tune(viral-video): VLM lite timeout 25s→45s (#2147)

Co-authored-by: xiaoxia <dev@xiaoxiajianji.com>
Co-committed-by: xiaoxia <dev@xiaoxiajianji.com>

.env.example

@@ -212,9 +212,14 @@ COSYVOICE_CLONE_MODEL=voice-enrollment
 
 DOUBAO_API_KEY=your-doubao-api-key
 DOUBAO_MODEL=doubao-seed-1-6-250615
+DOUBAO_FAST_MODEL=doubao-1-5-pro-32k-250115
 DOUBAO_BASE_URL=https://ark.cn-beijing.volces.com/api/v3
 DOUBAO_TIMEOUT=30
 DOUBAO_MAX_RETRIES=2
+# 视觉模型：pro 精度高，lite 速度快（viral-video 商品识别默认用 lite 提速）
+DOUBAO_VISION_MODEL=doubao-1-5-vision-pro-250328
+DOUBAO_VISION_LITE_MODEL=doubao-1-5-vision-lite-250315
+DOUBAO_VISION_USE_LITE=true
 
 # ==================== 积分/会员系统 (#1895) ====================
 # 积分系统总开关：默认 false（暂停积分系统）。

alembic/versions/090_viral_video_phase_msg.py

@@ -0,0 +1,35 @@
+"""viral video add phase_message column (#2134)
+
+Revision ID: 090_viral_video_phase_msg
+Revises: 089_viral_video_cols
+Create Date: 2026-10-02
+
+#2134 阶段细粒度提示：viral_video 表新增 phase_message 列（中文阶段提示文案）。
+current_stage 列已在之前版本存在，本迁移只补 phase_message。
+幂等 ADD COLUMN IF NOT EXISTS。
+"""
+
+import sqlalchemy as sa
+
+from alembic import op
+
+revision = "090_viral_video_phase_msg"
+down_revision = "089_viral_video_cols"
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    # SQLite/PostgreSQL 兼容的幂等添加列
+    conn = op.get_bind()
+    inspector = sa.inspect(conn)
+    cols = {c["name"] for c in inspector.get_columns("viral_video_jobs")}
+    if "phase_message" not in cols:
+        op.add_column(
+            "viral_video_jobs",
+            sa.Column("phase_message", sa.String(length=500), nullable=False, server_default=""),
+        )
+
+
+def downgrade() -> None:
+    op.drop_column("viral_video_jobs", "phase_message")

alembic/versions/091_viral_video_stage.py

@@ -0,0 +1,49 @@
+"""viral video add current_stage column (#2137 follow-up)
+
+Revision ID: 091_viral_video_stage
+Revises: 090_viral_video_phase_msg
+Create Date: 2026-10-02
+
+#2137 follow-up fix: 090 migration missed current_stage column on viral_video_jobs,
+causing UndefinedColumn errors and 500s on all authenticated viral-video endpoints.
+Idempotently add current_stage and double-check phase_message.
+"""
+
+import sqlalchemy as sa
+
+from alembic import op
+
+revision = "091_viral_video_stage"
+down_revision = "090_viral_video_phase_msg"
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    conn = op.get_bind()
+    inspector = sa.inspect(conn)
+    cols = {c["name"] for c in inspector.get_columns("viral_video_jobs")}
+    if "current_stage" not in cols:
+        op.add_column(
+            "viral_video_jobs",
+            sa.Column(
+                "current_stage",
+                sa.String(length=200),
+                nullable=False,
+                server_default="",
+            ),
+        )
+    if "phase_message" not in cols:
+        op.add_column(
+            "viral_video_jobs",
+            sa.Column(
+                "phase_message",
+                sa.String(length=500),
+                nullable=False,
+                server_default="",
+            ),
+        )
+
+
+def downgrade() -> None:
+    op.drop_column("viral_video_jobs", "current_stage")

alembic/versions/092_viral_video_heartbeat.py

@@ -0,0 +1,42 @@
+"""viral_video_jobs 增加 heartbeat_at 列（worker 心跳，用于僵尸任务超时回收）
+
+Revision ID: 092_viral_video_heartbeat
+Revises: 091_viral_video_stage
+Create Date: 2026-10-02
+"""
+
+import sqlalchemy as sa
+
+from alembic import op
+
+revision = "092_viral_video_heartbeat"
+down_revision = "091_viral_video_stage"
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    conn = op.get_bind()
+    inspector = sa.inspect(conn)
+    cols = {c["name"] for c in inspector.get_columns("viral_video_jobs")}
+    if "heartbeat_at" not in cols:
+        op.add_column("viral_video_jobs", sa.Column("heartbeat_at", sa.DateTime(), nullable=True))
+        op.execute(
+            "UPDATE viral_video_jobs SET heartbeat_at = updated_at " "WHERE status = 'running' AND heartbeat_at IS NULL"
+        )
+        try:
+            op.create_index("ix_viral_video_jobs_heartbeat_at", "viral_video_jobs", ["heartbeat_at"])
+        except Exception:
+            pass
+
+
+def downgrade() -> None:
+    conn = op.get_bind()
+    inspector = sa.inspect(conn)
+    cols = {c["name"] for c in inspector.get_columns("viral_video_jobs")}
+    if "heartbeat_at" in cols:
+        try:
+            op.drop_index("ix_viral_video_jobs_heartbeat_at", table_name="viral_video_jobs")
+        except Exception:
+            pass
+        op.drop_column("viral_video_jobs", "heartbeat_at")

apps/api/app/api/routes/viral_video.py

@@ -128,6 +128,8 @@ def _to_response(job) -> ViralVideoJobResponse:
         style_guide=job.style_guide,
         style_template_id=job.style_template_id,
         status=job.status,
+        current_stage=getattr(job, "current_stage", "") or "",
+        phase_message=getattr(job, "phase_message", "") or "",
         image_analysis=getattr(job, "image_analysis", None),
         storyboard=getattr(job, "storyboard", None),
         generated_copy_text=getattr(job, "generated_copy_text", "") or "",
@@ -398,28 +400,42 @@ def retry_viral_video_job(
     authenticated_user: AuthenticatedUser = Depends(get_current_user),
     session: Session = Depends(get_db_session),
 ) -> ViralVideoJobResponse:
-    """重试失败的爆款视频任务。"""
+    """重试失败的爆款视频任务（也支持对僵尸/超时 running 任务强制重置后重试）。"""
+    from datetime import datetime, timezone
+
     repo = _get_job_repo(session)
     job = repo.get(job_id)
     if job is None:
         raise HTTPException(status_code=404, detail="任务不存在")
     if job.user_id != authenticated_user.user.id:
         raise HTTPException(status_code=403, detail="无权操作此任务")
-    if job.status != ViralVideoStatus.FAILED:
-        raise HTTPException(status_code=409, detail="只有失败的任务可以重试")
+
+    # 判定是否为僵尸 running 任务：running 超过 10 分钟且心跳停止超过 2 分钟
+    now = datetime.now(timezone.utc)
+    is_stale_running = False
+    if job.status == ViralVideoStatus.RUNNING and job.started_at is not None:
+        hb = getattr(job, "heartbeat_at", None) or job.updated_at
+        if (now - job.started_at).total_seconds() > 10 * 60 and hb is not None and (now - hb).total_seconds() > 2 * 60:
+            is_stale_running = True
+
+    if job.status != ViralVideoStatus.FAILED and not is_stale_running:
+        raise HTTPException(status_code=409, detail="只有失败或超时的任务可以重试")
 
     # 重置状态
     job.retry_count += 1
     job.status = ViralVideoStatus.PENDING
-    job.error_msg = ""
+    job.error_msg = "" if not is_stale_running else "任务执行超时，已重置重试"
     job.started_at = None
     job.completed_at = None
+    job.current_stage = ""
+    job.phase_message = ""
+    job.heartbeat_at = None
     repo.update(job)
 
     # 重新入队
     try:
         celery_app.send_task("worker.run_viral_video_pipeline", args=[job.id])
-        logger.info("[爆款视频] 重试入队: job_id=%s retry_count=%d", job.id, job.retry_count)
+        logger.info("[爆款视频] 重试入队: job_id=%s retry_count=%d stale=%s", job.id, job.retry_count, is_stale_running)
     except Exception as e:
         logger.error("[爆款视频] 重试入队失败: %s", e, exc_info=True)
         job.mark_failed(f"重试入队失败: {e}")

apps/api/app/schemas/viral_video.py

@@ -201,6 +201,10 @@ class ViralVideoJobResponse(BaseModel):
     style_guide: dict | None = None
     style_template_id: str = ""
     status: str
+    current_stage: str = (
+        ""  # 细粒度阶段 snake_case（analyzing_images/parsing_intent/generating_script/reviewing/tts_synthesizing/rendering_video/uploading）
+    )
+    phase_message: str = ""  # 中文阶段提示文案（前端轮询/SSE 直接展示）
     image_analysis: dict | None = None
     # v1.6 编导脚本（推荐前端使用）
     copy_result: dict | None = None

apps/web/src/pages/viral-video/ViralVideoPage.tsx

@@ -9,8 +9,6 @@ import {
   CaretRightOutlined,
   VideoCameraOutlined,
   SoundOutlined,
-  BulbOutlined,
-  ThunderboltOutlined,
   ReloadOutlined,
   WarningFilled,
   SearchOutlined,
@@ -23,6 +21,7 @@ import {
   LoadingOutlined,
   CheckCircleFilled,
   EditOutlined,
+  HistoryOutlined,
 } from "@ant-design/icons"
 import { Select, Input, message } from "antd"
 import { uploadAssetDirect, getAssetLibraries, getAssetsByKind, type AssetItem } from "@/api/assets"
@@ -72,7 +71,7 @@ type UploadedImage = {
   localId: string
   url: string
   name: string
-  size: number
+  size: number | undefined
   uploading: boolean
   pct: number
   ossUrl?: string
@@ -122,6 +121,8 @@ type TabTask = {
   // STEP3 设置
   videoRatio: string
   videoModel: string
+  quality: string
+  upscale: string
   // UI 状态
   uiStep: UIStep
   imageAnalysis: ImageAnalysisResult | null
@@ -150,14 +151,61 @@ const TARGET_CUSTOMERS = [
   "都市蓝领",
   "都市银发",
 ]
-const PERSONAS = ["亲切邻家姐姐", "专业顾问", "搞笑段子手", "真实测评博主", "闺蜜种草", "行业专家"]
+const PERSONAS = [
+  "通用个人IP",
+  "老板型IP",
+  "专家型IP",
+  "顾问型IP",
+  "创始人IP",
+  "创业者IP",
+  "从业者经验派",
+  "避坑顾问型",
+  "知识科普型",
+  "测评种草型",
+]
 const STRUCTURES = [
-  "痛点开场→卖点→促单",
-  "反差对比→证据→促单",
-  "故事讲述→共鸣→促单",
-  "专家背书→数据→促单",
+  "反差破局+亮明观点+还原现状",
+  "痛点场景植入+给出通关秘籍",
+  "反常识开场+扎心现状+制造焦虑",
+  "核心难题+化解焦虑+描绘愿景",
+  "痛点前置+前后对比+给出期许",
+  "悬念拉满+揭露危害+明确目的",
+  "情绪共鸣+勾起好奇+拔高价值",
+  "颠覆认知+现状拆解+赋予价值",
+  "结果先行+消除顾虑+保留悬念",
+  "真实困境+认知反转+传授秘籍",
+  "反差事实+案例佐证+戳破真相",
+  "预埋悬念+标杆案例+引出核心",
+  "危机渲染+放大痛点+输出干货",
+  "逆袭开场+成功诀窍+低价引流",
+  "反常识观点+分析危害+描述现状",
+  "正反对比+解析原因+简易方法",
+  "真实场景+内心刻画+共情引导+展望未来",
+  "连续否定+代入感受+打破认知+精准戳痛",
+  "揭秘行业+类比施压+放大焦虑+精准锁客",
+]
+const PURPOSES = [
+  "获客引流",
+  "账号涨粉",
+  "活动通知",
+  "口播带货",
+  "促销转化",
+  "场景种草",
+  "痛点解决",
+  "功能演示",
+  "门店发现",
+  "到店实录",
+  "招牌体验",
+  "同城团购",
+  "品牌主张",
+  "品牌故事",
+  "生活方式",
+  "创意概念",
+  "对话短剧",
+  "反转短剧",
+  "悬念短剧",
+  "情绪短片",
 ]
-const PURPOSES = ["品牌种草", "新品上市", "促销转化", "活动引流"]
 const DURATIONS = [15, 20, 30, 45, 60]
 const RATIOS = [
   { v: "9:16", label: "9:16 竖屏（抖音/视频号）" },
@@ -168,6 +216,18 @@ const MODELS = [
   { v: "seedance-2.5", label: "Seedance 2.5（推荐）" },
   { v: "seedance-2.0", label: "Seedance 2.0" },
 ]
+const QUALITY_OPTIONS = [
+  { v: "480p", label: "480p（快速）" },
+  { v: "720p", label: "720p（清晰）" },
+  { v: "1080p", label: "1080p（高清）" },
+]
+const UPSCALE_OPTIONS = [
+  { v: "default", label: "默认" },
+  { v: "sd", label: "普清" },
+  { v: "hd", label: "高清" },
+  { v: "uhd", label: "超清" },
+  { v: "omnipotent", label: "全能" },
+]
 const VOICE_TIPS =
   "支持 MP3/WAV/M4A/AAC/OGG 格式，最大 10MB，时长 ≤30 秒。建议清晰人声、无背景音乐、环境安静；录音请保持距麦克风 15-20cm，音量适中。"
 
@@ -306,7 +366,8 @@ const MOCK_STORYBOARD: Storyboard = {
     "还在为餐桌选不到好桌子发愁？这张北美黑胡桃木餐桌，一家人坐下来吃饭刚刚好。全实木、无贴皮，纹理好看又耐刮。点小黄车，给家里添一张好桌子。",
 }
 
-const fmtSize = (bytes: number) => {
+const fmtSize = (bytes: number | undefined) => {
+  if (!bytes || bytes <= 0) return "--"
   if (bytes < 1024) return `${bytes}B`
   if (bytes < 1024 * 1024) return `${(bytes / 1024).toFixed(1)}KB`
   return `${(bytes / 1024 / 1024).toFixed(1)}MB`
@@ -327,11 +388,13 @@ const emptyTask = (id: string, title: string): TabTask => ({
   targetCustomer: "",
   language: "中文(普通话)",
   viralStructure: STRUCTURES[0],
-  marketingPurpose: PURPOSES[0],
+  marketingPurpose: "",
   duration: 15,
   persona: "",
   videoRatio: "9:16",
   videoModel: "seedance-2.5",
+  quality: "480p",
+  upscale: "default",
   uiStep: "step1_upload",
   imageAnalysis: null,
   storyboard: null,
@@ -343,6 +406,32 @@ const emptyTask = (id: string, title: string): TabTask => ({
 })
 
 /* ─────────── 主页面 ─────────── */
+
+function getCopyStageText(stage: string | undefined): string {
+  switch (stage) {
+    case "image_analysis":
+    case "video_analysis":
+    case "analyzing_images":
+      return "🔍 正在分析商品特征..."
+    case "intent_parsing":
+    case "generating_outline":
+    case "planning":
+      return "📋 正在构思脚本框架..."
+    case "script_generation":
+    case "generating_storyboard":
+    case "writing_shots":
+      return "🎬 正在生成分镜脚本..."
+    case "review":
+    case "generating_copy":
+    case "writing_copy":
+      return "✍️ 正在撰写口播文案..."
+    case "finalizing":
+      return "✨ 正在优化细节..."
+    default:
+      return "🎬 AI 正在创作视频分镜脚本..."
+  }
+}
+
 const ViralVideoPage: React.FC = () => {
   const [tasks, setTasks] = useState<TabTask[]>(() => [emptyTask("t1", "生成 1")])
   const [activeId, setActiveId] = useState<string>("t1")
@@ -493,8 +582,6 @@ const ViralVideoPage: React.FC = () => {
   )
 
