Merge branch 'develop' into refactor/viral-video-frontend-p1-component-split

1. _step_script_generation 直线化：
   - 新增 _validate_copy_result / _call_llm_once 两个小函数
   - 重试链统一为 for 循环：fast@0.8 → fast@0.6 → pro@0.7 → fallback
   - 删除冗余 compress-retry 分支
   - 删除 _try_gen / _try_gen_stream 两个重复函数，流式/同步统一由 _call_llm_once 根据 on_delta 自动选择
   - 删除 _validate_and_normalize_script（JSON fallback 已在 PR1 废弃，XML-only 路径）

2. 信任链公共函数：
   - 新增 _collect_person_portraits / _build_sparse_trusted_urls / _preheat_trust_chain_sync / _persist_pre_trusted_images
   - _start_trust_chain_preheat 与 _step_render 现场信任链复用同一套筛选/构建逻辑，消除 ~60 行重复

3. Celery task 样板抽取装饰器 with_viral_job：
   - 5个 Celery task（pipeline/resume/style_analysis/analyze/generate_copy/render）样板代码从 ~30 行/个降到 0
   - 支持 require_status / start_heartbeat / recover_stale / mark_running_on_entry / refund_on_failure 参数
   - bind=True 自适应（Celery self 或直接 job_id 字符串调用都兼容）
   - Retry 异常正确重新 raise，其余异常统一 mark_failed_and_notify
   - 阶段3 render 失败兜底退款由装饰器 refund_on_failure=True 统一处理

4. API 层积分结算提取 _settle_viral_video_credits：
   - 支持 deduct_first / recompute / delta 三种 action，confirm_copy / retry 共用
   - 消除 ~80 行重复
   - hard_constraints/negative_prompts fallback 统一从 constants 导入（API_FALLBACK_*），避免 worker/API 漂移
   - 入队失败处加 TODO 注释：应返回 503 而非 mark_failed+200（待后续 PR 修）

验证：python ast.parse OK，worker/api import OK，248 viral 单测全通过。
净减 112 行（constants 新增 36 行；worker -267 行；api -21 行；tests -54 行）

.ci-trigger

@@ -1,2 +1 @@
-CI trigger file - safe to delete
-retrigger at 2026-09-15 20:31:24 UTC
+trigger staging deploy for 3fa1fbf0 (MOCK_STORYBOARD fix) at 2026-10-12 13:40:16

.gitea/workflows/ci-pipeline.yml

@@ -437,7 +437,7 @@ jobs:
     if: "always() && needs.dedupe-check.outputs.skip_tests != 'true' && needs.check-frontend-only.outputs.skip_backend != 'true'"
     name: Unit Tests
     runs-on: ci-l2
-    timeout-minutes: 8
+    timeout-minutes: 20
     env:
       PIP_CACHE_DIR: /root/.cache/pip
       PIP_NO_CACHE_DIR: ''

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"
+        )
+    )

alembic/versions/105_narration_first.py

@@ -0,0 +1,123 @@
+# -*- coding: utf-8 -*-
+"""image_analysis v8 + storyboard v3 叙述优先重写版（架构大简化）
+
+Revision ID: 105_narration_first
+Revises: 104_v8_display_markdown
+Create Date: 2026-10-07
+
+变更：
+1. image_analysis v8：用「叙述优先」版整体替换 104 的 append 版——VLM 主交付物是
+   自然叙述 summary_markdown，结构化字段仅保留 type/name/brand/has_person，
+   顶层 products 改名 images；v8 active，其余 image_analysis 全部 deactivate。
+2. storyboard v3：整体替换为风格重写版（口播口语化、画面有画面感、
+   copy_display_markdown 流畅叙述）；v3 active，其余 storyboard deactivate。
+3. intent_parsing 类型模板全部 deactivate（意图解析步骤已删除）。
+模板内容直接取自 packages.application.viral_video.prompts.DEFAULT_TEMPLATES，
+保证代码默认值与 DB seed 完全一致。
+"""
+
+from sqlalchemy import text
+
+from alembic import op
+from packages.application.viral_video.prompts import DEFAULT_TEMPLATES
+
+revision = "105_narration_first"
+down_revision = "104_v8_display_markdown"
+branch_labels = None
+depends_on = None
+
+
+def _tpl(prompt_type: str, version: int) -> dict:
+    """取模板：先按指定版本找，找不到则取该类型最新版本（兼容 v3→v4 升级）。"""
+    # 先按指定版本找
+    for t in DEFAULT_TEMPLATES:
+        if t["prompt_type"] == prompt_type and t["version"] == version:
+            return t
+    # 找不到则取最新版本
+    candidates = [t for t in DEFAULT_TEMPLATES if t["prompt_type"] == prompt_type]
+    if candidates:
+        latest = max(candidates, key=lambda x: x["version"])
+        return latest
+    raise RuntimeError("default template missing: %s" % prompt_type)
+
+
+def _upsert(bind, t: dict) -> None:
+    existing = bind.execute(
+        text("SELECT id FROM viral_video_prompt_templates " "WHERE prompt_type = :pt AND version = :ver"),
+        {"pt": t["prompt_type"], "ver": t["version"]},
+    ).fetchone()
+    params = {
+        "pt": t["prompt_type"],
+        "ver": t["version"],
+        "name": t["name"],
+        "sys": t["system_prompt"],
+        "usr": t["user_prompt_template"],
+        "ex": t.get("example_output", "") or "",
+    }
+    if existing:
+        bind.execute(
+            text(
+                "UPDATE viral_video_prompt_templates SET name = :name, "
+                "system_prompt = :sys, user_prompt_template = :usr, "
+                "example_output = :ex, is_active = TRUE, updated_at = NOW() "
+                "WHERE prompt_type = :pt AND version = :ver"
+            ),
+            params,
+        )
+    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 (:pt, :ver, :name, :sys, :usr, :ex, TRUE, NOW(), NOW())"
+            ),
+            params,
+        )
+
+
+def upgrade() -> None:
+    bind = op.get_bind()
+
+    # 1. image_analysis：停用全部后写入叙述优先 v8
+    bind.execute(
+        text("UPDATE viral_video_prompt_templates SET is_active = FALSE " "WHERE prompt_type = 'image_analysis'")
+    )
+    _upsert(bind, _tpl("image_analysis", 8))
+
+    # 2. storyboard：停用全部后写入重写版 v3
+    bind.execute(text("UPDATE viral_video_prompt_templates SET is_active = FALSE " "WHERE prompt_type = 'storyboard'"))
+    _upsert(bind, _tpl("storyboard", 3))
+
+    # 3. intent_parsing 已废弃：全部停用
+    bind.execute(
+        text("UPDATE viral_video_prompt_templates SET is_active = FALSE " "WHERE prompt_type = 'intent_parsing'")
+    )
+
+    # 4. review 模板确保 active
+    bind.execute(text("UPDATE viral_video_prompt_templates SET is_active = TRUE " "WHERE prompt_type = 'review'"))
+
+
+def downgrade() -> None:
+    bind = op.get_bind()
+    # 恢复 104 的 v8/v3 无法重建（内容已替换），仅把版本 active 状态回退：
+    # 停用新版，尝试恢复 v7 / v2
+    bind.execute(
+        text(
+            "UPDATE viral_video_prompt_templates SET is_active = FALSE "
+            "WHERE prompt_type IN ('image_analysis','storyboard') "
+            "AND version IN (8, 3)"
+        )
+    )
+    bind.execute(
+        text(
+            "UPDATE viral_video_prompt_templates SET is_active = TRUE "
+            "WHERE prompt_type = 'image_analysis' AND version = 7"
+        )
+    )
+    bind.execute(
+        text(
+            "UPDATE viral_video_prompt_templates SET is_active = TRUE "
+            "WHERE prompt_type = 'storyboard' AND version = 2"
+        )
+    )

alembic/versions/106_system_settings.py

@@ -0,0 +1,43 @@
+# -*- coding: utf-8 -*-
+"""system_settings 正式建表（#2246）
+
+Revision ID: 106_system_settings
+Revises: 105_narration_first
+Create Date: 2026-10-08
+
+system_settings 表此前在 staging 手工创建（对应 034 占位迁移），
+此处补正式迁移保证其他环境一致。CREATE TABLE/INDEX 使用 IF NOT EXISTS，
+对已手工建表的环境幂等。
+"""
+
+from alembic import op
+
+revision = "106_system_settings"
+down_revision = "105_narration_first"
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    op.execute("""
+        CREATE TABLE IF NOT EXISTS system_settings (
+            id VARCHAR(36) NOT NULL,
+            setting_key VARCHAR(100) NOT NULL,
+            setting_value TEXT,
+            setting_type VARCHAR(20) NOT NULL,
+            description VARCHAR(255) NOT NULL DEFAULT '',
+            is_public BOOLEAN NOT NULL DEFAULT FALSE,
+            updated_by VARCHAR(36),
+            category VARCHAR(50) NOT NULL DEFAULT 'general',
+            created_at TIMESTAMP WITH TIME ZONE NOT NULL DEFAULT NOW(),
+            updated_at TIMESTAMP WITH TIME ZONE NOT NULL DEFAULT NOW(),
+            CONSTRAINT pk_system_settings PRIMARY KEY (id),
+            CONSTRAINT uq_system_settings_setting_key UNIQUE (setting_key)
+        )
+        """)
+    op.execute("CREATE INDEX IF NOT EXISTS ix_system_settings_category " "ON system_settings (category)")
+
+
+def downgrade() -> None:
+    op.execute("DROP INDEX IF EXISTS ix_system_settings_category")
+    op.execute("DROP TABLE IF EXISTS system_settings")

alembic/versions/107_add_language_to_viral_video_jobs.py

@@ -0,0 +1,23 @@
+"""add language column to viral_video_jobs
+
+Revision ID: 107
+Revises: 106_system_settings
+Create Date: 2026-10-09
+
+"""
+
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision = "107"
+down_revision = "106_system_settings"
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    op.execute("ALTER TABLE viral_video_jobs " "ADD COLUMN IF NOT EXISTS language VARCHAR(20) NOT NULL DEFAULT 'zh-CN'")
+
+
+def downgrade() -> None:
+    op.execute("ALTER TABLE viral_video_jobs DROP COLUMN IF EXISTS language")

alembic/versions/108_storyboard_v4_multivoice.py

@@ -0,0 +1,235 @@
+# -*- coding: utf-8 -*-
+"""storyboard v4 - 多角色对话 + 废除旁白 + visual 5要素
+
+Revision ID: 108_storyboard_v4_multivoice
+Revises: 107
+Create Date: 2026-10-10
+
+变更：
+1. storyboard v4: 废除旁白思维，所有voiceover必须是角色台词
+   - 增加<speaker>标签，每镜必须标注说话人
+   - visual强制5要素结构（景别/运镜/动作/环境/光线），每镜不少于30字
+   - voiceover_script用[speaker:xxx]标记格式
+2. 旧版storyboard模板is_active设为false
+3. 检查image_analysis和review是否有active模板，没有则插入保底版本
+"""
+
+from sqlalchemy import text
+
+from alembic import op
+
+revision = "108_storyboard_v4_multivoice"
+down_revision = "107"
+branch_labels = None
+depends_on = None
+
+# ── storyboard v4 system_prompt ──────────────────────────────────────
+V4_STORYBOARD_SYSTEM = """你是一名懂短视频的编导和口播文案高手。你会拿到图片的真实观察、营销目的和用户参数，请一次性完成对营销意图的理解，并产出可直接拍摄/生成的分镜脚本。不要单独输出"意图解析"，意图要直接体现在台词和分镜里。
+
+## 核心设计原则（必须严格遵守）
+1. **废除旁白思维**：所有视频类型——无论对话短剧/口播带货/获客引流/品牌故事——voiceover 必须是人物说的话（第一人称或角色对白），绝对不能出现第三人称旁白解说。观众看的是人在演、在说。
+2. **严禁第三人称解说性台词**：如"接下来展示...""这款产品..."这类上帝视角描述禁止出现在 voiceover 中。
+3. **短剧类营销目的**（对话短剧/反转短剧/悬念短剧/情绪短片）：双角色对话格式"甲：xxx 乙：xxx"，镜头在角色间切换。
+4. **口播类**（口播带货/促销转化/功能演示/痛点解决/获客引流/账号涨粉/活动通知/场景种草）：第一人称对镜头说话，像真人出镜。
+
+## 输出格式（XML，严格按结构输出，不要输出额外解释）
+<script>
+  <copy_display_markdown><![CDATA[直接展示给用户看的成片文案，用 Markdown 写成流畅叙述]]></copy_display_markdown>
+  <clips>
+    <clip index="1">
+      <time_range>0-3秒</time_range>
+      <speaker>说话人标识（如"店主""顾客""主播"）</speaker>
+      <voiceover>这一镜的角色台词（人物说的话，不是旁白）</voiceover>
+      <visual>【景别】【镜头运动】【人物动作/表情】【环境/道具】【光线氛围】5要素结构，不少于30字</visual>
+      <reference_image_index>0</reference_image_index>
+    </clip>
+  </clips>
+  <voiceover_script>把所有 clip 的 voiceover 连成完整台词稿，用 [speaker:xxx] 标记每个说话段落</voiceover_script>
+  <theme>一句话主题</theme>
+  <negative>【反套路化要求】
+禁止使用"家人们谁懂啊""绝绝子""宝子们""家人们""太绝了""yyds"等烂大街网络词；
+禁止固定模板化开头；语言要像真人朋友之间的分享，自然、具体、有信息量。</negative>
+</script>
+
+## 写作要求
+1. **台词（voiceover）**：像真人面对镜头说话或角色对白，短句、口语化、有停顿有情绪，开头 3 秒给出钩子；不要书面腔，不要机械报参数。严禁第三人称解说。
+2. **说话人（speaker）**：每个 clip 必须标注说话人标识，如"店主""顾客""主播""我"等。短剧类必须有至少 2 个不同角色。
+3. **画面描述（visual）**：强制 5 要素结构——【景别】【镜头运动】【人物动作/表情】【环境/道具】【光线氛围】，每镜 visual 不少于 30 字，要具体到"闭眼听台词能想象出画面"。
+4. **voiceover_script 格式**：用 [speaker:xxx] 标记每个说话段落，如"[speaker:店主]你是不是也觉得...[speaker:顾客]是啊，怎么回事？"
+5. copy_display_markdown：直接展示给最终用户的文案，用 Markdown 写成自然、流畅、有感染力的成片文案。
+6. 内容必须来自图片观察与用户给出的信息，不编造卖点、不夸大、不使用绝对化用语和虚假承诺。
+7. reference_image_index 填本镜参考图片序号（从 0 开始），没有合适参考图填 -1。
+8. 分镜数量与时长匹配总时长，节奏紧凑。
+9. **口播字数硬约束**（必须严格遵守）：按每秒约 2.5~3 个中文字（正常口播语速）计算：
+   - 5秒视频：voiceover_script 总字数 12~15 字
+   - 10秒视频：voiceover_script 总字数 25~30 字
+   - 15秒视频：voiceover_script 总字数 35~45 字
+   - 20秒视频：voiceover_script 总字数 50~60 字
+   - 30秒视频：voiceover_script 总字数 75~90 字
+   - 宁可少写也不要多写，超长会导致 TTS 音频超出视频时长限制
+10. **镜头数量硬约束**：5秒1~2镜、10秒3镜、15秒3~4镜、20秒4~5镜、30秒6~8镜
+11. **时间轴硬约束**：第一个clip从0秒开始，最后一个clip结束于total_duration秒，相邻clip首尾相接
+12. 必须严格按<marketing_purpose><target_audience><persona><viral_structure><language><industry>指定的参数写文案和分镜
+13. 镜头间动作衔接要自然，画面描述要具体到能直接拍摄/生成"""
+
+V4_STORYBOARD_USER = """<marketing_purpose>{marketing_purpose}</marketing_purpose>
+<industry>{industry}</industry>
+<image_analysis>
+{image_summary}
+</image_analysis>
+<user_parameters>
+<theme_hint>{theme_hint}</theme_hint>
+<duration>{duration}秒</duration>
+<aspect_ratio>{aspect_ratio}</aspect_ratio>
+<tone>{tone}</tone>
+<target_audience>{target_audience}</target_audience>
+<persona>{persona_hint}</persona>
+<viral_structure>{viral_structure_hint}</viral_structure>
+<language>{language_hint}</language>
+<extra_requirements>{extra_requirements}</extra_requirements>
+</user_parameters>
+{video_style_section}
+请严格按 XML 结构输出分镜脚本。"""
+
+V4_STORYBOARD_EXAMPLE = """<script>
+  <copy_display_markdown><![CDATA[# 在御众堂，把松弛的自己一点点找回来
+产后妈妈最懂那种力不从心，推开门，暖光和一杯热茶先接住了你……]]></copy_display_markdown>
+  <clips>
+    <clip index="1">
+      <time_range>0-3秒</time_range>
+      <speaker>店主</speaker>
+      <voiceover>生完娃，是不是连照镜子的勇气都没了？</voiceover>
+      <visual>【中近景】【缓推】【妈妈疲惫看向镜子】【暖光店内环境】【柔和暖光】</visual>
+      <reference_image_index>0</reference_image_index>
+    </clip>
+    <clip index="2">
+      <time_range>3-6秒</time_range>
+      <speaker>顾客</speaker>
+      <voiceover>是啊，怎么回事？</voiceover>
+      <visual>【近景】【固定】【顾客表情惊讶】【店内休息区】【暖色调】</visual>
+      <reference_image_index>0</reference_image_index>
+    </clip>
+  </clips>
+  <voiceover_script>[speaker:店主]生完娃，是不是连照镜子的勇气都没了？[speaker:顾客]是啊，怎么回事？</voiceover_script>
+  <theme>产后妈妈走进御众堂重拾状态</theme>
+  <negative>模糊、畸变、夸大疗效、绝对化用语</negative>
+</script>"""
+
+
+def upgrade() -> None:
+    conn = op.get_bind()
+
+    # 1. 查询 storyboard 当前最大 version
+    result = conn.execute(
+        text("SELECT MAX(version) FROM viral_video_prompt_templates WHERE prompt_type = :pt"),
+        {"pt": "storyboard"},
+    )
+    max_version = result.scalar() or 0
+    new_version = max_version + 1
+
+    # 2. 旧版 storyboard 模板 is_active 设为 false
+    conn.execute(
+        text("UPDATE viral_video_prompt_templates SET is_active = false WHERE prompt_type = :pt"),
+        {"pt": "storyboard"},
+    )
+
+    # 3. 防御性插入新版 storyboard 模板
+    existing = conn.execute(
+        text("SELECT id FROM viral_video_prompt_templates " "WHERE prompt_type = :pt AND version = :ver"),
+        {"pt": "storyboard", "ver": new_version},
+    ).fetchone()
+
+    if not existing:
+        conn.execute(
+            text(
+                "INSERT INTO viral_video_prompt_templates "
+                "(name, prompt_type, version, system_prompt, user_prompt_template, example_output, is_active) "
+                "VALUES (:name, :pt, :ver, :sys, :usr, :ex, :active)"
+            ),
+            {
+                "name": "编导分镜v4-多角色对话版",
+                "pt": "storyboard",
+                "ver": new_version,
+                "sys": V4_STORYBOARD_SYSTEM,
+                "usr": V4_STORYBOARD_USER,
+                "ex": V4_STORYBOARD_EXAMPLE,
+                "active": True,
+            },
+        )
+
+    # 4. 检查 image_analysis 是否有 active 模板，没有则插入保底
+    ia_active = conn.execute(
+        text("SELECT COUNT(*) FROM viral_video_prompt_templates WHERE prompt_type = :pt AND is_active = true"),
+        {"pt": "image_analysis"},
+    ).scalar()
+
+    if ia_active == 0:
+        # 插入保底 image_analysis 模板（简化版）
+        ia_max = (
+            conn.execute(
+                text("SELECT MAX(version) FROM viral_video_prompt_templates WHERE prompt_type = :pt"),
+                {"pt": "image_analysis"},
+            ).scalar()
+            or 0
+        )
+        conn.execute(
+            text(
+                "INSERT INTO viral_video_prompt_templates "
+                "(name, prompt_type, version, system_prompt, user_prompt_template, example_output, is_active) "
+                "VALUES (:name, :pt, :ver, :sys, :usr, :ex, :active)"
+            ),
+            {
+                "name": "图片分析保底版",
+                "pt": "image_analysis",
+                "ver": ia_max + 1,
+                "sys": "你是一名擅长观察和写作的品牌内容编导。分析图片并输出JSON。",
+                "usr": "请分析这张图片。图片地址：{image_url}",
+                "ex": '{"images": [{"type": "store", "name": "门店", "summary_markdown": "描述"}]}',
+                "active": True,
+            },
+        )
+
+    # 5. 检查 review 是否有 active 模板
+    review_active = conn.execute(
+        text("SELECT COUNT(*) FROM viral_video_prompt_templates WHERE prompt_type = :pt AND is_active = true"),
+        {"pt": "review"},
+    ).scalar()
+
+    if review_active == 0:
+        rv_max = (
+            conn.execute(
+                text("SELECT MAX(version) FROM viral_video_prompt_templates WHERE prompt_type = :pt"),
+                {"pt": "review"},
+            ).scalar()
+            or 0
+        )
+        conn.execute(
+            text(
+                "INSERT INTO viral_video_prompt_templates "
+                "(name, prompt_type, version, system_prompt, user_prompt_template, example_output, is_active) "
+                "VALUES (:name, :pt, :ver, :sys, :usr, :ex, :active)"
+            ),
+            {
+                "name": "文案审核保底版",
+                "pt": "review",
+                "ver": rv_max + 1,
+                "sys": "你是短视频广告合规审核专家。审核文案输出XML。",
+                "usr": "请审核：{fusion_text}",
+                "ex": "<review><passed>true</passed></review>",
+                "active": True,
+            },
+        )
+
+
+def downgrade() -> None:
+    # 恢复旧版 storyboard 为 active
+    conn = op.get_bind()
+    conn.execute(
+        text("UPDATE viral_video_prompt_templates SET is_active = true WHERE prompt_type = :pt AND version < :ver"),
+        {"pt": "storyboard", "ver": 4},
+    )
+    # 删除新版
+    conn.execute(
+        text("DELETE FROM viral_video_prompt_templates WHERE prompt_type = :pt AND version >= :ver"),
+        {"pt": "storyboard", "ver": 4},
+    )

apps/api/app/api/router.py

@@ -1,3 +1,4 @@
+from app.api.routes.admin.ditto_emotion import router as admin_ditto_emotion_router
 from app.api.routes.ai import router as ai_router
 from app.api.routes.ai_avatar_render import router as ai_avatar_render_router
 from app.api.routes.asset_diagnosis import router as asset_diagnosis_router
@@ -242,3 +243,6 @@ api_router.include_router(
     tags=["GPU Worker"],
 )
 api_router.include_router(viral_video_router, prefix="/viral-video", tags=["爆款视频"])
+
+# #2246：后台 Ditto 表情配置（router 自带 /admin/ditto-emotion 前缀）
+api_router.include_router(admin_ditto_emotion_router)

apps/api/app/api/routes/admin/__init__.py

@@ -0,0 +1 @@
+"""后台管理路由（#2246 起）."""

apps/api/app/api/routes/admin/ditto_emotion.py

@@ -0,0 +1,203 @@
+"""Ditto 数字人表情后台配置 — #2246.
+
+路由前缀 /api/v1/admin/ditto-emotion，全部使用 _verify_internal_api_key 鉴权
+（X-API-Key header）。仅开放 6 项白名单配置：
+- ditto_emotion_enabled / ditto_emotion_model / ditto_emotion_temperature
+- ditto_emotion_prompt / ditto_blend_frames / doubao_api_key
+注意：doubao_api_key 属于密钥，GET 只返回脱敏占位，绝不返回明文。
+ditto_emotion_model 为自由文本（支持火山方舟 ep- 接入点 ID），
+推荐选项 model_options 从 system_config 的 ditto_emotion_model_options（JSON 数组）读取。
+"""
+
+from __future__ import annotations
+
+import json
+import re
+from typing import Any
+
+from fastapi import APIRouter, Depends
+from pydantic import BaseModel
+
+from packages.application.system_config_service import get_system_config_service
+from packages.config import get_api_settings
+from packages.domain.system_setting import (
+    SETTING_TYPE_BOOL,
+    SETTING_TYPE_FLOAT,
+    SETTING_TYPE_INT,
+    SETTING_TYPE_STRING,
+)
+
+from ..auth import _verify_internal_api_key
+
+router = APIRouter(
+    prefix="/admin/ditto-emotion",
+    tags=["Admin"],
+    dependencies=[Depends(_verify_internal_api_key)],
+)
+
+# key → (类型, 分类)
+_WHITELIST: dict[str, str] = {
+    "ditto_emotion_enabled": SETTING_TYPE_BOOL,
+    "ditto_emotion_model": SETTING_TYPE_STRING,
+    "ditto_emotion_temperature": SETTING_TYPE_FLOAT,
+    "ditto_emotion_prompt": SETTING_TYPE_STRING,
+    "ditto_blend_frames": SETTING_TYPE_INT,
+    "doubao_api_key": SETTING_TYPE_STRING,
+}
+
+_DESCRIPTIONS: dict[str, str] = {
+    "ditto_emotion_enabled": "LLM 情绪分析开关。关闭时回退到原有关键词匹配模式，不影响正常出片。",
+    "ditto_emotion_model": "用于分析文案情绪的大模型。",
+    "ditto_emotion_temperature": "模型温度，0-1，越低越稳定保守。",
+    "ditto_emotion_prompt": "情绪分析提示词，核心调优入口，必须包含 {文案} 占位符。",
+    "ditto_blend_frames": "表情切换过渡帧数（6-30），越大越柔和。",
+    "doubao_api_key": "豆包/火山方舟 API Key（用于情绪分析LLM调用）",
+}
+
+# 密钥类配置：GET 脱敏、PUT 响应不回显明文
+_SECRET_KEYS = {"doubao_api_key"}
+_SECRET_PLACEHOLDER = "***已配置***"
+
+
+def _settings():
+    return get_api_settings()
+
+
+def _default_value(key: str) -> Any:
+    return getattr(_settings(), key)
+
+
+def _model_options_from_db(service) -> list[str]:
+    """模型推荐选项：从 system_config 读 ditto_emotion_model_options（JSON 数组字符串）。
+
+    未配置 / 非法 JSON / 非数组 JSON 均返回空数组；数组元素统一转字符串。
+    """
+    raw = service.get_config("ditto_emotion_model_options", "") or ""
+    if not str(raw).strip():
+        return []
+    try:
+        data = json.loads(raw)
+    except (TypeError, ValueError):
+        return []
+    if not isinstance(data, list):
+        return []
+    return [str(x) for x in data if isinstance(x, (str, int, float))]
+
+
+def _build_config_item(key: str) -> dict[str, Any]:
+    service = get_system_config_service()
+    if key in _SECRET_KEYS:
+        # 密钥不回明文、不暴露 env 默认值：DB 或 env 任一侧已配置即视为已配置
+        db_value = service.get_config(key, None)
+        configured = bool((db_value or "").strip()) or bool((getattr(_settings(), key, "") or "").strip())
+        return {
+            "key": key,
+            "type": _WHITELIST[key],
+            "description": _DESCRIPTIONS.get(key, ""),
+            "default": "",
+            "value": _SECRET_PLACEHOLDER if configured else "",
+        }
+    item: dict[str, Any] = {
+        "key": key,
+        "type": _WHITELIST[key],
+        "description": _DESCRIPTIONS.get(key, ""),
+        "default": _default_value(key),
+    }
+    item["value"] = service.get_config(key, _default_value(key))
+    if key == "ditto_emotion_model":
+        item["model_options"] = _model_options_from_db(service)
+    return item
+
+
+class ConfigUpdatePayload(BaseModel):
+    configs: dict[str, Any]
+
+
+class TestPayload(BaseModel):
+    test_text: str
+
+
+def _validate_value(key: str, value: Any) -> Any:
+    st = _WHITELIST[key]
+    if st == SETTING_TYPE_BOOL:
+        if not isinstance(value, bool):
+            raise ValueError(f"{key} 必须是布尔值")
+    elif st == SETTING_TYPE_INT:
+        if isinstance(value, bool) or not isinstance(value, int):
+            raise ValueError(f"{key} 必须是整数")
+        if not 6 <= value <= 30:
+            raise ValueError(f"{key} 必须在 6-30 之间")
+    elif st == SETTING_TYPE_FLOAT:
+        if isinstance(value, bool):
+            raise ValueError(f"{key} 必须是数字")
+        try:
+            value = float(value)
+        except (TypeError, ValueError) as exc:
+            raise ValueError(f"{key} 必须是数字") from exc
+        if not 0.0 <= value <= 1.0:
+            raise ValueError(f"{key} 必须在 0-1 之间")
+    elif st == SETTING_TYPE_STRING:
+        if not isinstance(value, str):
+            raise ValueError(f"{key} 必须是字符串")
+        if key == "ditto_emotion_prompt" and value.strip() and "{文案}" not in value:
+            raise ValueError("提示词必须包含 {文案} 占位符")
+        if key == "ditto_emotion_model":
+            if not value.strip():
+                raise ValueError(f"{key} 不能为空字符串")
+            if len(value) > 200:
+                raise ValueError(f"{key} 长度不能超过 200 字符")
+            if not re.fullmatch(r"[A-Za-z0-9_-]+", value):
+                raise ValueError(f"{key} 只能包含字母、数字、连字符和下划线")
+        if key == "doubao_api_key" and not value.strip():
+            raise ValueError(f"{key} 不能为空字符串")
+    return value
+
+
+@router.get("/config")
+def get_config() -> dict[str, Any]:
+    return {"configs": [_build_config_item(k) for k in _WHITELIST]}
+
+
+@router.put("/config")
+def update_config(
+    payload: ConfigUpdatePayload,
+    x_api_key: str = Depends(_verify_internal_api_key),
+) -> dict[str, Any]:
+    configs = payload.configs
+    illegal = [k for k in configs if k not in _WHITELIST]
+    if illegal:
+        return {
+            "ok": False,
+            "error": f"不允许修改的配置项：{', '.join(illegal)}",
+        }
+    service = get_system_config_service()
+    updated: dict[str, Any] = {}
+    for key, raw in configs.items():
+        try:
+            value = _validate_value(key, raw)
+        except ValueError as exc:
+            return {"ok": False, "error": str(exc)}
+        service.set_config(
+            key,
+            value,
+            setting_type=_WHITELIST[key],
+            updated_by=str(x_api_key)[:8] if x_api_key and x_api_key is not True else None,
+        )
+        updated[key] = _SECRET_PLACEHOLDER if key in _SECRET_KEYS else value
+    return {"ok": True, "updated": updated}
+
+
+@router.post("/config/test")
+def test_config(payload: TestPayload) -> dict[str, Any]:
+    text = (payload.test_text or "").strip()
+    if not text:
+        return {"ok": False, "error": "test_text 不能为空"}
+    from packages.application.ditto_emotion_service import get_ditto_emotion_service
+
+    service = get_ditto_emotion_service()
+    segments = service.analyze(text)
+    return {
+        "ok": True,
+        "enabled": service.enabled,
+        "segments": [s.to_dict() for s in segments],
+    }

apps/api/app/api/routes/points.py

@@ -89,6 +89,7 @@ def get_balance(
         is_member=_is_member(current_user),
         member_type=_member_type(current_user),
         member_expires_at=getattr(current_user.user, "member_expires_at", None),
+        credits_enabled=_credits_enabled(),
     )
 
 

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

@@ -49,7 +49,17 @@ from packages.adapters.sqlalchemy_impl.viral_video_repository import (
     SQLAlchemyViralVideoJobRepository,
     SQLAlchemyViralVideoStyleTemplateRepository,
 )
-from packages.domain.points_rules import list_viral_video_models
+from packages.application.viral_video.constants import (
+    API_FALLBACK_HARD_CONSTRAINTS,
+    API_FALLBACK_NEGATIVE_PROMPTS,
+)
+from packages.domain.points_rules import (
+    calculate_viral_video_credits,
+    calculate_viral_video_credits_with_breakdown,
+    list_viral_video_models,
+    resolve_video_dimensions,
+)
+from packages.domain.points_service import PointsService
 from packages.domain.viral_video import ViralVideoStatus
 from packages.shared.dashscope_client import get_dashscope_client
 
@@ -105,8 +115,8 @@ def _build_copy_result(job) -> dict | None:
         "overview": {"theme": title, "total_duration": getattr(job, "duration", 15), "aspect_ratio": ratio},
         "scene_and_lighting": "",
         "shots": shots,
-        "hard_constraints": ["无字幕", "无水印", "人物一致性"],
-        "negative_prompts": ["字幕", "水印", "错误文字", "五官崩坏"],
+        "hard_constraints": list(API_FALLBACK_HARD_CONSTRAINTS),
+        "negative_prompts": list(API_FALLBACK_NEGATIVE_PROMPTS),
         "voiceover_script": copy_text,
         "final_copy": copy_text,
         "suggested_copy": copy_text,
@@ -142,6 +152,7 @@ def _to_response(job) -> ViralVideoJobResponse:
         copy_result=_build_copy_result(job),
         voice_id=getattr(job, "voice_id", "") or "",
         voice_source=getattr(job, "voice_source", "") or "",
+        language=getattr(job, "language", "zh-CN") or "zh-CN",
         video_ratio=getattr(job, "video_ratio", "9:16") or "9:16",
         video_model=getattr(job, "video_model", "") or "",
         intent_result=job.intent_result,
@@ -167,6 +178,156 @@ def _get_style_repo(session: Session) -> SQLAlchemyViralVideoStyleTemplateReposi
     return SQLAlchemyViralVideoStyleTemplateRepository(session)
 
 
+def _settle_viral_video_credits(
+    user_id: int,
+    job,
+    session: Session,
+    *,
+    action: str,
+    old_duration: int | None = None,
+    old_resolution: str | None = None,
+    old_ratio: str | None = None,
+    old_model: str | None = None,
+    new_duration: int | None = None,
+    new_resolution: str | None = None,
+    new_ratio: str | None = None,
+    new_model: str | None = None,
+) -> dict:
+    """爆款视频积分结算统一入口（confirm-copy / retry 共用）。
+
+    action 取值：
+    - "deduct_first"   首次扣费（confirm 时还没付过）
+    - "recompute"      参数变更重算：退旧扣新（confirm_copy 时传 old/new）
+    - "delta"          参数变更差额结算：多退少补（retry 时传 old/new）
+    返回 credits_info dict，包含 old_est / new_est / action / diff 等。
+    积分不足时 raise HTTPException(402)，调用方无需再捕获。
+    """
+    from app.config import settings as _settings
+
+    info: dict = {"action": action, "old_est": 0.0, "new_est": 0.0, "diff": 0.0, "settled": False}
+    if not _settings.points_enabled:
+        return info
+
+    already_paid = (float(getattr(job, "credits_prepaid", 0) or 0) > 0) or (
+        float(getattr(job, "credits_cost", 0) or 0) > 0
+    )
+
+    if action == "deduct_first":
+        if already_paid:
+            return info
+        w, h = resolve_video_dimensions(
+            getattr(job, "video_resolution", "720p") or "720p", job.video_ratio or "9:16"
+        )
+        est = calculate_viral_video_credits(int(job.duration or 15), w, h, job.video_model or "seedance-2.5")
+        info["new_est"] = est
+        svc = PointsService()
+        res = svc.deduct_viral_video(user_id, est, job.id, session)
+        if not res.get("success"):
+            balance = res.get("balance", 0)
+            raise HTTPException(
+                status_code=402,
+                detail={
+                    "code": "INSUFFICIENT_POINTS",
+                    "message": f"积分不足，需要 {est} 积分，当前余额 {balance}",
+                    "required": est,
+                    "balance": balance,
+                },
+            )
+        job.credits_prepaid = est
+        job.credits_transaction_id = res.get("transaction_id", "") or ""
+        info["settled"] = True
+        info["diff"] = est
+        return info
+
+    if action == "recompute":
+        if not already_paid:
+            # 等价于首次扣费
+            return _settle_viral_video_credits(user_id, job, session, action="deduct_first")
+        old_w, old_h = resolve_video_dimensions(old_resolution or "720p", old_ratio or "9:16")
+        old_est = calculate_viral_video_credits(
+            int(old_duration or 15), old_w, old_h, old_model or "seedance-2.5"
+        )
+        new_w, new_h = resolve_video_dimensions(
+            getattr(job, "video_resolution", "720p") or "720p", job.video_ratio or "9:16"
+        )
+        new_est = calculate_viral_video_credits(
+            int(job.duration or 15), new_w, new_h, job.video_model or "seedance-2.5"
+        )
+        info.update({"old_est": old_est, "new_est": new_est, "diff": new_est - old_est})
+        svc = PointsService()
+        if getattr(job, "credits_transaction_id", None):
+            svc.refund_points(
+                user_id=user_id,
+                amount=float(job.credits_prepaid),
+                source="viral_video",
+                db=session,
+                ref_id=job.credits_transaction_id,
+                description="confirm-copy 参数变更退还旧预扣",
+            )
+        if new_est > 0:
+            res = svc.deduct_viral_video(user_id, new_est, job.id, session)
+            if not res.get("success"):
+                balance = res.get("balance", 0)
+                raise HTTPException(
+                    status_code=402,
+                    detail={
+                        "code": "INSUFFICIENT_POINTS",
+                        "message": f"积分不足，需要 {new_est} 积分，当前余额 {balance}",
+                        "required": new_est,
+                        "balance": balance,
+                    },
+                )
+            job.credits_prepaid = new_est
+            job.credits_transaction_id = res.get("transaction_id", "") or ""
+        info["settled"] = True
+        return info
+
+    if action == "delta":
+        eff_dur = int(new_duration if new_duration is not None else old_duration or 15)
+        eff_res = new_resolution if new_resolution is not None else (old_resolution or "720p")
+        eff_ratio = new_ratio if new_ratio is not None else (old_ratio or "9:16")
+        eff_model = new_model if new_model is not None else (old_model or "seedance-2.5")
+        w, h = resolve_video_dimensions(eff_res, eff_ratio)
+        new_est, _ = calculate_viral_video_credits_with_breakdown(eff_dur, w, h, eff_model)
+        old_prepaid = float(getattr(job, "credits_prepaid", 0) or 0)
+        diff = round(new_est - old_prepaid, 2)
+        info.update({"old_est": old_prepaid, "new_est": new_est, "diff": diff})
+        if abs(diff) < 0.01:
+            return info
+        svc = PointsService()
+        if diff > 0:
+            res = svc.deduct_viral_video(user_id, diff, job.id, session)
+            if not res.get("success"):
+                balance = res.get("balance", 0)
+                raise HTTPException(
+                    status_code=402,
+                    detail={
+                        "code": "INSUFFICIENT_POINTS",
+                        "message": f"重试参数变更后需补扣 {diff} 积分，余额不足（当前 {balance}，需 {new_est}）",
+                        "required": new_est,
+                        "balance": balance,
+                        "delta": diff,
+                    },
+                )
+            job.credits_prepaid = round(old_prepaid + diff, 2)
+        else:
+            refund = round(-diff, 2)
+            txn_id = getattr(job, "credits_transaction_id", "") or ""
+            svc.refund_points(
+                user_id=user_id,
+                amount=refund,
+                source="viral_video",
+                db=session,
+                ref_id=txn_id or job.id,
+                description="爆款视频重试参数变更退费",
+            )
+            job.credits_prepaid = round(old_prepaid - refund, 2)
+        info["settled"] = True
+        return info
+
+    raise ValueError(f"unknown action: {action}")
+
+
 # ── Endpoints ────────────────────────────────────────────────────────────
 
 
@@ -243,6 +404,7 @@ def analyze_images(
         style_strength=request.style_strength or "medium",
         voice_id=request.voice_id or "",
         voice_source=request.voice_source or "",
+        language=getattr(request, "language", "zh-CN") or "zh-CN",
         video_ratio=request.video_ratio or "9:16",
         video_model=request.video_model or "",
         video_resolution=getattr(request, "video_resolution", "720p") or "720p",
@@ -314,6 +476,7 @@ def generate_copy(
         job.style_guide = request.style_guide
     job.voice_id = request.voice_id or job.voice_id
     job.voice_source = request.voice_source or job.voice_source
+    job.language = getattr(request, "language", "") or job.language or "zh-CN"
     job.video_ratio = request.video_ratio or job.video_ratio or "9:16"
     job.video_model = request.video_model or job.video_model or ""
     job.video_resolution = getattr(request, "video_resolution", "") or job.video_resolution or "720p"
@@ -352,40 +515,55 @@ def confirm_copy(
     if not isinstance(job.copy_result, dict) or not job.copy_result:
         raise HTTPException(status_code=409, detail="文案数据缺失，请先点击「生成文案」")
 
-    # 积分预扣（已扣过/重试任务跳过）
-    from app.config import settings as _settings
+    # Bug1 fix: 用户 confirm 时允许修改 video_model/video_resolution/video_ratio/duration
+    old_duration = int(getattr(job, "duration", 15) or 15)
+    old_resolution = getattr(job, "video_resolution", "720p") or "720p"
+    old_ratio = getattr(job, "video_ratio", "9:16") or "9:16"
+    old_model = getattr(job, "video_model", None) or "seedance-2.5"
 
-    if _settings.points_enabled:
-        already_paid = (float(getattr(job, "credits_prepaid", 0) or 0) > 0) or (
-            float(getattr(job, "credits_cost", 0) or 0) > 0
+    if request.duration is not None:
+        job.duration = max(5, min(30, int(request.duration)))
+    if request.video_resolution is not None:
+        job.video_resolution = request.video_resolution
+    if request.video_ratio is not None:
+        job.video_ratio = request.video_ratio
+    if request.video_model is not None:
+        job.video_model = request.video_model
+    if request.voice_id is not None:
+        job.voice_id = request.voice_id
+    if request.voice_source is not None:
+        job.voice_source = request.voice_source
+    if getattr(request, "language", None) is not None:
+        job.language = request.language
+
+    param_changed = (
+        (request.duration is not None and int(request.duration) != old_duration)
+        or (request.video_resolution is not None and request.video_resolution != old_resolution)
+        or (request.video_ratio is not None and request.video_ratio != old_ratio)
+        or (request.video_model is not None and request.video_model != old_model)
+    )
+
+    # 积分预扣（PR2：统一走 _settle_viral_video_credits）
+    if param_changed:
+        credits_info = _settle_viral_video_credits(
+            authenticated_user.user.id,
+            job,
+            session,
+            action="recompute",
+            old_duration=old_duration,
+            old_resolution=old_resolution,
+            old_ratio=old_ratio,
+            old_model=old_model,
         )
-        if not already_paid:
-            from packages.domain.points_rules import calculate_viral_video_credits, resolve_video_dimensions
-            from packages.domain.points_service import PointsService
-
-            w, h = resolve_video_dimensions(
-                getattr(job, "video_resolution", "720p") or "720p",
-                job.video_ratio or "9:16",
-            )
-            est_credits = calculate_viral_video_credits(
-                int(job.duration or 15), w, h, job.video_model or "seedance-2.5"
-            )
-            svc = PointsService()
-            res = svc.deduct_viral_video(authenticated_user.user.id, est_credits, job.id, session)
-            if not res.get("success"):
-                balance = res.get("balance", 0)
-                raise HTTPException(
-                    status_code=402,
-                    detail={
-                        "code": "INSUFFICIENT_POINTS",
-                        "message": f"积分不足，需要 {est_credits} 积分，当前余额 {balance}",
-                        "required": est_credits,
-                        "balance": balance,
-                    },
-                )
-            job.credits_prepaid = est_credits
-            job.credits_transaction_id = res.get("transaction_id", "") or ""
-            repo.update(job)
+        logger.info(
+            "[爆款视频][confirm-copy] 参数变更，积分重算: old=%s new=%s job_id=%s",
+            credits_info.get("old_est"),
+            credits_info.get("new_est"),
+            job.id,
+        )
+    else:
+        _settle_viral_video_credits(authenticated_user.user.id, job, session, action="deduct_first")
+        repo.update(job)
 
     job.resume_from_copy_generated(edited_copy=request.edited_copy or None)
     repo.update(job)
@@ -395,6 +573,7 @@ def confirm_copy(
         logger.info("[爆款视频][阶段3] confirm-copy 入队: job_id=%s", job.id)
     except Exception as e:
         logger.error("[爆款视频][阶段3] confirm-copy 入队失败: %s", e, exc_info=True)
+        # TODO(PR2-followup): 入队失败应返回 HTTP 503 而非 mark_failed + 200，避免前端收到"成功"但 job 已 failed 的不一致状态
         job.mark_failed(f"任务入队失败: {e}")
         repo.update(job)
 
@@ -412,11 +591,6 @@ def estimate_credits(
     profit_multiplier / model_price / width / height / fps），便于前端展示计费明细。
     同时兼容前端传 model 或 video_model、resolution 或 video_resolution、ratio 或 video_ratio。
     """
-    from packages.domain.points_rules import (
-        calculate_viral_video_credits_with_breakdown,
-        resolve_video_dimensions,
-    )
-
     model = (request.model or "").strip() or "seedance-2.5"
     resolution = (request.resolution or "").strip() or "720p"
     ratio = (request.ratio or "").strip() or "9:16"
@@ -548,25 +722,7 @@ def retry_viral_video_job(
         or (new_model is not None and new_model != old_model)
     )
 
-    from app.config import settings as _settings
-
-    need_points_settle = False
-    new_est = 0.0
-    if _settings.points_enabled and param_changed:
-        from packages.domain.points_rules import (
-            calculate_viral_video_credits_with_breakdown,
-            resolve_video_dimensions,
-        )
-
-        eff_dur = int(new_duration if new_duration is not None else old_duration)
-        eff_res = new_resolution if new_resolution is not None else old_resolution
-        eff_ratio = new_ratio if new_ratio is not None else old_ratio
-        eff_model = new_model if new_model is not None else (old_model or "seedance-2.5")
-        w, h = resolve_video_dimensions(eff_res, eff_ratio)
-        new_est, _ = calculate_viral_video_credits_with_breakdown(eff_dur, w, h, eff_model or "seedance-2.5")
-        need_points_settle = True
-
-    # 写入新参数（即使不开 points 也要允许用户重试时改参数）
+    # 写入新参数（即使不开 points 也要允许用户重试时改参数；delta 结算依赖这些新值）
     if new_duration is not None:
         job.duration = max(5, min(30, int(new_duration)))
     if new_resolution is not None:
@@ -576,57 +732,31 @@ def retry_viral_video_job(
     if new_model is not None:
         job.video_model = new_model
 
-    if need_points_settle:
-        from packages.domain.points_service import PointsService
+    # PR2：统一走 _settle_viral_video_credits（delta 多退少补）
+    if param_changed:
+        credits_info = _settle_viral_video_credits(
+            authenticated_user.user.id,
+            job,
+            session,
+            action="delta",
+            old_duration=old_duration,
+            old_resolution=old_resolution,
+            old_ratio=old_ratio,
+            old_model=old_model,
+            new_duration=new_duration,
+            new_resolution=new_resolution,
+            new_ratio=new_ratio,
+            new_model=new_model,
+        )
+        if credits_info.get("settled"):
+            logger.info(
+                "[爆款视频][retry] 积分差额结算 job_id=%s diff=%.2f new_prepaid=%.2f",
+                job.id,
+                credits_info["diff"],
+                job.credits_prepaid,
+            )
 
-        old_prepaid = float(getattr(job, "credits_prepaid", 0) or 0)
-        svc = PointsService()
-        diff = round(new_est - old_prepaid, 2)
-        if abs(diff) >= 0.01:
-            if diff > 0:
-                # 新预扣更多：补扣差额
-                res = svc.deduct_viral_video(authenticated_user.user.id, diff, job.id, session)
-                if not res.get("success"):
-                    balance = res.get("balance", 0)
-                    raise HTTPException(
-                        status_code=402,
-                        detail={
-                            "code": "INSUFFICIENT_POINTS",
-                            "message": f"重试参数变更后需补扣 {diff} 积分，余额不足（当前 {balance}，需 {new_est}）",
-                            "required": new_est,
-                            "balance": balance,
-                            "delta": diff,
-                        },
-                    )
-                job.credits_prepaid = round(old_prepaid + diff, 2)
-                logger.info(
-                    "[爆款视频][retry] 补扣差额 job_id=%s diff=%.2f new_prepaid=%.2f",
-                    job.id,
-                    diff,
-                    job.credits_prepaid,
-                )
-            else:
-                # 新预扣更少：退还差额
-                refund = round(-diff, 2)
-                txn_id = getattr(job, "credits_transaction_id", "") or ""
-                svc.refund_points(
-                    user_id=authenticated_user.user.id,
-                    amount=refund,
-                    source="viral_video",
-                    db=session,
-                    ref_id=txn_id or job.id,
-                    description="爆款视频重试参数变更退费",
-                )
-                job.credits_prepaid = round(old_prepaid - refund, 2)
-                logger.info(
-                    "[爆款视频][retry] 退还差额 job_id=%s refund=%.2f new_prepaid=%.2f",
-                    job.id,
-                    refund,
-                    job.credits_prepaid,
-                )
-        # 差额为 0 则不调整
-
-    # 重置状态
+    # 重置状态（从阶段1重新开始，清空上一次产物避免脏数据）
     job.retry_count += 1
     job.status = ViralVideoStatus.PENDING
     job.error_msg = "" if not is_stale_running else "任务执行超时，已重置重试"
@@ -635,11 +765,19 @@ def retry_viral_video_job(
     job.current_stage = ""
     job.phase_message = ""
     job.heartbeat_at = None
+    # 重试时清空上次产物（参考 resume_from_image_analyzed 逻辑）
+    job.intent_result = None
+    job.copy_result = None
+    job.storyboard = None
+    job.generated_copy_text = ""
+    job.result_video_url = ""
+    job.pre_trusted_images = None
+    job.credits_cost = 0.0
     repo.update(job)
 
-    # 重新入队
+    # 重新入队：从图片分析阶段1重新跑
     try:
-        celery_app.send_task("worker.run_viral_video_pipeline", args=[job.id])
+        celery_app.send_task("worker.run_viral_video_analyze", args=[job.id])
         logger.info(
             "[爆款视频] 重试入队: job_id=%s retry_count=%d stale=%s params_changed=%s",
             job.id,
@@ -649,6 +787,7 @@ def retry_viral_video_job(
         )
     except Exception as e:
         logger.error("[爆款视频] 重试入队失败: %s", e, exc_info=True)
+        # TODO(PR2-followup): 入队失败应返回 HTTP 503 而非 mark_failed + 200
         job.mark_failed(f"重试入队失败: {e}")
         repo.update(job)
 
@@ -890,13 +1029,28 @@ async def viral_video_websocket(websocket: WebSocket, job_id: str) -> None:
         job = job_repo.get(job_id)
         if job is not None:
             status_val = job.status.value if hasattr(job.status, "value") else str(job.status)
+            # #P0: 初始快照必须包含前端重连/刷新所需的业务字段，
+            # 结构对齐 worker 推送的 image_analyzed / copy_generated 事件。
+            data: dict = {"status": status_val}
+            ia = getattr(job, "image_analysis", None)
+            if isinstance(ia, dict) and ia:
+                data["image_analysis"] = ia
+            cr = _build_copy_result(job)
+            if isinstance(cr, dict) and cr:
+                data["copy_result"] = cr
+            gct = getattr(job, "generated_copy_text", "") or ""
+            if gct:
+                data["generated_copy_text"] = gct
+            sb = getattr(job, "storyboard", None) or []
+            if sb:
+                data["storyboard"] = sb
             initial = {
                 "type": "viral_video:progress",
                 "job_id": job_id,
                 "stage": _stage_from_status(job),
                 "progress": _estimate_progress(job),
                 "message": _initial_message(job),
-                "data": {"status": status_val},
+                "data": data,
             }
             await websocket.send_json(initial)
             # 已经终态 → 再发一条终态事件后立即关闭，避免占连接

apps/api/app/schemas/points.py

@@ -19,6 +19,7 @@ class PointsBalanceResponse(BaseModel):
     is_member: bool = Field(default=False, description="是否付费会员")
     member_type: Optional[str] = Field(None, description="会员类型: monthly/quarterly/yearly")
     member_expires_at: Optional[datetime] = Field(None, description="会员到期时间")
+    credits_enabled: bool = Field(default=False, description="积分系统是否启用（false=免费放行不扣分）")
 
 
 # ============ 流水 ============

apps/api/app/schemas/viral_video.py

@@ -87,6 +87,7 @@ class CreateViralVideoRequest(BaseModel):
     style_template_id: str = ""
     voice_id: str = ""
     voice_source: str = ""
+    language: str = "zh-CN"
     video_ratio: str = "9:16"
     video_model: str = ""
     video_resolution: str = "720p"
@@ -121,6 +122,7 @@ class AnalyzeImagesRequest(BaseModel):
     video_model: str = ""
     video_resolution: str = "720p"
     duration: int = Field(default=15, ge=5, le=30)
+    language: str = "zh-CN"
 
 
 class GenerateCopyRequest(BaseModel):
@@ -142,6 +144,7 @@ class GenerateCopyRequest(BaseModel):
     style_guide: dict | None = None
     voice_id: str = ""
     voice_source: str = ""
+    language: str = "zh-CN"
     video_ratio: str = "9:16"
     video_model: str = ""
     video_resolution: str = "720p"
@@ -167,6 +170,13 @@ class ConfirmCopyRequest(BaseModel):
     """v1.5+ 阶段3：用户确认/编辑口播后开始渲染（TTS+单次Seedance）。"""
 
     edited_copy: str = Field(default="", description="用户编辑后的口播文案；为空则用 AI 生成的 voiceover_script")
+    video_model: str | None = Field(default=None, description="用户选定的视频生成模型（confirm时可选）")
+    video_resolution: str | None = Field(default=None, description="用户选定的分辨率（confirm时可选）")
+    video_ratio: str | None = Field(default=None, description="用户选定的比例（confirm时可选）")
+    duration: int | None = Field(default=None, ge=5, le=30, description="用户选定的时长秒数（confirm时可选，5~30）")
+    voice_id: str | None = Field(default=None, description="用户选定的音色ID（confirm时可选）")
+    voice_source: str | None = Field(default=None, description="用户选定的音色来源（confirm时可选）")
+    language: str | None = Field(default=None, description="用户选定的语言（confirm时可选）")
 
 
 class ConfirmIntentRequest(BaseModel):
@@ -218,6 +228,7 @@ class ViralVideoJobResponse(BaseModel):
     # 音色/视频参数
     voice_id: str = ""
     voice_source: str = ""
+    language: str = "zh-CN"
     video_ratio: str = "9:16"
     video_model: str = ""
     intent_result: dict | None = None

apps/api/app/services/lipsync_service.py

@@ -223,7 +223,47 @@ class LipsyncService:
         if timings:
             job.sentence_timings = timings
 
-        # 4. 检查是否走 GPU 路径：开关打开 + 有可用 Worker
+        # 4. 检查是否走 Ditto（蚂蚁数字人，#2076）：开关 + 配置完整
+        use_ditto = False
+        if self.settings.use_ditto_lipsync:
+            try:
+                from packages.application.ditto_service import get_ditto_client
+
+                ditto = get_ditto_client()
+                if ditto.is_configured:
+                    use_ditto = True
+                    logger.info("[lipsync] 优先走 Ditto 蚂蚁数字人: job_id=%s", job.id)
+                else:
+                    logger.info(
+                        "[lipsync] Ditto 开关已开但配置不完整（base_url=%s, template=%s），继续判断 GPU: job_id=%s",
+                        bool(ditto.base_url),
+                        bool(ditto.default_video_url),
+                        job.id,
+                    )
+            except Exception as exc:
+                logger.warning("[lipsync] Ditto 初始化失败，继续判断 GPU: job_id=%s err=%s", job.id, exc)
+
+        if use_ditto:
+            try:
+                # Ditto 使用预置人物模板视频，不用用户上传的 video_url；
+                # 但保留用户 video_url 以便失败回退到 GPU/MediaKit。
+                job.status = "processing"
+                job.mediakit_task_id = "ditto:submitted"
+                job.updated_at = datetime.now(UTC)
+                self.db.commit()
+                from app.tasks.lipsync_ditto import lipsync_ditto_process_async
+
+                lipsync_ditto_process_async.apply_async(args=(job.id, job.user_id))
+                logger.info("[lipsync] Ditto 任务已异步派发: job_id=%s", job.id)
+                return
+            except Exception as exc:
+                logger.warning("[lipsync] Ditto 派发失败，回退 GPU/MediaKit: job_id=%s err=%s", job.id, exc)
+                try:
+                    self.db.rollback()
+                except Exception:
+                    pass
+
+        # 5. 检查是否走 GPU 路径：开关打开 + 有可用 Worker
         use_gpu = False
         if self.settings.use_gpu_lipsync:
             try:
@@ -805,6 +845,30 @@ class LipsyncService:
         if job.status in (STATUS_COMPLETED, "failed"):
             return job
 
