fix(vision): #2203 V2快速路径两个P0 bug修复

E2E发现两个P0 bug：
1. as_completed(timeout=...)抛TimeoutError时未被捕获，异常直接冒泡到外层，
   跳过了pro VLM兜底路径，导致3/3返回空结果。
   修复：包裹try/except TimeoutError，cancel未完成futures，继续走pro兜底。

2. 外层_step_image_analysis设置client.max_retries=0不生效（ThreadPool线程
   里可能被其他任务覆盖/共享单例状态），实际仍重试1次（2次调用×6s=12s超时）。
   修复：在vlm_fast_json/vlm_fallback内部强制max_retries=0并在finally里恢复。

3. 收紧超时：fast总超时8→6s、lite JSON 6→5s、OCR 8→5s，让lite不够快时
   尽快走pro兜底（pro单次45s足够）。

本地mock验证：fast命中1.5s返回有效portrait_prompt；fast超时6s正确触发pro fallback。

CHANGES.md

@@ -0,0 +1,2 @@
+Mon Oct  5 04:09:11 PM CST 2026
+2198 lite/pro并行竞速 (commit 9699a1f) — CI rebuild trigger Mon Oct  5 08:09:11 AM UTC 2026

apps/web/src/api/viral-video/types.ts

@@ -176,7 +176,7 @@ export interface ViralVideoJob {
   fusion_level?: FusionLevel
   voice_id?: string
   voice_mode?: "global" | "per_video"
-  voice_source?: "preset" | "library" | "clone" | "upload"
+  voice_source?: "preset" | "library" | "clone" | "upload" | "my_voice"
   bgm_preference?: string
   intent_result?: IntentResult
   intent_text?: string
@@ -212,7 +212,7 @@ export interface GenerateViralVideoRequest {
   user_copy_text?: string
   fusion_level?: FusionLevel
   voice_id?: string
-  voice_source?: "preset" | "library" | "clone" | "upload"
+  voice_source?: "preset" | "library" | "clone" | "upload" | "my_voice"
   bgm_preference?: string
   industry?: string
   target_customer?: string
@@ -244,7 +244,7 @@ export interface AnalyzeImagesRequest {
   /** TTS 音色 ID（STEP1 已选音色时传） */
   voice_id?: string
   /** 音色来源：preset | library | clone | upload */
-  voice_source?: "preset" | "library" | "clone" | "upload"
+  voice_source?: "preset" | "library" | "clone" | "upload" | "my_voice"
   /** Seedance 视频比例：9:16 | 16:9 | 1:1 */
   video_ratio?: string
   /** Seedance 模型 ID（空则使用服务端默认） */
@@ -273,7 +273,7 @@ export interface GenerateCopyRequest {
   /** TTS 音色 ID（优先级高于 persona_id） */
   voice_id?: string
   /** 音色来源：preset | library | clone | upload */
-  voice_source?: "preset" | "library" | "clone" | "upload"
+  voice_source?: "preset" | "library" | "clone" | "upload" | "my_voice"
   /** Seedance 视频比例（9:16/16:9/1:1 等） */
   video_ratio?: string
   /** Seedance 模型 ID（空则使用服务端默认） */

apps/web/src/api/voice-clone/types.ts

@@ -18,6 +18,8 @@ export interface VoiceClone {
   language: string
   gender: string
   error_message: string | null
+  /** CosyVoice 实际使用的音色 ID（status=ready 时由后端填充，用于 TTS 调用） */
+  voice_id?: string | null
   created_at: string
   updated_at: string
 }

apps/web/src/api/voice-clone/utils.ts

@@ -18,6 +18,7 @@ export const toVoiceClone = (profile: VoiceCloneProfile): VoiceClone => ({
   language: profile.language || "",
   gender: profile.gender || "",
   error_message: profile.error_message || null,
+  voice_id: profile.voice_id,
   created_at: profile.created_at,
   updated_at: profile.updated_at,
 })

apps/web/src/components/common/DurationWheelPicker/DurationWheelPicker.css

@@ -0,0 +1,182 @@
+/* DurationWheelPicker —— 弹层式滚轮选择器（样式与表单一致） */
+
+/* 触发按钮：外观复用 .vv-select 风格 */
+.dw-trigger {
+  display: flex;
+  align-items: center;
+  justify-content: space-between;
+  width: 100%;
+  height: 36px;
+  padding: 0 12px;
+  background: #fff;
+  border: 1px solid #e0e0e8;
+  border-radius: 8px;
+  font-size: 13px;
+  color: #1f2937;
+  cursor: pointer;
+  box-sizing: border-box;
+  transition: all 0.15s;
+  user-select: none;
+}
+.dw-trigger:hover {
+  border-color: #c0c0d0;
+}
+.dw-trigger-open,
+.dw-trigger:focus-within {
+  border-color: #7c3aed !important;
+  box-shadow: 0 0 0 2px rgba(124, 58, 237, 0.12);
+}
+.dw-trigger-disabled {
+  opacity: 0.5;
+  pointer-events: none;
+  cursor: not-allowed;
+}
+.dw-trigger-val {
+  flex: 1;
+  overflow: hidden;
+  text-overflow: ellipsis;
+  white-space: nowrap;
+}
+.dw-trigger-placeholder {
+  color: #9ca3af;
+}
+.dw-trigger-arrow {
+  font-size: 10px;
+  color: #9ca3af;
+  margin-left: 8px;
+  transition: transform 0.2s;
+}
+.dw-trigger-arrow-up {
+  transform: rotate(180deg);
+}
+
+/* 弹层容器 */
+.dw-popup {
+  padding: 8px;
+  min-width: 140px;
+}
+
+/* 滚轮 */
+.dw-picker {
+  position: relative;
+  width: 100%;
+  overflow: hidden;
+  border-radius: 8px;
+  background: #fafafe;
+  border: 1px solid #e5e7eb;
+}
+.dw-picker-list {
+  margin: 0;
+  padding: 0;
+  list-style: none;
+  height: 100%;
+  overflow-y: scroll;
+  scroll-snap-type: y mandatory;
+  -webkit-overflow-scrolling: touch;
+  scrollbar-width: none;
+}
+.dw-picker-list::-webkit-scrollbar {
+  display: none;
+}
+.dw-picker-item {
+  display: flex;
+  align-items: baseline;
+  justify-content: center;
+  gap: 3px;
+  scroll-snap-align: center;
+  cursor: pointer;
+  font-size: 15px;
+  color: #9ca3af;
+  font-weight: 400;
+  transition:
+    color 0.15s,
+    transform 0.15s,
+    font-weight 0.15s;
+}
+.dw-picker-item-val {
+  font-variant-numeric: tabular-nums;
+}
+.dw-picker-item-unit {
+  font-size: 13px;
+  color: inherit;
+}
+.dw-picker-item-active {
+  color: #7c3aed;
+  font-weight: 600;
+}
+.dw-picker-item-active .dw-picker-item-val {
+  font-size: 18px;
+}
+.dw-picker-item-active .dw-picker-item-unit {
+  font-size: 14px;
+}
+
+/* 中心选中条 */
+.dw-picker-mask {
+  position: absolute;
+  left: 6px;
+  right: 6px;
+  pointer-events: none;
+  background: #f5f0ff;
+  border-radius: 6px;
+  z-index: 1;
+}
+.dw-picker-mask::before,
+.dw-picker-mask::after {
+  content: "";
+  position: absolute;
+  left: 0;
+  right: 0;
+  height: 1px;
+  background: #d8c4ff;
+}
+.dw-picker-mask::before {
+  top: 0;
+}
+.dw-picker-mask::after {
+  bottom: 0;
+}
+
+/* 上下渐变 */
+.dw-picker-fade {
+  position: absolute;
+  left: 0;
+  right: 0;
+  height: 40%;
+  pointer-events: none;
+  z-index: 2;
+}
+.dw-picker-fade-top {
+  top: 0;
+  background: linear-gradient(to bottom, #fafafe 25%, rgba(250, 250, 254, 0));
+}
+.dw-picker-fade-bottom {
+  bottom: 0;
+  background: linear-gradient(to top, #fafafe 25%, rgba(250, 250, 254, 0));
+}
+
+/* 弹层按钮区 */
+.dw-popup-actions {
+  display: flex;
+  gap: 8px;
+  justify-content: flex-end;
+  margin-top: 8px;
+}
+.dw-popup-actions .ant-btn {
+  border-radius: 6px;
+}
+.dw-popup-actions .ant-btn-primary {
+  background: #7c3aed;
+}
+.dw-popup-actions .ant-btn-primary:hover {
+  background: #6d28d9 !important;
+}
+
+/* 覆盖 antd Popover 默认内边距 */
+.dw-popover .ant-popover-inner {
+  padding: 0 !important;
+  overflow: hidden;
+}
+.dw-popover .ant-popover-arrow {
+  display: none;
+}

apps/web/src/components/common/DurationWheelPicker/index.tsx

@@ -0,0 +1,180 @@
+/**
+ * DurationWheelPicker —— 竖屏滚轮式时长选择器（弹层版）
+ *
+ * 设计：
+ * - 外观是和其他表单 Select 一致的输入框（白色底+1px灰边+紫色focus ring）
+ * - 点击输入框弹出 Popover，内部是滚轮 picker（原生 scroll-snap，零依赖）
+ * - 滚轮样式：白底容器，选中行 #7c3aed 紫字加粗+浅紫背景条
+ * - 支持触摸/鼠标滚轮/点击；松手吸附；底部"确认/取消"按钮
+ * - 默认范围 15–30 秒，步长 1 秒
+ */
+import React, { useEffect, useMemo, useRef, useState, useCallback } from "react"
+import { Popover, Button } from "antd"
+import { DownOutlined } from "@ant-design/icons"
+import "./DurationWheelPicker.css"
+
+export interface DurationWheelPickerProps {
+  value?: number
+  min?: number
+  max?: number
+  step?: number
+  unit?: string
+  onChange?: (value: number) => void
+  placeholder?: string
+  disabled?: boolean
+  /** 弹层宽度，默认 160px */
+  popupWidth?: number
+  /** 弹层内滚轮高度，默认 180px */
+  wheelHeight?: number
+}
+
+const ITEM_HEIGHT = 36
+
+const DurationWheelPicker: React.FC<DurationWheelPickerProps> = ({
+  value = 20,
+  min = 15,
+  max = 30,
+  step = 1,
+  unit = "秒",
+  onChange,
+  placeholder = "请选择时长",
+  disabled = false,
+  popupWidth = 160,
+  wheelHeight = 180,
+}) => {
+  const options = useMemo(() => {
+    const arr: number[] = []
+    for (let v = min; v <= max; v += step) arr.push(v)
+    return arr
+  }, [min, max, step])
+
+  const [open, setOpen] = useState(false)
+  // 弹层内暂存值，点确认才提交
+  const [draft, setDraft] = useState<number>(value)
+  const listRef = useRef<HTMLUListElement>(null)
+  const scrollTimerRef = useRef<ReturnType<typeof setTimeout> | null>(null)
+
+  useEffect(() => {
+    if (open) {
+      setDraft(value)
+      // 下一帧滚到当前值
+      requestAnimationFrame(() => scrollToValue(value, false))
+    }
+    // eslint-disable-next-line react-hooks/exhaustive-deps
+  }, [open])
+
+  const scrollToValue = useCallback(
+    (v: number, smooth = true) => {
+      const list = listRef.current
+      if (!list) return
+      const idx = options.indexOf(v)
+      if (idx < 0) return
+      list.scrollTo({ top: idx * ITEM_HEIGHT, behavior: smooth ? "smooth" : "auto" })
+    },
+    [options],
+  )
+
+  const handleScroll = () => {
+    if (scrollTimerRef.current) clearTimeout(scrollTimerRef.current)
+    scrollTimerRef.current = setTimeout(() => {
+      const list = listRef.current
+      if (!list) return
+      const idx = Math.round(list.scrollTop / ITEM_HEIGHT)
+      const clamped = Math.max(0, Math.min(options.length - 1, idx))
+      const targetTop = clamped * ITEM_HEIGHT
+      if (Math.abs(list.scrollTop - targetTop) > 1) {
+        list.scrollTo({ top: targetTop, behavior: "smooth" })
+      }
+      setDraft(options[clamped])
+    }, 100)
+  }
+
+  const handleConfirm = () => {
+    onChange?.(draft)
+    setOpen(false)
+  }
+
+  const handleCancel = () => {
+    setOpen(false)
+  }
+
+  const handleItemClick = (v: number) => {
+    setDraft(v)
+    scrollToValue(v, true)
+  }
+
+  const maskTop = wheelHeight / 2 - ITEM_HEIGHT / 2
+
+  const wheel = (
+    <div className="dw-popup">
+      <div
+        className="dw-picker"
+        style={{ height: wheelHeight, width: popupWidth - 24 /* padding */ }}
+      >
+        <div className="dw-picker-mask" style={{ top: maskTop, height: ITEM_HEIGHT }} aria-hidden />
+        <div className="dw-picker-fade dw-picker-fade-top" aria-hidden />
+        <div className="dw-picker-fade dw-picker-fade-bottom" aria-hidden />
+        <ul
+          ref={listRef}
+          className="dw-picker-list"
+          onScroll={handleScroll}
+          style={{
+            paddingTop: wheelHeight / 2 - ITEM_HEIGHT / 2,
+            paddingBottom: wheelHeight / 2 - ITEM_HEIGHT / 2,
+          }}
+        >
+          {options.map((v) => {
+            const isActive = v === draft
+            return (
+              <li
+                key={v}
+                className={`dw-picker-item${isActive ? " dw-picker-item-active" : ""}`}
+                style={{ height: ITEM_HEIGHT, lineHeight: `${ITEM_HEIGHT}px` }}
+                onClick={() => handleItemClick(v)}
+                aria-selected={isActive}
+                role="option"
+              >
+                <span className="dw-picker-item-val">{v}</span>
+                <span className="dw-picker-item-unit">{unit}</span>
+              </li>
+            )
+          })}
+        </ul>
+      </div>
+      <div className="dw-popup-actions">
+        <Button size="small" onClick={handleCancel}>
+          取消
+        </Button>
+        <Button size="small" type="primary" onClick={handleConfirm}>
+          确认
+        </Button>
+      </div>
+    </div>
+  )
+
+  return (
+    <Popover
+      open={!disabled && open}
+      onOpenChange={(v) => setOpen(v)}
+      content={wheel}
+      trigger="click"
+      placement="bottomLeft"
+      overlayClassName="dw-popover"
+      overlayStyle={{ padding: 0 }}
+      overlayInnerStyle={{ padding: 0, borderRadius: 10 }}
+      destroyTooltipOnHide
+    >
+      <div
+        className={`dw-trigger${disabled ? " dw-trigger-disabled" : ""}${open ? " dw-trigger-open" : ""}`}
+        style={{ height: 36 }}
+      >
+        <span className={`dw-trigger-val${value != null ? "" : " dw-trigger-placeholder"}`}>
+          {value != null ? `${value}${unit}` : placeholder}
+        </span>
+        <DownOutlined className={`dw-trigger-arrow${open ? " dw-trigger-arrow-up" : ""}`} />
+      </div>
+    </Popover>
+  )
+}
+
+export default DurationWheelPicker

