Merge pull request 'feat: 提示词控制展示格式 - summary_markdown + copy_display_markdown' (#2236) from feature/prompt-controlled-display-format into develop

- 引入 marked 轻量 markdown 库（gfm）
- 识别描述汇览：有 summary_markdown 时渲染 markdown，否则 fallback 到现有 vv-recog-line
- 分镜脚本区：新增文案预览区渲染 copy_display_markdown，编辑交互不变
- types 补充 summary_markdown / copy_display_markdown 字段
- 新增 vv-md-body markdown 排版样式

alembic/versions/104_v8_image_analysis_and_storyboard_v3.py

@@ -0,0 +1,242 @@
+# -*- coding: utf-8 -*-
+"""image_analysis v8 prompt + storyboard v3 prompt - 用户端展示格式 markdown 控制
+
+Revision ID: 104_v8_display_markdown
+Revises: 103_v7_prompt_and_tokens_3000
+Create Date: 2026-10-07
+
+变更：
+1. image_analysis v8: 在 v7 基础上 system_prompt 末尾追加「## 用户端展示格式」章节，
+   要求 VLM 在每张图的 JSON 里输出 summary_markdown 字段（markdown 格式的图片描述），
+   v8 设 is_active=true，v7 设 is_active=false。
+2. storyboard v3: 在 v2 基础上 system_prompt 追加要求 LLM 在 copy_result 中
+   输出 copy_display_markdown 字段（markdown 格式的完整文案展示），
+   v3 设 is_active=true，v2 设 is_active=false。
+"""
+
+from sqlalchemy import text
+
+from alembic import op
+
+revision = "104_v8_display_markdown"
+down_revision = "103_v7_prompt_and_tokens_3000"
+branch_labels = None
+depends_on = None
+
+# ── v8 追加的 system prompt 内容 ──────────────────────────────────────
+V8_SYSTEM_APPEND = """
+
+## 用户端展示格式
+
+对于每张分析的图片，在 JSON 中额外输出一个 **summary_markdown** 字段，用 markdown 格式写出给用户看的图片描述。
+
+格式要求（根据图片类型自适应）：
+
+**商品图（type=product）**示例：
+### 商品名称
+**品牌**：品牌名 | **类目**：服饰鞋包/美妆/数码/...
+**核心特征**
+- 特征1：描述
+- 特征2：描述
+**外观**：颜色+材质+设计描述
+**包装**：包装类型描述
+**文字信息**：包装上看到的文字
+
+**门店场景图（type=store）**示例：
+### 门店名称/类型
+**类型**：奶茶店/便利店/养生馆/...
+**品牌标识**：招牌文字描述
+**环境氛围**：店内整体感觉
+**陈列亮点**
+- 亮点1
+- 亮点2
+**氛围**：亲民/专业/时尚/...
+
+**人物图（type=person）**示例：
+### 人物描述
+**形象**：年龄段 + 风格
+**穿搭**
+- 上装：颜色+款式
+- 下装：颜色+款式
+- 配饰：...
+**气质**：表情+姿势+整体感觉
+
+**风景/场景图（type=scene）**示例：
+### 场景名称
+**类型**：自然风景/城市街景/动物/美食
+**主体**：画面主要元素
+**氛围**：整体感觉描述
+
+要求：
+- 内容真实具体，从实际图片分析得出
+- 用 markdown 语法：**加粗**、列表、标题
+- 控制在 100-200 字
+- 不要编造图片中没有的信息
+"""
+
+# ── storyboard v3 追加的 system prompt 内容 ──────────────────────────
+V3_STORYBOARD_APPEND = """
+
+## 用户端展示格式
+
+在输出分镜脚本的同时，在顶层输出一个 **copy_display_markdown** 字段（用 XML 标签 <copy_display_markdown> 包裹），用 markdown 格式写出完整文案展示。
+
+格式示例：
+# 标题/主题
+
+## 整体概要
+一句话描述视频内容
+
+## 分镜预览
+
+### 镜头1（0-3秒）
+**景别**：近景俯拍，缓慢推镜
+**画面**：场景描述
+**台词**：口播文本
+**动作**：人物动作描述
+
+### 镜头2（3-9秒）
+...
+
+## 完整口播
+完整口播文案文本
+
+要求：
+- 把所有分镜按时间顺序整理成易读的格式
+- 用 markdown 语法组织，**加粗**标签、##二级标题、列表等
+- 控制在 300-500 字
+- 让用户一眼看懂视频会拍成什么样
+"""
+
+
+def upgrade() -> None:
+    bind = op.get_bind()
+
+    # ── 1. image_analysis v8 ──────────────────────────────────────────
+    # 停用所有 active image_analysis prompt
+    bind.execute(
+        text(
+            "UPDATE viral_video_prompt_templates SET is_active = FALSE "
+            "WHERE prompt_type = 'image_analysis' AND is_active = TRUE"
+        )
+    )
+
+    # 读取 v7 的 prompt 内容作为基础
+    v7_row = bind.execute(
+        text(
+            "SELECT system_prompt, user_prompt_template, COALESCE(example_output, '') "
+            "FROM viral_video_prompt_templates "
+            "WHERE prompt_type = 'image_analysis' "
+            "ORDER BY version DESC LIMIT 1"
+        )
+    ).fetchone()
+
+    if v7_row:
+        v7_system = v7_row[0] or ""
+        v8_system = v7_system + V8_SYSTEM_APPEND
+        v8_user = v7_row[1] or "{image_url}"
+        v8_example = v7_row[2] or ""
+
+        # 幂等：已有 v8 则更新，否则插入
+        existing_v8 = bind.execute(
+            text("SELECT id FROM viral_video_prompt_templates " "WHERE prompt_type = 'image_analysis' AND version = 8")
+        ).fetchone()
+        if existing_v8:
+            bind.execute(
+                text(
+                    "UPDATE viral_video_prompt_templates SET is_active = TRUE, "
