imageUrls.map((url) => ({
+              type: "image_url" as const,
+              image_url: {
+                url,
+                detail: "high" as const,
+              },
+            })),
+          ],
+        },
+      ],
+      responseFormat: {
+        type: "json_schema",
+        json_schema: {
+          name: "pose_correction_multimodal",
+          strict: true,
+          schema: multimodalCorrectionSchema,
+        },
+      },
+      parse: (content) => {
+        if (typeof content === "string") {
+          return JSON.parse(content);
+        }
+        return content as Record<string, unknown>;
+      },
+    });
+
+    return {
+      kind: "pose_correction_multimodal" as const,
+      imageUrls: payload.imageUrls,
+      report,
+      corrections: renderMultimodalCorrectionMarkdown(report as Parameters<typeof renderMultimodalCorrectionMarkdown>[0]),
+      visionStatus: "ok" as const,
+    };
+  } catch (error) {
+    const fallback = await createTextCorrectionResult(payload);
+    return {
+      kind: "pose_correction_multimodal" as const,
+      imageUrls: payload.imageUrls,
+      report: null,
+      corrections: fallback.corrections,
+      visionStatus: "fallback" as const,
+      warning: error instanceof Error ? error.message : "Vision model unavailable",
+    };
+  }
+}
+
+async function runMediaFinalizeTask(task: NonNullable<TaskRow>) {
+  const payload = task.payload as {
+    sessionId: string;
+    title: string;
+    exerciseType?: string;
+  };
+  const session = await getRemoteMediaSession(payload.sessionId);
+
+  if (session.userId !== String(task.userId)) {
+    throw new Error("Media session does not belong to the task user");
+  }
+
+  if (session.archiveStatus === "queued") {
+    await db.rescheduleBackgroundTask(task.id, {
+      progress: 45,
+      message: "录制文件已入队，等待归档",
+      delayMs: 4_000,
+    });
+    return null;
+  }
+
+  if (session.archiveStatus === "processing") {
+    await db.rescheduleBackgroundTask(task.id, {
+      progress: 78,
+      message: "录制文件正在整理与转码",
+      delayMs: 4_000,
+    });
+    return null;
+  }
+
+  if (session.archiveStatus === "failed") {
+    throw new Error(session.lastError || "Media archive failed");
+  }
+
+  if (!session.playback.ready) {
+    await db.rescheduleBackgroundTask(task.id, {
+      progress: 70,
+      message: "等待回放文件就绪",
+      delayMs: 4_000,
+    });
+    return null;
+  }
+
+  const preferredUrl = session.playback.mp4Url || session.playback.webmUrl;
+  const format = session.playback.mp4Url ? "mp4" : "webm";
+  if (!preferredUrl) {
+    throw new Error("Media session did not expose a playback URL");
+  }
+
+  const fileKey = `media/sessions/${session.id}/recording.${format}`;
+  const existing = await db.getVideoByFileKey(task.userId, fileKey);
+  if (existing) {
+    return {
+      kind: "media_finalize" as const,
+      sessionId: session.id,
+      videoId: existing.id,
+      url: existing.url,
+      fileKey,
+      format,
+    };
+  }
+
+  const publicUrl = toPublicUrl(preferredUrl);
+  const videoId = await db.createVideo({
+    userId: task.userId,
+    title: payload.title || session.title,
+    fileKey,
+    url: publicUrl,
+    format,
+    fileSize: format === "mp4" ? (session.playback.mp4Size ?? null) : (session.playback.webmSize ?? null),
+    duration: null,
+    exerciseType: payload.exerciseType || "recording",
+    analysisStatus: "completed",
+  });
+
+  return {
+    kind: "media_finalize" as const,
+    sessionId: session.id,
+    videoId,
+    url: publicUrl,
+    fileKey,
+    format,
+  };
+}
+
+export async function processBackgroundTask(task: NonNullable<TaskRow>) {
+  switch (task.type) {
+    case "training_plan_generate":