apps/web/src/config/features.ts

@@ -10,4 +10,4 @@
  *   功能流程不做积分预校验，直接走生成。
  * - true：展示完整积分系统 UI。
  */
-export const ENABLE_CREDIT_SYSTEM = false
+export const ENABLE_CREDIT_SYSTEM = true

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

@@ -1834,3 +1834,117 @@
   color: #7c3aed;
   font-weight: 500;
 }
+
+/* ── v1.6 我的音色（默认主路径） ── */
+.vv-voice-section {
+  margin-top: 4px;
+  border: 1px solid #e5e7eb;
+  border-radius: 8px;
+  background: #fff;
+  padding: 10px;
+}
+.vv-voice-section-head {
+  display: flex;
+  align-items: center;
+  justify-content: space-between;
+  margin-bottom: 8px;
+}
+.vv-voice-section-title {
+  font-size: 13px;
+  font-weight: 600;
+  color: #1f2937;
+  display: inline-flex;
+  align-items: center;
+  gap: 6px;
+}
+.vv-voice-section-title .anticon {
+  color: #7c3aed;
+}
+.vv-link-btn-sm {
+  font-size: 12px;
+  padding: 2px 6px;
+  display: inline-flex;
+  align-items: center;
+  gap: 4px;
+}
+.vv-my-voice-list {
+  display: flex;
+  flex-direction: column;
+  gap: 4px;
+  max-height: 220px;
+  overflow-y: auto;
+}
+.vv-my-voice-item {
+  display: flex;
+  align-items: center;
+  gap: 10px;
+  padding: 8px 10px;
+  border-radius: 6px;
+  cursor: pointer;
+  transition: background 0.15s;
+}
+.vv-my-voice-item:hover {
+  background: #f5f0ff;
+}
+.vv-my-voice-item.selected {
+  background: #f5f0ff;
+}
+.vv-my-voice-item.selected .vv-voice-radio {
+  border-color: #7c3aed;
+  background: #7c3aed;
+  box-shadow: inset 0 0 0 2px #fff;
+}
+.vv-voice-empty {
+  padding: 16px 12px;
+  text-align: center;
+  color: #9ca3af;
+  font-size: 12px;
+  display: flex;
+  flex-direction: column;
+  align-items: center;
+  gap: 8px;
+}
+.vv-voice-empty-ic {
+  font-size: 28px;
+  color: #d1d5db;
+}
+.vv-voice-empty-text {
+  font-size: 12px;
+}
+.vv-voice-alt-row {
+  display: flex;
+  gap: 8px;
+  margin-top: 8px;
+}
+.vv-voice-alt-btn {
+  flex: 1;
+  display: inline-flex;
+  align-items: center;
+  justify-content: center;
+  gap: 6px;
+  padding: 8px 10px;
+  background: #fafafe;
+  border: 1px solid #e5e7eb;
+  border-radius: 6px;
+  color: #6b7280;
+  font-size: 12px;
+  cursor: pointer;
+  transition: all 0.15s;
+}
+.vv-voice-alt-btn:hover {
+  border-color: #7c3aed;
+  color: #7c3aed;
+  background: #f5f0ff;
+}
+.vv-voice-alt-btn.selected {
+  background: #f5f0ff;
+  border-color: #7c3aed;
+  color: #7c3aed;
+}
+.vv-voice-panel-actions {
+  display: flex;
+  gap: 12px;
+  margin-bottom: 6px;
+  padding-bottom: 6px;
+  border-bottom: 1px dashed #e5e7eb;
+}

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

@@ -23,10 +23,13 @@ import {
   CheckCircleFilled,
   EditOutlined,
   HistoryOutlined,
+  UserOutlined,
 } from "@ant-design/icons"
 import { Select, Input, message } from "antd"
 import { uploadAssetDirect, getAssetLibraries, getAssetsByKind, type AssetItem } from "@/api/assets"
 import { fetchPresetVoices } from "@/api/voices"
+import { getVoiceClones, getVoiceClonePreview } from "@/api/voice-clone"
+import type { VoiceClone } from "@/api/voice-clone"
 import {
   FUSION_LEVELS,
   STYLE_STRENGTHS,
@@ -94,13 +97,15 @@ type RefVideo = {
 } | null
 
 type RefAudio = {
-  source: "preset" | "library" | "upload" | "clone"
+  source: "preset" | "library" | "upload" | "clone" | "my_voice"
   id?: string
   name: string
   url?: string
+  /** 克隆音色的 CosyVoice voice_id（source=my_voice 时使用） */
+  ttsVoiceId?: string
 } | null
 
-type VoicePanel = "preset" | "upload" | "library" | "record" | "douyin" | null
+type VoicePanel = "my_voices" | "preset" | "upload" | "library" | "record" | "douyin" | null
 
 type TabTask = {
   id: string
@@ -213,12 +218,12 @@ const PURPOSES = [
   "悬念短剧",
   "情绪短片",
 ]
-const DURATIONS = [15, 20, 30, 45, 60]
 const RATIOS = [
   { v: "9:16", label: "9:16 竖屏（抖音/视频号）" },
   { v: "16:9", label: "16:9 横屏（B站/YouTube）" },
   { v: "1:1", label: "1:1 方形（小红书）" },
 ]
+const DURATIONS = Array.from({ length: 16 }, (_, i) => 15 + i)
 /** 兜底模型列表（接口未返回时使用，字段与 ViralVideoModel 对齐；后端返回后自动覆盖） */
 const FALLBACK_VIDEO_MODELS: ViralVideoModel[] = [
   {
@@ -432,7 +437,7 @@ const emptyTask = (id: string, title: string): TabTask => ({
   language: "中文(普通话)",
   viralStructure: STRUCTURES[0],
   marketingPurpose: "",
-  duration: 15,
+  duration: 20,
   persona: "",
   videoRatio: "9:16",
   videoModel: "seedance-2.5",
@@ -485,6 +490,9 @@ const ViralVideoPage: React.FC = () => {
     { id: string; name: string; url: string; desc?: string }[]
   >([])
   const [presetVoices, setPresetVoices] = useState<PresetVoice[]>([])
+  /** 我的克隆音色（GET /voice-clones?status=ready） */
+  const [myVoices, setMyVoices] = useState<VoiceClone[]>([])
+  const [myVoicesLoading, setMyVoicesLoading] = useState(false)
   const [voicePickerOpen, setVoicePickerOpen] = useState(false)
   const [videoModels, setVideoModels] = useState<ViralVideoModel[]>(FALLBACK_VIDEO_MODELS)
   const [assetPicker, setAssetPicker] = useState<{
@@ -514,8 +522,18 @@ const ViralVideoPage: React.FC = () => {
     [activeId],
   )
 
-  /* ── 加载音色 ── */
+  /* ── 加载音色（我的克隆音色 + 上传的配音素材 + 预设音色） ── */
+  const reloadMyVoices = useCallback(() => {
+    setMyVoicesLoading(true)
+    getVoiceClones({ status: "ready", limit: 50 })
+      .then((list) => {
+        setMyVoices(list.filter((v) => v.status === "ready"))
+      })
+      .catch(() => setMyVoices([]))
+      .finally(() => setMyVoicesLoading(false))
+  }, [])
   useEffect(() => {
+    reloadMyVoices()
     getAssetsByKind("voice", { limit: 50 })
       .then((list) =>
         setVoiceAssets(
@@ -523,7 +541,7 @@ const ViralVideoPage: React.FC = () => {
             id: a.id,
             name: a.name,
             url: a.file_url || "",
-            desc: a.duration ? `${Math.round(a.duration)}s` : "我的音色",
+            desc: a.duration ? `${Math.round(a.duration)}s` : "配音文件",
           })),
         ),
       )
@@ -544,7 +562,7 @@ const ViralVideoPage: React.FC = () => {
         // 接口失败兜底：弹窗内 MOCK_VOICES 会生效（voice_id 与后端 CosyVoice 一致）
         setPresetVoices([])
       })
-  }, [])
+  }, [reloadMyVoices])
 
   /* ── 加载视频模型列表 ── */
   useEffect(() => {
@@ -914,21 +932,36 @@ const ViralVideoPage: React.FC = () => {
     [setTask, uploadFile],
   )
 
-  const onCloneSuccess = (voice: { id?: string; name?: string }) => {
+  const onCloneSuccess = (
+    voice: VoiceClone | { id?: string; name?: string; voice_id?: string | null },
+  ) => {
     setCloneModalOpen(false)
+    // 克隆成功后刷新我的音色列表
+    reloadMyVoices()
+    const vid = "voice_id" in voice ? voice.voice_id : undefined
     if (voice?.id) {
       setTask({
-        refAudio: { source: "clone", id: voice.id, name: voice.name || "我录制的音色" },
+        refAudio: {
+          source: "my_voice",
+          id: voice.id,
+          name: voice.name || "我录制的音色",
+          ttsVoiceId: vid || undefined,
+        },
         voicePanel: null,
       })
       getAssetsByKind("voice", { limit: 50 })
         .then((list) =>
           setVoiceAssets(
-            list.map((a) => ({ id: a.id, name: a.name, url: a.file_url || "", desc: "我的音色" })),
+            list.map((a) => ({
+              id: a.id,
+              name: a.name,
+              url: a.file_url || "",
+              desc: a.duration ? `${Math.round(a.duration)}s` : "配音文件",
+            })),
           ),
         )
         .catch(() => {})
-      message.success("录音已完成，已自动选中")
+      message.success("音色克隆完成，已自动选中")
     }
   }
 
@@ -1014,7 +1047,12 @@ const ViralVideoPage: React.FC = () => {
             : undefined,
         fusion_level: task.fusionLevel,
         style_strength: task.refVideo?.ossUrl ? task.styleStrength : undefined,
-        voice_id: task.refAudio?.source === "preset" ? task.refAudio.id : undefined,
+        voice_id:
+          task.refAudio?.source === "preset"
+            ? task.refAudio.id
+            : task.refAudio?.source === "my_voice"
+              ? task.refAudio.ttsVoiceId || task.refAudio.id
+              : undefined,
         voice_source: task.refAudio?.source || undefined,
         video_ratio: task.videoRatio,
         video_model: task.videoModel,
@@ -1058,7 +1096,12 @@ const ViralVideoPage: React.FC = () => {
           style_strength: task.refVideo?.ossUrl ? task.styleStrength : undefined,
           user_copy_text: task.storyboard?.voiceover_script || task.userCopy || undefined,
           fusion_level: task.fusionLevel,
-          voice_id: task.refAudio?.id || undefined,
+          voice_id:
+            task.refAudio?.source === "preset"
+              ? task.refAudio.id
+              : task.refAudio?.source === "my_voice"
+                ? task.refAudio.ttsVoiceId || task.refAudio.id
+                : undefined,
           voice_source: task.refAudio?.source,
           industry: task.industry || undefined,
           target_customer: task.targetCustomer || undefined,
@@ -2033,43 +2076,119 @@ const ViralVideoPage: React.FC = () => {
                 )}
               </div>
 