+                    "system_prompt = :sys, user_prompt_template = :usr, "
+                    "example_output = :ex, name = 'v8 用户端展示格式', "
+                    "updated_at = NOW() "
+                    "WHERE prompt_type = 'image_analysis' AND version = 8"
+                ),
+                {"sys": v8_system, "usr": v8_user, "ex": v8_example},
+            )
+        else:
+            bind.execute(
+                text(
+                    "INSERT INTO viral_video_prompt_templates "
+                    "(prompt_type, version, name, system_prompt, user_prompt_template, "
+                    "example_output, is_active, created_at, updated_at) "
+                    "VALUES ('image_analysis', 8, 'v8 用户端展示格式', "
+                    ":sys, :usr, :ex, TRUE, NOW(), NOW())"
+                ),
+                {"sys": v8_system, "usr": v8_user, "ex": v8_example},
+            )
+
+    # ── 2. storyboard v3 ─────────────────────────────────────────────
+    # 停用所有 active storyboard prompt
+    bind.execute(
+        text(
+            "UPDATE viral_video_prompt_templates SET is_active = FALSE "
+            "WHERE prompt_type = 'storyboard' AND is_active = TRUE"
+        )
+    )
+
+    # 读取当前 storyboard prompt
+    sb_row = bind.execute(
+        text(
+            "SELECT system_prompt, user_prompt_template, COALESCE(example_output, '') "
+            "FROM viral_video_prompt_templates "
+            "WHERE prompt_type = 'storyboard' "
+            "ORDER BY version DESC LIMIT 1"
+        )
+    ).fetchone()
+
+    if sb_row:
+        sb_system = sb_row[0] or ""
+        v3_system = sb_system + V3_STORYBOARD_APPEND
+        v3_user = sb_row[1] or ""
+        v3_example = sb_row[2] or ""
+
+        existing_v3 = bind.execute(
+            text("SELECT id FROM viral_video_prompt_templates " "WHERE prompt_type = 'storyboard' AND version = 3")
+        ).fetchone()
+        if existing_v3:
+            bind.execute(
+                text(
+                    "UPDATE viral_video_prompt_templates SET is_active = TRUE, "
+                    "system_prompt = :sys, user_prompt_template = :usr, "
+                    "example_output = :ex, name = 'v3 用户端展示格式', "
+                    "updated_at = NOW() "
+                    "WHERE prompt_type = 'storyboard' AND version = 3"
+                ),
+                {"sys": v3_system, "usr": v3_user, "ex": v3_example},
+            )
+        else:
+            bind.execute(
+                text(
+                    "INSERT INTO viral_video_prompt_templates "
+                    "(prompt_type, version, name, system_prompt, user_prompt_template, "
+                    "example_output, is_active, created_at, updated_at) "
+                    "VALUES ('storyboard', 3, 'v3 用户端展示格式', "
+                    ":sys, :usr, :ex, TRUE, NOW(), NOW())"
+                ),
+                {"sys": v3_system, "usr": v3_user, "ex": v3_example},
+            )
+
+
+def downgrade() -> None:
+    bind = op.get_bind()
+
+    # 删除 v8
+    bind.execute(
+        text("DELETE FROM viral_video_prompt_templates " "WHERE prompt_type = 'image_analysis' AND version = 8")
+    )
+    # 恢复 v7 active
+    bind.execute(
+        text(
+            "UPDATE viral_video_prompt_templates SET is_active = TRUE, updated_at = NOW() "
+            "WHERE prompt_type = 'image_analysis' AND version = 7"
+        )
+    )
+
+    # 删除 v3
+    bind.execute(text("DELETE FROM viral_video_prompt_templates " "WHERE prompt_type = 'storyboard' AND version = 3"))
+    # 恢复 storyboard v2 active
+    bind.execute(
+        text(
+            "UPDATE viral_video_prompt_templates SET is_active = TRUE, updated_at = NOW() "
+            "WHERE prompt_type = 'storyboard' AND version = 2"
+        )
+    )