   /* ── 图片 ── */
-  const canEditAssets = task.uiStep === "step1_upload" || task.uiStep === "failed"
-
   const appendImagesFromFiles = useCallback(
     async (files: FileList | File[]) => {
       const remaining = 9 - task.images.length
@@ -530,7 +617,15 @@ const ViralVideoPage: React.FC = () => {
           setTask((t) => ({
             ...t,
             images: t.images.map((im) =>
-              im.localId === item.localId ? { ...im, uploading: false, pct: 100, ossUrl: url } : im,
+              im.localId === item.localId
+                ? {
+                    ...im,
+                    uploading: false,
+                    pct: 100,
+                    ossUrl: url,
+                    size: item.file!.size || im.size,
+                  }
+                : im,
             ),
           }))
         } catch {
@@ -551,7 +646,7 @@ const ViralVideoPage: React.FC = () => {
         localId: `lib_${a.id}_${Math.random().toString(36).slice(2, 6)}`,
         url: a.thumbnail_url || a.file_url || "",
         name: a.name || "未命名",
-        size: a.file_size || 0,
+        size: a.file_size || undefined,
         uploading: false,
         pct: 100,
         ossUrl: a.file_url || undefined,
@@ -788,7 +883,9 @@ const ViralVideoPage: React.FC = () => {
         duration: task.duration,
         persona_id: task.persona || undefined,
         user_copy_text:
-          task.fusionLevel === "user_primary" && task.userCopy ? task.userCopy : undefined,
+          (task.fusionLevel === "user_primary" || task.fusionLevel === "ai_polish") && task.userCopy
+            ? task.userCopy
+            : undefined,
         fusion_level: task.fusionLevel,
         style_strength: task.refVideo?.ossUrl ? task.styleStrength : undefined,
         voice_id: task.refAudio?.source === "preset" ? task.refAudio.id : undefined,
@@ -895,10 +992,6 @@ const ViralVideoPage: React.FC = () => {
   const outputUrl = task.job?.output_url || task.job?.result_video_url
   const errMsg = task.videoError || task.job?.error_message || task.job?.error_msg
 
-  const canEditStep2 =
-    task.uiStep !== "step2_generating" &&
-    task.uiStep !== "step3_generating" &&
-    task.uiStep !== "step3_done"
   const step2Enabled =
     task.uiStep === "step1_done" ||
     task.uiStep === "step2_generating" ||
@@ -943,24 +1036,61 @@ const ViralVideoPage: React.FC = () => {
         </div>
         {products.map((p, i) => (
           <div key={i} className="vv-recog-item">
-            {p.name && (
+            <div className="vv-recog-line">
+              <span className="vv-recog-k">图片{i + 1}：</span>
+              <span>
+                {p.name || "未识别"}
+                {p.spec && <span className="vv-recog-meta">（{p.spec}）</span>}
+                {p.brand && <span className="vv-recog-meta"> · {p.brand}</span>}
+                {p.category && <span className="vv-recog-meta"> · {p.category}</span>}
+              </span>
+            </div>
+            {featureText(p.key_features ?? p.features) && (
               <div className="vv-recog-line">
-                <span className="vv-recog-k">商品：</span>
-                <span>{p.name}</span>
-                {p.spec && <span style={{ color: "#6b7280" }}>（{p.spec}）</span>}
-                {p.brand && <span style={{ color: "#6b7280" }}> · {p.brand}</span>}
+                <span className="vv-recog-k">核心特征：</span>
+                <span className="vv-recog-v">{featureText(p.key_features ?? p.features)}</span>
               </div>
             )}
-            {featureText(p.features) && (
+            {p.colors && p.colors.length > 0 && (
               <div className="vv-recog-line">
-                <span className="vv-recog-k">视觉特征：</span>
-                <span>{featureText(p.features)}</span>
+                <span className="vv-recog-k">主色调：</span>
+                <span className="vv-recog-v">{p.colors.join(" / ")}</span>
               </div>
             )}
-            {p.label_text && (
+            {p.material_or_texture && (
               <div className="vv-recog-line">
-                <span className="vv-recog-k">品牌/标签文字：</span>
-                <span>{p.label_text}</span>
+                <span className="vv-recog-k">材质/纹理：</span>
+                <span className="vv-recog-v">{p.material_or_texture}</span>
+              </div>
+            )}
+            {p.visual_style && (
+              <div className="vv-recog-line">
+                <span className="vv-recog-k">视觉风格：</span>
+                <span className="vv-recog-v">{p.visual_style}</span>
+              </div>
+            )}
+            {p.scene && (
+              <div className="vv-recog-line">
+                <span className="vv-recog-k">场景：</span>
+                <span className="vv-recog-v">{p.scene}</span>
+              </div>
+            )}
+            {p.target_audience_hint && (
+              <div className="vv-recog-line">
+                <span className="vv-recog-k">目标人群：</span>
+                <span className="vv-recog-v">{p.target_audience_hint}</span>
+              </div>
+            )}
+            {(p.text_on_image || p.label_text) && (
+              <div className="vv-recog-line">
+                <span className="vv-recog-k">包装文字：</span>
+                <span className="vv-recog-v">{p.text_on_image || p.label_text}</span>
+              </div>
+            )}
+            {p.selling_points && (
+              <div className="vv-recog-line">
+                <span className="vv-recog-k">卖点：</span>
+                <span className="vv-recog-v">{p.selling_points}</span>
               </div>
             )}
           </div>
@@ -1009,8 +1139,11 @@ const ViralVideoPage: React.FC = () => {
     if (task.uiStep === "step2_generating") {
       return (
         <div className="vv-copy-box vv-copy-loading">
-          <LoadingOutlined style={{ color: "#7c3aed", fontSize: 18, marginRight: 8 }} />
-          <span>AI 正在创作视频分镜脚本…</span>
+          <span className="vv-spinner" />
+          <span className="vv-copy-loading-text">
+            {task.job?.progress_message || getCopyStageText(task.job?.progress_stage)}
+          </span>
+          <span className="vv-copy-loading-hint">预计 10-30 秒，请稍候</span>
         </div>
       )
     }
@@ -1049,160 +1182,160 @@ const ViralVideoPage: React.FC = () => {
     ) {
       return null
     }
-    const locked = task.uiStep === "step3_generating" || task.uiStep === "step3_done"
+    const locked = false
     const isMockPreview = sb === MOCK_STORYBOARD
 
     return (
       <div className="vv-copy-box vv-storyboard">
-        <div className="vv-copy-head">
-          <EditOutlined style={{ color: "#7c3aed" }} />
-          <span className="vv-copy-title">AI 分镜脚本</span>
-          {isMockPreview && (
-            <span className="vv-copy-tag vv-copy-tag-preview">
-              示例预览 · 等后端输出新结构后自动替换
-            </span>
-          )}
-        </div>
-
-        {/* 1. 总览卡片 */}
-        <div className="vv-sb-overview">
-          <div className="vv-sb-ov-field">
-            <label>主题</label>
+        <div className="vv-sb-doc">
+          {/* 视频总览 */}
+          <h4 className="vv-sb-h">视频总览</h4>
+          <div className="vv-sb-kv">
+            <span className="vv-sb-k">整体主题：</span>
             <input
-              className="vv-input vv-sb-input"
+              className="vv-sb-inline vv-sb-inline-input"
               value={sb.overview.theme}
               disabled={locked}
               onChange={(e) => updateOverview({ theme: e.target.value })}
             />
           </div>
-          <div className="vv-sb-ov-field">
-            <label>总时长</label>
+          <div className="vv-sb-kv">
+            <span className="vv-sb-k">总时长：</span>
             <input
-              className="vv-input vv-sb-input"
+              className="vv-sb-inline vv-sb-inline-input"
               value={sb.overview.total_duration}
               disabled={locked}
               onChange={(e) => updateOverview({ total_duration: e.target.value })}
             />
           </div>
-          <div className="vv-sb-ov-field">
-            <label>画幅</label>
+          <div className="vv-sb-kv">
+            <span className="vv-sb-k">画幅：</span>
             <Select
-              className="vv-select vv-select-purple vv-sb-input"
-              style={{ width: "100%" }}
+              className="vv-select vv-sb-inline-select"
+              style={{ width: 200 }}
               value={sb.overview.aspect_ratio}
               disabled={locked}
               onChange={(v) => updateOverview({ aspect_ratio: v })}
-              options={RATIOS.map((r) => ({ value: r.v, label: r.label }))}
+              options={RATIOS.map((r) => ({ value: r.v, label: r.v }))}
+              variant="borderless"
             />
           </div>
-        </div>
 
-        {/* 2. 场景与光线 */}
-        <div className="vv-sb-section">
-          <div className="vv-sb-section-title">
-            <BulbOutlined style={{ color: "#7c3aed" }} /> 场景与光线
-          </div>
+          {/* 场景与光线 */}
+          <h4 className="vv-sb-h">场景与光线</h4>
           <textarea
-            className="vv-textarea"
+            className="vv-textarea vv-sb-doc-ta"
             value={sb.scene_and_lighting}
             disabled={locked}
             onChange={(e) => updateStoryboard({ scene_and_lighting: e.target.value })}
             placeholder="描述整体场景氛围、光线方向与色温…"
           />
-        </div>
 
-        {/* 3. 逐镜头卡片 */}
-        <div className="vv-sb-section">
-          <div className="vv-sb-section-title">
-            <VideoCameraOutlined style={{ color: "#7c3aed" }} /> 逐镜头拆解（{sb.shots.length} 镜）
-          </div>
-          <div className="vv-sb-shots">
+          {/* 逐镜头 */}
+          <h4 className="vv-sb-h">逐秒镜头拆解</h4>
+          <div className="vv-sb-doc-shots">
             {sb.shots.map((sh, idx) => {
               const refImg =
                 typeof sh.reference_image_index === "number"
                   ? task.images[sh.reference_image_index]
                   : undefined
               return (
-                <div key={idx} className="vv-sb-shot">
-                  <div className="vv-sb-shot-head">
-                    <span className="vv-sb-shot-num">镜头 {idx + 1}</span>
+                <div key={idx} className="vv-sb-doc-shot">
+                  <div className="vv-sb-doc-shot-head">
                     <input
-                      className="vv-sb-time"
+                      className="vv-sb-time-doc"
                       value={sh.time_range}
                       disabled={locked}
                       onChange={(e) => updateShot(idx, { time_range: e.target.value })}
+                      placeholder="0-3秒"
                     />
                   </div>
-                  <div className="vv-sb-shot-grid">
-                    <div className="vv-sb-cell">
-                      <label>景别 / 运镜</label>
-                      <textarea
-                        className="vv-textarea vv-sb-cell-ta"
-                        value={sh.shot_type_angle_movement}
-                        disabled={locked}
-                        onChange={(e) =>
-                          updateShot(idx, { shot_type_angle_movement: e.target.value })
-                        }
-                      />
-                    </div>
-                    <div className="vv-sb-cell vv-sb-cell-wide">
-                      <label>场景与对白</label>
-                      <textarea
-                        className="vv-textarea vv-sb-cell-ta"
-                        value={sh.scene_and_dialogue}
-                        disabled={locked}
-                        onChange={(e) => updateShot(idx, { scene_and_dialogue: e.target.value })}
-                      />
-                    </div>
-                    <div className="vv-sb-cell">
-                      <label>动作细节</label>
-                      <textarea
-                        className="vv-textarea vv-sb-cell-ta"
-                        value={sh.action_details}
-                        disabled={locked}
-                        onChange={(e) => updateShot(idx, { action_details: e.target.value })}
-                      />
-                    </div>
-                    <div className="vv-sb-cell">
-                      <label>音效 / BGM</label>
-                      <textarea
-                        className="vv-textarea vv-sb-cell-ta"
-                        value={sh.audio_bgm}
-                        disabled={locked}
-                        onChange={(e) => updateShot(idx, { audio_bgm: e.target.value })}
-                      />
-                    </div>
-                    <div className="vv-sb-cell">
-                      <label>转场</label>
-                      <textarea
-                        className="vv-textarea vv-sb-cell-ta"
-                        value={sh.transition}
-                        disabled={locked}
-                        onChange={(e) => updateShot(idx, { transition: e.target.value })}
-                      />
-                    </div>
-                    <div className="vv-sb-cell vv-sb-ref">
-                      <label>参考图</label>
-                      <div className="vv-sb-ref-box">
-                        {refImg?.url ? (
-                          <img src={refImg.url} alt={`ref-${idx}`} className="vv-sb-ref-img" />
-                        ) : (
-                          <span className="vv-sb-ref-empty">未指定</span>
+                  <div className="vv-sb-kv">
+                    <span className="vv-sb-k">景别/角度与运镜：</span>
+                    <textarea
+                      className="vv-textarea vv-sb-doc-ta vv-sb-doc-ta-sm"
+                      value={sh.shot_type_angle_movement}
+                      disabled={locked}
+                      onChange={(e) =>
+                        updateShot(idx, { shot_type_angle_movement: e.target.value })
+                      }
+                    />
+                  </div>
+                  <div className="vv-sb-kv">
+                    <span className="vv-sb-k">场景与对白：</span>
+                    <textarea
+                      className="vv-textarea vv-sb-doc-ta"
+                      value={sh.scene_and_dialogue}
+                      disabled={locked}
+                      onChange={(e) => updateShot(idx, { scene_and_dialogue: e.target.value })}
+                    />
+                  </div>
+                  <div className="vv-sb-kv">
+                    <span className="vv-sb-k">动作与真人细节：</span>
+                    <textarea
+                      className="vv-textarea vv-sb-doc-ta vv-sb-doc-ta-sm"
+                      value={sh.action_details}
+                      disabled={locked}
+                      onChange={(e) => updateShot(idx, { action_details: e.target.value })}
+                    />
+                  </div>
+                  <div className="vv-sb-kv">
+                    <span className="vv-sb-k">音效/BGM：</span>
+                    <textarea
+                      className="vv-textarea vv-sb-doc-ta vv-sb-doc-ta-sm"
+                      value={sh.audio_bgm}
+                      disabled={locked}
+                      onChange={(e) => updateShot(idx, { audio_bgm: e.target.value })}
+                    />
+                  </div>
+                  <div className="vv-sb-kv">
+                    <span className="vv-sb-k">转场：</span>
+                    <textarea
+                      className="vv-textarea vv-sb-doc-ta vv-sb-doc-ta-sm"
+                      value={sh.transition}
+                      disabled={locked}
+                      onChange={(e) => updateShot(idx, { transition: e.target.value })}
+                    />
+                  </div>
+                  <div className="vv-sb-kv vv-sb-kv-ref">
+                    <span className="vv-sb-k">参考图片：</span>
+                    {refImg?.url ? (
+                      <span className="vv-sb-ref-badge">
+                        <img src={refImg.url} alt={`ref-${idx}`} className="vv-sb-ref-thumb" />
+                        {!locked && (
+                          <button
+                            className="vv-sb-ref-x"
+                            onClick={() => updateShot(idx, { reference_image_index: undefined })}
+                            aria-label="移除参考图"
+                          >
+                            ×
+                          </button>
                         )}
-                      </div>
-                    </div>
+                      </span>
+                    ) : !locked && task.images.length > 0 ? (
+                      <span className="vv-sb-ref-pick">
+                        {task.images.map((im, i) => (
+                          <button
+                            key={im.localId}
+                            className="vv-sb-ref-opt"
+                            onClick={() => updateShot(idx, { reference_image_index: i })}
+                            title={`@图片${i + 1}`}
+                          >
+                            <img src={im.url} alt={`img${i}`} />
+                          </button>
+                        ))}
+                      </span>
+                    ) : (
+                      <span className="vv-sb-muted">未指定</span>
+                    )}
                   </div>
                 </div>
               )
             })}
           </div>
-        </div>
 