-              {/* 参考音频 2x2 */}
+              {/* 配音选择（默认"我的音色"） */}
               <div className="vv-subblock">
                 <div className="vv-subblock-head">
-                  <span className="vv-subblock-title">参考音频</span>
+                  <span className="vv-subblock-title">AI 配音</span>
                   <span className="vv-opt-tag">选填</span>
-                  <span className="vv-subblock-hint">最大 10MB · 最长 30 秒</span>
+                  <span className="vv-subblock-hint">
+                    默认使用我的克隆音色，也可选择内置音色或上传配音
+                  </span>
                 </div>
-                <div className="vv-audio-grid">
+
+                {/* 我的音色（主路径） */}
+                <div className="vv-voice-section">
+                  <div className="vv-voice-section-head">
+                    <span className="vv-voice-section-title">
+                      <UserOutlined /> 我的音色
+                    </span>
+                    <button
+                      className="vv-link-btn vv-link-btn-sm"
+                      onClick={() => setCloneModalOpen(true)}
+                    >
+                      <PlusOutlined /> 录制新音色
+                    </button>
+                  </div>
+                  {myVoicesLoading ? (
+                    <div className="vv-muted vv-voice-empty">加载中…</div>
+                  ) : myVoices.length === 0 ? (
+                    <div className="vv-voice-empty">
+                      <AudioMutedOutlined className="vv-voice-empty-ic" />
+                      <div className="vv-voice-empty-text">还没有克隆音色</div>
+                      <button
+                        className="vv-btn vv-btn-primary vv-btn-sm"
+                        onClick={() => setCloneModalOpen(true)}
+                      >
+                        去录制一个
+                      </button>
+                    </div>
+                  ) : (
+                    <div className="vv-my-voice-list">
+                      {myVoices.map((v) => {
+                        const selected =
+                          task.refAudio?.source === "my_voice" && task.refAudio.id === v.id
+                        const playing = task.playingVoiceId === `my-${v.id}`
+                        return (
+                          <div
+                            key={v.id}
+                            className={`vv-my-voice-item ${selected ? "selected" : ""}`}
+                            onClick={() =>
+                              setTask({
+                                refAudio: {
+                                  source: "my_voice",
+                                  id: v.id,
+                                  name: v.name,
+                                  ttsVoiceId: v.voice_id || undefined,
+                                },
+                                voicePanel: null,
+                              })
+                            }
+                          >
+                            <div className="vv-voice-radio" />
+                            <div className="vv-voice-info">
+                              <div className="vv-voice-name">{v.name}</div>
+                              <div className="vv-voice-desc">
+                                AI 克隆声音
+                                {v.gender
+                                  ? ` · ${v.gender === "male" ? "男" : v.gender === "female" ? "女" : ""}`
+                                  : ""}
+                              </div>
+                            </div>
+                            <button
+                              className={`vv-voice-play ${playing ? "playing" : ""}`}
+                              onClick={(e) => {
+                                e.stopPropagation()
+                                // 克隆音色试听：优先 sample_url，否则调用 preview 接口
+                                if (v.sample_url) {
+                                  toggleVoice({ id: `my-${v.id}`, url: v.sample_url })
+                                } else {
+                                  message.info("正在获取试听音频…")
+                                  getVoiceClonePreview(v.id)
+                                    .then((r) => {
+                                      if (r.audio_url)
+                                        toggleVoice({ id: `my-${v.id}`, url: r.audio_url })
+                                    })
+                                    .catch(() => message.error("试听获取失败"))
+                                }
+                              }}
+                            >
+                              {playing ? <PauseCircleOutlined /> : <PlayCircleOutlined />}
+                            </button>
+                          </div>
+                        )
+                      })}
+                    </div>
+                  )}
+                </div>
+
+                {/* 次要选项：内置音色 / 上传配音 */}
+                <div className="vv-voice-alt-row">
                   <button
-                    className={`vv-audio-cell ${task.refAudio?.source === "preset" ? "selected" : ""}`}
+                    className={`vv-voice-alt-btn ${task.refAudio?.source === "preset" ? "selected" : ""}`}
                     onClick={() => setVoicePickerOpen(true)}
                   >
-                    <FileTextOutlined className="vv-audio-ic" />
-                    <span>选择内置音色</span>
+                    <FileTextOutlined /> 内置音色
                   </button>
                   <button
-                    className={`vv-audio-cell ${task.voicePanel === "upload" ? "active" : ""} ${task.refAudio?.source === "upload" ? "selected" : ""}`}
-                    onClick={() => voiceInputRef.current?.click()}
+                    className={`vv-voice-alt-btn ${task.refAudio?.source === "upload" || task.refAudio?.source === "library" ? "selected" : ""}`}
+                    onClick={() =>
+                      setTask({ voicePanel: task.voicePanel === "library" ? null : "library" })
+                    }
                   >
-                    <UploadOutlined className="vv-audio-ic" />
-                    <span>本地上传</span>
-                  </button>
-                  <button
-                    className={`vv-audio-cell ${task.voicePanel === "library" ? "active" : ""} ${task.refAudio?.source === "library" ? "selected" : ""}`}
-                    onClick={() => setAssetPicker({ open: true, kind: "voice", multiple: false })}
-                  >
-                    <FolderOpenOutlined className="vv-audio-ic" />
-                    <span>从素材库选择</span>
-                  </button>
-                  <button
-                    className={`vv-audio-cell ${task.voicePanel === "record" ? "active" : ""} ${task.refAudio?.source === "clone" ? "selected" : ""}`}
-                    onClick={() => setCloneModalOpen(true)}
-                  >
-                    <AudioMutedOutlined className="vv-audio-ic" />
-                    <span>直接录音</span>
+                    <FolderOpenOutlined /> 上传配音
                   </button>
                 </div>
+
                 <input
                   ref={voiceInputRef}
                   type="file"
@@ -2083,7 +2202,14 @@ const ViralVideoPage: React.FC = () => {
                 {task.refAudio && (
                   <div className="vv-selected-audio">
                     <SoundOutlined style={{ color: "#7c3aed", marginRight: 6 }} />
-                    <span style={{ flex: 1 }}>已选：{task.refAudio.name}</span>
+                    <span style={{ flex: 1 }}>
+                      已选：
+                      {task.refAudio.source === "my_voice"
+                        ? `我的音色 · ${task.refAudio.name}`
+                        : task.refAudio.source === "preset"
+                          ? `内置 · ${task.refAudio.name}`
+                          : task.refAudio.name}
+                    </span>
                     <button className="vv-link-btn" onClick={() => setTask({ refAudio: null })}>
                       移除
                     </button>
@@ -2091,16 +2217,32 @@ const ViralVideoPage: React.FC = () => {
                 )}
                 {task.voicePanel === "library" && (
                   <div className="vv-voice-panel">
+                    <div className="vv-voice-panel-actions">
+                      <button
+                        className="vv-link-btn vv-link-btn-sm"
+                        onClick={() => voiceInputRef.current?.click()}
+                      >
+                        <UploadOutlined /> 本地上传音频文件
+                      </button>
+                      <button
+                        className="vv-link-btn vv-link-btn-sm"
+                        onClick={() =>
+                          setAssetPicker({ open: true, kind: "voice", multiple: false })
+                        }
+                      >
+                        <FolderOpenOutlined /> 从素材库选择
+                      </button>
+                    </div>
                     {voiceAssets.length === 0 ? (
                       <div className="vv-muted" style={{ padding: 12, textAlign: "center" }}>
-                        暂无已上传的音频
+                        暂无已上传的配音文件
                       </div>
                     ) : (
                       <div className="vv-voice-list">
                         {voiceAssets.map((v) => (
                           <div
                             key={v.id}
-                            className={`vv-voice-item ${task.refAudio?.id === v.id && task.refAudio.source === "library" ? "selected" : ""}`}
+                            className={`vv-voice-item ${task.refAudio?.id === v.id && (task.refAudio.source === "library" || task.refAudio.source === "upload") ? "selected" : ""}`}
                             onClick={() =>
                               pickVoiceAsset({
                                 id: v.id,
@@ -2289,7 +2431,7 @@ const ViralVideoPage: React.FC = () => {
                     options={PURPOSES.map((i) => ({ value: i, label: i }))}
                   />
                 </div>
-                <div className="vv-form-row" style={{ gridColumn: "1 / -1" }}>
+                <div className="vv-form-row">
                   <label className="vv-label">文案视频时长</label>
                   <Select
                     className="vv-select"

apps/worker/worker_app/tasks/vision/__init__.py

@@ -0,0 +1,4 @@
+# -*- coding: utf-8 -*-
+"""V2 图片分析：火山OCR专用API + doubao-lite强约束JSON并行，单次pro VLM兜底。"""
+
+from .fast_path import analyze_image_v2, analyze_images_v2  # noqa: F401