+        # Ditto 异步路径：mediakit_task_id 以 "ditto:" 开头，由 Celery 任务异步更新
+        # 不做 MediaKit 轮询，只检查是否卡住太久（>10 分钟）则标失败
+        if job.mediakit_task_id and job.mediakit_task_id.startswith("ditto:"):
+            if job.status in ("processing", "submitted"):
+                _now = datetime.now(UTC)
+                _upd = job.updated_at
+                if _upd is not None and _upd.tzinfo is None:
+                    _upd = _upd.replace(tzinfo=UTC)
+                stale_minutes = 10
+                if _upd and (_now - _upd).total_seconds() > stale_minutes * 60:
+                    logger.warning(
+                        "Ditto 异步任务超时（>%d 分钟），标记失败: job_id=%s",
+                        stale_minutes,
+                        job_id,
+                    )
+                    job.status = "failed"
+                    job.error_message = f"Ditto 处理超时（>{stale_minutes} 分钟）"
+                    job.error_code = "DittoTimeout"
+                    job.completed_at = _now
+                    job.updated_at = _now
+                    self.db.commit()
+                    self._refund_lip_sync(job)
+            return job
+
         # GPU 异步路径：mediakit_task_id 以 "gpu:" 开头，由 Celery 任务异步更新
         # 不做 MediaKit 轮询，只检查是否卡住太久（>30 分钟）则标失败
         if job.mediakit_task_id and job.mediakit_task_id.startswith("gpu:"):

apps/api/app/tasks/lipsync_ditto.py

@@ -0,0 +1,552 @@
+"""Ditto 蚂蚁数字人口型异步任务 — #2076.
+
+把 Ditto 同步 HTTP 调用（30-120s）从 API 请求移到 Celery 后台执行：
+1. 加载 LipsyncJob
+2. 确定驱动视频：用户上传的 video_url 优先，无则用 settings.ditto_default_video_url 兜底
+3. 视频时长对齐：若视频 < 音频+2s，用 ffmpeg 循环视频到足够长度后上传临时文件
+4. 调 DittoClient.generate_and_persist
+5. 成功：标记 completed，写入 output_video_url（Ditto 输出自带音频，无需二次混流/超分）
+6. 失败：回退 GPU MuseTalk → 再失败回退 MediaKit
+
+注意：
+- 保留 MuseTalk 代码不动；Ditto 优先，失败按原链路兜底
+- Ditto 默认使用用户上传的视频作为驱动模板，default_video_url 仅作兜底
+- 用户视频短于音频时，ffmpeg stream_loop 循环到音频时长+2s余量
+- 不传 GFPGAN 超分，不需要 ffmpeg 音视频混流（Ditto 输出已带音视频）
+"""
+
+from __future__ import annotations
+
+import logging
+import os
+import subprocess
+import tempfile
+from datetime import UTC, datetime
+from typing import TYPE_CHECKING, Optional
+
+from celery import shared_task
+from sqlalchemy.orm import Session
+
+if TYPE_CHECKING:
+    from packages.adapters.sqlalchemy_impl.models import LipsyncJobModel
+
+logger = logging.getLogger(__name__)
+
+_DITTO_URL_TTL_SECONDS = 7 * 24 * 3600  # Ditto 结果 OSS URL 7 天有效
+_VIDEO_LOOP_MARGIN_SECONDS = 2.0  # 循环视频时比音频多留 2 秒余量
+_MAX_VIDEO_PREPROCESS_SIZE = 200 * 1024 * 1024  # 用户视频最大 200MB
+
+
+def _get_db_session() -> Session:
+    try:
+        from worker_app.db import SessionLocal  # type: ignore
+    except ImportError:
+        from app.db import SessionLocal  # type: ignore
+    return SessionLocal()
+
+
+def _sign_media_url(url: str) -> str:
+    """对自家 OSS URL 签 7 天预签名。"""
+    if not url:
+        return url
+    try:
+        from urllib.parse import urlparse
+
+        from packages.shared.storage import get_shared_storage_service
+
+        storage = get_shared_storage_service()
+        public_base = getattr(storage, "public_url", "")
+        if not isinstance(public_base, str) or not public_base:
+            return url
+        own_host = urlparse(public_base).netloc.lower()
+        host = urlparse(url).netloc.lower()
+        if not own_host or host != own_host:
+            return url
+        return storage.get_download_url(url, expires_seconds=_DITTO_URL_TTL_SECONDS)
+    except Exception:
+        return url
+
+
+def _probe_media_duration(path_or_bytes, *, is_bytes: bool = False) -> float:
+    """用 ffprobe 探测视频/音频时长（秒）；失败返回 0。
+
+    Args:
+        path_or_bytes: 文件路径(str) 或 字节数据(bytes)
+        is_bytes: 传入的是 bytes 还是文件路径
+    """
+    tmp_path = None
+    try:
+        if is_bytes:
+            with tempfile.NamedTemporaryFile(suffix=".bin", delete=False) as tmp:
+                tmp.write(path_or_bytes)
+                tmp_path = tmp.name
+            target = tmp_path
+        else:
+            target = path_or_bytes
+        out = subprocess.check_output(
+            [
+                "ffprobe",
+                "-v",
+                "error",
+                "-show_entries",
+                "format=duration",
+                "-of",
+                "default=noprint_wrappers=1:nokey=1",
+                target,
+            ],
+            stderr=subprocess.STDOUT,
+            timeout=15,
+        )
+        return float(out.decode().strip() or 0)
+    except Exception as exc:
+        logger.warning("[ditto_task] ffprobe 失败: %s", exc)
+        return 0.0
+    finally:
+        if tmp_path and os.path.exists(tmp_path):
+            try:
+                os.unlink(tmp_path)
+            except Exception:
+                pass
+
+
+def _probe_video_duration(video_bytes: bytes) -> float:
+    """用 ffprobe 探测视频时长（秒）；失败返回 0。"""
+    return _probe_media_duration(video_bytes, is_bytes=True)
+
+
+def _prepare_driver_video(
+    *,
+    user_video_url: str,
+    default_video_url: str,
+    audio_duration: float,
+    job_id: str,
+    user_id: str,
+) -> tuple[str, bool]:
+    """准备传给 Ditto 的驱动视频 URL。
+
+    逻辑：
+    1. 优先使用用户上传的视频（user_video_url），无则用 default_video_url 兜底
+    2. 下载视频，探测时长
+    3. 若视频时长 >= 音频时长+2s 余量：直接用原 URL（签名后）
+    4. 若视频时长 < 音频时长+2s：ffmpeg stream_loop 循环到目标时长，上传临时 OSS，返回临时 URL
+    5. 任何异常：回退到 default_video_url（兜底）
+
+    Returns:
+        (video_url_for_ditto, is_temporary) — is_temporary=True 表示 URL 是本次临时生成的
+    """
+    from packages.shared.storage import get_shared_storage_service
+    from packages.shared.url_security import safe_download_bytes
+
+    storage = get_shared_storage_service()
+
+    # 1. 选择源 URL（用户视频优先）
+    source_url = user_video_url or default_video_url
+    source_label = "user" if user_video_url else "default"
+    if not source_url:
+        raise RuntimeError("无可用驱动视频（用户视频和默认模板都为空）")
+
+    # 2. 下载视频
+    try:
+        # 用户视频需要签名才能下载
+        signed_source = _sign_media_url(source_url) if user_video_url else source_url
+        video_bytes = safe_download_bytes(
+            signed_source,
+            purpose="ditto-driver-video",
+            max_size=_MAX_VIDEO_PREPROCESS_SIZE,
+            allowed_mime_types=("video/mp4", "video/quicktime", "video/x-msvideo", "video/webm"),
+            timeout=60,
+        )
+    except Exception as dl_exc:
+        logger.warning(
+            "[ditto_task] 下载驱动视频失败（%s），回退默认模板: job=%s err=%s",
+            source_label,
+            job_id,
+            dl_exc,
+        )
+        if user_video_url and default_video_url:
+            return default_video_url, False
+        raise
+
+    # 3. 探测视频时长
+    video_duration = _probe_media_duration(video_bytes, is_bytes=True)
+    target_duration = audio_duration + _VIDEO_LOOP_MARGIN_SECONDS
+
+    # 4. 视频足够长：直接用签名后的原 URL
+    if video_duration >= target_duration:
+        logger.info(
+            "[ditto_task] 驱动视频足够长（%.2fs >= %.2fs），直接使用: job=%s source=%s",
+            video_duration,
+            target_duration,
+            job_id,
+            source_label,
+        )
+        return _sign_media_url(source_url) if user_video_url else source_url, False
+
+    # 5. 视频不够长：ffmpeg 循环到目标时长
+    logger.info(
+        "[ditto_task] 驱动视频不够长（%.2fs < %.2fs），ffmpeg 循环延长: job=%s",
+        video_duration,
+        target_duration,
+        job_id,
+    )
+    looped_path = None
+    try:
+        # 写临时文件
+        with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as src_tmp:
+            src_tmp.write(video_bytes)
+            src_path = src_tmp.name
+        looped_fd, looped_path = tempfile.mkstemp(suffix=".mp4")
+        os.close(looped_fd)
+
+        # ffmpeg: stream_loop -1 循环输入，-t 截到目标时长
+        # 使用 -c copy 快速复制流（不重编码），速度快无画质损失
+        cmd = [
+            "ffmpeg",
+            "-y",
+            "-stream_loop",
+            "-1",
+            "-i",
+            src_path,
+            "-t",
+            str(target_duration),
+            "-c",
+            "copy",
+            "-movflags",
+            "+faststart",
+            looped_path,
+        ]
+        try:
+            subprocess.run(cmd, check=True, capture_output=True, timeout=60)
+        except subprocess.CalledProcessError:
+            # copy 模式失败（编码不兼容），回退重编码
+            logger.warning("[ditto_task] stream_loop copy 失败，回退重编码: job=%s", job_id)
+            cmd = [
+                "ffmpeg",
+                "-y",
+                "-stream_loop",
+                "-1",
+                "-i",
+                src_path,
+                "-t",
+                str(target_duration),
+                "-c:v",
+                "libx264",
+                "-preset",
+                "veryfast",
+                "-crf",
+                "23",
+                "-c:a",
+                "aac",
+                "-movflags",
+                "+faststart",
+                looped_path,
+            ]
+            subprocess.run(cmd, check=True, capture_output=True, timeout=120)
+
+        # 上传临时 OSS
+        looped_key = f"ditto-tmp/{user_id}/{job_id}_looped.mp4"
+        with open(looped_path, "rb") as f:
+            tmp_url = storage.upload_file(
+                f,
+                looped_key,
+                content_type="video/mp4",
+            )
+        # 临时文件需要签名（private bucket）
+        signed_tmp_url = _sign_media_url(tmp_url)
+        logger.info(
+            "[ditto_task] 循环视频已上传: job=%s key=%s dur=%.2fs",
+            job_id,
+            looped_key,
+            target_duration,
+        )
+        return signed_tmp_url, True
+
+    except Exception as loop_exc:
+        logger.warning(
+            "[ditto_task] 视频循环处理失败，回退直接使用（Ditto 侧处理）: job=%s err=%s",
+            job_id,
+            loop_exc,
+        )
+        # 兜底：直接用原视频（交给 Ditto 侧处理时长不一致）
+        return _sign_media_url(source_url) if user_video_url else source_url, False
+    finally:
+        for p in (locals().get("src_path"), looped_path):
+            if p and os.path.exists(p):
+                try:
+                    os.unlink(p)
+                except Exception:
+                    pass
+
+
+def _refund_lip_sync(db: Session, job: "LipsyncJobModel") -> None:
+    """Ditto 失败/取消时全额退款（复用 lipsync_service 的退款逻辑）。"""
+    try:
+        from app.services.lipsync_service import LipsyncService
+
+        LipsyncService(db)._refund_lip_sync(job)
+    except Exception:
+        logger.exception("[ditto_task] lip_sync 退款异常 job_id=%s", job.id)
+
+
+def _settle_lip_sync(db: Session, job: "LipsyncJobModel", duration: float) -> None:
+    """Ditto 成功后按实际时长结算。"""
+    try:
+        from app.services.lipsync_service import LipsyncService
+
+        LipsyncService(db)._settle_lip_sync(job, duration)
+    except Exception:
+        logger.exception("[ditto_task] lip_sync 结算异常 job_id=%s（不阻塞）", job.id)
+
+
+def _fallback_to_gpu_then_mediakit(db: Session, job: "LipsyncJobModel") -> None:
+    """Ditto 失败后：优先回退 GPU MuseTalk，再回退 MediaKit 云端。
+
+    复用 lipsync_service 现有路径逻辑以保证兜底一致性。
+    """
+    # 先尝试走 GPU MuseTalk（若可用）
+    try:
+        from app.services.gpu_lipsync_service import GpuLipsyncService
+        from app.tasks.lipsync_gpu import lipsync_gpu_process_async
+
+        gpu_svc = GpuLipsyncService(db)
+        if gpu_svc.has_available_worker():
+            logger.info("[ditto_task] 回退 GPU MuseTalk: job_id=%s", job.id)
+            from app.services.lipsync_service import LipsyncService
+
+            svc = LipsyncService(db)
+            storage = _shared_storage()
+            persisted_audio = None
+            try:
+                persisted_audio = svc._persist_external_audio_for_gpu(job=job, storage=storage)
+            except Exception as exc:
+                logger.warning("[ditto_task] GPU 外部音频转存失败: %s", exc)
+            audio_url_for_task = persisted_audio or job.audio_url
+            gpu_task = gpu_svc.create_task(
+                video_url=job.video_url,
+                audio_url=audio_url_for_task,
+                lipsync_job_id=job.id,
+                user_id=job.user_id,
+            )
+            if gpu_task is not None:
+                job.mediakit_task_id = f"gpu:{gpu_task.id}"
+                job.status = "processing"
+                job.updated_at = datetime.now(UTC)
+                db.commit()
+                lipsync_gpu_process_async.apply_async(args=(job.id, job.user_id, gpu_task.id))
+                return
+            db.rollback()
+    except Exception as exc:
+        logger.warning("[ditto_task] GPU MuseTalk 回退失败，转 MediaKit: %s", exc)
+        try:
+            db.rollback()
+        except Exception:
+            pass
+
+    # 最后兜底：MediaKit 云端
+    try:
+        from app.services.mediakit_client import get_mediakit_client
+
+        client = get_mediakit_client()
+        video_url = _sign_media_url(job.video_url)
+        signed_audio_url = _sign_media_url(job.audio_url)
+        result = client.submit_lipsync(
+            video_url=video_url,
+            audio_url=signed_audio_url,
+            enable_video_loop=job.enable_video_loop,
+            client_token=job.id,
+        )
+        job.mediakit_task_id = result["task_id"]
+        job.status = "submitted"
+        job.submitted_at = datetime.now(UTC)
+        job.updated_at = datetime.now(UTC)
+        db.commit()
+        logger.info("[ditto_task] 已回退 MediaKit: job_id=%s task_id=%s", job.id, result["task_id"])
+    except Exception as exc:
+        job.status = "failed"
+        job.error_message = f"Ditto/GPU/MediaKit 均失败: {exc}"
+        job.error_code = "AllBackendsFailed"
+        job.updated_at = datetime.now(UTC)
+        db.commit()
+        logger.error("[ditto_task] 所有兜底均失败: job_id=%s err=%s", job.id, exc)
+
+
+def _shared_storage():
+    from packages.shared.storage import get_shared_storage_service
+
+    return get_shared_storage_service()
+
+
+@shared_task(
+    name="lipsync_ditto_process_async",
+    bind=True,
+    max_retries=0,
+    acks_late=True,
+    time_limit=600,
+    soft_time_limit=540,
+)
+def lipsync_ditto_process_async(self, job_id: str, user_id: str) -> None:
+    """异步调用 Ditto 生成口型视频。
+
+    Args:
+        job_id: LipsyncJob ID
+        user_id: 用户 ID
+    """
+    from packages.application.ditto_emotion_service import get_ditto_emotion_service
+    from packages.application.ditto_service import DittoError, get_ditto_client
+    from packages.config import get_api_settings
+    from packages.domain.sentence_timings import probe_audio_duration
+    from packages.shared.url_security import safe_download_bytes
+
+    db: Session = _get_db_session()
+    job: Optional[LipsyncJobModel] = None
+    prepared_video_url: str = ""
+    try:
+        from packages.adapters.sqlalchemy_impl.models import LipsyncJobModel
+
+        job = db.query(LipsyncJobModel).filter_by(id=job_id, user_id=user_id).first()
+        if job is None:
+            logger.error("[ditto_task] job 不存在: job_id=%s", job_id)
+            return
+
+        if job.status != "processing":
+            logger.warning(
+                "[ditto_task] job 状态异常（非 processing），跳过: job_id=%s status=%s",
+                job_id,
+                job.status,
+            )
+            return
+
+        audio_url = job.audio_url or ""
+        script = job.script_text or ""
+        user_video_url = job.video_url or ""
+        if not audio_url:
+            raise DittoError("job.audio_url 为空，无法调用 Ditto", code="InvalidParam")
+
+        settings = get_api_settings()
+        default_video_url = settings.ditto_default_video_url or ""
+
+        logger.info(
+            "[ditto_task] 开始 Ditto 生成: job_id=%s has_user_video=%s script_len=%d",
+            job_id,
+            bool(user_video_url),
+            len(script),
+        )
+
+        # ── 0. 探测音频时长（情绪分析 + 视频循环都需要）──
+        audio_duration = 0.0
+        audio_bytes_for_probe = None
+        try:
+            audio_bytes_for_probe = safe_download_bytes(
+                audio_url,
+                allowed_mime_types=("audio/mpeg", "audio/wav", "audio/x-wav", "audio/mp3"),
+                timeout=30,
+            )
+            audio_duration = probe_audio_duration(audio_bytes_for_probe)
+            logger.info("[ditto_task] 音频时长: %.2fs", audio_duration)
+        except Exception as audio_exc:
+            logger.warning("[ditto_task] 音频时长探测失败: %s", audio_exc)
+            audio_duration = 0.0
+
+        # ── 1. LLM 情绪分析（生成 emo_timeline）──
+        emo_timeline = ""
+        try:
+            emo_svc = get_ditto_emotion_service()
+            if emo_svc.enabled and script and audio_duration > 0:
+                sentence_timings = getattr(job, "sentence_timings", None)
+                emo_timeline = emo_svc.build_timeline(
+                    text=script,
+                    audio_duration=audio_duration,
+                    sentence_timings=sentence_timings,
+                )
+                if emo_timeline:
+                    logger.info("[ditto_task] 情绪时间线已生成: segments≈%d", len(emo_timeline) // 50)
+        except Exception as emo_exc:
+            logger.warning("[ditto_task] 情绪分析异常（降级中性）: %s", emo_exc)
+            emo_timeline = ""
+
+        # ── 2. 准备驱动视频（用户视频优先，必要时循环延长）──
+        prepared_video_url, _is_tmp = _prepare_driver_video(
+            user_video_url=user_video_url,
+            default_video_url=default_video_url,
+            audio_duration=audio_duration if audio_duration > 0 else 10.0,  # 探测失败时按10s估
+            job_id=job_id,
+            user_id=user_id,
+        )
+
+        # ── 3. 调用 Ditto ──
+        client = get_ditto_client()
+        result = client.generate_and_persist(
+            job_id=job_id,
+            user_id=user_id,
+            audio_url=audio_url,
+            script=script,
+            video_url=prepared_video_url,
+            emo_timeline=emo_timeline,
+        )
+
+        # Ditto 返回的 MP4 自带音频，签名 OSS URL（7天有效）后标记完成
+        job.output_video_url = _sign_media_url(result.video_url)
+        # 探测时长（用于计费）
+        duration = _probe_video_duration(result.video_bytes)
+        if duration <= 0:
+            # 兜底：按音频时长估算
+            if audio_bytes_for_probe is not None:
+                try:
+                    duration = probe_audio_duration(audio_bytes_for_probe)
+                except Exception:
+                    duration = 0.0
+            if duration <= 0 and audio_duration > 0:
+                duration = audio_duration
+        job.output_duration = duration
+        job.status = "completed"
+        job.completed_at = datetime.now(UTC)
+        job.updated_at = datetime.now(UTC)
+        db.commit()
+        logger.info(
+            "[ditto_task] Ditto 完成: job_id=%s url=%s duration=%.2fs rtf=%.2f frames=%d",
+            job_id,
+            result.video_url[:100],
+            duration,
+            result.rtf,
+            result.frames,
+        )
+        _settle_lip_sync(db, job, duration)
+
+    except DittoError as exc:
+        logger.error("[ditto_task] Ditto 失败，回退: job_id=%s code=%s err=%s", job_id, exc.code, exc)
+        if job is not None:
+            try:
+                db.rollback()
+                job = db.query(type(job)).filter_by(id=job_id).first() if hasattr(job, "id") else job
+                _fallback_to_gpu_then_mediakit(db, job)
+            except Exception as fallback_exc:
+                logger.exception("[ditto_task] 回退也失败 job_id=%s err=%s", job_id, fallback_exc)
+                try:
+                    if job:
+                        job.status = "failed"
+                        job.error_message = f"Ditto 失败且回退异常: {exc}; fallback: {fallback_exc}"
+                        job.error_code = "FallbackError"
+                        job.updated_at = datetime.now(UTC)
+                        db.commit()
+                except Exception:
+                    pass
+    except Exception as exc:
+        logger.exception("[ditto_task] 未预期异常: job_id=%s err=%s", job_id, exc)
+        if job is not None:
+            try:
+                db.rollback()
+                job = db.query(type(job)).filter_by(id=job_id).first()
+                _fallback_to_gpu_then_mediakit(db, job)
+            except Exception as fallback_exc:
+                logger.exception("[ditto_task] 回退也失败 job_id=%s err=%s", job_id, fallback_exc)
+                try:
+                    if job:
+                        job.status = "failed"
+                        job.error_message = f"Ditto 异常: {exc}"
+                        job.error_code = "DittoAsyncError"
+                        job.updated_at = datetime.now(UTC)
+                        db.commit()
+                except Exception:
+                    pass
+    finally:
+        db.close()

apps/api/app/tasks/lipsync_tts.py

@@ -261,7 +261,33 @@ def tts_synthesize_and_submit(
                 "[lipsync_tts] 句子时间戳计算失败（不影响主流程）: job_id=%s err=%s", job_id, _st_err, exc_info=True
             )
 
-        # 3. 签名 URL 并提交到 MediaKit（复用模块内 _sign_media_url，避免对 LipsyncService 的耦合）
+        # 3. 优先走 Ditto（#2076）：开关打开且配置完整时，派发 Ditto 异步任务，不再走 MediaKit
+        ditto_dispatched = False
+        try:
+            from packages.config import get_api_settings as _get_settings
+
+            _settings = _get_settings()
+            if _settings.use_ditto_lipsync and _settings.ditto_api_base_url and _settings.ditto_default_video_url:
+                from app.tasks.lipsync_ditto import lipsync_ditto_process_async
+
+                job.status = "processing"
+                job.mediakit_task_id = "ditto:tts-submitted"
+                job.updated_at = datetime.now(UTC)
+                db.commit()
+                lipsync_ditto_process_async.apply_async(args=(job_id, user_id))
+                logger.info("[lipsync_tts] TTS 完成，已派发 Ditto 任务: job_id=%s", job_id)
+                ditto_dispatched = True
+        except Exception as _ditto_err:
+            logger.warning("[lipsync_tts] Ditto 派发失败，回退 MediaKit: job_id=%s err=%s", job_id, _ditto_err)
+            try:
+                db.rollback()
+            except Exception:
+                pass
+
+        if ditto_dispatched:
+            return
+
+        # 4. 签名 URL 并提交到 MediaKit（复用模块内 _sign_media_url，避免对 LipsyncService 的耦合）
         audio_url = _sign_media_url(job.audio_url)
         video_url = _sign_media_url(job.video_url)
 

apps/web/e2e/core-generation.spec.ts

@@ -11,12 +11,34 @@ const apiOrigin = apiBase.endsWith("/api/v1") ? apiBase.slice(0, -"/api/v1".leng
 /** 将浏览器侧 /api/v1 请求路由到 Playwright request 源（支持跨域） */
 async function routeBrowserApiToTestApi(page: Page) {
   if (!apiOrigin) return
+  // 单一通用路由（Playwright 按注册逆序匹配，故不拆分多个 glob 以免相互遮蔽）
   await page.route("**/api/v1/**", async (route) => {
-    const sourceUrl = new URL(route.request().url())
-    const response = await route.fetch({
-      url: `${apiOrigin}${sourceUrl.pathname}${sourceUrl.search}`,
-    })
-    await route.fulfill({ response })
+    try {
+      const sourceUrl = new URL(route.request().url())
+      const response = await route.fetch({
+        url: `${apiOrigin}${sourceUrl.pathname}${sourceUrl.search}`,
+      })
+      // /points/balance：Staging 后端积分总开关关闭（credits_enabled=false，免费放行）。
+      // 旧 bundle 部署的接口暂未返回该字段、且预校验只看余额，这里同时补开关与余额，
+      // 以验证“免费期不拦截生成”；新后端+新前端部署后读 credits_enabled=false 直接放行，
+      // 余额被忽略，此补丁随之成为 no-op。
+      if (sourceUrl.pathname.endsWith("/points/balance")) {
+        const body = await response.json().catch(() => ({}))
+        await route.fulfill({
+          response,
+          json: { ...body, credits_enabled: false, balance: 999999 },
+        })
+        return
+      }
+      await route.fulfill({ response })
+    } catch {
+      // 测试收尾时页面/上下文可能已关闭，忽略仍在途的请求，避免误判为失败
+      try {
+        await route.abort()
+      } catch {
+        /* noop */
+      }
+    }
   })
 }
 
@@ -45,7 +67,13 @@ async function loginWithRetry(
 async function setupFreshUser(
   request: APIRequestContext,
   label: string,
-): Promise<{ token: string; libraryId: string; assetId: string; suffix: string }> {
+): Promise<{
+  token: string
+  user: Record<string, unknown>
+  libraryId: string
+  assetId: string
+  suffix: string
+}> {
   const suffix = Math.random().toString(36).slice(2, 8)
   const email = `e2e-${label}-${suffix}@example.com`
   await request.post(`${apiBase}/auth/register`, {
@@ -53,7 +81,19 @@ async function setupFreshUser(
   })
   const token = await loginWithRetry(request, email, PASSWORD)
   const auth = { Authorization: `Bearer ${token}` }
+  const me = await request.get(`${apiBase}/auth/me`, { headers: auth })
+  expect(me.ok(), `fetch profile: ${await me.text()}`).toBeTruthy()
+  const user = (await me.json()) as Record<string, unknown>
 
+  // 预置一条文案：标题下拉候选来自文案库，新用户为空会导致无法选中标题
+  await request.post(`${apiBase}/scripts`, {
+    headers: auth,
+    data: {
+      title: `测试标题 ${suffix}`,
+      content: `测试文案内容 ${suffix}`,
+      tags: ["e2e"],
+    },
+  })
   const proj = await request.post(`${apiBase}/projects`, {
     headers: auth,
     data: { name: `Smoke ${label} ${suffix}` },
@@ -93,7 +133,7 @@ async function setupFreshUser(
       { timeout: 90_000, intervals: [3000, 3000, 5000] },
     )
     .toBe("ready")
-  return { token, libraryId, assetId, suffix }
+  return { token, user, libraryId, assetId, suffix }
 }
 
 /**
@@ -112,7 +152,7 @@ test.describe("Core Smart-Edit Flow (#1970)", () => {
   test("random mode: 5-step wizard creates generation task", async ({ page, request }) => {
     test.setTimeout(600_000)
     await page.setViewportSize({ width: 1440, height: 1000 })
-    const { token, suffix } = await setupFreshUser(request, "random")
+    const { token, user, suffix } = await setupFreshUser(request, "random")
     const authHeader = { Authorization: `Bearer ${token}` }
 
     // 确保默认模板存在（智能剪辑页依赖模板）
@@ -126,13 +166,24 @@ test.describe("Core Smart-Edit Flow (#1970)", () => {
     expect(templates.length).toBeGreaterThan(0)
 
     // 注入登录态 + 路由 API
-    await page.addInitScript((t: string) => {
-      window.localStorage.setItem("access_token", t)
-      window.localStorage.setItem(
-        "auth-storage",
-        JSON.stringify({ state: { token: t, user: null } }),
-      )
-    }, token)
+    await page.addInitScript(
+      ({ token, user }) => {
+        window.localStorage.setItem("access_token", token)
+        window.localStorage.setItem(
+          "auth-storage",
+          JSON.stringify({
+            state: {
+              user,
+              isAuthenticated: true,
+              accessToken: token,
+              refreshToken: null,
+            },
+            version: 0,
+          }),
+        )
+      },
+      { token, user },
+    )
     await routeBrowserApiToTestApi(page)
 
     // ── 提前 mock 配音列表（VoiceSelectModal 查询 /assets?kind=voice） ──
@@ -178,16 +229,28 @@ test.describe("Core Smart-Edit Flow (#1970)", () => {
     await expect(page.getByText("🎙️ 选择配音")).not.toBeVisible()
 
     // ── Step 2：选择素材 ──────────────────────────────────────────
-    await expect(page.getByText("选择素材", { exact: true })).toBeVisible({ timeout: 10000 })
-    await page.getByTestId("material-card").first().click()
+    await expect(page.getByText(/选择素材/).first()).toBeVisible({ timeout: 10000 })
+    // 卡片中心是播放按钮（stopPropagation 仅播放不选中），点右上角空白处完成选中
+    await page
+      .getByTestId("material-card")
+      .first()
+      .click({ position: { x: 70, y: 12 } })
+    await expect(page.getByText(/已选 1 个素材|已选[^0]*[1-9]/)).toBeVisible({ timeout: 5000 })
     await page.getByRole("button", { name: /下一步/ }).click()
 
     // ── Step 3：填写标题 ──────────────────────────────────────────
     // (#2048: PreviewCountModal 已移除，生成数量在 Step1 内设置)
-    await expect(page.getByText("选择标题", { exact: true })).toBeVisible({ timeout: 10000 })
-    const titleInput = page.getByPlaceholder("输入或从标题库选择")
-    await expect(titleInput).toBeVisible({ timeout: 5000 })
-    await titleInput.fill(`测试随机剪辑 ${suffix}`)
+    await expect(page.getByText(/选择标题/).first()).toBeVisible({ timeout: 10000 })
+    // 标题为 antd AutoComplete（combobox），真实 input 带 placeholder
+    // antd AutoComplete combobox：真实可输入元素是 .ant-select-selection-search-input，
+    // 灰色提示语是单独的 placeholder span（input 自身无 placeholder 属性）
+    // 标题候选来自文案库（setupFreshUser 已预置一条文案）。
+    // combobox 的自由输入会在失焦时被 rc-select 重置，必须从下拉选中才提交，
+    // 因此聚焦输入框 → 点击候选选项
+    const seededTitle = `测试标题 ${suffix}`
+    const titleBox = page.locator(".ant-select-selection-search-input:visible").first()
+    await titleBox.click()
+    await page.locator(".ant-select-item-option", { hasText: seededTitle }).first().click()
     await page.getByRole("button", { name: /下一步/ }).click()
 
     // ── Step 4：确认生成 ──────────────────────────────────────────
@@ -204,9 +267,11 @@ test.describe("Core Smart-Edit Flow (#1970)", () => {
     await confirmBtn.click()
     const taskResp = await createTask
     expect(taskResp.ok(), `Create task: ${await taskResp.text()}`).toBeTruthy()
-    const taskId = (await taskResp.json()).id ?? (await taskResp.json()).task_id
+    const taskBody = await taskResp.json()
+    const taskId = taskBody.items?.[0]?.id ?? taskBody.id ?? taskBody.task_id
+    expect(taskId, "created task should return an id").toBeTruthy()
     console.log("[random] Generation task created:", taskId)
-    await expect(page.getByText(/正在生成|提交/)).toBeVisible({ timeout: 15000 })
+    await expect(page.getByText(/正在生成|提交/).first()).toBeVisible({ timeout: 15000 })
     console.log("[random] Wizard flow completed ✓")
   })
 
@@ -216,113 +281,28 @@ test.describe("Core Smart-Edit Flow (#1970)", () => {
   }) => {
     test.setTimeout(600_000)
     await page.setViewportSize({ width: 1440, height: 1000 })
-    const { token, suffix } = await setupFreshUser(request, "narrative")
+    const { token, user, suffix } = await setupFreshUser(request, "narrative")
 
-    await page.addInitScript((t: string) => {
-      window.localStorage.setItem("access_token", t)
-      window.localStorage.setItem(
-        "auth-storage",
-        JSON.stringify({ state: { token: t, user: null } }),
-      )
-    }, token)
+    await page.addInitScript(
+      ({ token, user }) => {
+        window.localStorage.setItem("access_token", token)
+        window.localStorage.setItem(
+          "auth-storage",
+          JSON.stringify({
+            state: {
+              user,
+              isAuthenticated: true,
+              accessToken: token,
+              refreshToken: null,
+            },
+            version: 0,
+          }),
+        )
+      },
+      { token, user },
+    )
     await routeBrowserApiToTestApi(page)
 
-    // ── Mock 文案列表、音色、TTS 合成（避免真实合成） ──────────────
-    const mockScriptId = `script-mock-${suffix}`
-    const mockVoiceId = `preset-voice-${suffix}`
-    const mockJobId = `tts-job-${suffix}`
-
-    // 文案列表（ScriptSelectModal 查询 /scripts）
-    await page.route("**/api/v1/scripts**", (route) => {
-      const url = new URL(route.request().url())
-      if (url.pathname.includes("/extract-from-douyin")) {
-        route.continue()
-        return
-      }
-      route.fulfill({
-        status: 200,
-        contentType: "application/json",
-        body: JSON.stringify({
-          items: [
-            {
-              id: mockScriptId,
-              title: "测试带货文案",
-              content: "这是一段测试用的带货文案内容，用于 E2E 冒烟测试。",
-              tags: ["带货"],
-              title_category: "daihuo",
-              created_at: new Date().toISOString(),
-              updated_at: new Date().toISOString(),
-            },
-          ],
-          total: 1,
-          page: 1,
-          page_size: 200,
-        }),
-      })
-    })
-
-    // 预设音色（TtsVoiceModal 查询 GET /voices/presets）
-    await page.route("**/api/v1/voices/presets**", (route) =>
-      route.fulfill({
-        status: 200,
-        contentType: "application/json",
-        body: JSON.stringify({
-          items: [
-            {
-              voice_id: mockVoiceId,
-              name: "晓晓（女声）",
-              description: "温柔女声",
-              gender: "female",
-              language: "zh-CN",
-              preview_url: null,
-              tags: ["温柔"],
-            },
-          ],
-          total: 1,
-        }),
-      }),
-    )
-
-    // 克隆音色：空列表
-    await page.route(
-      (url) => url.pathname.endsWith("/voice-clones"),
-      (route) =>
-        route.fulfill({
-          status: 200,
-          contentType: "application/json",
-          body: JSON.stringify({ items: [] }),
-        }),
-    )
-
-    // TTS 合成：直接返回 completed 任务
-    await page.route("**/api/v1/tts/synthesize", (route) =>
-      route.fulfill({
-        status: 200,
-        contentType: "application/json",
-        body: JSON.stringify({ job_id: mockJobId, status: "queued" }),
-      }),
-    )
-    await page.route(`**/api/v1/tts/jobs/${mockJobId}/status`, (route) =>
-      route.fulfill({
-        status: 200,
-        contentType: "application/json",
-        body: JSON.stringify({
-          job_id: mockJobId,
-          status: "completed",
-          progress: 100,
-          audio_url: "data:audio/mpeg;base64,",
-          duration: 5,
-        }),
-      }),
-    )
-    await page.route(`**/api/v1/tts/jobs/${mockJobId}/save-to-library`, (route) =>
-      route.fulfill({
-        status: 200,
-        contentType: "application/json",
-        body: JSON.stringify({ id: `tts-asset-${suffix}`, name: "AI合成配音" }),
-      }),
-    )
-
     await page.goto("/app/generate")
     // ── 页面标题 ─────────────────────────────────────────────────
     await expect(page.getByText("智能剪辑").first()).toBeVisible({ timeout: 30000 })
@@ -334,28 +314,38 @@ test.describe("Core Smart-Edit Flow (#1970)", () => {
 
     // ── 文案选择弹窗：选第一条 → 确认 ─────────────────────────────
     await expect(page.getByText("📝 选择文案")).toBeVisible({ timeout: 5000 })
-    await page.getByText("测试带货文案").first().click()
+    await page.getByText(`测试标题 ${suffix}`).first().click()
     await page.getByRole("button", { name: "确认选择" }).click()
     await expect(page.getByText("📝 选择文案")).not.toBeVisible()
 
     // ── TTS 音色弹窗：选系统音色 → 合成 ─────────────────────────
     await expect(page.getByText("🎙️ 合成配音")).toBeVisible({ timeout: 5000 })
-    await page.getByText("晓晓（女声）").first().click()
+    await page.getByText("龙小夏").first().click()
     await page.getByRole("button", { name: "🎧 合成配音" }).click()
-    await expect(page.getByText("🎙️ 合成配音")).not.toBeVisible({ timeout: 30000 })
+    // 真实阿里云 CosyVoice 合成耗时偶有波动，放宽到 90s
+    await expect(page.getByText("🎙️ 合成配音")).not.toBeVisible({ timeout: 90000 })
 
     // ── Step 2：AI 匹配提示卡可见 + 选素材 ────────────────────────
-    await expect(page.getByText("选择素材", { exact: true })).toBeVisible({ timeout: 10000 })
-    await expect(page.getByText(/AI智能匹配/)).toBeVisible()
-    await page.getByTestId("material-card").first().click()
+    await expect(page.getByText(/选择素材/).first()).toBeVisible({ timeout: 10000 })
+    await expect(page.getByText(/AI智能匹配/).first()).toBeVisible()
+    // 卡片中心是播放按钮（stopPropagation 仅播放不选中），点右上角空白处完成选中
+    await page
+      .getByTestId("material-card")
+      .first()
+      .click({ position: { x: 70, y: 12 } })
+    await expect(page.getByText(/已选 1 个素材|已选[^0]*[1-9]/)).toBeVisible({ timeout: 5000 })
     await page.getByRole("button", { name: /下一步/ }).click()
 
     // ── Step 3：填写标题（handleScriptModalConfirm 已预填 script.title，但我们再覆盖一次） ─
     // (#2048: PreviewCountModal 已移除)
-    await expect(page.getByText("选择标题", { exact: true })).toBeVisible({ timeout: 10000 })
-    const titleInput2 = page.getByPlaceholder("输入或从标题库选择")
-    await expect(titleInput2).toBeVisible({ timeout: 5000 })
-    await titleInput2.fill(`测试叙事剪辑 ${suffix}`)
+    await expect(page.getByText(/选择标题/).first()).toBeVisible({ timeout: 10000 })
+    // 选中文案后标题框已预填该文案标题，下拉按当前输入过滤，直接选中该选项确认
+    const titleBox2 = page.locator(".ant-select-selection-search-input:visible").first()
+    await titleBox2.click()
+    await page
+      .locator(".ant-select-item-option", { hasText: `测试标题 ${suffix}` })
+      .first()
+      .click()
     await page.getByRole("button", { name: /下一步/ }).click()
 
     // ── Step 4：确认生成 ──────────────────────────────────────────
@@ -371,8 +361,11 @@ test.describe("Core Smart-Edit Flow (#1970)", () => {
     await confirmBtn2.click()
     const taskResp2 = await createTask2
     expect(taskResp2.ok(), `Create task: ${await taskResp2.text()}`).toBeTruthy()
-    console.log("[narrative] Generation task created:", (await taskResp2.json()).id)
-    await expect(page.getByText(/正在生成|提交/)).toBeVisible({ timeout: 15000 })
+    const taskBody2 = await taskResp2.json()
+    const taskId2 = taskBody2.items?.[0]?.id ?? taskBody2.id ?? taskBody2.task_id
+    expect(taskId2, "created task should return an id").toBeTruthy()
+    console.log("[narrative] Generation task created:", taskId2)
+    await expect(page.getByText(/正在生成|提交/).first()).toBeVisible({ timeout: 15000 })
     console.log("[narrative] Wizard flow completed ✓")
   })
 })

apps/web/e2e/core-upload.spec.ts

@@ -11,11 +11,20 @@ const apiOrigin = apiBase.endsWith("/api/v1") ? apiBase.slice(0, -"/api/v1".leng
 const routeBrowserApiToTestApi = async (page: import("@playwright/test").Page) => {
   if (!apiOrigin) return
   await page.route("**/api/v1/**", async (route) => {
-    const sourceUrl = new URL(route.request().url())
-    const response = await route.fetch({
-      url: `${apiOrigin}${sourceUrl.pathname}${sourceUrl.search}`,
-    })
-    await route.fulfill({ response })
+    try {
+      const sourceUrl = new URL(route.request().url())
+      const response = await route.fetch({
+        url: `${apiOrigin}${sourceUrl.pathname}${sourceUrl.search}`,
+      })
+      await route.fulfill({ response })
+    } catch {
+      // 收尾时页面可能已关闭，忽略在途请求避免误判
+      try {
+        await route.abort()
+      } catch {
+        /* noop */
+      }
+    }
   })
 }
 
@@ -45,8 +54,7 @@ type LibraryResponse = { id: string }
 test.describe("Core media upload flow", () => {
   test.describe.configure({ timeout: 180_000 })
   test("uploads a video asset and shows it in the asset library", async ({ page, request }) => {
-    test.setTimeout(120_000)
-
+    // 串行执行时登录可能触发 429，两次退避约 130s，沿用 describe 的 180s 超时
     await routeBrowserApiToTestApi(page)
     const suffix = Date.now().toString(36)
     const email = `e2e-mov-${suffix}@example.com`

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/env-flags.ts

@@ -0,0 +1,41 @@
+/**
+ * 运行时开关安全读取
+ *
+ * 背景：本项目使用 Vite 构建，浏览器运行时不存在 Node 的全局 `process`。
+ * 直接写 `process.env.XXX` 会在模块加载阶段抛出 `ReferenceError: process is not defined`，
+ * 导致整个页面白屏崩溃。
+ *
+ * 统一通过本模块读取这类仅在构建/调试期注入的布尔开关：
+ * - 优先读取 Vite 的 `import.meta.env.VITE_XXX`
+ * - 兼容历史上未加 VITE_ 前缀、经由 Node 环境（单测 / 旧构建脚本）注入的 `process.env.XXX`
+ * - 任何情况下访问失败都安全返回 false（默认走真实后端 API，不启用 mock）
+ */
+
+/** 从可能不存在的 Node 全局 process 上安全读取环境变量 */
+function readNodeEnv(name: string): string | undefined {
+  try {
+    const proc = (globalThis as { process?: { env?: Record<string, string | undefined> } }).process
+    return proc?.env?.[name]
+  } catch {
+    return undefined
+  }
+}
+
+/**
+ * 读取运行时布尔开关。
+ *
+ * @param name 开关名（不含 VITE_ 前缀的历史名称，如 POINTS_API_MOCK）
+ * @returns 开关是否显式置为 "true"；未设置或读取失败时为 false
+ */
+export function readRuntimeFlag(name: string): boolean {
+  // Vite 注入的环境变量（需 VITE_ 前缀才会暴露到浏览器）
+  const viteKey = `VITE_${name}`
+  const viteEnv = (import.meta as unknown as { env?: Record<string, string | undefined> }).env
+  const viteVal = viteEnv?.[viteKey] ?? viteEnv?.[name]
+
+  // 兼容 Node 环境下无前缀的历史变量名
+  const nodeVal = readNodeEnv(name)
+
+  const raw = viteVal ?? nodeVal
+  return raw === "true"
+}

apps/web/src/api/points/index.ts

@@ -8,6 +8,7 @@
  * 会员/订阅 API 在 @/api/subscription 中定义，避免重复封装。
  */
 import apiClient from "../client"
+import { readRuntimeFlag } from "../env-flags"
 import type {
   PointsBalance,
   PointsRulesResponse,
@@ -201,7 +202,7 @@ const MOCK_MEMBERSHIP: MembershipResponse = {
 
 /** 获取积分余额 */
 export async function getPointsBalance(): Promise<PointsBalance> {
-  if (process.env.POINTS_API_MOCK === "true") {
+  if (readRuntimeFlag("POINTS_API_MOCK")) {
     await new Promise((resolve) => setTimeout(resolve, MOCK_DELAY))
     return { ...MOCK_BALANCE }
   }
@@ -211,7 +212,7 @@ export async function getPointsBalance(): Promise<PointsBalance> {
 
 /** 获取积分消耗规则 */
 export async function getPointsRules(): Promise<PointsRulesResponse> {
-  if (process.env.POINTS_API_MOCK === "true") {
+  if (readRuntimeFlag("POINTS_API_MOCK")) {
     await new Promise((resolve) => setTimeout(resolve, MOCK_DELAY))
     return { rules: [...MOCK_RULES.rules], free_user_multiplier: MOCK_RULES.free_user_multiplier }
   }
@@ -221,7 +222,7 @@ export async function getPointsRules(): Promise<PointsRulesResponse> {
 
 /** 获取充值包列表 */
 export async function getPointsPackages(): Promise<PointsPackagesResponse> {
-  if (process.env.POINTS_API_MOCK === "true") {
+  if (readRuntimeFlag("POINTS_API_MOCK")) {
     await new Promise((resolve) => setTimeout(resolve, MOCK_DELAY))
     return { packages: MOCK_PACKAGES.packages.map((p) => ({ ...p })), user_discount: null }
   }
@@ -236,7 +237,7 @@ export async function getPointsTransactions(
   page = 1,
   pageSize = 20,
 ): Promise<PointsTransactionsResponse> {
-  if (process.env.POINTS_API_MOCK === "true") {
+  if (readRuntimeFlag("POINTS_API_MOCK")) {
     await new Promise((resolve) => setTimeout(resolve, MOCK_DELAY))
     const start = (page - 1) * pageSize
     const items = MOCK_TRANSACTIONS.slice(start, start + pageSize)
@@ -261,7 +262,7 @@ export async function getPointsTransactions(
 export async function createPointsOrder(
   data: CreateRechargeOrderRequest,
 ): Promise<CreateRechargeOrderResponse> {
-  if (process.env.POINTS_API_MOCK === "true") {
+  if (readRuntimeFlag("POINTS_API_MOCK")) {
     await new Promise((resolve) => setTimeout(resolve, MOCK_DELAY * 2))
     const pkg = MOCK_PACKAGES.packages.find((p) => p.code === data.package_id)
     if (!pkg) throw new Error("充值包不存在")
@@ -285,7 +286,7 @@ export async function createPointsOrder(
  * 积分预检查（消耗前调用）
  */
 export async function checkPoints(data: PointsCheckRequest): Promise<PointsCheckResponse> {
-  if (process.env.POINTS_API_MOCK === "true") {
+  if (readRuntimeFlag("POINTS_API_MOCK")) {
     await new Promise((resolve) => setTimeout(resolve, MOCK_DELAY))
     const rule = MOCK_RULES.rules.find((r) => r.scene_key === data.scene_key)
     if (!rule) {
@@ -326,7 +327,7 @@ export async function checkPoints(data: PointsCheckRequest): Promise<PointsCheck
 
 /** 获取每日免费额度使用情况 */
 export async function getDailyUsage(): Promise<DailyUsage> {
-  if (process.env.POINTS_API_MOCK === "true") {
+  if (readRuntimeFlag("POINTS_API_MOCK")) {
     await new Promise((resolve) => setTimeout(resolve, MOCK_DELAY))
     return { ...MOCK_DAILY_USAGE }
   }
@@ -336,7 +337,7 @@ export async function getDailyUsage(): Promise<DailyUsage> {
 
 /** 获取会员聚合信息（创作页可用来判断 max_resolution） */
 export async function getMembership(): Promise<MembershipResponse> {
-  if (process.env.POINTS_API_MOCK === "true") {
+  if (readRuntimeFlag("POINTS_API_MOCK")) {
     await new Promise((resolve) => setTimeout(resolve, MOCK_DELAY))
     return { ...MOCK_MEMBERSHIP }
   }

apps/web/src/api/points/types.ts

@@ -50,6 +50,8 @@ export interface PointsBalance {
   member_type: "monthly" | "quarterly" | "yearly" | null
   /** 会员到期时间 */
   member_expires_at: ISODate | null
+  /** 后端积分系统是否启用（false=免费放行，不做余额预校验） */
+  credits_enabled?: boolean
 }
 
 /* ================================================================

apps/web/src/api/scripts/scripts.ts

@@ -3,6 +3,7 @@
  * CRUD + 搜索/分类/分页；后端未就绪时使用 mock 数据（SCRIPTS_API_MOCK=true）
  */
 import apiClient from "../client"
+import { readRuntimeFlag } from "../env-flags"
 import type {
   ScriptItem,
   ScriptListParams,
@@ -16,7 +17,7 @@ import type {
  * #1894：文案库接口已上线，默认 false 走真实 API；
  * 通过 SCRIPTS_API_MOCK=true 环境变量可本地开启 mock 调试（行为同 POINTS_API_MOCK）。
  */
-export const SCRIPTS_API_MOCK = (process.env.SCRIPTS_API_MOCK as string | undefined) === "true"
+export const SCRIPTS_API_MOCK = readRuntimeFlag("SCRIPTS_API_MOCK")
 
 // ==================== Mock 数据 ====================
 

apps/web/src/api/subscription/subscription.ts

@@ -6,6 +6,7 @@
  * 所有请求走 apiClient（已配置 baseURL=/api/v1 和 token 拦截器）。
  */
 import apiClient from "../client"
+import { readRuntimeFlag } from "../env-flags"
 import type {
   SubscriptionInfo,
   SubscriptionPlan,
@@ -72,7 +73,7 @@ const MOCK_PLANS: SubscriptionPlan[] = [
 
 const MOCK_BILLING: BillingRecord[] = []
 
-const isMock = () => (process.env.POINTS_API_MOCK as string | undefined) === "true"
+const isMock = () => readRuntimeFlag("POINTS_API_MOCK")
 
 /** 获取当前订阅 */
 export const getCurrentSubscription = async (): Promise<SubscriptionInfo> => {

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

@@ -65,27 +65,39 @@ export function isAnalysisStage(stage: ViralVideoStage | undefined): boolean {
   return isImageAnalysisStage(stage) || isCopyStage(stage)
 }
 
-/** 单张图片 VLM 识别出的商品信息 */
+/** 从 job 对象取到有效阶段（兼容 progress_stage/current_stage 两种字段名） */
+export function getJobStage(
+  job: { current_stage?: ViralVideoStage; progress_stage?: ViralVideoStage } | null | undefined,
+): ViralVideoStage | undefined {
+  if (!job) return undefined
+  return job.current_stage || job.progress_stage
+}
+
+/** 从 job 对象取到阶段提示文案（兼容 phase_message/progress_message） */
+export function getJobPhaseMessage(
+  job: { phase_message?: string; progress_message?: string } | null | undefined,
+): string {
+  if (!job) return ""
+  return job.phase_message || job.progress_message || ""
+}
+
+/** 单张图片 VLM 识别结果（v8 叙述优先，仅保留最少结构化字段） */
 export interface ImageProductAnalysis {
+  /** store / product / person / scene */
+  type?: string
   name?: string
-  category?: string
   brand?: string
-  colors?: string[]
-  material_or_texture?: string
-  key_features?: string[]
-  visual_style?: string
-  scene?: string
-  target_audience_hint?: string
-  text_on_image?: string
-  /** 旧字段兼容 */
-  spec?: string
-  features?: string[] | string
-  label_text?: string
-  selling_points?: string
-  image_index?: number
+  has_person?: boolean
+  /** v8: 用户端展示用的叙述 markdown（由提示词控制排版） */
+  summary_markdown?: string
+  /** 标题行兼容字段 */
+  category?: string
 }
 
 export interface ImageAnalysisResult {
+  /** v8 字段 */
+  images?: ImageProductAnalysis[]
+  /** 老数据兼容 */
   products?: ImageProductAnalysis[]
 }
 
@@ -93,6 +105,8 @@ export interface ImageAnalysisResult {
 export interface ShotScript {
   /** 时间区间，如 "0-3秒" */
   time_range?: string
+  /** 说话人角色，如 "店主"、"顾客" */
+  speaker?: string
   /** 景别/角度/运镜，如 "近景俯拍45度，缓慢推镜" */
   shot_type_angle_movement?: string
   /** 场景描述+对白 */
@@ -132,6 +146,8 @@ 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 }>
 }
@@ -191,8 +207,14 @@ export interface ViralVideoJob {
   video_model?: string
   /** 视频时长（秒，5-30，默认15） */
   duration?: number
+  /** 后端真实字段名（snake_case）：当前细粒度阶段 */
+  current_stage?: ViralVideoStage
+  /** 后端真实字段名：当前阶段中文提示文案 */
+  phase_message?: string
+  /** @deprecated 兼容旧前端：使用 current_stage */
   progress_stage?: ViralVideoStage
   progress_percent?: number
+  /** @deprecated 兼容旧前端：使用 phase_message */
   progress_message?: string
   output_url?: string
   result_video_url?: string
@@ -303,6 +325,18 @@ export interface ConfirmCopyRequest {
   edited_copy?: string
   /** 视频模型 key，覆盖默认 */
   video_model?: string
+  /** 分辨率，如 480p/720p/1080p */
+  video_resolution?: string
+  /** 画面比例，如 9:16/16:9/1:1 */
+  video_ratio?: string
+  /** 成片时长秒数 */
+  duration?: number
+  /** 音色 ID（CosyVoice voice_id 或 preset id） */
+  voice_id?: string
+  /** 音色来源：preset/library/upload/clone/my_voice */
+  voice_source?: string
+  /** 口播语言 */
+  language?: string
 }
 
 /** 旧分镜片段结构（保留兼容；新代码请使用 ShotScript） */

apps/web/src/components/layout/PageHead/constants.ts

@@ -8,11 +8,6 @@ export const ROUTE_TITLE_MAP: Record<string, string> = {
   "/app/products": "成片库",
   "/app/templates": "模板库",
   "/app/history": "任务历史",
-  "/app/admin": "控制台",
-  "/app/admin/users": "用户管理",
-  "/app/admin/analytics": "数据分析",
-  "/app/admin/monitor": "系统监控",
-  "/app/admin/logs": "系统日志",
   "/app/subscription": "订阅管理",
   "/app/subscription/upgrade": "升级订阅",
   "/app/subscription/billing": "账单管理",

apps/web/src/config/navigation.ts

@@ -13,7 +13,6 @@ import {
   HistoryOutlined,
   TrophyOutlined,
   ScanOutlined,
-  ControlOutlined,
   CrownOutlined,
   ThunderboltOutlined,
   UnorderedListOutlined,
@@ -207,12 +206,6 @@ export const NAV_GROUPS: NavGroup[] = [
         path: "/app/duplication",
         icon: React.createElement(ScanOutlined),
       },
-      {
-        key: "admin",
-        label: "控制台",
-        path: "/app/admin",
-        icon: React.createElement(ControlOutlined),
-      },
       {
         key: "subscription",
         label: "会员订阅",

apps/web/src/pages/admin/Admin.css

@@ -1,27 +0,0 @@
-/* Admin 页面样式（Phase 3 精简）
- *
- * 原始 477 行 → 精简至仅保留实际使用的 class。
- * 已迁移至 global.css / ui.css 的样式不再重复定义：
- *   .xx-page-head       → global.css
- *   .xx-primary-btn     → global.css
- *   .xx-tag / .xx-card  → ui.css / global.css
- *
- * 以下 class 仅被 AdminComingSoon.tsx 使用。
- */
-
-.admin-coming-soon-page {
-  padding: 32px;
-  max-width: 1680px;
-  margin: 0 auto;
-}
-
-.xx-result-center {
-  display: flex;
-  justify-content: center;
-  align-items: center;
-  min-height: 400px;
-}
-
-.xx-result-center .ant-result {
-  padding: 48px;
-}

apps/web/src/pages/admin/AdminComingSoon.tsx

@@ -1,35 +0,0 @@
-import React from "react"
-import { Button, Card, Result } from "antd"
-import { useNavigate } from "react-router-dom"
-import "./Admin.css"
-
-const AdminComingSoon: React.FC = () => {
-  const navigate = useNavigate()
-
-  return (
-    <div className="admin-coming-soon-page">
-      <div className="xx-result-center">
-        <Card className="xx-card">
-          <Result
-            status="info"
-            title="Admin 后台暂未开放"
-            subTitle="当前版本未接入后台用户、监控、日志、分析等后端服务，因此不展示模拟运营数据，也不会提供假操作入口。"
-            extra={[
-              <Button
-                type="primary"
-                onClick={() => navigate("/app/dashboard")}
-                className="xx-primary-btn"
-              >
-                返回首页
-              </Button>,
-            ]}
-          />
-        </Card>
-      </div>
-    </div>
-  )
-}
-
-export default AdminComingSoon
-
-export const Component = AdminComingSoon

apps/web/src/pages/generate/GeneratePage.tsx

@@ -38,6 +38,8 @@ const GeneratePage: React.FC = () => {
 
   /* ── 积分状态 ── */
   const { balance, dailyUsage, rules, init: initPoints } = usePointsStore()
+  // 后端积分总开关（由 /points/balance 返回；数据未就绪时保守按 UI 开关处理）
+  const creditsEnabled = balance?.credits_enabled ?? ENABLE_CREDIT_SYSTEM
   useEffect(() => {
     initPoints()
   }, [initPoints])
@@ -342,7 +344,8 @@ const GeneratePage: React.FC = () => {
   const handleConfirmGenerate = useCallback(async () => {
     // 积分预检查（积分系统关闭时跳过，直接走生成流程）
     let check: ReturnType<typeof hasEnoughPoints> = { sufficient: true, cost: 0 }
-    if (ENABLE_CREDIT_SYSTEM) {
+    // 仅当后端积分系统真实启用时才做余额预校验（后端免费放行时前端不得拦截）
+    if (creditsEnabled) {
       const units = isBatch ? Math.max(selectedVariantIds.length, 1) : 1
       check = hasEnoughPoints(
         balance ?? null,
@@ -388,6 +391,7 @@ const GeneratePage: React.FC = () => {
     balance,
     dailyUsage,
     rules,
+    creditsEnabled,
   ])
 
   /* ── 步骤导航 ── */
@@ -502,7 +506,7 @@ const GeneratePage: React.FC = () => {
   /* ── 积分消耗估算（步骤3确认生成展示用） ── */
   const unitsForCost = isBatch ? Math.max(selectedVariantIds.length, 1) : 1
   const pointsEstimate = useMemo(() => {
-    if (!ENABLE_CREDIT_SYSTEM) return { sufficient: true, cost: 0 }
+    if (!creditsEnabled) return { sufficient: true, cost: 0 }
     return hasEnoughPoints(
       balance ?? null,
       unitsForCost,
@@ -511,8 +515,8 @@ const GeneratePage: React.FC = () => {
       "free",
       rules?.free_user_multiplier ?? 1.15,
     )
-  }, [unitsForCost, balance, dailyUsage, rules])
-  const insufficientPoints = ENABLE_CREDIT_SYSTEM && !pointsEstimate.sufficient
+  }, [unitsForCost, balance, dailyUsage, rules, creditsEnabled])
+  const insufficientPoints = creditsEnabled && !pointsEstimate.sufficient
 