-        {/* 4. 硬性约束 */}
-        <div className="vv-sb-section">
-          <div className="vv-sb-section-title">
-            <ThunderboltOutlined style={{ color: "#7c3aed" }} /> 硬性约束
-          </div>
+          {/* 硬性约束 */}
+          <h4 className="vv-sb-h">硬性约束</h4>
           <div className="vv-sb-taglist">
             {sb.hard_constraints.map((c, i) => (
               <span key={i} className="vv-sb-tag vv-sb-tag-hard">
@@ -1229,13 +1362,9 @@ const ViralVideoPage: React.FC = () => {
               </button>
             )}
           </div>
-        </div>
 
-        {/* 5. 负面提示词 */}
-        <div className="vv-sb-section">
-          <div className="vv-sb-section-title">
-            <CloseOutlined style={{ color: "#ef4444" }} /> 负面提示词
-          </div>
+          {/* 负面提示词 */}
+          <h4 className="vv-sb-h">负面提示词</h4>
           <div className="vv-sb-taglist">
             {sb.negative_prompts.map((c, i) => (
               <span key={i} className="vv-sb-tag vv-sb-tag-neg">
@@ -1262,20 +1391,18 @@ const ViralVideoPage: React.FC = () => {
               </button>
             )}
           </div>
-        </div>
 
-        {/* 6. 完整口播稿（折叠） */}
-        <div className="vv-sb-section">
-          <button className="vv-sb-collapse" onClick={() => setVoOpen(!voOpen)}>
-            <span>
+          {/* 完整口播稿（可折叠） */}
+          <h4 className="vv-sb-h">
+            <button className="vv-sb-doc-collapse" onClick={() => setVoOpen(!voOpen)} type="button">
               <SoundOutlined style={{ color: "#7c3aed", marginRight: 6 }} />
               完整口播稿（TTS 合成使用）
-            </span>
-            <span className={`vv-sb-caret ${voOpen ? "open" : ""}`}>▾</span>
-          </button>
+              <span className={`vv-sb-caret ${voOpen ? "open" : ""}`}>▾</span>
+            </button>
+          </h4>
           {voOpen && (
             <textarea
-              className="vv-textarea vv-sb-vo"
+              className="vv-textarea vv-sb-doc-ta vv-sb-vo"
               value={sb.voiceover_script}
               disabled={locked}
               onChange={(e) => updateStoryboard({ voiceover_script: e.target.value })}
@@ -1284,16 +1411,18 @@ const ViralVideoPage: React.FC = () => {
           )}
         </div>
 
-        {/* 7. 重新生成按钮（ghost） */}
-        {task.uiStep === "step2_copy_ready" && (
-          <button
-            className="vv-btn vv-btn-ghost vv-btn-block"
-            onClick={handleGenerateCopy}
-            style={{ marginTop: 12 }}
-          >
-            <ReloadOutlined /> 重新生成文案
-          </button>
-        )}
+        <div className="vv-sb-actions">
+          {isMockPreview && (
+            <span className="vv-copy-tag vv-copy-tag-preview">
+              示例预览 · 后端新结构就绪后自动替换
+            </span>
+          )}
+          {task.uiStep === "step2_copy_ready" && (
+            <button className="vv-btn vv-btn-ghost" onClick={handleGenerateCopy}>
+              <ReloadOutlined /> 重新生成文案
+            </button>
+          )}
+        </div>
       </div>
     )
   }
@@ -1341,17 +1470,12 @@ const ViralVideoPage: React.FC = () => {
             <div className="vv-section-body" style={{ paddingTop: 4 }}>
               {/* 图片按钮 */}
               <div className="vv-action-row">
-                <button
-                  className="vv-action-btn"
-                  onClick={() => imgInputRef.current?.click()}
-                  disabled={!canEditAssets}
-                >
+                <button className="vv-action-btn" onClick={() => imgInputRef.current?.click()}>
                   <UploadOutlined /> 本地上传
                 </button>
                 <button
                   className="vv-action-btn"
                   onClick={() => setAssetPicker({ open: true, kind: "image", multiple: true })}
-                  disabled={!canEditAssets}
                 >
                   <FolderOpenOutlined /> 从素材库选择图片
                 </button>
@@ -1373,7 +1497,7 @@ const ViralVideoPage: React.FC = () => {
                   已选 <strong style={{ color: "#111827" }}>{task.images.length}</strong>/9
                 </span>
                 {task.images.length > 0 && (
-                  <button className="vv-link-btn" onClick={clearImages} disabled={!canEditAssets}>
+                  <button className="vv-link-btn" onClick={clearImages}>
                     清空图片
                   </button>
                 )}
@@ -1406,7 +1530,7 @@ const ViralVideoPage: React.FC = () => {
                         <button
                           className="vv-icon-btn"
                           onClick={() => moveImg(i, -1)}
-                          disabled={i === 0 || im.uploading || !canEditAssets}
+                          disabled={i === 0 || im.uploading}
                           title="上移"
                         >
                           <ArrowUpOutlined />
@@ -1414,7 +1538,7 @@ const ViralVideoPage: React.FC = () => {
                         <button
                           className="vv-icon-btn"
                           onClick={() => moveImg(i, 1)}
-                          disabled={i === task.images.length - 1 || im.uploading || !canEditAssets}
+                          disabled={i === task.images.length - 1 || im.uploading}
                           title="下移"
                         >
                           <ArrowDownOutlined />
@@ -1422,7 +1546,7 @@ const ViralVideoPage: React.FC = () => {
                         <button
                           className="vv-icon-btn vv-icon-btn-danger"
                           onClick={() => removeImg(im.localId)}
-                          disabled={im.uploading || !canEditAssets}
+                          disabled={!!im.uploading}
                           title="删除"
                         >
                           <CloseOutlined />
@@ -1441,11 +1565,7 @@ const ViralVideoPage: React.FC = () => {
                   <span className="vv-subblock-hint">最大 200MB · 最长 120 秒</span>
                 </div>
                 <div className="vv-action-row">
-                  <button
-                    className="vv-action-btn"
-                    onClick={() => videoInputRef.current?.click()}
-                    disabled={!canEditAssets}
-                  >
+                  <button className="vv-action-btn" onClick={() => videoInputRef.current?.click()}>
                     <UploadOutlined /> 本地上传视频
                   </button>
                   <button
@@ -1453,7 +1573,6 @@ const ViralVideoPage: React.FC = () => {
                     onClick={() =>
                       setTask({ voicePanel: task.voicePanel === "douyin" ? null : "douyin" })
                     }
-                    disabled={!canEditAssets}
                   >
                     <LinkOutlined /> 抖音链接识别
                   </button>
@@ -1504,11 +1623,7 @@ const ViralVideoPage: React.FC = () => {
                         </div>
                       )}
                     </div>
-                    <button
-                      className="vv-link-btn"
-                      onClick={removeRefVideo}
-                      disabled={!canEditAssets}
-                    >
+                    <button className="vv-link-btn" onClick={removeRefVideo}>
                       删除重传
                     </button>
                     {task.refVideo.ossUrl && (
@@ -1522,7 +1637,6 @@ const ViralVideoPage: React.FC = () => {
                               key={s.value}
                               className={`vv-pill ${task.styleStrength === s.value ? "active" : ""}`}
                               onClick={() => setTask({ styleStrength: s.value })}
-                              disabled={!canEditAssets}
                             >
                               {s.label}
                             </button>
@@ -1545,7 +1659,6 @@ const ViralVideoPage: React.FC = () => {
                   <button
                     className={`vv-audio-cell ${task.refAudio?.source === "preset" ? "selected" : ""}`}
                     onClick={() => setVoicePickerOpen(true)}
-                    disabled={!canEditAssets}
                   >
                     <FileTextOutlined className="vv-audio-ic" />
                     <span>选择内置音色</span>
@@ -1553,7 +1666,6 @@ const ViralVideoPage: React.FC = () => {
                   <button
                     className={`vv-audio-cell ${task.voicePanel === "upload" ? "active" : ""} ${task.refAudio?.source === "upload" ? "selected" : ""}`}
                     onClick={() => voiceInputRef.current?.click()}
-                    disabled={!canEditAssets}
                   >
                     <UploadOutlined className="vv-audio-ic" />
                     <span>本地上传</span>
@@ -1561,7 +1673,6 @@ const ViralVideoPage: React.FC = () => {
                   <button
                     className={`vv-audio-cell ${task.voicePanel === "library" ? "active" : ""} ${task.refAudio?.source === "library" ? "selected" : ""}`}
                     onClick={() => setAssetPicker({ open: true, kind: "voice", multiple: false })}
-                    disabled={!canEditAssets}
                   >
                     <FolderOpenOutlined className="vv-audio-ic" />
                     <span>从素材库选择</span>
@@ -1569,7 +1680,6 @@ const ViralVideoPage: React.FC = () => {
                   <button
                     className={`vv-audio-cell ${task.voicePanel === "record" ? "active" : ""} ${task.refAudio?.source === "clone" ? "selected" : ""}`}
                     onClick={() => setCloneModalOpen(true)}
-                    disabled={!canEditAssets}
                   >
                     <AudioMutedOutlined className="vv-audio-ic" />
                     <span>直接录音</span>
@@ -1589,11 +1699,7 @@ const ViralVideoPage: React.FC = () => {
                   <div className="vv-selected-audio">
                     <SoundOutlined style={{ color: "#7c3aed", marginRight: 6 }} />
                     <span style={{ flex: 1 }}>已选：{task.refAudio.name}</span>
-                    <button
-                      className="vv-link-btn"
-                      onClick={() => setTask({ refAudio: null })}
-                      disabled={!canEditAssets}
-                    >
+                    <button className="vv-link-btn" onClick={() => setTask({ refAudio: null })}>
                       移除
                     </button>
                   </div>
@@ -1689,7 +1795,7 @@ const ViralVideoPage: React.FC = () => {
         </div>
 
         {/* ═══════════ 中列 STEP 2 ═══════════ */}
-        <div className={`vv-col ${step2Enabled ? "" : "vv-col-disabled"}`}>
+        <div className="vv-col">
           <div className="vv-section">
             <div className="vv-step-subhead">
               <span className={`vv-step-badge ${step2Enabled ? "" : "vv-step-badge-dim"}`}>2</span>
@@ -1705,22 +1811,27 @@ const ViralVideoPage: React.FC = () => {
                     key={f.value}
                     className={`vv-fusion-btn ${task.fusionLevel === f.value ? "active" : ""}`}
                     onClick={() => setTask({ fusionLevel: f.value })}
-                    disabled={!canEditStep2 || !step2Enabled}
                   >
                     <strong>{f.label}</strong>
                     {f.desc}
                   </button>
                 ))}
               </div>
-              {task.fusionLevel === "user_primary" && (
+              {(task.fusionLevel === "ai_polish" || task.fusionLevel === "user_primary") && (
                 <div className="vv-form-row">
-                  <label className="vv-label">写下你的文案/草稿</label>
+                  <label className="vv-label">
+                    {task.fusionLevel === "ai_polish" ? "写下你的文案草稿" : "写下你的文案/草稿"}
+                  </label>
                   <textarea
+                    key={task.fusionLevel}
                     className="vv-textarea"
-                    placeholder="请输入你想要的原始文案，AI 将严格保留你的表达"
+                    placeholder={
+                      task.fusionLevel === "ai_polish"
+                        ? "请输入你的文案草稿，AI将在此基础上润色优化..."
+                        : "请输入你想要的原始文案，AI将严格保留你的表达..."
+                    }
                     value={task.userCopy}
                     onChange={(e) => setTask({ userCopy: e.target.value })}
-                    disabled={!canEditStep2 || !step2Enabled}
                   />
                 </div>
               )}
@@ -1736,77 +1847,71 @@ const ViralVideoPage: React.FC = () => {
                     placeholder="如：家具店、宠物店、装修公司"
                     value={task.industry}
                     onChange={(e) => setTask({ industry: e.target.value })}
-                    disabled={!canEditStep2 || !step2Enabled}
                   />
                 </div>
                 <div className="vv-form-row">
                   <label className="vv-label">目标客户</label>
                   <Select
-                    className="vv-select vv-select-purple"
+                    className="vv-select"
                     style={{ width: "100%" }}
                     placeholder="请选择目标客户"
                     value={task.targetCustomer || undefined}
                     onChange={(v) => setTask({ targetCustomer: v })}
                     options={TARGET_CUSTOMERS.map((i) => ({ value: i, label: i }))}
                     allowClear
-                    disabled={!canEditStep2 || !step2Enabled}
                   />
                 </div>
                 <div className="vv-form-row">
                   <label className="vv-label">台词语言</label>
                   <Select
-                    className="vv-select vv-select-purple"
+                    className="vv-select"
                     style={{ width: "100%" }}
                     value={task.language}
                     onChange={(v) => setTask({ language: v })}
                     options={LANGUAGES.map((i) => ({ value: i, label: i }))}
-                    disabled={!canEditStep2 || !step2Enabled}
                   />
                 </div>
                 <div className="vv-form-row">
                   <label className="vv-label">我的人设</label>
                   <Select
-                    className="vv-select vv-select-purple"
+                    className="vv-select"
                     style={{ width: "100%" }}
-                    placeholder="请选择人设"
+                    placeholder="请选择人设类型"
+                    optionFilterProp="children"
                     value={task.persona || undefined}
                     onChange={(v) => setTask({ persona: v })}
                     options={PERSONAS.map((i) => ({ value: i, label: i }))}
-                    allowClear
-                    disabled={!canEditStep2 || !step2Enabled}
                   />
                 </div>
                 <div className="vv-form-row">
                   <label className="vv-label">爆款结构</label>
                   <Select
-                    className="vv-select vv-select-purple"
+                    className="vv-select"
                     style={{ width: "100%" }}
                     value={task.viralStructure}
                     onChange={(v) => setTask({ viralStructure: v })}
                     options={STRUCTURES.map((i) => ({ value: i, label: i }))}
-                    disabled={!canEditStep2 || !step2Enabled}
                   />
                 </div>
                 <div className="vv-form-row">
                   <label className="vv-label">营销目的</label>
                   <Select
-                    className="vv-select vv-select-purple"
+                    className="vv-select"
                     style={{ width: "100%" }}
-                    value={task.marketingPurpose}
+                    placeholder="请选择营销目的"
+                    value={task.marketingPurpose || undefined}
                     onChange={(v) => setTask({ marketingPurpose: v })}
                     options={PURPOSES.map((i) => ({ value: i, label: i }))}
-                    disabled={!canEditStep2 || !step2Enabled}
                   />
                 </div>
                 <div className="vv-form-row" style={{ gridColumn: "1 / -1" }}>
                   <label className="vv-label">文案视频时长</label>
                   <Select
-                    className="vv-select vv-select-purple"
+                    className="vv-select"
                     style={{ width: "100%" }}
                     value={task.duration}
                     onChange={(v) => setTask({ duration: v })}
                     options={DURATIONS.map((n) => ({ value: n, label: `${n}秒` }))}
-                    disabled={!canEditStep2 || !step2Enabled}
                   />
                 </div>
               </div>
@@ -1860,43 +1965,68 @@ const ViralVideoPage: React.FC = () => {
         </div>
 