apps/web/package-lock.json

@@ -14,6 +14,7 @@
         "axios": "^1.7.2",
         "classnames": "^2.5.1",
         "dayjs": "^1.11.23",
+        "marked": "^12.0.2",
         "mp4box": "^2.4.1",
         "react": "^18.3.1",
         "react-dom": "^18.3.1",
@@ -4502,6 +4503,17 @@
         "url": "https://github.com/sponsors/sindresorhus"
       }
     },
+    "node_modules/marked": {
+      "version": "12.0.2",
+      "resolved": "https://registry.npmmirror.com/marked/-/marked-12.0.2.tgz",
+      "integrity": "sha512-qXUm7e/YKFoqFPYPa3Ukg9xlI5cyAtGmyEIzMfW//m6kXwCy2Ps9DYf5ioijFKQ8qyuscrHoY04iJGctu2Kg0Q==",
+      "bin": {
+        "marked": "bin/marked.js"
+      },
+      "engines": {
+        "node": ">= 18"
+      }
+    },
     "node_modules/math-intrinsics": {
       "version": "1.1.0",
       "resolved": "https://registry.npmjs.org/math-intrinsics/-/math-intrinsics-1.1.0.tgz",

apps/web/package.json

@@ -25,6 +25,7 @@
     "axios": "^1.7.2",
     "classnames": "^2.5.1",
     "dayjs": "^1.11.23",
+    "marked": "^12.0.2",
     "mp4box": "^2.4.1",
     "react": "^18.3.1",
     "react-dom": "^18.3.1",