apps/worker/worker_app/tasks/vision/assembler.py

@@ -0,0 +1,307 @@
+# -*- coding: utf-8 -*-
+"""把 fast_json VLM 输出 + OCR 文本组装为与旧 _normalize() 完全一致的 dict。
+
+目标：下游（信任链t2i/intent_parsing/script_generation）零改动。
+必出字段：name, brand, category, appearance, packaging, text_on_package,
+         key_features, scene, mood, portrait_prompt, summary, _source
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+# ---------- portrait_prompt 模板 ----------
+# 目标：60-100 字的人物穿搭描述，用于 Seedream 纯文生图。要求具体、风格化、视觉细节丰富。
+# 旧 VLM 输出格式参考："一位25岁左右的亚洲女性，身穿白色V领短袖T恤，黑色高腰阔腿裤，
+# 搭配银色项链，长发披肩，表情自信，街拍风格，阳光明媚的城市街头"
+
+
+def _join_parts(*parts: str | None) -> str:
+    return "".join(p for p in parts if p)
+
+
+_AGE_PREFIX = {
+    "青年": "年轻",
+    "中年": "中年",
+    "老年": "老年",
+}
+# gender 后缀
+_GENDER_WORD = {"男": "男性", "女": "女性"}
+
+
+def _person_subject(fj: dict[str, Any]) -> str:
+    """人物主语：年轻女性 / 中年男性 / 少女 / 小男孩 / 人物 等。"""
+    gender = fj.get("gender") or ""
+    age = fj.get("age_range") or ""
+    gw = _GENDER_WORD.get(gender, "")
+    if age == "儿童":
+        if gender == "女":
+            return "小女孩"
+        if gender == "男":
+            return "小男孩"
+        return "儿童"
+    if age == "青少年":
+        if gender == "女":
+            return "少女"
+        if gender == "男":
+            return "少年"
+        return "青少年"
+    prefix = _AGE_PREFIX.get(age, "")
+    if gw:
+        return f"{prefix}{gw}" if prefix else gw
+    return f"{prefix}人物" if prefix else "人物"
+
+
+def _build_wear_sentence(fj: dict[str, Any]) -> str:
+    """穿搭段：上装+下装/连衣裙，带颜色+材质+图案。"""
+    upper = fj.get("upper_wear") or ""
+    upper_color = fj.get("upper_color") or ""
+    lower = fj.get("lower_wear") or ""
+    lower_color = fj.get("lower_color") or ""
+    dress_color = fj.get("dress_color") or ""
+    material = fj.get("material") or ""
+    pattern = fj.get("pattern") or ""
+
+    is_dress = ("连衣裙" in upper) or ("裙" in upper and not lower)
+    if is_dress:
+        c = dress_color or upper_color
+        wear = f"{c}{upper}" if c else upper
+        if material and material not in wear:
+            wear = f"{material}{wear}"
+        if pattern and pattern not in wear and pattern != "纯色":
+            wear += f"，{pattern}图案"
+        return f"身穿{wear}"
+
+    parts: list[str] = []
+    if upper:
+        up = f"{upper_color}{upper}" if upper_color else upper
+        if material and material not in up:
+            up = f"{material}{up}"
+        if pattern and pattern != "纯色" and pattern not in up:
+            up += f"（{pattern}）"
+        parts.append(f"上身{up}" if up else "")
+    if lower:
+        lo = f"{lower_color}{lower}" if lower_color else lower
+        parts.append(f"下身{lo}" if lo else "")
+    return "，".join(p for p in parts if p)
+
+
+def _build_portrait_prompt(fj: dict[str, Any]) -> str:
+    """组装最终 portrait_prompt（目标 60-100 字，用于 Seedream 纯文生图）。"""
+    if not fj.get("has_person"):
+        # 非人像：用商品+场景+mood 拼一段
+        name = fj.get("product_name") or "商品"
+        brand = fj.get("brand") or ""
+        colors = fj.get("colors") or []
+        style = fj.get("style") or ""
+        scene = fj.get("scene") or ""
+        mood = fj.get("mood") or ""
+        pieces = []
+        if brand:
+            pieces.append(brand)
+        pieces.append(name)
+        if colors:
+            pieces.append("、".join(colors[:3]) + "配色")
+        if style:
+            pieces.append(style + "风格")
+        if mood:
+            pieces.append(mood + "氛围")
+        if scene and scene not in ("通用",):
+            pieces.append(scene + "场景")
+        pieces.append("产品特写")
+        prompt = "，".join(p for p in pieces if p)
+        return prompt if len(prompt) >= 10 else "产品展示图，特写镜头"
+
+    subject = _person_subject(fj)
+    wear = _build_wear_sentence(fj)
+
+    accessories = fj.get("accessories") or []
+    if isinstance(accessories, str):
+        accessories = [accessories]
+    acc_str = ""
+    if accessories:
+        acc_str = "，佩戴" + "、".join(str(a) for a in accessories if a)
+
+    hairstyle = fj.get("hairstyle") or ""
+    expression = fj.get("expression") or ""
+    pose = fj.get("pose") or ""
+    style = fj.get("style") or ""
+    scene = fj.get("scene") or ""
+    mood = fj.get("mood") or ""
+
+    detail_parts: list[str] = []
+    if hairstyle:
+        detail_parts.append(hairstyle)
+    if expression and expression not in ("自然", "平静"):
+        detail_parts.append(f"神情{expression}")
+    if pose and pose not in ("站立",):
+        detail_parts.append(pose)
+
+    style_parts: list[str] = []
+    if style:
+        style_parts.append(style)
+    if mood:
+        style_parts.append(mood)
+    if scene and scene not in ("通用",):
+        style_parts.append(scene)
+
+    pieces = [f"一位{subject}"]
+    if wear:
+        pieces.append(wear)
+    if acc_str:
+        pieces.append(acc_str.lstrip("，"))
+    if detail_parts:
+        pieces.append("，".join(detail_parts))
+    if style_parts:
+        # 风格词之间不用逗号，用空格紧凑
+        pieces.append("".join(style_parts) + "风格")
+    else:
+        pieces.append("人像写真")
+
+    full = "，".join(p for p in pieces if p)
+    # 过短补充镜头词
+    if len(full) < 40:
+        full += "，自然光线下人像特写，画面清晰"
+    # 过长截断
+    if len(full) > 120:
+        full = full[:120].rstrip("，") + "。"
+    return full
+
+
+# ---------- 商品字段 ----------
+
+
+def _infer_name(fj: dict[str, Any], ocr_texts: list[str]) -> str:
+    pname = fj.get("product_name")
+    if pname and pname != "未识别":
+        return str(pname)
+    # 人物图 → name 用穿搭主件
+    if fj.get("has_person"):
+        up = fj.get("upper_wear") or ""
+        if "连衣裙" in up:
+            return up
+        return up or "人物穿搭"
+    if ocr_texts:
+        # 商品名可能是 OCR 最长的一行（品牌/产品名）
+        return max(ocr_texts, key=len)
+    return "未识别"
+
+
+def _infer_brand(fj: dict[str, Any], ocr_texts: list[str]) -> str:
+    brand = fj.get("brand")
+    if brand:
+        return str(brand)
+    # OCR 里短的、纯字母/汉字短串可能是 brand
+    for t in ocr_texts:
+        if 1 < len(t) <= 12:
+            return t
+    return "无法判断"
+
+
+def _infer_category(fj: dict[str, Any]) -> str:
+    cat = fj.get("category")
+    if cat:
+        return str(cat)
+    if fj.get("has_person"):
+        return "服饰"
+    return "非产品图"
+
+
+def _build_appearance(fj: dict[str, Any]) -> str:
+    """外观描述：颜色+款式+材质+图案 拼成一段。"""
+    parts: list[str] = []
+    for key, _label in [
+        ("upper_color", "主色"),
+        ("upper_wear", "款式"),
+        ("material", "材质"),
+        ("pattern", "图案"),
+    ]:
+        v = fj.get(key)
+        if v and v not in ("无法判断", "未知", "纯色"):
+            parts.append(str(v))
+    if not parts:
+        if fj.get("has_person"):
+            return "人像穿搭整体造型"
+        return "无法判断"
+    return "、".join(parts)
+
+
+def _build_key_features(fj: dict[str, Any], ocr_texts: list[str]) -> list[str]:
+    feats: list[str] = []
+    for key in (
+        "upper_wear",
+        "lower_wear",
+        "upper_color",
+        "lower_color",
+        "dress_color",
+        "material",
+        "pattern",
+        "style",
+        "accessories",
+    ):
+        v = fj.get(key)
+        if not v:
+            continue
+        if isinstance(v, list):
+            feats.extend(str(x) for x in v if x)
+        elif isinstance(v, str) and v not in ("无法判断", "未知", "纯色"):
+            feats.append(v)
+    if ocr_texts:
+        feats.append(f"画面文字: {'/'.join(ocr_texts[:3])}")
+    # 去重
+    out: list[str] = []
+    seen: set[str] = set()
+    for f in feats:
+        f = f.strip()
+        if f and f not in seen and len(f) <= 30:
+            seen.add(f)
+            out.append(f)
+    return out[:6] if out else ["无法判断"]
+
+
+def assemble_result(
+    idx: int,
+    fast_json: dict[str, Any] | None,
+    ocr_texts: list[str],
+) -> dict[str, Any]:
+    """把 fast_json 结果 + OCR 文本组装成下游兼容的 product dict。"""
+    fj = fast_json or {}
+    ocr_texts = ocr_texts or []
+
+    portrait_prompt = _build_portrait_prompt(fj)
+    name = _infer_name(fj, ocr_texts)
+    brand = _infer_brand(fj, ocr_texts)
+    category = _infer_category(fj)
+    appearance = _build_appearance(fj)
+    key_features = _build_key_features(fj, ocr_texts)
+    scene = fj.get("scene") or "通用"
+    mood = fj.get("mood") or ""
+    packaging = "无法判断"  # 包装细节专用API无，保留占位
+    text_on_package = ocr_texts[:8]
+    summary = _build_summary(fj, name, brand, category)
+
+    return {
+        "name": name,
+        "brand": brand,
+        "category": category,
+        "appearance": appearance,
+        "packaging": packaging,
+        "text_on_package": text_on_package,
+        "key_features": key_features,
+        "scene": scene,
+        "mood": mood,
+        "portrait_prompt": portrait_prompt,
+        "summary": summary,
+        "_source": "v2_fast_json",
+    }
+
+
+def _build_summary(fj: dict, name: str, brand: str, category: str) -> str:
+    if fj.get("has_person"):
+        up = fj.get("upper_wear") or "穿搭"
+        style = fj.get("style") or ""
+        base = f"{style}{up}" if style and style not in up else up
+        return base
+    if brand != "无法判断" and name != brand:
+        return f"{brand} {name}"
+    return name

apps/worker/worker_app/tasks/vision/fast_path.py

@@ -0,0 +1,144 @@
+# -*- coding: utf-8 -*-
+"""V2 图片分析主路径：每图并行 OCR（火山专用API）+ lite JSON VLM，失败时单次 pro VLM 兜底。
+
+设计原则（灵应10-05要求）：
+- 主力路径简洁：单图2路并行，外层N图全并发
+- 兜底简单：单次 pro VLM 调用，无竞速/重试/复杂超时
+- 输出 dict 格式与旧版完全一致，下游零改动
+"""
+
+from __future__ import annotations
+
+import logging
+import os
+import time
+from concurrent.futures import ThreadPoolExecutor, as_completed
+from typing import Any
+
+from . import assembler, ocr_volc, vlm_fallback, vlm_fast_json
+
+logger = logging.getLogger(__name__)
+
+# 可通过环境变量调参（有默认值，无需配置即可跑）
+_IMG_WORKERS = int(os.environ.get("VISION_V2_IMG_WORKERS", "8"))
+_FAST_TIMEOUT = float(os.environ.get("VISION_V2_FAST_TIMEOUT", "6"))
+_FAST_JSON_TIMEOUT = float(os.environ.get("VISION_V2_FAST_JSON_TIMEOUT", "5"))
+_OCR_TIMEOUT = float(os.environ.get("VISION_V2_OCR_TIMEOUT", "5"))
+_PRO_TIMEOUT = float(os.environ.get("VISION_V2_PRO_TIMEOUT", "45"))
+
+_FALLBACK_RESULT = {
+    "name": "未识别",
+    "brand": "无法判断",
+    "category": "非产品图",
+    "appearance": "无法判断",
+    "packaging": "无法判断",
+    "text_on_package": [],
+    "key_features": ["无法判断"],
+    "scene": "通用",
+    "mood": "",
+    "portrait_prompt": "无法判断",
+    "summary": "未识别",
+}
+
+
+def _is_usable(r: dict[str, Any]) -> bool:
+    """结果可用判定：portrait_prompt 是核心，有效就算 usable。"""
+    pp = (r.get("portrait_prompt") or "").strip()
+    if pp and pp not in ("无人像", "无法判断", "未识别"):
+        return True
+    name = (r.get("name") or "").strip()
+    if name and name not in ("未识别", "无法判断", "未知"):
+        return True
+    return False
+
+
+def analyze_image_v2(idx: int, img_url: str) -> dict[str, Any]:
+    """单张图片 V2 分析。"""
+    t0 = time.time()
+
+    # 第1层：OCR + lite JSON VLM 并行
+    fj_result: dict[str, Any] | None = None
+    ocr_result: list[str] = []
+    with ThreadPoolExecutor(max_workers=2) as pool:
+        f_fj = pool.submit(vlm_fast_json.call_fast_json, img_url, timeout=_FAST_JSON_TIMEOUT)
+        f_ocr = pool.submit(ocr_volc.call_ocr, img_url, timeout=_OCR_TIMEOUT)
+        try:
+            for fut in as_completed([f_fj, f_ocr], timeout=_FAST_TIMEOUT):
+                try:
+                    res = fut.result(timeout=1)
+                except Exception as e:
+                    logger.warning("[vision.v2] 图片 #%d 子任务异常: %s", idx, e)
+                    continue
+                if fut is f_fj and isinstance(res, dict):
+                    fj_result = res
+                elif fut is f_ocr and isinstance(res, list):
+                    ocr_result = res
+        except TimeoutError:
+            # fast 整体超时，取消还没跑完的子任务，继续走 pro 兜底
+            for f in (f_fj, f_ocr):
+                if not f.done():
+                    f.cancel()
+            logger.warning("[vision.v2] 图片 #%d fast路径超时(%.0fs)，走pro兜底", idx, _FAST_TIMEOUT)
+
+    fast_elapsed = time.time() - t0
+
+    # 组装 fast 结果
+    if fj_result:
+        assembled = assembler.assemble_result(idx, fj_result, ocr_result)
+        if _is_usable(assembled):
+            assembled["_fast_elapsed"] = round(fast_elapsed, 2)
+            logger.info(
+                "[vision.v2] 图片 #%d fast命中 elapsed=%.2fs pp=%s",
+                idx,
+                fast_elapsed,
+                (assembled.get("portrait_prompt") or "")[:40],
+            )
+            return assembled
+
+    # 第2层：pro VLM 单次兜底
+    pro_t0 = time.time()
+    pro_result = vlm_fallback.call_pro_vlm(img_url, idx, timeout=_PRO_TIMEOUT)
+    if pro_result and _is_usable(pro_result):
+        pro_result["_fallback_used"] = True
+        pro_result["_fast_elapsed"] = round(fast_elapsed, 2)
+        pro_result["_pro_elapsed"] = round(time.time() - pro_t0, 2)
+        if ocr_result and not pro_result.get("text_on_package"):
+            pro_result["text_on_package"] = ocr_result[:8]
+        logger.info("[vision.v2] 图片 #%d pro兜底命中 total=%.2fs", idx, time.time() - t0)
+        return pro_result
+
+    # 最终：返回最小可用结果
+    logger.warning("[vision.v2] 图片 #%d 全路径失败 elapsed=%.2fs", idx, time.time() - t0)
+    out = dict(_FALLBACK_RESULT)
+    out["_source"] = "v2_all_failed"
+    out["text_on_package"] = ocr_result[:8]
+    out["_fast_elapsed"] = round(fast_elapsed, 2)
+    return out
+
+
+def analyze_images_v2(img_urls: list[str]) -> list[dict[str, Any]]:
+    """批量图片 V2 分析，外层全并发。"""
+    if not img_urls:
+        return []
+    workers = min(_IMG_WORKERS, len(img_urls), 16)
+    results: list[dict[str, Any] | None] = [None] * len(img_urls)
+
+    logger.info("[vision.v2] 开始图片分析 n=%d workers=%d fast_timeout=%.0fs", len(img_urls), workers, _FAST_TIMEOUT)
+    t0 = time.time()
+    with ThreadPoolExecutor(max_workers=workers) as pool:
+        future_to_idx = {pool.submit(analyze_image_v2, idx, url): idx for idx, url in enumerate(img_urls)}
+        for fut in as_completed(future_to_idx):
+            idx = future_to_idx[fut]
+            try:
+                results[idx] = fut.result()
+            except Exception as e:
+                logger.warning("[vision.v2] 图片 #%d future异常: %s", idx, e, exc_info=True)
+                r = dict(_FALLBACK_RESULT)
+                r["_source"] = "v2_future_exception"
+                results[idx] = r
+
+    elapsed = time.time() - t0
+    succ = sum(1 for r in results if r and _is_usable(r))
+    fb = sum(1 for r in results if r and r.get("_fallback_used"))
+    logger.info("[vision.v2] 完成 n=%d usable=%d pro_fallback=%d elapsed=%.2fs", len(img_urls), succ, fb, elapsed)
+    return [r for r in results if r is not None]

apps/worker/worker_app/tasks/vision/ocr_volc.py

@@ -0,0 +1,109 @@
+# -*- coding: utf-8 -*-
+"""火山引擎 AI MediaKit OCR（同步）调用封装。
+
+接口：POST {mediakit_base_url}/tools-sync/ocr
+鉴权：Bearer {mediakit_api_key}
+请求体：{"image_url": "<公网可访问URL>"}  （部分版本也支持 image_base64）
+响应：{"code":0,"data":{"texts":[{"text":"...","bbox":[x,y,w,h],...},...],...}}
+
+目标：识别商品包装/Logo/水印上的文字，作为 fast_json VLM 的补充。
+返回值：识别到的文本字符串列表（失败返回 []）。
+"""
+
+from __future__ import annotations
+
+import logging
+import time
+from typing import Any
+
+logger = logging.getLogger(__name__)
+
+DEFAULT_TIMEOUT = 8  # OCR 秒级返回，8s 绰绰有余
+
+
+def call_ocr(img_url: str, *, timeout: int = DEFAULT_TIMEOUT) -> list[str]:
+    """调用 MediaKit 同步 OCR，返回去重后的纯文本列表。
+
+    不做重试（外层降级逻辑负责）。失败/未配置返回空列表，不抛异常。
+    """
+    t0 = time.time()
+    try:
+        import httpx
+
+        from packages.shared.mediakit_client import get_mediakit_client
+
+        client = get_mediakit_client()
+        if not client.is_available:
+            logger.info("[vision.v2] mediakit 未配置，跳过 OCR")
+            return []
+
+        url = f"{client.base_url}/tools-sync/ocr"
+        headers = {
+            "Authorization": f"Bearer {client.api_key}",
+            "Content-Type": "application/json",
+        }
+        payload: dict[str, Any] = {"image_url": img_url}
+        # 部分文档版本用 image_base64，但公网 URL 场景下 image_url 最简
+        resp = httpx.post(url, headers=headers, json=payload, timeout=timeout)
+        elapsed = time.time() - t0
+        if resp.status_code != 200:
+            logger.warning(
+                "[vision.v2] OCR HTTP %d elapsed=%.1fs body=%s",
+                resp.status_code,
+                elapsed,
+                resp.text[:200],
+            )
+            return []
+        data = resp.json()
+        # 兼容几种可能的响应结构
+        code = data.get("code", data.get("status", 0))
+        if code not in (0, "OK", "success", 200):
+            logger.warning("[vision.v2] OCR 业务错误 code=%s elapsed=%.1fs resp=%s", code, elapsed, str(data)[:200])
+            return []
+        texts = _extract_texts(data)
+        # 去重 + 过滤空
+        seen: set[str] = set()
+        out: list[str] = []
+        for t in texts:
+            t = (t or "").strip()
+            if t and t not in seen and len(t) <= 100:  # 过滤过长的误识别
+                seen.add(t)
+                out.append(t)
+        logger.info("[vision.v2] OCR 完成 elapsed=%.1fs n=%d texts=%s", elapsed, len(out), out[:5])
+        return out
+    except Exception as e:
+        elapsed = time.time() - t0
+        logger.warning("[vision.v2] OCR 异常 elapsed=%.1fs err=%s", elapsed, e, exc_info=True)
+        return []
+
+
+def _extract_texts(data: dict) -> list[str]:
+    """从 OCR 响应中抽取文本，兼容多种结构。"""
+    out: list[str] = []
+    # 常见结构1: data.texts = [{"text": "..."}, ...]
+    d = data.get("data") or data
+    if isinstance(d, dict):
+        for key in ("texts", "lines", "words", "items", "result"):
+            items = d.get(key)
+            if isinstance(items, list):
+                for it in items:
+                    if isinstance(it, dict):
+                        txt = it.get("text") or it.get("content") or it.get("word")
+                        if txt:
+                            out.append(str(txt))
+                    elif isinstance(it, str):
+                        out.append(it)
+                break
+        # 结构2: data.text = "..."
+        if not out:
+            t = d.get("text")
+            if isinstance(t, str):
+                out.append(t)
+        # 结构3: data.ocr_text / data.content
+        if not out:
+            for key in ("ocr_text", "content", "raw_text"):
+                v = d.get(key)
+                if isinstance(v, str) and v.strip():
+                    out.append(v)
+                    break
+    return out

apps/worker/worker_app/tasks/vision/vlm_fallback.py

@@ -0,0 +1,226 @@
+# -*- coding: utf-8 -*-
+"""VLM 兜底：专用API路径失败时的最后一道防线，单次调用 doubao-seed-2.1-pro。
+
+设计原则：简单、直接、无竞速、无复杂超时逻辑。只在 fast_json 结果不可用时调用。
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+import re
+import time
+from typing import Any
+
+logger = logging.getLogger(__name__)
+
+DEFAULT_PRO_MODEL = "doubao-seed-2-1-pro-260915"
+DEFAULT_TIMEOUT = 45
+DEFAULT_MAX_TOKENS = 800
+
+
+def _strip_code_fence(s: str) -> str:
+    s = s.strip()
+    if s.startswith("```"):
+        lines = s.split("\n")
+        if lines and lines[0].startswith("```"):
+            lines = lines[1:]
+        if lines and lines[-1].strip().startswith("```"):
+            lines = lines[:-1]
+        s = "\n".join(lines).strip()
+    return s
+
+
+def _xml_text(tag: str, xml: str) -> str:
+    m = re.search(rf"<{tag}[^>]*>(.*?)</{tag}>", xml, re.S)
+    return (m.group(1) if m else "").strip()
+
+
+def _xml_attr(tag: str, attr: str, xml: str) -> str:
+    m = re.search(rf"<{tag}[^>]*\b{attr}\s*=\s*[\"']([^\"']*)[\"']", xml)
+    return (m.group(1) if m else "").strip()
+
+
+def _xml_to_product(raw: str, idx: int) -> dict[str, Any]:
+    """解析 VLM 输出的 XML 格式（简化版）。"""
+    scene = _xml_text("scene", raw) or "通用"
+    mood = _xml_text("mood", raw) or ""
+
+    portrait_prompt = "无人像"
+    p_has = _xml_attr("people", "has_person", raw)
+    if p_has and p_has.lower() != "false":
+        gender = _xml_attr("people", "gender", raw) or ""
+        age = _xml_attr("people", "age_range", raw) or ""
+        outfit = _xml_attr("people", "outfit", raw) or ""
+        hair = _xml_attr("people", "hair", raw) or "自然发型"
+        pose = _xml_attr("people", "pose", raw) or ""
+        expr = _xml_attr("people", "expression", raw) or "自然"
+        parts: list[str] = []
+        if gender:
+            parts.append(gender + ("性" if not gender.endswith("性") else ""))
+        if age:
+            parts.append(age)
+        parts.append("人物")
+        parts.append(hair)
+        if outfit:
+            parts.append(f"身着{outfit}")
+        if pose:
+            parts.append(f"姿态{pose}")
+        parts.append(f"表情{expr}")
+        portrait_prompt = "，".join(parts)
+
+    m = re.search(r"<product[^>]*>(.*?)</product>", raw, re.S)
+    if m:
+        pbody = m.group(1)
+        name = _xml_attr("product", "name", raw) or _xml_text("name", pbody) or "未识别"
+        brand = _xml_attr("product", "brand", raw) or _xml_text("brand", pbody) or "无法判断"
+        category = _xml_attr("product", "category", raw) or _xml_text("category", pbody) or "无法判断"
+        appearance = _xml_attr("product", "appearance", raw) or _xml_text("appearance", pbody) or "无法判断"
+        packaging = _xml_attr("product", "packaging", raw) or _xml_text("packaging", pbody) or "无法判断"
+        feat = _xml_attr("product", "features", raw) or _xml_text("features", pbody) or ""
+        feat_list = [x.strip() for x in re.split(r"[,，;；]", feat) if x.strip()] if feat else ["无法判断"]
+        top_text = _xml_attr("product", "text_on_package", raw) or _xml_text("text_on_package", pbody) or ""
+        text_list = [x.strip() for x in re.split(r"[,，;；]", top_text) if x.strip()] if top_text else []
+        summary = _xml_attr("product", "summary", raw) or _xml_text("summary", pbody) or f"{brand} {name}"
+        pp_attr = _xml_attr("product", "portrait_prompt", raw)
+        if pp_attr and pp_attr != "无人像":
+            portrait_prompt = pp_attr
+        return {
+            "name": name,
+            "brand": brand,
+            "category": category,
+            "appearance": appearance,
+            "packaging": packaging,
+            "text_on_package": text_list,
+            "key_features": feat_list,
+            "scene": scene,
+            "mood": mood,
+            "portrait_prompt": portrait_prompt,
+            "summary": summary,
+            "_source": "vlm_pro_xml",
+        }
+
+    if portrait_prompt != "无人像":
+        return {
+            "name": "未识别",
+            "brand": "无法判断",
+            "category": "无法判断",
+            "appearance": "无法判断",
+            "packaging": "无法判断",
+            "text_on_package": [],
+            "key_features": ["无法判断"],
+            "scene": scene,
+            "mood": mood,
+            "portrait_prompt": portrait_prompt,
+            "summary": "未识别",
+            "_source": "vlm_pro_no_product",
+        }
+    return {
+        "name": "未识别",
+        "brand": "无法判断",
+        "category": "无法判断",
+        "appearance": "无法判断",
+        "packaging": "无法判断",
+        "text_on_package": [],
+        "key_features": ["无法判断"],
+        "scene": scene,
+        "mood": mood,
+        "portrait_prompt": "无人像",
+        "summary": "未识别",
+        "_source": "vlm_pro_no_tag",
+    }
+
+
+def call_pro_vlm(
+    img_url: str,
+    idx: int,
+    *,
+    model: str | None = None,
+    timeout: int = DEFAULT_TIMEOUT,
+) -> dict[str, Any] | None:
+    """单次调用 pro VLM，解析后返回 product dict；失败返回 None。"""
+    t0 = time.time()
+    try:
+        from packages.application.viral_video.prompt_loader import (
+            get_template,
+            render_system_prompt,
+            render_user_prompt,
+        )
+        from packages.shared.ai_client import get_doubao_client
+    except ImportError as e:
+        logger.warning("[vision.vlm] 导入失败: %s", e)
+        return None
+
+    try:
+        template = get_template("image_analysis")
+        system = render_system_prompt(template)
+        user = render_user_prompt(template, image_count=1, industry="通用", image_urls=f"第1张：{img_url}")
+    except Exception as e:
+        logger.warning("[vision.vlm] 模板加载失败: %s", e)
+        return None
+
+    client = get_doubao_client()
+    if not client.is_available:
+        return None
+
+    use_model = model or DEFAULT_PRO_MODEL
+    _orig_retries = client.max_retries
+    client.max_retries = 0
+    try:
+        raw = client.vision_completion(
+            messages=[{"role": "system", "content": system}, {"role": "user", "content": user}],
+            images=[img_url],
+            temperature=0.3,
+            max_tokens=DEFAULT_MAX_TOKENS,
+            timeout=timeout,
+            model=use_model,
+        )
+    except Exception as e:
+        logger.warning("[vision.vlm] 图片 #%d pro VLM 调用失败 elapsed=%.1fs err=%s", idx, time.time() - t0, e)
+        client.max_retries = _orig_retries
+        return None
+    client.max_retries = _orig_retries
+
+    elapsed = time.time() - t0
+    if not raw:
+        logger.warning("[vision.vlm] 图片 #%d pro VLM 返回空 elapsed=%.1fs", idx, elapsed)
+        return None
+
+    text = _strip_code_fence(raw)
+    l, r = text.find("{"), text.rfind("}")
+    if l >= 0 and r > l:
+        try:
+            obj = json.loads(text[l : r + 1])
+            if isinstance(obj, dict):
+                logger.info("[vision.vlm] 图片 #%d pro VLM JSON 完成 elapsed=%.1fs", idx, elapsed)
+                return {
+                    "name": obj.get("name") or "未识别",
+                    "brand": obj.get("brand") or "无法判断",
+                    "category": obj.get("category") or "无法判断",
+                    "appearance": obj.get("appearance") or "无法判断",
+                    "packaging": obj.get("packaging") or "无法判断",
+                    "text_on_package": obj.get("text_on_package") or [],
+                    "key_features": obj.get("key_features") or obj.get("features") or ["无法判断"],
+                    "scene": obj.get("scene") or "通用",
+                    "mood": obj.get("mood") or "",
+                    "portrait_prompt": obj.get("portrait_prompt") or "无人像",
+                    "summary": obj.get("summary") or f"{obj.get('brand','')} {obj.get('name','')}",
+                    "_source": "vlm_pro_json",
+                }
+        except json.JSONDecodeError:
+            pass
+
+    try:
+        result = _xml_to_product(text, idx)
+        result["_fallback_used"] = True
+        result["_pro_elapsed"] = round(elapsed, 2)
+        logger.info(
+            "[vision.vlm] 图片 #%d pro VLM XML 完成 elapsed=%.2fs pp=%s",
+            idx,
+            elapsed,
+            (result.get("portrait_prompt") or "")[:40],
+        )
+        return result
+    except Exception as e:
+        logger.warning("[vision.vlm] 图片 #%d 解析失败 elapsed=%.1fs err=%s head=%s", idx, elapsed, e, raw[:200])
+        return None

apps/worker/worker_app/tasks/vision/vlm_fast_json.py

@@ -0,0 +1,148 @@
+# -*- coding: utf-8 -*-
+"""doubao-seed-2.1-lite 强约束 JSON-only 调用。
+
+目标：替代"人体属性/商品检测/图像标签"三个火山不存在的专用云端 API。
+设计要点：
+- system prompt 极致精简，只给字段 schema 和强约束（禁止自然语言、禁止 markdown）
+- max_tokens=350（比旧 VLM 的 1200 小很多，降低延迟）
+- temperature=0.1（极低，稳定输出 JSON）
+- timeout=8s（够快，失败则由外层走 pro VLM 兜底）
+- 期望返回纯 JSON object（无 ```json 包裹、无解释文字）
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+import time
+from typing import Any
+
+logger = logging.getLogger(__name__)
+
+# 极简 system prompt：只给字段定义 + 硬性输出要求
+_FAST_SYSTEM = (
+    "你是图片结构化识别器。严格按下方 JSON schema 返回一个对象，不要任何解释、"
+    "不要markdown、不要代码块、不要前后缀文字。字段值不确定时填 null 或空数组。\n"
+    "{\n"
+    '  "has_person": true/false,  // 图中是否有人\n'
+    '  "gender": "男"/"女"/null,\n'
+    '  "age_range": "儿童"/"青少年"/"青年"/"中年"/"老年"/null,\n'
+    '  "upper_wear": "上装款式，如T恤/衬衫/卫衣/毛衣/西装/夹克/连衣裙/吊带/背心/外套等",\n'
+    '  "upper_color": "上装主色",\n'
+    '  "lower_wear": "下装款式，如牛仔裤/休闲裤/短裙/长裙/短裤/西裤/运动裤等；穿连衣裙时填null",\n'
+    '  "lower_color": "下装主色",\n'
+    '  "dress_color": "连衣裙主色（穿连衣裙时填）",\n'
+    '  "accessories": ["眼镜"/"帽子"/"项链"/"耳环"/"背包"/"手表"等数组],\n'
+    '  "hairstyle": "发型，如短发/长发/马尾/卷发/丸子头/光头等",\n'
+    '  "expression": "表情，如微笑/严肃/酷/开心等",\n'
+    '  "pose": "姿势，如站立/坐姿/侧身/行走等",\n'
+    '  "scene": "场景，如室内/街拍/户外/办公室/家居/海边/雪景/森林等",\n'
+    '  "style": "风格，如休闲/商务/运动/复古/潮流/甜美/酷飒/优雅/街头/法式等",\n'
+    '  "has_product": true/false,  // 是否有明确商品展示\n'
+    '  "category": "产品类目：服饰/鞋包/美妆/数码/食品/家居/配饰/母婴/非产品图",\n'
+    '  "product_name": "产品名称，非产品图填null",\n'
+    '  "brand": "品牌或文字标识，无则null",\n'
+    '  "material": "材质，如棉质/牛仔/皮革/真丝/针织/涤纶等",\n'
+    '  "pattern": "图案，如纯色/条纹/波点/格子/印花/碎花/Logo等",\n'
+    '  "colors": ["主色数组"],\n'
+    '  "mood": "整体氛围/情绪，如清新/活力/高级/温暖/冷峻/甜美/复古等"\n'
+    "}"
+)
+
+_FAST_USER = "识别这张图片的人物穿搭与主体信息，只返回JSON对象。"
+
+# 默认模型
+DEFAULT_LITE_MODEL = "doubao-seed-2-1-lite-260915"
+DEFAULT_TIMEOUT = 8
+DEFAULT_MAX_TOKENS = 350
+
+
+def _strip_code_fence(s: str) -> str:
+    """剥离 ```json ... ``` 包裹（即使要求纯 JSON，模型偶尔仍会包代码块）。"""
+    s = s.strip()
+    if s.startswith("```"):
+        lines = s.split("\n")
+        # 去掉首行 ```json
+        if lines and lines[0].startswith("```"):
+            lines = lines[1:]
+        # 去掉尾行 ```
+        if lines and lines[-1].strip().startswith("```"):
+            lines = lines[:-1]
+        s = "\n".join(lines).strip()
+    return s
+
+
+def call_fast_json(
+    img_url: str,
+    *,
+    model: str | None = None,
+    timeout: int = DEFAULT_TIMEOUT,
+    max_tokens: int = DEFAULT_MAX_TOKENS,
+) -> dict[str, Any] | None:
+    """调用 lite VLM 返回结构化 dict；失败/非 JSON 返回 None。
+
+    注意：不做重试（外层竞速/降级逻辑负责），max_retries=0 由外层统一设置。
+    """
+    t0 = time.time()
+    try:
+        from packages.shared.ai_client import get_doubao_client
+
+        client = get_doubao_client()
+        if not client.is_available:
+            logger.warning("[vision.v2] doubao client 不可用，跳过 fast_json")
+            return None
+
+        use_model = model or DEFAULT_LITE_MODEL
+        # 强制不重试：lite 是快速路径，失败直接走外层 pro 兜底
+        _orig_retries = client.max_retries
+        client.max_retries = 0
+        try:
+            raw = client.vision_completion(
+                messages=[
+                    {"role": "system", "content": _FAST_SYSTEM},
+                    {"role": "user", "content": _FAST_USER},
+                ],
+                images=[img_url],
+                temperature=0.1,
+                max_tokens=max_tokens,
+                timeout=timeout,
+                model=use_model,
+            )
+        finally:
+            client.max_retries = _orig_retries
+        elapsed = time.time() - t0
+        if raw is None:
+            logger.warning("[vision.v2] fast_json 返回 None elapsed=%.1fs model=%s", elapsed, use_model)
+            return None
+
+        text = _strip_code_fence(raw)
+        # 截到第一个 { 和最后一个 } 之间，容忍前后偶发文字
+        l = text.find("{")
+        r = text.rfind("}")
+        if l >= 0 and r > l:
+            text = text[l : r + 1]
+        try:
+            obj = json.loads(text)
+        except json.JSONDecodeError:
+            logger.warning(
+                "[vision.v2] fast_json JSON 解析失败 elapsed=%.1fs head=%s",
+                elapsed,
+                raw[:200],
+            )
+            return None
+        if not isinstance(obj, dict):
+            logger.warning("[vision.v2] fast_json 非 dict: %s", type(obj))
+            return None
+        logger.info(
+            "[vision.v2] fast_json 完成 model=%s elapsed=%.1fs has_person=%s has_product=%s category=%s",
+            use_model,
+            elapsed,
+            obj.get("has_person"),
+            obj.get("has_product"),
+            obj.get("category"),
+        )
+        return obj
+    except Exception as e:
+        elapsed = time.time() - t0
+        logger.warning("[vision.v2] fast_json 异常 elapsed=%.1fs err=%s", elapsed, e, exc_info=True)
+        return None

packages/application/viral_video/prompts.py

@@ -50,14 +50,25 @@ _IMAGE_ANALYSIS_SYSTEM = f"""你是电商商品视觉分析师，负责从商品
 请严格按下面的标签格式输出，标签名一个都不能改，不要输出任何解释，不要用代码块：
 <products> 下面每个产品用一个 <product> 标签，属性 name 是产品名、features 是外观特征、position 是 main 或 secondary、image_index 是第几张图（从0开始）。
 <colors> 下面每个主要颜色用一个 <color> 标签，属性 hex 是色值、name 是颜色名、coverage 是占比小数。
-<people> 用一个标签，属性 has_person、count、gender、age_range、pose、expression 分别描述人物情况。
+<people> 用一个标签，属性 has_person、count、gender、age_range、hair（发型发色）、skin_tone（肤色）、face_shape（脸型）、outfit（穿着）、pose（姿态）、expression（表情）分别描述人物外貌。有人物时属性尽量具体（如hair="黑色长直发"、outfit="白色衬衫"），无人像时除has_person=false外其他填"无法判断"。
 <mood> 标签写画面整体情绪氛围。
 <visible_text> 下面每处可见文字用一个 <text_item> 标签，属性 text 是文字内容、position 是位置。
 <scene> 标签写场景描述。
 <quality> 用一个标签，属性 resolution、lighting、composition、blur 描述画质。
 <key_selling_points> 下面每个卖点用一个 <point> 标签。
 
-看不到或无法判断的内容，属性值填“无法判断”，布尔值填 false，不要留空标签。"""
+【人物属性硬性要求（has_person=true时必须遵守）】
+hair/skin_tone/face_shape/outfit四项绝对禁止填“无法判断”，必须基于图片可见特征给出具体中文描述：
+- hair：必须描述发型+发色，如“黑色齐肩直发”“棕色微卷中长发”“深棕色短发”
+- skin_tone：必须描述肤色，如“暖调自然肤色”“白皙肤色”“小麦色”
+- face_shape：必须描述脸型，如“鹅蛋脸”“圆脸”“瓜子脸”“方脸”
+- outfit：必须描述可见穿着，如“米色翻领衬衫”“白色T恤”“黑色连衣裙”
+即使局部被遮挡也要根据可见部分合理推断；确实看不清时按最接近的直观印象描述。
+
+其他非人物属性看不到或无法判断时填“无法判断”，布尔值填false，不要留空标签。
+
+【有人物场景输出参考（女性手持商品示例，必须写全10个属性，禁止省略）】
+<people has_person="true" count="1" gender="女" age_range="青年" hair="黑色齐肩直发" skin_tone="暖调自然肤色" face_shape="鹅蛋脸" outfit="米色翻领衬衫" pose="正面半身，手持商品" expression="面带微笑"/>"""
 
 _IMAGE_ANALYSIS_USER = """请分析以下商品图片，共 {image_count} 张。
 所属行业：{industry}
@@ -73,7 +84,7 @@ _IMAGE_ANALYSIS_EXAMPLE = """<products>
 <color hex="#D32F2F" name="红色" coverage="0.4"/>
 <color hex="#FFFFFF" name="白色" coverage="0.5"/>
 </colors>
-<people has_person="false" count="0" gender="无法判断" age_range="无法判断" pose="无法判断" expression="无法判断"/>
+<people has_person="false" count="0" gender="无法判断" age_range="无法判断" hair="无法判断" skin_tone="无法判断" face_shape="无法判断" outfit="无法判断" pose="无法判断" expression="无法判断"/>
 <mood>干净、实用</mood>
 <visible_text>
 <text_item text="多功能油污净" position="瓶身正面"/>
@@ -132,7 +143,10 @@ _FUSION_SYSTEM = """你负责为短视频生成营销文案。请按思维链分
 <hook> 开头3秒钩子，5到15字。
 <body_points> 每个要点用一个 <point> 标签，属性 elaboration 是展开说明、image_index 是对应第几张图（从0开始），标签内容写要点。
 <cta> 口语化的行动号召。
-<script_segments> 每段配音用一个 <segment> 标签，属性 duration_sec 是秒数、image_index 是对应图片，标签内容写配音文案。
+<script_segments> 每段配音用一个 <segment> 标签，属性 duration_sec 是秒数、image_index 是对应图片，标签内容写配音文案（纯口播文本，不加旁白标注、不加镜头标注、不加"主播："之类前缀）。
+<voiceover_script> 把所有 segment 的配音文案按顺序自然拼接成一段完整的纯口播文本（无标记、无括号、无前缀），长度要适配 {duration} 秒，约 {approx_chars} 字。
+<overview_theme> 视频主题（一句话概括）。
+<scene_and_lighting> 整体场景描述+光线设定（100-200字，要具体：在哪拍、什么光线、什么色调、什么氛围）。
 <word_count> 配音总字数，只写数字。
 <estimated_duration> 预计时长秒数，只写数字。
 
@@ -161,25 +175,38 @@ _FUSION_EXAMPLE = """<title>厨房重油污，别再用洗洁精硬擦了</title
 <segment duration_sec="6" image_index="0">后来换了这个大公鸡头油污净，喷上等几分钟，一擦就干净</segment>
 <segment duration_sec="4" image_index="0">39块钱625ml，厨房重油污的可以试一瓶</segment>
 </script_segments>
+<voiceover_script>这油污我真的忍很久了，用洗洁精擦半天都没用。后来换了这个大公鸡头油污净，喷上等几分钟，一擦就干净。39块钱625ml，厨房重油污的可以试一瓶。</voiceover_script>
+<overview_theme>厨房油污清洁好物分享</overview_theme>
+<scene_and_lighting>简洁明亮的厨房台面场景，自然光从窗户洒入，色调温暖柔和，突出产品白色瓶身与去油污对比效果。</scene_and_lighting>
 <word_count>58</word_count>
 <estimated_duration>13</estimated_duration>"""
 
 # ── 模板4：编导级分镜（LLM）────────────────────────────────────────────
-_STORYBOARD_SYSTEM = f"""你是短视频编导，负责把文案拆成可拍摄的分镜。
+_STORYBOARD_SYSTEM = """你是短视频编导，负责把文案拆成可拍摄的分镜，为 Seedance 2.5 视频模型写编导分镜脚本。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。
 
 工作方式：
 1. 按文案的 script_segments 顺序分配镜头。
-2. 每个镜头确定画面、运镜、时长、配音和字幕。
+2. 每个镜头确定景别/角度/运镜、画面场景与对白、人物动作细节、音效/BGM、转场。
 3. 检查所有镜头时长加起来接近目标时长，误差不超过2秒。
 4. image_index 必须在已上传图片范围内，第一张主图必须用在第一个镜头。
 
-{GLOBAL_CONSTRAINTS}
+{fusion_instruction}
+
+{global_constraints}
+
+{negative_rules}
 
 请严格按下面的标签格式输出，不要解释，不要用代码块：
-<clips> 下面每个镜头用一个 <clip> 标签，属性 image_index 是图片序号（从0开始）、transition 取 fade、cut、zoom_in、slide_left、dissolve、wipe 之一、zoom 取 in、out 或 null、duration_sec 是该镜头秒数、bgm_note 是该段BGM情绪。每个 <clip> 里面包含：
-<voice_text> 该镜头配音文本；
-<subtitle_text> 字幕文本，可与配音一致或更精简；
-<ken_burns> 用一个空标签，属性 start、end 写“x,y”坐标、ease 写缓动方式；不需要运镜时坐标相同。"""
+<clips> 下面每个镜头用一个 <clip> 标签，属性 image_index 是图片序号（从0开始）、transition 取 fade/cut/zoom_in/slide_left/dissolve/wipe 之一、zoom 取 in/out/null、duration_sec 是该镜头秒数、bgm_note 是该段BGM情绪。每个 <clip> 里面包含：
+  <voice_text> 该镜头配音文本（纯口播文本，不加旁白标注）；
+  <subtitle_text> 字幕文本，可与配音一致或更精简；
+  <shot_type_angle_movement> 景别+角度+运镜（例：近景俯拍45度，缓慢推镜；中景平视，固定镜头；特写平视，快速拉镜）；
+  <scene_and_dialogue> 画面场景描述 + 人物口播台词（对白要自然口语化，像朋友聊天，不要硬广推销腔）；
+  <action_details> 人物动作、表情、物品操作细节（手怎么动、表情变化、产品怎么展示）；
+  <audio_bgm> 环境音+BGM提示（例：轻快流行BGM，环境嘈杂咖啡店背景音）；
+  <transition> 硬切/淡入淡出/叠化（最后一镜写『结束』即可）；
+  <reference_image_index> 参考图片索引（0-based，对应第几张产品图，无则空）；
+  <ken_burns> 用一个空标签，属性 start、end 写"x,y"坐标、ease 写缓动方式；不需要运镜时坐标相同。"""
 
 _STORYBOARD_USER = """目标时长：{duration}秒
 上传图片数量：{image_count}张（第1张是主图/封面）
@@ -194,17 +221,35 @@ _STORYBOARD_EXAMPLE = """<clips>
 <clip image_index="0" transition="cut" zoom="null" duration_sec="3" bgm_note="日常、轻微烦躁">
 <voice_text>这油污我真的忍很久了</voice_text>
 <subtitle_text>这油污忍很久了</subtitle_text>
+<shot_type_angle_movement>近景俯拍45度，缓慢推镜</shot_type_angle_movement>
+<scene_and_dialogue>厨房台面，主妇皱眉看着灶台油污。对白：这油污我真的忍很久了</scene_and_dialogue>
+<action_details>右手拿着脏抹布，无奈摇头</action_details>
+<audio_bgm>轻快日常BGM，带一点烦躁感</audio_bgm>
+<transition>硬切</transition>
+<reference_image_index>0</reference_image_index>
 <ken_burns start="0,0" end="0,0" ease="linear"/>
 </clip>
 <clip image_index="0" transition="zoom_in" zoom="in" duration_sec="6" bgm_note="轻快、出现转机">
 <voice_text>后来换了大公鸡头油污净，喷上等几分钟，一擦就干净</voice_text>
 <subtitle_text>喷上等几分钟，一擦就干净</subtitle_text>
+<shot_type_angle_movement>特写平视，固定镜头</shot_type_angle_movement>
+<scene_and_dialogue>手部特写，喷油污净在油污处。对白：后来换了这个大公鸡头油污净，喷上等几分钟，一擦就干净</scene_and_dialogue>
+<action_details>左手拿产品瓶身，右手按压喷头，等待片刻后用抹布轻擦</action_details>
+<audio_bgm>轻快转折BGM，带清爽感</audio_bgm>
+<transition>淡入淡出</transition>
+<reference_image_index>0</reference_image_index>
 <ken_burns start="20,20" end="80,80" ease="ease-in-out"/>
 </clip>
 <clip image_index="0" transition="fade" zoom="null" duration_sec="4" bgm_note="温暖、推荐">
 <voice_text>39块钱625ml，厨房重油污的可以试一瓶</voice_text>
 <subtitle_text>39元625ml，可以试一瓶</subtitle_text>
-<ken_burns start="0,0" end="0,0" ease="linear"/>
+<shot_type_angle_movement>中景平视，缓慢拉镜</shot_type_angle_movement>
+<scene_and_dialogue>产品正面展示，明亮背景。对白：39块钱625ml，厨房重油污的可以试一瓶</scene_and_dialogue>
+<action_details>产品置于画面中央，轻微转动展示瓶身</action_details>
+<audio_bgm>温暖收尾BGM</audio_bgm>
+<transition>结束</transition>
+<reference_image_index>0</reference_image_index>
+<ken_burns start="50,50" end="20,20" ease="ease-in-out"/>
 </clip>
 </clips>"""
 

packages/config/base.py

@@ -103,7 +103,7 @@ class SharedSettings(BaseSettings):
     doubao_vision_lite_model: str = (
         "doubao-seed-2-1-lite-260915"  # 快速视觉（Seed 2.1 Lite 原生多模态；原 vision-lite-250315 不可用）
     )
-    doubao_vision_use_lite: bool = False  # #2181: lite视觉模型100%超时，默认关闭走pro（25-38s稳定返回）
+    doubao_vision_use_lite: bool = True  # #2188: lite恢复稳定，爆款视频默认lite-first提速（20-30s）
     doubao_embedding_model: str = "doubao-embedding-vision-251215"  # 多模态向量化（原 large-text-240915 已 Retiring）
     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/shared/ai_client.py

@@ -585,7 +585,7 @@ class DoubaoClient:
         # #2172/#2174: 信任链——使用预热好的 Seedream t2i 文生图（纯模型生成人像，是方舟信任产物，
         # 不会触发肖像审核）。预热在 VLM 分析后由 daemon 线程后台完成，结果通过 pre_trusted_images 传入。
         # - 预热结果有效 → 替换原参考图，走 omni_ref 模式
-        # - 预热结果不可用 → 直接用原图（若被400肖像拦截，#2166自动降级纯t2v），避免现场跑t2i阻塞渲染
+        # - 预热结果不可用 → 直接用原图（上层 _step_render 已现场同步跑 Seedream t2i 兜底；若再被400拦截，下方自动降级纯t2v）
         # 信任链只作用于 doubao provider；DashScope(Wan) 保持原行为。
         trust_chain_applied = False
         if provider == "doubao" and getattr(self, "trust_chain_enabled", True) and pre_trusted_images:
@@ -763,6 +763,26 @@ class DoubaoClient:
                 # 保留第二次的错误信息
                 sc, body = sc2, body2
 
+        # #2183: portrait_intercept / 真人肖像审核拦截 → 去掉所有参考图（含image_url首帧），纯 t2v 重试一次
+        # 信任链预热或现场 t2i 都失败时的最后兜底，保证能出片
+        if not task_id and sc == 400:
+            _err_code_for_400, _ = _classify_video_error(sc, body, last_err)
+            if _err_code_for_400 == "portrait_intercept" and (image_url or ref_imgs):
+                logger.warning(
+                    "Seedance 创建因 portrait_intercept 失败，降级纯 t2v（移除所有参考图）重试: img=%d ref=%d",
+                    1 if image_url else 0,
+                    len(ref_imgs),
+                )
+                _t2v_payload = dict(create_payload)
+                _t2v_payload["content"] = [{"type": "text", "text": prompt.strip()}]
+                _t2v_payload["ratio"] = ratio or "9:16"
+                task_id, last_err, sc3, body3 = _do_create(_t2v_payload)
+                if task_id:
+                    sc, body = sc3, body3
+                    logger.info("[trust-chain] portrait_intercept 降级纯 t2v 成功 task_id=%s", task_id)
+                else:
+                    sc, body = sc3, body3
+
         if not task_id:
             err_code, user_msg = _classify_video_error(sc, body, last_err)
             self.last_video_error = {

tests/unit/test_viral_video.py

@@ -403,33 +403,30 @@ class TestViralVideoPipeline:
         """v1.6: _step_script_generation 返回 dict 形式的 CopyResult，含 voiceover_script + shots。"""
         from apps.worker.worker_app.tasks.viral_video import _step_script_generation
 
-        mock_llm.return_value = {
-            "overview": {"theme": "口红推荐", "total_duration": 15, "aspect_ratio": "9:16"},
-            "scene_and_lighting": "明亮化妆台，柔和自然光",
-            "shots": [
-                {
-                    "time_range": "0-5秒",
-                    "shot_type_angle_movement": "近景平视，缓慢推镜",
-                    "scene_and_dialogue": "女主微笑展示口红：大家好，今天分享一款口红",
-                    "action_details": "手持口红特写",
-                    "audio_bgm": "轻快流行BGM",
-                    "transition": "硬切",
-                    "reference_image_index": 0,
-                },
-                {
-                    "time_range": "5-15秒",
-                    "shot_type_angle_movement": "特写，固定镜头",
-                    "scene_and_dialogue": "涂抹口红：颜色特别好看很显白",
-                    "action_details": "嘴唇涂抹特写",
-                    "audio_bgm": "轻快BGM继续",
-                    "transition": "结束",
-                    "reference_image_index": 1,
-                },
-            ],
-            "hard_constraints": ["无字幕无水印"],
-            "negative_prompts": ["字幕", "水印"],
-            "voiceover_script": "大家好，今天分享一款口红，颜色特别好看很显白。",
-        }
+        mock_llm.return_value = """<clips>
+<clip image_index="0" transition="cut" zoom="null" duration_sec="5" bgm_note="轻快流行BGM">
+<voice_text>大家好，今天分享一款口红</voice_text>
+<subtitle_text>大家好，今天分享一款口红</subtitle_text>
+<shot_type_angle_movement>近景平视，缓慢推镜</shot_type_angle_movement>
+<scene_and_dialogue>女主微笑展示口红：大家好，今天分享一款口红</scene_and_dialogue>
+<action_details>手持口红特写</action_details>
+<audio_bgm>轻快流行BGM</audio_bgm>
+<transition>硬切</transition>
+<reference_image_index>0</reference_image_index>
+<ken_burns start="0,0" end="0,0" ease="linear"/>
+</clip>
+<clip image_index="0" transition="fade" zoom="null" duration_sec="10" bgm_note="轻快BGM">
+<voice_text>颜色特别好看很显白</voice_text>
+<subtitle_text>颜色特别好看很显白</subtitle_text>
+<shot_type_angle_movement>特写，固定镜头</shot_type_angle_movement>
+<scene_and_dialogue>涂抹口红：颜色特别好看很显白</scene_and_dialogue>
+<action_details>嘴唇涂抹特写</action_details>
+<audio_bgm>轻快BGM继续</audio_bgm>
+<transition>结束</transition>
+<reference_image_index>1</reference_image_index>
+<ken_burns start="0,0" end="0,0" ease="linear"/>
+</clip>
+</clips>"""
         result = _step_script_generation(
             mock_job, {"intent": "推广口红", "key_messages": [], "tone": "亲切"}, {"products": []}
         )
@@ -452,12 +449,13 @@ class TestViralVideoPipeline:
         assert result["voiceover_script"]
         assert len(result["shots"]) >= 1
 
-    @patch("packages.shared.ai_service.call_llm")
-    def test_review_pass_v16(self, mock_llm, mock_job):
+    @patch("packages.application.viral_video.reviewer.Reviewer.review")
+    def test_review_pass_v16(self, mock_review, mock_job):
         """v1.6 _step_review 接收 copy_result dict。"""
         from apps.worker.worker_app.tasks.viral_video import _step_review
+        from packages.application.viral_video.reviewer import ReviewIssue, ReviewResult
 