+      return runTrainingPlanGenerateTask(task);
+    case "training_plan_adjust":
+      return runTrainingPlanAdjustTask(task);
+    case "analysis_corrections":
+      return runTextCorrectionTask(task);
+    case "pose_correction_multimodal":
+      return runMultimodalCorrectionTask(task);
+    case "media_finalize":
+      return runMediaFinalizeTask(task);
+    default:
+      throw new Error(`Unsupported task type: ${String(task.type)}`);
+  }
+}
--- a/server/worker.ts
+++ b/server/worker.ts
@@ -0,0 +1,47 @@
+import "dotenv/config";
+import { ENV } from "./_core/env";
+import * as db from "./db";
+import { processBackgroundTask } from "./taskWorker";
+
+const workerId = `app-worker-${process.pid}`;
+
+function sleep(ms: number) {
+  return new Promise((resolve) => setTimeout(resolve, ms));
+}
+
+async function workOnce() {
+  await db.requeueStaleBackgroundTasks(new Date(Date.now() - ENV.backgroundTaskStaleMs));
+
+  const task = await db.claimNextBackgroundTask(workerId);
+  if (!task) {
+    return false;
+  }
+
+  try {
+    const result = await processBackgroundTask(task);
+    if (result !== null) {
+      await db.completeBackgroundTask(task.id, result, "任务执行完成");
+    }
+  } catch (error) {
+    const message = error instanceof Error ? error.message : "Unknown background task error";
+    await db.failBackgroundTask(task.id, message);
+    console.error(`[worker] task ${task.id} failed:`, error);
+  }
+
+  return true;
+}
+
+async function main() {
+  console.log(`[worker] ${workerId} started`);
+  for (;;) {
+    const hasWorked = await workOnce();
+    if (!hasWorked) {
+      await sleep(ENV.backgroundTaskPollMs);
+    }
+  }
+}
+
+main().catch((error) => {
+  console.error("[worker] fatal error", error);
+  process.exit(1);
+});
--- a/tests/e2e/helpers/mockApp.ts
+++ b/tests/e2e/helpers/mockApp.ts
@@ -59,6 +59,7 @@ type MockAppState = {
  user: MockUser;
  videos: any[];
  analyses: any[];
+  tasks: any[];
  activePlan: {
    id: number;
    title: string;
@@ -79,6 +80,7 @@ type MockAppState = {
  } | null;
  mediaSession: MockMediaSession | null;
  nextVideoId: number;
+  nextTaskId: number;
  authMeNullResponsesAfterLogin: number;
 };

@@ -159,6 +161,32 @@ function buildMediaSession(user: MockUser, title: string): MockMediaSession {
  };
 }

+function createTask(state: MockAppState, input: {
+  type: string;
+  title: string;
+  status?: string;
+  progress?: number;
+  message?: string;
+  result?: any;
+  error?: string | null;
+}) {
+  const task = {
+    id: `task-${state.nextTaskId++}`,
+    userId: state.user.id,
+    type: input.type,
+    status: input.status ?? "succeeded",
+    title: input.title,
+    message: input.message ?? "任务执行完成",
+    progress: input.progress ?? 100,
+    result: input.result ?? null,
+    error: input.error ?? null,
+    createdAt: nowIso(),
+    updatedAt: nowIso(),
+  };
+  state.tasks = [task, ...state.tasks];
+  return task;
+}
+
 async function fulfillJson(route: Route, body: unknown) {
  await route.fulfill({
    status: 200,
@@ -218,11 +246,112 @@ async function handleTrpc(route: Route, state: MockAppState) {
            },
          ],
        };
-        return trpcResult({ planId: state.activePlan.id, plan: state.activePlan });
+        return trpcResult({
+          taskId: createTask(state, {
+            type: "training_plan_generate",
+            title: "7天训练计划生成",
+            result: {
+              kind: "training_plan_generate",
+              planId: state.activePlan.id,
+              plan: state.activePlan,
+            },
+          }).id,
+        });
+      case "plan.adjust":
+        return trpcResult({
+          taskId: createTask(state, {
+            type: "training_plan_adjust",
+            title: "训练计划调整",