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

@@ -83,6 +83,8 @@ export interface ImageProductAnalysis {
   label_text?: string
   selling_points?: string
   image_index?: number
+  /** v8: 用户端展示用的 markdown 描述（由提示词控制排版） */
+  summary_markdown?: string
 }
 
 export interface ImageAnalysisResult {
@@ -132,6 +134,8 @@ export interface CopyResult {
   /** 向后兼容：= voiceover_script */
   suggested_copy?: string
   title?: string
+  /** v3 storyboard: 用户端展示用的 markdown 文案（由提示词控制排版） */
+  copy_display_markdown?: string
   /** v1.5 旧字段兼容（老数据降级时可能出现） */
   scenes?: Array<{ shot: string; narration: string; duration?: number }>
 }

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

@@ -1973,3 +1973,99 @@
   padding-bottom: 6px;
   border-bottom: 1px dashed #e5e7eb;
 }
+
+/* ─────────── markdown 渲染（提示词控制展示格式） ─────────── */
+.vv-recog-md {
+  padding: 4px 0;
+}
+.vv-copy-preview {
+  margin-bottom: 14px;
+  padding: 12px 14px;
+  background: linear-gradient(180deg, #faf7ff 0%, #f6f2ff 100%);
+  border: 1px solid #ece4fb;
+  border-radius: 10px;
+}
+.vv-copy-preview-h {
+  margin: 0 0 8px;
+  border-bottom: none;
+  padding-bottom: 0;
+}
+.vv-md-body {
+  font-size: 13px;
+  line-height: 1.7;
+  color: #374151;
+  word-break: break-word;
+}
+.vv-md-body h1,
+.vv-md-body h2,
+.vv-md-body h3,
+.vv-md-body h4 {
+  margin: 10px 0 6px;
+  font-weight: 600;
+  color: #1f2937;
+  line-height: 1.4;
+}
+.vv-md-body h1 {
+  font-size: 18px;
+}
+.vv-md-body h2 {
+  font-size: 16px;
+}
+.vv-md-body h3 {
+  font-size: 15px;
+}
+.vv-md-body h4 {
+  font-size: 14px;
+}
+.vv-md-body p {
+  margin: 6px 0;
+}
+.vv-md-body ul,
+.vv-md-body ol {
+  margin: 6px 0;
+  padding-left: 20px;
+}
+.vv-md-body li {
+  margin: 3px 0;
+}
+.vv-md-body strong {
+  color: #111827;
+  font-weight: 600;
+}
+.vv-md-body blockquote {
+  margin: 8px 0;
+  padding: 4px 12px;
+  border-left: 3px solid #7c3aed;
+  background: rgba(124, 58, 237, 0.05);
+  color: #4b5563;
+}
+.vv-md-body code {
+  padding: 1px 5px;
+  background: #f3f4f6;
+  border-radius: 4px;
+  font-size: 12px;
+  color: #be185d;
+}
+.vv-md-body a {
+  color: #7c3aed;
+  text-decoration: none;
+}
+.vv-md-body a:hover {
+  text-decoration: underline;
+}
+.vv-md-body table {
+  border-collapse: collapse;
+  margin: 8px 0;
+  width: 100%;
+}
+.vv-md-body th,
+.vv-md-body td {
+  border: 1px solid #e5e7eb;
+  padding: 6px 10px;
+  text-align: left;
+}
+.vv-md-body hr {
+  border: none;
+  border-top: 1px solid #e5e7eb;
+  margin: 12px 0;
+}

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

@@ -1,5 +1,6 @@
 import React, { useCallback, useEffect, useRef, useState } from "react"
 import axios from "axios"
+import { marked } from "marked"
 import {
   PlusOutlined,
   CloseOutlined,
@@ -150,6 +151,16 @@ type TabTask = {
   audioInst: HTMLAudioElement | null
 }
 