   /* ================================================================
      渲染

apps/web/src/pages/viral-video/ViralVideoPage.constants.ts

@@ -0,0 +1,128 @@
+import type { ViralVideoModel } from "@/api/viral-video/types"
+
+export const LANGUAGES = ["中文(普通话)", "粤语", "英语", "日语", "韩语"]
+export const TARGET_CUSTOMERS = [
+  "小镇青年",
+  "小镇中老年",
+  "Z世代",
+  "精致妈妈",
+  "新锐白领",
+  "资深中产",
+  "都市蓝领",
+  "都市银发",
+]
+export const PERSONAS = [
+  "通用个人IP",
+  "老板型IP",
+  "专家型IP",
+  "顾问型IP",
+  "创始人IP",
+  "创业者IP",
+  "从业者经验派",
+  "避坑顾问型",
+  "知识科普型",
+  "测评种草型",
+]
+export const STRUCTURES = [
+  "反差破局+亮明观点+还原现状",
+  "痛点场景植入+给出通关秘籍",
+  "反常识开场+扎心现状+制造焦虑",
+  "核心难题+化解焦虑+描绘愿景",
+  "痛点前置+前后对比+给出期许",
+  "悬念拉满+揭露危害+明确目的",
+  "情绪共鸣+勾起好奇+拔高价值",
+  "颠覆认知+现状拆解+赋予价值",
+  "结果先行+消除顾虑+保留悬念",
+  "真实困境+认知反转+传授秘籍",
+  "反差事实+案例佐证+戳破真相",
+  "预埋悬念+标杆案例+引出核心",
+  "危机渲染+放大痛点+输出干货",
+  "逆袭开场+成功诀窍+低价引流",
+  "反常识观点+分析危害+描述现状",
+  "正反对比+解析原因+简易方法",
+  "真实场景+内心刻画+共情引导+展望未来",
+  "连续否定+代入感受+打破认知+精准戳痛",
+  "揭秘行业+类比施压+放大焦虑+精准锁客",
+]
+export const PURPOSES = [
+  "获客引流",
+  "账号涨粉",
+  "活动通知",
+  "口播带货",
+  "促销转化",
+  "场景种草",
+  "痛点解决",
+  "功能演示",
+  "门店发现",
+  "到店实录",
+  "招牌体验",
+  "同城团购",
+  "品牌主张",
+  "品牌故事",
+  "生活方式",
+  "创意概念",
+  "对话短剧",
+  "反转短剧",
+  "悬念短剧",
+  "情绪短片",
+]
+export 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）" },
+]
+export const DURATIONS = Array.from({ length: 16 }, (_, i) => 15 + i)
+/** 兜底模型列表（接口未返回时使用） */
+export const FALLBACK_VIDEO_MODELS: ViralVideoModel[] = [
+  {
+    key: "seedance-2.5",
+    display_name: "Seedance 2.5 — 最新最强",
+    supports_audio: true,
+    supported_resolutions: ["480p", "720p", "1080p"],
+    max_duration: 30,
+    is_default: true,
+  },
+  {
+    key: "seedance-2.0",
+    display_name: "Seedance 2.0 — 正式首选",
+    supports_audio: true,
+    supported_resolutions: ["480p", "720p", "1080p"],
+    max_duration: 15,
+  },
+  {
+    key: "seedance-2.0-fast",
+    display_name: "Seedance 2.0 Fast — 快速测试",
+    supports_audio: true,
+    supported_resolutions: ["480p", "720p"],
+    max_duration: 15,
+  },
+  {
+    key: "seedance-2.0-mini",
+    display_name: "Seedance 2.0 Mini — 低成本",
+    supports_audio: true,
+    supported_resolutions: ["480p", "720p"],
+    max_duration: 15,
+  },
+  {
+    key: "wan-3.0",
+    display_name: "Wan 3.0 — 通义万相全能",
+    supports_audio: true,
+    supported_resolutions: ["480p", "720p", "1080p"],
+    max_duration: 15,
+  },
+]
+export const RESOLUTION_ORDER = ["480p", "720p", "1080p", "4k"]
+export const QUALITY_OPTIONS = [
+  { v: "480p", label: "480p（快速）" },
+  { v: "720p", label: "720p（清晰）" },
+  { v: "1080p", label: "1080p（高清）" },
+]
+export const UPSCALE_OPTIONS = [
+  { v: "default", label: "默认" },
+  { v: "sd", label: "普清" },
+  { v: "hd", label: "高清" },
+  { v: "uhd", label: "超清" },
+  { v: "omnipotent", label: "全能" },
+]
+export const VOICE_TIPS =
+  "支持 MP3/WAV/M4A/AAC/OGG 格式，最大 10MB，时长 ≤30 秒。建议清晰人声、无背景音乐、环境安静；录音请保持距麦克风 15-20cm，音量适中。"

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

@@ -1020,6 +1020,11 @@
   border-radius: 10px;
   padding: 10px 12px;
   margin-top: 8px;
+  display: flex;
+  flex-direction: column;
+  flex: 1 1 auto;
+  min-height: 260px;
+  overflow: hidden;
 }
 .vv-copy-head {
   display: flex;
@@ -1050,7 +1055,8 @@
   flex-direction: column;
   align-items: center;
   justify-content: center;
-  height: 360px;
+  flex: 1 1 auto;
+  min-height: 260px;
   padding: 28px 16px;
   gap: 10px;
   background: #fff;
@@ -1087,7 +1093,8 @@
 .vv-storyboard {
   display: flex;
   flex-direction: column;
-  height: 360px;
+  flex: 1 1 auto;
+  min-height: 260px;
   padding: 10px 12px;
   background: #fff;
   border: 1px solid #e5e7eb;
@@ -1440,7 +1447,8 @@
 /* ── Preview ──────────────── */
 .vv-preview {
   aspect-ratio: 9 / 16;
-  max-height: 360px;
+  max-height: 100%;
+  max-width: 100%;
   background: #fff;
   border: 1px dashed #e5e7eb;
   border-radius: 10px;
@@ -1973,3 +1981,294 @@
   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;
+}
+
+/* ── 分镜脚本流式预览（WebSocket delta 实时输出） ── */
+.vv-streaming-box {
+  outline: 2px solid #ede9fe;
+  background: linear-gradient(180deg, #fafafe 0%, #ffffff 100%);
+}
+.vv-streaming-header {
+  display: flex;
+  align-items: center;
+  gap: 8px;
+  padding: 4px 0 10px 0;
+  border-bottom: 1px dashed #ede9fe;
+  margin-bottom: 8px;
+  flex-shrink: 0;
+}
+.vv-streaming-title {
+  font-size: 13px;
+  color: #7c3aed;
+  font-weight: 500;
+  flex: 1 1 auto;
+}
+.vv-streaming-count {
+  font-size: 12px;
+  color: #9ca3af;
+  background: #f5f0ff;
+  padding: 1px 8px;
+  border-radius: 10px;
+}
+.vv-streaming-doc {
+  flex: 1 1 auto;
+  overflow-y: auto;
+  padding: 4px 2px 8px 2px;
+  /* 复用 vv-sb-doc 的滚动条样式 */
+  scrollbar-width: thin;
+  scrollbar-color: #d8c4ff transparent;
+}
+.vv-streaming-doc::-webkit-scrollbar {
+  width: 6px;
+}
+.vv-streaming-doc::-webkit-scrollbar-thumb {
+  background: #d8c4ff;
+  border-radius: 3px;
+}
+.vv-streaming-pre {
+  margin: 0;
+  padding: 4px 6px;
+  white-space: pre-wrap;
+  word-break: break-word;
+  font-family: ui-monospace, SFMono-Regular, Menlo, Consolas, "Liberation Mono", monospace;
+  font-size: 13px;
+  line-height: 1.65;
+  color: #1f2937;
+  tab-size: 2;
+}
+.vv-streaming-caret {
+  display: inline-block;
+  width: 2px;
+  height: 14px;
+  background: #7c3aed;
+  vertical-align: middle;
+  margin-left: 2px;
+  animation: vv-caret-blink 1s steps(2, start) infinite;
+}
+@keyframes vv-caret-blink {
+  to {
+    visibility: hidden;
+  }
+}
+.vv-streaming-actions {
+  flex-shrink: 0;
+  background: #fff;
+  padding-top: 8px;
+  border-top: 1px solid #f3f4f6;
+}
+.vv-streaming-hint {
+  font-size: 12px;
+  color: #9ca3af;
+}
+
+/* ── 三列等高布局（2026-10-10 fix） ──
+ * vv-page 作为纵向 flex 容器，tabs 固定 + grid 占满剩余高度；
+ * vv-grid 改用 grid-auto-rows:1fr 强制三列等高；
+ * 每列的卡片使用 flex column + min-height:0，正文区域 flex:1 并内部滚动。
+ */
+.vv-page {
+  display: flex;
+  flex-direction: column;
+  /* 减去顶部导航 56px 和自身 padding 16*2=32px，给 Tabs + 三列区留一屏 */
+  height: calc(100vh - 56px);
+  min-height: 640px; /* 小窗口保底 */
+  overflow: hidden;
+}
+.vv-tabs {
+  flex-shrink: 0;
+}
+.vv-grid {
+  flex: 1 1 auto;
+  min-height: 0;
+  align-items: stretch; /* 覆盖原来的 start，强制行内格子撑满行高 */
+  grid-auto-rows: 1fr; /* 每行所有格子等高 */
+  overflow: hidden;
+}
+.vv-col {
+  display: flex;
+  flex-direction: column;
+  gap: 12px;
+  min-height: 0;
+  min-width: 0;
+}
+/* 列内的卡片（.vv-section）默认允许收缩，正文区域接管滚动 */
+.vv-col > .vv-section {
+  display: flex;
+  flex-direction: column;
+  min-height: 0;
+  flex-shrink: 1;
+}
+/* STEP1/STEP2 唯一卡片：撑满列高 */
+.vv-col > .vv-section:only-child {
+  flex: 1 1 auto;
+}
+/* STEP3 右侧两卡片：配置固定不缩，预览区吸剩余空间 */
+.vv-col > .vv-section-preview {
+  flex: 1 1 auto;
+  min-height: 0;
+}
+.vv-section-preview .vv-section-body.vv-preview-body {
+  flex: 1 1 auto;
+  min-height: 0;
+  display: flex;
+  align-items: center;
+  justify-content: center;
+  padding: 12px;
+  overflow: hidden;
+}
+.vv-section-preview .vv-preview {
+  max-height: 100%;
+  height: 100%;
+  aspect-ratio: auto;
+  width: 100%;
+}
+
+/* 列内 .vv-credits / .vv-error / 主按钮 不参与收缩，永远固定在底部可见 */
+.vv-col > .vv-credits,
+.vv-col > .vv-error,
+.vv-col > .vv-btn-primary {
+  flex-shrink: 0;
+}
+.vv-col > .vv-credits {
+  margin-top: 0;
+}
+
+/* 可滚动 section 正文 */
+.vv-section-body.vv-section-body-scroll {
+  flex: 1 1 auto;
+  min-height: 0;
+  display: flex;
+  flex-direction: column;
+  padding: 14px 16px 14px;
+  overflow: hidden;
+}
+.vv-scroll-inner {
+  flex: 1 1 auto;
+  min-height: 0;
+  overflow-y: auto;
+  padding-right: 6px;
+  display: flex;
+  flex-direction: column;
+  /* 美化滚动条（紫白主题） */
+  scrollbar-width: thin;
+  scrollbar-color: #d8c4ff transparent;
+}
+.vv-scroll-inner::-webkit-scrollbar {
+  width: 6px;
+}
+.vv-scroll-inner::-webkit-scrollbar-thumb {
+  background: #d8c4ff;
+  border-radius: 3px;
+}
+.vv-scroll-inner::-webkit-scrollbar-thumb:hover {
+  background: #c4a8ff;
+}
+.vv-scroll-inner::-webkit-scrollbar-track {
+  background: transparent;
+}
+
+/* 主按钮始终顶在 section-body 底部（在 vv-scroll-inner 之外） */
+.vv-section-body.vv-section-body-scroll > .vv-btn-primary {
+  margin-top: 10px;
+  flex-shrink: 0;
+}
+
+/* STEP1 文件列表改为根据可用空间自适应，不再用 max-height:260px 硬编码 */
+.vv-file-list {
+  max-height: none;
+}

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

@@ -0,0 +1,33 @@
+import React from "react"
+
+export interface EmptyBlockProps {
+  icon?: React.ReactNode
+  title?: string
+  desc?: React.ReactNode
+  action?: React.ReactNode
+  className?: string
+  style?: React.CSSProperties
+}
+
+/**
+ * 统一空态组件：icon + title + desc + action，避免散落的 vv-muted 空文案
+ */
+export const EmptyBlock: React.FC<EmptyBlockProps> = ({
+  icon = "🎬",
+  title,
+  desc,
+  action,
+  className,
+  style,
+}) => {
+  return (
+    <div className={`vv-empty-block ${className || ""}`} style={style}>
+      <div className="vv-empty-icon">{icon}</div>
+      {title && <div className="vv-empty-title">{title}</div>}
+      {desc && <div className="vv-empty-desc vv-muted">{desc}</div>}
+      {action && <div className="vv-empty-action">{action}</div>}
+    </div>
+  )
+}
+
+export default EmptyBlock

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

@@ -0,0 +1,43 @@
+import React from "react"
+import { WarningFilled, ReloadOutlined } from "@ant-design/icons"
+
+export interface ErrorRetryBarProps {
+  label?: string
+  message: string
+  onRetry?: () => void
+  retryText?: string
+  className?: string
+  style?: React.CSSProperties
+}
+
+/**
+ * 统一错误条：警告图标 + 标签前缀 + 错误消息 + 可选重试按钮
+ * 替换主页面里散落的 .vv-error 块
+ */
+export const ErrorRetryBar: React.FC<ErrorRetryBarProps> = ({
+  label,
+  message,
+  onRetry,
+  retryText = "重试",
+  className,
+  style,
+}) => {
+  return (
+    <div className={`vv-error ${className || ""}`} style={style}>
+      <WarningFilled style={{ marginRight: 6 }} />
+      {label ? `${label}：` : ""}
+      {message}
+      {onRetry && (
+        <button
+          className="vv-btn vv-btn-warn vv-btn-sm"
+          style={{ marginLeft: 10 }}
+          onClick={onRetry}
+        >
+          <ReloadOutlined /> {retryText}
+        </button>
+      )}
+    </div>
+  )
+}
+
+export default ErrorRetryBar

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

@@ -0,0 +1,111 @@
+import React, { useState, useRef, useEffect, useCallback } from "react"
+
+export interface InlineTextProps {
+  value: string
+  onChange?: (v: string) => void // 受控模式
+  onCommit: (v: string) => void // blur/Enter 提交
+  onCancel?: () => void
+  multiline?: boolean
+  textareaClass?: string
+  textClass?: string
+  placeholder?: string
+  editing: boolean
+  onStartEdit: () => void
+  rows?: number
+  maxLength?: number
+  disabled?: boolean
+}
+
+/**
+ * 通用"点文字→textarea→blur/Enter保存→Esc取消"内联编辑组件
+ * 替换 ViralVideoPage 里 11 处反射式 textarea/span 复制粘贴
+ */
+export const InlineText: React.FC<InlineTextProps> = ({
+  value,
+  onChange,
+  onCommit,
+  onCancel,
+  multiline,
+  textareaClass,
+  textClass,
+  placeholder,
+  editing,
+  onStartEdit,
+  rows = 2,
+  maxLength,
+  disabled,
+}) => {
+  const taRef = useRef<HTMLTextAreaElement>(null)
+  const [local, setLocal] = useState(value)
+
+  useEffect(() => {
+    setLocal(value)
+  }, [value])
+
+  useEffect(() => {
+    if (editing && taRef.current) {
+      taRef.current.focus()
+      taRef.current.select()
+    }
+  }, [editing])
+
+  const commit = useCallback(() => {
+    const v = (onChange ? value : local).trim()
+    const cur = (value ?? "").trim()
+    if (v !== cur) onCommit(v)
+    else onCancel?.()
+  }, [local, value, onCommit, onCancel, onChange])
+
+  const handleKey = useCallback(
+    (e: React.KeyboardEvent) => {
+      if (e.key === "Escape") {
+        e.preventDefault()
+        onCancel?.()
+      }
+      if (!multiline && e.key === "Enter" && !e.shiftKey) {
+        e.preventDefault()
+        commit()
+      }
+    },
+    [multiline, commit, onCancel],
+  )
+
+  if (editing) {
+    return (
+      <textarea
+        ref={taRef}
+        className={textareaClass}
+        value={onChange ? value : local}
+        onChange={(e) => (onChange ? onChange(e.target.value) : setLocal(e.target.value))}
+        onBlur={commit}
+        onKeyDown={handleKey}
+        rows={rows}
+        maxLength={maxLength}
+        placeholder={placeholder}
+      />
+    )
+  }
+
+  const displayText = value || placeholder || "点击编辑"
+  const isPlaceholder = !value
+
+  return (
+    <span
+      className={`${textClass || ""} ${isPlaceholder ? "vv-inline-placeholder" : ""} ${
+        disabled ? "vv-inline-disabled" : ""
+      }`}
+      onClick={() => {
+        if (!disabled) onStartEdit()
+      }}
+      style={
+        isPlaceholder
+          ? { color: "#9ca3af", cursor: disabled ? "not-allowed" : "text" }
+          : { cursor: disabled ? "not-allowed" : "text" }
+      }
+    >
+      {displayText}
+    </span>
+  )
+}
+
+export default InlineText

apps/web/src/pages/viral-video/components/Step1UploadPanel.tsx

@@ -0,0 +1,347 @@
+import React, { useMemo } from "react"
+import {
+  UploadOutlined,
+  FolderOpenOutlined,
+  ArrowUpOutlined,
+  ArrowDownOutlined,
+  CloseOutlined,
+  LinkOutlined,
+} from "@ant-design/icons"
+import { Input } from "antd"
+import { VVSpinner } from "./VVSpinner"
+import { renderMarkdown } from "../utils/renderMarkdown"
+import { CheckCircleFilled } from "@ant-design/icons"
+import { VoicePickerSection } from "./VoicePickerSection"
+import type { UploadedImage, RefVideo, RefAudio, VoicePanel, ImageAnalysisResult } from "../types"
+import type { ImageProductAnalysis, StyleStrength } from "@/api/viral-video/types"
+import { STYLE_STRENGTHS } from "@/api/viral-video/types"
+import type { VoiceClone } from "@/api/voice-clone"
+import type { AssetItem } from "@/api/assets"
+
+export interface Step1UploadPanelProps {
+  images: UploadedImage[]
+  refVideo: RefVideo
+  douyinInput: string
+  voicePanel: VoicePanel
+  refAudio: RefAudio
+  voiceTipsOpen: boolean
+  myVoices: VoiceClone[]
+  myVoicesLoading: boolean
+  voiceAssets: { id: string; name: string; url: string; desc?: string }[]
+  playingVoiceId: string | null
+  styleStrength: StyleStrength
+  imageAnalysis: ImageAnalysisResult | null
+  isAnalyzing: boolean
+  analysisError?: string
+  canAnalyze: boolean
+  analyzeBtnText: React.ReactNode
+  // callbacks - images
+  onPickLocalImages: () => void
+  onPickAssetImages: () => void
+  onRemoveImage: (localId: string) => void
+  onMoveImage: (idx: number, dir: -1 | 1) => void
+  onClearImages: () => void
+  onImagesSelected: (files: FileList) => void
+  onAnalyze: () => void
+  // callbacks - ref video
+  onPickLocalVideo: () => void
+  onToggleDouyinPanel: () => void
+  onDouyinInputChange: (v: string) => void
+  onDouyinSubmit: () => void
+  onVideoSelected: (file: File) => void
+  onRemoveRefVideo: () => void
+  onStyleStrengthChange: (v: StyleStrength) => void
+  // callbacks - voice (delegated to VoicePickerSection)
+  onPickLocalAudio: () => void
+  onPickAssetAudio: () => void
+  onRemoveRefAudio: () => void
+  onToggleVoicePanel: (p: VoicePanel) => void
+  onSelectMyVoice: (v: VoiceClone) => void
+  onOpenCloneModal: () => void
+  onSelectVoiceAsset: (a: AssetItem) => void
+  onToggleVoicePlay: (v: { id: string; url: string }) => void
+  onVoiceSelected: (file: File) => void
+  onToggleVoiceTips: () => void
+  // refs
+  imgInputRef: React.Ref<HTMLInputElement>
+  videoInputRef: React.Ref<HTMLInputElement>
+  voiceInputRef: React.Ref<HTMLInputElement>
+}
+
+const fmtSize = (bytes: number | undefined) => {
+  if (!bytes || bytes <= 0) return "--"
+  if (bytes < 1024) return `${bytes}B`
+  if (bytes < 1024 * 1024) return `${(bytes / 1024).toFixed(1)}KB`
+  return `${(bytes / 1024 / 1024).toFixed(1)}MB`
+}
+
+function RecognitionBlock({
+  imageAnalysis,
+  isAnalyzing,
+}: {
+  imageAnalysis: ImageAnalysisResult | null
+  isAnalyzing: boolean
+}) {
+  const images: ImageProductAnalysis[] = useMemo(() => {
+    return (
+      (imageAnalysis?.images as ImageProductAnalysis[] | undefined) ||
+      (imageAnalysis?.products as ImageProductAnalysis[] | undefined) ||
+      []
+    )
+  }, [imageAnalysis])
+
+  if (isAnalyzing) {
+    return (
+      <div className="vv-recog">
+        <div className="vv-recog-title">
+          <VVSpinner
+            color="#7c3aed"
+            style={{ marginRight: 6, display: "inline-block", verticalAlign: "middle" }}
+          />
+          识别描述汇览
+        </div>
+        <div className="vv-muted">AI 正在识别画面…</div>
+      </div>
+    )
+  }
+  if (images.length === 0) return null
+  return (
+    <div className="vv-recog">
+      <div className="vv-recog-title">
+        <CheckCircleFilled style={{ color: "#10b981" }} />
+        识别描述汇览
+      </div>
+      {images.map((p, i) => {
+        const meta = [p.name || "未识别", p.brand, p.category].filter(Boolean)
+        return (
+          <div key={i} className="vv-recog-item vv-recog-md">
+            <div className="vv-recog-line">
+              <span className="vv-recog-k">图片{i + 1}：</span>
+              <span>{meta.join(" · ")}</span>
+            </div>
+            {p.summary_markdown ? (
+              <div
+                className="vv-md-body"
+                dangerouslySetInnerHTML={{ __html: renderMarkdown(p.summary_markdown) }}
+              />
+            ) : (
+              <div className="vv-muted">（暂无叙述描述）</div>
+            )}
+          </div>
+        )
+      })}
+    </div>
+  )
+}
+
+export const Step1UploadPanel: React.FC<Step1UploadPanelProps> = (p) => {
+  return (
+    <div className="vv-section">
+      <div className="vv-step1-head">
+        <span className="vv-step-badge">1</span>
+        <span className="vv-step1-title">STEP 1 上传素材并分析图片</span>
+      </div>
+      <div className="vv-section-body vv-section-body-scroll">
+        <div className="vv-scroll-inner">
+          {/* 图片按钮 */}
+          <div className="vv-action-row">
+            <button className="vv-action-btn" onClick={p.onPickLocalImages}>
+              <UploadOutlined /> 本地上传
+            </button>
+            <button className="vv-action-btn" onClick={p.onPickAssetImages}>
+              <FolderOpenOutlined /> 从素材库选择图片
+            </button>
+          </div>
+          <input
+            ref={p.imgInputRef}
+            type="file"
+            multiple
+            accept="image/*"
+            hidden
+            onChange={(e) => {
+              if (e.target.files) p.onImagesSelected(e.target.files)
+              e.target.value = ""
+            }}
+          />
+
+          <div className="vv-count-row">
+            <span className="vv-muted">
+              已选 <strong style={{ color: "#111827" }}>{p.images.length}</strong>/9
+            </span>
+            {p.images.length > 0 && (
+              <button className="vv-link-btn" onClick={p.onClearImages}>
+                清空图片
+              </button>
+            )}
+          </div>
+
+          {p.images.length > 0 && (
+            <div className="vv-file-list">
+              {p.images.map((im, i) => (
+                <div key={im.localId} className="vv-file-item">
+                  <div className="vv-file-thumb">
+                    <span className="vv-img-badge">{i + 1}</span>
+                    <img src={im.ossUrl || im.url} alt={im.name} />
+                  </div>
+                  <div className="vv-file-info">
+                    <div className="vv-file-name" title={im.name}>
+                      {im.name}
+                    </div>
+                    <div className="vv-file-sub">
+                      {im.uploading ? (
+                        <>
+                          <span className="vv-progress-mini-inline" />
+                          上传中 {im.pct}%
+                        </>
+                      ) : (
+                        <>{fmtSize(im.size)}</>
+                      )}
+                    </div>
+                  </div>
+                  <div className="vv-file-actions">
+                    <button
+                      className="vv-icon-btn"
+                      onClick={() => p.onMoveImage(i, -1)}
+                      disabled={i === 0 || im.uploading}
+                      title="上移"
+                    >
+                      <ArrowUpOutlined />
+                    </button>
+                    <button
+                      className="vv-icon-btn"
+                      onClick={() => p.onMoveImage(i, 1)}
+                      disabled={i === p.images.length - 1 || im.uploading}
+                      title="下移"
+                    >
+                      <ArrowDownOutlined />
+                    </button>
+                    <button
+                      className="vv-icon-btn vv-icon-btn-danger"
+                      onClick={() => p.onRemoveImage(im.localId)}
+                      disabled={!!im.uploading}
+                      title="删除"
+                    >
+                      <CloseOutlined />
+                    </button>
+                  </div>
+                </div>
+              ))}
+            </div>
+          )}
+
+          {/* 参考视频 */}
+          <div className="vv-subblock">
+            <div className="vv-subblock-head">
+              <span className="vv-subblock-title">参考视频</span>
+              <span className="vv-opt-tag">选填</span>
+              <span className="vv-subblock-hint">最大 200MB · 最长 120 秒</span>
+            </div>
+            <div className="vv-action-row">
+              <button className="vv-action-btn" onClick={p.onPickLocalVideo}>
+                <UploadOutlined /> 本地上传视频
+              </button>
+              <button className="vv-action-btn" onClick={p.onToggleDouyinPanel}>
+                <LinkOutlined /> 抖音链接识别
+              </button>
+            </div>
+            <input
+              ref={p.videoInputRef}
+              type="file"
+              accept="video/*"
+              hidden
+              onChange={(e) => {
+                if (e.target.files?.[0]) p.onVideoSelected(e.target.files[0])
+                e.target.value = ""
+              }}
+            />
+            {p.voicePanel === "douyin" && !p.refVideo && (
+              <div className="vv-douyin-input">
+                <Input
+                  placeholder="粘贴抖音分享链接，例：https://v.douyin.com/xxxxx/"
+                  value={p.douyinInput}
+                  onChange={(e) => p.onDouyinInputChange(e.target.value)}
+                  onPressEnter={p.onDouyinSubmit}
+                  suffix={
+                    <button className="vv-link-btn" onClick={p.onDouyinSubmit}>
+                      识别
+                    </button>
+                  }
+                />
+                <div className="vv-muted" style={{ marginTop: 4, fontSize: 11 }}>
+                  粘贴后自动解析下载参考视频（敬请期待）
+                </div>
+              </div>
+            )}
+            {p.refVideo && (
+              <div className="vv-ref-video-box">
+                <video
+                  src={p.refVideo.ossUrl || p.refVideo.url}
+                  controls
+                  className="vv-ref-video-el"
+                />
+                <div className="vv-file-info" style={{ marginTop: 6 }}>
+                  <div className="vv-file-name">{p.refVideo.name}</div>
+                  {p.refVideo.uploading && (
+                    <div className="vv-progress-bar" style={{ margin: "6px 0 4px" }}>
+                      <div className="vv-progress-fill" style={{ width: `${p.refVideo.pct}%` }} />
+                    </div>
+                  )}
+                </div>
+                <button className="vv-link-btn" onClick={p.onRemoveRefVideo}>
+                  删除重传
+                </button>
+                {p.refVideo.ossUrl && (
+                  <div className="vv-mt-8">
+                    <div className="vv-label" style={{ marginBottom: 6 }}>
+                      风格借鉴强度
+                    </div>
+                    <div className="vv-pill-row">
+                      {STYLE_STRENGTHS.map((s) => (
+                        <button
+                          key={s.value}
+                          className={`vv-pill ${p.styleStrength === s.value ? "active" : ""}`}
+                          onClick={() => p.onStyleStrengthChange(s.value)}
+                        >
+                          {s.label}
+                        </button>
+                      ))}
+                    </div>
+                  </div>
+                )}
+              </div>
+            )}
+          </div>
+
+          <VoicePickerSection
+            refAudio={p.refAudio}
+            voicePanel={p.voicePanel}
+            voiceTipsOpen={p.voiceTipsOpen}
+            myVoices={p.myVoices}
+            myVoicesLoading={p.myVoicesLoading}
+            voiceAssets={p.voiceAssets}
+            playingVoiceId={p.playingVoiceId}
+            voiceInputRef={p.voiceInputRef}
+            onOpenCloneModal={p.onOpenCloneModal}
+            onToggleVoicePanel={p.onToggleVoicePanel}
+            onSelectMyVoice={p.onSelectMyVoice}
+            onSelectVoiceAsset={p.onSelectVoiceAsset}
+            onToggleVoicePlay={p.onToggleVoicePlay}
+            onPickLocalAudio={p.onPickLocalAudio}
+            onPickAssetAudio={p.onPickAssetAudio}
+            onVoiceSelected={p.onVoiceSelected}
+            onRemoveRefAudio={p.onRemoveRefAudio}
+            onToggleVoiceTips={p.onToggleVoiceTips}
+          />
+
+          <RecognitionBlock imageAnalysis={p.imageAnalysis} isAnalyzing={p.isAnalyzing} />
+        </div>
+
+        <button className="vv-btn vv-btn-primary" disabled={!p.canAnalyze} onClick={p.onAnalyze}>
+          {p.analyzeBtnText}
+        </button>
+      </div>
+    </div>
+  )
+}
+
+export default Step1UploadPanel

apps/web/src/pages/viral-video/components/Step2CopyPanel.tsx

@@ -0,0 +1,225 @@
+import React from "react"
+import { EditOutlined } from "@ant-design/icons"
+import { Select } from "antd"
+import { FUSION_LEVELS } from "@/api/viral-video/types"
+import type { FusionLevel } from "@/api/viral-video/types"
+import { VVSpinner } from "./VVSpinner"
+import { StoryboardCard } from "./StoryboardCard"
+import type { SbDoc, StoryboardEditor } from "../hooks/useStoryboardEditor"
+import type { UploadedImage, UIStep, ViralVideoJob } from "../types"
+import {
+  LANGUAGES,
+  TARGET_CUSTOMERS,
+  PERSONAS,
+  STRUCTURES,
+  PURPOSES,
+  DURATIONS,
+} from "../ViralVideoPage.constants"
+import { getJobPhaseMessage } from "@/api/viral-video/types"
+
+export interface Step2CopyPanelProps {
+  // 启用状态
+  step2Enabled: boolean
+  uiStep: UIStep
+  // 文案参数
+  fusionLevel: FusionLevel
+  userCopy: string
+  industry: string
+  targetCustomer: string
+  persona: string
+  language: string
+  viralStructure: string
+  marketingPurpose: string
+  duration: number
+  // 回调
+  onFusionLevelChange: (v: FusionLevel) => void
+  onUserCopyChange: (v: string) => void
+  onParamChange: (patch: {
+    industry?: string
+    targetCustomer?: string
+    persona?: string
+    language?: string
+    viralStructure?: string
+    marketingPurpose?: string
+    duration?: number
+  }) => void
+  onGenerateCopy: () => void
+  // 分镜/流式预览
+  storyboard: SbDoc | null
+  scriptStreamText: string
+  scriptStreamMessage?: string
+  editor: StoryboardEditor
+  images: UploadedImage[]
+  locked: boolean
+  job: ViralVideoJob | null
+}
+
+export const Step2CopyPanel: React.FC<Step2CopyPanelProps> = (p) => {
+  const isGenerating = p.uiStep === "step2_generating"
+  const showGenerateBtn =
+    p.uiStep !== "step2_copy_ready" && p.uiStep !== "step3_generating" && p.uiStep !== "step3_done"
+
+  return (
+    <div className="vv-section">
+      <div className="vv-step-subhead">
+        <span className={`vv-step-badge ${p.step2Enabled ? "" : "vv-step-badge-dim"}`}>2</span>
+        <span className="vv-step-subtitle">STEP 2 生成视频文案</span>
+        {!p.step2Enabled && <span className="vv-step-lock">请先完成 STEP 1</span>}
+      </div>
+      <div className="vv-section-body vv-section-body-scroll">
+        <div className="vv-scroll-inner">
+          {/* 文案融合方式 */}
+          <div className="vv-section-subtitle">文案融合方式</div>
+          <div className="vv-fusion-grid">
+            {FUSION_LEVELS.map((f) => (
+              <button
+                key={f.value}
+                className={`vv-fusion-btn ${p.fusionLevel === f.value ? "active" : ""}`}
+                onClick={() => p.onFusionLevelChange(f.value)}
+              >
+                <strong>{f.label}</strong>
+                {f.desc}
+              </button>
+            ))}
+          </div>
+          {(p.fusionLevel === "ai_polish" || p.fusionLevel === "user_primary") && (
+            <div className="vv-form-row">
+              <label className="vv-label">
+                {p.fusionLevel === "ai_polish" ? "写下你的文案草稿" : "写下你的文案/草稿"}
+              </label>
+              <textarea
+                key={p.fusionLevel}
+                className="vv-textarea"
+                placeholder={
+                  p.fusionLevel === "ai_polish"
+                    ? "请输入你的文案草稿，AI将在此基础上润色优化..."
+                    : "请输入你想要的原始文案，AI将严格保留你的表达..."
+                }
+                value={p.userCopy}
+                onChange={(e) => p.onUserCopyChange(e.target.value)}
+              />
+            </div>
+          )}
+
+          {/* 参数（2列网格） */}
+          <div className="vv-form-grid">
+            <div className="vv-form-row">
+              <label className="vv-label">
+                行业与产品 <span className="vv-label-optional">（选填）</span>
+              </label>
+              <input
+                className="vv-input"
+                placeholder="如：家具店、宠物店、装修公司"
+                value={p.industry}
+                onChange={(e) => p.onParamChange({ industry: e.target.value })}
+              />
+            </div>
+            <div className="vv-form-row">
+              <label className="vv-label">目标客户</label>
+              <Select
+                className="vv-select"
+                style={{ width: "100%" }}
+                placeholder="请选择目标客户"
+                value={p.targetCustomer || undefined}
+                onChange={(v) => p.onParamChange({ targetCustomer: v })}
+                options={TARGET_CUSTOMERS.map((i) => ({ value: i, label: i }))}
+                allowClear
+              />
+            </div>
+            <div className="vv-form-row">
+              <label className="vv-label">台词语言</label>
+              <Select
+                className="vv-select"
+                style={{ width: "100%" }}
+                value={p.language}
+                onChange={(v) => p.onParamChange({ language: v })}
+                options={LANGUAGES.map((i) => ({ value: i, label: i }))}
+              />
+            </div>
+            <div className="vv-form-row">
+              <label className="vv-label">我的人设</label>
+              <Select
+                className="vv-select"
+                style={{ width: "100%" }}
+                placeholder="请选择人设类型"
+                optionFilterProp="children"
+                value={p.persona || undefined}
+                onChange={(v) => p.onParamChange({ persona: v })}
+                options={PERSONAS.map((i) => ({ value: i, label: i }))}
+              />
+            </div>
+            <div className="vv-form-row">
+              <label className="vv-label">爆款结构</label>
+              <Select
+                className="vv-select"
+                style={{ width: "100%" }}
+                value={p.viralStructure}
+                onChange={(v) => p.onParamChange({ viralStructure: v })}
+                options={STRUCTURES.map((i) => ({ value: i, label: i }))}
+              />
+            </div>
+            <div className="vv-form-row">
+              <label className="vv-label">营销目的</label>
+              <Select
+                className="vv-select"
+                style={{ width: "100%" }}
+                placeholder="请选择营销目的"
+                value={p.marketingPurpose || undefined}
+                onChange={(v) => p.onParamChange({ marketingPurpose: v })}
+                options={PURPOSES.map((i) => ({ value: i, label: i }))}
+              />
+            </div>
+            <div className="vv-form-row">
+              <label className="vv-label">文案视频时长</label>
+              <Select
+                className="vv-select"
+                style={{ width: "100%" }}
+                value={p.duration}
+                onChange={(v) => p.onParamChange({ duration: v })}
+                options={DURATIONS.map((n) => ({ value: n, label: `${n}秒` }))}
+              />
+            </div>
+          </div>
+
+          {/* 分镜结果区 / 流式预览 */}
+          <StoryboardCard
+            uiStep={p.uiStep}
+            scriptStreamText={p.scriptStreamText}
+            scriptStreamMessage={p.scriptStreamMessage}
+            jobPhaseMessage={getJobPhaseMessage(p.job)}
+            storyboard={p.storyboard}
+            editor={p.editor}
+            images={p.images}
+            locked={p.locked}
+            onRegenerateCopy={p.onGenerateCopy}
+          />
+        </div>
+
+        {showGenerateBtn && (
+          <button
+            className={`vv-btn vv-btn-primary ${!p.step2Enabled ? "vv-btn-disabled-gray" : ""}`}
+            onClick={p.onGenerateCopy}
+            disabled={isGenerating || !p.step2Enabled}
+            title={!p.step2Enabled ? "请先完成图片分析" : ""}
+          >
+            {isGenerating ? (
+              <>
+                <VVSpinner /> AI 生成文案中…
+              </>
+            ) : !p.step2Enabled ? (
+              <>
+                <EditOutlined /> 生成文案（请先完成图片分析）
+              </>
+            ) : (
+              <>
+                <EditOutlined /> 生成文案
+              </>
+            )}
+          </button>
+        )}
+      </div>
+    </div>
+  )
+}
+
+export default Step2CopyPanel

apps/web/src/pages/viral-video/components/Step3VideoPanel.tsx

@@ -0,0 +1,218 @@
+import React from "react"
+import {
+  VideoCameraOutlined,
+  SoundOutlined,
+  ReloadOutlined,
+  HistoryOutlined,
+  LoadingOutlined,
+} from "@ant-design/icons"
+import { Select } from "antd"
+import { VVSpinner } from "./VVSpinner"
+import { ErrorRetryBar } from "./ErrorRetryBar"
+import type { UIStep, ViralVideoJob, ViralVideoModel } from "../types"
+import { getJobPhaseMessage } from "@/api/viral-video/types"
+import { RATIOS, QUALITY_OPTIONS, UPSCALE_OPTIONS } from "../ViralVideoPage.constants"
+
+export interface Step3VideoPanelProps {
+  step3Enabled: boolean
+  uiStep: UIStep
+  videoRatio: string
+  videoModel: string
+  quality: string
+  upscale: string
+  duration: number
+  estimatedCredits: number | null
+  estimatedCreditsError: boolean
+  isGenerating: boolean
+  outputUrl: string | null
+  videoError?: string
+  creditsReady: boolean
+  job: ViralVideoJob | null
+  videoModels: ViralVideoModel[]
+  locked: boolean
+  // callbacks
+  onRatioChange: (v: string) => void
+  onModelChange: (v: string) => void
+  onQualityChange: (v: string) => void
+  onUpscaleChange: (v: string) => void
+  onGenerate: () => void
+  onReset: () => void
+  onRetry: () => void
+}
+
+export const Step3VideoPanel: React.FC<Step3VideoPanelProps> = (p) => {
+  const isDone = p.uiStep === "step3_done"
+  const isGenerating = p.uiStep === "step3_generating"
+  const isFailed = p.uiStep === "failed"
+
+  const curM = p.videoModels.find((m) => m && m.key === p.videoModel)
+  const supportedRes = Array.isArray(curM?.supported_resolutions)
+    ? curM!.supported_resolutions
+    : QUALITY_OPTIONS.map((o) => o.v)
+  const qualityPool =
+    supportedRes && supportedRes.length > 0 ? supportedRes : QUALITY_OPTIONS.map((o) => o.v)
+  const qualityOptions = QUALITY_OPTIONS.filter((o) => qualityPool.includes(o.v))
+
+  return (
+    <>
+      <div className="vv-section">
+        <div className="vv-step-subhead vv-step3-head">
+          <div className="vv-step3-head-left">
+            <span className={`vv-step-badge ${p.step3Enabled ? "" : "vv-step-badge-dim"}`}>3</span>
+            <span className="vv-step-subtitle">STEP 3 生成视频</span>
+          </div>
+          {p.step3Enabled && (
+            <button
+              className="vv-btn vv-btn-ghost vv-btn-xs vv-history-btn"
+              onClick={() => {
+                // 历史作品功能待开发，此处保留原行为（message 由父组件触发）
+              }}
+              title="历史作品"
+            >
+              <HistoryOutlined /> 历史作品
+            </button>
+          )}
+          {!p.step3Enabled && <span className="vv-step-lock">请先完成 STEP 2</span>}
+        </div>
+        <div className="vv-step3-divider" />
+        <div className="vv-section-body vv-section-body-scroll vv-step3-config-body">
+          <div className="vv-scroll-inner">
+            <div className="vv-step3-grid">
+              <div className="vv-form-row">
+                <label className="vv-label">视频比例</label>
+                <Select
+                  className="vv-select vv-select-step3"
+                  style={{ width: "100%" }}
+                  value={p.videoRatio}
+                  onChange={p.onRatioChange}
+                  disabled={p.locked}
+                  options={RATIOS.map((r) => ({ value: r.v, label: r.v }))}
+                />
+              </div>
+              <div className="vv-form-row">
+                <label className="vv-label">视频模型</label>
+                <Select
+                  className="vv-select vv-select-step3"
+                  style={{ width: "100%" }}
+                  value={p.videoModel}
+                  disabled={p.locked}
+                  onChange={(v) => p.onModelChange(v)}
+                  options={p.videoModels
+                    .filter((m) => m && typeof m === "object" && m.key)
+                    .map((m) => ({ value: m.key, label: m.display_name || m.key }))}
+                />
+              </div>
+              <div className="vv-form-row">
+                <label className="vv-label">清晰度</label>
+                <Select
+                  className="vv-select vv-select-step3"
+                  style={{ width: "100%" }}
+                  value={p.quality}
+                  onChange={p.onQualityChange}
+                  disabled={p.locked}
+                  options={qualityOptions.map((o) => ({ value: o.v, label: o.label }))}
+                />
+              </div>
+              <div className="vv-form-row">
+                <label className="vv-label">成片超分</label>
+                <Select
+                  className="vv-select vv-select-step3"
+                  style={{ width: "100%" }}
+                  value={p.upscale}
+                  onChange={p.onUpscaleChange}
+                  disabled={p.locked}
+                  options={UPSCALE_OPTIONS.map((m) => ({ value: m.v, label: m.label }))}
+                />
+              </div>
+            </div>
+          </div>
+        </div>
+      </div>
+
+      {p.creditsReady && (
+        <div className="vv-credits">
+          <span>
+            <SoundOutlined style={{ marginRight: 4 }} />
+            预计消耗
+          </span>
+          <strong>
+            {p.estimatedCreditsError
+              ? "-- 积分"
+              : p.estimatedCredits === null
+                ? "… 积分"
+                : `${p.estimatedCredits.toFixed(2)} 积分`}
+          </strong>
+        </div>
+      )}
+
+      {/* 视频预览 */}
+      <div className="vv-section vv-section-preview">
+        <div className="vv-section-body vv-preview-body">
+          <div className="vv-preview">
+            {isDone && p.outputUrl ? (
+              <video src={p.outputUrl} controls />
+            ) : isGenerating ? (
+              <div className="vv-preview-placeholder">
+                <div className="ph-icon">
+                  <LoadingOutlined style={{ fontSize: 36, color: "#7c3aed" }} />
+                </div>
+                <div className="ph-txt">
+                  AI 正在生成视频中…
+                  <br />
+                  预计 1-3 分钟，请勿关闭页面
+                </div>
+                {getJobPhaseMessage(p.job) && (
+                  <div className="vv-muted" style={{ marginTop: 6, fontSize: 12 }}>
+                    {getJobPhaseMessage(p.job)}
+                  </div>
+                )}
+              </div>
+            ) : (
+              <div className="vv-preview-placeholder">
+                <div className="ph-icon">🎬</div>
+                <div className="ph-txt">完成前面步骤后视频将在此处预览</div>
+              </div>
+            )}
+          </div>
+        </div>
+      </div>
+
+      {/* 错误条 */}
+      {isFailed && p.videoError && (
+        <ErrorRetryBar label="视频生成失败" message={p.videoError} onRetry={p.onRetry} />
+      )}
+
+      {/* 主按钮 */}
+      {!isDone ? (
+        <button
+          className={`vv-btn vv-btn-primary ${!p.step3Enabled ? "vv-btn-disabled-gray" : ""}`}
+          onClick={p.onGenerate}
+          disabled={isGenerating || !p.step3Enabled}
+          title={!p.step3Enabled ? "请先在 STEP 2 生成文案" : ""}
+        >
+          {isGenerating ? (
+            <>
+              <VVSpinner /> AI 正在生成视频中…
+            </>
+          ) : !p.step3Enabled ? (
+            <>
+              <VideoCameraOutlined /> 开始生成视频（请先生成文案）
+            </>
+          ) : (
+            <>
+              <VideoCameraOutlined /> 开始生成视频
+            </>
+          )}
+        </button>
+      ) : (
+        p.outputUrl && (
+          <button className="vv-btn vv-btn-ghost vv-btn-block" onClick={p.onReset}>
+            <ReloadOutlined /> 再生成一个
+          </button>
+        )
+      )}
+    </>
+  )
+}
+
+export default Step3VideoPanel

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

@@ -0,0 +1,327 @@
+import React from "react"
+import { Select } from "antd"
+import { FileTextOutlined, SoundOutlined, ReloadOutlined } from "@ant-design/icons"
+import { InlineText } from "./InlineEditTextarea"
+import { StoryboardShotItem } from "./StoryboardShotItem"
+import { VVSpinner } from "./VVSpinner"
+import type { StoryboardEditor, SbDoc } from "../hooks/useStoryboardEditor"
+import type { UploadedImage, UIStep } from "../types"
+import { RATIOS } from "../ViralVideoPage.constants"
+
+export interface StoryboardCardProps {
+  /** 当前 uiStep（仅用于按钮/锁定判定） */
+  uiStep: UIStep
+  /** 当前流式预览文本（step2_generating 时优先展示） */
+  scriptStreamText: string
+  scriptStreamMessage?: string
+  jobPhaseMessage?: string
+  /** 真实 storyboard（已补默认值的 SbDoc） */
+  storyboard: SbDoc | null
+  editor: StoryboardEditor
+  images: UploadedImage[]
+  /** 是否锁定编辑（视频生成中/完成时为 true） */
+  locked: boolean
+  onRegenerateCopy?: () => void
+}
+
+import { renderMarkdown } from "../utils/renderMarkdown"
+
+/** 解析 voiceover_script 中的 [speaker:xxx] 标记，渲染为带颜色的 React 片段 */
+function renderVoiceoverWithSpeakers(text: string) {
+  if (!text) return null
+  const parts = text.split(/\[speaker:([^\]]+)\]/)
+  if (parts.length === 1) return <span>{text}</span>
+  const elements: React.ReactNode[] = []
+  const MAIN_SPEAKERS = ["店主", "主播", "老板", "我", "主讲人"]
+  for (let i = 1; i < parts.length; i += 2) {
+    const speaker = parts[i]
+    const segText = parts[i + 1] || ""
+    const isMain = MAIN_SPEAKERS.includes(speaker)
+    elements.push(
+      <span key={i}>
+        <strong style={{ color: isMain ? "#1890ff" : "#fa8c16" }}>{speaker}</strong>
+        <span>{segText}</span>
+      </span>,
+    )
+  }
+  return <>{elements}</>
+}
+
+export const StoryboardCard: React.FC<StoryboardCardProps> = ({
+  uiStep,
+  scriptStreamText,
+  scriptStreamMessage,
+  jobPhaseMessage,
+  storyboard,
+  editor,
+  images,
+  locked,
+  onRegenerateCopy,
+}) => {
+  const {
+    editingField,
+    startEdit,
+    cancelEdit,
+    commitField,
+    updateOverview,
+    updateConstraint,
+    addConstraint,
+    removeConstraint,
+    updateSbField,
+  } = editor
+
+  /* ── 生成中：流式预览 or 骨架 ── */
+  if (uiStep === "step2_generating") {
+    const hasStream = scriptStreamText.trim().length > 0
+    const stageMsg = scriptStreamMessage || jobPhaseMessage || "🎬 AI 正在创作视频分镜脚本..."
+    if (!hasStream) {
+      return (
+        <div className="vv-copy-box vv-copy-loading">
+          <VVSpinner />
+          <span className="vv-copy-loading-text">{stageMsg}</span>
+          <span className="vv-copy-loading-hint">预计 10-30 秒，请稍候</span>
+        </div>
+      )
+    }
+    return (
+      <div className="vv-copy-box vv-storyboard vv-streaming-box">
+        <div className="vv-streaming-header">
+          <VVSpinner />
+          <span className="vv-streaming-title">{stageMsg}</span>
+          <span className="vv-streaming-count">{scriptStreamText.length} 字</span>
+        </div>
+        <div className="vv-sb-doc vv-streaming-doc">
+          <pre className="vv-streaming-pre">
+            {scriptStreamText
+              .replace(/<\/?[a-zA-Z_][\w:-]*(?:\s[^>]*)?>/g, "")
+              .replace(/\s{3,}/g, "\n\n")
+              .trim()}
+            <span className="vv-streaming-caret" aria-hidden="true" />
+          </pre>
+        </div>
+        <div className="vv-sb-actions vv-streaming-actions">
+          <span className="vv-streaming-hint">AI 实时输出中，完成后可编辑分镜与口播文案</span>
+        </div>
+      </div>
+    )
+  }
+
+  /* ── 没有 storyboard 时的空态 ── */
+  if (!storyboard) {
+    if (uiStep === "step2_copy_ready" || uiStep === "step3_generating" || uiStep === "step3_done") {
+      return (
+        <div className="vv-copy-box vv-copy-loading">
+          <span style={{ color: "#9ca3af" }}>暂无分镜数据</span>
+        </div>
+      )
+    }
+    return null
+  }
+
+  const sb = storyboard
+  const textCls = "vv-sb-field-val"
+  const taCls = "vv-textarea vv-sb-doc-ta vv-sb-inline-edit-ta"
+
+  return (
+    <div className="vv-copy-box vv-storyboard">
+      <div className="vv-sb-doc">
+        {/* 文案预览 markdown（只读） */}
+        {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">
+          <strong className="vv-sb-field-k">整体主题：</strong>
+          <InlineText
+            value={sb.overview.theme}
+            editing={!locked && editingField === "overview.theme"}
+            onStartEdit={() => startEdit("overview.theme")}
+            onCommit={(v) => commitField("overview.theme", v)}
+            onCancel={cancelEdit}
+            textClass={textCls}
+            textareaClass={`${taCls} vv-sb-inline-edit-ta-sm`}
+            disabled={locked}
+            rows={1}
+          />
+        </p>
+        <p className="vv-sb-inline-row">
+          <strong className="vv-sb-field-k">总时长：</strong>
+          <InlineText
+            value={sb.overview.total_duration}
+            editing={!locked && editingField === "overview.total_duration"}
+            onStartEdit={() => startEdit("overview.total_duration")}
+            onCommit={(v) => commitField("overview.total_duration", v)}
+            onCancel={cancelEdit}
+            textClass={textCls}
+            textareaClass={`${taCls} vv-sb-inline-edit-ta-sm`}
+            disabled={locked}
+            rows={1}
+          />
+        </p>
+        <p className="vv-sb-inline-row">
+          <strong className="vv-sb-field-k">画幅：</strong>
+          <Select
+            className="vv-select vv-sb-inline-select"
+            style={{ width: 90 }}
+            value={sb.overview.aspect_ratio}
+            disabled={locked}
+            onChange={(v) => updateOverview({ aspect_ratio: v })}
+            options={RATIOS.map((r) => ({ value: r.v, label: r.v }))}
+            variant="borderless"
+          />
+        </p>
+
+        {/* 场景与光线 */}
+        <h4 className="vv-sb-h">场景与光线</h4>
+        <p className="vv-sb-para">
+          <InlineText
+            value={sb.scene_and_lighting}
+            editing={!locked && editingField === "scene_and_lighting"}
+            onStartEdit={() => startEdit("scene_and_lighting")}
+            onCommit={(v) => commitField("scene_and_lighting", v)}
+            onCancel={cancelEdit}
+            textClass={textCls}
+            textareaClass={taCls}
+            placeholder="描述整体场景氛围、光线方向与色温…"
+            disabled={locked}
+            multiline
+            rows={2}
+          />
+        </p>
+
+        {/* 逐镜头 */}
+        <h4 className="vv-sb-h">逐秒镜头拆解</h4>
+        <div className="vv-sb-doc-shots">
+          {sb.shots.map((sh, idx) => (
+            <StoryboardShotItem
+              key={idx}
+              index={idx}
+              shot={sh}
+              editor={editor}
+              images={images}
+              locked={locked}
+            />
+          ))}
+        </div>
+
+        {/* 硬性约束 */}
+        <h4 className="vv-sb-h">硬性约束</h4>
+        <div className="vv-sb-tags-block">
+          <div className="vv-sb-taglist">
+            {sb.hard_constraints.map((c, i) => (
+              <span key={i} className="vv-sb-tag vv-sb-tag-hard">
+                <input
+                  className="vv-sb-tag-input"
+                  value={c}
+                  disabled={locked}
+                  onChange={(e) => updateConstraint("hard_constraints", i, e.target.value)}
+                />
+                {!locked && (
+                  <button
+                    className="vv-sb-tag-del"
+                    onClick={() => removeConstraint("hard_constraints", i)}
+                    aria-label="删除"
+                  >
+                    ×
+                  </button>
+                )}
+              </span>
+            ))}
+            {!locked && (
+              <button className="vv-sb-tag-add" onClick={() => addConstraint("hard_constraints")}>
+                + 添加
+              </button>
+            )}
+          </div>
+        </div>
+
+        {/* 负面提示词 */}
+        <h4 className="vv-sb-h">负面提示词</h4>
+        <div className="vv-sb-tags-block">
+          <div className="vv-sb-taglist">
+            {sb.negative_prompts.map((c, i) => (
+              <span key={i} className="vv-sb-tag vv-sb-tag-neg">
+                <input
+                  className="vv-sb-tag-input"
+                  value={c}
+                  disabled={locked}
+                  onChange={(e) => updateConstraint("negative_prompts", i, e.target.value)}
+                />
+                {!locked && (
+                  <button
+                    className="vv-sb-tag-del"
+                    onClick={() => removeConstraint("negative_prompts", i)}
+                    aria-label="删除"
+                  >
+                    ×
+                  </button>
+                )}
+              </span>
+            ))}
+            {!locked && (
+              <button className="vv-sb-tag-add" onClick={() => addConstraint("negative_prompts")}>
+                + 添加
+              </button>
+            )}
+          </div>
+        </div>
+
+        {/* 完整口播稿 */}
+        <p className="vv-sb-field">
+          <strong className="vv-sb-field-k">
+            <SoundOutlined style={{ color: "#7c3aed", marginRight: 4 }} />
+            口播稿：
+          </strong>
+          {editingField === "voiceover_script" ? (
+            <textarea
+              className={taCls}
+              value={sb.voiceover_script}
+              autoFocus
+              onChange={(e) => updateSbField("voiceover_script", e.target.value)}
+              onBlur={() => {
+                // 受控模式：blur 即退出编辑（内容已实时同步）
+                if (editingField === "voiceover_script") cancelEdit()
+              }}
+              onKeyDown={(e) => {
+                if (e.key === "Escape") cancelEdit()
+              }}
+              placeholder="AI 合成配音用的完整口播稿…"
+              rows={4}
+            />
+          ) : (
+            <span
+              className={textCls}
+              onClick={() => {
+                if (!locked) startEdit("voiceover_script")
+              }}
+              style={{ cursor: locked ? "not-allowed" : "text" }}
+            >
+              {renderVoiceoverWithSpeakers(sb.voiceover_script)}
+            </span>
+          )}
+        </p>
+      </div>
+
+      <div className="vv-sb-actions">
+        {uiStep === "step2_copy_ready" && onRegenerateCopy && (
+          <button className="vv-btn vv-btn-ghost" onClick={onRegenerateCopy}>
+            <ReloadOutlined /> 重新生成文案
+          </button>
+        )}
+      </div>
+    </div>
+  )
+}
+
+export default StoryboardCard