         {/* ═══════════ 右列 STEP 3 ═══════════ */}
-        <div className={`vv-col ${step3Enabled ? "" : "vv-col-disabled"}`}>
+        <div className="vv-col">
           <div className="vv-section">
-            <div className="vv-step-subhead">
-              <span className={`vv-step-badge ${step3Enabled ? "" : "vv-step-badge-dim"}`}>3</span>
-              <span className="vv-step-subtitle">STEP 3 视频设置并生成视频</span>
+            <div className="vv-step-subhead vv-step3-head">
+              <div className="vv-step3-head-left">
+                <span className={`vv-step-badge ${step3Enabled ? "" : "vv-step-badge-dim"}`}>
+                  3
+                </span>
+                <span className="vv-step-subtitle">STEP 3 生成视频</span>
+              </div>
+              {step3Enabled && (
+                <button
+                  className="vv-btn vv-btn-ghost vv-btn-xs vv-history-btn"
+                  onClick={() => message.info("历史作品功能开发中")}
+                >
+                  <HistoryOutlined /> 历史作品
+                </button>
+              )}
               {!step3Enabled && <span className="vv-step-lock">请先完成 STEP 2</span>}
             </div>
+            <div className="vv-step3-divider" />
             <div className="vv-section-body">
-              <div className="vv-form-row">
-                <label className="vv-label">视频比例</label>
-                <Select
-                  className="vv-select"
-                  style={{ width: "100%" }}
-                  value={task.videoRatio}
-                  onChange={(v) => setTask({ videoRatio: v })}
-                  options={RATIOS.map((r) => ({ value: r.v, label: r.label }))}
-                  disabled={
-                    !step3Enabled ||
-                    task.uiStep === "step3_generating" ||
-                    task.uiStep === "step3_done"
-                  }
-                />
-              </div>
-              <div className="vv-form-row">
-                <label className="vv-label">生成模型</label>
-                <Select
-                  className="vv-select"
-                  style={{ width: "100%" }}
-                  value={task.videoModel}
-                  onChange={(v) => setTask({ videoModel: v })}
-                  options={MODELS.map((m) => ({ value: m.v, label: m.label }))}
-                  disabled={
-                    !step3Enabled ||
-                    task.uiStep === "step3_generating" ||
-                    task.uiStep === "step3_done"
-                  }
-                />
+              <div className="vv-step3-grid">
+                <div className="vv-form-row">
+                  <label className="vv-label">视频比例</label>
+                  <Select
+                    className="vv-select vv-select-step3"
+                    style={{ width: "100%" }}
+                    value={task.videoRatio}
+                    onChange={(v) => setTask({ videoRatio: v })}
+                    options={RATIOS.map((r) => ({ value: r.v, label: r.v }))}
+                  />
+                </div>
+                <div className="vv-form-row">
+                  <label className="vv-label">视频模型</label>
+                  <Select
+                    className="vv-select vv-select-step3"
+                    style={{ width: "100%" }}
+                    value={task.videoModel}
+                    onChange={(v) => setTask({ videoModel: v })}
+                    options={MODELS.map((m) => ({ value: m.v, label: m.label }))}
+                  />
+                </div>
+                <div className="vv-form-row">
+                  <label className="vv-label">清晰度</label>
+                  <Select
+                    className="vv-select vv-select-step3"
+                    style={{ width: "100%" }}
+                    value={task.quality}
+                    onChange={(v) => setTask({ quality: v })}
+                    options={QUALITY_OPTIONS.map((m) => ({ value: m.v, label: m.label }))}
+                  />
+                </div>
+                <div className="vv-form-row">
+                  <label className="vv-label">成片超分</label>
+                  <Select
+                    className="vv-select vv-select-step3"
+                    style={{ width: "100%" }}
+                    value={task.upscale}
+                    onChange={(v) => setTask({ upscale: v })}
+                    options={UPSCALE_OPTIONS.map((m) => ({ value: m.v, label: m.label }))}
+                  />
+                </div>
               </div>
             </div>
           </div>

apps/web/src/pages/viral-video/components/PresetVoicePickerModal.tsx

@@ -32,9 +32,10 @@ interface Props {
   onConfirm: (voice: PresetVoice) => void
 }
 
-/** 兜底 mock 音色（仅当后端 /voices/presets 接口失败时使用，voice_id 必须与后端 CosyVoice 预置音色一致）
- *  参考: packages/domain/preset_voices.py — PRESET_VOICES（8 个 v3 系统音色） */
+/** 兜底 mock 音色（后端 /api/v1/tts/presets 返回字段不够时使用） */
 const MOCK_VOICES: PresetVoice[] = [
+  // ⚠️ 兜底 mock，仅在 /voices/presets 接口不可达时使用；ID 必须与后端
+  // packages/domain/preset_voices.py PRESET_VOICES 的 voice_id 对齐（v3后缀）
   {
     id: "longxiaochun_v3",
     name: "龙小淳",
@@ -157,8 +158,9 @@ const PresetVoicePickerModal: React.FC<Props> = ({
 
   // 合并真实数据和 mock：如果真实数据 gender/category 缺失，用 mock 兜底
   const allVoices: PresetVoice[] = useMemo(() => {
+    // 真实 API 返回的 voice_id 以 API 为准（如 longxiaochun_v3），前端不做硬编码覆盖
     const realList: PresetVoice[] = (voices || []).map((v) => {
-      // 按 voice_id 精确匹配 mock 获取补充信息
+      // 按 id 精确匹配 mock 获取补充元信息（id 即 voice_id，唯一稳定键）
       const mockMatch = MOCK_VOICES.find((m) => m.id === v.id)
       return {
         ...v,
@@ -166,12 +168,12 @@ const PresetVoicePickerModal: React.FC<Props> = ({
         category:
           v.category ||
           mockMatch?.category ||
-          (v.gender === "female" ? "女声" : v.gender === "male" ? "男声" : undefined),
+          (v.gender === "female" ? "女声" : v.gender === "male" ? "男声" : "其他"),
         desc: v.desc || mockMatch?.desc,
         sample_audio_url: v.sample_audio_url,
       }
     })
-    // 如果没有真实数据，使用 mock
+    // 如果没有真实数据，使用兜底 mock（接口失败时）
     return realList.length > 0 ? realList : MOCK_VOICES
   }, [voices])
 

apps/worker/worker_app/tasks/viral_video.py

@@ -23,6 +23,9 @@ import json
 import logging
 import os
 import tempfile
+import threading
+import time
+from concurrent.futures import ThreadPoolExecutor, as_completed
 from pathlib import Path
 
 from celery import Task, shared_task
@@ -40,6 +43,7 @@ from packages.domain.viral_video import (
     ViralVideoStage,
     ViralVideoStatus,
 )
+from packages.shared import get_shared_settings
 
 logger = logging.getLogger(__name__)
 
@@ -88,6 +92,112 @@ def _save_job(repo, job, session):
     session.commit()
 
 
+def _set_stage(job, repo, session, stage: str, message: str, persist: bool = True) -> None:
+    """更新细粒度阶段并持久化到 DB，同时通过 Redis 推送进度事件。
+
+    stage 用 ViralVideoStage.value（snake_case，与前端 phase 对齐）。
+    message 为中文提示文案，前端轮询/SSE 直接展示给用户。
+    """
+    job.current_stage = stage or ""
+    job.phase_message = message or ""
+    _emit_progress(job.id, stage, 0.0, message)
+    if persist and repo is not None and session is not None:
+        try:
+            _save_job(repo, job, session)
+        except Exception as e:  # 阶段持久化失败不阻塞主流程
+            logger.warning("[爆款视频] 阶段持久化失败 stage=%s err=%s", stage, e)
+
+
+# ── worker 心跳（僵尸任务检测） ─────────────────────────────────────────
+
+# 心跳间隔（秒）；超过此时间未更新 heartbeat_at 视为 worker 异常
+_HEARTBEAT_INTERVAL_SEC = 25
+# 任务整体超时：running 超过此时长且心跳停止，则判定为僵尸并回收
+_STALE_RUNNING_TIMEOUT_SEC = 10 * 60  # 10 分钟
+# 心跳过期窗口：heartbeat_at 距 now 超过此时长视为失效
+_HEARTBEAT_EXPIRE_SEC = 2 * 60  # 2 分钟
+
+
+def _heartbeat_once(job_id: str) -> None:
+    """在独立 session 中更新一次 heartbeat_at（不捕获主流程事务状态）。"""
+    ssn = None
+    try:
+        from datetime import datetime, timezone
+
+        ssn = SessionLocal()
+        ssn.execute(
+            __import__("sqlalchemy").text(
+                "UPDATE viral_video_jobs SET heartbeat_at = :now, updated_at = :now "
+                "WHERE id = :jid AND status = 'running'"
+            ),
+            {"now": datetime.now(timezone.utc), "jid": job_id},
+        )
+        ssn.commit()
+    except Exception as e:
+        logger.warning("[爆款视频] 心跳更新失败 job=%s err=%s", job_id, e)
+    finally:
+        if ssn is not None:
+            try:
+                ssn.close()
+            except Exception:
+                pass
+
+
+def _start_heartbeat_thread(job_id: str) -> tuple[threading.Event, threading.Thread]:
+    """启动后台心跳线程，每 _HEARTBEAT_INTERVAL_SEC 秒更新一次 heartbeat_at。
+    返回 (stop_event, thread)；任务结束时调用 stop_event.set() 停止心跳。
+    """
+    stop = threading.Event()
+
+    def _loop():
+        # 立即打一次心跳
+        _heartbeat_once(job_id)
+        while not stop.wait(_HEARTBEAT_INTERVAL_SEC):
+            _heartbeat_once(job_id)
+
+    t = threading.Thread(target=_loop, name=f"vv-heartbeat-{job_id[:8]}", daemon=True)
+    t.start()
+    return stop, t
+
+
+def _recover_stale_jobs() -> int:
+    """启动/定时扫描：把僵尸任务（running 超时且心跳停止）标记为 failed。
+    返回本次回收的任务数。可由 celery beat 周期性调用，也可在任务启动前顺带扫一次。
+    """
+    from datetime import datetime, timedelta, timezone
+
+    ssn = None
+    try:
+        ssn = SessionLocal()
+        now = datetime.now(timezone.utc)
+        # 判定条件：status=running 且 (started_at 距今 > 10min) 且 (heartbeat_at < now-2min 或 heartbeat_at IS NULL 且 updated_at < now-2min)
+        cutoff_beat = now - timedelta(seconds=_HEARTBEAT_EXPIRE_SEC)
+        cutoff_start = now - timedelta(seconds=_STALE_RUNNING_TIMEOUT_SEC)
+        sql = __import__("sqlalchemy").text(
+            "UPDATE viral_video_jobs "
+            "SET status='failed', error_msg='任务执行超时，请重试', updated_at=:now "
+            "WHERE status='running' "
+            "  AND started_at IS NOT NULL AND started_at < :cutoff_start "
+            "  AND (heartbeat_at IS NULL OR heartbeat_at < :cutoff_beat) "
+            "  AND (heartbeat_at IS NOT NULL OR updated_at < :cutoff_beat)"
+        )
+        result = ssn.execute(sql, {"now": now, "cutoff_start": cutoff_start, "cutoff_beat": cutoff_beat})
+        ssn.commit()
+        cnt = result.rowcount or 0
+        if cnt > 0:
+            logger.warning("[爆款视频] 回收 %d 个僵尸 running 任务", cnt)
+        return cnt
+    except Exception as e:
+        logger.warning("[爆款视频] 僵尸任务扫描失败: %s", e)
+        return 0
+    finally:
+        if ssn is not None:
+            try:
+                ssn.close()
+            except Exception:
+                pass
+
+
 # ── 默认结构 ─────────────────────────────────────────────────────────────
 
 _DEFAULT_HARD_CONSTRAINTS = [
@@ -136,37 +246,42 @@ def _empty_copy_result(duration: int = 15, ratio: str = "9:16") -> dict:
 # ── 流水线各步骤 ────────────────────────────────────────────────────────
 
 
-_IMAGE_ANALYSIS_PROMPT = """请仔细观察这张图片，只基于图片中真实可见的内容进行分析，不要凭空想象。
-
-必须输出严格的 JSON（不要 Markdown 代码块，不要额外解释），字段如下：
+_IMAGE_ANALYSIS_SYSTEM_PROMPT = """你是电商商品视觉分析师，从商品图片中提取关键商品信息。严格规则：
+1. 只说图片里真实可见的内容，看不清/没有的填「无法判断」，不要瞎猜。
+2. 输出必须是严格 JSON（不要 Markdown 代码块，不要额外解释文字）。
+3. 字段说明：
 {
-  "category": "产品大类，如护肤品/彩妆/食品/数码/服饰/家居等，若无法识别填『无法判断』",
-  "name": "产品名称（从包装/品牌/logo/文字推断；没有品牌时描述外观如『粉色包装面霜』）",
-  "brand": "品牌名（看 logo/包装文字；看不清填『未知』）",
-  "colors": ["主体颜色"],
-  "material_or_texture": "材质/质地描述（如玻璃瓶装/塑料软管/哑光质感/金属外壳等；无法判断填『无法判断』）",
+  "name": "商品全名（品牌+产品名+规格，如『大公鸡头管家 多功能油污净 625ml』，从包装 OCR 读出）",
+  "brand": "品牌名（从 Logo/包装文字读出，看不清填『无法判断』）",
+  "category": "商品品类（如『家用清洁/油污清洁剂』『日化/洗衣液』；非产品图填『非产品图』）",
+  "appearance": "外观特征（50-100字：瓶身形状、颜色、瓶盖、标签颜色、尺寸感）",
+  "packaging": "包装细节（50-100字：标签分区、图案元素、瓶盖/泵头样式、塑封状态）",
+  "text_on_package": ["包装上清晰可见的文字列表（品牌、产品名、卖点、规格等，看不清的不列）"],
   "key_features": [
-    "3-5 条**图片中确实能看到**的外观特征/卖点描述（如『按压式泵头』『瓶身有金色装饰线』等），不要编图片里没有的功效"
+    "3-5 条图片中能看到的外观/视觉特征（如『红色瓶盖白色瓶身』『鸡头图案 Logo』等）"
   ],
-  "visual_style": "视觉风格（如简约高端/粉嫩少女/国潮/科技感/生活方式实拍等）",
-  "scene": "图片中的使用/展示场景（如白底棚拍/浴室场景/户外街拍/桌面静物等；纯白底填『白底产品图』）",
-  "target_audience_hint": "从视觉推断的目标人群（如年轻女性/男性商务/亲子家庭等；不确定填『通用』）",
-  "text_on_image": "图片上出现的可读文字（品牌名/Slogan/产品名等，没有则填『无』）"
-}
+  "scene": "图片场景（如白底棚拍/浴室实拍/桌面静物/手持实拍等）",
+  "summary": "100-180字中文导购描述，连贯自然段落，像电商详情页介绍，前端直接展示，必须提到品牌/品名/核心外观特征，不能写『无法判断』"
+}"""
 