+/* ── marked 配置：禁用 mangle/headerIds，输出干净 HTML ── */
+marked.setOptions({ gfm: true, breaks: false })
+const renderMarkdown = (md: string): string => {
+  try {
+    return marked.parse(md ?? "", { async: false }) as string
+  } catch {
+    return (md ?? "").replace(/&/g, "&amp;").replace(/</g, "&lt;")
+  }
+}
+
 /* ─────────── 常量 ─────────── */
 
 const LANGUAGES = ["中文(普通话)", "粤语", "英语", "日语", "韩语"]
@@ -296,6 +307,8 @@ interface Storyboard {
   hard_constraints: string[]
   negative_prompts: string[]
   voiceover_script: string
+  /** v3: 用户端展示用 markdown 文案（由提示词控制排版） */
+  copy_display_markdown: string
 }
 
 /** 兼容旧 copy_result（final_copy/title/scenes）→ 新 Storyboard 结构 */
@@ -323,6 +336,7 @@ function copyResultToStoryboard(cr: CopyResult | null | undefined): Storyboard |
       hard_constraints: Array.isArray(cr.hard_constraints) ? cr.hard_constraints : [],
       negative_prompts: Array.isArray(cr.negative_prompts) ? cr.negative_prompts : [],
       voiceover_script: cr.voiceover_script || cr.final_copy || cr.suggested_copy || "",
+      copy_display_markdown: cr.copy_display_markdown || "",
     }
   }
   // 兜底：旧结构转简单分镜
@@ -358,6 +372,7 @@ function copyResultToStoryboard(cr: CopyResult | null | undefined): Storyboard |
     hard_constraints: [],
     negative_prompts: [],
     voiceover_script: finalCopy,
+    copy_display_markdown: cr.copy_display_markdown || "",
   }
 }
 
@@ -412,6 +427,7 @@ const MOCK_STORYBOARD: Storyboard = {
   negative_prompts: ["冷色调", "模糊", "变形", "水印文字", "卡通风格", "空无一人"],
   voiceover_script:
     "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。",
+  copy_display_markdown: "",
 }
 
 const fmtSize = (bytes: number | undefined) => {
@@ -1217,67 +1233,76 @@ const ViralVideoPage: React.FC = () => {
           <CheckCircleFilled style={{ color: "#10b981" }} />
           识别描述汇览
         </div>
-        {products.map((p, i) => (
-          <div key={i} className="vv-recog-item">
-            <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>
+        {products.map((p, i) =>
+          p.summary_markdown ? (
+            <div key={i} className="vv-recog-item vv-recog-md">
+              <div
+                className="vv-md-body"
+                dangerouslySetInnerHTML={{ __html: renderMarkdown(p.summary_markdown) }}
+              />
             </div>
-            {featureText(p.key_features ?? p.features) && (
+          ) : (
+            <div key={i} className="vv-recog-item">
               <div className="vv-recog-line">
-                <span className="vv-recog-k">核心特征：</span>
-                <span className="vv-recog-v">{featureText(p.key_features ?? p.features)}</span>
+                <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>
-            )}
-            {p.colors && p.colors.length > 0 && (
-              <div className="vv-recog-line">
-                <span className="vv-recog-k">主色调：</span>
-                <span className="vv-recog-v">{p.colors.join(" / ")}</span>
-              </div>
-            )}
-            {p.material_or_texture && (
-              <div className="vv-recog-line">
-                <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>
-        ))}
+              {featureText(p.key_features ?? p.features) && (
+                <div className="vv-recog-line">
+                  <span className="vv-recog-k">核心特征：</span>
+                  <span className="vv-recog-v">{featureText(p.key_features ?? p.features)}</span>
+                </div>
+              )}
+              {p.colors && p.colors.length > 0 && (
+                <div className="vv-recog-line">
+                  <span className="vv-recog-k">主色调：</span>
+                  <span className="vv-recog-v">{p.colors.join(" / ")}</span>
+                </div>
+              )}
+              {p.material_or_texture && (
+                <div className="vv-recog-line">
+                  <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>
+          ),
+        )}
       </div>
     )
   }
@@ -1416,6 +1441,19 @@ const ViralVideoPage: React.FC = () => {
     return (
       <div className="vv-copy-box vv-storyboard">
         <div className="vv-sb-doc">
+          {/* 文案预览（提示词控制排版，只读；编辑在下方分镜字段中进行） */}
+          {sb.copy_display_markdown && (
+            <div className="vv-copy-preview">
+              <h4 className="vv-sb-h vv-copy-preview-h">
+                <FileTextOutlined style={{ color: "#7c3aed", marginRight: 6 }} />
+                文案预览
+              </h4>
+              <div
+                className="vv-md-body"
+                dangerouslySetInnerHTML={{ __html: renderMarkdown(sb.copy_display_markdown) }}
+              />
+            </div>
+          )}
           {/* 视频总览 */}
           <h4 className="vv-sb-h">视频总览</h4>
           <p className="vv-sb-inline-row">