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

@@ -0,0 +1,193 @@
+import React from "react"
+import { InlineText } from "./InlineEditTextarea"
+import type { StoryboardEditor, SbShot } from "../hooks/useStoryboardEditor"
+import type { UploadedImage } from "../types"
+
+export interface StoryboardShotItemProps {
+  index: number
+  shot: SbShot
+  editor: StoryboardEditor
+  images: UploadedImage[]
+  locked: boolean
+}
+
+const MAIN_SPEAKERS = ["店主", "主播", "老板", "我", "主讲人"]
+
+const SHOT_FIELD_KEY = (idx: number, field: string) => `shot.${idx}.${field}`
+
+/** 单个镜头行：时间/说话人/景别/场景/动作/BGM/转场/参考图 */
+export const StoryboardShotItem: React.FC<StoryboardShotItemProps> = ({
+  index,
+  shot,
+  editor,
+  images,
+  locked,
+}) => {
+  const { editingField, startEdit, cancelEdit, commitField, updateShot } = editor
+  const refImg =
+    typeof shot.reference_image_index === "number" ? images[shot.reference_image_index] : undefined
+
+  const textCls = "vv-sb-field-val"
+  const taCls = "vv-textarea vv-sb-doc-ta vv-sb-inline-edit-ta"
+
+  const isMainSpeaker = shot.speaker ? MAIN_SPEAKERS.includes(shot.speaker) : false
+
+  return (
+    <div className="vv-sb-doc-shot">
+      {/* 时间 */}
+      <InlineText
+        value={shot.time_range}
+        editing={!locked && editingField === SHOT_FIELD_KEY(index, "time_range")}
+        onStartEdit={() => startEdit(SHOT_FIELD_KEY(index, "time_range"))}
+        onCommit={(v) => commitField(SHOT_FIELD_KEY(index, "time_range"), v)}
+        onCancel={cancelEdit}
+        textClass="vv-sb-time-doc"
+        textareaClass={`${taCls} vv-sb-time-ta`}
+        placeholder="0-3秒"
+        disabled={locked}
+        rows={1}
+      />
+
+      {/* 说话人标签（有 speaker 时展示，纯展示不加编辑） */}
+      {shot.speaker && (
+        <span
+          className="vv-sb-speaker-tag"
+          style={{
+            display: "inline-block",
+            padding: "2px 8px",
+            marginRight: 8,
+            marginLeft: 8,
+            borderRadius: 4,
+            fontSize: 12,
+            fontWeight: 500,
+            backgroundColor: isMainSpeaker ? "#e6f7ff" : "#fff7e6",
+            color: isMainSpeaker ? "#1890ff" : "#fa8c16",
+          }}
+        >
+          {shot.speaker}
+        </span>
+      )}
+
+      {/* 景别/角度与运镜 */}
+      <p className="vv-sb-field">
+        <strong className="vv-sb-field-k">景别/角度与运镜：</strong>
+        <InlineText
+          value={shot.shot_type_angle_movement}
+          editing={!locked && editingField === SHOT_FIELD_KEY(index, "shot_type_angle_movement")}
+          onStartEdit={() => startEdit(SHOT_FIELD_KEY(index, "shot_type_angle_movement"))}
+          onCommit={(v) => commitField(SHOT_FIELD_KEY(index, "shot_type_angle_movement"), v)}
+          onCancel={cancelEdit}
+          textClass={textCls}
+          textareaClass={taCls}
+          disabled={locked}
+          multiline
+          rows={2}
+        />
+      </p>
+
+      {/* 场景与对白 */}
+      <p className="vv-sb-field">
+        <strong className="vv-sb-field-k">场景与对白：</strong>
+        <InlineText
+          value={shot.scene_and_dialogue}
+          editing={!locked && editingField === SHOT_FIELD_KEY(index, "scene_and_dialogue")}
+          onStartEdit={() => startEdit(SHOT_FIELD_KEY(index, "scene_and_dialogue"))}
+          onCommit={(v) => commitField(SHOT_FIELD_KEY(index, "scene_and_dialogue"), v)}
+          onCancel={cancelEdit}
+          textClass={textCls}
+          textareaClass={taCls}
+          disabled={locked}
+          multiline
+          rows={2}
+        />
+      </p>
+
+      {/* 动作与真人细节 */}
+      <p className="vv-sb-field">
+        <strong className="vv-sb-field-k">动作与真人细节：</strong>
+        <InlineText
+          value={shot.action_details}
+          editing={!locked && editingField === SHOT_FIELD_KEY(index, "action_details")}
+          onStartEdit={() => startEdit(SHOT_FIELD_KEY(index, "action_details"))}
+          onCommit={(v) => commitField(SHOT_FIELD_KEY(index, "action_details"), v)}
+          onCancel={cancelEdit}
+          textClass={textCls}
+          textareaClass={taCls}
+          disabled={locked}
+          multiline
+          rows={2}
+        />
+      </p>
+
+      {/* 音效/BGM */}
+      <p className="vv-sb-field">
+        <strong className="vv-sb-field-k">音效/BGM：</strong>
+        <InlineText
+          value={shot.audio_bgm}
+          editing={!locked && editingField === SHOT_FIELD_KEY(index, "audio_bgm")}
+          onStartEdit={() => startEdit(SHOT_FIELD_KEY(index, "audio_bgm"))}
+          onCommit={(v) => commitField(SHOT_FIELD_KEY(index, "audio_bgm"), v)}
+          onCancel={cancelEdit}
+          textClass={textCls}
+          textareaClass={taCls}
+          disabled={locked}
+          multiline
+          rows={1}
+        />
+      </p>
+
+      {/* 转场 */}
+      <p className="vv-sb-field">
+        <strong className="vv-sb-field-k">转场：</strong>
+        <InlineText
+          value={shot.transition}
+          editing={!locked && editingField === SHOT_FIELD_KEY(index, "transition")}
+          onStartEdit={() => startEdit(SHOT_FIELD_KEY(index, "transition"))}
+          onCommit={(v) => commitField(SHOT_FIELD_KEY(index, "transition"), v)}
+          onCancel={cancelEdit}
+          textClass={textCls}
+          textareaClass={taCls}
+          disabled={locked}
+          multiline
+          rows={1}
+        />
+      </p>
+
+      {/* 参考图片 */}
+      <p className="vv-sb-field vv-sb-ref-row">
+        <strong className="vv-sb-field-k">参考图片：</strong>
+        {refImg?.url ? (
+          <span className="vv-sb-ref-badge">
+            <img src={refImg.url} alt={`ref-${index}`} className="vv-sb-ref-thumb" />
+            {!locked && (
+              <button
+                className="vv-sb-ref-x"
+                onClick={() => updateShot(index, { reference_image_index: undefined })}
+                aria-label="移除参考图"
+              >
+                ×
+              </button>
+            )}
+          </span>
+        ) : !locked && images.length > 0 ? (
+          <span className="vv-sb-ref-pick">
+            {images.map((im, i) => (
+              <button
+                key={im.localId}
+                className="vv-sb-ref-opt"
+                onClick={() => updateShot(index, { reference_image_index: i })}
+                title={`@图片${i + 1}`}
+              >
+                <img src={im.url} alt={`img${i}`} />
+              </button>
+            ))}
+          </span>
+        ) : (
+          <span className="vv-sb-muted">未指定</span>
+        )}
+      </p>
+    </div>
+  )
+}
+
+export default StoryboardShotItem

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

@@ -0,0 +1,26 @@
+import React from "react"
+
+export interface VVSpinnerProps {
+  size?: number
+  color?: string
+  style?: React.CSSProperties
+  className?: string
+}
+
+/**
+ * 统一 spinner 组件：复用 vv-spinner CSS 动画类，避免多处散落的 LoadingOutlined
+ */
+export const VVSpinner: React.FC<VVSpinnerProps> = ({ size, color, style, className }) => {
+  return (
+    <span
+      className={`vv-spinner ${className || ""}`}
+      style={{
+        ...(size ? { width: size, height: size, borderWidth: Math.max(2, size / 8) } : {}),
+        ...(color ? { borderTopColor: color } : {}),
+        ...style,
+      }}
+    />
+  )
+}
+
+export default VVSpinner

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

@@ -0,0 +1,218 @@
+import React from "react"
+import {
+  UserOutlined,
+  AudioMutedOutlined,
+  PlayCircleOutlined,
+  PauseCircleOutlined,
+  FileTextOutlined,
+  FolderOpenOutlined,
+  UploadOutlined,
+  CaretDownOutlined,
+  CaretRightOutlined,
+  PlusOutlined,
+} from "@ant-design/icons"
+import { message } from "antd"
+import type { RefAudio, VoicePanel } from "../types"
+import type { VoiceClone } from "@/api/voice-clone"
+import { getVoiceClonePreview } from "@/api/voice-clone"
+import type { AssetItem } from "@/api/assets"
+import { VOICE_TIPS } from "../ViralVideoPage.constants"
+
+export interface VoicePickerSectionProps {
+  refAudio: RefAudio
+  voicePanel: VoicePanel
+  voiceTipsOpen: boolean
+  myVoices: VoiceClone[]
+  myVoicesLoading: boolean
+  voiceAssets: { id: string; name: string; url: string; desc?: string }[]
+  playingVoiceId: string | null
+  voiceInputRef: React.Ref<HTMLInputElement>
+  onOpenCloneModal: () => void
+  onToggleVoicePanel: (p: VoicePanel) => void
+  onSelectMyVoice: (v: VoiceClone) => void
+  onSelectVoiceAsset: (a: AssetItem) => void
+  onToggleVoicePlay: (v: { id: string; url: string }) => void
+  onPickLocalAudio: () => void
+  onPickAssetAudio: () => void
+  onVoiceSelected: (file: File) => void
+  onRemoveRefAudio: () => void
+  onToggleVoiceTips: () => void
+}
+
+export const VoicePickerSection: React.FC<VoicePickerSectionProps> = (p) => {
+  return (
+    <div className="vv-subblock">
+      <div className="vv-subblock-head">
+        <span className="vv-subblock-title">AI 配音</span>
+        <span className="vv-opt-tag">选填</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>
+
+      {/* 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={p.onOpenCloneModal}>
+            <PlusOutlined /> 录制新音色
+          </button>
+        </div>
+        {p.myVoicesLoading ? (
+          <div className="vv-muted vv-voice-empty">加载中…</div>
+        ) : p.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={p.onOpenCloneModal}>
+              去录制一个
+            </button>
+          </div>
+        ) : (
+          <div className="vv-my-voice-list">
+            {p.myVoices.map((v) => {
+              const selected = p.refAudio?.source === "my_voice" && p.refAudio.id === v.id
+              const playing = p.playingVoiceId === `my-${v.id}`
+              return (
+                <div
+                  key={v.id}
+                  className={`vv-my-voice-item ${selected ? "selected" : ""}`}
+                  onClick={() => p.onSelectMyVoice(v)}
+                >
+                  <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 ${playing ? "playing" : ""}`}
+                    onClick={(e) => {
+                      e.stopPropagation()
+                      if (v.sample_url) {
+                        p.onToggleVoicePlay({ id: `my-${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)
+                              p.onToggleVoicePlay({ 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 */}
+      <div className="vv-voice-alt-row">
+        <button
+          className={`vv-voice-alt-btn ${p.refAudio?.source === "preset" ? "selected" : ""}`}
+          onClick={() => p.onToggleVoicePanel("preset")}
+        >
+          <FileTextOutlined /> 内置音色
+        </button>
+        <button
+          className={`vv-voice-alt-btn ${p.refAudio?.source === "upload" || p.refAudio?.source === "library" ? "selected" : ""}`}
+          onClick={() => p.onToggleVoicePanel("library")}
+        >
+          <FolderOpenOutlined /> 上传配音
+        </button>
+      </div>
+
+      <input
+        ref={p.voiceInputRef}
+        type="file"
+        accept="audio/*"
+        hidden
+        onChange={(e) => {
+          if (e.target.files?.[0]) p.onVoiceSelected(e.target.files[0])
+          e.target.value = ""
+        }}
+      />
+      {p.refAudio && (
+        <div className="vv-selected-audio">
+          <span style={{ marginRight: 6 }}>🔊</span>
+          <span style={{ flex: 1 }}>
+            已选：
+            {p.refAudio.source === "my_voice"
+              ? `我的音色 · ${p.refAudio.name}`
+              : p.refAudio.source === "preset"
+                ? `内置 · ${p.refAudio.name}`
+                : p.refAudio.name}
+          </span>
+          <button className="vv-link-btn" onClick={p.onRemoveRefAudio}>
+            移除
+          </button>
+        </div>
+      )}
+      {p.voicePanel === "library" && (
+        <div className="vv-voice-panel">
+          <div className="vv-voice-panel-actions">
+            <button className="vv-link-btn vv-link-btn-sm" onClick={p.onPickLocalAudio}>
+              <UploadOutlined /> 本地上传音频文件
+            </button>
+            <button className="vv-link-btn vv-link-btn-sm" onClick={p.onPickAssetAudio}>
+              <FolderOpenOutlined /> 从素材库选择
+            </button>
+          </div>
+          {p.voiceAssets.length === 0 ? (
+            <div className="vv-muted" style={{ padding: 12, textAlign: "center" }}>
+              暂无已上传的配音文件
+            </div>
+          ) : (
+            <div className="vv-voice-list">
+              {p.voiceAssets.map((v) => (
+                <div
+                  key={v.id}
+                  className={`vv-voice-item ${p.refAudio?.id === v.id && (p.refAudio.source === "library" || p.refAudio.source === "upload") ? "selected" : ""}`}
+                  onClick={() =>
+                    p.onSelectVoiceAsset({ id: v.id, name: v.name, file_url: v.url } as AssetItem)
+                  }
+                >
+                  <div className="vv-voice-radio" />
+                  <div className="vv-voice-info">
+                    <div className="vv-voice-name">{v.name}</div>
+                    <div className="vv-voice-desc">{v.desc || ""}</div>
+                  </div>
+                  {v.url && (
+                    <button
+                      className={`vv-voice-play ${p.playingVoiceId === v.id ? "playing" : ""}`}
+                      onClick={(e) => {
+                        e.stopPropagation()
+                        p.onToggleVoicePlay(v)
+                      }}
+                    >
+                      {p.playingVoiceId === v.id ? <PauseCircleOutlined /> : <PlayCircleOutlined />}
+                    </button>
+                  )}
+                </div>
+              ))}
+            </div>
+          )}
+        </div>
+      )}
+
+      <div className="vv-tips">
+        <button className="vv-tips-head" onClick={p.onToggleVoiceTips}>
+          {p.voiceTipsOpen ? <CaretDownOutlined /> : <CaretRightOutlined />}
+          支持格式与录音建议
+        </button>
+        {p.voiceTipsOpen && <div className="vv-tips-body">{VOICE_TIPS}</div>}
+      </div>
+    </div>
+  )
+}
+
+export default VoicePickerSection

apps/web/src/pages/viral-video/hooks/useStoryboardEditor.ts

@@ -0,0 +1,140 @@
+import { useCallback, useState } from "react"
+
+// 最小化的 Storyboard 类型，避免循环依赖（字段名与 api/viral-video/types 对齐）
+export interface SbShot {
+  time_range: string
+  speaker?: string
+  shot_type_angle_movement: string
+  scene_and_dialogue: string
+  action_details: string
+  audio_bgm: string
+  transition: string
+  reference_image_index?: number
+}
+export interface SbOverview {
+  theme: string
+  total_duration: string
+  aspect_ratio: string
+}
+export interface SbDoc {
+  overview: SbOverview
+  scene_and_lighting: string
+  shots: SbShot[]
+  hard_constraints: string[]
+  negative_prompts: string[]
+  voiceover_script: string
+  copy_display_markdown: string
+}
+
+export type SbUpdater = (patch: (sb: SbDoc) => SbDoc) => void
+
+export interface StoryboardEditor {
+  editingField: string | null
+  startEdit: (field: string) => void
+  cancelEdit: () => void
+  updateShot: (idx: number, patch: Partial<SbShot>) => void
+  updateOverview: (patch: Partial<SbOverview>) => void
+  updateSbField: <K extends keyof SbDoc>(key: K, value: SbDoc[K]) => void
+  updateConstraint: (list: "hard_constraints" | "negative_prompts", idx: number, v: string) => void
+  addConstraint: (list: "hard_constraints" | "negative_prompts") => void
+  removeConstraint: (list: "hard_constraints" | "negative_prompts", idx: number) => void
+  commitField: (field: string, value: string) => void
+}
+
+/**
+ * 分镜脚本编辑器 hook：管理 editingField 状态，以及 storyboard 字段级更新方法。
+ * 父组件传入 updateSb(fn) 即可，fn 接收当前 sb 返回新 sb（immer 风格）。
+ */
+export function useStoryboardEditor(updateSb: SbUpdater): StoryboardEditor {
+  const [editingField, setEditingField] = useState<string | null>(null)
+
+  const startEdit = useCallback((field: string) => {
+    setEditingField(field)
+  }, [])
+
+  const cancelEdit = useCallback(() => {
+    setEditingField(null)
+  }, [])
+
+  const updateShot = useCallback(
+    (idx: number, patch: Partial<SbShot>) => {
+      updateSb((sb) => ({
+        ...sb,
+        shots: sb.shots.map((s, i) => (i === idx ? { ...s, ...patch } : s)),
+      }))
+    },
+    [updateSb],
+  )
+
+  const updateOverview = useCallback(
+    (patch: Partial<SbOverview>) => {
+      updateSb((sb) => ({ ...sb, overview: { ...sb.overview, ...patch } }))
+    },
+    [updateSb],
+  )
+
+  const updateSbField = useCallback(
+    <K extends keyof SbDoc>(key: K, value: SbDoc[K]) => {
+      updateSb((sb) => ({ ...sb, [key]: value }))
+    },
+    [updateSb],
+  )
+
+  const updateConstraint = useCallback(
+    (list: "hard_constraints" | "negative_prompts", idx: number, v: string) => {
+      updateSb((sb) => {
+        const arr = [...sb[list]]
+        arr[idx] = v
+        return { ...sb, [list]: arr }
+      })
+    },
+    [updateSb],
+  )
+
+  const addConstraint = useCallback(
+    (list: "hard_constraints" | "negative_prompts") => {
+      updateSb((sb) => ({ ...sb, [list]: [...sb[list], ""] }))
+    },
+    [updateSb],
+  )
+
+  const removeConstraint = useCallback(
+    (list: "hard_constraints" | "negative_prompts", idx: number) => {
+      updateSb((sb) => ({ ...sb, [list]: sb[list].filter((_, i) => i !== idx) }))
+    },
+    [updateSb],
+  )
+
+  const commitField = useCallback(
+    (field: string, value: string) => {
+      if (field.startsWith("shot.")) {
+        const parts = field.split(".")
+        const idx = Number(parts[1])
+        const shotField = parts.slice(2).join(".") as keyof SbShot
+        updateShot(idx, { [shotField]: value } as Partial<SbShot>)
+      } else if (field.startsWith("overview.")) {
+        const f = field.split(".")[1] as keyof SbOverview
+        updateOverview({ [f]: value } as Partial<SbOverview>)
+      } else if (field === "scene_and_lighting" || field === "voiceover_script") {
+        updateSbField(field, value)
+      }
+      setEditingField(null)
+    },
+    [updateShot, updateOverview, updateSbField],
+  )
+
+  return {
+    editingField,
+    startEdit,
+    cancelEdit,
+    updateShot,
+    updateOverview,
+    updateSbField,
+    updateConstraint,
+    addConstraint,
+    removeConstraint,
+    commitField,
+  }
+}
+
+export default useStoryboardEditor

apps/web/src/pages/viral-video/hooks/useViralVideoHandlers.ts

@@ -0,0 +1,404 @@
+import { useCallback, useEffect } from "react"
+import axios from "axios"
+import { message } from "antd"
+import { isVideoStage, isCopyStage, getJobStage } from "@/api/viral-video/types"
+import type { UIStep, TabTask } from "../types"
+import {
+  retryViralVideo,
+  analyzeViralImages,
+  generateViralCopy,
+  confirmViralCopy,
+} from "@/api/viral-video"
+import type { CopyResult } from "@/api/viral-video/types"
+import type { ViralVideoJob } from "@/api/viral-video/types"
+import { applyPollUpdate } from "../utils/applyPollUpdate"
+import {
+  useViralVideoMedia,
+  type UseViralVideoMediaOptions,
+  type MediaActions,
+} from "./useViralVideoMedia"
+import { resolveModelSwitch } from "../utils/resolveModelSwitch"
+
+export interface ViralVideoActionsApi extends MediaActions {
+  resetTask: () => void
+  handleAnalyzeImages: () => void
+  handleGenerateCopy: () => void
+  handleGenerateVideo: () => void
+  handleRetry: () => void
+  onPollUpdate: (job: ViralVideoJob) => void
+  onScriptDelta: (ev: { data: { full_text?: string }; progress?: number; message?: string }) => void
+  handleModelChange: (
+    v: string,
+    videoModels: { key: string; supported_resolutions?: string[]; max_duration?: number }[],
+  ) => void
+  setTask: (patch: Partial<TabTask> | ((t: TabTask) => TabTask)) => void
+  setAssetPicker: React.Dispatch<
+    React.SetStateAction<{ open: boolean; kind: "image" | "video" | "voice"; multiple: boolean }>
+  >
+  setCloneModalOpen: React.Dispatch<React.SetStateAction<boolean>>
+  setPresetVoicePickerOpen: React.Dispatch<React.SetStateAction<boolean>>
+  refreshVoiceAssets: () => void
+  reloadMyVoices: () => void
+}
+
+export interface UseViralVideoHandlersOptions extends Omit<UseViralVideoMediaOptions, "setTask"> {
+  activeId: string
+  setTasks: React.Dispatch<React.SetStateAction<TabTask[]>>
+  videoModels: {
+    key: string
+    supported_resolutions?: string[]
+    max_duration?: number
+    display_name?: string
+  }[]
+}
+
+function emptyTask(id: string, title: string): TabTask {
+  return {
+    id,
+    title,
+    images: [],
+    refVideo: null,
+    douyinInput: "",
+    voicePanel: null,
+    refAudio: null,
+    fusionLevel: "ai_full",
+    styleStrength: "medium",
+    userCopy: "",
+    industry: "",
+    targetCustomer: "",
+    language: "中文(普通话)",
+    viralStructure: "反差破局+亮明观点+还原现状",
+    marketingPurpose: "",
+    duration: 20,
+    persona: "",
+    videoRatio: "9:16",
+    videoModel: "seedance-2.5",
+    quality: "480p",
+    upscale: "default",
+    estimatedCredits: null,
+    estimatedCreditsError: false,
+    uiStep: "step1_upload",
+    imageAnalysis: null,
+    storyboard: null,
+    voiceTipsOpen: false,
+    job: null,
+    jobId: null,
+    playingVoiceId: null,
+    scriptStreamText: "",
+    scriptStreamProgress: 0,
+    scriptStreamMessage: "",
+  }
+}
+
+const CLEAR_STREAM = {
+  scriptStreamText: "",
+  scriptStreamProgress: 0,
+  scriptStreamMessage: "",
+} as const
+
+/**
+ * 业务 action hook：聚合素材操作 + 三步骤 submit/retry/polling/model switch 等业务动作。
+ */
+export function useViralVideoHandlers(opts: UseViralVideoHandlersOptions): ViralVideoActionsApi {
+  const {
+    task,
+    activeId,
+    setTasks,
+    setAssetPicker,
+    setCloneModalOpen,
+    refreshVoiceAssets,
+    reloadMyVoices,
+    setPresetVoicePickerOpen,
+  } = opts
+
+  const setTask = useCallback(
+    (patch: Partial<TabTask> | ((t: TabTask) => TabTask)) => {
+      setTasks((ts) =>
+        ts.map((t) =>
+          t.id === activeId ? (typeof patch === "function" ? patch(t) : { ...t, ...patch }) : t,
+        ),
+      )
+    },
+    [activeId, setTasks],
+  )
+
+  const media = useViralVideoMedia({
+    task,
+    setTask,
+    setAssetPicker,
+    setCloneModalOpen,
+    setPresetVoicePickerOpen,
+    refreshVoiceAssets,
+    reloadMyVoices,
+  })
+
+  // 卸载时停掉正在播放的音频
+  useEffect(() => () => media.audioRef.current?.pause(), [media.audioRef])
+
+  const resetTask = () => {
+    media.audioRef.current?.pause()
+    setTasks((ts) => ts.map((t) => (t.id === activeId ? emptyTask(t.id, t.title) : t)))
+  }
+
+  const handleAnalyzeImages = async () => {
+    const readyImgs = task.images.filter((i) => i.ossUrl && !i.uploading)
+    if (readyImgs.length === 0) {
+      message.warning("请至少上传 1 张产品图")
+      return
+    }
+    if (task.images.some((i) => i.uploading)) {
+      message.warning("图片上传中，请稍候")
+      return
+    }
+    if (task.refVideo?.uploading) {
+      message.warning("参考视频上传中，请稍候")
+      return
+    }
+    setTask({
+      uiStep: "step1_analyzing",
+      jobId: null,
+      job: null,
+      analysisError: undefined,
+      copyError: undefined,
+      videoError: undefined,
+      imageAnalysis: undefined,
+      storyboard: undefined,
+      ...CLEAR_STREAM,
+      estimatedCredits: null,
+      estimatedCreditsError: false,
+    })
+    try {
+      const job = await analyzeViralImages({
+        images: readyImgs.map((i) => i.ossUrl as string),
+        reference_video_url: task.refVideo?.ossUrl || undefined,
+        style_strength: task.refVideo?.ossUrl ? task.styleStrength : undefined,
+      })
+      setTask({ job, jobId: job.id })
+      message.success("已提交图片分析…")
+    } catch (err: unknown) {
+      message.error(err instanceof Error ? err.message : "分析失败")
+      setTask({
+        uiStep: "failed",
+        analysisError: err instanceof Error ? err.message : "分析失败",
+        copyError: undefined,
+        videoError: undefined,
+      })
+    }
+  }
+
+  const handleGenerateCopy = async () => {
+    if (!task.jobId) {
+      message.warning("请先完成图片分析")
+      return
+    }
+    setTask({
+      uiStep: "step2_generating",
+      copyError: undefined,
+      videoError: undefined,
+      analysisError: undefined,
+      ...CLEAR_STREAM,
+    })
+    try {
+      const job = await generateViralCopy(task.jobId, {
+        industry: task.industry || undefined,
+        target_customer: task.targetCustomer || undefined,
+        viral_structure: task.viralStructure || undefined,
+        marketing_purpose: task.marketingPurpose || undefined,
+        duration: task.duration,
+        persona_id: task.persona || undefined,
+        user_copy_text:
+          (task.fusionLevel === "user_primary" || task.fusionLevel === "ai_polish") && task.userCopy
+            ? task.userCopy
+            : undefined,
+        fusion_level: task.fusionLevel,
+        style_strength: task.refVideo?.ossUrl ? task.styleStrength : undefined,
+        voice_id:
+          task.refAudio?.source === "preset"
+            ? task.refAudio.id
+            : 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,
+      })
+      setTask({ job })
+    } catch (err: unknown) {
+      message.error(err instanceof Error ? err.message : "文案生成失败")
+      setTask({
+        uiStep: "failed",
+        copyError: err instanceof Error ? err.message : "文案生成失败",
+        videoError: undefined,
+        analysisError: undefined,
+      })
+    }
+  }
+
+  const handleGenerateVideo = async () => {
+    setTask({
+      uiStep: "step3_generating",
+      videoError: undefined,
+      copyError: undefined,
+      analysisError: undefined,
+      ...CLEAR_STREAM,
+    })
+    try {
+      let job
+      if (task.jobId && task.storyboard && task.storyboard.voiceover_script.trim()) {
+        const edited = task.storyboard.voiceover_script.trim()
+        const originalCopy =
+          (task.job?.copy_result as CopyResult | null | undefined)?.voiceover_script ||
+          task.job?.copy_result?.final_copy ||
+          ""
+        job = await confirmViralCopy(task.jobId, {
+          edited_copy: edited && edited !== originalCopy.trim() ? edited : undefined,
+          video_model: task.videoModel,
+          video_resolution: task.quality,
+          video_ratio: task.videoRatio,
+          duration: task.duration,
+          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,
+          language: task.language || undefined,
+        })
+      } else {
+        message.warning("请先生成文案后再生成视频")
+        setTask({
+          uiStep: task.storyboard
+            ? "step2_copy_ready"
+            : task.imageAnalysis
+              ? "step1_done"
+              : "step1_upload",
+        })
+        return
+      }
+      setTask({ job, jobId: job.id })
+      message.success("视频已提交生成，预计 1-3 分钟…")
+    } catch (err: unknown) {
+      if (axios.isAxiosError(err) && err.response?.status === 402) {
+        const detail =
+          (err.response.data as { detail?: string; message?: string; msg?: string } | undefined)
+            ?.detail ||
+          (err.response.data as { detail?: string; message?: string; msg?: string } | undefined)
+            ?.message ||
+          "积分不足，请充值"
+        message.error(detail)
+        setTask({
+          uiStep: "step2_copy_ready",
+          videoError: detail,
+          copyError: undefined,
+          analysisError: undefined,
+          ...CLEAR_STREAM,
+        })
+        return
+      }
+      const msg = err instanceof Error ? err.message : "提交失败"
+      message.error(msg)
+      setTask({ uiStep: "failed", videoError: msg, copyError: undefined, analysisError: undefined })
+    }
+  }
+
+  const handleRetry = async () => {
+    if (task.jobId) {
+      try {
+        const j = await retryViralVideo(task.jobId)
+        const retryStage = getJobStage(j)
+        let retryStep: UIStep = "step1_analyzing"
+        if (isVideoStage(retryStage)) retryStep = "step3_generating"
+        else if (isCopyStage(retryStage)) retryStep = "step2_generating"
+        else if (task.storyboard) retryStep = "step2_copy_ready"
+        else if (task.imageAnalysis) retryStep = "step1_done"
+        setTask({
+          job: j,
+          jobId: j.id,
+          uiStep: retryStep,
+          videoError: undefined,
+          copyError: undefined,
+          analysisError: undefined,
+          imageAnalysis: undefined,
+          storyboard: undefined,
+          ...CLEAR_STREAM,
+          estimatedCredits: null,
+          estimatedCreditsError: false,
+        })
+        message.success("已重试")
+      } catch (err: unknown) {
+        message.error(err instanceof Error ? err.message : "重试失败")
+      }
+    } else {
+      setTask({
+        uiStep: task.storyboard ? "step2_copy_ready" : "step1_done",
+        jobId: null,
+        job: null,
+        videoError: undefined,
+        copyError: undefined,
+        analysisError: undefined,
+        imageAnalysis: undefined,
+        storyboard: undefined,
+        ...CLEAR_STREAM,
+        estimatedCredits: null,
+        estimatedCreditsError: false,
+      })
+    }
+  }
+
+  const onPollUpdate = useCallback(
+    (job: ViralVideoJob) => {
+      setTask((t) => applyPollUpdate(t, job))
+    },
+    [setTask],
+  )
+
+  const onScriptDelta = useCallback(
+    (ev: { data: { full_text?: string }; progress?: number; message?: string }) => {
+      setTask((t) => ({
+        ...t,
+        scriptStreamText: ev.data.full_text || t.scriptStreamText,
+        scriptStreamProgress:
+          typeof ev.progress === "number" ? ev.progress : t.scriptStreamProgress,
+        scriptStreamMessage: ev.message || t.scriptStreamMessage,
+      }))
+    },
+    [setTask],
+  )
+
+  const handleModelChange = (
+    v: string,
+    vms: { key: string; supported_resolutions?: string[]; max_duration?: number }[],
+  ) => {
+    const m = vms.find((x) => x && x.key === v)
+    if (!m) {
+      setTask({ videoModel: v })
+      return
+    }
+    const { quality, duration } = resolveModelSwitch(
+      { quality: task.quality, duration: task.duration },
+      m,
+    )
+    setTask({ videoModel: v, quality, duration })
+  }
+
+  return {
+    ...media,
+    resetTask,
+    handleAnalyzeImages,
+    handleGenerateCopy,
+    handleGenerateVideo,
+    handleRetry,
+    onPollUpdate,
+    onScriptDelta,
+    handleModelChange,
+    setTask,
+    setAssetPicker,
+    setCloneModalOpen,
+    setPresetVoicePickerOpen,
+    refreshVoiceAssets,
+    reloadMyVoices,
+  }
+}
+
+export default useViralVideoHandlers

apps/web/src/pages/viral-video/hooks/useViralVideoMedia.ts

@@ -0,0 +1,365 @@
+import { useCallback, useRef } from "react"
+import { message } from "antd"
+import { getAssetLibraries, uploadAssetDirect, type AssetItem } from "@/api/assets"
+import type { UploadedImage, RefVideo, RefAudio, VoicePanel, TabTask } from "../types"
+import type { VoiceClone } from "@/api/voice-clone"
+
+export interface UseViralVideoMediaOptions {
+  task: TabTask
+  setTask: (patch: Partial<TabTask> | ((t: TabTask) => TabTask)) => void
+  setAssetPicker: React.Dispatch<
+    React.SetStateAction<{ open: boolean; kind: "image" | "video" | "voice"; multiple: boolean }>
+  >
+  setCloneModalOpen: React.Dispatch<React.SetStateAction<boolean>>
+  setPresetVoicePickerOpen: React.Dispatch<React.SetStateAction<boolean>>
+  refreshVoiceAssets: () => void
+  reloadMyVoices: () => void
+}
+
+export interface MediaActions {
+  imgInputRef: React.MutableRefObject<HTMLInputElement | null>
+  videoInputRef: React.MutableRefObject<HTMLInputElement | null>
+  voiceInputRef: React.MutableRefObject<HTMLInputElement | null>
+  audioRef: React.MutableRefObject<HTMLAudioElement | null>
+  appendImagesFromFiles: (files: FileList | File[]) => void
+  appendImagesFromAssets: (assets: AssetItem[]) => void
+  clearImages: () => void
+  removeImg: (localId: string) => void
+  moveImg: (idx: number, dir: -1 | 1) => void
+  handleLocalRefVideo: (file: File) => void
+  handleDouyinUrl: () => void
+  removeRefVideo: () => void
+  handleVoiceFile: (file: File) => void
+  onCloneSuccess: (v: VoiceClone | { id?: string; name?: string; voice_id?: string | null }) => void
+  pickVoiceAsset: (a: AssetItem) => void
+  selectMyVoice: (v: VoiceClone) => void
+  toggleVoicePanel: (p: VoicePanel) => void
+  toggleVoice: (v: { id: string; url: string }) => void
+  openAssetPicker: (kind: "image" | "video" | "voice", multiple: boolean) => void
+}
+
+/**
+ * 素材上传/音色选择/图片列表/参考视频等操作：独立 hook，保持每个文件 ≤400 行。
+ */
+export function useViralVideoMedia(opts: UseViralVideoMediaOptions): MediaActions {
+  const {
+    task,
+    setTask,
+    setAssetPicker,
+    setCloneModalOpen,
+    refreshVoiceAssets,
+    setPresetVoicePickerOpen,
+  } = opts
+  const imgInputRef = useRef<HTMLInputElement>(null)
+  const videoInputRef = useRef<HTMLInputElement>(null)
+  const voiceInputRef = useRef<HTMLInputElement>(null)
+  const audioRef = useRef<HTMLAudioElement | null>(null)
+
+  const uploadFile = useCallback(
+    async (file: File, kind: "image" | "video" | "voice", onProgress?: (pct: number) => void) => {
+      const libs = await getAssetLibraries()
+      const lib = (libs as { kind: string; id: string }[]).find((l) => l.kind === kind) || libs[0]
+      if (!lib) throw new Error("未找到素材库")
+      const res = await uploadAssetDirect({
+        file,
+        library_id: lib.id,
+        onProgress: (p) => onProgress?.(Math.round(p)),
+      })
+      return res.url
+    },
+    [],
+  )
+
+  const appendImagesFromFiles = useCallback(
+    async (files: FileList | File[]) => {
+      const remaining = 9 - task.images.length
+      if (remaining <= 0) {
+        message.warning("最多上传 9 张产品图")
+        return
+      }
+      const arr = Array.from(files)
+        .filter((f) => f.type.startsWith("image/"))
+        .slice(0, remaining)
+      if (arr.length === 0) return
+      const newImgs: UploadedImage[] = arr.map((f) => ({
+        localId: `${Date.now()}_${Math.random().toString(36).slice(2, 7)}`,
+        url: URL.createObjectURL(f),
+        name: f.name,
+        size: f.size,
+        uploading: true,
+        pct: 0,
+        file: f,
+      }))
+      setTask((t) => ({
+        ...t,
+        images: [...t.images, ...newImgs.map(({ file: _f, ...rest }) => rest)],
+      }))
+      for (const item of newImgs) {
+        try {
+          const url = await uploadFile(item.file!, "image", (p) => {
+            setTask((t) => ({
+              ...t,
+              images: t.images.map((im) => (im.localId === item.localId ? { ...im, pct: p } : im)),
+            }))
+          })
+          setTask((t) => ({
+            ...t,
+            images: t.images.map((im) =>
+              im.localId === item.localId
+                ? {
+                    ...im,
+                    uploading: false,
+                    pct: 100,
+                    ossUrl: url,
+                    size: item.file!.size || im.size,
+                  }
+                : im,
+            ),
+          }))
+        } catch {
+          message.error(`${item.name} 上传失败`)
+          setTask((t) => ({ ...t, images: t.images.filter((im) => im.localId !== item.localId) }))
+        }
+      }
+    },
+    [task.images.length, setTask, uploadFile],
+  )
+
+  const appendImagesFromAssets = useCallback(
+    (assets: AssetItem[]) => {
+      const remaining = 9 - task.images.length
+      const picked = assets.slice(0, remaining)
+      if (picked.length === 0) return
+      const newImgs: UploadedImage[] = picked.map((a) => ({
+        localId: `lib_${a.id}_${Math.random().toString(36).slice(2, 6)}`,
+        url: a.thumbnail_url || a.file_url || "",
+        name: a.name || "未命名",
+        size: a.file_size || undefined,
+        uploading: false,
+        pct: 100,
+        ossUrl: a.file_url || undefined,
+      }))
+      setTask((t) => ({ ...t, images: [...t.images, ...newImgs] }))
+      message.success(`已添加 ${picked.length} 张图片`)
+    },
+    [task.images.length, setTask],
+  )
+
+  const clearImages = () => {
+    setTask((t) => {
+      t.images.forEach((im) => {
+        if (im.url?.startsWith("blob:")) URL.revokeObjectURL(im.url)
+      })
+      return {
+        ...t,
+        images: [],
+        imageAnalysis: null,
+        uiStep: "step1_upload",
+        storyboard: null,
+        job: null,
+        jobId: null,
+      }
+    })
+  }
+
+  const removeImg = (localId: string) => {
+    setTask((t) => {
+      const im = t.images.find((i) => i.localId === localId)
+      if (im?.url?.startsWith("blob:")) URL.revokeObjectURL(im.url)
+      const newImgs = t.images.filter((i) => i.localId !== localId)
+      return {
+        ...t,
+        images: newImgs,
+        imageAnalysis: newImgs.length === 0 ? null : t.imageAnalysis,
+        storyboard: newImgs.length === 0 ? null : t.storyboard,
+      }
+    })
+  }
+
+  const moveImg = (idx: number, dir: -1 | 1) => {
+    setTask((t) => {
+      const imgs = [...t.images]
+      const target = idx + dir
+      if (target < 0 || target >= imgs.length) return t
+      ;[imgs[idx], imgs[target]] = [imgs[target], imgs[idx]]
+      return { ...t, images: imgs }
+    })
+  }
+
+  const handleLocalRefVideo = useCallback(
+    async (file: File) => {
+      if (file.size > 200 * 1024 * 1024) {
+        message.warning("视频不能超过 200MB")
+        return
+      }
+      const localUrl = URL.createObjectURL(file)
+      setTask({
+        refVideo: {
+          source: "local",
+          url: localUrl,
+          name: file.name,
+          uploading: true,
+          pct: 0,
+          file,
+        } as RefVideo,
+      })
+      try {
+        const url = await uploadFile(file, "video", (p) => {
+          setTask((t) => (t.refVideo ? { ...t, refVideo: { ...t.refVideo, pct: p } } : t))
+        })
+        setTask((t) =>
+          t.refVideo
+            ? { ...t, refVideo: { ...t.refVideo, uploading: false, pct: 100, ossUrl: url } }
+            : t,
+        )
+      } catch {
+        message.error("参考视频上传失败")
+        if (localUrl.startsWith("blob:")) URL.revokeObjectURL(localUrl)
+        setTask({ refVideo: null })
+      }
+    },
+    [setTask, uploadFile],
+  )
+
+  const handleDouyinUrl = () => {
+    const url = task.douyinInput.trim()
+    if (!url) {
+      message.warning("请粘贴抖音分享链接")
+      return
+    }
+    if (!/(douyin\.com|iesdouyin\.com)/.test(url)) {
+      message.warning("请输入有效的抖音链接")
+      return
+    }
+    message.info("抖音链接识别功能即将上线，当前请使用本地上传参考视频")
+    setTask({ douyinInput: "" })
+  }
+
+  const removeRefVideo = () => {
+    setTask((t) => {
+      if (t.refVideo?.url.startsWith("blob:")) URL.revokeObjectURL(t.refVideo.url)
+      return { ...t, refVideo: null }
+    })
+  }
+
+  const handleVoiceFile = useCallback(
+    async (file: File) => {
+      if (file.size > 10 * 1024 * 1024) {
+        message.warning("音频不能超过 10MB")
+        return
+      }
+      const localUrl = URL.createObjectURL(file)
+      try {
+        const url = await uploadFile(file, "voice")
+        setTask({
+          refAudio: { source: "upload", name: file.name, url, id: url } as RefAudio,
+          voicePanel: null,
+        })
+        message.success("音频已上传")
+        refreshVoiceAssets()
+      } catch {
+        message.error("音频上传失败")
+        URL.revokeObjectURL(localUrl)
+      }
+    },
+    [setTask, uploadFile, refreshVoiceAssets],
+  )
+
+  const onCloneSuccess = (
+    v: VoiceClone | { id?: string; name?: string; voice_id?: string | null },
+  ) => {
+    setCloneModalOpen(false)
+    opts.reloadMyVoices()
+    const vid = "voice_id" in v ? v.voice_id : undefined
+    if (v?.id) {
+      setTask({
+        refAudio: {
+          source: "my_voice",
+          id: v.id,
+          name: v.name || "我录制的音色",
+          ttsVoiceId: vid || undefined,
+        } as RefAudio,
+        voicePanel: null,
+      })
+      refreshVoiceAssets()
+      message.success("音色克隆完成，已自动选中")
+    }
+  }
+
+  const pickVoiceAsset = (asset: AssetItem) => {
+    setTask({
+      refAudio: {
+        source: "library",
+        id: asset.id,
+        name: asset.name,
+        url: asset.file_url || "",
+      } as RefAudio,
+      voicePanel: null,
+    })
+  }
+
+  const selectMyVoice = (v: VoiceClone) => {
+    setTask({
+      refAudio: {
+        source: "my_voice",
+        id: v.id,
+        name: v.name,
+        ttsVoiceId: v.voice_id || undefined,
+      } as RefAudio,
+      voicePanel: null,
+    })
+  }
+
+  const toggleVoicePanel = (p: VoicePanel) => {
+    if (p === "preset") {
+      setPresetVoicePickerOpen(true)
+      return
+    }
+    setTask({ voicePanel: task.voicePanel === p ? null : p })
+  }
+
+  const openAssetPicker = (kind: "image" | "video" | "voice", multiple: boolean) => {
+    setAssetPicker({ open: true, kind, multiple })
+  }
+
+  const toggleVoice = (v: { id: string; url: string }) => {
+    if (!v.url) return
+    if (task.playingVoiceId === v.id) {
+      audioRef.current?.pause()
+      audioRef.current = null
+      setTask({ playingVoiceId: null })
+      return
+    }
+    audioRef.current?.pause()
+    const a = new Audio(v.url)
+    a.dataset.playingId = v.id
+    a.onended = () => setTask({ playingVoiceId: null })
+    a.play().catch(() => {})
+    audioRef.current = a
+    setTask({ playingVoiceId: v.id })
+  }
+
+  return {
+    imgInputRef,
+    videoInputRef,
+    voiceInputRef,
+    audioRef,
+    appendImagesFromFiles,
+    appendImagesFromAssets,
+    clearImages,
+    removeImg,
+    moveImg,
+    handleLocalRefVideo,
+    handleDouyinUrl,
+    removeRefVideo,
+    handleVoiceFile,
+    onCloneSuccess,
+    pickVoiceAsset,
+    selectMyVoice,
+    toggleVoicePanel,
+    toggleVoice,
+    openAssetPicker,
+  }
+}
+
+export default useViralVideoMedia

apps/web/src/pages/viral-video/hooks/useViralVideoPolling.ts

@@ -1,6 +1,11 @@
 import { useCallback, useEffect, useRef } from "react"
 import { getViralVideoJob } from "@/api/viral-video"
-import { isAnalysisStage, type ViralVideoJob, type ViralVideoStatus } from "@/api/viral-video/types"
+import {
+  isAnalysisStage,
+  getJobStage,
+  type ViralVideoJob,
+  type ViralVideoStatus,
+} from "@/api/viral-video/types"
 
 const TERMINAL: ViralVideoStatus[] = ["completed", "failed", "cancelled"]
 
@@ -11,8 +16,9 @@ export interface UseViralVideoPollingOptions {
 
 /**
  * 爆款视频任务 HTTP 轮询 hook。
- * 负责持续拉取任务状态并回调给上层；上层负责根据状态/阶段切换 UI 文案。
- * 任务进入终态（completed/failed/cancelled）后自动停止。
+ * - onUpdate 通过 ref 持有，避免 useCallback 依赖变化导致重启轮询。
+ * - 任务进入终态（completed/failed/cancelled）后自动停止。
+ * - 连续失败 20 次（约 2 分钟）后熔断。
  */
 export function useViralVideoPolling(
   jobId: string | null | undefined,
@@ -23,6 +29,9 @@ export function useViralVideoPolling(
   const timerRef = useRef<ReturnType<typeof setTimeout> | null>(null)
   const stoppedRef = useRef(false)
   const failCountRef = useRef(0)
+  // ref 持有最新 onUpdate，避免依赖变化导致 pollOnce 重建、轮询被重启
+  const onUpdateRef = useRef(onUpdate)
+  onUpdateRef.current = onUpdate
 
   const stop = useCallback(() => {
     stoppedRef.current = true
@@ -37,26 +46,30 @@ export function useViralVideoPolling(
       try {
         const job = await getViralVideoJob(id)
         failCountRef.current = 0
-        onUpdate(job)
+        onUpdateRef.current(job)
         if (TERMINAL.includes(job.status)) {
           stop()
           return
         }
         if (stoppedRef.current) return
-        // 视频渲染阶段（Seedance 多段视频生成较慢）拉长轮询间隔
-        const inRender = job.progress_stage === "rendering"
-        // 分析阶段走默认间隔即可
-        const isAnalyzing = isAnalysisStage(job.progress_stage)
+        const stage = getJobStage(job)
+        const inRender = stage === "rendering"
+        const isAnalyzing = isAnalysisStage(stage)
         const nextDelay = inRender ? 3000 : isAnalyzing ? 2000 : intervalMs
         timerRef.current = setTimeout(() => pollOnce(id), nextDelay)
       } catch (_err) {
         failCountRef.current += 1
         if (stoppedRef.current) return
+        if (failCountRef.current >= 20) {
+          console.warn("[viral-video] 轮询连续失败 20 次，停止轮询，请刷新页面重试")
+          stop()
+          return
+        }
         const delay = Math.min(intervalMs * 2 ** Math.min(failCountRef.current, 3), 10000)
         timerRef.current = setTimeout(() => pollOnce(id), delay)
       }
     },
-    [intervalMs, onUpdate, stop],
+    [intervalMs, stop],
   )
 
   useEffect(() => {

apps/web/src/pages/viral-video/hooks/useViralVideoWebSocket.ts

@@ -0,0 +1,151 @@
+import { useCallback, useEffect, useRef } from "react"
+
+export interface ViralVideoDeltaEvent {
+  type: "viral_video:script_delta"
+  job_id: string
+  stage: string
+  progress: number
+  message?: string
+  data: {
+    delta: string
+    full_text: string
+    text_length: number
+  }
+}
+
+export interface ViralVideoWSError {
+  message: string
+}
+
+export interface ViralVideoTerminalEvent {
+  type: "viral_video:completed" | "viral_video:failed" | "viral_video:copy_generated"
+  job_id: string
+}
+
+export interface UseViralVideoWSOptions {
+  /** delta 事件回调（script_delta 推送时触发） */
+  onDelta?: (ev: ViralVideoDeltaEvent) => void
+  /** 连接关闭/失败回调 */
+  onError?: (err: ViralVideoWSError) => void
+  /** 连接建立回调 */
+  onOpen?: () => void
+  /** P2-6: 收到终态事件（copy_generated/completed/failed）时回调，上层应立即主动拉取一次最新 job，减少终态延迟 */
+  onTerminal?: (ev: ViralVideoTerminalEvent) => void
+}
+
+const WS_SCHEME =
+  typeof window !== "undefined" && window.location.protocol === "https:" ? "wss:" : "ws:"
+
+function buildWsUrl(jobId: string): string {
+  const token = localStorage.getItem("access_token") || ""
+  const host = window.location.host
+  return `${WS_SCHEME}//${host}/api/v1/viral-video/ws/${jobId}?token=${encodeURIComponent(token)}`
+}
+
+/**
+ * 爆款视频 WebSocket 订阅 hook。
+ * 仅在 jobId 非空时建立连接；组件卸载/ jobId 变更时自动关闭。
+ * 失败静默：后端已说明流式异常会自动 fallback 到 HTTP 轮询的 copy_generated 终态，
+ * 前端不需要中断主流程，连接失败时只记录不影响轮询继续推进 UI。
+ */
+export function useViralVideoWS(
+  jobId: string | null | undefined,
+  { onDelta, onError, onOpen, onTerminal }: UseViralVideoWSOptions = {},
+) {
+  const wsRef = useRef<WebSocket | null>(null)
+  const manualCloseRef = useRef(false)
+  const onDeltaRef = useRef(onDelta)
+  const onErrorRef = useRef(onError)
+  const onOpenRef = useRef(onOpen)
+  const onTerminalRef = useRef<UseViralVideoWSOptions["onTerminal"]>()
+  onDeltaRef.current = onDelta
+  onErrorRef.current = onError
+  onOpenRef.current = onOpen
+  onTerminalRef.current = onTerminal
+
+  const close = useCallback(() => {
+    manualCloseRef.current = true
+    if (wsRef.current) {
+      try {
+        wsRef.current.close()
+      } catch {
+        /* noop */
+      }
+      wsRef.current = null
+    }
+  }, [])
+
+  useEffect(() => {
+    if (!jobId) {
+      close()
+      return
+    }
+    manualCloseRef.current = false
+    let closed = false
+    let retryTimer: ReturnType<typeof setTimeout> | null = null
+    let startTimer: ReturnType<typeof setTimeout> | null = null
+    let retries = 0
+    const MAX_RETRIES = 3
+
+    const connect = () => {
+      if (closed) return
+      try {
+        const ws = new WebSocket(buildWsUrl(jobId))
+        wsRef.current = ws
+
+        ws.onopen = () => {
+          if (closed) return
+          retries = 0
+          onOpenRef.current?.()
+        }
+
+        ws.onmessage = (ev) => {
+          try {
+            const payload = JSON.parse(ev.data)
+            if (payload?.type === "viral_video:script_delta" && payload.data) {
+              onDeltaRef.current?.(payload as ViralVideoDeltaEvent)
+            } else if (
+              payload?.type === "viral_video:completed" ||
+              payload?.type === "viral_video:failed" ||
+              payload?.type === "viral_video:copy_generated"
+            ) {
+              // P2-6: 收到终态事件时主动通知上层立即拉取，降低终态延迟（不用等下一个轮询周期）
+              onTerminalRef.current?.(payload as ViralVideoTerminalEvent)
+            }
+            // 心跳/其他事件忽略
+          } catch {
+            // 心跳/非 JSON 消息忽略
+          }
+        }
+
+        ws.onerror = () => {
+          onErrorRef.current?.({ message: "websocket error" })
+        }
+
+        ws.onclose = () => {
+          wsRef.current = null
+          if (closed || manualCloseRef.current) return
+          if (retries < MAX_RETRIES) {
+            retries += 1
+            const delay = 500 * 2 ** (retries - 1)
+            retryTimer = setTimeout(connect, delay)
+          }
+        }
+      } catch (e) {
+        onErrorRef.current?.({ message: (e as Error).message })
+      }
+    }
+
+    // 延迟 100ms 再连，给后端 job 初始化留一点时间
+    startTimer = setTimeout(connect, 100)
+
+    return () => {
+      closed = true
+      if (startTimer) clearTimeout(startTimer)
+      if (retryTimer) clearTimeout(retryTimer)
+      close()
+    }
+  }, [jobId, close])
+
+  return { close }
+}