-        mock_llm.return_value = {"passed": True, "score": 90, "details": {}}
+        mock_review.return_value = ReviewResult(passed=True, score=90, issues=[], rewrite_suggestions=[])
         cr = {"voiceover_script": "大家好", "shots": []}
         result = _step_review(mock_job, cr)
         assert result["passed"] is True

tests/unit/test_viral_video_wiring.py

@@ -0,0 +1,302 @@
+"""#2040 接线集成测试：验证运行中的 viral_video 任务使用 prompt_loader 从 DB 读取模板。
+
+mock LLM/Vision 调用，验证：
+1. image_analysis 走 loader 模板 + XML 解析
+2. intent_parsing 走 loader 模板 + XML 解析
+3. script_generation 走 storyboard 模板 + XML 解析，输出兼容 Seedance 的 copy_result
+4. review 走 Reviewer（review 模板）带自动重写
+5. 三档融合（ai_full / ai_polish / user_primary）注入不同 FUSION_INSTRUCTIONS
+"""
+
+from __future__ import annotations
+
+import sys
+from pathlib import Path as _Path
+
+_WORKER_ROOT = _Path(__file__).resolve().parents[2] / "apps" / "worker"
+if str(_WORKER_ROOT) not in sys.path:
+    sys.path.insert(0, str(_WORKER_ROOT))
+
+from unittest.mock import MagicMock, patch
+
+import pytest
+
+from packages.domain.viral_video import ViralVideoJob
+
+
+@pytest.fixture
+def job():
+    j = ViralVideoJob(
+        user_id="u1",
+        images=["https://img/1.jpg", "https://img/2.jpg"],
+        industry="美妆",
+        duration=15,
+        user_copy_text="这款口红真的太绝了，显白又持久，姐妹们冲！",
+        fusion_level="ai_polish",
+    )
+    return j
+
+
+# ── Mock LLM/Vision 返回的 XML 文本 ─────────────────────────────────
+
+IMAGE_XML = """
+<analysis>
+<scene>室内桌面拍摄，柔和自然光</scene>
+<mood>清新温暖</mood>
+<product name="lipstick" brand="品牌X" category="唇部彩妆"
+         appearance="管状红色膏体" packaging="黑色金属管"
+         features="显白,持久,滋润" portrait_prompt="无人像"
+         summary="品牌X红色口红">
+  <text_on_package>品牌X,211</text_on_package>
+</product>
+</analysis>
+""".strip()
+
+INTENT_XML = """
+<intent>
+<intent_summary>推广显白持久口红</intent_summary>
+<core_messages>
+  <message must_keep="true">显白</message>
+  <message must_keep="true">持久</message>
+</core_messages>
+<personal_brands>
+  <brand text="品牌X" category="brand"/>
+</personal_brands>
+<emotion_tone>亲切自然</emotion_tone>
+<suggested_title>显白持久口红推荐</suggested_title>
+</intent>
+""".strip()
+
+STORYBOARD_XML = """
+<clips>
+<clip image_index="0" transition="cut" zoom="null" duration_sec="5" bgm_note="轻快BGM">
+  <voice_text>这款口红真的太绝了</voice_text>
+  <subtitle_text>显白又持久</subtitle_text>
+  <shot_type_angle_movement>近景俯拍45度，缓慢推镜</shot_type_angle_movement>
+  <scene_and_dialogue>厨房台面，主妇展示口红。对白：这款口红真的太绝了</scene_and_dialogue>
+  <action_details>右手持口红展示膏体</action_details>
+  <audio_bgm>轻快BGM</audio_bgm>
+  <transition>硬切</transition>
+  <reference_image_index>0</reference_image_index>
+  <ken_burns start="0,0" end="0,0" ease="linear"/>
+</clip>
+</clips>
+""".strip()
+
+
+@pytest.fixture(autouse=True)
+def invalidate_loader_cache():
+    from packages.application.viral_video import prompt_loader as pl
+
+    pl.invalidate()
+    yield
+    pl.invalidate()
+
+
+# ── 1) 图片分析走模板 ───────────────────────────────────────────────
+
+
+class TestImageAnalysisWiring:
+    def test_uses_loader_template_and_xml_parse(self, job):
+        from apps.worker.worker_app.tasks import viral_video as vv
+
+        with patch("packages.shared.ai_service.call_vision", return_value=IMAGE_XML) as mock_v:
+            result = vv._analyze_single_image(0, "https://img/1.jpg", "vlm-lite", 15)
+
+        mock_v.assert_called_once()
+        # 验证调用时传入了 system_prompt（说明走了 loader 渲染的模板）
+        call_kwargs = mock_v.call_args.kwargs
+        assert "system_prompt" in call_kwargs and call_kwargs["system_prompt"]
+        # 结果包含从 XML 解析出的产品信息
+        assert result["name"] == "lipstick"
+        assert result["brand"] == "品牌X"
+        assert "显白" in result["key_features"]
+        assert result["text_on_package"] == ["品牌X", "211"]
+
+
+# ── 2) 意图解析走模板 ───────────────────────────────────────────────
+
+
+class TestIntentParsingWiring:
+    def test_uses_loader_and_parses_xml(self, job):
+        from apps.worker.worker_app.tasks import viral_video as vv
+
+        img_result = {"products": [{"name": "lipstick", "brand": "品牌X", "key_features": ["显白", "持久"]}]}
+        with patch("packages.shared.ai_service.call_llm", return_value=INTENT_XML) as mock_llm:
+            result = vv._step_intent_parsing(job, img_result)
+
+        mock_llm.assert_called_once()
+        assert result["intent"] == "推广显白持久口红"
+        assert "显白" in result["key_messages"]
+        assert result["suggested_title"] == "显白持久口红推荐"
+
+
+# ── 3) 脚本生成：storyboard 模板 + XML 解析 + fusion_level 注入 ────
+
+
+class TestScriptGenerationWiring:
+    @pytest.mark.parametrize("level", ["ai_full", "ai_polish", "user_primary"])
+    def test_fusion_level_injected(self, job, level):
+        """三档融合水平被注入到 storyboard 模板的 system_prompt"""
+        from apps.worker.worker_app.tasks import viral_video as vv
+        from packages.application.viral_video.prompts import FUSION_INSTRUCTIONS
+
+        job.fusion_level = level
+        intent = {"intent": "推广", "key_messages": ["显白"], "tone": "亲切"}
+
+        captured_system = {}
+
+        def fake_call_llm(messages, **kw):
+            captured_system["final"] = messages[0]["content"]
+            return STORYBOARD_XML
+
+        with patch("packages.shared.ai_service.call_llm", side_effect=fake_call_llm):
+            result = vv._step_script_generation(job, intent, {})
+
+        # fusion_level 对应的指令文本被注入到 system prompt 中
+        assert FUSION_INSTRUCTIONS[level] in captured_system["final"], f"fusion_level {level} 指令未注入 system_prompt"
+        # 输出保持 Seedance 兼容结构
+        assert "overview" in result
+        assert "shots" in result
+        assert len(result["shots"]) >= 1
+        assert result["shots"][0]["shot_type_angle_movement"]
+        assert result["voiceover_script"]
+
+    def test_fallback_when_xml_and_json_unparseable(self, job):
+        """XML 解析失败且无法解析为 JSON 时，回退到兜底脚本"""
+        from apps.worker.worker_app.tasks import viral_video as vv
+
+        job.fusion_level = "ai_polish"
+        intent = {"intent": "推广", "key_messages": [], "tone": "亲切"}
+        with patch("packages.shared.ai_service.call_llm", return_value="not xml not json"):
+            result = vv._step_script_generation(job, intent, {})
+        assert isinstance(result, dict)
+        assert "voiceover_script" in result
+        assert "shots" in result
+
+
+# ── 4) Review 使用 Reviewer + 自动重写 ─────────────────────────────
+
+
+class TestReviewWiring:
+    def test_pass_path(self, job):
+        from apps.worker.worker_app.tasks import viral_video as vv
+        from packages.application.viral_video.reviewer import Reviewer, ReviewResult
+
+        copy_result = {
+            "title": "口红推荐",
+            "overview": {"theme": "口红推荐"},
+            "voiceover_script": "这款口红显白又持久",
+            "shots": [{"scene_and_dialogue": "展示口红"}],
+        }
+        job.intent_result = {"key_messages": ["显白", "持久"], "intent": "推广"}
+
+        pass_result = ReviewResult(passed=True, score=90, issues=[], rewrite_suggestions=[])
+        with patch.object(Reviewer, "review", return_value=pass_result):
+            out = vv._step_review(job, copy_result)
+        assert out["passed"] is True
+
+    def test_rewrite_path(self, job):
+        """审核不通过时触发自动重写，并更新 job.copy_result"""
+        from apps.worker.worker_app.tasks import viral_video as vv
+        from packages.application.viral_video.reviewer import Reviewer, ReviewResult
+        from packages.application.viral_video.schemas import FusionResult, ReviewIssue, ScriptSegment
+
+        copy_result = {
+            "title": "原标题",
+            "overview": {"theme": "原标题"},
+            "voiceover_script": "这款口红绝了",
+            "shots": [{"scene_and_dialogue": "展示"}],
+        }
+        job.intent_result = {"key_messages": ["显白"], "intent": "推广"}
+
+        fail_result = ReviewResult(
+            passed=False,
+            score=50,
+            issues=[ReviewIssue(dimension="违规词", severity="high", location="开头", text="绝了")],
+            rewrite_suggestions=["去掉夸大词"],
+        )
+        rewritten = FusionResult(
+            title="新标题",
+            hook="修改后钩子",
+            script_segments=[ScriptSegment(text="修改后口播正文")],
+            cta="行动号召",
+            word_count=10,
+            estimated_duration=10,
+        )
+        pass_after = ReviewResult(passed=True, score=88, issues=[], rewrite_suggestions=[])
+
+        with (
+            patch.object(Reviewer, "review", side_effect=[fail_result, pass_after]),
+            patch.object(Reviewer, "rewrite", return_value=rewritten),
+        ):
+            out = vv._step_review(job, copy_result)
+
+        assert out["passed"] is True
+        assert "rewritten_copy" in out
+        assert job.generated_copy_text == "修改后口播正文"
+
+
+# ── 5) 端到端：每个 step 调用 loader 对应 prompt_type ──────────────
+
+
+class TestEndToEndLoaderUsed:
+    def test_each_step_calls_loader(self, job):
+        from apps.worker.worker_app.tasks import viral_video as vv
+        from packages.application.viral_video import prompt_loader as pl
+
+        called_types = []
+        real_get = pl.get_template
+
+        def spy_get(prompt_type, **kwargs):
+            called_types.append(prompt_type)
+            return real_get(prompt_type, **kwargs)
+
+        with (
+            patch.object(pl, "get_template", side_effect=spy_get),
+            patch("packages.shared.ai_service.call_vision", return_value=IMAGE_XML),
+            patch("packages.shared.ai_service.call_llm", return_value=INTENT_XML),
+        ):
+            # 1) image
+            img_res = vv._analyze_single_image(0, "https://img/1.jpg", "vlm", 15)
+            # 2) intent
+            intent_res = vv._step_intent_parsing(job, {"products": [img_res]})
+
+        # 前两步分别调用了 image_analysis 和 intent_parsing
+        assert "image_analysis" in called_types
+        assert "intent_parsing" in called_types
+
+        # script 和 review 单独验证（需要不同的 LLM 返回）
+        called_types_2 = []
+
+        def spy_get_2(prompt_type, **kwargs):
+            called_types_2.append(prompt_type)
+            return real_get(prompt_type, **kwargs)
+
+        with (
+            patch.object(pl, "get_template", side_effect=spy_get_2),
+            patch("packages.shared.ai_service.call_llm", return_value=STORYBOARD_XML),
+        ):
+            copy_res = vv._step_script_generation(job, intent_res, {"products": [img_res]})
+        assert "storyboard" in called_types_2
+
+        called_types_3 = []
+
+        def spy_get_3(prompt_type, **kwargs):
+            called_types_3.append(prompt_type)
+            return real_get(prompt_type, **kwargs)
+
+        from packages.application.viral_video.reviewer import Reviewer, ReviewResult
+
+        pass_result = ReviewResult(passed=True, score=90, issues=[], rewrite_suggestions=[])
+        job.intent_result = intent_res
+        job.copy_result = copy_res
+        with (
+            patch.object(pl, "get_template", side_effect=spy_get_3),
+            patch.object(Reviewer, "review", return_value=pass_result) as mock_review,
+        ):
+            review_res = vv._step_review(job, copy_res)
+        # review 步骤内部直接调用 Reviewer.review，该方法被 mock，因此 get_template 不会被调用；
+        # 此处验证 Reviewer.review 被调用即可说明 review 步骤走通了。
+        assert mock_review.called, "_step_review 未调用 Reviewer.review"
+        assert isinstance(review_res, dict) and "passed" in review_res