apps/worker/worker_app/tasks/viral_video.py

@@ -948,6 +948,10 @@ def _script_from_xml(raw: str, job: ViralVideoJob) -> dict | None:
     base["final_copy"] = joined
     base["suggested_copy"] = joined
     base["title"] = base["overview"]["theme"]
+    # 提取 copy_display_markdown（用户端展示格式）
+    copy_display_md = xp.text_of(raw, "copy_display_markdown")
+    if copy_display_md:
+        base["copy_display_markdown"] = copy_display_md
     return base
 
 
@@ -2072,18 +2076,27 @@ def _run_render_pipeline(job_id: str, session, repo, job) -> dict:
     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...")
-            # #2040: Reviewer 已在 _step_review 内完成 1 次自动重写
-            rewritten = review_result.get("rewritten_copy")
-            if isinstance(rewritten, dict) and rewritten:
-                copy_result = rewritten
-            else:
-                # #2218: 审核重写失败不再从意图解析重跑，直接报错让用户重新生成文案
-                logger.error(
-                    "[爆款视频][阶段3] 合规审核未通过且自动重写失败 job_id=%s，终止渲染",
+            # #2233: 如果issues为空但passed=False，说明是LLM超时/异常导致的误判，降级放行
+            _issues = review_result.get("issues") or []
+            if not _issues:
+                logger.warning(
+                    "[爆款视频][阶段3] 审核未通过但无具体问题（可能LLM超时），降级放行 job_id=%s",
                     job_id,
                 )
-                raise ValueError("文案合规审核未通过，请修改文案后重试或重新生成文案")
+                review_result["passed"] = True
+            else:
+                _emit_progress(job_id, ViralVideoStage.REVIEW, 67.0, "审核未通过，正在自动重写...")
+                # #2040: Reviewer 已在 _step_review 内完成 1 次自动重写
+                rewritten = review_result.get("rewritten_copy")
+                if isinstance(rewritten, dict) and rewritten:
+                    copy_result = rewritten
+                else:
+                    # #2218: 审核重写失败不再从意图解析重跑，直接报错让用户重新生成文案
+                    logger.error(
+                        "[爆款视频][阶段3] 合规审核未通过且自动重写失败 job_id=%s，终止渲染",
+                        job_id,
+                    )
+                    raise ValueError("文案合规审核未通过，请修改文案后重试或重新生成文案")
             job.copy_result = copy_result
             job.generated_copy_text = copy_result.get("voiceover_script", "") or ""
             _save_job(repo, job, session)

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

@@ -22,6 +22,90 @@ _AGE_PREFIX = {"青年": "年轻", "中年": "中年", "老年": "老年"}
 _GENDER_WORD = {"男": "男性", "女": "女性"}
 