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

@@ -0,0 +1,134 @@
+import type {
+  FusionLevel,
+  StyleStrength,
+  ViralVideoJob,
+  ImageAnalysisResult,
+  CopyResult,
+  ShotScript,
+  ViralVideoModel,
+} from "@/api/viral-video/types"
+
+/* ─────────── UI 阶段（三步骤独立交互） ─────────── */
+export type UIStep =
+  | "step1_upload" // 初始：上传素材
+  | "step1_analyzing" // 图片分析中
+  | "step1_done" // 图片分析完成
+  | "step2_generating" // 文案生成中
+  | "step2_copy_ready" // 文案已生成，等待确认
+  | "step3_generating" // 视频生成中
+  | "step3_done" // 视频完成
+  | "failed" // 失败
+
+export const TERMINAL_UI_STEPS: UIStep[] = ["step3_done", "failed"]
+
+export type UploadedImage = {
+  localId: string
+  url: string
+  name: string
+  size: number | undefined
+  uploading: boolean
+  pct: number
+  ossUrl?: string
+  file?: File
+}
+
+export type RefVideo = {
+  source: "local" | "douyin"
+  url: string
+  name: string
+  uploading: boolean
+  pct: number
+  ossUrl?: string
+  file?: File
+  douyinUrl?: string
+} | null
+
+export type RefAudio = {
+  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
+
+export type VoicePanel = "my_voices" | "preset" | "upload" | "library" | "record" | "douyin" | null
+
+// ── 分镜脚本数据模型（新后端 copy_result 结构） ──
+export interface StoryboardShot {
+  time_range: string
+  speaker?: string
+  shot_type_angle_movement: string
+  scene_and_dialogue: string
+  action_details: string
+  audio_bgm: string
+  transition: string
+  reference_image_index?: number
+}
+export interface Storyboard {
+  overview: { theme: string; total_duration: string; aspect_ratio: string }
+  scene_and_lighting: string
+  shots: StoryboardShot[]
+  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
+}
+
+export type TabTask = {
+  id: string
+  title: string
+  // 素材
+  images: UploadedImage[]
+  refVideo: RefVideo
+  douyinInput: string
+  voicePanel: VoicePanel
+  refAudio: RefAudio
+  // STEP2 参数
+  fusionLevel: FusionLevel
+  styleStrength: StyleStrength
+  userCopy: string
+  industry: string
+  targetCustomer: string
+  persona: string
+  language: string
+  viralStructure: string
+  marketingPurpose: string
+  duration: number
+  // STEP3 设置
+  videoRatio: string
+  videoModel: string
+  quality: string
+  upscale: string
+  // 积分预估
+  estimatedCredits: number | null
+  estimatedCreditsError: boolean
+  // UI 状态
+  uiStep: UIStep
+  imageAnalysis: ImageAnalysisResult | null
+  storyboard: Storyboard | null
+  voiceTipsOpen: boolean
+  copyError?: string
+  videoError?: string
+  analysisError?: string
+  // 运行态（后端 job）
+  job: ViralVideoJob | null
+  jobId: string | null
+  playingVoiceId: string | null
+  // 分镜脚本流式预览
+  scriptStreamText: string
+  scriptStreamProgress: number
+  scriptStreamMessage: string
+}
+
+// re-export api types for convenience
+export type {
+  FusionLevel,
+  StyleStrength,
+  ViralVideoJob,
+  ImageAnalysisResult,
+  CopyResult,
+  ShotScript,
+  ViralVideoModel,
+}

apps/web/src/pages/viral-video/utils/applyPollUpdate.ts

@@ -0,0 +1,102 @@
+import { message } from "antd"
+import {
+  isVideoStage,
+  isCopyStage,
+  isImageAnalysisStage,
+  getJobStage,
+  type ViralVideoJob,
+} from "@/api/viral-video/types"
+import type { UIStep, TabTask } from "../types"
+import { TERMINAL_UI_STEPS } from "../types"
+import { copyResultToStoryboard } from "./copyResultToStoryboard"
+import { classifyJobError, cleanErrorMessage } from "./errorClassification"
+
+/**
+ * 根据 job 状态计算下一个 TabTask 快照：单调性守卫 + 阶段推进 + 错误分类。
+ * 返回 null 表示丢弃此次更新（终态守卫命中，使用旧状态即可）。
+ */
+export function applyPollUpdate(t: TabTask, job: ViralVideoJob): TabTask {
+  if (
+    TERMINAL_UI_STEPS.includes(t.uiStep) &&
+    job.status !== "completed" &&
+    job.status !== "failed"
+  ) {
+    return t
+  }
+  let nextStep: UIStep = t.uiStep
+  let copyError = t.copyError
+  let videoError = t.videoError
+  let analysisError = t.analysisError
+  const err = job.error_message || job.error_msg
+  const nextImageAnalysis = job.image_analysis || t.imageAnalysis
+  const nextCopy = job.copy_result
+  const nextStoryboard = nextCopy ? copyResultToStoryboard(nextCopy) || t.storyboard : t.storyboard
+  const stage = getJobStage(job)
+
+  switch (job.status) {
+    case "image_analyzed":
+      nextStep = "step1_done"
+      copyError = undefined
+      break
+    case "copy_generated":
+      nextStep = "step2_copy_ready"
+      copyError = undefined
+      break
+    case "wait_user_confirm":
+      nextStep = t.uiStep === "step3_generating" ? t.uiStep : "step2_copy_ready"
+      break
+    case "running":
+      if (isVideoStage(stage)) nextStep = "step3_generating"
+      else if (isCopyStage(stage)) nextStep = "step2_generating"
+      else if (isImageAnalysisStage(stage)) nextStep = "step1_analyzing"
+      copyError = undefined
+      videoError = undefined
+      break
+    case "completed":
+      nextStep = "step3_done"
+      copyError = undefined
+      videoError = undefined
+      break
+    case "failed": {
+      if (stage === "review" && err && (err.includes("合规") || err.includes("审核未通过"))) {
+        nextStep = "step2_copy_ready"
+        copyError = err
+        videoError = undefined
+        analysisError = undefined
+        message.warning(err)
+        break
+      }
+      nextStep = "failed"
+      const cleaned = cleanErrorMessage(err || "")
+      const kind = classifyJobError(cleaned, stage)
+      if (kind === "video") {
+        videoError = cleaned || "视频生成失败"
+        copyError = undefined
+        analysisError = undefined
+      } else if (kind === "image") {
+        analysisError = cleaned || "图片分析失败"
+        copyError = undefined
+        videoError = undefined
+      } else {
+        copyError = cleaned || "文案生成失败"
+        videoError = undefined
+        analysisError = undefined
+      }
+      break
+    }
+    case "pending":
+      break
+  }
+  return {
+    ...t,
+    job,
+    uiStep: nextStep,
+    imageAnalysis: nextImageAnalysis,
+    storyboard: nextStoryboard,
+    copyError,
+    videoError,
+    analysisError,
+  }
+}
+
+export default applyPollUpdate

apps/web/src/pages/viral-video/utils/copyResultToStoryboard.ts

@@ -0,0 +1,71 @@
+import type { Storyboard } from "../types"
+import type { CopyResult, ShotScript } from "@/api/viral-video/types"
+
+export function copyResultToStoryboard(cr: CopyResult | null | undefined): Storyboard | null {
+  if (!cr) return null
+  if (Array.isArray(cr.shots) && cr.overview) {
+    return {
+      overview: {
+        theme: cr.overview.theme || cr.title || "",
+        total_duration: (() => {
+          const d = cr.overview.total_duration
+          if (!d) return ""
+          return String(d).endsWith("s") ? String(d) : `${d}s`
+        })(),
+        aspect_ratio: cr.overview.aspect_ratio || "9:16",
+      },
+      scene_and_lighting: cr.scene_and_lighting || "",
+      shots: ((cr.shots ?? []) as ShotScript[]).map((s) => ({
+        time_range: s.time_range || "",
+        speaker: s.speaker || "",
+        shot_type_angle_movement: s.shot_type_angle_movement || "",
+        scene_and_dialogue: s.scene_and_dialogue || "",
+        action_details: s.action_details || "",
+        audio_bgm: s.audio_bgm || "",
+        transition: s.transition || "",
+        reference_image_index:
+          typeof s.reference_image_index === "number" ? s.reference_image_index : undefined,
+      })),
+      hard_constraints: Array.isArray(cr.hard_constraints) ? cr.hard_constraints : [],
+      negative_prompts: Array.isArray(cr.negative_prompts) ? cr.negative_prompts : [],
+      voiceover_script: cr.voiceover_script || cr.final_copy || cr.suggested_copy || "",
+      copy_display_markdown: cr.copy_display_markdown || "",
+    }
+  }
+  const finalCopy = cr.final_copy || cr.suggested_copy || ""
+  const title = cr.title || ""
+  return {
+    overview: { theme: title || "AI 分镜脚本", total_duration: "", aspect_ratio: "9:16" },
+    scene_and_lighting: "",
+    shots:
+      cr.scenes && cr.scenes.length
+        ? cr.scenes.map((sc, i: number) => ({
+            time_range: sc.duration
+              ? `${i * sc.duration}-${(i + 1) * sc.duration}s`
+              : `镜头${i + 1}`,
+            shot_type_angle_movement: sc.shot || "",
+            scene_and_dialogue: sc.narration || "",
+            action_details: "",
+            audio_bgm: "",
+            transition: "",
+          }))
+        : finalCopy
+          ? [
+              {
+                time_range: "0-15s",
+                shot_type_angle_movement: "",
+                scene_and_dialogue: finalCopy,
+                action_details: "",
+                audio_bgm: "",
+                transition: "",
+              },
+            ]
+          : [],
+    hard_constraints: [],
+    negative_prompts: [],
+    voiceover_script: finalCopy,
+    copy_display_markdown: cr.copy_display_markdown || "",
+  }
+}
+
+export default copyResultToStoryboard

apps/web/src/pages/viral-video/utils/errorClassification.ts

@@ -0,0 +1,55 @@
+import {
+  isCopyStage,
+  isImageAnalysisStage,
+  isVideoStage,
+  getJobStage,
+} from "@/api/viral-video/types"
+import type { ViralVideoJob, ViralVideoStage } from "@/api/viral-video/types"
+
+/**
+ * 统一的视频/文案/图片阶段错误分类。
+ * 优先按 stage 字段判定；stage 丢失时基于错误消息关键词兜底。
+ */
+export type JobErrorKind = "image" | "copy" | "video"
+
+const VIDEO_ERR_RE =
+  /wan 3\.0|wan-?3|seedance|视频生成|视频下载|视频渲染|tts|配音|render|musetalk|lip.?sync/i
+const IMAGE_ERR_RE = /vlm|图片分析|vision|商品识别|ocr|image analysis|图片识别|图像分析/i
+
+export function classifyJobError(err: string, stage?: string | null): JobErrorKind {
+  const msg = err || ""
+  if (stage) {
+    const s = stage as ViralVideoStage
+    if (isVideoStage(s)) return "video"
+    if (isImageAnalysisStage(s)) return "image"
+    if (isCopyStage(s)) return "copy"
+  }
+  if (VIDEO_ERR_RE.test(msg)) return "video"
+  if (IMAGE_ERR_RE.test(msg)) return "image"
+  return "copy"
+}
+
+/**
+ * 去掉后端错误消息中常见的"XX失败："前缀，避免前端 UI 重复前缀。
+ */
+const ERROR_PREFIX_RE =
+  /^(视频生成失败|文案生成失败|图片分析失败|Wan 3\.0 视频生成失败|Seedance[0-9. ]*视频生成失败)[：:]\s*/g
+
+export function cleanErrorMessage(err: string): string {
+  if (!err) return ""
+  return err.replace(ERROR_PREFIX_RE, "").trim()
+}
+
+/**
+ * 基于 job 统一提取错误分类和清理后的消息；无错误返回 null。
+ */
+export function extractJobError(
+  job: ViralVideoJob | null | undefined,
+): { kind: JobErrorKind; msg: string } | null {
+  if (!job) return null
+  const raw = job.error_message || job.error_msg
+  if (!raw) return null
+  const stage = getJobStage(job)
+  const kind = classifyJobError(raw, stage)
+  return { kind, msg: cleanErrorMessage(raw) }
+}

apps/web/src/pages/viral-video/utils/renderMarkdown.ts

@@ -0,0 +1,70 @@
+/**
+ * 极简 markdown 渲染器：仅支持粗体、斜体、标题、列表、换行。
+ * 不依赖 marked 包（node_modules 中暂未安装），满足 viral-video 文案预览的
+ * 简单排版需求。
+ */
+function escapeHtml(s: string): string {
+  return s
+    .replace(/&/g, "&")
+    .replace(/</g, "&lt;")
+    .replace(/>/g, "&gt;")
+    .replace(/"/g, "&quot;")
+}
+
+function inline(md: string): string {
+  let s = escapeHtml(md)
+  // bold **xxx**
+  s = s.replace(/\*\*([^*]+)\*\*/g, "<strong>$1</strong>")
+  // italic *xxx*
+  s = s.replace(/(^|[^*])\*([^*\n]+)\*/g, "$1<em>$2</em>")
+  // inline code `xxx`
+  s = s.replace(/`([^`]+)`/g, "<code>$1</code>")
+  return s
+}
+
+export function renderMarkdown(md: string): string {
+  if (!md) return ""
+  const lines = md.split(/\r?\n/)
+  const out: string[] = []
+  let inList = false
+  for (const raw of lines) {
+    const line = raw.trimEnd()
+    if (!line.trim()) {
+      if (inList) {
+        out.push("</ul>")
+        inList = false
+      }
+      continue
+    }
+    // headings
+    const h = line.match(/^(#{1,4})\s+(.*)$/)
+    if (h) {
+      if (inList) {
+        out.push("</ul>")
+        inList = false
+      }
+      const lvl = h[1].length
+      out.push(`<h${lvl}>${inline(h[2])}</h${lvl}>`)
+      continue
+    }
+    // unordered list
+    const li = line.match(/^\s*[-*]\s+(.*)$/)
+    if (li) {
+      if (!inList) {
+        out.push("<ul>")
+        inList = true
+      }
+      out.push(`<li>${inline(li[1])}</li>`)
+      continue
+    }
+    if (inList) {
+      out.push("</ul>")
+      inList = false
+    }
+    out.push(`<p>${inline(line)}</p>`)
+  }
+  if (inList) out.push("</ul>")
+  return out.join("\n")
+}
+
+export default renderMarkdown

apps/web/src/pages/viral-video/utils/resolveModelSwitch.ts

@@ -0,0 +1,30 @@
+import { RESOLUTION_ORDER } from "../ViralVideoPage.constants"
+
+/**
+ * resolveModelSwitch 所需的最小模型信息（结构性类型，兼容 ViralVideoModel）。
+ */
+export interface ModelSwitchInfo {
+  supported_resolutions?: string[]
+  max_duration?: number
+}
+
+/**
+ * 切换视频模型时，联动修正分辨率/时长：新模型不支持当前分辨率则回退到最低支持，
+ * 当前时长超过新模型 max_duration 则截断到 max_duration。
+ */
+export function resolveModelSwitch(
+  cur: { quality: string; duration: number },
+  m: ModelSwitchInfo,
+): { quality: string; duration: number } {
+  const mRes = Array.isArray(m.supported_resolutions)
+    ? m.supported_resolutions
+    : ["480p", "720p", "1080p"]
+  const sortedRes = mRes.slice().sort((a, b) => {
+    const ra = RESOLUTION_ORDER.indexOf(a)
+    const rb = RESOLUTION_ORDER.indexOf(b)
+    return (rb === -1 ? -1 : rb) - (ra === -1 ? -1 : ra)
+  })
+  const newRes = mRes.includes(cur.quality) ? cur.quality : sortedRes[0] || "480p"
+  const newDur = Math.min(cur.duration, typeof m.max_duration === "number" ? m.max_duration : 15)
+  return { quality: newRes, duration: newDur }
+}

apps/web/src/router/appRoutes.tsx

@@ -116,31 +116,6 @@ const appChildren: RouteObject[] = [
     path: "profile",
     lazy: lazyRoute(() => import("@/pages/profile/Settings")),
   },
-  {
-    path: "admin",
-    children: [
-      {
-        index: true,
-        lazy: lazyRoute(() => import("@/pages/admin/AdminComingSoon")),
-      },
-      {
-        path: "users",
-        lazy: lazyRoute(() => import("@/pages/admin/AdminComingSoon")),
-      },
-      {
-        path: "analytics",
-        lazy: lazyRoute(() => import("@/pages/admin/AdminComingSoon")),
-      },
-      {
-        path: "monitor",
-        lazy: lazyRoute(() => import("@/pages/admin/AdminComingSoon")),
-      },
-      {
-        path: "logs",
-        lazy: lazyRoute(() => import("@/pages/admin/AdminComingSoon")),
-      },
-    ],
-  },
 ]
 
 export const appRoutes: RouteObject = {

apps/web/src/test/pages/AdminComingSoon.test.tsx

@@ -1,32 +0,0 @@
-import { describe, expect, it, vi } from "vitest"
-import { render, screen } from "@testing-library/react"
-import { MemoryRouter } from "react-router-dom"
-import AdminComingSoon from "@/pages/admin/AdminComingSoon"
-
-vi.mock("react-router-dom", async () => {
-  const actual = await vi.importActual("react-router-dom")
-  return {
-    ...actual,
-    useNavigate: () => vi.fn(),
-  }
-})
-
-describe("AdminComingSoon Page", () => {
-  it("should render without crashing", () => {
-    render(
-      <MemoryRouter>
-        <AdminComingSoon />
-      </MemoryRouter>,
-    )
-    expect(screen.getByText("Admin 后台暂未开放")).toBeTruthy()
-  })
-
-  it("should render back button", () => {
-    render(
-      <MemoryRouter>
-        <AdminComingSoon />
-      </MemoryRouter>,
-    )
-    expect(screen.getByText("返回首页")).toBeTruthy()
-  })
-})

apps/worker/worker_app/celery_app.py

@@ -53,6 +53,9 @@ celery_app.conf.imports = (
     # #1998 GPU MuseTalk 异步推理：wait_for_result→签名 URL→回写 lipsync_jobs
     # 必须在 Worker 侧注册，否则 apply_async 消息无人消费，job 永远卡在 processing
     "app.tasks.lipsync_gpu",
+    # #2076 Ditto 蚂蚁数字人异步推理：同步 HTTP 调用 Ditto → MP4 流转存 OSS → 回写 lipsync_jobs
+    # 必须在 Worker 侧注册；失败回退 GPU MuseTalk → MediaKit
+    "app.tasks.lipsync_ditto",
 )
 
 # Celery Beat 定时任务调度

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

@@ -1,13 +1,13 @@
 # -*- coding: utf-8 -*-
-"""V2 prompt 解析：优先读后台 viral_video_prompt_templates 表（prompt_type='image_analysis'
-且 is_active=true），30s TTL 热加载；DB 无有效记录/异常时，fallback 到纯硬编码 JSON schema prompt。
+"""V2 prompt 解析：优先读后台 viral_video_prompt_templates（prompt_type='image_analysis'
+且 is_active=true），30s TTL 热加载；DB 无有效记录/异常时，fallback 到 prompts.py 的
+image_analysis v8 默认 system/user。
 
-规则（简单直接，不做字符串匹配判断）：
-- DB 有 is_active=true 的 image_analysis 记录（含种子版本和用户修改后的版本）：
-  * system = DB.system_prompt（DB prompt 自带完整输出格式，不追加硬编码 schema，
-    避免 DB 写 XML、调用强制 json_object 造成的格式冲突）
-  * user   = DB.user_prompt_template 渲染后使用；渲染后为空则用硬编码默认
-- DB 无记录/连接异常/返回空：system/user 全部用纯硬编码 JSON schema prompt
+规则：
+- DB 有 is_active=true 的 image_analysis 记录：system 原样用 DB.system_prompt
+  （自带完整输出格式，不追加任何硬编码 schema），user 用 DB.user_prompt_template
+  渲染（填入 image_url / ocr_text）；
+- DB 无记录/异常：system/user 用 prompts.py 里的 v8 默认模板。
 """
 
 from __future__ import annotations
@@ -15,192 +15,79 @@ from __future__ import annotations
 import logging
 import threading
 import time
-from typing import Any
 
 logger = logging.getLogger(__name__)
 
-# ---- 纯硬编码 JSON schema（DB 无有效配置时全量使用） ----
 
-_FAST_JSON_SCHEMA = (
-    "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,\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；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,\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'
-    "}\n\n"
-    "你必须只返回一个合法的JSON对象，不要输出任何其他文字、解释、XML标签或markdown。"
-)
-DEFAULT_FAST_USER = "识别这张图片的人物穿搭与主体信息，只返回JSON对象。"
+def _default_template() -> dict:
+    # 延迟导入：避免模块加载时拉起整个 packages 依赖链（也便于旧 Python 收集测试）
+    from packages.application.viral_video.prompts import DEFAULT_TEMPLATES
 
-_PRO_JSON_SCHEMA = (
-    "你是图片分析专家。严格按下方 JSON schema 返回一个对象，不要解释、不要markdown、不要代码块、不要XML标签。\n"
-    "{\n"
-    '  "has_person": true/false,\n'
-    '  "gender": "男"/"女"/null,\n'
-    '  "age_range": "儿童"/"青少年"/"青年"/"中年"/"老年"/null,\n'
-    '  "outfit": "整体穿着描述（含颜色款式）",\n'
-    '  "hair": "发型发色",\n'
-    '  "pose": "姿势",\n'
-    '  "expression": "表情",\n'
-    '  "scene": "场景",\n'
-    '  "mood": "氛围",\n'
-    '  "has_product": true/false,\n'
-    '  "category": "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，非产品图填null",\n'
-    '  "brand": "品牌，无则null",\n'
-    '  "key_features": ["核心特征数组，3-6个短语"]\n'
-    "}\n\n"
-    "你必须只返回一个合法的JSON对象，不要输出任何其他文字、解释、XML标签或markdown。"
-)
-DEFAULT_PRO_USER = "分析这张图片，返回符合schema的JSON。"
+    for item in DEFAULT_TEMPLATES:
+        if item["prompt_type"] == "image_analysis":
+            return item
+    raise RuntimeError("image_analysis 默认模板缺失")
 
-# 保留旧 JSON schema 追加文本作为常量（DB prompt 完全控制输出格式后不再使用，
-# 保留以便排查历史行为）。
-_FAST_JSON_APPEND = (
-    "\n\n【输出格式要求】无论上文如何要求，最终你必须只返回一个合法的JSON对象，"
-    "严格包含以下字段（字段值不确定时填null或空数组）：\n"
-    "{\n"
-    '  "has_person": true/false,\n'
-    '  "gender": "男"/"女"/null,\n'
-    '  "age_range": "儿童"/"青少年"/"青年"/"中年"/"老年"/null,\n'
-    '  "upper_wear": "上装款式字符串",\n'
-    '  "upper_color": "上装主色",\n'
-    '  "lower_wear": "下装款式（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": ["配饰数组"],\n'
-    '  "hairstyle": "发型",\n'
-    '  "expression": "表情",\n'
-    '  "pose": "姿势",\n'
-    '  "scene": "场景",\n'
-    '  "style": "风格",\n'
-    '  "has_product": true/false,\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",\n'
-    '  "pattern": "图案",\n'
-    '  "colors": ["主色数组"],\n'
-    '  "mood": "整体氛围"\n'
-    "}\n"
-    "不要输出任何其他文字、解释、XML标签或markdown。"
-)
-
-_PRO_JSON_APPEND = (
-    "\n\n【输出格式要求】无论上文如何要求，最终你必须只返回一个合法的JSON对象，"
-    "严格包含以下字段（字段值不确定时填null或空数组）：\n"
-    "{\n"
-    '  "has_person": true/false,\n'
-    '  "gender": "男"/"女"/null,\n'
-    '  "age_range": "儿童"/"青少年"/"青年"/"中年"/"老年"/null,\n'
-    '  "outfit": "整体穿着描述（含颜色款式）",\n'
-    '  "hair": "发型发色",\n'
-    '  "pose": "姿势",\n'
-    '  "expression": "表情",\n'
-    '  "scene": "场景",\n'
-    '  "mood": "氛围",\n'
-    '  "has_product": true/false,\n'
-    '  "category": "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，非产品图填null",\n'
-    '  "brand": "品牌，无则null",\n'
-    '  "key_features": ["核心特征3-6个短语"]\n'
-    "}\n"
-    "不要输出任何其他文字、解释、XML标签或markdown。"
-)
 
 _cache_lock = threading.Lock()
-_cache: dict[str, tuple[float, Any]] = {}
+_cache: dict[str, tuple[float, tuple[str, str]]] = {}
 _CACHE_TTL = 30.0
 
 
-def _load_db_template() -> Any | None:
-    """直接查DB viral_video_prompt_templates 中 is_active=true 的 image_analysis 记录；
-    DB不可达/无记录/异常返回None。
-    复用 prompt_loader._load_from_db，它只查DB不做DEFAULT_TEMPLATES fallback，
-    返回None表示DB无记录或异常。"""
+def _load_db_template():
+    """查 DB is_active=true 的 image_analysis 记录；不可达/无记录返回 None。"""
     try:
         from packages.application.viral_video.prompt_loader import _load_from_db
 
         return _load_from_db("image_analysis")
-    except Exception as e:
-        logger.warning("[vision.v2] 查询DB prompt配置失败: %s", e)
+    except Exception as e:  # noqa: BLE001
+        logger.warning("[vision.v2] 查询DB image_analysis prompt失败: %s", e)
         return None
 
 
-def _render_user(tpl: Any | None, default_user: str) -> str:
-    if not tpl:
-        return default_user
-    tpl_str = getattr(tpl, "user_prompt_template", "") or ""
-    if not tpl_str.strip():
-        return default_user
-    rendered = tpl_str.replace("{image_count}", "1").replace("{industry}", "通用").replace("{image_urls}", "").strip()
-    return rendered or default_user
+def _render_user(user_tpl: str, image_url: str, ocr_text: str) -> str:
+    try:
+        return user_tpl.format(image_url=image_url, ocr_text=ocr_text or "无")
+    except Exception:  # noqa: BLE001
+        return user_tpl
 
 
-def resolve_fast_prompt() -> tuple[str, str]:
-    return _resolve("fast")
-
-
-def resolve_pro_prompt() -> tuple[str, str]:
-    return _resolve("pro")
-
-
-def _resolve(kind: str) -> tuple[str, str]:
+def _resolve(kind: str, image_url: str = "", ocr_text: str = "") -> tuple[str, str]:
     now = time.time()
     cache_key = f"prompt_{kind}"
     with _cache_lock:
         hit = _cache.get(cache_key)
         if hit and now - hit[0] < _CACHE_TTL:
-            return hit[1]
+            sys_prompt, usr_prompt = hit[1]
+            return sys_prompt, _render_user(usr_prompt, image_url, ocr_text)
 
-    default_sys = _FAST_JSON_SCHEMA if kind == "fast" else _PRO_JSON_SCHEMA
-    default_user = DEFAULT_FAST_USER if kind == "fast" else DEFAULT_PRO_USER
+    default = _default_template()
+    sys_prompt = default["system_prompt"]
+    usr_prompt = default["user_prompt_template"]
 
-    sys_prompt = default_sys
-    usr_prompt = default_user
-    try:
-        tpl = _load_db_template()
-        if tpl is not None:
-            db_sys = (getattr(tpl, "system_prompt", "") or "").strip()
-            if db_sys:
-                sys_prompt = db_sys  # DB prompt自带完整输出格式，不追加硬编码schema避免冲突
-                usr_prompt = _render_user(tpl, default_user)
-                logger.info(
-                    "[vision.v2] 使用DB image_analysis prompt (kind=%s version=%s sys_len=%d)",
-                    kind,
-                    getattr(tpl, "version", "?"),
-                    len(db_sys),
-                )
-            else:
-                logger.debug("[vision.v2] DB image_analysis system_prompt为空，使用默认JSON (kind=%s)", kind)
-        else:
-            logger.debug("[vision.v2] DB无image_analysis记录/不可达，使用默认JSON prompt (kind=%s)", kind)
-    except Exception as e:
-        logger.warning("[vision.v2] 解析DB prompt异常，使用默认: %s", e)
+    tpl = _load_db_template()
+    if tpl is not None:
+        db_sys = (getattr(tpl, "system_prompt", "") or "").strip()
+        if db_sys:
+            sys_prompt = db_sys
+            db_usr = getattr(tpl, "user_prompt_template", "") or usr_prompt
+            usr_prompt = db_usr or usr_prompt
+            logger.info(
+                "[vision.v2] 使用DB image_analysis prompt version=%s",
+                getattr(tpl, "version", "?"),
+            )
 
     with _cache_lock:
         _cache[cache_key] = (now, (sys_prompt, usr_prompt))
-    return sys_prompt, usr_prompt
+    return sys_prompt, _render_user(usr_prompt, image_url, ocr_text)
+
+
+def resolve_fast_prompt(image_url: str = "", ocr_text: str = "") -> tuple[str, str]:
+    return _resolve("fast", image_url, ocr_text)
+
+
+def resolve_pro_prompt(image_url: str = "", ocr_text: str = "") -> tuple[str, str]:
+    return _resolve("pro", image_url, ocr_text)
 
 
 def invalidate_cache() -> None:

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

@@ -1,12 +1,11 @@
 # -*- coding: utf-8 -*-
-"""V2 图片分析主路径：每图并行 OCR（火山MediaKit，未配置时自动跳过）+ qwen3.8-flash JSON VLM，
-失败时单次 qwen3.7-plus 兜底。
+"""V2 图片分析主路径（v8 叙述优先）：每图并行 OCR（火山 MediaKit，未配置自动跳过）
++ fast VLM 强约束 JSON；失败时单次 pro VLM 兜底。
 
-架构（灵应10-05确认）：
-- 唯一后端：阿里云百炼 DashScope，qwen3.8-flash 做快速路径、qwen3.7-plus 做兜底
-- 主力：单图2路并行（OCR + fast VLM），外层N图全并发（workers=8）
-- 兜底：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
+架构：
+- 单图 2 路并行（OCR + fast VLM），外层 N 图全并发（workers=8）；
+- 兜底单次 pro VLM，无竞速/复杂重试；
+- 输出统一为 5 字段 image dict（type/name/brand/has_person/summary_markdown）。
 """
 
 from __future__ import annotations
@@ -21,36 +20,24 @@ from . import assembler, ocr_volc, vlm_fallback, vlm_fast_json
 
 logger = logging.getLogger(__name__)
 
-# 超时（可通过环境变量覆盖）
 _IMG_WORKERS = int(os.environ.get("VISION_V2_IMG_WORKERS", "8"))
 _FAST_TIMEOUT = float(os.environ.get("VISION_V2_FAST_TIMEOUT", "20"))
 _FAST_JSON_TIMEOUT = float(os.environ.get("VISION_V2_FAST_JSON_TIMEOUT", "20"))
 _OCR_TIMEOUT = float(os.environ.get("VISION_V2_OCR_TIMEOUT", "6"))
 _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] | None) -> bool:
+    if not isinstance(r, dict):
+        return False
+    return bool((r.get("summary_markdown") or "").strip())
 
 
-def _is_usable(r: dict[str, Any]) -> bool:
-    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 _basic_failure(ocr_result: list[str], fast_elapsed: float, source: str) -> dict[str, Any]:
+    image = assembler.assemble_result(-1, {}, ocr_result)
+    image["_source"] = source
+    image["_fast_elapsed"] = round(fast_elapsed, 2)
+    return image
 
 
 def analyze_image_v2(idx: int, img_url: str) -> dict[str, Any]:
@@ -58,7 +45,6 @@ def analyze_image_v2(idx: int, img_url: str) -> dict[str, Any]:
 
     fj_result: dict[str, Any] | None = None
     ocr_result: list[str] = []
-    fast_elapsed = 0.0
     pool = ThreadPoolExecutor(max_workers=2)
     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)
@@ -66,7 +52,7 @@ def analyze_image_v2(idx: int, img_url: str) -> dict[str, Any]:
         for fut in as_completed([f_fj, f_ocr], timeout=_FAST_TIMEOUT):
             try:
                 res = fut.result(timeout=1)
-            except Exception as e:
+            except Exception as e:  # noqa: BLE001
                 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):
@@ -80,37 +66,24 @@ def analyze_image_v2(idx: int, img_url: str) -> dict[str, Any]:
         logger.warning("[vision.v2] 图片 #%d fast路径超时(%.0fs)，走pro兜底", idx, _FAST_TIMEOUT)
     finally:
         fast_elapsed = time.time() - t0
-        pool.shutdown(wait=False)  # 不等待未完成的线程，避免计时膨胀
+        pool.shutdown(wait=False)
 
     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],
-            )
+            logger.info("[vision.v2] 图片 #%d fast命中 elapsed=%.2fs", idx, fast_elapsed)
             return assembled
 
-    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 = vlm_fallback.call_pro_vlm(img_url, idx, ocr_hint=ocr_result, timeout=_PRO_TIMEOUT)
+    if _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
 
     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
+    return _basic_failure(ocr_result, fast_elapsed, "v2_all_failed")
 
 
 def analyze_images_v2(img_urls: list[str]) -> list[dict[str, Any]]:
@@ -133,14 +106,19 @@ def analyze_images_v2(img_urls: list[str]) -> list[dict[str, Any]]:
             idx = future_to_idx[fut]
             try:
                 results[idx] = fut.result()
-            except Exception as e:
+            except Exception as e:  # noqa: BLE001
                 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
+                results[idx] = assembler.assemble_result(idx, {}, [])
+                results[idx]["_source"] = "v2_future_exception"  # type: ignore[index]
 
     elapsed = time.time() - t0
-    succ = sum(1 for r in results if r and _is_usable(r))
+    succ = sum(1 for r in results if _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]
+    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]  # type: ignore[misc]

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

@@ -1,14 +1,8 @@
 # -*- coding: utf-8 -*-
-"""V2 兜底路径：image_analysis（默认 qwen-vl-plus 视觉模型，fallback qwen3.7-plus / DashScope）单图调用。
+"""V2 兜底路径：vision client（fallback 变体）单图调用，走 v8 叙述优先 prompt。
 
-fast_json 超时/返回非 JSON/识别为空时，本路径单次调用兜底。
-设计要点：
-- 通过 ai_router.get_vision_client() 获取 DoubaoClient 实例，不再自己拼 httpx 请求
-- enable_thinking=False + response_format=json_object
-- system prompt 优先读后台 viral_video_prompt_templates 配置，DB不可用时fallback到硬编码JSON schema
-- max_tokens 不传，使用 client 中 capability 的 DB 配置（避免硬编码截断 JSON）
-- timeout=30s
-- 返回 dict 统一走 assembler.assemble_result 组装，与 fast 路径输出格式完全一致
+fast 超时/非 JSON/为空时单次调用；输出统一走 assembler.assemble_result 组装，
+与 fast 路径同为 5 字段 image dict。
 """
 
 from __future__ import annotations
@@ -28,11 +22,12 @@ def call_pro_vlm(
     img_url: str,
     idx: int,
     *,
+    ocr_hint: list[str] | None = None,
     timeout: int = _DEFAULT_TIMEOUT,
     max_tokens: int | None = None,
 ) -> dict[str, Any] | None:
-    """max_tokens 默认 None：不显式传参，使用 client 内 capability 的 DB 配置。"""
     t0 = time.time()
+    ocr_text = "、".join(t for t in (ocr_hint or []) if t)[:200]
 
     try:
         from packages.shared.ai_router import ai_router
@@ -41,13 +36,13 @@ def call_pro_vlm(
         if not client or not client.is_available:
             logger.warning("[vision.v2] pro vision client 不可用，跳过")
             return None
-    except Exception as e:
+    except Exception as e:  # noqa: BLE001
         logger.warning("[vision.v2] ai_router 获取失败: %s", e)
         return None
 
-    system_prompt, user_prompt = _prompt.resolve_pro_prompt()
+    system_prompt, user_prompt = _prompt.resolve_pro_prompt(img_url, ocr_text)
 
-    messages = [
+    messages: list[dict[str, Any]] = [
         {"role": "system", "content": system_prompt},
         {
             "role": "user",
@@ -61,31 +56,31 @@ def call_pro_vlm(
     try:
         call_kwargs: dict[str, Any] = {
             "messages": messages,
-            "images": None,  # 图片已在 messages 中
+            "images": None,
             "temperature": 0.3,
             "timeout": timeout,
             "enable_thinking": False,
             "response_format": {"type": "json_object"},
+            "max_tokens": max_tokens if max_tokens is not None else 4000,
         }
-        # pro fallback：显式4000 tokens给复杂门店图留足空间
-        call_kwargs["max_tokens"] = max_tokens if max_tokens is not None else 4000
 
         from .json_utils import extract_json_object
 
-        raw = None
         obj = None
         for _outer in range(2):
             kw = dict(call_kwargs)
             if _outer == 1:
                 kw.pop("response_format", None)
                 msgs2 = [dict(messages[0]), dict(messages[1])]
-                cont = [dict(c) for c in list(msgs2[1]["content"])]
-                cont[-1] = {"type": "text", "text": user_prompt + "\n严格只输出JSON对象，不要解释或markdown。"}
+                cont = [dict(c) for c in msgs2[1]["content"]]
+                cont[-1] = {
+                    "type": "text",
+                    "text": user_prompt + "\n严格只输出JSON对象，不要解释或markdown。",
+                }
                 msgs2[1] = {"role": "user", "content": cont}
                 kw["messages"] = msgs2
             raw = client.vision_completion(**kw)
             if not raw:
-                logger.warning("[vision.v2] pro 返回空 outer=%s", _outer)
                 continue
             obj = extract_json_object(raw)
             if obj is not None:
@@ -97,20 +92,18 @@ def call_pro_vlm(
             logger.warning("[vision.v2] pro 两次均未得到JSON elapsed=%.1fs", elapsed)
             return None
         if obj.get("_partial"):
-            logger.warning("[vision.v2] pro 返回截断JSON(partial) elapsed=%.1fs", elapsed)
-        logger.info(
-            "[vision.v2] pro 完成 model=%s elapsed=%.1fs type=%s",
-            client.model,
-            elapsed,
-            obj.get("type"),
-        )
+            logger.warning("[vision.v2] pro 截断JSON(partial) elapsed=%.1fs", elapsed)
 
-        # 通过assembler统一组装，兼容v4嵌套schema和旧扁平schema
-        result = assembler.assemble_result(idx, obj, [])
+        result = assembler.assemble_result(idx, obj, ocr_hint or [])
         result["_source"] = "vlm_pro"
         result["_fallback_used"] = True
+        logger.info("[vision.v2] pro 完成 model=%s elapsed=%.1fs", client.model, elapsed)
         return result
-    except Exception as e:
-        elapsed = time.time() - t0
-        logger.warning("[vision.v2] pro 异常 elapsed=%.1fs err=%s", elapsed, e, exc_info=True)
+    except Exception as e:  # noqa: BLE001
+        logger.warning(
+            "[vision.v2] pro 异常 elapsed=%.1fs err=%s",
+            time.time() - t0,
+            e,
+            exc_info=True,
+        )
         return None

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

@@ -1,15 +1,11 @@
 # -*- coding: utf-8 -*-
-"""V2 快速路径：image_analysis capability（默认 qwen-vl-plus 视觉模型 / DashScope）强约束 JSON-only 调用。
+"""V2 快速路径：vision client（默认 image_analysis 能力）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：
-- 通过 ai_router.get_vision_client() 获取 DoubaoClient 实例，不再自己拼 httpx 请求
-- enable_thinking=False 关闭推理链（reasoning 是延迟主因）
-- response_format=json_object 强约束JSON输出
-- system prompt 优先读后台 viral_video_prompt_templates 配置，DB不可用时fallback到硬编码JSON schema
-- max_tokens 不传，使用 client 中 capability 的 DB 配置（避免硬编码截断 JSON）
-- temperature=0.1（稳定输出 JSON）
-- timeout=15s（失败由外层走 pro 兜底）
+要点：
+- 通过 ai_router.get_vision_client() 获取 client；
+- enable_thinking=False 关闭推理链，response_format=json_object 强约束 JSON；
+- system/user prompt 优先读后台模板（v8 叙述优先），DB 不可用时用 prompts.py 默认；
+- temperature=0.1（稳定输出 JSON）；两次尝试（第二次去 json_object 约束）。
 """
 
 from __future__ import annotations
@@ -31,11 +27,6 @@ def call_fast_json(
     timeout: int = _DEFAULT_TIMEOUT,
     max_tokens: int | None = None,
 ) -> dict[str, Any] | None:
-    """调用 vision client 返回结构化 dict；失败/非 JSON 返回 None。
-
-    max_tokens 默认 None：不显式传参，使用 client 内 capability 的 DB 配置；
-    显式传入时作为覆盖。
-    """
     t0 = time.time()
 
     try:
@@ -45,13 +36,13 @@ def call_fast_json(
         if not client or not client.is_available:
             logger.warning("[vision.v2] vision client 不可用，跳过 fast_json")
             return None
-    except Exception as e:
+    except Exception as e:  # noqa: BLE001
         logger.warning("[vision.v2] ai_router 获取失败: %s", e)
         return None
 
-    system_prompt, user_prompt = _prompt.resolve_fast_prompt()
+    system_prompt, user_prompt = _prompt.resolve_fast_prompt(img_url, "")
 
-    messages = [
+    messages: list[dict[str, Any]] = [
         {"role": "system", "content": system_prompt},
         {
             "role": "user",
@@ -65,7 +56,7 @@ def call_fast_json(
     try:
         call_kwargs: dict[str, Any] = {
             "messages": messages,
-            "images": None,  # 图片已在 messages 中
+            "images": None,
             "temperature": 0.1,
             "timeout": timeout,
             "enable_thinking": False,
@@ -74,50 +65,47 @@ def call_fast_json(
         if max_tokens is not None:
             call_kwargs["max_tokens"] = max_tokens
 
-        # 双重防护：第1次正常调用；第2次去掉json_object强约束（部分模型在该约束下
-        # 反而幻觉），并加严格指令。解析全部走 json_utils，截断partial产物可用。
         from .json_utils import extract_json_object
 
-        raw = None
         obj = None
         for _outer in range(2):
             kw = dict(call_kwargs)
             if _outer == 1:
                 kw.pop("response_format", None)
                 msgs2 = [dict(messages[0]), dict(messages[1])]
-                cont = list(msgs2[1]["content"])
-                cont = [dict(c) for c in cont]
-                cont[-1] = {"type": "text", "text": user_prompt + "\n严格只输出JSON对象，不要解释或markdown。"}
+                cont = [dict(c) for c in msgs2[1]["content"]]
+                cont[-1] = {
+                    "type": "text",
+                    "text": user_prompt + "\n严格只输出JSON对象，不要解释或markdown。",
+                }
                 msgs2[1] = {"role": "user", "content": cont}
                 kw["messages"] = msgs2
             raw = client.vision_completion(**kw)
             if not raw:
-                logger.warning("[vision.v2] fast_json 返回空 outer=%s", _outer)
                 continue
             obj = extract_json_object(raw)
             if obj is not None:
                 break
-            logger.warning(
-                "[vision.v2] fast_json 非JSON(100字) outer=%s: %s",
-                _outer,
-                raw[:100],
-            )
+            logger.warning("[vision.v2] fast_json 非JSON(100字) outer=%s: %s", _outer, raw[:100])
 
         elapsed = time.time() - t0
         if obj is None:
             logger.warning("[vision.v2] fast_json 两次均未得到JSON elapsed=%.1fs", elapsed)
             return None
         if obj.get("_partial"):
-            logger.warning("[vision.v2] fast_json 返回截断JSON(partial) elapsed=%.1fs", elapsed)
+            logger.warning("[vision.v2] fast_json 截断JSON(partial) elapsed=%.1fs", elapsed)
         logger.info(
-            "[vision.v2] fast_json 完成 model=%s elapsed=%.1fs has_person=%s type=%s",
+            "[vision.v2] fast_json 完成 model=%s elapsed=%.1fs type=%s",
             client.model,
             elapsed,
-            obj.get("has_person"),
             obj.get("type"),
         )
         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)
+    except Exception as e:  # noqa: BLE001
+        logger.warning(
+            "[vision.v2] fast_json 异常 elapsed=%.1fs err=%s",
+            time.time() - t0,
+            e,
+            exc_info=True,
+        )
         return None

deploy/configs/.env.staging

@@ -311,3 +311,10 @@ GPU_ENCODE_CRF=23
 GPU_ENCODE_FALLBACK_CPU=true
 GPU_ENCODE_MEZZANINE_TRANSPORT=oss
 GPU_ENCODE_OSS_TMP_PREFIX=tmp/gpu-mezzanine/
+
+# ==================== Ditto 蚂蚁数字人口型 ====================
+# 注意：这些值必须写死在模板里（不是 CI Secret），否则每次 CI 重新渲染 .env 都会被丢弃，
+# 导致 staging 发版后 Ditto 口型服务静默降级到 GPU/MediaKit（P0 防复发）。
+USE_DITTO_LIPSYNC=true
+DITTO_API_BASE_URL=http://100.76.80.23:8000
+DITTO_DEFAULT_VIDEO_URL=https://xiaoxia-autocut.oss-cn-hangzhou.aliyuncs.com/uploads/default_avatar.mp4

infra/docker/nginx-production.conf

@@ -40,7 +40,9 @@ server {
         proxy_set_header X-Real-IP $remote_addr;
         proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
         proxy_set_header X-Forwarded-Proto $scheme;
-        proxy_read_timeout 300s;
+        proxy_set_header Upgrade $http_upgrade;
+        proxy_set_header Connection "upgrade";
+        proxy_read_timeout 3600s;
         proxy_send_timeout 300s;
         proxy_request_buffering off;
     }

infra/docker/nginx-staging.conf

@@ -39,7 +39,9 @@ server {
         proxy_set_header X-Real-IP $remote_addr;
         proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
         proxy_set_header X-Forwarded-Proto $scheme;
-        proxy_read_timeout 300s;
+        proxy_set_header Upgrade $http_upgrade;
+        proxy_set_header Connection "upgrade";
+        proxy_read_timeout 3600s;
         proxy_send_timeout 300s;
         proxy_request_buffering off;
     }

packages/adapters/sqlalchemy_impl/models.py

@@ -961,6 +961,7 @@ class ViralVideoJobModel(Base):
     # v1.5 音频/视频参数
     voice_id = Column(String(200), nullable=False, default="")
     voice_source = Column(String(20), nullable=False, default="")
+    language = Column(String(20), nullable=False, default="zh-CN")
     video_ratio = Column(String(10), nullable=False, default="9:16")
     video_model = Column(String(100), nullable=False, default="")
     # 结果与状态
@@ -1019,3 +1020,20 @@ class ViralVideoPromptTemplateModel(Base):
     is_active = Column(Boolean, nullable=False, default=True)
     created_at = Column(DateTime(timezone=True), nullable=False, default=lambda: datetime.now(UTC))
     updated_at = Column(DateTime(timezone=True), nullable=False, default=lambda: datetime.now(UTC))
+
+
+class SystemSettingModel(Base):
+    """系统配置表 ORM 模型（#2246：后台可配置项，表已手工存在于 staging）."""
+
+    __tablename__ = "system_settings"
+
+    id = Column(String(36), primary_key=True)
+    setting_key = Column(String(100), nullable=False, unique=True)
+    setting_value = Column(Text, nullable=True)
+    setting_type = Column(String(20), nullable=False)
+    description = Column(String(255), nullable=False, default="", server_default="")
+    is_public = Column(Boolean, nullable=False, default=False, server_default="false")
+    updated_by = Column(String(36), nullable=True)
+    category = Column(String(50), nullable=False, default="general", server_default="general")
+    created_at = Column(DateTime(timezone=True), nullable=False, default=lambda: datetime.now(UTC))
+    updated_at = Column(DateTime(timezone=True), nullable=False, default=lambda: datetime.now(UTC))

packages/adapters/sqlalchemy_impl/system_setting_repository.py

@@ -0,0 +1,65 @@
+"""system_settings 表 SQLAlchemy Repository — #2246."""
+
+from __future__ import annotations
+
+from sqlalchemy.orm import Session
+
+from packages.adapters.sqlalchemy_impl.models import SystemSettingModel
+from packages.domain.system_setting import SystemSetting
+
+
+class SQLAlchemySystemSettingRepository:
+    def __init__(self, session: Session):
+        self.session = session
+
+    def get_by_key(self, setting_key: str) -> SystemSetting | None:
+        model = self.session.query(SystemSettingModel).filter(SystemSettingModel.setting_key == setting_key).first()
+        if model is None:
+            return None
+        return self._to_domain(model)
+
+    def list_all(self, category: str | None = None) -> list[SystemSetting]:
+        query = self.session.query(SystemSettingModel)
+        if category is not None:
+            query = query.filter(SystemSettingModel.category == category)
+        return [self._to_domain(m) for m in query.all()]
+
+    def upsert(self, setting: SystemSetting) -> SystemSetting:
+        model = (
+            self.session.query(SystemSettingModel).filter(SystemSettingModel.setting_key == setting.setting_key).first()
+        )
+        if model is None:
+            model = SystemSettingModel(id=setting.id)
+            self.session.add(model)
+        model.setting_key = setting.setting_key
+        model.setting_value = setting.setting_value
+        model.setting_type = setting.setting_type
+        model.description = setting.description
+        model.is_public = setting.is_public
+        model.category = setting.category
+        model.updated_by = setting.updated_by
+        self.session.commit()
+        return setting
+
+    def delete_by_key(self, setting_key: str) -> bool:
+        model = self.session.query(SystemSettingModel).filter(SystemSettingModel.setting_key == setting_key).first()
+        if model is None:
+            return False
+        self.session.delete(model)
+        self.session.commit()
+        return True
+
+    @staticmethod
+    def _to_domain(model: SystemSettingModel) -> SystemSetting:
+        return SystemSetting(
+            id=model.id,
+            setting_key=model.setting_key,
+            setting_value=model.setting_value,
+            setting_type=model.setting_type,
+            description=model.description or "",
+            is_public=bool(model.is_public),
+            category=model.category or "general",
+            updated_by=model.updated_by,
+            created_at=model.created_at,
+            updated_at=model.updated_at,
+        )

packages/adapters/sqlalchemy_impl/viral_video_repository.py

@@ -49,6 +49,7 @@ def _to_domain(model: ViralVideoJobModel) -> ViralVideoJob:
         style_template_id=getattr(model, "style_template_id", "") or "",
         voice_id=getattr(model, "voice_id", "") or "",
         voice_source=getattr(model, "voice_source", "") or "",
+        language=getattr(model, "language", "zh-CN") or "zh-CN",
         video_ratio=getattr(model, "video_ratio", "9:16") or "9:16",
         video_model=getattr(model, "video_model", "") or "",
         status=ViralVideoStatus(model.status) if model.status else ViralVideoStatus.PENDING,
@@ -102,6 +103,7 @@ class SQLAlchemyViralVideoJobRepository:
             style_template_id=job.style_template_id,
             voice_id=job.voice_id,
             voice_source=job.voice_source,
+            language=getattr(job, "language", "zh-CN") or "zh-CN",
             video_ratio=job.video_ratio,
             video_model=job.video_model,
             status=job.status,
@@ -169,6 +171,7 @@ class SQLAlchemyViralVideoJobRepository:
         model.style_template_id = job.style_template_id
         model.voice_id = job.voice_id or ""
         model.voice_source = job.voice_source or ""
+        model.language = getattr(job, "language", "zh-CN") or "zh-CN"
         model.video_ratio = job.video_ratio or "9:16"
         model.video_model = job.video_model or ""
         model.updated_at = datetime.now(timezone.utc)

packages/application/auth/password_hasher.py

@@ -3,10 +3,24 @@
 使用 bcrypt 安全存储密码
 """
 
+import hashlib
 from typing import Optional
 
 import bcrypt
 
+# bcrypt 只对前 72 字节有效，且 bcrypt>=4.1 会对超长输入直接抛 ValueError。
+# 超长密码先做一次 SHA-256（定长 hex），再交给 bcrypt，
+# 既绕过长度限制又保持对超长不同密码的区分度。
+_BCRYPT_MAX_BYTES = 72
+
+
+def _prepare_password_bytes(password: str) -> bytes:
+    raw = password.encode("utf-8")
+    if len(raw) > _BCRYPT_MAX_BYTES:
+        return hashlib.sha256(raw).hexdigest().encode("utf-8")
+    return raw
+
+
 from packages.domain.auth.password_hasher import PasswordHasherPort, PasswordValidatorPort
 
 
@@ -42,8 +56,8 @@ class PasswordHasher(PasswordHasherPort):
         if not password:
             raise ValueError("Password cannot be empty")
 
-        # bcrypt 需要 bytes
-        password_bytes = password.encode("utf-8")
+        # bcrypt 需要 bytes（超长密码先 SHA-256 以兼容 72 字节限制）
+        password_bytes = _prepare_password_bytes(password)
 
         # 生成 salt 并哈希
         salt = bcrypt.gensalt(rounds=self.rounds)
@@ -67,7 +81,7 @@ class PasswordHasher(PasswordHasherPort):
             return False
 
         try:
-            password_bytes = password.encode("utf-8")
+            password_bytes = _prepare_password_bytes(password)
             hashed_bytes = hashed_password.encode("utf-8")
 
             return bcrypt.checkpw(password_bytes, hashed_bytes)

packages/application/ditto_emotion_service.py

@@ -0,0 +1,401 @@
+"""Ditto LLM 情绪分析服务 — #2076 后续：根据文案生成 emo_timeline.
+
+职责：
+1. 分句：先按 。！？； 句末标点切分；整段无句末标点且超长时按逗号等语义停顿二次切分
+2. 调 DoubaoClient.chat_completion 分析每句表情（emo: 0-7 八种全部开放, intensity: 0-1）
+3. 结果 LRU 缓存（文案 hash → 情绪列表）
+4. LLM 失败/超时/格式错 → 返回空列表（降级中性表情，不阻塞生成）
+5. TTS 完成后按字数比例或 sentence_timings 对齐成秒级 timeline
+"""
+
+from __future__ import annotations
+
+import hashlib
+import json
+import logging
+import re
+from functools import lru_cache
+from pathlib import Path
+from typing import Any, Optional
+
+logger = logging.getLogger(__name__)
+
+# ── 表情常量 ─────────────────────────────────────────────────────
+EMO_ANGER = 0
+EMO_DISGUST = 1
+EMO_FEAR = 2
+EMO_HAPPY = 3
+EMO_NEUTRAL = 4
+EMO_SAD = 5
+EMO_SURPRISE = 6
+EMO_CONTEMPT = 7
+ALLOWED_EMOS = set(range(8))  # 8 种表情全部开放（0-7）；超出范围的非法值仍会被降级为中性
+
+# ── 分句正则 ─────────────────────────────────────────────────────
+_SENT_SPLIT_RE = re.compile(r"(?<=[。！？；!?;])\s*")
+# 二次切分：逗号/分号/顿号/冒号等语义停顿（lookbehind 保留标点在前句末尾）
+_CLAUSE_SPLIT_RE = re.compile(r"(?<=[，,、：:])")
+# 整段无句末标点且超过该长度时才做逗号二次切分
+_SINGLE_SENTENCE_THRESHOLD = 30
+# 二次切分后过短片段（<5 字，如"家人们，"）并入相邻句避免切太碎；
+# 4 字以上语义完整的短句（如"心中有爱，"）保留独立成段
+_MIN_CLAUSE_LEN = 5
+
+# ── 默认 prompt 模板文件路径 ──────────────────────────────────────
+_DEFAULT_PROMPT_PATH = Path(__file__).parent / "prompts" / "ditto_emotion.txt"
+
+
+def _load_default_prompt() -> str:
+    try:
+        return _DEFAULT_PROMPT_PATH.read_text(encoding="utf-8").strip()
+    except Exception:
+        # 文件不存在时用极简兜底
+        return (
+            "分析文案每句话表情，输出JSON数组："
+            '[{"text":"句子","emo":4,"intensity":0.2}]，emo:0-7八种表情全部开放，'
+            "至少30%句子用非中性表情。\n【文案】\n{文案}"
+        )
+
+
+# ── 数据结构 ─────────────────────────────────────────────────────
+class EmotionSegment:
+    """单句情绪结果（LLM 输出的原始结构）."""
+
+    __slots__ = ("text", "emo", "intensity")
+
+    def __init__(self, text: str, emo: int, intensity: float):
+        self.text = text
+        self.emo = emo
+        self.intensity = intensity
+
+    def to_dict(self) -> dict[str, Any]:
+        return {"text": self.text, "emo": self.emo, "intensity": self.intensity}
+
+
+class EmotionTimelineEntry:
+    """对齐到音频时间轴后的情绪片段（传给 Ditto）."""
+
+    __slots__ = ("start", "end", "emo", "intensity")
+
+    def __init__(self, start: float, end: float, emo: int, intensity: float):
+        self.start = round(start, 2)
+        self.end = round(end, 2)
+        self.emo = emo
+        self.intensity = round(intensity, 2)
+
+    def to_dict(self) -> dict[str, Any]:
+        return {
+            "start": self.start,
+            "end": self.end,
+            "emo": self.emo,
+            "intensity": self.intensity,
+        }
+
+
+# ── 分句 ─────────────────────────────────────────────────────────
+def _split_by_clauses(sentence: str) -> list[str]:
+    """按逗号/分号等语义停顿二次切分；过短片段（<5 字）并入相邻句，避免切太碎.
+
+    每句尽量控制在 10-50 字左右（50 字为软上限，无更细停顿点时保持语义完整不硬切）。
+    """
+    raw = [c for c in _CLAUSE_SPLIT_RE.split(sentence) if c and c.strip()]
+    if len(raw) <= 1:
+        return [sentence]
+    merged: list[str] = []
+    buf = ""
+    for clause in raw:
+        if not buf:
+            buf = clause
+            continue
+        if len(buf) < _MIN_CLAUSE_LEN:
+            buf += clause
+            continue
+        merged.append(buf)
+        buf = clause
+    if buf:
+        if merged and len(buf) < _MIN_CLAUSE_LEN:
+            merged[-1] += buf
+        else:
+            merged.append(buf)
+    return [m.strip() for m in merged if m.strip()]
+
+
+def split_sentences(text: str) -> list[str]:
+    """按中文句末标点切分，过滤空串.
+
+    Bug1 修复：若按句末标点切分后只有 1 段且文本超过 30 字（文案全是逗号、
+    结尾无标点），再按逗号/分号等语义停顿做二次切分，避免整段只返回 1 个
+    segment 导致 LLM 只能对整段打一个表情。
+    """
+    if not text:
+        return []
+    parts = [p.strip() for p in _SENT_SPLIT_RE.split(text.strip()) if p and p.strip()]
+    if len(parts) <= 1 and parts and len(parts[0]) > _SINGLE_SENTENCE_THRESHOLD:
+        parts = _split_by_clauses(parts[0])
+    return parts
+
+
+# ── 解析 LLM 返回的 JSON ─────────────────────────────────────────
+def _parse_emotion_json(raw: str) -> list[EmotionSegment]:
+    """解析 LLM 返回，容错处理：
+    - 去掉 markdown 代码块包裹
+    - 只取第一个 JSON 数组
+    - 逐行校验 emo/intensity 合法性，过滤无效项
+    """
+    if not raw:
+        return []
+    text = raw.strip()
+    # 去掉 ```json ... ``` 包裹
+    if text.startswith("```"):
+        text = re.sub(r"^```(?:json)?\s*", "", text)
+        text = re.sub(r"\s*```$", "", text)
+    # 找第一个 [ 到最后一个 ]
+    lb = text.find("[")
+    rb = text.rfind("]")
+    if lb == -1 or rb == -1 or rb <= lb:
+        return []
+    try:
+        data = json.loads(text[lb : rb + 1])
+    except (json.JSONDecodeError, ValueError):
+        return []
+    if not isinstance(data, list):
+        return []
+
+    results: list[EmotionSegment] = []
+    for item in data:
+        if not isinstance(item, dict):
+            continue
+        try:
+            emo = int(item.get("emo", EMO_NEUTRAL))
+            intensity = float(item.get("intensity", 0.2))
+        except (TypeError, ValueError):
+            continue
+        if emo not in ALLOWED_EMOS:
+            emo = EMO_NEUTRAL
+        intensity = max(0.05, min(1.0, intensity))
+        sent_text = str(item.get("text", "")).strip()
+        if not sent_text:
+            continue
+        results.append(EmotionSegment(text=sent_text, emo=emo, intensity=intensity))
+    return results
+
+
+# ── 时间对齐（按字数比例）────────────────────────────────────────
+def align_timeline_by_length(
+    segments: list[EmotionSegment],
+    audio_duration: float,
+) -> list[EmotionTimelineEntry]:
+    """按各句字数占总字数比例分配 audio_duration 时长."""
+    if not segments or audio_duration <= 0:
+        return []
+    total_chars = sum(len(s.text) for s in segments)
+    if total_chars <= 0:
+        return []
+    entries: list[EmotionTimelineEntry] = []
+    pos = 0.0
+    for i, seg in enumerate(segments):
+        if i == len(segments) - 1:
+            end = audio_duration  # 最后一段到结尾，避免浮点误差
+        else:
+            end = pos + (len(seg.text) / total_chars) * audio_duration
+        if end > pos:
+            entries.append(
+                EmotionTimelineEntry(
+                    start=pos,
+                    end=end,
+                    emo=seg.emo,
+                    intensity=seg.intensity,
+                )
+            )
+        pos = end
+    return entries
+
+
+def align_timeline_by_timings(
+    segments: list[EmotionSegment],
+    sentence_timings: list[dict[str, Any]],
+    audio_duration: float,
+) -> list[EmotionTimelineEntry]:
+    """使用 TTS sentence_timings 精确对齐（优先方案）.
+
+    sentence_timings 格式：[{"start":0.0,"end":1.2,"text":"句子"}, ...]
+    按句序匹配 segments 和 timings，长度不一致时回退到按字数比例。
+    """
+    if not sentence_timings or len(sentence_timings) != len(segments):
+        return align_timeline_by_length(segments, audio_duration)
+    entries: list[EmotionTimelineEntry] = []
+    for seg, timing in zip(segments, sentence_timings, strict=False):
+        try:
+            start = float(timing.get("start", 0))
+            end = float(timing.get("end", 0))
+        except (TypeError, ValueError):
+            return align_timeline_by_length(segments, audio_duration)
+        if end <= start:
+            continue
+        entries.append(
+            EmotionTimelineEntry(
+                start=start,
+                end=end,
+                emo=seg.emo,
+                intensity=seg.intensity,
+            )
+        )
+    return entries
+
+
+# ── LLM 结果合理性校验（提示词遵循度排查）─────────────────────────
+# 感叹号或明显情绪关键词出现时，LLM 不应返回全中性/低强度
+_EMO_SIGNAL_RE = re.compile(r"[！!]|是不是|有没有|居然|竟然|只要|太坑|千万别|免费|赶紧|推荐|惊艳")
+
+
+def _warn_if_suspicious_all_neutral(text: str, segments: list[EmotionSegment]) -> None:
+    """文案含明显情绪信号但 LLM 返回全部 emo=4 且 intensity<=0.1 时打 warning 日志."""
+    if not segments:
+        return
+    all_flat = all(s.emo == EMO_NEUTRAL and s.intensity <= 0.1 for s in segments)
+    if all_flat and _EMO_SIGNAL_RE.search(text):
+        logger.warning(
+            "[ditto_emotion] 文案含情绪信号(感叹号/关键词)但 LLM 返回全中性/低强度，"
+            "疑似提示词遵循度差: text=%.80s, segments=%d",
+            text,
+            len(segments),
+        )
+
+
+# ── LLM 情绪分析服务 ─────────────────────────────────────────────
+class DittoEmotionService:
+    """Ditto 情绪分析服务（带 LRU 缓存）."""
+
+    def __init__(self, settings=None):
+        from packages.config import get_api_settings
+
+        self.settings = settings or get_api_settings()
+        self._client = None
+        self._client_api_key: str | None = None
+
+    def _cfg(self, key: str) -> Any:
+        """优先读后台 system_config，未配置则回退到 settings（env 默认）."""
+        try:
+            from packages.application.system_config_service import get_config
+
+            return get_config(key, getattr(self.settings, key, None))
+        except Exception:
+            return getattr(self.settings, key, None)
+
+    @property
+    def enabled(self) -> bool:
+        return bool(self._cfg("ditto_emotion_enabled"))
+
+    def _get_prompt_template(self) -> str:
+        """优先用配置（环境变量），否则读文件."""
+        cfg_prompt = self._cfg("ditto_emotion_prompt") or ""
+        if cfg_prompt.strip():
+            return cfg_prompt.strip()
+        return _load_default_prompt()
+
+    def _cache_key(self, text: str) -> str:
+        return hashlib.md5(text.strip().encode("utf-8")).hexdigest()
+
+    def _get_llm_client(self):
+        """创建/复用 DoubaoClient；API Key 优先读后台配置（DB），fallback 环境变量。
+
+        后台可能随时改 doubao_api_key，每次调用前对比 key，变化则重建 client。
+        """
+        api_key = (self._cfg("doubao_api_key") or "").strip()
+        if self._client is None or self._client_api_key != api_key:
+            from packages.shared.ai_client import DoubaoClient
+
+            self._client = DoubaoClient(api_key=api_key)
+            self._client_api_key = api_key
+        return self._client
+
+    def _call_llm(self, text: str) -> list[EmotionSegment]:
+        """调 LLM 分析情绪，失败返回空列表."""
+        template = self._get_prompt_template()
+        prompt = template.replace("{文案}", text)
+        messages = [{"role": "user", "content": prompt}]
+        model = self._cfg("ditto_emotion_model") or None
+        temperature = self._cfg("ditto_emotion_temperature")
+        timeout = getattr(self.settings, "ditto_emotion_timeout", 10)
+        max_tokens = getattr(self.settings, "ditto_emotion_max_tokens", 1024)
+        try:
+            client = self._get_llm_client()
+            result = client.chat_completion(
+                messages=messages,
+                model=model,
+                temperature=temperature,
+                max_tokens=max_tokens,
+                timeout=timeout,
+            )
+        except Exception as exc:
+            logger.warning("[ditto_emotion] LLM 调用异常: %s", exc)
+            return []
+        if not result:
+            return []
+        segments = _parse_emotion_json(result)
+        if not segments:
+            logger.warning("[ditto_emotion] LLM 返回解析失败: %s", result[:200])
+        else:
+            _warn_if_suspicious_all_neutral(text, segments)
+        return segments
+
+    def analyze(self, text: str) -> list[EmotionSegment]:
+        """分析文案情绪（带缓存），失败返回空列表."""
+        if not self.enabled or not text or not text.strip():
+            return []
+        key = self._cache_key(text)
+        return _cached_analyze(self, key, text)
+
+    def build_timeline(
+        self,
+        text: str,
+        audio_duration: float,
+        sentence_timings: Optional[list[dict[str, Any]]] = None,
+    ) -> str:
+        """完整流程：分句→LLM分析→时间对齐→序列化为JSON字符串.
+
+        返回: JSON 字符串（可直接传 Ditto emo_timeline 参数）；空字符串表示降级中性。
+        """
+        segments = self.analyze(text)
+        if not segments:
+            return ""
+        if sentence_timings:
+            entries = align_timeline_by_timings(segments, sentence_timings, audio_duration)
+        else:
+            entries = align_timeline_by_length(segments, audio_duration)
+        if not entries:
+            return ""
+        return json.dumps([e.to_dict() for e in entries], ensure_ascii=False)
+
+
+# ── 模块级 LRU 缓存实例 ─────────────────────────────────────────
+# 每个 service 实例共享缓存（按 cache_key 区分）
+@lru_cache(maxsize=512)
+def _cached_analyze(service: DittoEmotionService, cache_key: str, text: str) -> list[EmotionSegment]:
+    """LRU 缓存包装：cache_key 由文案 hash 生成，maxsize 从配置读."""
+    # 注意：service 参数仅用于传递调用，缓存由 cache_key 驱动
+    segments = service._call_llm(text)
+    # 如果 LLM 返回空（比如分句数量不匹配），尝试直接对预分句结果分析
+    if not segments:
+        pre_splits = split_sentences(text)
+        if len(pre_splits) >= 1:
+            # 用预分句结果兜底：全中性低强度
+            segments = [EmotionSegment(text=s, emo=EMO_NEUTRAL, intensity=0.1) for s in pre_splits]
+    return segments
+
+
+_singleton: Optional[DittoEmotionService] = None
+
+
+def get_ditto_emotion_service() -> DittoEmotionService:
+    global _singleton
+    if _singleton is None:
+        _singleton = DittoEmotionService()
+    return _singleton
+
+
+def reset_ditto_emotion_service() -> None:
+    """#2246：后台配置变更后重置单例并清空 LLM 结果 LRU 缓存."""
+    global _singleton
+    _singleton = None
+    _cached_analyze.cache_clear()

packages/application/ditto_service.py

@@ -0,0 +1,308 @@
+"""蚂蚁 Ditto 数字人口型 API 客户端 — #2076.
+
+封装 Ditto FastAPI（部署在 5060Ti GPU 节点，Tailscale 内网可达）：
+- GET  /health               健康检查
+- POST /generate             生成口型视频（同步返回 MP4 流）
+
+关键特性：
+- 入参：video_url（人物驱动视频 URL，用户上传优先；未传时用 default_video_url 兜底）
+        + audio_url（TTS 音频 URL） + script（文案原文）
+        + emo_timeline（可选，LLM 情绪时间线 JSON 字符串）
+- 出参：直接返回 video/mp4 字节流（自带音频，无需二次混流）
+- 429 时指数退避重试（最多 ditto_max_retries 次）
+- 500/超时/网络不可达视为失败
+- 输出 MP4 字节流转存到自家 OSS，返回公网 URL
+
+注意：
+- 保留 MuseTalk/GPU 路径不变；本服务作为更高优先级的第三条口型路径
+- video_url 时长对齐（视频<音频时循环延长）在 Celery 任务层用 ffmpeg 预处理
+- Ditto 输出自带音视频，不需要 GFPGAN 超分，不需要 ffmpeg 音视频混流
+"""
+
+from __future__ import annotations
+
+import io
+import logging
+import time
+from dataclasses import dataclass
+from typing import Optional
+
+import httpx
+
+from packages.config import get_api_settings
+
+logger = logging.getLogger(__name__)
+
+
+class DittoError(Exception):
+    """Ditto API 调用失败."""
+
+    def __init__(self, message: str, code: str = "DittoError", status_code: int = 0):
+        self.code = code
+        self.status_code = status_code
+        super().__init__(message)
+
+
+@dataclass
+class DittoResult:
+    """Ditto 生成结果."""
+
+    video_bytes: bytes
+    video_url: str = ""  # 转存 OSS 后填充
+    elapsed_seconds: float = 0.0
+    rtf: float = 0.0  # 实时率（响应头 X-RTF）
+    frames: int = 0  # 帧数（响应头 X-Frames）
+
+
+class DittoClient:
+    """蚂蚁 Ditto 数字人口型 API 客户端."""
+
+    def __init__(
+        self,
+        base_url: Optional[str] = None,
+        default_video_url: Optional[str] = None,
+        max_retries: Optional[int] = None,
+        timeout: Optional[int] = None,
+    ):
+        s = get_api_settings()
+        self.base_url = (base_url or s.ditto_api_base_url or "").rstrip("/")
+        self.default_video_url = default_video_url or s.ditto_default_video_url or ""
+        self.max_retries = int(max_retries if max_retries is not None else s.ditto_max_retries)
+        self.timeout = int(timeout if timeout is not None else s.ditto_request_timeout)
+        try:
+            from packages.application.system_config_service import get_config
+
+            self.blend_frames = int(get_config("ditto_blend_frames", s.ditto_blend_frames))
+        except Exception:
+            self.blend_frames = int(s.ditto_blend_frames)
+
+    @property
+    def is_configured(self) -> bool:
+        """配置是否完整（base_url 必填 + 默认模板视频兜底 URL 有值）。
+
+        注意：即使 is_configured=True，实际生成时优先使用用户上传的 video_url；
+        default_video_url 仅作为用户未上传视频时的兜底。
+        """
+        return bool(self.base_url) and bool(self.default_video_url)
+
+    def health(self) -> bool:
+        """健康检查；成功返回 True，失败返回 False（不抛异常）."""
+        if not self.base_url:
+            return False
+        url = f"{self.base_url}/health"
+        try:
+            with httpx.Client(timeout=5.0) as client:
+                resp = client.get(url)
+                ok = resp.status_code == 200
+                if ok:
+                    logger.info("[ditto] health check OK: %s", url)
+                else:
+                    logger.warning("[ditto] health check status=%d: %s", resp.status_code, url)
+                return ok
+        except Exception as exc:
+            logger.warning("[ditto] health check failed: %s", exc)
+            return False
+
+    def generate(
+        self,
+        *,
+        audio_url: str,
+        script: str,
+        video_url: Optional[str] = None,
+        emo_global: int = 4,
+        use_script_emo: bool = True,
+        blend_frames: Optional[int] = None,
+        emo_timeline: str = "",
+    ) -> DittoResult:
+        """调用 Ditto /generate 接口，返回 MP4 字节流结果.
+
+        Raises DittoError on failure.
+        """
+        if not self.base_url:
+            raise DittoError("DITTO_API_BASE_URL 未配置", code="ConfigMissing")
+        driver_url = video_url or self.default_video_url
+        if not driver_url:
+            raise DittoError("Ditto 人物模板视频 URL 未配置", code="ConfigMissing")
+        if not audio_url:
+            raise DittoError("audio_url 不能为空", code="InvalidParam")
+        if not script:
+            script = " "
+
+        _blend = blend_frames if blend_frames is not None else self.blend_frames
+        payload = {
+            "video_url": driver_url,
+            "audio_url": audio_url,
+            "script": script,
+            "emo_global": emo_global,
+            "use_script_emo": use_script_emo,
+            "blend_frames": _blend,
+        }
+        if emo_timeline:
+            payload["emo_timeline"] = emo_timeline
+        url = f"{self.base_url}/generate"
+
+        last_exc: Optional[Exception] = None
+        for attempt in range(self.max_retries + 1):
+            try:
+                start = time.monotonic()
+                # 精细化超时：connect=10s（网络不通快速失败），read=120s（最长音频~45s按RTF=2.8推算）
+                _timeout = httpx.Timeout(connect=10.0, read=self.timeout, write=30.0, pool=10.0)
+                with httpx.Client(timeout=_timeout, follow_redirects=True) as client:
+                    resp = client.post(url, json=payload)
+                elapsed = time.monotonic() - start
+
+                if resp.status_code == 429:
+                    wait = min(2**attempt, 30)
+                    logger.warning(
+                        "[ditto] GPU 繁忙 (429)，%ds 后重试 (%d/%d)",
+                        wait,
+                        attempt + 1,
+                        self.max_retries,
+                    )
+                    if attempt >= self.max_retries:
+                        raise DittoError(
+                            f"Ditto GPU 繁忙，重试 {self.max_retries} 次仍失败",
+                            code="BusyRetriesExhausted",
+                            status_code=429,
+                        )
+                    time.sleep(wait)
+                    continue
+
+                if resp.status_code != 200:
+                    _text = (resp.text or "")[:300]
+                    logger.error(
+                        "[ditto] generate 失败 status=%d attempt=%d body=%s",
+                        resp.status_code,
+                        attempt + 1,
+                        _text,
+                    )
+                    if resp.status_code >= 500 and attempt < self.max_retries:
+                        time.sleep(min(2**attempt, 15))
+                        continue
+                    raise DittoError(
+                        f"Ditto 返回 {resp.status_code}: {_text}",
+                        code="DittoAPIError",
+                        status_code=resp.status_code,
+                    )
+
+                video_bytes = resp.content
+                if not video_bytes or len(video_bytes) < 1024:
+                    raise DittoError(
+                        f"Ditto 返回内容异常（size={len(video_bytes) if video_bytes else 0}）",
+                        code="EmptyResponse",
+                    )
+                try:
+                    rtf = float(resp.headers.get("X-RTF", "0") or 0)
+                except ValueError:
+                    rtf = 0.0
+                try:
+                    frames = int(resp.headers.get("X-Frames", "0") or 0)
+                except ValueError:
+                    frames = 0
+                try:
+                    x_time = float(resp.headers.get("X-Time", "0") or 0)
+                    if x_time > 0:
+                        elapsed = x_time
+                except ValueError:
+                    pass
+
+                logger.info(
+                    "[ditto] generate 成功 size=%d rtf=%.2f frames=%d elapsed=%.1fs attempt=%d",
+                    len(video_bytes),
+                    rtf,
+                    frames,
+                    elapsed,
+                    attempt + 1,
+                )
+                return DittoResult(
+                    video_bytes=video_bytes,
+                    elapsed_seconds=elapsed,
+                    rtf=rtf,
+                    frames=frames,
+                )
+
+            except DittoError:
+                raise
+            except (httpx.ConnectError, httpx.NetworkError, ConnectionError, OSError) as exc:
+                # 网络不通/连接被拒（如 GPU 断网/Tailscale 掉线），不重试，直接快速回退
+                logger.warning("[ditto] 网络不可达 attempt=%d err=%s", attempt + 1, exc)
+                raise DittoError(
+                    f"Ditto 网络不可达: {exc}",
+                    code="NetworkUnreachable",
+                ) from exc
+            except httpx.TimeoutException as exc:
+                last_exc = exc
+                logger.warning("[ditto] 请求超时 attempt=%d err=%s", attempt + 1, exc)
+                if attempt < self.max_retries:
+                    time.sleep(min(2**attempt, 15))
+                    continue
+                raise DittoError(
+                    f"Ditto 请求超时（read={self.timeout}s），重试耗尽",
+                    code="Timeout",
+                ) from exc
+            except Exception as exc:
+                last_exc = exc
+                logger.warning("[ditto] 请求异常 attempt=%d err=%s", attempt + 1, exc)
+                if attempt < self.max_retries:
+                    time.sleep(min(2**attempt, 10))
+                    continue
+                raise DittoError(f"Ditto 调用异常: {exc}", code="NetworkError") from exc
+
+        raise DittoError("Ditto 未知错误", code="Unknown") from last_exc
+
+    def generate_and_persist(
+        self,
+        *,
+        job_id: str,
+        user_id: str,
+        audio_url: str,
+        script: str,
+        video_url: Optional[str] = None,
+        emo_timeline: str = "",
+        blend_frames: Optional[int] = None,
+    ) -> DittoResult:
+        """调用 generate 并把 MP4 转存到自家 OSS，返回带 video_url 的结果."""
+        result = self.generate(
+            audio_url=audio_url,
+            script=script,
+            video_url=video_url,
+            emo_timeline=emo_timeline,
+            blend_frames=blend_frames,
+        )
+        try:
+            from packages.shared.storage import get_shared_storage_service
+
+            storage = get_shared_storage_service()
+            storage_key = f"ditto-output/{user_id}/{job_id}.mp4"
+            public_url = storage.upload_file(
+                io.BytesIO(result.video_bytes),
+                storage_key,
+                content_type="video/mp4",
+            )
+            result.video_url = public_url
+            logger.info(
+                "[ditto] 转存 OSS 完成 job=%s key=%s",
+                job_id,
+                storage_key,
+            )
+        except Exception as exc:
+            logger.error("[ditto] 转存 OSS 失败 job=%s err=%s", job_id, exc, exc_info=True)
+            raise DittoError(f"Ditto 结果转存 OSS 失败: {exc}", code="StorageError") from exc
+        return result
+
+
+_ditto_client_singleton: Optional[DittoClient] = None
+
+
+def get_ditto_client() -> DittoClient:
+    """获取 DittoClient 单例（简易工厂，便于单测 mock）."""
+    global _ditto_client_singleton
+    if _ditto_client_singleton is None:
+        _ditto_client_singleton = DittoClient()
+    return _ditto_client_singleton
+
+
+def reset_ditto_client() -> None:
+    """#2246：后台配置变更后重置 DittoClient 单例."""
+    global _ditto_client_singleton
+    _ditto_client_singleton = None

packages/application/prompts/ditto_emotion.txt

@@ -0,0 +1,42 @@
+你是一个数字人短视频表情导演。为每句话分配一个最贴切的表情，让数字人表情丰富、有情绪感染力，像真人说话一样有起伏变化。
+
+【8种表情】
+0=愤怒：怒斥、骂坑、强烈不满、控诉
+1=厌恶：嫌弃、吐槽劣质、鄙视套路
+2=害怕：紧张、警告、危险提示、悬念
+3=开心：夸赞、推荐、好消息、福利、号召
+4=中性：陈述事实、客观描述、数据、过渡衔接
+5=伤心：共情痛点、遗憾、难过、踩坑经历
+6=惊讶：惊喜、超预期、居然、竟然、只要、数字冲击
+7=轻蔑：嘲讽、冷笑、拆穿套路、不屑
+
+【分句说明】
+- 输入文案已按句末标点预分句；若某句内部仍有明显语义停顿（逗号、分号），你应根据语义进一步拆分成更细的句子分别打分
+- 若输入没有句末标点，要自己根据语义在逗号处分句
+
+【强度】0.1几乎无变化 0.3轻微自然 0.5明显情绪 0.7强烈 0.9爆发
+感叹号结尾0.4-0.7，句号0.2-0.4，问号0.3-0.6。
+
+【示例1】
+文案："家人们！你们是不是经常买东西踩坑？花了大价钱买回来的东西根本没法用！但是今天这款产品真的惊艳到我了！"
+输出：
+[{"text":"家人们！","emo":6,"intensity":0.5},{"text":"你们是不是经常买东西踩坑？","emo":5,"intensity":0.3},{"text":"花了大价钱买回来的东西根本没法用！","emo":0,"intensity":0.5},{"text":"但是今天这款产品真的惊艳到我了！","emo":3,"intensity":0.6}]
+
+【示例2】
+文案："做公益不一定是一个人去做惊天动地的大事，而是一群人扎扎实实的干好身边的小事，心中有爱，眼里有光。"
+输出：
+[{"text":"做公益不一定是一个人去做惊天动地的大事，","emo":4,"intensity":0.2},{"text":"而是一群人扎扎实实的干好身边的小事，","emo":4,"intensity":0.2},{"text":"心中有爱，","emo":3,"intensity":0.4},{"text":"眼里有光。","emo":3,"intensity":0.5}]
+
+【关键要求】
+- 每句话都必须独立判断情绪，至少有30%的句子使用非中性(4)表情！感叹号结尾的句子intensity不得低于0.4！
+- 严禁全部句子都返回中性(4)！严禁全部强度都是0.1！
+- 情绪要有变化：痛点段伤心或愤怒→转折→卖点开心→福利惊讶→号召开心
+- 语义优先：根据每句话的实际语义选择表情，不要固定开头结尾情绪
+- 8种表情全部开放，该愤怒就愤怒(0)，该厌恶就厌恶(1)，该害怕就害怕(2)，该轻蔑就轻蔑(7)
+- 开头第一句就根据内容定情绪，不要强行用中性开场
+
+【输出】严格JSON数组，只输出JSON，不要其他文字。
+[{"text":"原文","emo":3,"intensity":0.5}]
+
+【文案】
+{文案}

packages/application/system_config_service.py

@@ -0,0 +1,189 @@
+"""系统配置应用服务 — #2246.
+
+- get_config(key, default) / set_config(...) / list_configs(category)
+- 首次访问时从 DB 加载并进程内缓存；set_config 后失效缓存
+- set_config 成功后重置 Ditto 情绪服务与 Ditto 客户端单例（含 LRU 缓存），
+  保证后台改动立即生效；worker 不直接 HTTP 读配置，统一通过本模块。
+"""
+
+from __future__ import annotations
+
+import logging
+import threading
+from typing import Any
+
+from packages.adapters.sqlalchemy_impl import session as db_session
+from packages.adapters.sqlalchemy_impl.system_setting_repository import (
+    SQLAlchemySystemSettingRepository,
+)
+from packages.domain.system_setting import (
+    SystemSetting,
+    infer_setting_type,
+    serialize_setting_value,
+)
+
+logger = logging.getLogger(__name__)
+
+
+class SystemConfigService:
+    """进程内缓存的系统配置服务."""
+
+    def __init__(self, session_factory=None):
+        self._session_factory = session_factory
+        self._cache: dict[str, Any] | None = None
+        self._lock = threading.Lock()
+
+    # ── 会话 ─────────────────────────────────────────────────────
+    def _get_session_factory(self):
+        # 必须运行时读取模块属性：模块导入时 SessionLocal 还是 None，
+        # initialize_database() 之后才被赋值，import 时绑定会拿到旧值。
+        factory = self._session_factory or db_session.SessionLocal
+        if factory is None:
+            raise RuntimeError("数据库会话工厂未初始化")
+        return factory
+
+    # ── 缓存 ─────────────────────────────────────────────────────
+    def _load_cache(self) -> dict[str, Any]:
+        with self._lock:
+            if self._cache is not None:
+                return self._cache
+            cache: dict[str, Any] = {}
+            factory = self._get_session_factory()
+            session = factory()
+            try:
+                repo = SQLAlchemySystemSettingRepository(session)
+                for setting in repo.list_all():
+                    try:
+                        cache[setting.setting_key] = setting.get_typed_value()
+                    except Exception as exc:  # 损坏配置不阻塞启动
+                        logger.warning(
+                            "[system_config] 跳过损坏配置 %s: %s",
+                            setting.setting_key,
+                            exc,
+                        )
+            finally:
+                session.close()
+            self._cache = cache
+            return cache
+
+    def reload(self) -> dict[str, Any]:
+        """强制重新从 DB 加载配置，返回新缓存."""
+        with self._lock:
+            self._cache = None
+        return self._load_cache()
+
+    def invalidate(self) -> None:
+        with self._lock:
+            self._cache = None
+
+    # ── 读 ───────────────────────────────────────────────────────
+    def get_config(self, key: str, default: Any = None) -> Any:
+        cache = self._load_cache()
+        return cache.get(key, default)
+
+    def list_configs(self, category: str | None = None) -> list[SystemSetting]:
+        factory = self._get_session_factory()
+        session = factory()
+        try:
+            return SQLAlchemySystemSettingRepository(session).list_all(category)
+        finally:
+            session.close()
+
+    # ── 写 ───────────────────────────────────────────────────────
+    def set_config(
+        self,
+        key: str,
+        value: Any,
+        setting_type: str | None = None,
+        updated_by: str | None = None,
+        *,
+        description: str | None = None,
+        category: str = "general",
+        is_public: bool = False,
+    ) -> Any:
+        st = setting_type or infer_setting_type(value)
+        serialized = serialize_setting_value(value, st)
+
+        factory = self._get_session_factory()
+        session = factory()
+        try:
+            repo = SQLAlchemySystemSettingRepository(session)
+            existing = repo.get_by_key(key)
+            if existing is not None:
+                existing.setting_value = serialized
+                existing.setting_type = st
+                if updated_by is not None:
+                    existing.updated_by = updated_by
+                if description is not None:
+                    existing.description = description
+                setting = existing
+            else:
+                setting = SystemSetting(
+                    setting_key=key,
+                    setting_value=serialized,
+                    setting_type=st,
+                    description=description or "",
+                    is_public=is_public,
+                    category=category,
+                    updated_by=updated_by,
+                )
+            repo.upsert(setting)
+        finally:
+            session.close()
+
+        self.invalidate()
+        self._reset_ditto_singletons(key)
+        return value
+
+    def delete_config(self, key: str) -> bool:
+        factory = self._get_session_factory()
+        session = factory()
+        try:
+            deleted = SQLAlchemySystemSettingRepository(session).delete_by_key(key)
+        finally:
+            session.close()
+        if deleted:
+            self.invalidate()
+            self._reset_ditto_singletons(key)
+        return deleted
+
+    # ── Ditto 联动 ───────────────────────────────────────────────
+    @staticmethod
+    def _reset_ditto_singletons(key: str) -> None:
+        try:
+            from packages.application import ditto_emotion_service as emo_mod
+            from packages.application import ditto_service as ditto_mod
+
+            emo_mod.reset_ditto_emotion_service()
+            ditto_mod.reset_ditto_client()
+            logger.info("[system_config] %s 更新，已重置 Ditto 单例", key)
+        except Exception as exc:  # 联动失败不影响配置落库
+            logger.warning("[system_config] 重置 Ditto 单例失败: %s", exc)
+
+
+_service_singleton: SystemConfigService | None = None
+
+
+def get_system_config_service() -> SystemConfigService:
+    global _service_singleton
+    if _service_singleton is None:
+        _service_singleton = SystemConfigService()
+    return _service_singleton
+
+
+def get_config(key: str, default: Any = None) -> Any:
+    """便捷读取：优先 DB 配置，未配置时返回 default（调用方传 env 值兜底）."""
+    try:
+        return get_system_config_service().get_config(key, default)
+    except Exception as exc:
+        logger.warning("[system_config] 读取 %s 失败，使用默认值: %s", key, exc)
+        return default
+
+
+def set_config(
+    key: str,
+    value: Any,
+    setting_type: str | None = None,
+    updated_by: str | None = None,
+) -> Any:
+    return get_system_config_service().set_config(key, value, setting_type=setting_type, updated_by=updated_by)

packages/application/viral_video/constants.py

@@ -0,0 +1,86 @@
+# -*- coding: utf-8 -*-
+"""爆款视频时长-字数-镜头数统一规范。
+
+prompt 模板与代码校验共用同一份数据，避免不一致。"""
+
+from __future__ import annotations
+
+# 时长规格：key=duration(秒)，value=对应约束
+DURATION_SPEC: dict[int, dict[str, int]] = {
+    5: {"min_chars": 12, "max_chars": 20, "min_shots": 1, "max_shots": 2},
+    10: {"min_chars": 25, "max_chars": 40, "min_shots": 2, "max_shots": 3},
+    15: {"min_chars": 38, "max_chars": 60, "min_shots": 3, "max_shots": 4},
+    20: {"min_chars": 50, "max_chars": 80, "min_shots": 4, "max_shots": 5},
+    30: {"min_chars": 75, "max_chars": 120, "min_shots": 6, "max_shots": 8},
+}
+
+
+def get_duration_spec(dur: int) -> dict[str, int]:
+    """获取时长规格；clamp 到 [5,30] 并选取最近档位。
+
+    注意：dur=0/falsey 时默认取 15 秒档位（兜底）；dur 为正数时按实际值 clamp。
+    """
+    if dur is None:
+        dur = 15
+    try:
+        d = int(dur)
+    except (TypeError, ValueError):
+        d = 15
+    if d <= 0:
+        d = 15
+    clamped = max(5, min(30, d))
+    nearest = min(DURATION_SPEC.keys(), key=lambda k: abs(k - clamped))
+    return DURATION_SPEC[nearest]
+
+
+def build_duration_constraint_text(dur: int) -> str:
+    """生成注入到 LLM user prompt 的字数/镜头数约束文本。"""
+    spec = get_duration_spec(dur)
+    return (
+        f"视频时长{dur}秒。口播总字数严格控制在{spec['min_chars']}~{spec['max_chars']}字之间"
+        f"（中文正常口播语速约2.5~4字/秒，{dur}秒={dur * 4}字为硬上限，宁可少写也不要多写，"
+        f"超长会导致TTS配音超出视频时长而失败）。"
+        f"分镜镜头数{spec['min_shots']}~{spec['max_shots']}个，每个镜头的time_range首尾相接，"
+        f"不得跳秒或重叠，每个镜头时长≥2秒。"
+    )
+
+
+def get_expected_shot_count(dur: int) -> tuple[int, int]:
+    """返回 (min_shots, max_shots)，供代码校验使用。"""
+    spec = get_duration_spec(dur)
+    return spec["min_shots"], spec["max_shots"]
+
+
+# ── Seedance 生成默认约束（worker/API 共用，避免硬编码漂移） ──────────────
+DEFAULT_HARD_CONSTRAINTS: list[str] = [
+    "无字幕、无水印、无任何自动生成文字、无 logo",
+    "严格还原参考图片中的真实场景、门店环境、商品陈列、人物外貌服装特征，不得凭空生成与参考图无关的人物、场景或物品",
+    "同一人物全程保持一致的五官、发型、服装、身材，不得换脸或变形",
+    "口播语音必须在指定时长内自然念完，语速自然，口型与语音同步",
+    "画面流畅无闪烁、无多余肢体、无扭曲变形、无穿模",
+    "色彩自然、曝光正确、电影级质感、高清细节",
+]
+
+DEFAULT_NEGATIVE_PROMPTS: list[str] = [
+    "字幕",
+    "自动字幕",
+    "水印",
+    "logo",
+    "图标",
+    "错误文字",
+    "乱码文字",
+    "男女声错配",
+    "中途换声",
+    "五官崩坏",
+    "脸部变形",
+    "多余手指",
+    "肢体扭曲",
+    "闪烁",
+    "画面抖动",
+    "模糊",
+    "低分辨率",
+]
+
+# API _build_copy_result 兼容老数据的兜底（比 worker 的短，仅用于老字段拼装）
+API_FALLBACK_HARD_CONSTRAINTS: list[str] = ["无字幕", "无水印", "人物一致性"]
+API_FALLBACK_NEGATIVE_PROMPTS: list[str] = ["字幕", "水印", "错误文字", "五官崩坏"]

packages/application/viral_video/generator.py

@@ -1,427 +1,12 @@
-"""爆款视频 5 步编排：图片分析 → 意图解析 → 文案融合 → 分镜 → 审核重写。
+# -*- coding: utf-8 -*-
+"""爆款视频旧 5 步编排器（已废弃）。
 
-所有 LLM 调用走 DoubaoClient，单测通过 client 参数注入 mock，不真调 API。
-任何一步解析失败都走规则 fallback，不抛异常阻断。
+PR1 (refactor/viral-video-backend-p0-deadcode-and-consistency)：CopyGenerator 整体删除。
+- 图片分析走 worker_app/tasks/vision 下的 V2 fast_path（analyze_images_v2）；
+- 文案/分镜/审核走 worker_app/tasks/viral_video.py 的 _step_script_generation / _step_review；
+- 旧的 intent_parsing / copy_fusion 两步早已在叙述优先重构中删除，不再有调用点。
+
+本文件保留为空壳，避免误导入；任何新代码请不要从这里 import。
 """
 
 from __future__ import annotations
-
-import logging
-from typing import Optional
-
-from packages.application.viral_video import xml_parser as xp
-from packages.application.viral_video.prompt_loader import (
-    PromptTemplate,
-    get_template,
-    render_system_prompt,
-    render_user_prompt,
-)
-from packages.application.viral_video.prompts import (
-    FUSION_INSTRUCTIONS,
-    GLOBAL_CONSTRAINTS,
-    NEGATIVE_RULES,
-)
-from packages.application.viral_video.reviewer import Reviewer
-from packages.application.viral_video.schemas import (
-    BodyPoint,
-    Clip,
-    ColorItem,
-    CoreMessage,
-    FusionResult,
-    ImageAnalysis,
-    IntentResult,
-    KenBurns,
-    PersonalBrand,
-    ProductItem,
-    ReviewResult,
-    ScriptSegment,
-    Storyboard,
-    TextItem,
-)
-
-logger = logging.getLogger(__name__)
-
-
-class CopyGenerator:
-    """5 步 Prompt 编排器。"""
-
-    def __init__(self, client=None, reviewer: Optional[Reviewer] = None):
-        if client is None:
-            from packages.shared.ai_client import get_doubao_client
-
-            client = get_doubao_client()
-        self.client = client
-        self.reviewer = reviewer or Reviewer(client)
-
-    # ── 底层调用 ────────────────────────────────────────────────────────
-    def _chat(self, template: PromptTemplate, system_kwargs: dict | None, **user_kwargs) -> str:
-        system = render_system_prompt(template, **(system_kwargs or {}))
-        user = render_user_prompt(template, **user_kwargs)
-        result = self.client.chat_completion(
-            [
-                {"role": "system", "content": system},
-                {"role": "user", "content": user},
-            ],
-            temperature=0.7,
-            max_tokens=2048,
-        )
-        return result or ""
-
-    # ── 步骤1：图片多模态分析 ───────────────────────────────────────────
-    def analyze_images(self, images: list[str], industry: str = "") -> ImageAnalysis:
-        template = get_template("image_analysis")
-        image_urls = "\n".join(f"第{i + 1}张：{url}" for i, url in enumerate(images))
-        system = render_system_prompt(template)
-        user = render_user_prompt(template, image_count=len(images), industry=industry or "通用", image_urls=image_urls)
-        raw = self.client.vision_completion(
-            [
-                {"role": "system", "content": system},
-                {"role": "user", "content": user},
-            ],
-            images=images,
-            max_tokens=2048,
-            temperature=0.3,
-        )
-        analysis = self._parse_image_analysis(raw or "")
-        if not analysis.products and not analysis.key_selling_points:
-            logger.warning("图片分析标签解析失败，走规则 fallback")
-            return self._fallback_image_analysis(images, raw or "")
-        return analysis
-
-    def _parse_image_analysis(self, raw: str) -> ImageAnalysis:
-        products = [
-            ProductItem(
-                name=n["attrs"].get("name", "无法判断"),
-                features=n["attrs"].get("features", "无法判断"),
-                position=n["attrs"].get("position", "secondary"),
-                image_index=xp.attr_int(n["attrs"].get("image_index"), 0),
-            )
-            for n in xp.find_all(raw, "product")
-        ]
-        colors = [
-            ColorItem(
-                hex=c["attrs"].get("hex", "#000000"),
-                name=c["attrs"].get("name", "无法判断"),
-                coverage=xp.attr_float(c["attrs"].get("coverage"), 0.0),
-            )
-            for c in xp.find_all(raw, "color")
-        ]
-        people = xp.find_first(raw, "people")
-        visible_text = [
-            TextItem(text=t["attrs"].get("text", ""), position=t["attrs"].get("position", ""))
-            for t in xp.find_all(raw, "text_item")
-        ]
-        quality_node = xp.find_first(raw, "quality")
-        selling_points = [n["text"] or n["attrs"].get("text", "") for n in xp.find_all(raw, "point")]
-        return ImageAnalysis(
-            products=products,
-            colors=colors,
-            has_person=xp.attr_bool(people["attrs"].get("has_person")) if people else False,
-            person_count=xp.attr_int(people["attrs"].get("count"), 0) if people else 0,
-            people=people["attrs"] if people else {},
-            mood=xp.text_of(raw, "mood"),
-            visible_text=visible_text,
-            scene=xp.text_of(raw, "scene"),
-            quality=quality_node["attrs"] if quality_node else {},
-            key_selling_points=[p for p in selling_points if p],
-            raw=raw,
-        )
-
-    def _fallback_image_analysis(self, images: list[str], raw: str) -> ImageAnalysis:
-        return ImageAnalysis(
-            products=[ProductItem(name="无法判断（视觉分析不可用）", image_index=0)],
-            scene="无法判断",
-            raw=raw,
-        )
-
-    # ── 步骤2：意图解析 ─────────────────────────────────────────────────
-    def parse_intent(self, user_copy_text: str, image_analysis: ImageAnalysis, industry: str = "") -> IntentResult:
-        template = get_template("intent_parsing")
-        raw = self._chat(
-            template,
-            None,
-            user_copy_text=user_copy_text or "（用户没有提供文案）",
-            industry=industry or "通用",
-            image_analysis=self._image_brief(image_analysis),
-        )
-        intent = self._parse_intent(raw)
-        if not intent.intent_summary and not intent.core_messages:
-            logger.warning("意图解析标签解析失败，走规则 fallback")
-            return self._fallback_intent(user_copy_text, raw)
-        return intent
-
-    def _parse_intent(self, raw: str) -> IntentResult:
-        messages = [
-            CoreMessage(
-                text=n["text"],
-                must_keep=xp.attr_bool(n["attrs"].get("must_keep"), default=False),
-                confidence=xp.attr_float(n["attrs"].get("confidence"), 0.0),
-            )
-            for n in xp.find_all(raw, "message")
-            if n["text"]
-        ]
-        brands = [
-            PersonalBrand(text=n["text"], category=n["attrs"].get("category", "brand"))
-            for n in xp.find_all(raw, "brand")
-            if n["text"]
-        ]
-        missing = [n["text"] for n in xp.find_all(raw, "info") if n["text"]]
-        return IntentResult(
-            intent_summary=xp.text_of(raw, "intent_summary"),
-            core_messages=messages,
-            personal_brands=brands,
-            emotion_tone=xp.text_of(raw, "emotion_tone"),
-            missing_info=missing,
-            raw=raw,
-        )
-
-    def _fallback_intent(self, user_copy_text: str, raw: str) -> IntentResult:
-        text = (user_copy_text or "").strip()
-        messages = [CoreMessage(text=text[:80], must_keep=True, confidence=1.0)] if text else []
-        return IntentResult(
-            intent_summary=text[:30] or "未提供文案，按产品图片自由创作",
-            core_messages=messages,
-            personal_brands=[],
-            raw=raw,
-        )
-
-    # ── 步骤3：文案融合生成（三档）──────────────────────────────────────
-    def fuse(
-        self,
-        fusion_level: str,
-        image_analysis: ImageAnalysis,
-        intent: IntentResult,
-        industry: str = "",
-        target_customer: str = "",
-        marketing_purpose: str = "",
-        duration: int = 15,
-    ) -> FusionResult:
-        template = get_template("copy_fusion")
-        system_kwargs = {
-            "fusion_instruction": FUSION_INSTRUCTIONS.get(fusion_level, FUSION_INSTRUCTIONS["ai_polish"]),
-            "global_constraints": GLOBAL_CONSTRAINTS,
-            "negative_rules": NEGATIVE_RULES,
-        }
-        raw = self._chat(
-            template,
-            system_kwargs,
-            industry=industry or "通用",
-            target_customer=target_customer or "通用消费者",
-            marketing_purpose=marketing_purpose or "产品种草",
-            duration=duration,
-            image_analysis=self._image_brief(image_analysis),
-            intent_result=self._intent_brief(intent),
-        )
-        result = self._parse_fusion(raw)
-        if not result.title and not result.script_segments:
-            logger.warning("文案融合标签解析失败（fusion=%s），走规则 fallback", fusion_level)
-            return self._fallback_fusion(fusion_level, image_analysis, intent, duration, raw)
-        return result
-
-    def _parse_fusion(self, raw: str) -> FusionResult:
-        body_points = [
-            BodyPoint(
-                text=n["text"],
-                elaboration=n["attrs"].get("elaboration", ""),
-                image_index=xp.attr_int(n["attrs"].get("image_index"), 0),
-            )
-            for n in xp.find_all(raw, "point")
-            if n["text"]
-        ]
-        segments = [
-            ScriptSegment(
-                text=n["text"],
-                duration_sec=xp.attr_float(n["attrs"].get("duration_sec"), 0.0),
-                image_index=xp.attr_int(n["attrs"].get("image_index"), 0),
-            )
-            for n in xp.find_all(raw, "segment")
-            if n["text"]
-        ]
-        return FusionResult(
-            title=xp.text_of(raw, "title"),
-            hook=xp.text_of(raw, "hook"),
-            body_points=body_points,
-            cta=xp.text_of(raw, "cta"),
-            script_segments=segments,
-            word_count=xp.attr_int(xp.text_of(raw, "word_count"), 0),
-            estimated_duration=xp.attr_int(xp.text_of(raw, "estimated_duration"), 0),
-            raw=raw,
-        )
-
-    def _fallback_fusion(
-        self,
-        fusion_level: str,
-        image_analysis: ImageAnalysis,
-        intent: IntentResult,
-        duration: int,
-        raw: str,
-    ) -> FusionResult:
-        product_name = image_analysis.products[0].name if image_analysis.products else "这款产品"
-        selling = image_analysis.key_selling_points[:2]
-        if fusion_level == "ai_full":
-            title = f"{product_name}，很多人用完都回购了"
-            hook = f"这个{product_name}，我想认真说说"
-            body = selling or ["图片可见的产品卖点"]
-            cta = "感兴趣的可以了解一下"
-        elif fusion_level == "user_primary":
-            user_text = intent.intent_summary or product_name
-            title = user_text[:20]
-            hook = user_text[:15]
-            body = [m.text for m in intent.core_messages] or [user_text]
-            cta = "想了解的可以看看"
-        else:
-            title = intent.intent_summary[:20] or product_name
-            hook = intent.core_messages[0].text[:15] if intent.core_messages else product_name
-            body = [m.text for m in intent.core_messages] or selling or [product_name]
-            cta = "有需要的可以了解一下"
-
-        brand_texts = [b.text for b in intent.personal_brands]
-        points = [BodyPoint(text=b) for b in body]
-        lines = [hook] + body + brand_texts[:2] + [cta]
-        joined = "，".join(lines)
-        per = max(3, duration // max(1, len(lines)))
-        segments = [ScriptSegment(text=line, duration_sec=per, image_index=0) for line in lines]
-        return FusionResult(
-            title=title,
-            hook=hook,
-            body_points=points,
-            cta=cta,
-            script_segments=segments,
-            word_count=len(joined),
-            estimated_duration=duration,
-            raw=raw,
-        )
-
-    # ── 步骤4：编导级分镜 ───────────────────────────────────────────────
-    def storyboard(
-        self, fusion: FusionResult, image_analysis: ImageAnalysis, images: list[str], duration: int
-    ) -> Storyboard:
-        template = get_template("storyboard")
-        raw = self._chat(
-            template,
-            None,
-            duration=duration,
-            image_count=len(images),
-            fusion_result=self._fusion_brief(fusion),
-            image_analysis=self._image_brief(image_analysis),
-        )
-        board = self._parse_storyboard(raw)
-        if not board.clips:
-            logger.warning("分镜标签解析失败，走规则 fallback")
-            return self._fallback_storyboard(fusion, duration, raw)
-        return board
-
-    def _parse_storyboard(self, raw: str) -> Storyboard:
-        clips: list[Clip] = []
-        for node in xp.find_all(raw, "clip"):
-            attrs = node["attrs"]
-            body = node["text"]
-            kb = xp.find_first(node["text"] and f"<root>{node['text']}</root>", "ken_burns")
-            clips.append(
-                Clip(
-                    image_index=xp.attr_int(attrs.get("image_index"), 0),
-                    transition=attrs.get("transition", "cut"),
-                    zoom=(None if attrs.get("zoom") in (None, "null", "None", "") else attrs.get("zoom")),
-                    duration_sec=xp.attr_float(attrs.get("duration_sec"), 0.0),
-                    bgm_note=attrs.get("bgm_note", ""),
-                    voice_text=xp.text_of(body and f"<root>{body}</root>", "voice_text"),
-                    subtitle_text=xp.text_of(body and f"<root>{body}</root>", "subtitle_text"),
-                    ken_burns=KenBurns(
-                        start=kb["attrs"].get("start", "0,0") if kb else "0,0",
-                        end=kb["attrs"].get("end", "0,0") if kb else "0,0",
-                        ease=kb["attrs"].get("ease", "linear") if kb else "linear",
-                    ),
-                )
-            )
-        return Storyboard(clips=clips, raw=raw)
-
-    def _fallback_storyboard(self, fusion: FusionResult, duration: int, raw: str) -> Storyboard:
-        segments = fusion.script_segments or [ScriptSegment(text=fusion.hook or fusion.title, duration_sec=duration)]
-        total = sum(s.duration_sec for s in segments) or duration
-        clips = [
-            Clip(
-                image_index=min(s.image_index, 0),
-                transition="cut",
-                duration_sec=max(2.0, s.duration_sec * duration / total if total else duration / len(segments)),
-                voice_text=s.text,
-                subtitle_text=s.text[:20],
-            )
-            for s in segments
-        ]
-        return Storyboard(clips=clips, raw=raw)
-
-    # ── 步骤5：审核（不通过自动重写1次）─────────────────────────────────
-    def review_and_rewrite(
-        self, fusion: FusionResult, intent: IntentResult, fusion_level: str
-    ) -> tuple[FusionResult, ReviewResult, int]:
-        """返回最终文案、最后一次审核结果、重写次数（0或1）。"""
-        review = self.reviewer.review(fusion, intent, fusion_level)
-        if review.passed:
-            return fusion, review, 0
-
-        logger.info("文案审核不通过，自动重写 1 次：%s", [i.text for i in review.issues])
-        rewritten = self.reviewer.rewrite(fusion, review, intent, fusion_level)
-        second = self.reviewer.review(rewritten, intent, fusion_level)
-        if second.passed:
-            return rewritten, second, 1
-        # 二次仍不通过：带上重写结果和问题返回，由上游决定是否交给前端
-        return rewritten, second, 1
-
-    # ── 全流程编排 ──────────────────────────────────────────────────────
-    def generate(
-        self,
-        images: list[str],
-        *,
-        industry: str = "",
-        target_customer: str = "",
-        marketing_purpose: str = "",
-        duration: int = 15,
-        user_copy_text: str = "",
-        fusion_level: str = "ai_polish",
-    ) -> dict:
-        image_analysis = self.analyze_images(images, industry)
-        intent = self.parse_intent(user_copy_text, image_analysis, industry)
-        fusion = self.fuse(
-            fusion_level,
-            image_analysis,
-            intent,
-            industry=industry,
-            target_customer=target_customer,
-            marketing_purpose=marketing_purpose,
-            duration=duration,
-        )
-        fusion, review, rewrites = self.review_and_rewrite(fusion, intent, fusion_level)
-        board = self.storyboard(fusion, image_analysis, images, duration)
-        return {
-            "image_analysis": image_analysis,
-            "intent_result": intent,
-            "fusion_result": fusion,
-            "review_result": review,
-            "storyboard": board,
-            "rewrite_count": rewrites,
-        }
-
-    # ── 简报工具 ────────────────────────────────────────────────────────
-    @staticmethod
-    def _image_brief(a) -> str:
-        if a is None:
-            return "无图片分析信息"
-        lines = [f"产品：{p.name}（{p.features}）" for p in a.products]
-        lines += [f"卖点：{s}" for s in a.key_selling_points]
-        lines.append(f"场景：{a.scene}")
-        return "\n".join(lines) or "无图片分析信息"
-
-    @staticmethod
-    def _intent_brief(i: IntentResult) -> str:
-        lines = [f"意图：{i.intent_summary}"]
-        lines += [f"核心信息[must_keep={m.must_keep}]：{m.text}" for m in i.core_messages]
-        lines += [f"事实（{b.category}）：{b.text}" for b in i.personal_brands]
-        return "\n".join(lines)
-
-    @staticmethod
-    def _fusion_brief(f: FusionResult) -> str:
-        lines = [f"标题：{f.title}", f"钩子：{f.hook}"]
-        lines += [f"要点：{p.text}" for p in f.body_points]
-        lines += [f"配音：{s.text}" for s in f.script_segments]
-        lines.append(f"行动号召：{f.cta}")
-        return "\n".join(lines)

packages/application/viral_video/prompt_loader.py

@@ -17,7 +17,7 @@ from packages.application.viral_video.prompts import DEFAULT_TEMPLATES
 
 CACHE_TTL_SECONDS = 30.0
 
-_VALID_TYPES = {"image_analysis", "intent_parsing", "copy_fusion", "storyboard", "review"}
+_VALID_TYPES = {"image_analysis", "storyboard", "review"}
 
 
 @dataclass

packages/application/viral_video/prompts.py

@@ -1,339 +1,246 @@
-"""爆款视频 5 套 Prompt 模板默认值（#2040 核心资产）。
+# -*- coding: utf-8 -*-
+"""爆款视频 Prompt 模板默认值（v8 / v3 叙述优先重构，PR1 清理版）。
 
-重要约定（用户明确要求）：
-- 所有 system_prompt / user_prompt_template / example_output 都是**纯文本自然语言 + XML 标签**，
-  运营可直接看懂和编辑，禁止 JSON、禁止 ```json 代码块。
-- LLM 按 XML 标签输出字段，程序用正则解析（见 xml_parser.py）。
-- user_prompt_template 中花括号占位符（如 {user_copy_text}）在运行时填充。
+设计原则（灵应 2026-10-07）：LLM 直接输出最终给用户看的文案，代码尽量薄。
+- image_analysis v8：VLM 主交付物是自然叙述风格的 summary_markdown，结构化
+  字段仅保留 type/name/brand/has_person，顶层 products 改名 images；
+- storyboard v3：口播台词口语化、画面描述有画面感，copy_display_markdown 是
+  LLM 直接写给用户看的流畅叙述文案，代码只做解析不改写；
+- intent_parsing / copy_fusion 两步整体删除，意图理解并入 storyboard 一次调用。
+
+模板字段与 DB 表 viral_video_prompt_templates、prompt_loader 完全对应：
+name / prompt_type / version(int) / system_prompt / user_prompt_template /
+example_output / is_active。
 """
 
 from __future__ import annotations
 
-TEMPLATE_VERSION = 1
-
 # 所有文案类 Prompt 自动注入的硬约束
 GLOBAL_CONSTRAINTS = """【必须遵守的硬约束】
-1. 不编造时间：不写“今年最新”“2024 爆款”等会过时的时间表述。
-2. 不承诺效果：不写“保证”“一定”“100%有效”“包治百病”等绝对化用语。
+1. 不编造时间：不写"今年最新""2024 爆款"等会过时的时间表述。
+2. 不承诺效果：不写"保证""一定""100%有效""包治百病"等绝对化用语。
 3. 不编造价格、销量、认证、奖项：除非用户在文案中明确给出，否则一律不写。
 4. 符合广告法及平台社区规范。
 5. 只描述图片中真实可见的内容，看不到的不瞎猜。"""
 
-# 反套路化要求
+# 负向提示（注入 storyboard / 视频生成负面词）
 NEGATIVE_RULES = """【反套路化要求】
-禁止使用“家人们谁懂啊”“绝绝子”“宝子们”“家人们”“太绝了”“yyds”等烂大街网络词；
+禁止使用"家人们谁懂啊""绝绝子""宝子们""家人们""太绝了""yyds"等烂大街网络词；
+禁止使用"很近"等错别字，应使用"最近"；
 禁止固定模板化开头；语言要像真人朋友之间的分享，自然、具体、有信息量。"""
 
-# 输出禁用套路词（测试会检查）
-BANNED_PHRASES = ["家人们谁懂啊", "绝绝子", "宝子们", "yyds", "太绝了"]
+# 输出禁用套路词（测试/reviewer/worker 三处共享）
+# 注：广告法绝对化用语（如"最""第一""100%"）在 worker 侧 _quick_compliance_blacklist_check 单独做本地替换，
+# 这里只列出明显的网络烂梗/套路词。
+COMPLIANCE_BLACKLIST: list[str] = [
+    "家人们谁懂啊",
+    "绝绝子",
+    "宝子们",
+    "yyds",
+    "太绝了",
+    "家人们",
+]
 
-# 文案融合三档独立指令段
-FUSION_INSTRUCTIONS = {
-    "ai_full": """【本次创作模式：AI 全权创作】
-你是资深短视频编导。用户只提供了产品图片，没有给出具体文案方向。请根据图片内容和营销参数，自由发挥创作完整的爆款短视频文案。充分挖掘产品真实可见的卖点，使用爆款结构，抓人眼球。""",
-    "ai_polish": """【本次创作模式：AI 辅助润色】
-你是用户的文案助理。用户已经写了草稿/关键词/碎碎念，表达了他想讲的核心意思，但表达不完整、不够吸引人。你的任务是：以用户的意思为主，保留他想表达的所有核心信息点，在此基础上润色扩写、调整语序、增加衔接、优化表达，让文案更流畅更有吸引力。绝对不能改变用户想表达的核心意思，不能把用户的观点换成相反的，不能添加用户没提到的产品卖点。用户提到的品牌名、价格、人名、具体事实必须原样保留。""",
-    "user_primary": """【本次创作模式：以用户原文为主】
-你是文案润色助手。用户已经写好了明确的文案，这是他最终想表达的内容。你的任务是最小化修改：只做必要的错别字修正、标点调整、语句通顺度优化，以及添加必要的衔接词让口播更自然。用户的核心句子、关键表述、事实信息一律不改。如果用户文案本身已经很好，直接返回，不要为了改而改。personal_brands 中的事实信息必须逐字保留。""",
+# 向后兼容别名
+BANNED_PHRASES = COMPLIANCE_BLACKLIST
+
+# ── 模板1：图片多模态分析 v8（叙述优先）───────────────────────────────
+_IMAGE_ANALYSIS_SYSTEM = """你是一名擅长观察和写作的品牌内容编导。面对一张真实图片，先用眼睛仔细看，再用自然、流畅、具体的中文把画面写成一段可以直接读给人听的描述。
+
+## 输出格式（严格 JSON，不要输出 JSON 以外的任何内容）
+{
+  "images": [
+    {
+      "type": "store 或 product 或 person 或 scene，四选一",
+      "name": "主体名称，看不出就写"未识别"",
+      "brand": "品牌名，看不出就留空字符串",
+      "has_person": false,
+      "summary_markdown": "用 Markdown 写成的自然叙述，这是最主要的交付物"
+    }
+  ]
 }
 
-# ── 模板1：图片多模态分析（VLM）────────────────────────────────────────
-_IMAGE_ANALYSIS_SYSTEM = f"""你是电商商品视觉分析师，负责从商品图片中提取真实可见的商品信息。
+## summary_markdown 写作要求（最重要）
+1. 写成完整、通顺的句子，像在跟朋友认真描述你看到的画面；不要用分号堆砌关键词，不要罗列"核心特征：xxx""主色调：xxx"这类填表式标签。
+2. 开头先给一句整体定性，让读者立刻明白这是什么场景、什么主体。
+3. 颜色、材质、形状、部件要具体可感，写到位置和搭配；画面里出现的文字原样读出并自然融进句子，数字、规格、价格精确引用，看不清的不要编造。
+4. 只写真实看到的内容，不脑补功能、疗效、销量或画面之外的信息。
+5. 长度控制在 200-500 字。
 
-工作方式（分步骤看，不要跳步）：
-1. 先看整体：有哪些产品、什么场景、有没有人物。
-2. 再看细节：包装文字、颜色构成、人物状态、画面质感。
-3. 最后提炼卖点：只总结图片里能看到的卖点。
+## 按类型组织内容
+- type=store（门店/店内环境）：用以下小标题分段，小标题下写连贯的句子而不是清单：
+###店铺主体
+###周边物品
+1.家具陈设
+2.商品与标识
+- type=product（商品）：按自然段从整体到局部描写——先说是什么、什么品牌，再写包装/外形、颜色与材质、标签文字、可见部件与规格。
+- type=person（人物）：描述人物身份感、姿态、穿着（上下装/颜色/款式）、动作与所处环境；用于品牌宣传时突出其精神状态。
+- type=scene（纯场景/风景）：描述空间或风景的构成、色彩、光线、氛围与关键物件。
 
-{GLOBAL_CONSTRAINTS}
+## 判断规则
+- has_person：画面中出现可辨识的真实人物（脸或完整上半身）才为 true，海报/模特立牌/照片里的人不算。
+- 一张图只描述其本身；多张图属于同一场景时可呼应，但不编造对应关系。
+- 输出必须是严格 JSON，summary_markdown 是字符串，内部换行用 \\n 表示。"""
 
-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、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。
+_IMAGE_ANALYSIS_USER = """请分析这张图片。
+图片地址：{image_url}
+OCR 辅助文字（可能为空，仅供参考，不要照抄错误识别）：{ocr_text}
 
-【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。
+严格按系统要求只输出 JSON。"""
 
-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。
+_IMAGE_ANALYSIS_EXAMPLE = """{
+  "images": [
+    {
+      "type": "store",
+      "name": "御众堂门店",
+      "brand": "御众堂",
+      "has_person": false,
+      "summary_markdown": "###店铺主体\\n这是一家名为"御众堂"的线下门店内部，整体暖木色调……"
+    }
+  ]
+}"""
 
-【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"/>"""
+# ── 模板2：编导级分镜 v3（意图理解 + 分镜一次完成）────────────────────
+# 字数/镜头数硬约束在 user_prompt 通过 {duration_constraint_text} 占位符注入，
+# 由 constants.build_duration_constraint_text(dur) 统一生成，避免模板与代码校验不一致。
+_STORYBOARD_SYSTEM = (
+    """你是一名懂短视频的编导和口播文案高手。你会拿到图片的真实观察、营销目的和用户参数，请一次性完成对营销意图的理解，并产出可直接拍摄/生成的分镜脚本。不要单独输出"意图解析"，意图要直接体现在台词和分镜里。
 
-_IMAGE_ANALYSIS_USER = """请分析以下商品图片，共 {image_count} 张。
-所属行业：{industry}
-图片地址：
-{image_urls}
+## 输出格式（XML，严格按结构输出，不要输出额外解释）
+<script>
+  <copy_display_markdown><![CDATA[直接展示给用户看的成片文案，用 Markdown 写成流畅叙述]]></copy_display_markdown>
+  <clips>
+    <clip index="1">
+      <time_range>0-3秒</time_range>
+      <voiceover>这一镜的口播台词</voiceover>
+      <visual>具体、有画面感的镜头描述（主体/动作/镜头运动/景别/光线）</visual>
+      <reference_image_index>0</reference_image_index>
+    </clip>
+  </clips>
+  <voiceover_script>把所有 clip 的 voiceover 连成完整口播稿</voiceover_script>
+  <theme>一句话主题</theme>
+  <negative>"""
+    + NEGATIVE_RULES
+    + """</negative>
+</script>
 
-按约定的标签格式输出分析结果。"""
+## 写作要求
+1. 口播台词：像真人面对镜头说话，短句、口语化、有停顿有情绪，开头 3 秒给出钩子；不要书面腔，不要机械报参数。
+2. 画面描述：写清"观众会看到什么"，有动作、有镜头运动、有景别和光线，具体可拍；不堆砌形容词，不写无法实现的画面。
+3. copy_display_markdown：直接展示给最终用户的文案，用 Markdown 写成自然、流畅、有感染力的成片文案，可用小标题与短句组织；不要做字段列表，不要出现"镜头一/台词："这类制作说明。
+4. 内容必须来自图片观察与用户给出的信息，不编造卖点、不夸大、不使用绝对化用语和虚假承诺。
+5. reference_image_index 填本镜参考图片序号（从 0 开始），没有合适参考图填 -1。
+6. 分镜数量与时长匹配总时长，节奏紧凑。
+7. 【字数/镜头数/时间轴硬约束】：严格遵守 user_prompt 里给出的 {duration_constraint_text}，宁少勿多。
+8. 每个 clip 必须分配一个 reference_image_index（从 0 开始的图片序号），没有合适图片填 -1。
+9. 必须严格按<marketing_purpose><target_audience><persona><viral_structure><language><industry>指定的参数写文案和分镜，不能忽略任何一项用户参数。
+10. 禁止使用"很近"等错别字，统一使用"最近"。"""
+)
 
-_IMAGE_ANALYSIS_EXAMPLE = """<products>
-<product name="大公鸡头 多功能油污净 625ml" features="红色瓶盖白色瓶身，鸡头图案Logo" position="main" image_index="0"/>
-</products>
-<colors>
-<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" 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"/>
-</visible_text>
-<scene>白底棚拍产品图</scene>
-<quality resolution="高清" lighting="均匀柔和" composition="主体居中" blur="false"/>
-<key_selling_points>
-<point>针对重油污设计</point>
-<point>大容量625ml</point>
-</key_selling_points>"""
+_STORYBOARD_USER = """<marketing_purpose>{marketing_purpose}</marketing_purpose>
+<industry>{industry}</industry>
+<image_analysis>
+{image_summary}
+</image_analysis>
+<user_parameters>
+<theme_hint>{theme_hint}</theme_hint>
+<duration>{duration}秒</duration>
+<aspect_ratio>{aspect_ratio}</aspect_ratio>
+<tone>{tone}</tone>
+<target_audience>{target_audience}</target_audience>
+<persona>{persona_hint}</persona>
+<viral_structure>{viral_structure_hint}</viral_structure>
+<language>{language_hint}</language>
+<extra_requirements>{extra_requirements}</extra_requirements>
+</user_parameters>
+{video_style_section}
+<duration_constraints>
+{duration_constraint_text}
+</duration_constraints>
+请严格按 XML 结构输出分镜脚本。"""
 
-# ── 模板2：用户文案意图解析（LLM）──────────────────────────────────────
-_INTENT_SYSTEM = f"""你负责理解用户的营销意图。用户给的文案可能只是几个关键词、碎碎念或者不完整的短句，你要读懂他真正想讲什么。
+_STORYBOARD_EXAMPLE = """<script>
+  <copy_display_markdown><![CDATA[# 在御众堂，把松弛的自己一点点找回来
+产后妈妈最懂那种力不从心，推开门，暖光和一杯热茶先接住了你。
+最近连抱娃都觉得腰发硬，过来躺一会儿，按完整个背都松了，带娃也有力气了。]]></copy_display_markdown>
+  <clips>
+    <clip index="1">
+      <time_range>0-4秒</time_range>
+      <voiceover>生完娃，是不是连照镜子的勇气都没了？</voiceover>
+      <visual>中近景，暖光下一位妈妈略显疲惫地看向镜中，镜头缓缓推近</visual>
+      <reference_image_index>0</reference_image_index>
+    </clip>
+    <clip index="2">
+      <time_range>4-10秒</time_range>
+      <voiceover>推开门，暖光和一杯热茶先接住了你。</voiceover>
+      <visual>中景，门店前台暖光木色装修，店员微笑递上茶水，镜头平移扫过环境</visual>
+      <reference_image_index>1</reference_image_index>
+    </clip>
+    <clip index="3">
+      <time_range>10-15秒</time_range>
+      <voiceover>来御众堂躺一会儿，整个人都松了。</voiceover>
+      <visual>特写，妈妈闭眼放松做护理，暖光打在侧脸，最后镜头拉远露出品牌LOGO</visual>
+      <reference_image_index>2</reference_image_index>
+    </clip>
+  </clips>
+  <voiceover_script>生完娃，是不是连照镜子的勇气都没了？推开门，暖光和一杯热茶先接住了你。来御众堂躺一会儿，整个人都松了。</voiceover_script>
+  <theme>产后妈妈走进御众堂重拾状态</theme>
+  <negative>模糊、畸变、夸大疗效、绝对化用语、"很近"等错别字</negative>
+</script>"""
 
-{GLOBAL_CONSTRAINTS}
+# ── 模板3：文案审核（合规/质量门禁）───────────────────────────────────
+_REVIEW_SYSTEM = """你是一名短视频广告合规审核与文案优化专家。请审核下面这段口播文案：
+1) 广告法与平台合规（绝对化用语、虚假承诺、医疗功效宣称、导流违规、套路烂梗词）；
+2) 口播是否自然、口语化、无书面腔；
+3) 是否有明显错别字（例如把"最近"写成"很近"）。
 
-请严格按下面的标签格式输出，不要解释，不要用代码块：
-<intent_summary> 用用户的语言风格，一句话、30字以内概括核心意图。
-<core_messages> 下面每个核心信息点用一个 <message> 标签，属性 must_keep 为 true 或 false、confidence 为 0 到 1 的小数，标签内容写信息点。
-<personal_brands> 把用户提到的具体事实——品牌名、价格、人名、地名、时间、产品名——每条用一个 <brand> 标签，属性 category 取 brand、price、person、place、time、product 之一。这些事实必须原样引用，一个字都不能改。
-<emotion_tone> 写文案的情绪调性。
-<missing_info> 把你认为缺失、后续生成时需要合理推断的信息，每条用一个 <info> 标签；没有就输出空标签。"""
+只输出 XML，结构：
+<review>
+  <passed>true 或 false</passed>
+  <issues>
+    <issue dimension="违规词/夸大承诺/错别字/其他" severity="error或warning">具体问题</issue>
+  </issues>
+</review>
+没有问题时 issues 留空、passed 为 true。不要输出 rewrite/rewrite_suggestion 字段。"""
 
-_INTENT_USER = """用户原始文案：{user_copy_text}
-所属行业：{industry}
-营销目的：{marketing_purpose}
-图片分析结果（供参考）：
-{image_analysis}
-图片类型推断：{image_category_hint}
+_REVIEW_USER = """<voiceover_script>
+{voiceover_script}
+</voiceover_script>
 
-请理解用户意图，按标签格式输出。注意：theme和emotion_tone应与图片类型和营销目的匹配——门店类图片偏向"门店探店/到店体验"，商品图偏向"好物分享/产品种草"，人物图偏向"穿搭/人物故事"。"""
+请审核以上口播文案。"""
 
-_INTENT_EXAMPLE = """<intent_summary>一款厨房去油污神器，喷一喷油污就掉</intent_summary>
-<core_messages>
-<message must_keep="true" confidence="0.97">去油污效果好，喷上等几分钟再擦</message>
-<message must_keep="false" confidence="0.7">适合厨房重油污场景</message>
-</core_messages>
-<personal_brands>
-<brand category="product">大公鸡头多功能油污净</brand>
-<brand category="price">39块钱一瓶</brand>
-</personal_brands>
-<emotion_tone>亲切、真实、带分享感</emotion_tone>
-<missing_info>
-<info>没有说明具体容量，按图片读出的625ml处理</info>
-</missing_info>"""
-
-# ── 模板3：文案融合生成（LLM）──────────────────────────────────────────
-_FUSION_SYSTEM = """你负责为短视频生成营销文案。请按思维链分步完成：先定人设和目标客户，再找卖点，再搭结构，再安排情绪，最后写行动号召，不要一步到位乱写。
-
-{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：
-<title> 视频标题。
-<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（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。
-
-用户在 personal_brands 中提到的品牌名、价格、人名、地名、时间、产品名等事实信息，必须原样出现在文案里，一个字都不能改。"""
-
-_FUSION_USER = """所属行业：{industry}
-目标客户：{target_customer}
-营销目的：{marketing_purpose}
-视频时长：{duration}秒
-图片分析结果：
-{image_analysis}
-用户意图解析结果：
-{intent_result}
-
-请按标签格式生成文案。"""
-
-_FUSION_EXAMPLE = """<title>厨房重油污，别再用洗洁精硬擦了</title>
-<hook>这油污，我真的忍很久了</hook>
-<body_points>
-<point elaboration="喷在油污上等几分钟，一擦就干净" image_index="0">大公鸡头油污净去油快</point>
-<point elaboration="39块钱625ml，能用很久" image_index="0">39块钱一瓶，性价比高</point>
-</body_points>
-<cta>厨房油污重的，真的可以试一瓶</cta>
-<script_segments>
-<segment duration_sec="3" image_index="0">这油污我真的忍很久了，用洗洁精擦半天都没用</segment>
-<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>厨房好物分享·产品种草</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 = """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。
-
-{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/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）
-文案内容：
-{fusion_result}
-图片分析结果：
-{image_analysis}
-
-重要：overview_theme 必须与图片实际内容和营销目的匹配。门店/餐饮/服务类图片用"门店探店·到店体验"；商品图用"好物分享·产品种草"；人物图用"穿搭分享·人物故事"；场景图用"空间体验·场景氛围"。不要对所有图片都使用"好物分享"。
-
-请按标签格式输出分镜。"""
-
-_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>
-<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>"""
-
-# ── 模板5：文案审核（LLM）──────────────────────────────────────────────
-_REVIEW_SYSTEM = f"""你是短视频文案合规审核员，从6个维度逐条检查文案：
-1. 违规词：有没有平台禁用词、敏感词。
-2. 夸大承诺：有没有“包治百病”“100%有效”“保证赚钱”等绝对化、夸大表述。
-3. 事实一致性：有没有编造价格、数据、认证，或者用户没提到的产品特性。
-4. 用户意图保留：在 ai_polish 和 user_primary 模式下，core_messages 中 must_keep=true 的点是否都保留了。
-5. 结构完整性：标题、钩子、正文、行动号召是否齐全。
-6. 语气人设：是否符合选定的人设语气，有没有“家人们谁懂啊”“绝绝子”“宝子们”等套路词。
-
-{GLOBAL_CONSTRAINTS}
-
-请严格按下面的标签格式输出，不要解释，不要用代码块：
-<passed> 整体是否通过，只写 true 或 false。
-<issues> 每个问题用一个 <issue> 标签，属性 dimension 是维度名、severity 取 error 或 warning、location 是问题所在（如 hook、body_points、cta），标签内容写问题描述；没有问题就输出空标签。
-<rewrite_suggestions> 每条具体修改建议用一个 <suggestion> 标签；没有就输出空标签。"""
-
-_REVIEW_USER = """本次创作模式：{fusion_level}
-待审核文案：
-{fusion_result}
-用户意图解析（用于核对核心信息是否保留）：
-{intent_result}
-
-请按6个维度审核，按标签格式输出。"""
-
-_REVIEW_EXAMPLE = """<passed>false</passed>
-<issues>
-<issue dimension="夸大承诺" severity="error" location="body_points">出现了“一喷100%掉光”的绝对化表述，违反广告法</issue>
-<issue dimension="用户意图保留" severity="warning" location="cta">用户强调的“39块钱”没有保留</issue>
-</issues>
-<rewrite_suggestions>
-<suggestion>把“一喷100%掉光”改为“喷上等几分钟，大部分油污能擦掉”</suggestion>
-<suggestion>在结尾补回“39块钱625ml”</suggestion>
-</rewrite_suggestions>"""
+_REVIEW_EXAMPLE = """<review>
+  <passed>false</passed>
+  <issues>
+    <issue dimension="夸大承诺" severity="error">使用绝对化用语"100%有效"，违反广告法</issue>
+    <issue dimension="违规词" severity="warning">使用套路词"绝绝子"</issue>
+  </issues>
+</review>"""
 
 
-# 5 套模板默认数据（seed 数据源与 loader 的兜底）
 DEFAULT_TEMPLATES: list[dict] = [
     {
-        "name": "图片多模态分析",
+        "name": "图片多模态分析 v8",
         "prompt_type": "image_analysis",
-        "version": TEMPLATE_VERSION,
+        "version": 8,
         "system_prompt": _IMAGE_ANALYSIS_SYSTEM,
         "user_prompt_template": _IMAGE_ANALYSIS_USER,
         "example_output": _IMAGE_ANALYSIS_EXAMPLE,
         "is_active": True,
     },
     {
-        "name": "用户文案意图解析",
-        "prompt_type": "intent_parsing",
-        "version": TEMPLATE_VERSION,
-        "system_prompt": _INTENT_SYSTEM,
-        "user_prompt_template": _INTENT_USER,
-        "example_output": _INTENT_EXAMPLE,
-        "is_active": True,
-    },
-    {
-        "name": "文案融合生成",
-        "prompt_type": "copy_fusion",
-        "version": TEMPLATE_VERSION,
-        "system_prompt": _FUSION_SYSTEM,
-        "user_prompt_template": _FUSION_USER,
-        "example_output": _FUSION_EXAMPLE,
-        "is_active": True,
-    },
-    {
-        "name": "编导级分镜",
+        "name": "编导级分镜 v3.1",
         "prompt_type": "storyboard",
-        "version": TEMPLATE_VERSION,
+        "version": 4,
         "system_prompt": _STORYBOARD_SYSTEM,
         "user_prompt_template": _STORYBOARD_USER,
         "example_output": _STORYBOARD_EXAMPLE,
         "is_active": True,
     },
     {
-        "name": "文案审核",
+        "name": "文案审核 v2",
         "prompt_type": "review",
-        "version": TEMPLATE_VERSION,
+        "version": 2,
         "system_prompt": _REVIEW_SYSTEM,
         "user_prompt_template": _REVIEW_USER,
         "example_output": _REVIEW_EXAMPLE,

packages/application/viral_video/reviewer.py

@@ -1,7 +1,12 @@
-"""文案审核 + 自动重写（#2040 第5套 Prompt）。
+# -*- coding: utf-8 -*-
+"""文案审核（PR1 简化版）。
 
-6 维度：违规词 / 夸大承诺 / 事实一致性 / 用户意图保留 / 结构完整性 / 语气人设。
-LLM 审核之外叠加本地规则预检（保证即使 LLM 不可用也能兜住广告法红线）。
+只做两件事：
+1. 本地规则预检：广告法绝对化用语 + 套路烂梗词（COMPLIANCE_BLACKLIST）；
+2. LLM 审核（单参数 voiceover_script），LLM 不可用时降级放行并塞 warning issue。
+
+不再有 rewrite / rule_fix / intent 保留检查等死路径（intent 在新链路里永远是空的，
+personal_brands/must_keep 检查在原链路里也从未真正生效）。
 """
 
 from __future__ import annotations
@@ -16,16 +21,12 @@ from packages.application.viral_video.prompt_loader import (
     render_system_prompt,
     render_user_prompt,
 )
-from packages.application.viral_video.schemas import (
-    FusionResult,
-    IntentResult,
-    ReviewIssue,
-    ReviewResult,
-)
+from packages.application.viral_video.prompts import COMPLIANCE_BLACKLIST
+from packages.application.viral_video.schemas import ReviewIssue, ReviewResult
 
 logger = logging.getLogger(__name__)
 
-# 本地规则：绝对化/夸大词
+# 本地规则：绝对化/夸大词（广告法红线，兜底挡住最明显违规）
 _EXAGGERATION_PATTERNS = [
     r"100\s*%",
     r"百分百",
@@ -39,61 +40,87 @@ _EXAGGERATION_PATTERNS = [
     r"一喷(就|全|100)",
 ]
 
-# 本地规则：平台违规/套路词
-_VIOLATION_PHRASES = [
-    "家人们谁懂啊",
-    "绝绝子",
-    "宝子们",
-    "yyds",
-    "最(好|强|牛|便宜)",  # 广告法极限词
-    "第一(名|品牌)?",
+# 本地规则：套路烂梗/明显错别字（与 prompts.COMPLIANCE_BLACKLIST 同步）
+_VIOLATION_PATTERNS = [re.escape(p) for p in COMPLIANCE_BLACKLIST] + [
+    r"最(好|强|牛|便宜|美|棒)",  # 广告法极限词（最X）
+    r"第一(名|品牌)?",
+    r"很近",  # 常见错别字，应使用"最近"
 ]
 
-_LOCATIONS = ["title", "hook", "body_points", "cta", "script_segments"]
-
 
 class Reviewer:
+    """口播文案审核器。"""
+
     def __init__(self, client=None):
         if client is None:
             try:
                 from packages.shared.ai_router import ai_router
 
                 client = ai_router.get_llm_client("copy_review")
-            except Exception:
-                from packages.shared.ai_client import get_doubao_client
+            except Exception:  # noqa: BLE001
+                try:
+                    from packages.shared.ai_client import get_doubao_client
 
-                client = get_doubao_client()
+                    client = get_doubao_client()
+                except Exception:  # noqa: BLE001
+                    client = None
         self.client = client
 
     # ── 审核 ────────────────────────────────────────────────────────────
-    def review(self, fusion: FusionResult, intent: IntentResult, fusion_level: str) -> ReviewResult:
-        local = self._rule_check(fusion, intent, fusion_level)
-        llm_result = self._llm_review(fusion, intent, fusion_level)
+    def review(
+        self,
+        voiceover_script: str,
+        copy_display_markdown: str = "",
+    ) -> ReviewResult:
+        """审核口播文案。
+
+        Args:
+            voiceover_script: 完整口播稿字符串（必传）。
+            copy_display_markdown: 展示给用户看的 Markdown 文案（可选，保留接口）。
+
+        Returns:
+            ReviewResult: passed=True 表示放行；LLM 不可用时降级放行（带 warning issue）。
+        """
+        text = voiceover_script or ""
+        local = self._check_compliance(text)
+        llm_result = self._llm_review(text)
         if llm_result is None:
+            # LLM 审核失败（超时/网络/不可用），降级放行
+            degraded_issue = ReviewIssue(
+                dimension="审核降级",
+                severity="warning",
+                location="voiceover_script",
+                text="LLM 审核不可用，已降级为本地规则审核放行，请人工复核",
+            )
             return ReviewResult(
-                passed=not local,
-                issues=local,
+                passed=not any(i.severity == "error" for i in local),
+                issues=local + [degraded_issue],
                 rewrite_suggestions=[],
                 raw="",
             )
         # LLM 与本地规则合并去重
         issues = self._merge_issues(llm_result.issues, local)
+        has_error = any(i.severity == "error" for i in issues)
         return ReviewResult(
-            passed=llm_result.passed and not local,
+            passed=llm_result.passed and not has_error,
             issues=issues,
             rewrite_suggestions=llm_result.rewrite_suggestions,
             raw=llm_result.raw,
         )
 
-    def _llm_review(self, fusion: FusionResult, intent: IntentResult, fusion_level: str) -> Optional[ReviewResult]:
+    def _llm_review(self, voiceover_script: str) -> Optional[ReviewResult]:
+        try:
+            return self._llm_review_inner(voiceover_script)
+        except Exception as e:  # noqa: BLE001
+            logger.warning("[Reviewer] LLM审核调用异常，降级放行: %s", e)
+            return None
+
+    def _llm_review_inner(self, voiceover_script: str) -> Optional[ReviewResult]:
+        if not self.client or not getattr(self.client, "is_available", True):
+            return None
         template = get_template("review")
         system = render_system_prompt(template)
-        user = render_user_prompt(
-            template,
-            fusion_level=fusion_level,
-            fusion_result=self._fusion_text(fusion),
-            intent_result=self._intent_text(intent),
-        )
+        user = render_user_prompt(template, voiceover_script=voiceover_script)
         raw = self.client.chat_completion(
             [
                 {"role": "system", "content": system},
@@ -105,84 +132,52 @@ class Reviewer:
         )
         if not raw:
             return None
-        passed = xp.text_of(raw, "passed").strip().lower()
+        passed_text = xp.text_of(raw, "passed").strip().lower()
         issues = [
             ReviewIssue(
                 dimension=n["attrs"].get("dimension", "未知维度"),
                 severity=n["attrs"].get("severity", "warning"),
-                location=n["attrs"].get("location", ""),
+                location=n["attrs"].get("location", "voiceover_script"),
                 text=n["text"],
             )
             for n in xp.find_all(raw, "issue")
             if n["text"]
         ]
         suggestions = [n["text"] for n in xp.find_all(raw, "suggestion") if n["text"]]
-        parsed = ReviewResult(
-            passed=passed == "true" and not issues,
+        return ReviewResult(
+            passed=passed_text == "true" and not issues,
             issues=issues,
             rewrite_suggestions=suggestions,
             raw=raw,
         )
-        return parsed
 
     # ── 本地规则预检 ────────────────────────────────────────────────────
-    def _rule_check(self, fusion: FusionResult, intent, fusion_level: str) -> list[ReviewIssue]:
+    def _check_compliance(self, text: str) -> list[ReviewIssue]:
+        """扫描口播文本中的广告法红线和套路词。"""
         issues: list[ReviewIssue] = []
-        for location, text in self._segments(fusion):
-            for pattern in _EXAGGERATION_PATTERNS:
-                if re.search(pattern, text):
-                    issues.append(
-                        ReviewIssue(
-                            dimension="夸大承诺",
-                            severity="error",
-                            location=location,
-                            text=f"出现夸大/绝对化表述：{self._hit(text, pattern)}",
-                        )
+        if not text:
+            return issues
+        for pattern in _EXAGGERATION_PATTERNS:
+            if re.search(pattern, text):
+                issues.append(
+                    ReviewIssue(
+                        dimension="夸大承诺",
+                        severity="error",
+                        location="voiceover_script",
+                        text="出现夸大/绝对化表述：" + self._hit(text, pattern),
                     )
-            for phrase in _VIOLATION_PHRASES:
-                if re.search(phrase, text, flags=re.IGNORECASE):
-                    issues.append(
-                        ReviewIssue(
-                            dimension="违规词",
-                            severity="error",
-                            location=location,
-                            text=f"出现违规或套路词：{self._hit(text, phrase)}",
-                        )
-                    )
-
-        # 结构完整性
-        if not fusion.title:
-            issues.append(ReviewIssue(dimension="结构完整性", severity="warning", location="title", text="缺少标题"))
-        if not fusion.hook:
-            issues.append(ReviewIssue(dimension="结构完整性", severity="warning", location="hook", text="缺少开头钩子"))
-        if not fusion.cta:
-            issues.append(ReviewIssue(dimension="结构完整性", severity="warning", location="cta", text="缺少行动号召"))
-
-        # 用户意图保留（must_keep）
-        full_text = self._fusion_text(fusion)
-        if fusion_level in {"ai_polish", "user_primary"} and intent is not None:
-            for message in intent.core_messages:
-                if message.must_keep:
-                    key = self._compact(message.text)
-                    if key and key[:10] not in self._compact(full_text):
-                        issues.append(
-                            ReviewIssue(
-                                dimension="用户意图保留",
-                                severity="warning",
-                                location="script_segments",
-                                text=f"用户核心信息被丢失：{message.text[:30]}",
-                            )
-                        )
-            for brand in intent.personal_brands:
-                if brand.text and brand.text not in full_text:
-                    issues.append(
-                        ReviewIssue(
-                            dimension="事实一致性",
-                            severity="error",
-                            location="script_segments",
-                            text=f"personal_brands 事实信息未原样保留：{brand.text[:30]}",
-                        )
+                )
+        for pattern in _VIOLATION_PATTERNS:
+            if re.search(pattern, text, flags=re.IGNORECASE):
+                sev = "warning" if pattern == "很近" else "error"
+                issues.append(
+                    ReviewIssue(
+                        dimension="违规词",
+                        severity=sev,
+                        location="voiceover_script",
+                        text="出现违规/套路词或错别字：" + self._hit(text, pattern),
                     )
+                )
         return issues
 
     @staticmethod
@@ -192,7 +187,7 @@ class Reviewer:
 
     @staticmethod
     def _compact(text: str) -> str:
-        return re.sub(r"[\s，。！？、,.!?；;：:\"'“”‘’（）()【】\[\]]", "", text)
+        return re.sub(r"[\s，。！？、,.!?；;：:\"'\u201c\u201d\u2018\u2019（）()【】\[\]]", "", text)
 
     @staticmethod
     def _merge_issues(llm_issues: list[ReviewIssue], local: list[ReviewIssue]) -> list[ReviewIssue]:
@@ -204,116 +199,3 @@ class Reviewer:
                 merged.append(issue)
                 seen.add(key)
         return merged
-
-    # ── 自动重写（1 次）─────────────────────────────────────────────────
-    def rewrite(
-        self,
-        fusion: FusionResult,
-        review: ReviewResult,
-        intent: IntentResult,
-        fusion_level: str,
-    ) -> FusionResult:
-        from packages.application.viral_video.generator import CopyGenerator
-
-        template = get_template("copy_fusion")
-        system_kwargs = {
-            "fusion_instruction": (
-                "【本次任务：按审核意见修正文案】只修改指出的问题，其他内容尽量原样保留；"
-                "personal_brands 事实信息逐字保留；修正后按原标签格式完整输出。"
-            ),
-            "global_constraints": "",
-            "negative_rules": "",
-        }
-        issue_text = "\n".join(f"- [{i.dimension}/{i.location}] {i.text}" for i in review.issues)
-        suggestion_text = "\n".join(f"- {s}" for s in review.rewrite_suggestions)
-        user = render_user_prompt(
-            template,
-            industry="",
-            target_customer="",
-            marketing_purpose="",
-            duration=fusion.estimated_duration or 15,
-            image_analysis="（沿用原图片分析）",
-            intent_result=self._intent_text(intent),
-        )
-        user = (
-            f"{user}\n\n原文案：\n{self._fusion_text(fusion)}\n\n"
-            f"审核发现的问题：\n{issue_text}\n\n修改建议：\n{suggestion_text or '（无）'}\n"
-            "请输出修正后的完整文案。"
-        )
-        system = render_system_prompt(template, **system_kwargs)
-        raw = self.client.chat_completion(
-            [
-                {"role": "system", "content": system},
-                {"role": "user", "content": user},
-            ],
-            temperature=0.5,
-            max_tokens=2048,
-            timeout=25,
-        )
-        if not raw:
-            return self._rule_fix(fusion, review)
-        rewritten = CopyGenerator._parse_fusion(CopyGenerator(self.client), raw)
-        if not rewritten.title and not rewritten.script_segments:
-            return self._rule_fix(fusion, review)
-        # 保底：personal_brands 必须保留
-        full = self._fusion_text(rewritten)
-        for brand in intent.personal_brands:
-            if brand.text and brand.text not in full:
-                rewritten.cta = (rewritten.cta + brand.text).strip()
-        return rewritten
-
-    def _rule_fix(self, fusion: FusionResult, review: ReviewResult) -> FusionResult:
-        """LLM 重写不可用时的本地兜底：删除/替换明显违规表述。"""
-        replacements = [
-            (re.compile(r"100\s*%|百分百"), "大部分"),
-            (re.compile(r"绝对(有效|安全|靠谱)"), "比较\\1"),
-            (re.compile(r"包治百病"), "适用多种情况"),
-            (re.compile(r"立刻见效"), "坚持使用会有改善"),
-            (re.compile(r"一喷(就|全|100%)"), "喷上等一会儿可以"),
-            (re.compile(r"家人们谁懂啊|绝绝子|宝子们|yyds", re.IGNORECASE), ""),
-            (re.compile(r"最好|最强|最牛|最便宜"), "很不错"),
-        ]
-
-        def fix(text: str) -> str:
-            for pattern, repl in replacements:
-                text = pattern.sub(repl, text)
-            return text
-
-        fusion.title = fix(fusion.title)
-        fusion.hook = fix(fusion.hook)
-        fusion.cta = fix(fusion.cta)
-        for point in fusion.body_points:
-            point.text = fix(point.text)
-            point.elaboration = fix(point.elaboration)
-        for segment in fusion.script_segments:
-            segment.text = fix(segment.text)
-        fusion.raw = ""
-        return fusion
-
-    # ── 文本工具 ────────────────────────────────────────────────────────
-    @staticmethod
-    def _segments(fusion: FusionResult):
-        yield "title", fusion.title
-        yield "hook", fusion.hook
-        for point in fusion.body_points:
-            yield "body_points", f"{point.text} {point.elaboration}"
-        yield "cta", fusion.cta
-        for segment in fusion.script_segments:
-            yield "script_segments", segment.text
-
-    @staticmethod
-    def _fusion_text(fusion: FusionResult) -> str:
-        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)
-        return "\n".join(p for p in parts if p)
-
-    @staticmethod
-    def _intent_text(intent) -> str:
-        if intent is None:
-            return "无意图信息"
-        parts = [f"意图：{intent.intent_summary}"]
-        parts += [f"核心信息[must_keep={m.must_keep}]：{m.text}" for m in intent.core_messages]
-        parts += [f"事实（{b.category}）：{b.text}" for b in intent.personal_brands]
-        return "\n".join(parts)

packages/application/viral_video/xml_parser.py

@@ -13,6 +13,13 @@ from typing import Optional
 _OPEN_RE = re.compile(r"<(?P<tag>[\w-]+)(?P<attrs>(?:\s(?:[^>]*?\S)?)?)(?P<self>/?)>")
 _CLOSE_RE = re.compile(r"</(?P<tag>[\w-]+)\s*>")
 _ATTR_RE = re.compile(r"""([\w:-]+)\s*=\s*(?:"([^"]*)"|'([^']*)')""")
+_CDATA_RE = re.compile(r"^<!\[CDATA\[(.*)\]\]>$", re.DOTALL)
+
+
+def _strip_cdata(s: str) -> str:
+    """剥离 LLM 可能照抄示例输出的 ``<![CDATA[...]]>`` 包裹层。"""
+    m = _CDATA_RE.match(s.strip())
+    return m.group(1) if m else s
 
 
 def parse_attributes(raw: str) -> dict[str, str]:
@@ -58,6 +65,7 @@ def parse_tags(text: Optional[str]) -> list[dict]:
                 if stack[idx]["tag"] == tag:
                     node = stack[idx]
                     node["text"] = unescape(text[node["_start"] : token.start()].strip())
+                    node["text"] = _strip_cdata(node["text"])
                     node.pop("_start", None)
                     del stack[idx:]
                     break
@@ -65,6 +73,7 @@ def parse_tags(text: Optional[str]) -> list[dict]:
     for node in stack:
         if "_start" in node:
             node["text"] = unescape(text[node["_start"] :].strip())
+            node["text"] = _strip_cdata(node["text"])
             node.pop("_start", None)
     return results
 

packages/config/base.py

@@ -173,6 +173,73 @@ class SharedSettings(BaseSettings):
     # 判断 Worker 可用的心跳新鲜度窗口（秒）—— last_heartbeat_at 在窗口内视为在线
     gpu_worker_stale_seconds: int = 300
 
+    # ── Ditto 蚂蚁数字人口型 API（#2076）─────────────────────────────────
+    # 是否优先使用 Ditto（蚂蚁数字人，替代 MuseTalk）。开关开启且 base_url 配置
+    # 非空时，对口型任务优先走 Ditto；失败后回退 MuseTalk/MediaKit。
+    use_ditto_lipsync: bool = Field(
+        default=False,
+        validation_alias=AliasChoices("USE_DITTO_LIPSYNC", "use_ditto_lipsync"),
+    )
+    # Ditto FastAPI 内网地址（Tailscale），如 http://100.x.x.x:8000
+    ditto_api_base_url: str = Field(
+        default="",
+        validation_alias=AliasChoices("DITTO_API_BASE_URL", "ditto_api_base_url"),
+    )
+    # 默认人物模板视频 URL（兜底用：用户未上传视频时使用，或视频预处理失败时回退）。
+    # 正面 5-10 秒、光线均匀、半身 1080x1920 竖版；正常流程下 Ditto 优先使用用户上传的 video_url。
+    ditto_default_video_url: str = Field(
+        default="",
+        validation_alias=AliasChoices("DITTO_DEFAULT_VIDEO_URL", "ditto_default_video_url"),
+    )
+    # 429 GPU 繁忙时指数退避最大重试次数
+    ditto_max_retries: int = Field(
+        default=3,
+        validation_alias=AliasChoices("DITTO_MAX_RETRIES", "ditto_max_retries"),
+    )
+    # Ditto 单次请求 read 超时（秒）：数字人半身视频推理通常 30-120s（RTF≈2.8，40s音频约112s）
+    # connect 超时固定 10s（代码硬编码，网络不通快速失败）
+    ditto_request_timeout: int = Field(
+        default=120,
+        validation_alias=AliasChoices("DITTO_REQUEST_TIMEOUT", "ditto_request_timeout"),
+    )
+    # Ditto 句间过渡帧数（平滑表情/口型切换）
+    ditto_blend_frames: int = Field(
+        default=12,
+        validation_alias=AliasChoices("DITTO_BLEND_FRAMES", "ditto_blend_frames"),
+    )
+
+    # ── Ditto LLM 情绪分析（emo_timeline）──────────────────────────────
+    # 总开关；关闭或 LLM 失败时走 GPU 端关键词匹配兜底
+    ditto_emotion_enabled: bool = Field(
+        default=False,
+        validation_alias=AliasChoices("DITTO_EMOTION_ENABLED", "ditto_emotion_enabled"),
+    )
+    ditto_emotion_model: str = Field(
+        default="doubao-seed-2-1-lite-250915",
+        validation_alias=AliasChoices("DITTO_EMOTION_MODEL", "ditto_emotion_model"),
+    )
+    ditto_emotion_temperature: float = Field(
+        default=0.8,
+        validation_alias=AliasChoices("DITTO_EMOTION_TEMPERATURE", "ditto_emotion_temperature"),
+    )
+    ditto_emotion_timeout: int = Field(
+        default=10,
+        validation_alias=AliasChoices("DITTO_EMOTION_TIMEOUT", "ditto_emotion_timeout"),
+    )
+    ditto_emotion_max_tokens: int = Field(
+        default=1024,
+        validation_alias=AliasChoices("DITTO_EMOTION_MAX_TOKENS", "ditto_emotion_max_tokens"),
+    )
+    ditto_emotion_cache_size: int = Field(
+        default=500,
+        validation_alias=AliasChoices("DITTO_EMOTION_CACHE_SIZE", "ditto_emotion_cache_size"),
+    )
+    # 提示词模板：必须包含 {文案} 占位符；后台可通过环境变量覆盖
+    ditto_emotion_prompt: str = Field(
+        default="",
+        validation_alias=AliasChoices("DITTO_EMOTION_PROMPT", "ditto_emotion_prompt"),
+    )
+
     # ── P4000 NVENC 硬件编码 ────────────────────────────────────────────
     # GPU 编码总开关；关闭或 endpoint 为空时始终走本机 CPU libx264
     enable_gpu_encode: bool = Field(

packages/domain/asset_atom_clip.py

@@ -8,7 +8,12 @@ from __future__ import annotations
 
 import uuid
 from dataclasses import dataclass, field
-from datetime import UTC, datetime
+from datetime import datetime, timezone
+
+try:
+    from datetime import UTC
+except ImportError:
+    UTC = timezone.utc
 
 
 @dataclass

packages/domain/system_setting.py

@@ -0,0 +1,134 @@
+"""系统配置领域实体 — #2246.
+
+承载 setting_type（bool/int/float/string/json）及 setting_value 的
+序列化/反序列化规则；与 DB、框架无关。
+"""
+
+from __future__ import annotations
+
+import json
+from dataclasses import dataclass, field
+from datetime import UTC, datetime
+from typing import Any
+from uuid import uuid4
+
+SETTING_TYPE_BOOL = "bool"
+SETTING_TYPE_INT = "int"
+SETTING_TYPE_FLOAT = "float"
+SETTING_TYPE_STRING = "string"
+SETTING_TYPE_JSON = "json"
+
+VALID_SETTING_TYPES = {
+    SETTING_TYPE_BOOL,
+    SETTING_TYPE_INT,
+    SETTING_TYPE_FLOAT,
+    SETTING_TYPE_STRING,
+    SETTING_TYPE_JSON,
+}
+
+
+class SystemSettingError(ValueError):
+    """系统配置类型或序列化错误."""
+
+
+def infer_setting_type(value: Any) -> str:
+    """根据 Python 值推断 setting_type（bool 必须先于 int 判断）."""
+    if isinstance(value, bool):
+        return SETTING_TYPE_BOOL
+    if isinstance(value, int):
+        return SETTING_TYPE_INT
+    if isinstance(value, float):
+        return SETTING_TYPE_FLOAT
+    if isinstance(value, str):
+        return SETTING_TYPE_STRING
+    return SETTING_TYPE_JSON
+
+
+def serialize_setting_value(value: Any, setting_type: str) -> str:
+    """把 Python 值按 setting_type 序列化为可入库的字符串."""
+    if setting_type == SETTING_TYPE_BOOL:
+        if not isinstance(value, bool):
+            raise SystemSettingError(f"bool 配置值必须是布尔类型，收到 {value!r}")
+        return "true" if value else "false"
+    if setting_type == SETTING_TYPE_INT:
+        if isinstance(value, bool) or not isinstance(value, int):
+            raise SystemSettingError(f"int 配置值必须是整数，收到 {value!r}")
+        return str(value)
+    if setting_type == SETTING_TYPE_FLOAT:
+        if isinstance(value, bool):
+            raise SystemSettingError(f"float 配置值不能是布尔类型，收到 {value!r}")
+        try:
+            return repr(float(value))
+        except (TypeError, ValueError) as exc:
+            raise SystemSettingError(f"float 配置值非法：{value!r}") from exc
+    if setting_type == SETTING_TYPE_STRING:
+        if not isinstance(value, str):
+            raise SystemSettingError(f"string 配置值必须是字符串，收到 {value!r}")
+        return value
+    if setting_type == SETTING_TYPE_JSON:
+        try:
+            return json.dumps(value, ensure_ascii=False)
+        except (TypeError, ValueError) as exc:
+            raise SystemSettingError(f"json 配置值无法序列化：{value!r}") from exc
+    raise SystemSettingError(f"未知 setting_type: {setting_type}")
+
+
+def deserialize_setting_value(raw: str | None, setting_type: str) -> Any:
+    """把入库字符串按 setting_type 反序列化为 Python 值."""
+    if raw is None:
+        return None
+    if setting_type == SETTING_TYPE_BOOL:
+        return str(raw).strip().lower() in {"1", "true", "yes", "on"}
+    if setting_type == SETTING_TYPE_INT:
+        try:
+            return int(str(raw).strip())
+        except ValueError as exc:
+            raise SystemSettingError(f"int 配置值损坏：{raw!r}") from exc
+    if setting_type == SETTING_TYPE_FLOAT:
+        try:
+            return float(str(raw).strip())
+        except ValueError as exc:
+            raise SystemSettingError(f"float 配置值损坏：{raw!r}") from exc
+    if setting_type == SETTING_TYPE_STRING:
+        return raw
+    if setting_type == SETTING_TYPE_JSON:
+        try:
+            return json.loads(raw)
+        except (TypeError, ValueError) as exc:
+            raise SystemSettingError(f"json 配置值损坏：{raw!r}") from exc
+    raise SystemSettingError(f"未知 setting_type: {setting_type}")
+
+
+@dataclass(slots=True)
+class SystemSetting:
+    """系统配置领域实体."""
+
+    setting_key: str
+    setting_value: str | None = None
+    setting_type: str = SETTING_TYPE_STRING
+    description: str = ""
+    is_public: bool = False
+    category: str = "general"
+    updated_by: str | None = None
+    id: str = field(default_factory=lambda: uuid4().hex)
+    created_at: datetime = field(default_factory=lambda: datetime.now(UTC))
+    updated_at: datetime = field(default_factory=lambda: datetime.now(UTC))
+
+    def get_typed_value(self) -> Any:
+        return deserialize_setting_value(self.setting_value, self.setting_type)
+
+    @classmethod
+    def from_value(
+        cls,
+        setting_key: str,
+        value: Any,
+        setting_type: str | None = None,
+        **kwargs: Any,
+    ) -> "SystemSetting":
+        st = setting_type or infer_setting_type(value)
+        return cls(
+            setting_key=setting_key,
+            setting_value=serialize_setting_value(value, st),
+            setting_type=st,
+            **kwargs,
+        )

packages/domain/viral_video.py

@@ -107,6 +107,7 @@ class ViralVideoJob:
     # v1.5.1 音频/视频参数
     voice_id: str = ""
     voice_source: str = ""
+    language: str = "zh-CN"
     video_ratio: str = "9:16"
     video_model: str = ""
     # v1.4+ 产物
@@ -243,6 +244,13 @@ class ViralVideoJob:
         self.heartbeat_at = _now
         self.status = ViralVideoStatus.RUNNING
         self.updated_at = _now
+        # 清空上一次渲染留下的错误/产物（参考 resume_from_image_analyzed 对齐）
+        self.error_msg = ""
+        self.result_video_url = ""
+        self.current_stage = ""
+        self.phase_message = ""
+        self.completed_at = None
+        self.credits_cost = 0.0
 
     def resume_from_confirm(self) -> None:
         if self.status != ViralVideoStatus.WAIT_USER_CONFIRM:
@@ -259,9 +267,12 @@ class ViralVideoJob:
         self.completed_at = datetime.now(timezone.utc)
         self.updated_at = datetime.now(timezone.utc)
 
-    def mark_failed(self, error_msg: str) -> None:
+    def mark_failed(self, error_msg: str, stage: str = "") -> None:
         self.status = ViralVideoStatus.FAILED
         self.error_msg = error_msg
+        # 失败时持久化失败阶段，避免前端错误归因
+        if stage:
+            self.current_stage = stage
         self.completed_at = datetime.now(timezone.utc)
         self.updated_at = datetime.now(timezone.utc)
 

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:
@@ -363,6 +363,111 @@ class DoubaoClient:
         logger.error("豆包API调用最终失败: elapsed=%.1fs err=%s", time.time() - _t0, last_error)
         return None
 
+    def chat_completion_stream(
+        self,
+        messages: list[dict[str, str]],
+        temperature: float = 0.7,
+        max_tokens: int | None = None,
+        model: str | None = None,
+        timeout: int | None = None,
+        **kwargs,
+    ):
+        """流式调用 Chat Completion 接口（SSE），逐块 yield delta 文本。
+
+        Yields:
+            str: 增量文本片段（delta.content）；全部结束后 StopIteration。
+            失败时 yield 空并返回（由调用方决定是否降级到同步调用）。
+
+        注意：
+        - 流式不做 finish_reason=length 自动扩容（流式难以拼接重试）；
+          如果调用方需要 length 截断处理，建议自行 fallback 到同步 chat_completion。
+        - 重试只在连接建立阶段（首包之前）有效；一旦开始 yield，错误直接抛出。
+        """
+        if not self.is_available:
+            return
+
+        import json as _json
+
+        url = f"{self.base_url}/chat/completions"
+        headers = {
+            "Authorization": f"Bearer {self.api_key}",
+            "Content-Type": "application/json",
+            "Accept": "text/event-stream",
+        }
+        effective_max_tokens = max_tokens if max_tokens is not None else (self.max_tokens or 1024)
+        payload: dict[str, Any] = {
+            "model": model or self.model,
+            "messages": messages,
+            "temperature": temperature,
+            "max_tokens": effective_max_tokens,
+            "stream": True,
+        }
+        if self.extra_params:
+            payload.update(self.extra_params)
+        if kwargs:
+            payload.update(kwargs)
+
+        _t0 = time.time()
+        for attempt in range(self.max_retries + 1):
+            try:
+                _req_timeout = httpx.Timeout(connect=10.0, read=120.0, write=30.0, pool=10.0)
+                if timeout:
+                    _req_timeout = httpx.Timeout(connect=10.0, read=max(int(timeout), 30), write=30.0, pool=10.0)
+                with httpx.stream(
+                    "POST",
+                    url,
+                    headers=headers,
+                    json=payload,
+                    timeout=_req_timeout,
+                ) as resp:
+                    resp.raise_for_status()
+                    for line in resp.iter_lines():
+                        if not line:
+                            continue
+                        line = line.strip()
+                        if not line.startswith("data:"):
+                            continue
+                        data_str = line[5:].strip()
+                        if data_str == "[DONE]":
+                            break
+                        try:
+                            chunk = _json.loads(data_str)
+                        except (_json.JSONDecodeError, ValueError):
+                            continue
+                        choices = chunk.get("choices") or []
+                        if not choices:
+                            continue
+                        delta = choices[0].get("delta") or {}
+                        content_piece = delta.get("content") or ""
+                        if content_piece:
+                            yield content_piece
+                        finish_reason = choices[0].get("finish_reason")
+                        if finish_reason:
+                            self.last_finish_reason = finish_reason
+                _elapsed = time.time() - _t0
+                logger.info(
+                    "[doubao] chat_completion_stream 完成 model=%s elapsed=%.1fs attempt=%d",
+                    payload.get("model"),
+                    _elapsed,
+                    attempt + 1,
+                )
+                return
+            except Exception as e:
+                if attempt < self.max_retries:
+                    wait = 0.5 * (2**attempt)
+                    logger.warning(
+                        "豆包流式API失败，%.1fs后重试 (%d/%d, elapsed=%.1fs): %s",
+                        wait,
+                        attempt + 1,
+                        self.max_retries + 1,
+                        time.time() - _t0,
+                        e,
+                    )
+                    time.sleep(wait)
+                    continue
+                logger.error("豆包流式API最终失败: elapsed=%.1fs err=%s", time.time() - _t0, e)
+                return
+
     def vision_completion(
         self,
         messages: list[dict],
@@ -636,6 +741,10 @@ class DoubaoClient:
                     resolution=resolution,
                     output_dir=output_dir,
                     model=video_model,
+                    generate_audio=bool(generate_audio),
+                    reference_images=reference_images,
+                    reference_audios=reference_audios,
+                    reference_videos=reference_videos,
                 )
                 if not result and hasattr(ds, "last_video_error") and ds.last_video_error:
                     self.last_video_error = dict(ds.last_video_error)

packages/shared/celery_queues.py

@@ -51,6 +51,8 @@ task_routes = {
     "ai_avatar_render.execute": {"queue": QUEUE_GENERATION},
     # GPU MuseTalk 口型同步（用户等成片，链路子任务全部走 generation 避免跨队列阻塞）
     "lipsync_gpu_process_async": {"queue": QUEUE_GENERATION},
+    # #2076 Ditto 蚂蚁数字人口型同步（走 generation 队列，避免跨队列阻塞）
+    "lipsync_ditto_process_async": {"queue": QUEUE_GENERATION},
     "lipsync_tts.synthesize_and_submit": {"queue": QUEUE_GENERATION},
     "lipsync_tts.poll_mediakit_status": {"queue": QUEUE_GENERATION},
     "lipsync_tts.persist_output_video": {"queue": QUEUE_GENERATION},

packages/shared/dashscope_client.py

@@ -67,10 +67,12 @@ def _classify_dashscope_error(status_code: int, body: str, task_msg: str = "") -
         return "invalid_param", f"Wan 3.0 参数错误：{msg_in_body or body[:200]}"
     if status_code == 0:
         return "network_error", "Wan 3.0 服务连接失败（网络超时），请稍后重试。"
-    # 任务内失败
+    # 任务内失败（异步轮询结果）：此时 HTTP 200 只是查询成功，不代表任务成功，提示文案不要带 HTTP 200 以免误导
     if task_msg and any(kw in task_msg.lower() for kw in ("portrait", "真人", "人脸", "violation", "blocked")):
         return "portrait_intercept", "Wan 3.0 视频内容被安全策略拦截，请调整文案或参考图后重试。"
     detail = msg_in_body or body[:200]
+    if task_msg and status_code == 200:
+        return "task_failed", f"Wan 3.0 视频生成失败：{detail}"
     return "unknown", f"Wan 3.0 视频生成失败（HTTP {status_code}）：{detail}"
 
 
@@ -115,10 +117,20 @@ class DashScopeClient:
         watermark: bool = False,
         output_dir: str | None = None,
         model: str = "wan3.0-video",
+        generate_audio: bool = True,
+        reference_images: list[str] | None = None,
+        reference_audios: list[str] | None = None,
+        reference_videos: list[str] | None = None,
     ) -> dict | None:
-        """调用 DashScope 异步视频合成接口，轮询完成后下载到本地。
+        """调用 DashScope Wan 3.0 异步视频合成接口，轮询完成后下载到本地。
 
-        返回 {"video_path": str, "usage": dict | None}；失败返回 None，错误详情写入 self.last_video_error。
+        官方协议（input.media 数组 + parameters.audio）：
+        - 仅 1 张图且无其它参考 -> type=first_frame（首帧模式，严格从该帧起）。
+        - 有参考音频 / 多张图 -> 图片全部走 type=reference_image（全能参考模式，
+          可与 reference_audio 共存）；prompt 用"图1/图2/音频1"按 media 顺序引用。
+        - parameters.audio 控制输出是否含音轨；参考音频通过 media 传入。
+
+        返回 {"video_path": str, "usage": dict | None}；失败返回 None，错误写入 self.last_video_error。
         """
         self.last_video_error = {}
         if not self.is_available:
@@ -130,7 +142,7 @@ class DashScopeClient:
             return None
 
         # DashScope 分辨率参数：720P / 1080P / 480P（大写 P）
-        res_upper = (resolution or "720p").upper().replace("P", "P")
+        res_upper = (resolution or "720p").upper()
         if res_upper == "480P":
             ds_res = "480P"
         elif res_upper == "1080P":
@@ -138,16 +150,35 @@ class DashScopeClient:
         else:
             ds_res = "720P"
 
-        # 构造 input+parameters
+        # ── 构造官方 media 数组 ────────────────────────────────────────
+        ref_imgs = [u for u in (reference_images or [])[:10] if u]
+        ref_auds = [u for u in (reference_audios or [])[:5] if u]
+        ref_vids = [u for u in (reference_videos or [])[:5] if u]
+
+        media: list[dict[str, Any]] = []
+        all_imgs = ([image_url] if image_url else []) + [u for u in ref_imgs if u != image_url]
+        use_first_frame = bool(image_url) and len(all_imgs) == 1 and not (ref_auds or ref_vids)
+        if use_first_frame:
+            media.append({"type": "first_frame", "url": image_url})
+        else:
+            for u in all_imgs:
+                media.append({"type": "reference_image", "url": u})
+            for u in ref_vids:
+                media.append({"type": "reference_video", "url": u})
+            for u in ref_auds:
+                media.append({"type": "reference_audio", "url": u})
+
+        # ── input + parameters ─────────────────────────────────────────
         input_obj: dict[str, Any] = {"prompt": prompt.strip()}
-        if image_url:
-            input_obj["img_url"] = image_url
+        if media:
+            input_obj["media"] = media
+
         params: dict[str, Any] = {
             "resolution": ds_res,
-            "duration": str(float(duration)),
+            "duration": int(duration),
             "watermark": bool(watermark),
+            "audio": bool(generate_audio),
         }
-        # 比例透传：Wan 支持 "9:16" / "16:9" / "1:1" 等
         if ratio and ratio != "adaptive":
             params["aspect_ratio"] = ratio
 
@@ -163,14 +194,15 @@ class DashScopeClient:
         }
         create_url = f"{self.base_url}/services/aigc/video-generation/video-synthesis"
         logger.info(
-            "[dashscope] 创建任务: model=%s dur=%ds ratio=%s res=%s img=%s",
+            "[dashscope] 创建任务: model=%s dur=%ds ratio=%s res=%s media=%d audio=%s",
             model,
             duration,
             ratio,
             ds_res,
-            bool(image_url),
+            len(media),
+            generate_audio,
         )
-        logger.info("[dashscope] 创建任务 payload: model=%s params=%s", model, params)
+        logger.info("[dashscope] media types: %s", [m["type"] for m in media])
 
         # 创建任务
         task_id: str | None = None
@@ -196,27 +228,14 @@ class DashScopeClient:
                 if tid:
                     task_id = tid
                     break
-                # 部分情况下 code != 错误
                 code = data.get("code")
-                if code and code != "":
-                    err_code, user_msg = _classify_dashscope_error(400, body_text, str(code))
+                if code:
+                    logger.error("[dashscope] 创建任务返回 code=%s body=%s", code, body_text)
+                    err_code, user_msg = _classify_dashscope_error(sc, body_text)
                     self._set_error(err_code, user_msg, sc, body_text, model=model)
                     return None
-                else:
-                    self._set_error("unknown", "Wan 3.0 响应格式异常，未返回任务ID", sc, str(data)[:500], model=model)
-                    return None
-            except _HTTP_NETWORK_ERRORS as ne:
-                last_sc = 0
-                last_body = f"network error: {ne}"
-                logger.warning(
-                    "[dashscope] 网络异常 %s，重试 %d/%d", type(ne).__name__, attempt + 1, self.max_retries + 1
-                )
-                if attempt < self.max_retries:
-                    time.sleep(0.5 * (2**attempt))
-                    continue
-                self._set_error("network_error", "Wan 3.0 服务连接失败（网络超时），请稍后重试。", 0, str(ne))
-                return None
             except Exception as _e:
+                logger.warning("[dashscope] 创建任务异常(attempt=%d): %s", attempt, _e)
                 if attempt < self.max_retries:
                     time.sleep(0.5 * (2**attempt))
                     continue
@@ -258,7 +277,6 @@ class DashScopeClient:
                     video_url = out.get("video_url") or ""
                     usage = d.get("usage")
                     if not video_url:
-                        # 结果在 results 数组
                         results = out.get("results") or []
                         if results and isinstance(results, list):
                             video_url = results[0].get("url") or results[0].get("video_url")
@@ -284,7 +302,6 @@ class DashScopeClient:
                     logger.warning("[dashscope] 任务 %s 被取消", task_id)
                     self._set_error("unknown", "Wan 3.0 任务被取消。", 200, "task cancelled", task_id=task_id)
                     return None
-                # PENDING / RUNNING / SUSPENDED → 继续轮询
                 if poll_count % 5 == 0:
                     logger.info("[dashscope] 轮询中 task=%s status=%s polls=%d", task_id, task_status, poll_count)
             except Exception as e:

packages/shared/storage.py

@@ -334,7 +334,7 @@ class SharedStorageService(StoragePort):
 
         storage_key = self.normalize_storage_key(storage_key_or_url)
         try:
-            signed = sign_bucket.sign_url("GET", storage_key, expires_seconds)
+            signed = sign_bucket.sign_url("GET", storage_key, expires_seconds, slash_safe=True)
             logger.info(
                 "signed URL generated for key=%s prefix=%s",
                 storage_key[:80],

scripts/ci/ci_env.sh

@@ -12,3 +12,22 @@ export CI_LOCAL_PG_PORT="${CI_LOCAL_PG_PORT:-5432}"
 
 # === 默认数据库名 ===
 export CI_DEFAULT_DB="${CI_DEFAULT_DB:-xiaoxia_saas}"
+
+# === Python 版本保障：本项目依赖 datetime.UTC，需要 Python >= 3.11 ===
+_ensure_python311() {
+  if python3 -c "import sys; assert sys.version_info >= (3, 11)" 2>/dev/null; then
+    return 0
+  fi
+  for cand in python3.12 python3.11 /opt/python3.12/bin/python3; do
+    if command -v "$cand" >/dev/null 2>&1 && "$cand" -c "import sys; assert sys.version_info >= (3, 11)" 2>/dev/null; then
+      _d="$(dirname "$(command -v "$cand")")"
+      export PATH="$_d:$PATH"
+      hash -r
+      echo "✅ ci_env: 切换到 $cand ($("$cand" -c 'import sys; print(sys.version.split()[0])'))"
+      return 0
+    fi
+  done
+  echo "❌ ci_env: 未找到 Python >= 3.11（datetime.UTC 需要），请安装 Python 3.11/3.12" >&2
+  return 1
+}
+_ensure_python311

scripts/ci/mypy_check.sh

@@ -4,8 +4,28 @@
 
 set -e
 
+SCRIPT_DIR="$(dirname "${BASH_SOURCE[0]}")"
+# shellcheck source=ci_env.sh
+source "${SCRIPT_DIR}/ci_env.sh"
+
+
 echo "=== Installing mypy ==="
-python3 -m pip install -q mypy
+# pip源容错: 默认源(阿里云)缺文件时fallback到清华/官方源(2026-10-10 librt-0.6.0 metadata 404)
+MYPY_SPEC="mypy<1.19"
+INSTALL_OK=0
+python3 -m pip install -q "$MYPY_SPEC" && INSTALL_OK=1 || true
+if [ "$INSTALL_OK" != "1" ]; then
+  echo "WARN: default pip index failed, retry tsinghua mirror..."
+  python3 -m pip install -q -i https://pypi.tuna.tsinghua.edu.cn/simple "$MYPY_SPEC" && INSTALL_OK=1 || true
+fi
+if [ "$INSTALL_OK" != "1" ]; then
+  echo "WARN: tsinghua mirror failed, retry pypi.org..."
+  python3 -m pip install -q -i https://pypi.org/simple "$MYPY_SPEC" && INSTALL_OK=1 || true
+fi
+if [ "$INSTALL_OK" != "1" ]; then
+  echo "ERROR: pip install mypy failed on all indexes"
+  exit 1
+fi
 mypy --version
 echo ""
 echo "=== Running mypy type check (hard gate mode) ==="

scripts/ci/run_unit_tests.sh

@@ -7,9 +7,29 @@ JOB_NAME="${1:-Unit Tests}"
 
 echo "=== CI Unit Tests 开始 ==="
 
-# --- 依赖缓存检查 ---
+# --- Python 版本选择（必须 >= 3.11，代码使用 datetime.UTC）---
+if ! python3 -c "import sys; assert sys.version_info >= (3, 11)" 2>/dev/null; then
+  for cand in python3.12 python3.11 /opt/python3.12/bin/python3; do
+    if command -v "$cand" >/dev/null 2>&1 && "$cand" -c "import sys; assert sys.version_info >= (3, 11)" 2>/dev/null; then
+      PY3_DIR="$(dirname "$(command -v "$cand")")"
+      export PATH="$PY3_DIR:$PATH"
+      hash -r
+      echo "✅ python3 版本过低，改用 $cand ($("$cand" -c "import sys; print(sys.version.split()[0])"))"
+      break
+    fi
+  done
+  if ! python3 -c "import sys; assert sys.version_info >= (3, 11)" 2>/dev/null; then
+    echo "❌ 未找到 Python >= 3.11，本项目要求 Python 3.11+（使用 datetime.UTC）" >&2
+    exit 1
+  fi
+fi
+
+PYVER=$(python3 -c 'import sys; print(f"{sys.version_info.major}.{sys.version_info.minor}")')
+echo "使用 Python 版本: $(python3 --version)"
+
+# --- 依赖缓存检查（按 Python 版本区分缓存文件，避免跨版本复用）---
 # 如果 requirements 文件未变化且依赖已安装，跳过 pip install（持久 runner 优化）
-REQ_HASH_FILE="/tmp/.ci_unit_tests_req_hash"
+REQ_HASH_FILE="/tmp/.ci_unit_tests_req_hash_py${PYVER}"
 CURRENT_REQ_HASH=""
 if [ -f requirements-base.txt ] && [ -f requirements.txt ] && [ -f requirements-dev.txt ]; then
   CURRENT_REQ_HASH=$(cat requirements-base.txt requirements.txt requirements-dev.txt | md5sum | cut -d' ' -f1)

scripts/ci/validate_code_quality.sh

@@ -77,6 +77,11 @@ set +e
 bandit -r apps packages -q -ll
 BANDIT_EXIT=$?
 set -e
+
+SCRIPT_DIR="$(dirname "${BASH_SOURCE[0]}")"
+# shellcheck source=ci_env.sh
+source "${SCRIPT_DIR}/ci_env.sh"
+
 if [ "$BANDIT_EXIT" -ne 0 ]; then
   echo "⚠️  Bandit found security issues (advisory mode - not blocking CI)"
 else

scripts/ci_staging_deploy.sh

@@ -108,8 +108,10 @@ server {
         proxy_set_header X-Real-IP $remote_addr;
         proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
         proxy_set_header X-Forwarded-Proto $scheme;
-        proxy_read_timeout 300s;
-        proxy_send_timeout 300s;
+        proxy_set_header Upgrade $http_upgrade;
+        proxy_set_header Connection "upgrade";
+        proxy_read_timeout 3600s;
+        proxy_send_timeout 3600s;
         proxy_request_buffering off;
     }
 

scripts/seed_viral_video_prompts.py

@@ -1,4 +1,4 @@
-"""爆款视频 5 套 Prompt 模板种子脚本（#2040）。
+"""爆款视频 Prompt 模板种子脚本（#2040，PR1 清理至 3 套：image_analysis/storyboard/review）。
 
 幂等：以 (prompt_type, version) 为唯一键，存在则更新（UPSERT），重复执行结果一致。
 用法：
@@ -72,7 +72,7 @@ def seed(engine) -> int:
 def main() -> int:
     engine = _engine()
     count = seed(engine)
-    print(f"seed 完成：{count} 套模板已写入/更新（image_analysis/intent_parsing/copy_fusion/storyboard/review）")
+    print(f"seed 完成：{count} 套模板已写入/更新（image_analysis/storyboard/review）")
     return 0
 