-严格要求：
-1. 任何字段无法确认时填『无法判断』或『未知』，不要猜。
-2. key_features 只能描述图片里肉眼可见的物理外观，不要写『补水保湿』『抗衰老』这类功效词（除非包装上明确印了）。
-3. 如果图片完全不是产品图（比如风景/人像/截图），category 填『非产品图』，name 填实际看到的内容。
-"""
+_IMAGE_ANALYSIS_USER_PROMPT = """请分析这张商品图片，输出严格 JSON。重点：
+1. name/brand/text_on_package 从图片包装 OCR 读取，不编造；
+2. appearance/packaging 各写 50-100 字，要具体；
+3. summary 必须是 100-180 字连贯中文段落，说清商品是什么、长什么样、适合谁用，不要写「无法判断」；
+4. 非产品图时 category 填「非产品图」，name 填实际看到的内容；
+5. 看不清的字段填「无法判断」。"""
 
 
 def _vision_fallback(idx: int, reason: str, extra: dict | None = None) -> dict:
     d = {
         "name": "未识别",
         "category": "无法判断",
+        "appearance": "无法判断",
+        "packaging": "无法判断",
+        "text_on_package": [],
         "key_features": [],
         "scene": "通用",
+        "summary": "",
         "_source": reason,
     }
     if extra:
@@ -174,10 +289,118 @@ def _vision_fallback(idx: int, reason: str, extra: dict | None = None) -> dict:
     return d
 
 
+def _is_vision_result_usable(result: dict) -> bool:
+    """判断 VLM 返回是否有效：name/summary 不能为未识别/无法判断/空，summary 要够长。"""
+    if not isinstance(result, dict):
+        return False
+    name = (result.get("name") or "").strip()
+    if not name or name in ("未识别", "无法判断", "未知"):
+        return False
+    summary = (result.get("summary") or "").strip()
+    if len(summary) < 30 or summary in ("无法判断", "未识别"):
+        return False
+    category = (result.get("category") or "").strip()
+    if category == "非产品图":
+        return True
+    feats = result.get("key_features") or []
+    if not isinstance(feats, list) or len(feats) == 0:
+        return False
+    return True
+
+
+def _analyze_single_image(
+    idx: int,
+    img_url: str,
+    vision_model: str,
+    timeout: int,
+    *,
+    pro_fallback_model: str | None = None,
+) -> dict:
+    """单张图片 VLM 分析（线程池并行调用）。
+    - lite 失败/结果不可用 时自动用 pro 模型降级重试 1 次。
+    - 失败/None/不可用最终返回含默认字段的 dict（不会让用户看到「未识别·无法判断」裸结果）。
+    """
+    from packages.shared.ai_service import call_vision
+
+    if not img_url or not isinstance(img_url, str):
+        return _vision_fallback(idx, "invalid_url")
+
+    def _call(model: str, tmo: int):
+        try:
+            return call_vision(
+                image_url=img_url,
+                prompt=_IMAGE_ANALYSIS_USER_PROMPT,
+                model=model,
+                max_tokens=800,
+                temperature=0.1,
+                timeout=tmo,
+                system_prompt=_IMAGE_ANALYSIS_SYSTEM_PROMPT,
+            )
+        except Exception as e:
+            logger.warning("[爆款视频] 图片 #%d call_vision(%s) 异常 err=%s", idx, model, e)
+            return None
+
+    def _normalize(raw, source: str) -> dict:
+        if raw is None:
+            return _vision_fallback(idx, f"{source}_none")
+        if isinstance(raw, str):
+            logger.warning("[爆款视频] 图片 #%d VLM(%s) 返回非 JSON: %s", idx, source, raw[:200])
+            return _vision_fallback(idx, f"{source}_text", {"_raw": raw[:500]})
+        if not isinstance(raw, dict):
+            return _vision_fallback(idx, f"{source}_badtype")
+        raw.setdefault("_source", source)
+        raw.setdefault("name", "未识别")
+        raw.setdefault("brand", "无法判断")
+        raw.setdefault("category", "无法判断")
+        raw.setdefault("appearance", "无法判断")
+        raw.setdefault("packaging", "无法判断")
+        raw.setdefault("text_on_package", [])
+        raw.setdefault("key_features", [])
+        raw.setdefault("scene", "通用")
+        raw.setdefault("summary", "")
+        if not isinstance(raw.get("text_on_package"), list):
+            raw["text_on_package"] = []
+        if not isinstance(raw.get("key_features"), list):
+            raw["key_features"] = []
+        return raw
+
+    # 第一次：传入模型（通常是 lite）
+    first_raw = _call(vision_model, timeout)
+    tag1 = vision_model.split("/")[-1] if "/" in vision_model else vision_model
+    first_result = _normalize(first_raw, tag1)
+    if _is_vision_result_usable(first_result):
+        return first_result
+
+    logger.warning(
+        "[爆款视频] 图片 #%d VLM(%s) 结果不可用 name=%r summary_len=%d，尝试 pro 降级",
+        idx,
+        vision_model,
+        first_result.get("name"),
+        len(first_result.get("summary") or ""),
+    )
+
+    # 第二次：pro 降级重试
+    if pro_fallback_model and pro_fallback_model != vision_model:
+        pro_raw = _call(pro_fallback_model, max(60, timeout))
+        pro_result = _normalize(pro_raw, "pro_fallback")
+        if _is_vision_result_usable(pro_result):
+            pro_result["_fallback_used"] = True
+            return pro_result
+        logger.warning(
+            "[爆款视频] 图片 #%d pro 降级仍不可用 name=%r summary_len=%d",
+            idx,
+            pro_result.get("name"),
+            len(pro_result.get("summary") or ""),
+        )
+        return pro_result
+
+    return first_result
+
+
 def _step_image_analysis(job: ViralVideoJob) -> dict:
-    """步骤 1: 图片 VLM 分析 — 识别产品特征、场景、卖点（加 None 防护）。"""
+    """步骤 1: 图片 VLM 分析 — 识别产品特征（v1.6 优化：并行 + lite 模型提速）。"""
     try:
-        from packages.shared.ai_service import call_vision
+        from packages.shared.ai_service import call_vision  # noqa: F401
     except ImportError:
         logger.warning("[爆款视频] ai_service.call_vision 不可用，使用占位结果")
         return {"products": [_vision_fallback(0, "fallback_import_error")]}
@@ -186,36 +409,46 @@ def _step_image_analysis(job: ViralVideoJob) -> dict:
         logger.warning("[爆款视频] 任务无 images，跳过图片分析")
         return {"products": []}
 
-    results = []
-    for idx, img_url in enumerate(job.images):
-        if not img_url or not isinstance(img_url, str):
-            logger.warning("[爆款视频] 图片 #%d URL 非法", idx)
-            results.append(_vision_fallback(idx, "invalid_url"))
-            continue
-        logger.info("[爆款视频] 图片分析 #%d img=%s", idx, img_url[:160])
-        try:
-            result = call_vision(image_url=img_url, prompt=_IMAGE_ANALYSIS_PROMPT)
-            if result is None:
-                logger.warning("[爆款视频] 图片 #%d call_vision 返回 None", idx)
-                results.append(_vision_fallback(idx, "vision_none"))
-            elif isinstance(result, str):
-                # VLM 返回了非 JSON 文本（JSON 解析失败），记录原始文本但不要让 None 传播
-                logger.warning("[爆款视频] 图片 #%d VLM 返回非 JSON 文本: %s", idx, result[:200])
-                results.append(_vision_fallback(idx, "vision_text", {"_raw": result[:500]}))
-            elif isinstance(result, dict):
-                result.setdefault("_source", "vision")
-                # 防御：关键字段缺失则补默认
-                result.setdefault("name", "未识别")
-                result.setdefault("category", "无法判断")
-                result.setdefault("key_features", [])
-                result.setdefault("scene", "通用")
-                results.append(result)
-            else:
-                logger.warning("[爆款视频] 图片 #%d VLM 返回意外类型 %s", idx, type(result))
-                results.append(_vision_fallback(idx, "vision_unexpected_type"))
-        except Exception as e:
-            logger.warning("[爆款视频] 图片分析失败 img=%s err=%s", img_url[:120], e, exc_info=True)
-            results.append(_vision_fallback(idx, "vision_exception", {"_error": str(e)[:200]}))
+    # 选择视觉模型：lite 速度优先（默认），pro 作为降级备用
+    try:
+        _s = get_shared_settings()
+        if _s.doubao_vision_use_lite:
+            vision_model = _s.doubao_vision_lite_model
+            pro_model = _s.doubao_vision_model
+            vision_timeout = 45  # lite 给到 45s 避免首轮就 timeout 降级 pro
+        else:
+            vision_model = _s.doubao_vision_model
+            pro_model = None  # 已经是 pro，不再降级
+            vision_timeout = 60
+    except Exception:
+        vision_model = "doubao-1-5-vision-lite-250315"
+        pro_model = "doubao-1-5-vision-pro-250328"
+        vision_timeout = 45
+
+    results: list[dict] = [None] * len(job.images)  # type: ignore
+    max_workers = min(4, max(1, len(job.images)))
+    logger.info(
+        "[爆款视频] 开始并行图片分析 n=%d model=%s pro_fallback=%s timeout=%d workers=%d",
+        len(job.images),
+        vision_model,
+        pro_model,
+        vision_timeout,
+        max_workers,
+    )
+    with ThreadPoolExecutor(max_workers=max_workers) as pool:
+        future_to_idx = {
+            pool.submit(
+                _analyze_single_image, idx, url, vision_model, vision_timeout, pro_fallback_model=pro_model
+            ): idx
+            for idx, url in enumerate(job.images)
+        }
+        for fut in as_completed(future_to_idx):
+            idx = future_to_idx[fut]
+            try:
+                results[idx] = fut.result()
+            except Exception as e:
+                logger.warning("[爆款视频] 图片 #%d future 异常 err=%s", idx, e, exc_info=True)
+                results[idx] = _vision_fallback(idx, "future_exception", {"_error": str(e)[:200]})
 
     return {"products": results}
 
@@ -289,8 +522,11 @@ def _step_intent_parsing(job: ViralVideoJob, image_analysis: dict) -> dict:
   "suggested_title": "视频主题标题（5-15字）"
 }}"""
 
+    _s = get_shared_settings()
+    _fast = _s.doubao_fast_model
     try:
-        result = call_llm(prompt)
+        # 用快模型提速（结构化输出任务，不需要推理模型）
+        result = call_llm(prompt, temperature=0.4, max_tokens=800, model=_fast)
         return (
             result
             if isinstance(result, dict)
@@ -304,26 +540,47 @@ def _step_intent_parsing(job: ViralVideoJob, image_analysis: dict) -> dict:
 # ── 编导分镜脚本生成（核心，v1.6 新 prompt） ──────────────────────────────
 
 
-_SCRIPT_GENERATION_PROMPT = """你是一名资深短视频导演，擅长为 AI 视频生成模型（Seedance 2.5）撰写专业编导分镜脚本。
+# 人设 IP 类型 → 文案/出镜风格指导（前端下拉 10 个 IP 类型）
+_PERSONA_STYLE_GUIDE = {
+    "通用个人IP": "亲切自然、像朋友分享好物，第一人称口语化，不端着",
+    "老板型IP": "沉稳大气、有行业格局感，适度使用『我做了XX年』『我一直坚持』等老板视角，语气自信不夸张",
+    "专家型IP": "专业权威、讲原理和数据支撑，用词严谨，少用网梗，像行业专家做科普",
+    "顾问型IP": "贴心周到、给建议给方案，多用『建议你』『可以试试』『我帮你梳理』",
+    "创始人IP": "真诚有温度、讲品牌故事和创业初心，带点情怀和个人观点，不端老板架子",
+    "创业者IP": "真实接地气、讲踩坑经验和创业心路，带点自嘲和韧劲，像身边的创业者朋友",
+    "从业者经验派": "内行视角、讲行业内幕/实操经验/踩坑教训，多用『干了X年我发现』『内行都知道』",
+    "避坑顾问型": "直接点出痛点和雷区，先讲『别买XX』『很多人踩过的坑』再给正确选择，节奏感强",
+    "知识科普型": "清晰讲原理、讲知识点，条理分明、信息密度高，像做一期小科普",
+    "测评种草型": "真实测评感、讲使用体验和优缺点对比，带『亲测』『我用了XX天』『实测下来』真实感词汇",
+}
 
-## 产品信息
+
+def _persona_style_hint(persona_id: str) -> str:
+    """根据 persona_id 查文案风格指导；未命中/空值返回通用提示。"""
+    pid = (persona_id or "").strip()
+    if pid in _PERSONA_STYLE_GUIDE:
+        return f"【人设风格：{pid}】{_PERSONA_STYLE_GUIDE[pid]}"
+    if pid:
+        # 前端传了自由值，照直提示，不阻塞
+        return f"【人设风格：{pid}】按该人设的口吻、话术习惯组织口播和出镜动作"
+    return "【人设风格：未指定】亲切自然、像朋友分享好物"
+
+
+_SCRIPT_GENERATION_PROMPT = """你是资深短视频导演，为 Seedance 2.5（单次生成最多{duration}秒）写编导分镜脚本。脚本将整体作为 prompt 一次性传给视频模型，必须让模型在连贯镜头流中清楚每段时间拍什么、画面如何、人物说什么。
+
+## 产品
 {products_summary}
 
 ## 营销参数
-- 视频主题/意图：{intent}
+- 主题/意图：{intent}
 - 关键信息：{key_messages}
 - 调性：{tone}
 - 目标客户：{target_customer}
-- 用户原始文案/卖点（必须融入口播）：{user_copy}
-- 视频时长：{duration} 秒（单次生成）
-- 画幅比例：{ratio}
-- 产品图片数量：{n_images} 张（将作为 reference_images 传给视频模型，第1张通常作为首帧/主产品图）
-- 参考风格（可选）：{style_hint}
-
-## 任务
-请撰写**一段完整的编导分镜脚本**，包含视频总览、场景光线、逐镜头时间轴、硬性约束、负面提示词，以及自然口语化的口播对白。
-
-这段脚本会**整个拼成一个长 prompt**一次性传给 Seedance 2.5（单次生成最多30秒视频），所以你的描述必须让模型在一个长镜头/连贯镜头流里理解每个时间段该拍什么、画面如何、人物说什么做什么。
+- 用户原始卖点（必须融入口播）：{user_copy}
+- 时长：{duration}秒 / 画幅：{ratio} / 产品图：{n_images}张（第1张通常是主图/首帧）
+- 风格参考：{style_hint}
+- 爆款结构（必须严格遵循节奏/段落顺序）：{viral_structure_block}
+- 人设/出镜口吻（必须贯穿全部对白和动作描写）：{persona_hint}
 
 ## 输出格式（必须输出严格 JSON，不要 Markdown，不要解释，字段一个都不能少）
 
@@ -364,17 +621,21 @@ _SCRIPT_GENERATION_PROMPT = """你是一名资深短视频导演，擅长为 AI
 ```
 
 ## 关键要求
-1. **镜头感**：每镜必须写清景别（特写/近景/中景/全景）、角度（平视/俯拍/仰拍/45度侧拍）、运镜（推/拉/摇/移/跟/固定），不能笼统说"展示产品"。
-2. **画面具体**：描述主体是谁（性别/年龄/穿着风格）、在什么场景、做什么动作、光线从哪来、镜头怎么动，让 AI 能画出来。
-3. **对白自然**：像真人说话，不要"家人们谁懂啊""宝子们"这种浮夸腔，也不要"今天给大家推荐一款XX真的太好用了"这种硬广推销腔。要像朋友自然分享好物。
-4. **参考图片分配**：reference_image_index 填 0-based 索引，产品特写镜头用产品图（索引0通常是主图），人像/场景镜头可留 null。
-5. **时长控制**：所有 shots 的 time_range 加起来必须等于 {duration} 秒，单镜 2-8 秒。
-6. **硬性约束和负面词必须包含**：不要删减，可根据产品类型追加。
-7. **voiceover_script 必须是纯口播文本**：不含任何标记、括号、说明，字数按中文每秒 3-4 字估算（{duration}秒约{approx_chars}字）。
-"""
+1. 每镜写清景别/角度/运镜（特写/近景/中景+平视/俯拍+推/拉/固定）。
+2. 画面具体：主体（性别/年龄/穿着）、场景、动作、光线、镜头运动要可落地。
+3. 对白自然口语化，像朋友分享好物；拒绝"家人们""宝子们""太好用了"等浮夸/硬广腔。
+4. reference_image_index 填 0-based 索引（产品特写用索引0主图），人像/场景可 null。
+5. shots time_range 累计={duration}秒，单镜2-8秒。
+6. hard_constraints/negative_prompts 保留默认项可追加，不要删减。
+7. voiceover_script 为纯口播文本（无标记/括号/前缀），{duration}秒约{approx_chars}字。
+8. 严格按上方「爆款结构」的节奏/段落顺序编排（钩子/痛点/反转/案例/行动号召与结构对齐）。
+9. 输出前自检：口播对白禁止错别字和语病（特别注意"很/最"等常见误用），同音字错误一律修正。
+10. 必须使用产品信息中真实的品牌、品名和外观特征，不要编造与产品无关的内容。"""
 
 
 def _build_products_summary(image_analysis: dict) -> str:
+    """把 VLM 返回的商品分析结果拼给文案/分镜生成 prompt 用。
+    优先用 summary（自然段落）；没有时用结构化字段兜底拼一段。"""
     products = (image_analysis or {}).get("products", []) or []
     if not products:
         return "- （无图片信息，请自由创作自然生活化场景）"
@@ -383,34 +644,60 @@ def _build_products_summary(image_analysis: dict) -> str:
         if not isinstance(p, dict):
             continue
         name = p.get("name") or "产品"
+        # 优先 VLM 生成的 summary 段（自然语言，给编导模型看效果最好）
+        summary = (p.get("summary") or "").strip()
+        if summary and len(summary) >= 30:
+            lines.append(f"- 图{i+1} {name}：{summary}")
+            continue
+        # 结构化字段兜底
         brand = p.get("brand") or ""
         cat = p.get("category") or ""
+        spec = p.get("spec") or ""
+        appearance = p.get("appearance") or ""
+        packaging = p.get("packaging") or ""
         colors = p.get("colors") or []
         mat = p.get("material_or_texture") or ""
         style = p.get("visual_style") or ""
         scene = p.get("scene") or ""
-        audience = p.get("target_audience_hint") or ""
-        text_on_img = p.get("text_on_image") or ""
+        audience = p.get("target_audience") or p.get("target_audience_hint") or ""
+        # text_on_package 可能是数组（新格式）或字符串（旧格式）
+        text_list = p.get("text_on_package") or []
+        if isinstance(text_list, str):
+            text_on_img = text_list
+        else:
+            text_on_img = "；".join([str(x) for x in text_list[:8]]) if text_list else (p.get("text_on_image") or "")
         feats = p.get("key_features") or p.get("features") or []
+        sellings = p.get("selling_points") or []
+        scenes = p.get("suitable_scenes") or []
         parts = [f"图{i+1} {name}"]
         if brand and brand not in ("未知", "无法判断"):
             parts.append(f"品牌={brand}")
         if cat and cat not in ("无法判断", "非产品图"):
             parts.append(f"品类={cat}")
+        if spec and spec != "无法判断":
+            parts.append(f"规格={spec}")
+        if appearance and appearance != "无法判断":
+            parts.append(f"外观={appearance}")
+        if packaging and packaging != "无法判断":
+            parts.append(f"包装={packaging}")
         if colors:
             parts.append(f"颜色={','.join(colors)}")
-        if mat and mat not in ("无法判断",):
+        if mat and mat != "无法判断":
             parts.append(f"材质={mat}")
         if style:
             parts.append(f"风格={style}")
         if scene and scene not in ("通用",):
-            parts.append(f"场景={scene}")
-        if audience and audience != "通用":
+            parts.append(f"展示场景={scene}")
+        if scenes:
+            parts.append(f"适用场景={','.join([str(x) for x in scenes[:4]])}")
+        if audience and audience not in ("通用", "无法判断"):
             parts.append(f"目标人群={audience}")
         if text_on_img and text_on_img not in ("无",):
-            parts.append(f"图片文字={text_on_img}")
+            parts.append(f"包装文字={text_on_img[:300]}")
         if feats:
             parts.append("外观特征=" + "；".join([str(x) for x in feats[:6]]))
+        if sellings:
+            parts.append("营销卖点=" + "；".join([str(x) for x in sellings[:5]]))
         lines.append("- " + "，".join(parts))
     return "\n".join(lines)
 
@@ -594,6 +881,14 @@ def _step_script_generation(job: ViralVideoJob, intent: dict, image_analysis: di
         intent_str = "推广产品"
         tone = "亲切自然"
 
+    # 爆款结构：用户在 STEP1/STEP2 选的中文结构名，必须严格注入 prompt 指导 AI 编排
+    _vs = (job.viral_structure or "").strip()
+    if _vs:
+        viral_structure_block = f"【{_vs}】—— 请严格按照这个爆款结构的节奏/段落顺序编排镜头、台词和情绪节点（开场钩子、痛点、反转、案例、行动号召等按结构走），不要打乱顺序"
+    else:
+        viral_structure_block = "未指定（自由编排，但仍需有钩子开头+产品展示+行动号召的基本节奏）"
+
+    persona_hint = _persona_style_hint(getattr(job, "persona_id", ""))
     prompt = _SCRIPT_GENERATION_PROMPT.format(
         products_summary=products_summary,
         intent=intent_str,
@@ -606,14 +901,58 @@ def _step_script_generation(job: ViralVideoJob, intent: dict, image_analysis: di
         n_images=len(job.images or []),
         style_hint=style_hint,
         approx_chars=approx_chars,
+        viral_structure_block=viral_structure_block,
+        persona_hint=persona_hint,
     )
 
+    _s = get_shared_settings()
+    _fast = _s.doubao_fast_model
+    _pro = getattr(_s, "doubao_model", None) or _fast
+
+    def _try_gen(model: str, temp: float, max_tok: int, label: str):
+        logger.info("[爆款视频] 编导脚本生成 model=%s label=%s", model, label)
+        r = call_llm(prompt, temperature=temp, max_tokens=max_tok, model=model)
+        if r is None:
+            logger.warning("[爆款视频] 编导脚本返回None label=%s", label)
+            return None
+        parsed = _safe_json_loads(r)
+        normalized = _validate_and_normalize_script(parsed, job)
+        voiceover = (normalized or {}).get("voiceover_script") or ""
+        voiceover_len = len(voiceover)
+        shots_cnt = len((normalized or {}).get("shots") or [])
+        # 判定是否"退化到兜底质量"：口播过短（<20字）或镜头数<1；正常的短口播（如15s视频~40字）不视为兜底
+        fallback_marker = "我最近在用的好物" in voiceover  # _fallback_script 的特征串
+        is_fallback = fallback_marker or shots_cnt < 1 or voiceover_len < 20
+        logger.info(
+            "[爆款视频] 编导脚本结果 label=%s voiceover_len=%d shots=%d fallback=%s raw_type=%s",
+            label,
+            voiceover_len,
+            shots_cnt,
+            is_fallback,
+            type(r).__name__,
+        )
+        if is_fallback:
+            return None  # 触发重试
+        return normalized
+
     try:
-        result = call_llm(prompt)
-        parsed = _safe_json_loads(result)
-        return _validate_and_normalize_script(parsed, job)
+        # 第一次：快模型
+        normalized = _try_gen(_fast, 0.8, 2500, "fast-first")
+        if normalized is not None:
+            return normalized
+        # 第二次：快模型降温度+加大 max_tokens
+        normalized = _try_gen(_fast, 0.6, 3200, "fast-retry")
+        if normalized is not None:
+            return normalized
+        # 第三次：用主力模型兜底
+        if _pro and _pro != _fast:
+            normalized = _try_gen(_pro, 0.7, 3500, "pro-fallback")
+            if normalized is not None:
+                return normalized
+        logger.warning("[爆款视频] 编导脚本三次都未生成合格结果，使用兜底脚本")
+        return _fallback_script(job)
     except Exception as e:
-        logger.warning("[爆款视频] 编导脚本生成失败: %s，使用兜底脚本", e, exc_info=True)
+        logger.warning("[爆款视频] 编导脚本生成异常: %s，使用兜底脚本", e, exc_info=True)
         return _fallback_script(job)
 
 
@@ -638,8 +977,11 @@ def _step_review(job: ViralVideoJob, copy_result: dict) -> dict:
 - score: int（0-100分）
 - details: 各维度评分和说明
 - issues: 需要修改的问题列表（如有）"""
+    _s = get_shared_settings()
+    _fast = _s.doubao_fast_model
     try:
-        result = call_llm(prompt)
+        # 合规审核用快模型 + 短输出（结构化判断）
+        result = call_llm(prompt, temperature=0.1, max_tokens=500, model=_fast)
         return result if isinstance(result, dict) else {"passed": True, "score": 80, "details": {}}
     except Exception as e:
         logger.warning("[爆款视频] 合规审核失败: %s", e)
@@ -654,7 +996,7 @@ def _step_tts(job: ViralVideoJob, voiceover_script: str):
         from apps.worker.services.tts_service_factory import get_tts_service
 
         tts_service = get_tts_service()
-        voice_id = (getattr(job, "voice_id", "") or job.persona_id or "").strip()
+        voice_id = (getattr(job, "voice_id", "") or "").strip()
         text = (voiceover_script or "").strip()
         if not text:
             logger.warning("[爆款视频] voiceover_script 为空，跳过 TTS")
@@ -819,6 +1161,27 @@ def _step_upload(job: ViralVideoJob, video_path: str) -> str:
     return video_url
 
 
+def _wait_oss_ready(url: str, timeout_sec: int = 10) -> bool:
+    """轮询 OSS 公网 URL，直到 HEAD 返回 200 或超时。
+    用于缓解 OSS 上传后 1-5s 公网 eventual consistency 导致的 NoSuchKey。
+    """
+    import httpx
+
+    deadline = time.monotonic() + timeout_sec
+    last_status = 0
+    while time.monotonic() < deadline:
+        try:
+            r = httpx.head(url, follow_redirects=True, timeout=3.0)
+            last_status = r.status_code
+            if r.status_code == 200 and int(r.headers.get("content-length", "0") or 0) > 0:
+                return True
+        except Exception as e:
+            logger.debug("[爆款视频] OSS head 轮询失败: %s", e)
+        time.sleep(1.0)
+    logger.warning("[爆款视频] OSS 成片在 %ds 内未就绪 last_status=%s url=%s", timeout_sec, last_status, url[:120])
+    return False
+
+
 # ── 主编排器 ────────────────────────────────────────────────────────────
 
 
@@ -827,15 +1190,17 @@ def run_viral_video_pipeline(self: Task, job_id: str) -> dict:
     """旧一键流水线（保留兼容）：图片分析→风格分析→意图解析→WAIT_USER_CONFIRM。"""
     session = None
     try:
+        _recover_stale_jobs()  # 顺带回收僵尸任务
+        _hb_stop = None
         session, repo, job = _get_repo_and_job(job_id)
         if job is None:
             return {"ok": False, "error": "job not found"}
 
         job.mark_running()
         _save_job(repo, job, session)
-        _emit_progress(job_id, ViralVideoStage.IMAGE_ANALYSIS, 5.0, "开始图片分析")
+        _hb_stop, _hb_thread = _start_heartbeat_thread(job_id)
+        _set_stage(job, repo, session, ViralVideoStage.IMAGE_ANALYSIS, "正在分析商品特征...")
 
-        _emit_progress(job_id, ViralVideoStage.IMAGE_ANALYSIS, 10.0, "正在分析产品图片...")
         image_analysis = _step_image_analysis(job)
         job.image_analysis = image_analysis
         _save_job(repo, job, session)
@@ -843,16 +1208,18 @@ def run_viral_video_pipeline(self: Task, job_id: str) -> dict:
 
         style_guide = None
         if job.reference_video_url or job.style_template_id:
-            _emit_progress(job_id, ViralVideoStage.VIDEO_ANALYSIS, 20.0, "正在分析参考视频风格...")
+            _set_stage(job, repo, session, ViralVideoStage.VIDEO_ANALYSIS, "正在分析参考视频风格...")
             style_guide = _step_video_analysis(job)
             job.style_guide = style_guide
             _save_job(repo, job, session)
             _emit_progress(job_id, ViralVideoStage.VIDEO_ANALYSIS, 25.0, "风格分析完成", {"style_guide": style_guide})
 
-        _emit_progress(job_id, ViralVideoStage.INTENT_PARSING, 30.0, "正在解析文案意图...")
+        _set_stage(job, repo, session, ViralVideoStage.INTENT_PARSING, "正在解析文案意图...")
         intent_result = _step_intent_parsing(job, image_analysis)
 
         job.mark_wait_user_confirm(intent_result)
+        job.current_stage = ViralVideoStage.INTENT_PARSING
+        job.phase_message = "意图解析完成，等待用户确认"
         _save_job(repo, job, session)
         _emit_progress(
             job_id,
@@ -878,6 +1245,8 @@ def run_viral_video_pipeline(self: Task, job_id: str) -> dict:
         _mark_failed_and_notify(job_id, session, None, None, str(e), "")
         return {"ok": False, "job_id": job_id, "error": str(e)}
     finally:
+        if _hb_stop is not None:
+            _hb_stop.set()
         if session:
             session.close()
 
@@ -960,15 +1329,17 @@ def run_viral_video_analyze(self: Task, job_id: str) -> dict:
     """v1.5+ 阶段1：图片 VLM 分析 + 可选视频风格分析。"""
     session = None
     try:
+        _recover_stale_jobs()  # 顺带回收僵尸任务
+        _hb_stop = None
         session, repo, job = _get_repo_and_job(job_id)
         if job is None:
             return {"ok": False, "error": "job not found"}
 
         job.mark_running()
         _save_job(repo, job, session)
-        _emit_progress(job_id, ViralVideoStage.IMAGE_ANALYSIS, 5.0, "开始图片分析")
+        _hb_stop, _hb_thread = _start_heartbeat_thread(job_id)
+        _set_stage(job, repo, session, ViralVideoStage.IMAGE_ANALYSIS, "正在分析商品特征...")
 
-        _emit_progress(job_id, ViralVideoStage.IMAGE_ANALYSIS, 10.0, "正在分析产品图片...")
         image_analysis = _step_image_analysis(job)
         job.image_analysis = image_analysis
         _save_job(repo, job, session)
@@ -976,7 +1347,7 @@ def run_viral_video_analyze(self: Task, job_id: str) -> dict:
 
         style_guide = None
         if job.reference_video_url or job.style_template_id:
-            _emit_progress(job_id, ViralVideoStage.VIDEO_ANALYSIS, 70.0, "正在分析参考视频风格...")
+            _set_stage(job, repo, session, ViralVideoStage.VIDEO_ANALYSIS, "正在分析参考视频风格...")
             style_guide = _step_video_analysis(job)
             job.style_guide = style_guide
             _save_job(repo, job, session)
@@ -989,6 +1360,8 @@ def run_viral_video_analyze(self: Task, job_id: str) -> dict:
             )
 
         job.mark_image_analyzed()
+        job.current_stage = ViralVideoStage.IMAGE_ANALYSIS
+        job.phase_message = "图片分析完成，请填写营销参数以生成编导脚本"
         _save_job(repo, job, session)
         _emit_progress(
             job_id,
@@ -1007,55 +1380,72 @@ def run_viral_video_analyze(self: Task, job_id: str) -> dict:
         _mark_failed_and_notify(job_id, session, None, None, str(e), ViralVideoStage.IMAGE_ANALYSIS)
         return {"ok": False, "job_id": job_id, "error": str(e)}
     finally:
+        if _hb_stop is not None:
+            _hb_stop.set()
         if session:
             session.close()
 
 
 @shared_task(bind=True, max_retries=1, name="worker.run_viral_video_generate_copy")
 def run_viral_video_generate_copy(self: Task, job_id: str) -> dict:
-    """v1.6 阶段2：意图解析 → 编导分镜脚本生成 → 合规审核，完成后状态=copy_generated。"""
+    """v1.6 阶段2（v1.6.1 提速版）：意图解析 → 编导分镜脚本生成 → 直接返回，合规审核后置到出片前。
+
+    优化点（#2134 问题7）：
+    - 意图解析/编导脚本均使用快模型（doubao_fast_model，非推理模型），max_tokens 收紧
+    - _SCRIPT_GENERATION_PROMPT 精简冗余描述
+    - 合规审核改为异步后置：不阻塞前端，在 confirm-copy（阶段3 TTS前）再做最终审核
+    - 每个阶段通过 _set_stage 持久化 current_stage/phase_message 到 DB（问题8）
+    """
     session = None
     try:
+        _recover_stale_jobs()  # 顺带回收僵尸任务
+        _hb_stop = None
         session, repo, job = _get_repo_and_job(job_id)
         if job is None:
             return {"ok": False, "error": "job not found"}
         if job.status != ViralVideoStatus.RUNNING:
             return {"ok": False, "error": f"unexpected status: {job.status}"}
+        job.touch_heartbeat()
+        _save_job(repo, job, session)
+        _hb_stop, _hb_thread = _start_heartbeat_thread(job_id)
 
+        # 阶段：意图解析
+        _set_stage(job, repo, session, ViralVideoStage.INTENT_PARSING, "正在解析文案意图...")
         image_analysis = job.image_analysis or {"products": []}
-
-        _emit_progress(job_id, ViralVideoStage.INTENT_PARSING, 20.0, "正在解析文案意图...")
         intent_result = _step_intent_parsing(job, image_analysis)
         job.intent_result = intent_result
         _save_job(repo, job, session)
         _emit_progress(job_id, ViralVideoStage.INTENT_PARSING, 35.0, "意图解析完成")
 
-        _emit_progress(job_id, ViralVideoStage.SCRIPT_GENERATION, 40.0, "正在生成编导分镜脚本...")
+        # 阶段：编导脚本生成（核心耗时环节，已用快模型）
+        _set_stage(job, repo, session, ViralVideoStage.SCRIPT_GENERATION, "正在编排分镜脚本...")
         copy_result = _step_script_generation(job, intent_result, image_analysis)
         _emit_progress(
             job_id,
             ViralVideoStage.SCRIPT_GENERATION,
-            60.0,
-            "编导脚本生成完成",
+            85.0,
+            "分镜脚本生成完成",
             {"shots": len(copy_result.get("shots", []))},
         )
 
-        _emit_progress(job_id, ViralVideoStage.REVIEW, 65.0, "正在进行合规审核...")
-        review_result = _step_review(job, copy_result)
-        if not review_result.get("passed", True):
-            _emit_progress(job_id, ViralVideoStage.REVIEW, 67.0, "审核未通过，正在自动重写...")
-            copy_result = _step_script_generation(job, intent_result, image_analysis)
-            _step_review(job, copy_result)
-        _emit_progress(job_id, ViralVideoStage.REVIEW, 70.0, "合规审核完成")
+        # 合规审核后置：不再阻塞前端返回；在阶段3（confirm-copy 出片前）_run_render_pipeline 里再做最终审核。
+        # 这里只做一个快速轻量检查（关键字黑名单），发现明显违规再触发重写；LLM 深度审核放到出片前。
+        _set_stage(job, repo, session, ViralVideoStage.REVIEW, "正在快速检查脚本合规性...")
+        voiceover = (copy_result or {}).get("voiceover_script", "") or ""
+        _quick_compliance_blacklist_check(copy_result)
+        _emit_progress(job_id, ViralVideoStage.REVIEW, 95.0, "脚本合规初检完成")
 
+        # 标记 copy_generated 并持久化
         job.mark_copy_generated(copy_result)
+        job.current_stage = ViralVideoStage.REVIEW
+        job.phase_message = "分镜脚本已生成，请确认或编辑口播文案"
         _save_job(repo, job, session)
-        voiceover = copy_result.get("voiceover_script", "")
+
         _emit_progress(
             job_id,
             ViralVideoStage.REVIEW,
             100.0,
-            "编导分镜脚本已生成，请确认或编辑口播文案",
+            "分镜脚本已生成，请确认或编辑口播文案",
             {
                 "copy_result": copy_result,
                 "generated_copy_text": voiceover,
@@ -1086,43 +1476,113 @@ def run_viral_video_generate_copy(self: Task, job_id: str) -> dict:
         _mark_failed_and_notify(job_id, session, None, None, str(e), ViralVideoStage.SCRIPT_GENERATION)
         return {"ok": False, "job_id": job_id, "error": str(e)}
     finally:
+        if _hb_stop is not None:
+            _hb_stop.set()
         if session:
             session.close()
 
 
+def _quick_compliance_blacklist_check(copy_result: dict) -> None:
+    """阶段2快速黑名单检查：不调用 LLM，只扫描高风险关键词；命中则在 voiceover 中就地替换。
+
+    LLM 深度合规审核（_step_review）在阶段3 confirm-copy 出片前执行。
+    """
+    if not isinstance(copy_result, dict):
+        return
+    voiceover = copy_result.get("voiceover_script", "") or ""
+    # 广告法绝对化用语黑名单（常见速查，远非完整，仅挡住最明显违规）
+    BLACKLIST = {
+        "最": "很",
+        "第一": "领先",
+        "国家级": "高品质",
+        "世界级": "高品质",
+        "顶级": "优质",
+        "极品": "优质",
+        "独家": "特色",
+        "绝无仅有": "少见",
+        "100%": "大幅",
+        "百分百": "大幅",
+        "永久": "长久",
+        "万能": "多用途",
+        "特效": "效果好",
+        "速效": "快速见效",
+        "根治": "改善",
+        "包治": "改善",
+        "药到病除": "缓解不适",
+    }
+    changed = False
+    for k, v in BLACKLIST.items():
+        if k in voiceover:
+            voiceover = voiceover.replace(k, v)
+            changed = True
+    if changed:
+        copy_result["voiceover_script"] = voiceover
+        # 同步 final_copy/suggested_copy（如果存在）
+        for k in ("final_copy", "suggested_copy"):
+            if isinstance(copy_result.get(k), str) and copy_result[k]:
+                for bk, bv in BLACKLIST.items():
+                    copy_result[k] = copy_result[k].replace(bk, bv)
+
+
 def _run_render_pipeline(job_id: str, session, repo, job) -> dict:
-    """v1.6 阶段3 / 旧 resume 共用：TTS → 单次 Seedance → Upload → Completed。"""
+    """v1.6.1 阶段3：出片前合规审核（LLM 深度）→ TTS → Seedance → Upload → Completed。
+
+    阶段2 generate-copy 已把 LLM 深度审核后置，这里在 TTS 前做最终审核（不通过则自动重写1次）。
+    所有阶段通过 _set_stage 持久化 current_stage/phase_message。
+    """
     image_analysis = job.image_analysis or {"products": []}
 
-    # 如果没有 copy_result（旧数据/失败重试），现场补生成
+    # 如果没有 copy_result（旧数据/失败重试），现场补生成（意图+脚本，不走 LLM 审核，出片前会统一做）
     copy_result = job.copy_result
     if not isinstance(copy_result, dict) or not copy_result:
-        _emit_progress(job_id, ViralVideoStage.SCRIPT_GENERATION, 40.0, "正在补生成编导脚本...")
+        _set_stage(job, repo, session, ViralVideoStage.SCRIPT_GENERATION, "正在补生成编导脚本...")
         intent = job.intent_result or _step_intent_parsing(job, image_analysis)
         copy_result = _step_script_generation(job, intent, image_analysis)
-        _step_review(job, copy_result)
         job.mark_copy_generated(copy_result)
         _save_job(repo, job, session)
 
+    # 出片前 LLM 深度合规审核（#2134 问题7：审核从阶段2后置到这里，不阻塞前端预览脚本）
+    _set_stage(job, repo, session, ViralVideoStage.REVIEW, "正在进行出片前合规审核...")
+    try:
+        review_result = _step_review(job, copy_result)
+        if not review_result.get("passed", True):
+            _emit_progress(job_id, ViralVideoStage.REVIEW, 67.0, "审核未通过，正在自动重写...")
+            intent = job.intent_result or _step_intent_parsing(job, image_analysis)
+            copy_result = _step_script_generation(job, intent, image_analysis)
+            _step_review(job, copy_result)  # 二次审核，不通过也继续出片（避免反复循环）
+            job.copy_result = copy_result
+            job.generated_copy_text = copy_result.get("voiceover_script", "") or ""
+            _save_job(repo, job, session)
+    except Exception as e:
+        logger.warning("[爆款视频][阶段3] 合规审核异常，继续出片: %s", e)
+    _emit_progress(job_id, ViralVideoStage.REVIEW, 70.0, "合规审核完成")
+
     voiceover = copy_result.get("voiceover_script", "") or job.effective_copy_text
 
     # Step 5: TTS 整段合成
-    _emit_progress(job_id, ViralVideoStage.TTS, 72.0, "正在生成AI配音...")
+    _set_stage(job, repo, session, ViralVideoStage.TTS, "正在合成AI配音...")
     tts_path = _step_tts(job, voiceover)
     tts_url = _upload_tts_to_oss(job, tts_path)
     _emit_progress(job_id, ViralVideoStage.TTS, 78.0, "配音完成", {"has_tts": tts_url is not None})
 
     # Step 6: 单次 Seedance
-    _emit_progress(job_id, ViralVideoStage.RENDERING, 80.0, "正在调用AI生成视频（约1-3分钟）...")
+    _set_stage(job, repo, session, ViralVideoStage.RENDERING, "正在生成视频（约1-3分钟）...")
     video_path = _step_render(job, copy_result, tts_url)
     _emit_progress(job_id, ViralVideoStage.RENDERING, 92.0, "视频生成完成")
 
     # Step 7: Upload
-    _emit_progress(job_id, ViralVideoStage.UPLOADING, 95.0, "正在上传视频...")
+    _set_stage(job, repo, session, ViralVideoStage.UPLOADING, "正在上传视频...")
     video_url = _step_upload(job, video_path)
 
+    # P2-2 OSS 一致性：上传后循环 head 确认公网可访问（最多等 10s），
+    # 避免前端拿到 completed 立即下载时命中 NoSuchKey。
+    if video_url:
+        _wait_oss_ready(video_url, timeout_sec=10)
+
     job.credits_cost = CREDITS_VIRAL_VIDEO_COST
     job.mark_completed(video_url)
+    job.current_stage = ViralVideoStage.UPLOADING
+    job.phase_message = "视频生成完成"
     _save_job(repo, job, session)
     _emit_progress(job_id, ViralVideoStage.UPLOADING, 100.0, "视频生成完成！", {"video_url": video_url})
     _emit_progress(
@@ -1142,11 +1602,16 @@ def run_viral_video_render(self: Task, job_id: str) -> dict:
     """v1.6 阶段3：TTS + 单次 Seedance 生成 + 上传。"""
     session = None
     try:
+        _recover_stale_jobs()  # 顺带回收僵尸任务
+        _hb_stop = None
         session, repo, job = _get_repo_and_job(job_id)
         if job is None:
             return {"ok": False, "error": "job not found"}
         if job.status != ViralVideoStatus.RUNNING:
             return {"ok": False, "error": f"unexpected status: {job.status}"}
+        job.touch_heartbeat()
+        _save_job(repo, job, session)
+        _hb_stop, _hb_thread = _start_heartbeat_thread(job_id)
         return _run_render_pipeline(job_id, session, repo, job)
     except Retry:
         raise
@@ -1155,5 +1620,7 @@ def run_viral_video_render(self: Task, job_id: str) -> dict:
         _mark_failed_and_notify(job_id, session, None, None, str(e), ViralVideoStage.RENDERING)
         return {"ok": False, "job_id": job_id, "error": str(e)}
     finally:
+        if _hb_stop is not None:
+            _hb_stop.set()
         if session:
             session.close()

packages/adapters/sqlalchemy_impl/models.py

@@ -950,6 +950,9 @@ class ViralVideoJobModel(Base):
     video_model = Column(String(100), nullable=False, default="")
     # 结果与状态
     status = Column(String(30), nullable=False, default="pending", index=True)
+    current_stage = Column(String(200), nullable=False, default="")  # 细粒度阶段 snake_case
+    phase_message = Column(String(500), nullable=False, default="")  # 阶段中文提示文案
+    heartbeat_at = Column(DateTime, nullable=True, index=True)  # worker 心跳，用于僵尸任务超时回收
     intent_result = Column(JSON, nullable=True)
     image_analysis = Column(JSON, nullable=True)
     storyboard = Column(JSON, nullable=True)

packages/adapters/sqlalchemy_impl/viral_video_repository.py

@@ -37,6 +37,9 @@ def _to_domain(model: ViralVideoJobModel) -> ViralVideoJob:
         video_ratio=getattr(model, "video_ratio", "9:16") or "9:16",
         video_model=getattr(model, "video_model", "") or "",
         status=ViralVideoStatus(model.status) if model.status else ViralVideoStatus.PENDING,
+        current_stage=getattr(model, "current_stage", "") or "",
+        phase_message=getattr(model, "phase_message", "") or "",
+        heartbeat_at=getattr(model, "heartbeat_at", None),
         intent_result=dict(model.intent_result) if model.intent_result else None,
         image_analysis=dict(model.image_analysis) if getattr(model, "image_analysis", None) else None,
         storyboard=list(model.storyboard) if getattr(model, "storyboard", None) else None,
@@ -83,6 +86,9 @@ class SQLAlchemyViralVideoJobRepository:
             video_ratio=job.video_ratio,
             video_model=job.video_model,
             status=job.status,
+            current_stage=job.current_stage or "",
+            phase_message=job.phase_message or "",
+            heartbeat_at=job.heartbeat_at,
             intent_result=job.intent_result,
             image_analysis=job.image_analysis,
             storyboard=job.storyboard,
@@ -106,6 +112,9 @@ class SQLAlchemyViralVideoJobRepository:
         if model is None:
             raise ValueError(f"ViralVideoJob {job.id} not found")
         model.status = job.status
+        model.current_stage = job.current_stage or ""
+        model.phase_message = job.phase_message or ""
+        model.heartbeat_at = job.heartbeat_at
         model.intent_result = job.intent_result
         model.image_analysis = job.image_analysis
         model.storyboard = job.storyboard

packages/config/base.py

@@ -90,11 +90,14 @@ class SharedSettings(BaseSettings):
 
     # ── 豆包大模型（火山引擎方舟） ────────────────────────────────────────