 
+def _build_summary_markdown(product: dict) -> str:
+    """根据结构化字段生成 summary_markdown 兜底（VLM未返回时调用）。"""
+    p = product or {}
+    name = p.get("name") or "未识别"
+    brand = p.get("brand") or ""
+    category = p.get("category") or ""
+    appearance = p.get("appearance") or ""
+    packaging = p.get("packaging") or ""
+    scene = p.get("scene") or ""
+    mood = p.get("mood") or ""
+    key_features = p.get("key_features") or []
+    text_on_package = p.get("text_on_package") or []
+    store_type = p.get("store_type") or ""
+    brand_signage = p.get("brand_signage") or ""
+    visual_elements = p.get("visual_elements") or []
+    atmosphere = p.get("atmosphere") or ""
+    outfit_style = p.get("outfit_style") or ""
+    upper_wear = p.get("upper_wear") or ""
+    lower_wear = p.get("lower_wear") or ""
+    expression = p.get("expression") or ""
+    has_person = p.get("has_person", False)
+    person_count = p.get("person_count") or 0
+
+    lines = []
+
+    # 门店场景
+    if store_type or "门店" in str(category) or "店铺" in str(category):
+        lines.append(f"### {brand_signage or store_type or '门店'}")
+        if store_type:
+            lines.append(f"**类型**：{store_type}")
+        if brand_signage:
+            lines.append(f"**品牌标识**：{brand_signage}")
+        if atmosphere:
+            lines.append(f"**氛围**：{atmosphere}")
+        if visual_elements:
+            lines.append("**陈列亮点**")
+            for e in visual_elements[:5]:
+                lines.append(f"- {e}")
+        return "\n".join(lines)
+
+    # 人物
+    if has_person or person_count or outfit_style:
+        lines.append("### 人物描述")
+        if person_count:
+            lines.append(f"**人数**：{person_count}人")
+        if outfit_style:
+            lines.append(f"**风格**：{outfit_style}")
+        wear_parts = []
+        if upper_wear:
+            wear_parts.append(f"上装：{upper_wear}")
+        if lower_wear:
+            wear_parts.append(f"下装：{lower_wear}")
+        if wear_parts:
+            lines.append("**穿搭**")
+            for w in wear_parts:
+                lines.append(f"- {w}")
+        if expression:
+            lines.append(f"**气质**：{expression}")
+        return "\n".join(lines)
+
+    # 商品（默认）
+    lines.append(f"### {name}")
+    if brand and brand != "无法判断":
+        lines.append(f"**品牌**：{brand}")
+    if category and category != "非产品图" and category != "无法判断":
+        lines.append(f"**类目**：{category}")
+    if key_features:
+        lines.append("**核心特征**")
+        for kf in key_features[:5]:
+            lines.append(f"- {kf}")
+    if appearance:
+        lines.append(f"**外观**：{appearance[:100]}")
+    if packaging and packaging != "无法判断":
+        lines.append(f"**包装**：{packaging}")
+    if text_on_package:
+        lines.append(f"**文字信息**：{'、'.join(text_on_package[:3])}")
+    if scene and scene != "通用":
+        lines.append(f"**场景**：{scene}")
+    if mood:
+        lines.append(f"**氛围**：{mood}")
+
+    return "\n".join(lines) if lines else f"### {name}\n暂无详细描述"
+
+
 def _person_subject(gender: str, age: str) -> str:
     gw = _GENDER_WORD.get(gender, "")
     if age == "儿童":
@@ -384,7 +468,11 @@ def assemble_result(idx: int, fast_json: dict | None, ocr_texts: list[str]) -> d
         result = _assemble_v4(idx, fj, ocr_texts)
     else:
         result = _assemble_old(idx, fj, ocr_texts)
-    return _apply_partial_fallback(result, fj)
+    result = _apply_partial_fallback(result, fj)
+    # summary_markdown: 优先用VLM输出，否则兜底生成
+    if "summary_markdown" not in result or not result.get("summary_markdown"):
+        result["summary_markdown"] = _build_summary_markdown(result)
+    return result
 