+            result: {
+              kind: "training_plan_adjust",
+              adjustmentNotes: "已根据最近分析结果调整训练重点。",
+            },
+          }).id,
+        });
      case "video.list":
        return trpcResult(state.videos);
      case "analysis.list":
        return trpcResult(state.analyses);
+      case "task.list":
+        return trpcResult(state.tasks);
+      case "task.get": {
+        const rawInput = url.searchParams.get("input");
+        const parsedInput = rawInput ? JSON.parse(rawInput) : {};
+        const taskId = parsedInput.json?.taskId || parsedInput[0]?.json?.taskId;
+        return trpcResult(state.tasks.find((task) => task.id === taskId) || null);
+      }
+      case "task.retry": {
+        const rawInput = url.searchParams.get("input");
+        const parsedInput = rawInput ? JSON.parse(rawInput) : {};
+        const taskId = parsedInput.json?.taskId || parsedInput[0]?.json?.taskId;
+        const task = state.tasks.find((item) => item.id === taskId);
+        if (task) {
+          task.status = "succeeded";
+          task.progress = 100;
+          task.error = null;
+          task.message = "任务执行完成";
+        }
+        return trpcResult({ task });
+      }
+      case "task.createMediaFinalize": {
+        if (state.mediaSession) {
+          state.mediaSession.status = "archived";
+          state.mediaSession.archiveStatus = "completed";
+          state.mediaSession.playback = {
+            ready: true,
+            webmUrl: "/media/assets/sessions/session-e2e/recording.webm",
+            mp4Url: "/media/assets/sessions/session-e2e/recording.mp4",
+            webmSize: 2_400_000,
+            mp4Size: 1_800_000,
+            previewUrl: "/media/assets/sessions/session-e2e/recording.webm",
+          };
+          state.videos = [
+            {
+              id: state.nextVideoId++,
+              title: state.mediaSession.title,
+              url: state.mediaSession.playback.webmUrl,
+              format: "webm",
+              fileSize: state.mediaSession.playback.webmSize,
+              exerciseType: "recording",
+              analysisStatus: "completed",
+              createdAt: nowIso(),
+            },
+            ...state.videos,
+          ];
+        }
+        return trpcResult({
+          taskId: createTask(state, {
+            type: "media_finalize",
+            title: "录制归档",
+            result: {
+              kind: "media_finalize",
+              sessionId: state.mediaSession?.id,
+              videoId: state.videos[0]?.id,
+              url: state.videos[0]?.url,
+            },
+          }).id,
+        });
+      }
+      case "analysis.getCorrections":
+        return trpcResult({
+          taskId: createTask(state, {
+            type: "pose_correction_multimodal",
+            title: "动作纠正",
+            result: {
+              corrections: "## 动作概览\n整体节奏稳定，建议继续优化击球点前置。",
+              report: {
+                priorityFixes: [
+                  {
+                    title: "击球点前置",
+                    why: "击球点略靠后会影响挥拍连贯性。",
+                    howToPractice: "每组 8 次影子挥拍，刻意在身体前侧完成触球动作。",
+                    successMetric: "连续 3 组都能稳定在身体前侧完成挥拍。",
+                  },
+                ],
+              },
+            },
+          }).id,
+        });
      case "video.registerExternal":
        if (state.mediaSession?.playback.webmUrl || state.mediaSession?.playback.mp4Url) {
          state.videos = [
@@ -366,9 +495,11 @@ export async function installAppMocks(
        createdAt: nowIso(),
      },
    ],
+    tasks: [],
    activePlan: null,
    mediaSession: null,
    nextVideoId: 100,
+    nextTaskId: 1,
    authMeNullResponsesAfterLogin: options?.authMeNullResponsesAfterLogin ?? 0,
  };