 

tests/unit/test_ai_client.py

@@ -196,3 +196,159 @@ class TestGetDoubaoClient:
         """返回 DoubaoClient 实例"""
         client = get_doubao_client()
         assert isinstance(client, DoubaoClient)
+
+
+class TestChatCompletionStream:
+    """chat_completion_stream 方法测试"""
+
+    def _fake_sse(self, chunks: list[str]):
+        """构造一个伪装的 httpx stream 响应，按 chunks 逐行返回 SSE。"""
+        import json as _json
+
+        lines = []
+        for piece in chunks:
+            evt = {"choices": [{"delta": {"content": piece}, "finish_reason": None}]}
+            lines.append("data: " + _json.dumps(evt, ensure_ascii=False))
+        lines.append("data: " + _json.dumps({"choices": [{"delta": {}, "finish_reason": "stop"}]}))
+        lines.append("data: [DONE]")
+
+        m = MagicMock()
+        m.status_code = 200
+        m.__enter__ = MagicMock(return_value=m)
+        m.__exit__ = MagicMock(return_value=False)
+        m.raise_for_status = MagicMock()
+        m.iter_lines = MagicMock(return_value=iter(lines))
+        return m
+
+    def test_stream_yields_delta_content(self, client_with_key):
+        """流式调用逐段 yield delta.content"""
+        import json as _json
+
+        fake = self._fake_sse(["你好", "，", "世界"])
+        with patch("packages.shared.ai_client.httpx.stream", return_value=fake) as mock_stream:
+            out = list(
+                client_with_key.chat_completion_stream(
+                    messages=[{"role": "user", "content": "hi"}], max_tokens=100, timeout=30
+                )
+            )
+        assert out == ["你好", "，", "世界"]
+        call_kwargs = mock_stream.call_args[1]
+        assert call_kwargs["json"]["stream"] is True
+        assert call_kwargs["json"]["max_tokens"] == 100
+
+    def test_stream_unavailable_returns_empty(self, client_without_key):
+        """不可用时返回空生成器（不调用 httpx.stream）"""
+        with patch("packages.shared.ai_client.httpx.stream") as mock_stream:
+            out = list(client_without_key.chat_completion_stream(messages=[{"role": "user", "content": "hi"}]))
+        assert out == []
+        mock_stream.assert_not_called()
+
+    def test_stream_handles_done_marker(self, client_with_key):
+        """遇到 [DONE] 正确终止，不把它当内容"""
+        import json as _json
+
+        lines = [
+            "data: " + _json.dumps({"choices": [{"delta": {"content": "A"}}]}),
+            "data: [DONE]",
+            "data: " + _json.dumps({"choices": [{"delta": {"content": "NEVER"}}]}),  # 应被忽略
+        ]
+        m = MagicMock()
+        m.status_code = 200
+        m.__enter__ = MagicMock(return_value=m)
+        m.__exit__ = MagicMock(return_value=False)
+        m.raise_for_status = MagicMock()
+        m.iter_lines = MagicMock(return_value=iter(lines))
+        with patch("packages.shared.ai_client.httpx.stream", return_value=m):
+            out = list(
+                client_with_key.chat_completion_stream(
+                    messages=[{"role": "user", "content": "hi"}], max_tokens=50, timeout=30
+                )
+            )
+        assert out == ["A"]
+
+    def test_stream_skips_empty_delta(self, client_with_key):
+        """空 delta（role 等 metadata）不应产出内容"""
+        import json as _json
+
+        lines = [
+            "data: " + _json.dumps({"choices": [{"delta": {"role": "assistant"}}]}),
+            "data: " + _json.dumps({"choices": [{"delta": {"content": "hi"}}]}),
+            "data: " + _json.dumps({"choices": [{"delta": {}}]}),
+            "data: [DONE]",
+        ]
+        m = MagicMock()
+        m.status_code = 200
+        m.__enter__ = MagicMock(return_value=m)
+        m.__exit__ = MagicMock(return_value=False)
+        m.raise_for_status = MagicMock()
+        m.iter_lines = MagicMock(return_value=iter(lines))
+        with patch("packages.shared.ai_client.httpx.stream", return_value=m):
+            out = list(
+                client_with_key.chat_completion_stream(
+                    messages=[{"role": "user", "content": "x"}], max_tokens=50, timeout=30
+                )
+            )
+        assert out == ["hi"]
+
+    def test_stream_invalid_json_lines_skipped(self, client_with_key):
+        """SSE 行里脏数据/非 JSON 不应中断流"""
+        import json as _json
+
+        lines = [
+            "data: " + _json.dumps({"choices": [{"delta": {"content": "ok"}}]}),
+            "data: not-a-json",
+            ":comment line",
+            "",
+            "event: ping",
+            "data: " + _json.dumps({"choices": [{"delta": {"content": "2"}}]}),
+            "data: [DONE]",
+        ]
+        m = MagicMock()
+        m.status_code = 200
+        m.__enter__ = MagicMock(return_value=m)
+        m.__exit__ = MagicMock(return_value=False)
+        m.raise_for_status = MagicMock()
+        m.iter_lines = MagicMock(return_value=iter(lines))
+        with patch("packages.shared.ai_client.httpx.stream", return_value=m):
+            out = list(
+                client_with_key.chat_completion_stream(
+                    messages=[{"role": "user", "content": "x"}], max_tokens=50, timeout=30
+                )
+            )
+        assert out == ["ok", "2"]
+
+    def test_stream_network_error_yields_empty(self, client_with_key):
+        """网络错误（重试耗尽）yield 空，不抛异常给调用方"""
+        import httpx as _httpx
+
+        # max_retries=2 → 3 次总尝试
+        client = client_with_key
+        client.max_retries = 1  # 只重试 1 次，缩短测试
+        with patch("packages.shared.ai_client.httpx.stream", side_effect=_httpx.ConnectError("boom")):
+            out = list(
+                client.chat_completion_stream(messages=[{"role": "user", "content": "x"}], max_tokens=50, timeout=5)
+            )
+        assert out == []
+
+    def test_stream_records_finish_reason(self, client_with_key):
+        """流式结束后 last_finish_reason 被正确记录"""
+        import json as _json
+
+        lines = [
+            "data: " + _json.dumps({"choices": [{"delta": {"content": "x"}, "finish_reason": None}]}),
+            "data: " + _json.dumps({"choices": [{"delta": {}, "finish_reason": "stop"}]}),
+            "data: [DONE]",
+        ]
+        m = MagicMock()
+        m.status_code = 200
+        m.__enter__ = MagicMock(return_value=m)
+        m.__exit__ = MagicMock(return_value=False)
+        m.raise_for_status = MagicMock()
+        m.iter_lines = MagicMock(return_value=iter(lines))
+        with patch("packages.shared.ai_client.httpx.stream", return_value=m):
+            list(
+                client_with_key.chat_completion_stream(
+                    messages=[{"role": "user", "content": "x"}], max_tokens=50, timeout=30
+                )
+            )
+        assert client_with_key.last_finish_reason == "stop"

tests/unit/test_ai_router.py

@@ -1,209 +1,210 @@
-"""AI Router 单元测试 — 23 cases covering routing/cache/fallback/client construction."""
+"""AI Router 单元测试 — routing/cache/fallback/client construction.
+
+本文件只做*用例级* mock：通过 autouse fixture 在每个用例内 patch
+``packages.shared.config.get_shared_settings`` / ``packages.shared.ai_router.get_shared_settings``
+并在退出时自动恢复，绝不在模块顶层替换 ``sys.modules``，因此不会污染同进程的
+其他测试模块（如 test_ai_client.py）。
+
+在 Python 3.12 且依赖齐全的 CI 环境中，直接 import 真实模块即可；Redis / DB
+会话通过 patch 隔离。
+"""
 
 from __future__ import annotations
 
-import sys
 import unittest
-from dataclasses import dataclass
-from typing import Optional
 from unittest.mock import MagicMock, patch
 
-# ── Pre-mock heavy import chain to avoid pulling in full app ──
-_mock_config = MagicMock()
-_mock_settings = MagicMock()
-_mock_settings.doubao_model = "doubao-seed-2-1-pro-260915"
-_mock_settings.doubao_fast_model = "doubao-seed-2-1-pro-260915"
-_mock_settings.doubao_base_url = "https://ark.cn-beijing.volces.com/api/v3"
-_mock_settings.doubao_api_key = "test-key"
-_mock_settings.doubao_timeout = 45
-_mock_settings.doubao_max_retries = 1
-_mock_settings.doubao_image_model = "doubao-seedream-5-0-flash-260915"
-_mock_settings.doubao_image_timeout = 60
-_mock_settings.doubao_video_model = "doubao-seedance-2-5-260628"
-_mock_settings.doubao_video_timeout = 600
-_mock_settings.dashscope_api_key = "ds-key"
-_mock_settings.cosyvoice_api_key = "cv-key"
-_mock_settings.cosyvoice_base_url = "https://dashscope.aliyuncs.com/api/v1"
-_mock_settings.cosyvoice_model = "cosyvoice-v3-flash"
-_mock_settings.redis_url = "redis://localhost:6379/0"
-_mock_settings.celery_broker_url = "redis://localhost:6379/0"
-_mock_config.get_shared_settings.return_value = _mock_settings
+import pytest
 
-# Prevent the full packages.shared from loading
-for mod_name in list(sys.modules.keys()):
-    if "packages.shared" in mod_name and "ai_router" not in mod_name and "ai_config_version" not in mod_name:
-        pass  # don't remove, just prevent new imports
+from packages.shared import ai_config_version as _config_version_mod
+from packages.shared import ai_router as ai_router_mod
 
-# Direct import of our modules (bypassing __init__.py)
-import importlib.util
-import os
+# ── 统一的假配置（等价于旧文件里的 _mock_settings）──────────────────────────
 
 
-def _load_module_from_file(name, path):
-    spec = importlib.util.spec_from_file_location(name, path)
-    mod = importlib.util.module_from_spec(spec)
-    sys.modules[name] = mod
-    spec.loader.exec_module(mod)
-    return mod
+def _make_mock_settings() -> MagicMock:
+    s = MagicMock()
+    s.doubao_model = "doubao-seed-2-1-pro-260915"
+    s.doubao_fast_model = "doubao-seed-2-1-pro-260915"
+    s.doubao_base_url = "https://ark.cn-beijing.volces.com/api/v3"
+    s.doubao_api_key = "test-key"
+    s.doubao_timeout = 45
+    s.doubao_max_retries = 1
+    s.doubao_image_model = "doubao-seedream-5-0-flash-260915"
+    s.doubao_image_timeout = 60
+    s.doubao_vision_model = "doubao-seed-1-6-vision-250615"
+    s.doubao_video_model = "doubao-seedance-2-5-260628"
+    s.doubao_video_timeout = 600
+    s.dashscope_api_key = "ds-key"
+    s.dashscope_base_url = "https://dashscope.aliyuncs.com/api/v1"
+    s.dashscope_model = "qwen-vl-max"
+    s.cosyvoice_api_key = "cv-key"
+    s.cosyvoice_base_url = "https://dashscope.aliyuncs.com/api/v1"
+    s.cosyvoice_model = "cosyvoice-v3-flash"
+    s.redis_url = "redis://localhost:6379/0"
+    s.celery_broker_url = "redis://localhost:6379/0"
+    return s
 
 
-# Load ai_config_version
-_ai_config_version = _load_module_from_file(
-    "packages.shared.ai_config_version",
-    os.path.join(os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__)))), "packages", "shared", "ai_config_version.py"),
-)
-# Patch get_shared_settings in the loaded module
-_ai_config_version.get_shared_settings = lambda: _mock_settings
+@pytest.fixture(autouse=True)
+def _mock_settings_fixture():
+    """每个用例内 patch 配置来源，退出即恢复，不污染 sys.modules。"""
+    settings = _make_mock_settings()
+    with (
+        patch("packages.shared.config.get_shared_settings", return_value=settings),
+        patch.object(ai_router_mod, "get_shared_settings", return_value=settings),
+    ):
+        yield settings
 
-# Load ai_router - needs packages.shared.config to be available
-sys.modules["packages.shared.config"] = MagicMock()
-sys.modules["packages.shared.config"].get_shared_settings = lambda: _mock_settings
 
-# Mock packages.shared.ai_client to avoid triggering packages.shared.__init__ chain
-# (which fails on Python 3.10 due to datetime.UTC import in packages.domain)
-_mock_ai_client = MagicMock()
+def _capability_row() -> MagicMock:
+    row = MagicMock()
+    row.capability_key = "intent_parsing"
+    row.capability_name = "文案意图解析"
+    row.timeout_seconds = 45
+    row.max_retries = 1
+    row.max_tokens = None
+    row.temperature = None
+    row.concurrency = 2
+    row.extra_params = {}
+    row.is_enabled = True
+    row.pm_id = "model-1"
+    row.pm_name = "豆包"
+    row.pm_provider = "volcengine"
+    row.pm_model_key = "doubao-seed-1-6-250615"
+    row.pm_api_key = "test-key"
+    row.pm_api_base = "https://ark.test.com"
+    row.pm_api_version = None
+    row.pm_status = "active"
+    row.lm_id = None
+    row.fm_id = None
+    return row
 
-class _FakeDoubaoClient:
-    """Fake DoubaoClient for testing - mimics the real interface."""
-    def __init__(self, api_key="", base_url="", model="", timeout=0, max_retries=0,
-                 max_tokens=None, temperature=None, extra_params=None, provider="volcengine"):
-        self.api_key = api_key
-        self.base_url = base_url
-        self.model = model
-        self.timeout = timeout
-        self.max_retries = max_retries
-        self.max_tokens = max_tokens
-        self.temperature = temperature
-        self.extra_params = extra_params or {}
-        self.provider = provider
-        self.vision_model = model
 
-    @property
-    def is_available(self):
-        return bool(self.api_key)
+def _model_config(**overrides):
+    kwargs = dict(
+        id="m1",
+        name="test",
+        provider="volcengine",
+        model_key="test-model",
+        api_key="key",
+        api_base="https://test.com",
+        api_version=None,
+        status="active",
+    )
+    kwargs.update(overrides)
+    return ai_router_mod.ModelConfig(**kwargs)
 
-    def chat_completion(self, messages, **kwargs):
-        return None
 
-    def vision_completion(self, messages, **kwargs):
-        return None
+def _capability_config(**overrides):
+    kwargs = dict(
+        capability_key="test",
+        capability_name="test",
+        primary_model=None,
+        lite_model=None,
+        fallback_model=None,
+        timeout_seconds=30,
+        max_retries=1,
+        max_tokens=None,
+        temperature=None,
+        concurrency=2,
+        extra_params={},
+        is_enabled=True,
+    )
+    kwargs.update(overrides)
+    return ai_router_mod.CapabilityConfig(**kwargs)
 
-_mock_ai_client.DoubaoClient = _FakeDoubaoClient
-sys.modules["packages.shared.ai_client"] = _mock_ai_client
 
-_ai_router = _load_module_from_file(
-    "packages.shared.ai_router",
-    os.path.join(os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__)))), "packages", "shared", "ai_router.py"),
-)
+# ── Redis 版本号机制 ────────────────────────────────────────────────────────
 
 
 class TestAIConfigVersion(unittest.TestCase):
     """Redis 版本号机制测试"""
 
-    @patch.object(_ai_config_version, "_get_redis_client")
+    @patch.object(_config_version_mod, "_get_redis_client")
     def test_bump_version_success(self, mock_redis_fn):
         mock_r = MagicMock()
         mock_r.set.return_value = True
         mock_redis_fn.return_value = mock_r
-        ver = _ai_config_version.bump_version()
+        ver = _config_version_mod.bump_version()
         self.assertTrue(ver)
         self.assertTrue(ver.isdigit())
         mock_r.set.assert_called_once()
 
-    @patch.object(_ai_config_version, "_get_redis_client")
+    @patch.object(_config_version_mod, "_get_redis_client")
     def test_bump_version_redis_unavailable(self, mock_redis_fn):
         mock_redis_fn.return_value = None
-        ver = _ai_config_version.bump_version()
+        ver = _config_version_mod.bump_version()
         self.assertEqual(ver, "")
 
-    @patch.object(_ai_config_version, "_get_redis_client")
+    @patch.object(_config_version_mod, "_get_redis_client")
     def test_get_version_success(self, mock_redis_fn):
         mock_r = MagicMock()
         mock_r.get.return_value = "1234567890"
         mock_redis_fn.return_value = mock_r
-        ver = _ai_config_version.get_version()
+        ver = _config_version_mod.get_version()
         self.assertEqual(ver, "1234567890")
 
-    @patch.object(_ai_config_version, "_get_redis_client")
+    @patch.object(_config_version_mod, "_get_redis_client")
     def test_get_version_redis_down(self, mock_redis_fn):
         mock_redis_fn.return_value = None
-        ver = _ai_config_version.get_version()
+        ver = _config_version_mod.get_version()
         self.assertIsNone(ver)
 
-    @patch.object(_ai_config_version, "_get_redis_client")
+    @patch.object(_config_version_mod, "_get_redis_client")
     def test_get_version_exception(self, mock_redis_fn):
         mock_r = MagicMock()
         mock_r.get.side_effect = Exception("connection refused")
         mock_redis_fn.return_value = mock_r
-        ver = _ai_config_version.get_version()
+        ver = _config_version_mod.get_version()
         self.assertIsNone(ver)
 
 
+# ── AIRouter 路由/缓存/fallback ────────────────────────────────────────────
+
+
 class TestAIRouter(unittest.TestCase):
     """AIRouter 路由/缓存/fallback 测试"""
 
     def setUp(self):
-        self.router = _ai_router.AIRouter()
+        self.router = ai_router_mod.AIRouter()
 
-    @patch.object(_ai_config_version, "get_version", return_value=None)
-    def test_get_capability_db_unavailable(self, mock_ver):
-        with patch.object(_ai_router, "_get_session", return_value=None):
-            cap = self.router.get_capability("intent_parsing")
-            self.assertIsNone(cap)
+    def _freeze_version(self, value=None):
+        """让 get_capability 的版本比对固定，避免走 Redis。"""
+        return patch.object(_config_version_mod, "get_version", return_value=value)
 
-    @patch.object(_ai_config_version, "get_version", return_value=None)
-    def test_get_capability_from_db(self, mock_ver):
+    def test_get_capability_db_unavailable(self):
+        with self._freeze_version(None):
+            with patch.object(ai_router_mod, "_get_session", return_value=None):
+                cap = self.router.get_capability("intent_parsing")
+                self.assertIsNone(cap)
+
+    def test_get_capability_from_db(self):
         mock_session = MagicMock()
-        mock_row = MagicMock()
-        mock_row.capability_key = "intent_parsing"
-        mock_row.capability_name = "文案意图解析"
-        mock_row.timeout_seconds = 45
-        mock_row.max_retries = 1
-        mock_row.max_tokens = None
-        mock_row.temperature = None
-        mock_row.concurrency = 2
-        mock_row.extra_params = {}
-        mock_row.is_enabled = True
-        mock_row.pm_id = "model-1"
-        mock_row.pm_name = "豆包"
-        mock_row.pm_provider = "volcengine"
-        mock_row.pm_model_key = "doubao-seed-1-6-250615"
-        mock_row.pm_api_key = "test-key"
-        mock_row.pm_api_base = "https://ark.test.com"
-        mock_row.pm_api_version = None
-        mock_row.pm_status = "active"
-        mock_row.lm_id = None
-        mock_row.fm_id = None
-        mock_session.execute.return_value.first.return_value = mock_row
+        mock_session.execute.return_value.first.return_value = _capability_row()
+        with self._freeze_version(None):
+            with patch.object(ai_router_mod, "_get_session", return_value=mock_session):
+                cap = self.router.get_capability("intent_parsing")
+                self.assertIsNotNone(cap)
+                self.assertEqual(cap.capability_key, "intent_parsing")
+                self.assertEqual(cap.primary_model.model_key, "doubao-seed-1-6-250615")
 
-        with patch.object(_ai_router, "_get_session", return_value=mock_session):
-            cap = self.router.get_capability("intent_parsing")
-            self.assertIsNotNone(cap)
-            self.assertEqual(cap.capability_key, "intent_parsing")
-            self.assertEqual(cap.primary_model.model_key, "doubao-seed-1-6-250615")
-
-    @patch.object(_ai_config_version, "get_version", side_effect=[None, "v2"])
-    def test_cache_invalidation_on_version_change(self, mock_ver):
-        with patch.object(self.router, "_load_from_db", return_value=None):
-            self.router.get_capability("test_key")
+    def test_cache_invalidation_on_version_change(self):
+        with self._freeze_version(None):
+            with patch.object(self.router, "_load_from_db", return_value=None):
+                self.router.get_capability("test_key")
         self.router._local_ver = "v1"
-        self.assertTrue(self.router._check_version())
+        with patch.object(_config_version_mod, "get_version", return_value="v2"):
+            self.assertTrue(self.router._check_version())
 
-    @patch.object(_ai_config_version, "get_version", return_value="same_ver")
-    def test_cache_hit_same_version(self, mock_ver):
-        model = _ai_router.ModelConfig(
-            id="m1", name="test", provider="volcengine", model_key="test-model",
-            api_key="key", api_base="https://test.com", api_version=None, status="active",
-        )
-        cap = _ai_router.CapabilityConfig(
-            capability_key="test", capability_name="test", primary_model=model,
-            lite_model=None, fallback_model=None, timeout_seconds=30,
-            max_retries=1, max_tokens=None, temperature=None, concurrency=2,
-            extra_params={}, is_enabled=True,
+    def test_cache_hit_same_version(self):
+        cap = _capability_config(
+            primary_model=_model_config(model_key="test-model"),
         )
         self.router._cache["test"] = cap
         self.router._local_ver = "same_ver"
-        result = self.router.get_capability("test")
+        with patch.object(_config_version_mod, "get_version", return_value="same_ver"):
+            result = self.router.get_capability("test")
         self.assertEqual(result, cap)
 
     def test_invalidate_clears_cache(self):
@@ -213,181 +214,163 @@ class TestAIRouter(unittest.TestCase):
         self.assertEqual(len(self.router._cache), 0)
         self.assertIsNone(self.router._local_ver)
 
-    @patch.object(_ai_router, "_get_session", return_value=None)
-    @patch.object(_ai_config_version, "get_version", return_value=None)
-    def test_get_llm_client_fallback(self, mock_ver, mock_session):
-        _ai_router.get_shared_settings = lambda: _mock_settings
-        client = self.router.get_llm_client("intent_parsing")
+    def test_get_llm_client_fallback(self):
+        with self._freeze_version(None):
+            with patch.object(ai_router_mod, "_get_session", return_value=None):
+                client = self.router.get_llm_client("intent_parsing")
         self.assertIsNotNone(client)
         self.assertEqual(client.model, "doubao-seed-2-1-pro-260915")
         self.assertEqual(client.api_key, "test-key")
 
-    @patch.object(_ai_config_version, "get_version", return_value=None)
-    def test_get_llm_client_from_db(self, mock_ver):
-        model = _ai_router.ModelConfig(
-            id="m1", name="test", provider="dashscope", model_key="qwen3.8-flash",
-            api_key="db-key", api_base="https://dashscope.test.com", api_version=None, status="active",
-        )
-        cap = _ai_router.CapabilityConfig(
-            capability_key="image_analysis", capability_name="图片分析",
-            primary_model=model, lite_model=None, fallback_model=None,
-            timeout_seconds=15, max_retries=1, max_tokens=350, temperature=0.1,
-            concurrency=2, extra_params={}, is_enabled=True,
+    def test_get_llm_client_from_db(self):
+        cap = _capability_config(
+            capability_key="image_analysis",
+            capability_name="图片分析",
+            max_tokens=350,
+            temperature=0.1,
+            primary_model=_model_config(
+                provider="dashscope",
+                model_key="qwen3.8-flash",
+                api_key="db-key",
+                api_base="https://dashscope.test.com",
+            ),
         )
         with patch.object(self.router, "get_capability", return_value=cap):
             client = self.router.get_llm_client("image_analysis")
-            self.assertIsNotNone(client)
-            self.assertEqual(client.model, "qwen3.8-flash")
-            self.assertEqual(client.provider, "dashscope")
+        self.assertIsNotNone(client)
+        self.assertEqual(client.model, "qwen3.8-flash")
+        self.assertEqual(client.provider, "dashscope")
 
-    @patch.object(_ai_config_version, "get_version", return_value=None)
-    def test_get_vision_client(self, mock_ver):
-        model = _ai_router.ModelConfig(
-            id="m1", name="test", provider="dashscope", model_key="qwen3.8-flash",
-            api_key="key", api_base="https://dashscope.test.com", api_version=None, status="active",
-        )
-        cap = _ai_router.CapabilityConfig(
-            capability_key="image_analysis", capability_name="图片分析",
-            primary_model=model, lite_model=None, fallback_model=None,
-            timeout_seconds=15, max_retries=1, max_tokens=None, temperature=None,
-            concurrency=2, extra_params={}, is_enabled=True,
+    def test_get_vision_client(self):
+        cap = _capability_config(
+            capability_key="image_analysis",
+            capability_name="图片分析",
+            primary_model=_model_config(
+                provider="dashscope",
+                model_key="qwen3.8-flash",
+                api_key="key",
+                api_base="https://dashscope.test.com",
+            ),
         )
         with patch.object(self.router, "get_capability", return_value=cap):
             client = self.router.get_vision_client("image_analysis")
-            self.assertIsNotNone(client)
-            # #2220: vision client is now DoubaoClient with vision_completion
-            self.assertTrue(hasattr(client, "vision_completion"))
+        self.assertIsNotNone(client)
+        # #2220: vision client is now DoubaoClient with vision_completion
+        self.assertTrue(hasattr(client, "vision_completion"))
 
-    @patch.object(_ai_config_version, "get_version", return_value=None)
-    def test_get_tts_client(self, mock_ver):
-        model = _ai_router.ModelConfig(
-            id="m1", name="test", provider="dashscope", model_key="cosyvoice-v3-flash",
-            api_key="key", api_base="https://dashscope.test.com", api_version=None, status="active",
-        )
-        cap = _ai_router.CapabilityConfig(
-            capability_key="tts", capability_name="语音合成",
-            primary_model=model, lite_model=None, fallback_model=None,
-            timeout_seconds=60, max_retries=1, max_tokens=None, temperature=None,
-            concurrency=2, extra_params={}, is_enabled=True,
+    def test_get_tts_client(self):
+        cap = _capability_config(
+            capability_key="tts",
+            capability_name="语音合成",
+            primary_model=_model_config(
+                provider="dashscope",
+                model_key="cosyvoice-v3-flash",
+                api_key="key",
+                api_base="https://dashscope.test.com",
+            ),
         )
         with patch.object(self.router, "get_capability", return_value=cap):
             client = self.router.get_tts_client()
-            self.assertIsNotNone(client)
-            self.assertEqual(client.model, "cosyvoice-v3-flash")
+        self.assertIsNotNone(client)
+        self.assertEqual(client.model, "cosyvoice-v3-flash")
 
-    @patch.object(_ai_config_version, "get_version", return_value=None)
-    def test_get_image_gen_client(self, mock_ver):
-        model = _ai_router.ModelConfig(
-            id="m1", name="test", provider="volcengine", model_key="seedream-5.0-flash",
-            api_key="key", api_base="https://ark.test.com", api_version=None, status="active",
-        )
-        cap = _ai_router.CapabilityConfig(
-            capability_key="image_generation", capability_name="图片生成",
-            primary_model=model, lite_model=None, fallback_model=None,
-            timeout_seconds=60, max_retries=1, max_tokens=None, temperature=None,
-            concurrency=2, extra_params={"size": "1K"}, is_enabled=True,
+    def test_get_image_gen_client(self):
+        cap = _capability_config(
+            capability_key="image_generation",
+            capability_name="图片生成",
+            extra_params={"size": "1K"},
+            primary_model=_model_config(
+                model_key="seedream-5.0-flash",
+                api_key="key",
+                api_base="https://ark.test.com",
+            ),
         )
         with patch.object(self.router, "get_capability", return_value=cap):
             client = self.router.get_image_gen_client()
-            self.assertIsNotNone(client)
-            self.assertEqual(client.model, "seedream-5.0-flash")
+        self.assertIsNotNone(client)
+        self.assertEqual(client.model, "seedream-5.0-flash")
 
-    @patch.object(_ai_config_version, "get_version", return_value=None)
-    def test_get_video_gen_client(self, mock_ver):
-        model = _ai_router.ModelConfig(
-            id="m1", name="test", provider="volcengine", model_key="seedance-2.5",
-            api_key="key", api_base="https://ark.test.com", api_version=None, status="active",
-        )
-        cap = _ai_router.CapabilityConfig(
-            capability_key="video_generation", capability_name="视频生成",
-            primary_model=model, lite_model=None, fallback_model=None,
-            timeout_seconds=600, max_retries=1, max_tokens=None, temperature=None,
-            concurrency=1, extra_params={}, is_enabled=True,
+    def test_get_video_gen_client(self):
+        cap = _capability_config(
+            capability_key="video_generation",
+            capability_name="视频生成",
+            concurrency=1,
+            primary_model=_model_config(
+                model_key="seedance-2.5",
+                api_key="key",
+                api_base="https://ark.test.com",
+            ),
         )
         with patch.object(self.router, "get_capability", return_value=cap):
             client = self.router.get_video_gen_client()
-            self.assertIsNotNone(client)
-            self.assertEqual(client.model, "seedance-2.5")
+        self.assertIsNotNone(client)
+        self.assertEqual(client.model, "seedance-2.5")
 
-    @patch.object(_ai_config_version, "get_version", return_value=None)
-    def test_lite_variant_preference(self, mock_ver):
-        primary = _ai_router.ModelConfig(id="p1", name="pro", provider="volcengine", model_key="pro-model", api_key="k", api_base="u", api_version=None, status="active")
-        lite = _ai_router.ModelConfig(id="l1", name="lite", provider="volcengine", model_key="lite-model", api_key="k", api_base="u", api_version=None, status="active")
-        cap = _ai_router.CapabilityConfig(
-            capability_key="image_analysis", capability_name="图片分析",
-            primary_model=primary, lite_model=lite, fallback_model=None,
-            timeout_seconds=15, max_retries=1, max_tokens=None, temperature=None,
-            concurrency=2, extra_params={}, is_enabled=True,
+    def test_lite_variant_preference(self):
+        cap = _capability_config(
+            capability_key="image_analysis",
+            capability_name="图片分析",
+            primary_model=_model_config(id="p1", name="pro", model_key="pro-model", api_key="k", api_base="u"),
+            lite_model=_model_config(id="l1", name="lite", model_key="lite-model", api_key="k", api_base="u"),
         )
         model = self.router._get_model_or_fallback(cap, "lite")
         self.assertEqual(model.model_key, "lite-model")
         model_primary = self.router._get_model_or_fallback(cap, "primary")
         self.assertEqual(model_primary.model_key, "pro-model")
 
-    @patch.object(_ai_config_version, "get_version", return_value=None)
-    def test_disabled_capability_returns_fallback(self, mock_ver):
-        cap = _ai_router.CapabilityConfig(
-            capability_key="test", capability_name="test",
-            primary_model=None, lite_model=None, fallback_model=None,
-            timeout_seconds=30, max_retries=1, max_tokens=None, temperature=None,
-            concurrency=2, extra_params={}, is_enabled=False,
-        )
-        _ai_router.get_shared_settings = lambda: _mock_settings
+    def test_disabled_capability_returns_fallback(self):
+        cap = _capability_config(is_enabled=False)
         with patch.object(self.router, "get_capability", return_value=cap):
             client = self.router.get_llm_client("test")
-            self.assertIsNotNone(client)
-            self.assertEqual(client.model, "doubao-seed-2-1-pro-260915")
+        self.assertIsNotNone(client)
+        self.assertEqual(client.model, "doubao-seed-2-1-pro-260915")
 
-    @patch.object(_ai_config_version, "get_version", return_value=None)
-    def test_fallback_chain_primary_none(self, mock_ver):
+    def test_fallback_chain_primary_none(self):
         """primary_model 为 None 时 fallback 到 fallback_model"""
-        fb = _ai_router.ModelConfig(id="f1", name="fb", provider="volcengine", model_key="fb-model", api_key="k", api_base="u", api_version=None, status="active")
-        cap = _ai_router.CapabilityConfig(
-            capability_key="test", capability_name="test",
-            primary_model=None, lite_model=None, fallback_model=fb,
-            timeout_seconds=30, max_retries=1, max_tokens=None, temperature=None,
-            concurrency=2, extra_params={}, is_enabled=True,
+        cap = _capability_config(
+            fallback_model=_model_config(id="f1", name="fb", model_key="fb-model", api_key="k", api_base="u"),
         )
         model = self.router._get_model_or_fallback(cap, "primary")
         self.assertEqual(model.model_key, "fb-model")
 
 
+# ── 数据类冻结 ──────────────────────────────────────────────────────────────
+
+
 class TestModelConfig(unittest.TestCase):
     """数据类测试"""
 
     def test_model_config_frozen(self):
-        m = _ai_router.ModelConfig(id="1", name="t", provider="p", model_key="k", api_key="a", api_base="b", api_version=None, status="active")
+        m = _model_config(id="1", name="t", provider="p", model_key="k", api_key="a", api_base="b")
         with self.assertRaises(AttributeError):
             m.model_key = "new"
 
     def test_capability_config_frozen(self):
-        c = _ai_router.CapabilityConfig(
-            capability_key="k", capability_name="n", primary_model=None,
-            lite_model=None, fallback_model=None, timeout_seconds=30,
-            max_retries=1, max_tokens=None, temperature=None, concurrency=2,
-            extra_params={}, is_enabled=True,
-        )
+        c = _capability_config()
         with self.assertRaises(AttributeError):
             c.is_enabled = False
 
 
+# ── 客户端可用性 ────────────────────────────────────────────────────────────
+
+
 class TestClientAvailability(unittest.TestCase):
     """客户端可用性测试"""
 
     def test_tts_client_available(self):
-        c = _ai_router.TTSClient(provider="p", api_key="k", base_url="u", model="m")
+        c = ai_router_mod.TTSClient(provider="p", api_key="k", base_url="u", model="m")
         self.assertTrue(c.is_available)
 
     def test_tts_client_unavailable_no_model(self):
-        c = _ai_router.TTSClient(provider="p", api_key="k", base_url="u", model="")
+        c = ai_router_mod.TTSClient(provider="p", api_key="k", base_url="u", model="")
         self.assertFalse(c.is_available)
 
     def test_image_gen_client_unavailable_no_url(self):
-        c = _ai_router.ImageGenClient(provider="p", api_key="k", base_url="", model="m")
+        c = ai_router_mod.ImageGenClient(provider="p", api_key="k", base_url="", model="m")
         self.assertFalse(c.is_available)
 
     def test_video_gen_client_available(self):
-        c = _ai_router.VideoGenClient(provider="p", api_key="k", base_url="u", model="m")
+        c = ai_router_mod.VideoGenClient(provider="p", api_key="k", base_url="u", model="m")
         self.assertTrue(c.is_available)
 
 

tests/unit/test_ditto_emotion_service.py

@@ -0,0 +1,352 @@
+"""Ditto LLM 情绪分析服务单元测试."""
+
+from __future__ import annotations
+
+import json
+from unittest.mock import MagicMock, patch
+
+import pytest
+
+from packages.application.ditto_emotion_service import (
+    EMO_HAPPY,
+    EMO_NEUTRAL,
+    DittoEmotionService,
+    EmotionSegment,
+    _parse_emotion_json,
+    _warn_if_suspicious_all_neutral,
+    align_timeline_by_length,
+    align_timeline_by_timings,
+    split_sentences,
+)
+
+
+# ── 分句 ─────────────────────────────────────────────────────────
+class TestSplitSentences:
+    def test_empty(self):
+        assert split_sentences("") == []
+
+    def test_single(self):
+        assert split_sentences("你好。") == ["你好。"]
+
+    def test_multi(self):
+        sents = split_sentences("大家好！今天给大家推荐一款超棒的产品。它真的很好用；不信你试试？")
+        assert len(sents) == 4
+        assert "大家好！" in sents[0]
+
+    def test_english_punct(self):
+        sents = split_sentences("Hello! How are you? I'm fine.")
+        assert len(sents) == 3
+
+    # ── Bug1 修复：无句末标点长文案按逗号二次切分 ──
+    def test_no_ending_punct_long_text_split_by_comma(self):
+        """公益文案：全是逗号、结尾无标点，整段超过30字 → 逗号二次切分."""
+        text = "做公益不一定是一个人去做惊天动地的大事，而是一群人扎扎实实的干好身边的小事，心中有爱，眼里有光"
+        sents = split_sentences(text)
+        assert len(sents) >= 3
+        assert any("心中有爱" in s for s in sents)
+        assert any("眼里有光" in s for s in sents)
+
+    def test_short_text_no_split(self):
+        """不超过30字的无句末标点文案不二次切分."""
+        text = "大家好，欢迎收看"
+        assert split_sentences(text) == [text]
+
+    def test_short_clauses_merged(self):
+        """过短片段（<10字）并入相邻句，避免切太碎."""
+        text = "家人们，你们是不是经常买东西踩坑呢，花了大价钱买回来的东西根本没法用啊"
+        sents = split_sentences(text)
+        assert len(sents) == 2
+        assert all(len(s) >= 10 for s in sents)
+
+    def test_normal_punct_no_clause_split(self):
+        """有句末标点正常切分的文案不做逗号二次切分."""
+        text = "大家好！今天给大家推荐一款超棒的产品，它真的很好用。不信你试试？"
+        sents = split_sentences(text)
+        assert len(sents) == 3
+        assert "今天给大家推荐一款超棒的产品，它真的很好用。" in sents
+
+
+# ── JSON 解析 ────────────────────────────────────────────────────
+class TestParseEmotionJson:
+    def test_valid(self):
+        raw = json.dumps([{"text": "你好", "emo": 4, "intensity": 0.2}])
+        segs = _parse_emotion_json(raw)
+        assert len(segs) == 1
+        assert segs[0].emo == 4
+        assert segs[0].intensity == 0.2
+
+    def test_markdown_wrapped(self):
+        raw = "```json\n" + json.dumps([{"text": "好", "emo": 3, "intensity": 0.5}]) + "\n```"
+        segs = _parse_emotion_json(raw)
+        assert len(segs) == 1
+        assert segs[0].emo == 3
+
+    def test_all_eight_emos_allowed(self):
+        """8 种表情全部开放：0-7 原样保留，不再强制转中性."""
+        for emo in range(8):
+            raw = json.dumps([{"text": "句", "emo": emo, "intensity": 0.5}])
+            segs = _parse_emotion_json(raw)
+            assert len(segs) == 1
+            assert segs[0].emo == emo
+
+    def test_out_of_range_emo_becomes_neutral(self):
+        """超出 0-7 范围的非法 emo 值仍降级为中性."""
+        raw = json.dumps([{"text": "句", "emo": 9, "intensity": 0.5}])
+        segs = _parse_emotion_json(raw)
+        assert len(segs) == 1
+        assert segs[0].emo == EMO_NEUTRAL
+
+    def test_invalid_json(self):
+        assert _parse_emotion_json("not json") == []
+
+    def test_empty(self):
+        assert _parse_emotion_json("") == []
+
+    def test_intensity_clamp(self):
+        raw = json.dumps([{"text": "a", "emo": 3, "intensity": 1.5}])
+        segs = _parse_emotion_json(raw)
+        assert segs[0].intensity == 1.0
+
+    def test_missing_text_skipped(self):
+        raw = json.dumps([{"emo": 3, "intensity": 0.4}])
+        segs = _parse_emotion_json(raw)
+        assert len(segs) == 0
+
+
+# ── 时间对齐（按字数比例）────────────────────────────────────────
+class TestAlignTimelineByLength:
+    def test_basic(self):
+        segs = [
+            EmotionSegment("ab", EMO_NEUTRAL, 0.2),
+            EmotionSegment("cd", EMO_HAPPY, 0.5),
+        ]
+        entries = align_timeline_by_length(segs, 4.0)
+        assert len(entries) == 2
+        assert entries[0].start == 0.0
+        assert entries[0].end == 2.0
+        assert entries[1].start == 2.0
+        assert entries[1].end == 4.0
+        assert entries[0].emo == EMO_NEUTRAL
+        assert entries[1].emo == EMO_HAPPY
+
+    def test_empty_segments(self):
+        assert align_timeline_by_length([], 5.0) == []
+
+    def test_zero_duration(self):
+        segs = [EmotionSegment("ab", EMO_NEUTRAL, 0.2)]
+        assert align_timeline_by_length(segs, 0) == []
+
+    def test_unequal_length(self):
+        segs = [
+            EmotionSegment("a" * 3, EMO_HAPPY, 0.5),
+            EmotionSegment("b" * 1, EMO_NEUTRAL, 0.2),
+        ]
+        entries = align_timeline_by_length(segs, 4.0)
+        assert entries[0].end == 3.0
+        assert entries[1].start == 3.0
+        assert entries[1].end == 4.0
+
+
+# ── 时间对齐（sentence_timings）──────────────────────────────────
+class TestAlignTimelineByTimings:
+    def test_exact_match(self):
+        segs = [
+            EmotionSegment("hello", EMO_HAPPY, 0.4),
+            EmotionSegment("world", EMO_NEUTRAL, 0.2),
+        ]
+        timings = [
+            {"start": 0.0, "end": 1.5},
+            {"start": 1.5, "end": 3.0},
+        ]
+        entries = align_timeline_by_timings(segs, timings, 3.0)
+        assert len(entries) == 2
+        assert entries[0].start == 0.0
+        assert entries[0].end == 1.5
+        assert entries[1].start == 1.5
+        assert entries[1].end == 3.0
+
+    def test_length_mismatch_fallback(self):
+        segs = [EmotionSegment("hello", EMO_HAPPY, 0.4)]
+        timings = [{"start": 0, "end": 1}, {"start": 1, "end": 2}]
+        entries = align_timeline_by_timings(segs, timings, 2.0)
+        assert len(entries) == 1
+        assert entries[0].end == 2.0
+
+
+# ── DittoEmotionService ──────────────────────────────────────────
+def _make_service(enabled=True, model=None, temperature=0.1, timeout=10, max_tokens=1024, prompt=""):
+    s = MagicMock()
+    s.ditto_emotion_enabled = enabled
+    s.ditto_emotion_model = model or ""
+    s.ditto_emotion_temperature = temperature
+    s.ditto_emotion_timeout = timeout
+    s.ditto_emotion_max_tokens = max_tokens
+    s.ditto_emotion_cache_size = 100
+    s.ditto_emotion_prompt = prompt
+    return DittoEmotionService(settings=s)
+
+
+class TestDittoEmotionService:
+    def test_disabled_returns_empty(self):
+        svc = _make_service(enabled=False)
+        assert svc.analyze("你好世界") == []
+
+    def test_empty_text_returns_empty(self):
+        svc = _make_service(enabled=True)
+        assert svc.analyze("") == []
+
+    def test_llm_success(self):
+        svc = _make_service(enabled=True)
+        fake_reply = json.dumps([{"text": "你好", "emo": 4, "intensity": 0.2}])
+        with patch.object(svc, "_call_llm", return_value=_parse_emotion_json(fake_reply)):
+            segs = svc.analyze("你好")
+            assert len(segs) == 1
+            assert segs[0].emo == 4
+
+    def test_cache_hit(self):
+        svc = _make_service(enabled=True)
+        fake_reply = json.dumps([{"text": "你好世界", "emo": 3, "intensity": 0.5}])
+        with patch.object(svc, "_call_llm", return_value=_parse_emotion_json(fake_reply)) as mock_call:
+            svc.analyze("你好世界")
+            svc.analyze("你好世界")
+            assert mock_call.call_count == 1
+
+    def test_build_timeline_empty_when_disabled(self):
+        svc = _make_service(enabled=False)
+        assert svc.build_timeline("test", 5.0) == ""
+
+    def test_build_timeline_returns_json(self):
+        svc = _make_service(enabled=True)
+        fake_reply = json.dumps(
+            [
+                {"text": "ab", "emo": 4, "intensity": 0.2},
+                {"text": "cd", "emo": 3, "intensity": 0.4},
+            ]
+        )
+        with patch.object(svc, "_call_llm", return_value=_parse_emotion_json(fake_reply)):
+            result = svc.build_timeline("ab。cd。", 4.0)
+            data = json.loads(result)
+            assert len(data) == 2
+            assert data[0]["emo"] == 4
+            assert data[1]["emo"] == 3
+
+    def test_build_timeline_with_sentence_timings(self):
+        svc = _make_service(enabled=True)
+        fake_reply = json.dumps(
+            [
+                {"text": "hello", "emo": 3, "intensity": 0.4},
+                {"text": "world", "emo": 4, "intensity": 0.2},
+            ]
+        )
+        timings = [
+            {"start": 0.0, "end": 1.0},
+            {"start": 1.0, "end": 3.0},
+        ]
+        with patch.object(svc, "_call_llm", return_value=_parse_emotion_json(fake_reply)):
+            result = svc.build_timeline("hello world", 3.0, sentence_timings=timings)
+            data = json.loads(result)
+            assert data[0]["start"] == 0.0
+            assert data[0]["end"] == 1.0
+            assert data[1]["end"] == 3.0
+
+
+class TestPromptLoading:
+    def test_default_prompt_contains_placeholder(self):
+        from packages.application.ditto_emotion_service import _load_default_prompt
+
+        prompt = _load_default_prompt()
+        assert "{文案}" in prompt
+
+    def test_config_prompt_override(self):
+        custom = "分析情绪: {文案}"
+        svc = _make_service(enabled=True, prompt=custom)
+        assert svc._get_prompt_template() == custom
+
+
+# ── _get_llm_client：DB 配置 api_key 优先 + 变化重建 ─────────────
+class TestGetLLMClient:
+    def test_uses_config_api_key(self):
+        svc = _make_service(enabled=True)
+        with patch.object(svc, "_cfg", return_value="sk-from-db") as mock_cfg:
+            with patch("packages.shared.ai_client.DoubaoClient") as MockClient:
+                svc._get_llm_client()
+                mock_cfg.assert_called_with("doubao_api_key")
+                assert MockClient.call_args.kwargs["api_key"] == "sk-from-db"
+
+    def test_rebuilds_client_when_key_changes(self):
+        svc = _make_service(enabled=True)
+        keys = iter(["sk-key-v1", "sk-key-v2"])
+        with patch.object(svc, "_cfg", side_effect=lambda k: next(keys)):
+            with patch("packages.shared.ai_client.DoubaoClient") as MockClient:
+                MockClient.side_effect = lambda **kw: MagicMock(api_key=kw["api_key"])
+                c1 = svc._get_llm_client()
+                c2 = svc._get_llm_client()
+                assert c1 is not c2
+                assert MockClient.call_count == 2
+
+    def test_reuses_client_when_key_unchanged(self):
+        svc = _make_service(enabled=True)
+        with patch.object(svc, "_cfg", return_value="sk-same"):
+            with patch("packages.shared.ai_client.DoubaoClient") as MockClient:
+                MockClient.side_effect = lambda **kw: MagicMock(api_key=kw["api_key"])
+                c1 = svc._get_llm_client()
+                c2 = svc._get_llm_client()
+                assert c1 is c2
+                assert MockClient.call_count == 1
+
+    def test_empty_key_falls_back_to_doubao_client_default(self):
+        """DB/env 都没配时传空串，由 DoubaoClient 内部 fallback env / 标记不可用."""
+        svc = _make_service(enabled=True)
+        with patch.object(svc, "_cfg", return_value=""):
+            with patch("packages.shared.ai_client.DoubaoClient") as MockClient:
+                svc._get_llm_client()
+                assert MockClient.call_args.kwargs["api_key"] == ""
+
+
+# ── 单句兜底：LLM 返回空时单句也走中性兜底（>= 1）────────────────
+class TestSingleSentenceFallback:
+    def test_single_sentence_gets_neutral_fallback(self):
+        import packages.application.ditto_emotion_service as m
+
+        svc = _make_service(enabled=True)
+        m._cached_analyze.cache_clear()
+        try:
+            with patch.object(svc, "_call_llm", return_value=[]):
+                segs = svc.analyze("独一无二单句兜底测试文案abc123")
+                assert len(segs) == 1
+                assert segs[0].emo == EMO_NEUTRAL
+                assert segs[0].intensity == 0.1
+        finally:
+            m._cached_analyze.cache_clear()
+
+
+# ── LLM 结果合理性校验（Bug2）────────────────────────────────────
+class TestWarnIfSuspiciousAllNeutral:
+    def test_warns_when_all_neutral_with_signals(self, caplog):
+        """文案含感叹号但全中性/低强度 → warning."""
+        segs = [EmotionSegment(text="家人们！", emo=4, intensity=0.1)]
+        with caplog.at_level("WARNING", logger="packages.application.ditto_emotion_service"):
+            _warn_if_suspicious_all_neutral("家人们！你们是不是经常踩坑？", segs)
+        assert any("全中性" in r.message or "全中性" in r.getMessage() for r in caplog.records)
+
+    def test_no_warn_when_has_emotion(self, caplog):
+        """有非中性表情 → 不 warning."""
+        segs = [
+            EmotionSegment(text="家人们！", emo=6, intensity=0.5),
+            EmotionSegment(text="你们是不是经常踩坑？", emo=5, intensity=0.3),
+        ]
+        with caplog.at_level("WARNING", logger="packages.application.ditto_emotion_service"):
+            _warn_if_suspicious_all_neutral("家人们！你们是不是经常踩坑？", segs)
+        assert not caplog.records
+
+    def test_no_warn_when_no_signal(self, caplog):
+        """文案无情绪信号，即使全中性也不 warning."""
+        segs = [EmotionSegment(text="今天天气不错。", emo=4, intensity=0.1)]
+        with caplog.at_level("WARNING", logger="packages.application.ditto_emotion_service"):
+            _warn_if_suspicious_all_neutral("今天天气不错。", segs)
+        assert not caplog.records
+
+    def test_no_warn_when_empty(self, caplog):
+        with caplog.at_level("WARNING", logger="packages.application.ditto_emotion_service"):
+            _warn_if_suspicious_all_neutral("家人们！", [])
+        assert not caplog.records

tests/unit/test_ditto_service.py

@@ -0,0 +1,229 @@
+"""Ditto 蚂蚁数字人客户端单元测试 — #2076."""
+
+from __future__ import annotations
+
+import time
+from unittest.mock import MagicMock, patch
+
+import httpx
+import pytest
+
+from packages.application.ditto_service import DittoClient, DittoError, DittoResult
+
+
+class _FakeResponse:
+    def __init__(self, status_code=200, content=b"\x00\x01" * 1000, headers=None, text=""):
+        self.status_code = status_code
+        self.content = content
+        self.headers = headers or {}
+        self.text = text
+
+
+def _make_client(base_url="http://ditto:8000", default_video_url="http://oss/tpl.mp4", max_retries=2, timeout=60):
+    with patch("packages.application.ditto_service.get_api_settings") as mock_settings:
+        s = MagicMock()
+        s.ditto_api_base_url = base_url
+        s.ditto_default_video_url = default_video_url
+        s.ditto_max_retries = max_retries
+        s.ditto_request_timeout = timeout
+        mock_settings.return_value = s
+        return DittoClient()
+
+
+def test_is_configured_true():
+    c = _make_client()
+    assert c.is_configured is True
+
+
+def test_is_configured_false_without_base():
+    c = _make_client(base_url="")
+    assert c.is_configured is False
+
+
+def test_is_configured_false_without_template():
+    c = _make_client(default_video_url="")
+    assert c.is_configured is False
+
+
+def test_health_ok():
+    c = _make_client()
+    with patch("httpx.Client") as mock_cls:
+        client = MagicMock()
+        client.get.return_value = _FakeResponse(200)
+        mock_cls.return_value.__enter__.return_value = client
+        assert c.health() is True
+        client.get.assert_called_once()
+
+
+def test_health_fail_status():
+    c = _make_client()
+    with patch("httpx.Client") as mock_cls:
+        client = MagicMock()
+        client.get.return_value = _FakeResponse(500)
+        mock_cls.return_value.__enter__.return_value = client
+        assert c.health() is False
+
+
+def test_health_network_error():
+    c = _make_client()
+    with patch("httpx.Client") as mock_cls:
+        client = MagicMock()
+        client.get.side_effect = httpx.ConnectError("fail")
+        mock_cls.return_value.__enter__.return_value = client
+        assert c.health() is False
+
+
+def test_generate_missing_base():
+    c = _make_client(base_url="")
+    with pytest.raises(DittoError, match="DITTO_API_BASE_URL"):
+        c.generate(audio_url="http://x/a.mp3", script="你好")
+
+
+def test_generate_missing_audio():
+    c = _make_client()
+    with pytest.raises(DittoError, match="audio_url"):
+        c.generate(audio_url="", script="你好")
+
+
+def test_generate_success_with_headers():
+    c = _make_client(max_retries=0)
+    fake_resp = _FakeResponse(
+        status_code=200,
+        content=b"\x00" * 99999,
+        headers={"X-RTF": "0.35", "X-Frames": "125", "X-Time": "12.5"},
+    )
+    with patch("httpx.Client") as mock_cls, patch("time.monotonic", side_effect=[0, 1]):
+        client = MagicMock()
+        client.post.return_value = fake_resp
+        mock_cls.return_value.__enter__.return_value = client
+        result = c.generate(audio_url="http://x/a.mp3", script="你好")
+    assert isinstance(result, DittoResult)
+    assert len(result.video_bytes) == 99999
+    assert result.rtf == 0.35
+    assert result.frames == 125
+    assert result.elapsed_seconds == 12.5
+
+
+def test_generate_uses_default_template_when_video_url_empty():
+    c = _make_client(max_retries=0)
+    fake_resp = _FakeResponse(200, b"1" * 99999)
+    with patch("httpx.Client") as mock_cls:
+        client = MagicMock()
+        client.post.return_value = fake_resp
+        mock_cls.return_value.__enter__.return_value = client
+        c.generate(audio_url="http://x/a.mp3", script="你好")
+        call_kwargs = client.post.call_args
+        payload = call_kwargs.kwargs.get("json") or call_kwargs[1].get("json")
+        assert payload["video_url"] == "http://oss/tpl.mp4"
+        assert payload["audio_url"] == "http://x/a.mp3"
+        assert payload["script"] == "你好"
+        assert payload["emo_global"] == 4
+        assert payload["use_script_emo"] is True
+
+
+def test_generate_retries_on_429_then_success():
+    c = _make_client(max_retries=2)
+    busy = _FakeResponse(429, b"", text="busy")
+    ok = _FakeResponse(200, b"v" * 99999)
+    with patch("httpx.Client") as mock_cls, patch("time.sleep") as mock_sleep:
+        client = MagicMock()
+        client.post.side_effect = [busy, ok]
+        mock_cls.return_value.__enter__.return_value = client
+        result = c.generate(audio_url="http://x/a.mp3", script="你好")
+    assert len(result.video_bytes) == 99999
+    assert mock_sleep.called
+    assert client.post.call_count == 2
+
+
+def test_generate_429_exhausted():
+    c = _make_client(max_retries=1)
+    with patch("httpx.Client") as mock_cls, patch("time.sleep"):
+        client = MagicMock()
+        client.post.return_value = _FakeResponse(429, b"", text="busy")
+        mock_cls.return_value.__enter__.return_value = client
+        with pytest.raises(DittoError, match="重试"):
+            c.generate(audio_url="http://x/a.mp3", script="你好")
+
+
+def test_generate_400_no_retry():
+    c = _make_client(max_retries=2)
+    with patch("httpx.Client") as mock_cls:
+        client = MagicMock()
+        client.post.return_value = _FakeResponse(400, b"", text="bad request")
+        mock_cls.return_value.__enter__.return_value = client
+        with pytest.raises(DittoError, match="Ditto 返回 400"):
+            c.generate(audio_url="http://x/a.mp3", script="你好")
+        assert client.post.call_count == 1  # 400 不重试
+
+
+def test_generate_small_response_raises():
+    c = _make_client(max_retries=0)
+    with patch("httpx.Client") as mock_cls:
+        client = MagicMock()
+        client.post.return_value = _FakeResponse(200, b"xx")
+        mock_cls.return_value.__enter__.return_value = client
+        with pytest.raises(DittoError) as exc_info:
+            c.generate(audio_url="http://x/a.mp3", script="你好")
+        assert exc_info.value.code == "EmptyResponse"
+
+
+def test_generate_and_persist_uploads_to_storage():
+    c = _make_client(max_retries=0)
+    fake_resp = _FakeResponse(200, b"v" * 99999)
+    fake_storage = MagicMock()
+    fake_storage.upload_file.return_value = "http://oss/ditto/x.mp4"
+    with (
+        patch("httpx.Client") as mock_cls,
+        patch("packages.shared.storage.get_shared_storage_service", return_value=fake_storage),
+    ):
+        client = MagicMock()
+        client.post.return_value = fake_resp
+        mock_cls.return_value.__enter__.return_value = client
+        result = c.generate_and_persist(job_id="j1", user_id="u1", audio_url="http://x/a.mp3", script="hi")
+    assert result.video_url == "http://oss/ditto/x.mp4"
+    fake_storage.upload_file.assert_called_once()
+    call_args = fake_storage.upload_file.call_args
+    assert call_args.args[1].startswith("ditto-output/u1/j1")
+
+
+def test_empty_script_replaced_with_space():
+    c = _make_client(max_retries=0)
+    fake_resp = _FakeResponse(200, b"v" * 99999)
+    with patch("httpx.Client") as mock_cls:
+        client = MagicMock()
+        client.post.return_value = fake_resp
+        mock_cls.return_value.__enter__.return_value = client
+        c.generate(audio_url="http://x/a.mp3", script="")
+        payload = client.post.call_args.kwargs["json"]
+        assert payload["script"] == " "
+
+
+def test_generate_network_error_fails_fast(monkeypatch):
+    """网络不通（ConnectError）时不重试，直接快速抛 NetworkUnreachable，避免用户等5分钟"""
+    import httpx
+
+    from packages.application import ditto_service as ds_mod
+
+    calls = {"n": 0}
+
+    def _fake_post(self, url, json=None):
+        calls["n"] += 1
+        raise httpx.ConnectError("[Errno 113] No route to host")
+
+    monkeypatch.setattr(httpx.Client, "post", _fake_post)
+
+    client = _make_client(
+        base_url="http://100.76.80.23:8000",
+        default_video_url="http://oss/tpl.mp4",
+        max_retries=2,
+        timeout=120,
+    )
+
+    t0 = time.monotonic()
+    with pytest.raises(ds_mod.DittoError) as exc:
+        client.generate(audio_url="http://oss/a.wav", script="你好")
+    elapsed = time.monotonic() - t0
+
+    assert exc.value.code == "NetworkUnreachable"
+    assert calls["n"] == 1  # 不重试
+    assert elapsed < 5  # 快速失败<5秒