     doubao_api_key: str = ""
-    doubao_model: str = "doubao-seed-1-6-250615"
+    doubao_model: str = "doubao-seed-1-6-250615"  # 推理模型（通用兜底）
+    doubao_fast_model: str = "doubao-1-5-pro-32k-250115"  # 快速结构化输出模型（编导脚本/意图解析/审核）
     doubao_base_url: str = "https://ark.cn-beijing.volces.com/api/v3"
     doubao_timeout: int = 30
     doubao_max_retries: int = 2
-    doubao_vision_model: str = "doubao-1-5-vision-pro-250915"
+    doubao_vision_model: str = "doubao-1-5-vision-pro-250328"  # 高精度视觉（备用）
+    doubao_vision_lite_model: str = "doubao-1-5-vision-lite-250315"  # 快速视觉（商品识别默认，速度优先）
+    doubao_vision_use_lite: bool = True  # viral-video 图片分析默认用 lite 提速
     doubao_embedding_model: str = "doubao-embedding-large-text-240915"
     doubao_video_model: str = "doubao-seedance-2-5-260628"
     doubao_video_timeout: int = 600  # 视频生成轮询总超时（秒）

packages/domain/viral_video.py

@@ -117,6 +117,9 @@ class ViralVideoJob:
     # 状态
     id: str = field(default_factory=lambda: uuid4().hex)
     status: ViralVideoStatus = ViralVideoStatus.PENDING
+    current_stage: str = ""  # 细粒度阶段（ViralVideoStage.value，snake_case）
+    phase_message: str = ""  # 阶段中文提示文案，前端轮询直接展示
+    heartbeat_at: datetime | None = None  # worker 心跳时间，用于超时僵尸任务检测
     result_video_url: str = ""
     credits_cost: int = 0
     error_msg: str = ""
@@ -138,9 +141,19 @@ class ViralVideoJob:
         ):
             raise ValueError(f"Cannot transition from {self.status} to running")
         self.status = ViralVideoStatus.RUNNING
+        now = datetime.now(timezone.utc)
         if self.started_at is None:
-            self.started_at = datetime.now(timezone.utc)
-        self.updated_at = datetime.now(timezone.utc)
+            self.started_at = now
+        self.heartbeat_at = now
+        self.updated_at = now
+
+    def touch_heartbeat(self) -> None:
+        """更新心跳时间（worker 在长任务中周期性调用，用于超时检测）。"""
+        now = datetime.now(timezone.utc)
+        if self.started_at is None:
+            self.started_at = now
+        self.heartbeat_at = now
+        self.updated_at = now
 
     def mark_image_analyzed(self) -> None:
         if self.status not in (ViralVideoStatus.PENDING, ViralVideoStatus.RUNNING):

packages/shared/ai_client.py

@@ -40,6 +40,8 @@ class DoubaoClient:
         self.timeout: int = settings.doubao_timeout
         self.max_retries: int = settings.doubao_max_retries
         self.vision_model: str = settings.doubao_vision_model
+        self.vision_lite_model: str = settings.doubao_vision_lite_model
+        self.fast_model: str = settings.doubao_fast_model
 
     def embed_text(self, text: str, timeout: int | None = None) -> list[float] | None:
         """调用豆包文本 Embedding API，返回浮点向量；失败返回 None。"""
@@ -92,6 +94,7 @@ class DoubaoClient:
         messages: list[dict[str, str]],
         temperature: float = 0.7,
         max_tokens: int = 1024,
+        model: str | None = None,
     ) -> Optional[str]:
         """调用 Chat Completion 接口.
 
@@ -112,7 +115,7 @@ class DoubaoClient:
             "Content-Type": "application/json",
         }
         payload: dict[str, Any] = {
-            "model": self.model,
+            "model": model or self.model,
             "messages": messages,
             "temperature": temperature,
             "max_tokens": max_tokens,
@@ -154,11 +157,12 @@ class DoubaoClient:
         max_tokens: int = 2048,
         temperature: float = 0.3,
         timeout: int | None = None,
+        model: str | None = None,
     ) -> Optional[str]:
         """调用豆包视觉理解 API（OpenAI 兼容多模态格式）.
 
         将 images 附加到最后一条 user message 的 content 中，
-        使用 vision_model（默认 doubao-1-5-vision-pro-250915）。
+        使用 vision_model（默认 doubao-1-5-vision-pro-250328）。
 
         Args:
             messages: 对话消息列表。最后一条 user message 会被注入图片内容。
@@ -204,7 +208,7 @@ class DoubaoClient:
             "Content-Type": "application/json",
         }
         payload: dict[str, Any] = {
-            "model": self.vision_model,
+            "model": model or self.vision_model,
             "messages": vision_messages,
             "temperature": temperature,
             "max_tokens": max_tokens,
@@ -264,9 +268,9 @@ class DoubaoClient:
 
         Args:
             prompt: 文本提示词（含完整编导脚本：总览+场景光线+逐镜头时间轴+硬约束+负面词）
-            image_url: 首帧参考图 URL（可选，提供则走图生视频首帧模式，ratio 跟随首帧）
+            image_url: 首帧参考图 URL（可选，提供则走图生视频首帧模式，ratio 仍按参数传，模型会自动居中裁剪到目标比例）
             duration: 视频时长 4~30 秒
-            ratio: 宽高比 16:9/9:16/1:1/4:3/3:4/21:9/adaptive；image_url 存在时自动忽略
+            ratio: 宽高比 16:9/9:16/1:1/4:3/3:4/21:9/adaptive；图生视频时模型会自动居中裁剪首帧到目标比例
             resolution: 480p/720p/1080p
             generate_audio: 是否让模型原生合成音效/BGM（v1.6 默认 True，配合 reference_audios 做口型驱动）
             watermark: 是否加水印
@@ -315,8 +319,9 @@ class DoubaoClient:
         # v1.6: 参考视频（风格参考）
         if reference_videos:
             create_payload["reference_videos"] = [{"url": u} for u in reference_videos[:5] if u and isinstance(u, str)]
-        # Bug #2110 / v1.6: ratio=None 时不传（首帧图生视频跟随原图比例）
-        if ratio and not image_url:
+        # 图生视频也必须传 ratio，否则模型默认 adaptive 可能输出 1:1（首帧方形商品图会导致 960x960）
+        # 官方文档：图生视频会自动居中裁剪首帧到目标比例，支持 16:9/9:16/1:1/4:3/3:4/21:9/adaptive
+        if ratio:
             create_payload["ratio"] = ratio
 
         headers = {

packages/shared/ai_service.py

@@ -496,16 +496,32 @@ def run_generate_cover(
 # ── 通用 LLM / Vision 调用（#2039 ViralVideoOrchestrator 使用，复用现有豆包客户端）──
 
 
-def call_llm(prompt: str, temperature: float = 0.7) -> object:
-    """调用豆包大模型（文本对话），返回解析后的 JSON（dict/list）或原文字符串；失败返回 None。"""
+def call_llm(
+    prompt: str,
+    temperature: float = 0.7,
+    max_tokens: int = 2048,
+    model: str | None = None,
+    system_prompt: str | None = None,
+) -> object:
+    """调用豆包大模型（文本对话），返回解析后的 JSON（dict/list）或原文字符串；失败返回 None。
+
+    Args:
+        prompt: 用户侧提示。
+        temperature: 采样温度。
+        max_tokens: 输出上限（结构化任务默认 2048，长文案可按需加大）。
+        model: 覆盖默认模型（如 fast_model 提速用），None 走配置默认推理模型。
+        system_prompt: 覆盖默认 system prompt。
+    """
     client = get_doubao_client()
     if not client.is_available:
         return None
+    if system_prompt is None:
+        system_prompt = "你是专业的短视频内容策划助手。需要结构化输出时请严格使用 JSON。"
     messages = [
-        {"role": "system", "content": "你是专业的短视频内容策划助手。需要结构化输出时请严格使用 JSON。"},
+        {"role": "system", "content": system_prompt},
         {"role": "user", "content": prompt},
     ]
-    raw = client.chat_completion(messages, temperature=temperature, max_tokens=4096)
+    raw = client.chat_completion(messages, temperature=temperature, max_tokens=max_tokens, model=model)
     if raw is None:
         return None
     try:
@@ -514,14 +530,26 @@ def call_llm(prompt: str, temperature: float = 0.7) -> object:
         return raw
 
 
-def call_vision(image_url: str, prompt: str) -> object:
+def call_vision(
+    image_url: str,
+    prompt: str,
+    *,
+    model: str | None = None,
+    max_tokens: int = 1024,
+    temperature: float = 0.2,
+    timeout: int = 45,
+    system_prompt: str | None = None,
+) -> object:
     """调用豆包视觉大模型分析图片，返回解析后的 JSON 或原文字符串；失败返回 None。
 
-    Bug #2114 (VLM 牛头不对马嘴根因修复):
-    之前误走 client.chat_completion(用文本模型 doubao-seed-1.6)，多模态 content list 被当成
-    纯文本发给文本模型 → 模型要么看不到图、要么抛 400，静默被 except 吞掉 → 返回 None →
-    _step_image_analysis fallback 到 {"name":"未识别"} → 后续文案/分镜完全没图的信息。
-    现改走 vision_completion，走视觉模型 doubao-1-5-vision-pro-250915。
+    Args:
+        image_url: 可公网访问的图片 URL（直接传给豆包视觉模型，无需本地下载）。
+        prompt: 用户侧文本提示。
+        model: 覆盖默认视觉模型（如 vision_lite_model 提速用），None 走配置默认。
+        max_tokens: 输出上限，商品识别用 800~1200 足够，避免长输出拖慢首 token。
+        temperature: 温度。
+        timeout: 单次请求超时（秒）。
+        system_prompt: 覆盖默认 system prompt（viral-video 商品分析会传专门的详细 prompt）。
     """
     client = get_doubao_client()
     if not client.is_available:
@@ -531,28 +559,33 @@ def call_vision(image_url: str, prompt: str) -> object:
         logger.warning("[call_vision] 空 image_url，跳过视觉分析")
         return None
 
-    system_prompt = (
-        "你是资深电商视觉分析师。请严格基于用户提供的图片观察回答，"
-        "图片里没有的信息不要凭空想象或编造；看不清或无法判断时明确说"
-        "「图片中无法判断」，不要猜测。输出必须是严格 JSON，不要附加 Markdown 或解释文字。"
-    )
+    if system_prompt is None:
+        system_prompt = (
+            "你是资深电商视觉分析师。请严格基于用户提供的图片观察回答，"
+            "图片里没有的信息不要凭空想象或编造；看不清或无法判断时明确说"
+            "「无法判断」，不要猜测。输出必须是严格 JSON，不要附加 Markdown 或解释文字。"
+        )
     messages = [
         {"role": "system", "content": system_prompt},
         {"role": "user", "content": prompt},
     ]
 
+    used_model = model or getattr(client, "vision_model", "?")
     logger.info(
-        "[call_vision] 调用豆包视觉模型 vision_model=%s image_url=%s prompt_len=%d",
-        getattr(client, "vision_model", "?"),
+        "[call_vision] 调用豆包视觉模型 model=%s image_url=%s prompt_len=%d max_tokens=%d timeout=%d",
+        used_model,
         image_url[:120],
         len(prompt),
+        max_tokens,
+        timeout,
     )
     raw = client.vision_completion(
         messages=messages,
         images=[image_url],
-        temperature=0.2,
-        max_tokens=2048,
-        timeout=60,
+        temperature=temperature,
+        max_tokens=max_tokens,
+        timeout=timeout,
+        model=model,
     )
     if raw is None:
         logger.warning("[call_vision] 视觉模型返回 None (image_url=%s)", image_url[:80])
@@ -592,13 +625,14 @@ def call_video_generation(
     - reference_audios 传 TTS 音频 URL 数组做口型驱动；
     - reference_images 传产品素材 URL 数组做视觉参考；
     - 单次最长 30 秒，不分段不拼接；
-    - image_url 存在时为「首帧图生视频」模式，自动不传 ratio（Bug #2110）。
+    - 图生视频（image_url 存在）也强制传 ratio，避免商品方图导致默认输出 1:1。
+      官方文档：图生视频时 ratio 由参数决定，模型会居中裁剪首帧到目标比例。
     """
     client = get_doubao_client()
     if not client.is_available:
         logger.warning("[ai_service] 豆包客户端未配置，跳过视频生成")
         return None
-    effective_ratio = None if image_url else ratio
+    effective_ratio = ratio or "9:16"
     try:
         kwargs: dict = dict(
             prompt=prompt,

scripts/render_env.sh

@@ -57,7 +57,7 @@ if [ "$TARGET_ENV" = "staging" ]; then
 fi
 
 # 共用 secrets 直接导出（如果存在）
-SHARED_SECRETS="OSS_ACCESS_KEY_ID OSS_ACCESS_KEY_SECRET COSYVOICE_API_KEY DASHSCOPE_API_KEY MEDIAKIT_API_KEY DOUBAO_API_KEY DOUBAO_MODEL DOUBAO_BASE_URL DOUBAO_VISION_MODEL WECHAT_APP_ID WECHAT_APP_SECRET TIKHUB_API_KEY APIZERO_API_KEY GPU_WORKER_TOKEN"
+SHARED_SECRETS="OSS_ACCESS_KEY_ID OSS_ACCESS_KEY_SECRET COSYVOICE_API_KEY DASHSCOPE_API_KEY MEDIAKIT_API_KEY DOUBAO_API_KEY DOUBAO_MODEL DOUBAO_FAST_MODEL DOUBAO_BASE_URL DOUBAO_VISION_MODEL DOUBAO_VISION_LITE_MODEL DOUBAO_VISION_USE_LITE WECHAT_APP_ID WECHAT_APP_SECRET TIKHUB_API_KEY APIZERO_API_KEY GPU_WORKER_TOKEN"
 for var in $SHARED_SECRETS; do
   value="${!var:-}"
   # 已经在环境中了，无需额外操作

tests/unit/test_2035_coverage.py

@@ -14,10 +14,13 @@ from packages.shared.ai_client import DoubaoClient
 class _FakeSettings:
     doubao_api_key = "test-key"
     doubao_model = "test-model"
+    doubao_fast_model = "test-fast-model"
     doubao_base_url = "https://ark.cn-beijing.volces.com/api/v3"
     doubao_timeout = 10
     doubao_max_retries = 0
     doubao_vision_model = "test-vision"
+    doubao_vision_lite_model = "test-vision-lite"
+    doubao_vision_use_lite = False
     doubao_embedding_model = "test-embedding"
 
 

tests/unit/test_health_routes.py

@@ -44,7 +44,7 @@ class TestCheckDatabase:
         assert result["type"] == "postgresql"
         assert result["message"] == "Database connection successful"
         mock_psycopg.connect.assert_called_once_with(
-            "postgresql+psycopg://test:test@localhost/test", connect_timeout=3
+            "postgresql://test:test@localhost/test", connect_timeout=3
         )
         mock_cur.execute.assert_called_once_with("SELECT 1")
         mock_conn.close.assert_called_once()
@@ -96,7 +96,7 @@ class TestCheckMigrations:
         assert result["status"] == "healthy"
         assert result["message"] == "Database migrations applied"
         mock_psycopg.connect.assert_called_once_with(
-            "postgresql+psycopg://test:test@localhost/test", connect_timeout=3
+            "postgresql://test:test@localhost/test", connect_timeout=3
         )
         mock_conn.close.assert_called_once()
 

tests/unit/test_viral_video_p0.py

@@ -253,8 +253,8 @@ class TestCallVideoGenerationV16:
             )
             assert result == str(out)
             kwargs = mock_client.video_generation.call_args.kwargs
-            # 首帧模式不传 ratio（Bug #2110）
-            assert "ratio" not in kwargs
+            # 图生视频也必须传 ratio（避免首帧方图导致默认输出 1:1）
+            assert kwargs.get("ratio") == "9:16", f"ratio 应透传，got {kwargs.get('ratio')!r}"
             assert kwargs["image_url"] == "https://img/x.jpg"
             assert kwargs["reference_audios"] == ["https://oss/tts.mp3"]
             assert kwargs["reference_images"] == ["https://img/r1.jpg"]

tests/unit/test_viral_video_routes.py

@@ -61,6 +61,8 @@ def _make_job(job_id: str = "job-1", user_id: str = "u1", status: str = "pending
         "video_ratio": "9:16",
         "video_model": "",
         "credits_cost": 0,
+        "current_stage": "",
+        "phase_message": "",
         "updated_at": None,
         "is_terminal": False,
         "effective_copy_text": "",