 
 def _apply_partial_fallback(result: dict[str, Any], fj: dict) -> dict[str, Any]:
@@ -681,8 +769,13 @@ def _assemble_v4(idx: int, fj: dict, ocr_texts: list[str]) -> dict[str, Any]:
                 brand = "无法判断"
         else:
             brand = brand_raw
-        # name兜底：store_type为空时用brand
-        name = store_type if store_type != "店铺" else (brand if brand != "无法判断" else store_type)
+        # name兜底：更保守的策略
+        # - store_type有具体值（非"店铺"）时直接用store_type
+        # - store_type为默认"店铺"时，用"门店门头"而非brand（避免与brand字段重复）
+        if store_type and store_type != "店铺":
+            name = store_type
+        else:
+            name = "门店门头"
         category = "门店场景"
         # appearance: store_layout + furnishings + 陈设色调
         appearance_parts = []

packages/application/viral_video/reviewer.py

@@ -53,6 +53,9 @@ _LOCATIONS = ["title", "hook", "body_points", "cta", "script_segments"]
 
 
 class Reviewer:
+    # markdown展示字段不参与合规审核（避免格式字符误判）
+    _MARKDOWN_FIELDS = {"summary_markdown", "copy_display_markdown"}
+
     def __init__(self, client=None):
         if client is None:
             try:
@@ -70,8 +73,9 @@ class Reviewer:
         local = self._rule_check(fusion, intent, fusion_level)
         llm_result = self._llm_review(fusion, intent, fusion_level)
         if llm_result is None:
+            # LLM审核失败（超时/网络错误等），降级放行，不阻断渲染
             return ReviewResult(
-                passed=not local,
+                passed=True,
                 issues=local,
                 rewrite_suggestions=[],
                 raw="",
@@ -86,6 +90,17 @@ class Reviewer:
         )
 
     def _llm_review(self, fusion: FusionResult, intent: IntentResult, fusion_level: str) -> Optional[ReviewResult]:
+        try:
+            return self._llm_review_inner(fusion, intent, fusion_level)
+        except Exception as e:
+            import logging
+
+            logging.getLogger(__name__).warning("[Reviewer] LLM审核调用异常，降级放行: %s", e)
+            return None
+
+    def _llm_review_inner(
+        self, fusion: FusionResult, intent: IntentResult, fusion_level: str
+    ) -> Optional[ReviewResult]:
         template = get_template("review")
         system = render_system_prompt(template)
         user = render_user_prompt(
@@ -303,10 +318,13 @@ class Reviewer:
 
     @staticmethod
     def _fusion_text(fusion: FusionResult) -> str:
+        _MARKDOWN_FIELDS = {"summary_markdown", "copy_display_markdown"}
         parts = [fusion.title, fusion.hook]
         parts += [p.text for p in fusion.body_points]
         parts += [s.text for s in fusion.script_segments]
         parts.append(fusion.cta)
+        # 过滤掉markdown展示字段，避免格式字符被误判
+        parts = [p for p in parts if not any(mk in p for mk in _MARKDOWN_FIELDS)]
         return "\n".join(p for p in parts if p)
 
     @staticmethod

packages/shared/ai_client.py

@@ -337,13 +337,13 @@ class DoubaoClient:
                 self.last_finish_reason = finish_reason
                 _elapsed = time.time() - _t0
                 logger.info(
-                    "[doubao] chat_completion 完成 model=%s tokens_in=%d tokens_out=%d elapsed=%.1fs attempt=%d timeout=%d",
+                    "[doubao] chat_completion 完成 model=%s tokens_in=%d tokens_out=%d elapsed=%.1fs attempt=%d timeout=%s",
                     payload.get("model"),
                     data.get("usage", {}).get("prompt_tokens", 0),
                     data.get("usage", {}).get("completion_tokens", 0),
                     _elapsed,
                     attempt + 1,
-                    _req_timeout,
+                    getattr(_req_timeout, "read", _req_timeout),
                 )
                 return content.strip()
             except Exception as e: