Merge pull request 'fix: _assemble_v4路由优先级修复，scene/store有人物时不再误入person分支' (#2230) from fix/assembler-v4-routing-priority into develop

问题：person分支条件 vtype==person or has_person 过于宽泛，
当v6 prompt返回type=scene/store但画面有人时，has_person=true
导致直接进入person分支，scene/store的专有字段全部丢失。

修复：
1. person分支条件收窄为 vtype==person，去掉 or has_person
2. 新增向后兼容逻辑：type不在已知枚举时按has_person/products兜底
3. scene/store分支has_person=true时，提取人物穿着/姿势等信息
   追加到key_features（前缀人物:），不丢失人物数据
4. store返回的has_person字段改为反映真实值

Closes #2265

alembic/versions/100_fix_capability_model_bindings.py

@@ -0,0 +1,107 @@
+# -*- coding: utf-8 -*-
+"""100: 修正已有 capability 的模型绑定.
+
+幂等：仅当 primary_model_id 当前绑定到旧模型 (doubao-seed-1-6) 时才更新，
+避免覆盖用户在后台的自定义配置。
+
+- 更新 5 个 LLM capability (intent_parsing, copy_fusion, storyboard, copy_review, asset_classify)
+  的 primary_model_id 从 doubao-seed-1-6 改为 doubao-seed-2-1-pro-260915
+- 更新 image_analysis 的 primary/lite/fallback 模型绑定
+"""
+
+import sqlalchemy as sa
+
+from alembic import op
+
+revision = "100_fix_capability_model_bindings"
+down_revision = "099_ai_model_router_seed"
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    conn = op.get_bind()
+
+    # Check tables exist
+    table_check = conn.execute(sa.text("SELECT to_regclass('public.ai_models')")).scalar()
+    if not table_check:
+        return
+
+    config_table_check = conn.execute(sa.text("SELECT to_regclass('public.ai_capability_configs')")).scalar()
+    if not config_table_check:
+        return
+
+    # Look up model IDs by model_key (not hardcoded UUIDs)
+    pro_model_row = conn.execute(
+        sa.text(
+            "SELECT id FROM ai_models WHERE model_key = 'doubao-seed-2-1-pro-260915' AND deleted_at IS NULL LIMIT 1"
+        )
+    ).first()
+    if not pro_model_row:
+        return
+    pro_model_id = pro_model_row[0]
+
+    old_model_row = conn.execute(
+        sa.text("SELECT id FROM ai_models WHERE model_key = 'doubao-seed-1-6-250615' LIMIT 1")
+    ).first()
+    old_model_id = old_model_row[0] if old_model_row else None
+
+    llm_capabilities = [
+        "intent_parsing",
+        "copy_fusion",
+        "storyboard",
+        "copy_review",
+        "asset_classify",
+    ]
+
+    for cap_key in llm_capabilities:
+        if old_model_id:
+            conn.execute(
+                sa.text(
+                    "UPDATE ai_capability_configs SET primary_model_id = :new_id, updated_at = NOW() "
+                    "WHERE capability_key = :cap_key AND primary_model_id = :old_id"
+                ),
+                {"new_id": pro_model_id, "old_id": old_model_id, "cap_key": cap_key},
+            )
+
+    # Update image_analysis
+    qwen38_row = conn.execute(
+        sa.text("SELECT id FROM ai_models WHERE model_key = 'qwen3.8-flash' AND deleted_at IS NULL LIMIT 1")
+    ).first()
+    qwen37_row = conn.execute(
+        sa.text("SELECT id FROM ai_models WHERE model_key = 'qwen3.7-plus' AND deleted_at IS NULL LIMIT 1")
+    ).first()
+
+    if qwen38_row and qwen37_row:
+        qwen38_id = qwen38_row[0]
+        qwen37_id = qwen37_row[0]
+
+        current_ia = conn.execute(
+            sa.text(
+                "SELECT primary_model_id, lite_model_id, fallback_model_id "
+                "FROM ai_capability_configs WHERE capability_key = 'image_analysis'"
+            )
+        ).first()
+
+        if current_ia:
+            current_primary, current_lite, current_fallback = current_ia
+            updates = {}
+            if current_primary != qwen38_id:
+                updates["primary_model_id"] = qwen38_id
+            if current_lite != qwen38_id:
+                updates["lite_model_id"] = qwen38_id
+            if current_fallback != qwen37_id:
+                updates["fallback_model_id"] = qwen37_id
+
+            if updates:
+                set_clause = ", ".join([f"{k} = :{k}" for k in updates.keys()])
+                set_clause += ", updated_at = NOW()"
+                updates["cap_key"] = "image_analysis"
+                conn.execute(
+                    sa.text(f"UPDATE ai_capability_configs SET {set_clause} WHERE capability_key = :cap_key"),
+                    updates,
+                )
+
+
+def downgrade() -> None:
+    pass

alembic/versions/101_qwen_vl_plus_and_max_tokens.py

@@ -0,0 +1,153 @@
+# -*- coding: utf-8 -*-
+"""101: 补齐 qwen-vl-plus 视觉模型并修正 image_analysis 绑定与 max_tokens.
+
+背景：
+- qwen-vl-plus 做图片识别时返回 JSON 约 500-600 tokens，旧硬编码
+  max_tokens=350 导致 JSON 被截断、解析失败返回"未识别"。
+- 代码侧已移除硬编码，改由 capability 的 DB 配置决定 max_tokens。
+
+幂等：
+- qwen-vl-plus 已存在则不插入；
+- 仅当 image_analysis 当前 primary_model 不是 qwen-vl-plus 时才更新绑定，
+  避免覆盖后台手动配置。
+"""
+
+import sqlalchemy as sa
+
+from alembic import op
+
+revision = "101_qwen_vl_plus_and_max_tokens"
+down_revision = "100_fix_capability_model_bindings"
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    conn = op.get_bind()
+
+    models_table = conn.execute(sa.text("SELECT to_regclass('public.ai_models')")).scalar()
+    if not models_table:
+        return
+
+    caps_table = conn.execute(sa.text("SELECT to_regclass('public.ai_capability_configs')")).scalar()
+    if not caps_table:
+        return
+
+    # ── c. 补全其他 capability 的 max_tokens 默认值（幂等）──────────────────
+    # 放在 image_analysis 特定逻辑之前，确保任何分支 return 都不会跳过本段。
+    # 仅在当前值为 NULL 或过小 (<100) 时更新，不覆盖已有合理配置。
+    # embedding / tts / voice_clone 不走 chat 接口，无需设置。
+    default_max_tokens = {
+        "intent_parsing": 500,
+        "copy_fusion": 2500,
+        "storyboard": 4000,
+        "copy_review": 1000,
+        "asset_classify": 500,
+        "image_generation": 500,
+        "video_generation": 500,
+    }
+    for cap_key, mt in default_max_tokens.items():
+        conn.execute(
+            sa.text(
+                "UPDATE ai_capability_configs "
+                "SET max_tokens = :mt, updated_at = now() "
+                "WHERE capability_key = :key "
+                "AND (max_tokens IS NULL OR max_tokens < 100)"
+            ),
+            {"mt": mt, "key": cap_key},
+        )
+
+    # ── a. 确保 qwen-vl-plus 模型存在 ────────────────────────────────────────
+    conn.execute(sa.text("""
+            INSERT INTO ai_models (id, name, provider, model_key, api_key, api_base,
+                                   description, status, is_default, usage_today,
+                                   created_at, updated_at)
+            SELECT gen_random_uuid()::text,
+                   '通义千问VL Plus',
+                   'dashscope',
+                   'qwen-vl-plus',
+                   COALESCE(
+                       (SELECT api_key FROM ai_models
+                        WHERE provider = 'dashscope' AND deleted_at IS NULL
+                          AND api_key IS NOT NULL AND api_key != ''
+                        LIMIT 1),
+                       ''
+                   ),
+                   'https://dashscope.aliyuncs.com/compatible-mode/v1',
+                   '阿里云视觉理解模型（图片识别/分析）',
+                   'active', false, 0, now(), now()
+            WHERE NOT EXISTS (
+                SELECT 1 FROM ai_models
+                WHERE model_key = 'qwen-vl-plus' AND deleted_at IS NULL
+            )
+            """))
+
+    qwen_vl_row = conn.execute(
+        sa.text(
+            "SELECT id FROM ai_models WHERE model_key = 'qwen-vl-plus' "
+            "AND deleted_at IS NULL AND status = 'active' LIMIT 1"
+        )
+    ).first()
+    if not qwen_vl_row:
+        return
+    qwen_vl_id = qwen_vl_row[0]
+
+    qwen37_row = conn.execute(
+        sa.text(
+            "SELECT id FROM ai_models WHERE model_key = 'qwen3.7-plus' "
+            "AND deleted_at IS NULL AND status = 'active' LIMIT 1"
+        )
+    ).first()
+    qwen37_id = qwen37_row[0] if qwen37_row else None
+
+    # ── b. 仅当当前 primary 不是 qwen-vl-plus 时修正绑定与 max_tokens ───────
+    current = conn.execute(
+        sa.text(
+            "SELECT primary_model_id, lite_model_id, fallback_model_id, max_tokens "
+            "FROM ai_capability_configs WHERE capability_key = 'image_analysis'"
+        )
+    ).first()
+
+    if current is None:
+        # capability 不存在则创建
+        conn.execute(
+            sa.text("""
+                INSERT INTO ai_capability_configs
+                    (id, capability_key, capability_name, primary_model_id,
+                     lite_model_id, fallback_model_id, timeout_seconds,
+                     max_retries, max_tokens, concurrency, extra_params,
+                     is_enabled, created_at, updated_at)
+                VALUES (gen_random_uuid()::text, 'image_analysis', '图片分析',
+                        :primary, :primary, :fallback, 30, 1, 1000, 2,
+                        '{}'::jsonb, true, now(), now())
+                """),
+            {"primary": qwen_vl_id, "fallback": qwen37_id},
+        )
+        return
+
+    current_primary = current[0]
+    if current_primary == qwen_vl_id:
+        # 已经绑定 qwen-vl-plus：视为后台/数据迁移已处理，不覆盖任何配置
+        return
+
+    set_parts = [
+        "primary_model_id = :vl_id",
+        "lite_model_id = :vl_id",
+        "max_tokens = 1000",
+        "updated_at = now()",
+    ]
+    params: dict = {"vl_id": qwen_vl_id}
+    if qwen37_id is not None:
+        set_parts.insert(2, "fallback_model_id = :qwen37_id")
+        params["qwen37_id"] = qwen37_id
+
+    conn.execute(
+        sa.text(
+            "UPDATE ai_capability_configs SET " + ", ".join(set_parts) + " WHERE capability_key = 'image_analysis'"
+        ),
+        params,
+    )
+
+
+def downgrade() -> None:
+    pass

alembic/versions/102_image_analysis_max_tokens_1500.py

@@ -0,0 +1,36 @@
+# -*- coding: utf-8 -*-
+"""102: image_analysis max_tokens 1200 -> 1500.
+
+v6 prompt 更长、字段更多，旧 max_tokens 容易截断 JSON。
+仅在 image_analysis 当前 max_tokens < 1500 时更新（幂等，不覆盖后台已调到 >=1500 的配置）。
+"""
+
+import sqlalchemy as sa
+
+from alembic import op
+
+revision = "102_image_analysis_max_tokens_1500"
+down_revision = "101_qwen_vl_plus_and_max_tokens"
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    conn = op.get_bind()
+
+    caps_table = conn.execute(sa.text("SELECT to_regclass('public.ai_capability_configs')")).scalar()
+    if not caps_table:
+        return
+
+    conn.execute(
+        sa.text(
+            "UPDATE ai_capability_configs "
+            "SET max_tokens = 1500, updated_at = now() "
+            "WHERE capability_key = 'image_analysis' "
+            "AND (max_tokens IS NULL OR max_tokens < 1500)"
+        )
+    )
+
+
+def downgrade() -> None:
+    pass

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

@@ -1050,10 +1050,13 @@
   flex-direction: column;
   align-items: center;
   justify-content: center;
+  height: 360px;
   padding: 28px 16px;
   gap: 10px;
   background: #fff;
   border: 1px solid #e5e7eb;
+  border-radius: 10px;
+  margin-top: 8px;
 }
 .vv-copy-loading .vv-spinner {
   width: 28px;
@@ -1082,18 +1085,38 @@
 
 /* ── Storyboard (linear doc style) ── */
 .vv-storyboard {
-  padding: 6px 2px;
-  background: transparent;
-  border: none;
+  display: flex;
+  flex-direction: column;
+  height: 360px;
+  padding: 10px 12px;
+  background: #fff;
+  border: 1px solid #e5e7eb;
+  border-radius: 10px;
+  margin-top: 8px;
+  overflow: hidden;
 }
 
 .vv-sb-doc {
+  flex: 1 1 auto;
   display: flex;
   flex-direction: column;
   gap: 3px;
   color: #1f2937;
   font-size: 13px;
   line-height: 1.55;
+  overflow-y: auto;
+  padding-right: 4px;
+  margin-right: -4px;
+}
+.vv-sb-doc::-webkit-scrollbar {
+  width: 6px;
+}
+.vv-sb-doc::-webkit-scrollbar-thumb {
+  background: #d8c4ff;
+  border-radius: 3px;
+}
+.vv-sb-doc::-webkit-scrollbar-track {
+  background: transparent;
 }
 .vv-sb-h {
   margin: 6px 0 2px;
@@ -1383,12 +1406,14 @@
 /* 口播稿 —— 复用 vv-sb-field 样式，无额外需求 */
 
 .vv-sb-actions {
+  flex-shrink: 0;
   display: flex;
   align-items: center;
   gap: 8px;
   margin-top: 8px;
   padding-top: 8px;
   border-top: 1px solid #e5e7eb;
+  background: #fff;
 }
 .vv-sb-actions .vv-btn-ghost {
   padding: 6px 14px;

apps/worker/worker_app/tasks/viral_video.py

@@ -264,8 +264,8 @@ def _recover_stale_jobs() -> int:
             "SET status='failed', error_msg='任务执行超时，请重试', updated_at=:now "
             "WHERE status='running' "
             "  AND started_at IS NOT NULL AND started_at < :cutoff_start "
-            "  AND (heartbeat_at IS NULL OR heartbeat_at < :cutoff_beat) "
-            "  AND (heartbeat_at IS NOT NULL OR updated_at < :cutoff_beat)"
+            "  AND updated_at < :cutoff_beat "
+            "  AND (heartbeat_at IS NULL OR heartbeat_at < :cutoff_beat)"
         )
         result = ssn.execute(sql, {"now": now, "cutoff_start": cutoff_start, "cutoff_beat": cutoff_beat})
         ssn.commit()

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

@@ -170,7 +170,7 @@ def _is_v4_schema(fj: dict) -> bool:
     """判断是v4嵌套schema还是旧扁平schema"""
     return (
         isinstance(fj.get("products"), list)
-        or fj.get("type") in ("product", "store", "person", "other")
+        or fj.get("type") in ("product", "store", "person", "scene", "other")
         or isinstance(fj.get("people"), dict)
     )
 
@@ -418,8 +418,31 @@ def _assemble_v4(idx: int, fj: dict, ocr_texts: list[str]) -> dict[str, Any]:
 
     has_person = fj.get("has_person", False)
 
-    # 人物类
-    if vtype == "person" or has_person:
+    # 兼容老schema：无type字段或type不在已知枚举时，按has_person/products兜底
+    _KNOWN_V4_TYPES = ("product", "store", "person", "scene", "other")
+    if vtype not in _KNOWN_V4_TYPES:
+        if has_person:
+            vtype = "person"
+        elif products:
+            vtype = "product"
+        else:
+            vtype = "other"
+
+    # 人物信息提取辅助（scene/store分支有人物时追加到key_features）
+    def _extract_person_features(source: dict) -> list[str]:
+        parts = []
+        for pk in ("outfit_style", "upper_wear", "lower_wear", "dress_wear", "outerwear", "pose", "expression"):
+            pv = source.get(pk)
+            if pv and str(pv) not in ("null", None, "无法判断") and len(str(pv)) <= 40:
+                parts.append(f"人物:{pv}")
+        for pk2 in ("shoes", "bag", "hairstyle"):
+            pv2 = source.get(pk2)
+            if pv2 and str(pv2) not in ("null", None) and len(str(pv2)) <= 40:
+                parts.append(f"人物:{pv2}")
+        return parts
+
+    # ── 人物类 ──
+    if vtype == "person":
         # 取第一个人物信息（v5 schema人物信息在顶层）
         person_info = fj
         # 兼容people嵌套
@@ -479,6 +502,12 @@ def _assemble_v4(idx: int, fj: dict, ocr_texts: list[str]) -> dict[str, Any]:
             v = person_info.get(k)
             if v and v not in ("null", None):
                 kf.append(str(v))
+        env_obj = person_info.get("environment_objects") or []
+        if not isinstance(env_obj, list):
+            env_obj = [env_obj]
+        for o in env_obj:
+            if o and str(o) not in kf and len(str(o)) <= 40:
+                kf.append(str(o))
         if text_on_package:
             kf.append(f"文字:{'/'.join(text_on_package[:3])}")
         kf = kf[:8] or ["无法判断"]
@@ -539,9 +568,15 @@ def _assemble_v4(idx: int, fj: dict, ocr_texts: list[str]) -> dict[str, Any]:
             for s in sell[:2]:
                 if s and len(str(s)) <= 30 and str(s) not in kf:
                     kf.append(f"卖点:{s}")
+        env_obj = fj.get("environment_objects") or []
+        if not isinstance(env_obj, list):
+            env_obj = [env_obj]
+        for o in env_obj:
+            if o and str(o) not in kf and len(str(o)) <= 40:
+                kf.append(str(o))
         if text_on_package:
             kf.append(f"文字: {'/'.join(text_on_package[:3])}")
-        kf = kf[:6] or ["无法判断"]
+        kf = kf[:8] or ["无法判断"]
         portrait_prompt = _build_product_prompt_from_v4(main, fj)
         if brand != "无法判断" and brand not in name:
             summary = f"{brand} {name}"
@@ -563,57 +598,188 @@ def _assemble_v4(idx: int, fj: dict, ocr_texts: list[str]) -> dict[str, Any]:
             "_source": "v2_fast_json_v4",
         }
 
-    # 门店类或其他
+    # 门店类
     if vtype == "store":
         store_type = fj.get("store_type") or "店铺"
         name = store_type
         brand = fj.get("brand_signage") or "无法判断"
         category = "门店场景"
-        visual = fj.get("visual_elements") or []
-        if isinstance(visual, str):
-            visual = [visual]
-        atmosphere = fj.get("atmosphere") or mood
+        # appearance: store_layout + furnishings + 陈设色调
         appearance_parts = []
         if fj.get("store_layout"):
             appearance_parts.append(str(fj["store_layout"]))
-        if visual:
-            appearance_parts.append("、".join(str(v) for v in visual[:3]))
+        furnishings = fj.get("furnishings") or []
+        if isinstance(furnishings, list) and furnishings:
+            appearance_parts.append("陈设：" + "、".join(str(f) for f in furnishings[:4] if f))
         if fj.get("cleanliness"):
             appearance_parts.append(str(fj["cleanliness"]))
         appearance = "；".join(appearance_parts) if appearance_parts else "门店环境"
+        # key_features: 招牌细节+海报+外部物品+视觉元素+陈列商品+环境物件
         kf = []
-        if isinstance(visual, list):
-            kf.extend(str(v) for v in visual if v and len(str(v)) <= 30)
+        for arr_key in (
+            "signage_details",
+            "signage_posters",
+            "exterior_items",
+            "visual_elements",
+            "products_on_display",
+            "environment_objects",
+        ):
+            arr = fj.get(arr_key) or []
+            if not isinstance(arr, list):
+                arr = [arr]
+            for item in arr:
+                if item and str(item) not in kf and len(str(item)) <= 40:
+                    kf.append(str(item))
         prods_vis = fj.get("product_categories_visible") or []
         if isinstance(prods_vis, list):
-            kf.extend(str(c) for c in prods_vis[:3] if c)
+            for c in prods_vis[:3]:
+                if c and str(c) not in kf:
+                    kf.append(str(c))
         promo = fj.get("promotion_elements") or []
         if isinstance(promo, list) and promo:
             kf.append("促销活动:" + "、".join(str(p) for p in promo[:2]))
+        if has_person:
+            for pf in _extract_person_features(fj):
+                if pf not in kf:
+                    kf.append(pf)
         if text_on_package:
-            kf.append(f"文字: {'/'.join(text_on_package[:3])}")
-        kf = kf[:6] or ["门店场景"]
-        portrait_prompt = f"{brand if brand!='无法判断' else ''}{store_type}，{atmosphere}，{scene}场景，{('、'.join(color_names[:3])+'配色，') if color_names else ''}产品陈列丰富，门店实拍"
-        portrait_prompt = portrait_prompt.strip("，")
-        summary = f"{store_type}场景"
+            kf.append("文字:" + "/".join(text_on_package[:3]))
+        kf = kf[:8] or ["门店场景"]
+        atmosphere = fj.get("atmosphere") or fj.get("mood") or mood
+        # portrait_prompt: 品牌+store_type+主色+atmosphere+scene+核心视觉元素
+        pieces = []
+        if brand != "无法判断":
+            pieces.append(brand)
+        pieces.append(store_type)
+        if color_names:
+            pieces.append("、".join(color_names[:3]) + "配色")
+        if atmosphere:
+            pieces.append(atmosphere)
+        if scene and scene != "通用":
+            pieces.append(scene)
+        visual = fj.get("visual_elements") or []
+        if isinstance(visual, list) and visual:
+            pieces.append("、".join(str(v) for v in visual[:3] if v))
+        pieces.append("门店实拍")
+        portrait_prompt = "，".join(p for p in pieces if p)
+        if len(portrait_prompt) > 200:
+            portrait_prompt = portrait_prompt[:200].rstrip("，")
+        summary = f"{brand} {store_type}" if brand != "无法判断" else f"{store_type}场景"
         return {
             "name": name[:30],
             "brand": str(brand)[:30],
             "category": category,
-            "appearance": appearance[:200],
+            "appearance": appearance[:300],
             "packaging": "门店场景无包装",
             "text_on_package": text_on_package,
             "key_features": kf,
             "scene": scene,
-            "mood": atmosphere or mood,
+            "mood": atmosphere,
             "portrait_prompt": portrait_prompt[:200],
             "summary": summary[:40],
-            "has_person": False,
-            "_source": "v2_fast_json_v4",
+            "has_person": bool(fj.get("has_person", False)),
+            "_source": "v2_fast_json_v6_store",
         }
 
-    # other 兜底
-    desc = fj.get("description") or "未识别"
+    # 场景类（纯场景图，无产品/人物）
+    if vtype == "scene":
+        main_subject = fj.get("main_subject") or ""
+        scene_type = fj.get("scene_type") or "场景图"
+        name = main_subject or scene_type
+        brand = "无法判断"
+        category = scene_type or "场景图"
+        # appearance: 空间布局 + 关键元素
+        app_parts = []
+        if fj.get("spatial_layout"):
+            app_parts.append(str(fj["spatial_layout"]))
+        key_elements = fj.get("key_elements") or []
+        if not isinstance(key_elements, list):
+            key_elements = [key_elements]
+        if key_elements:
+            app_parts.append("关键元素：" + "、".join(str(k) for k in key_elements[:4] if k))
+        if fj.get("lighting"):
+            app_parts.append("光线：" + str(fj["lighting"]))
+        appearance = "；".join(app_parts) if app_parts else "场景环境"
+        # key_features: key_elements + environment_objects（去重，最多8）
+        kf = []
+        for k in key_elements:
+            if k and str(k) not in kf and len(str(k)) <= 40:
+                kf.append(str(k))
+        env_obj = fj.get("environment_objects") or []
+        if not isinstance(env_obj, list):
+            env_obj = [env_obj]
+        for o in env_obj:
+            if o and str(o) not in kf and len(str(o)) <= 40:
+                kf.append(str(o))
+        if has_person:
+            for pf in _extract_person_features(fj):
+                if pf not in kf:
+                    kf.append(pf)
+        if text_on_package:
+            kf.append("文字:" + "/".join(text_on_package[:3]))
+        kf = kf[:8] or ["场景元素"]
+        scene_name = fj.get("scene") or scene_type
+        scene_mood = fj.get("mood") or mood
+        packaging = "场景无包装"
+        # portrait_prompt: main_subject + key_elements + lighting + atmosphere + composition + style
+        pieces = []
+        if main_subject:
+            pieces.append(main_subject)
+        if key_elements:
+            pieces.append("、".join(str(k) for k in key_elements[:3] if k))
+        if fj.get("lighting"):
+            pieces.append(str(fj["lighting"]) + "光线")
+        if scene_mood:
+            pieces.append(scene_mood + "氛围")
+        if fj.get("composition"):
+            pieces.append(str(fj["composition"]))
+        if fj.get("style"):
+            pieces.append(str(fj["style"]))
+        pieces.append("场景实拍，摄影级画质")
+        portrait_prompt = "，".join(p for p in pieces if p)
+        if len(portrait_prompt) > 200:
+            portrait_prompt = portrait_prompt[:200].rstrip("，")
+        summary = main_subject or scene_type
+        return {
+            "name": str(name)[:30],
+            "brand": brand,
+            "category": str(category)[:20],
+            "appearance": appearance[:300],
+            "packaging": packaging,
+            "text_on_package": text_on_package,
+            "key_features": kf,
+            "scene": scene_name,
+            "mood": scene_mood,
+            "portrait_prompt": portrait_prompt[:200],
+            "summary": str(summary)[:40],
+            "has_person": bool(fj.get("has_person", False)),
+            "_source": "v2_fast_json_v6_scene",
+        }
+
+    # other 兜底：当 description 为空时，用 environment_objects/scene 拼基本描述
+    desc = fj.get("description") or ""
+    if not desc.strip():
+        env_obj = fj.get("environment_objects") or []
+        if not isinstance(env_obj, list):
+            env_obj = [env_obj]
+        env_names = [str(e) for e in env_obj if e and len(str(e)) <= 25][:5]
+        scene_text = fj.get("scene") or ""
+        if env_names:
+            desc = f"{scene_text}场景中" + "、".join(env_names) if scene_text else "、".join(env_names)
+        elif scene_text and scene_text != "通用":
+            desc = f"{scene_text}场景"
+        else:
+            desc = "未识别"
+    kf_other = []
+    if desc != "未识别":
+        kf_other.append(desc[:30])
+    env_obj = fj.get("environment_objects") or []
+    if not isinstance(env_obj, list):
+        env_obj = [env_obj]
+    for o in env_obj:
+        if o and str(o) not in kf_other and len(str(o)) <= 40:
+            kf_other.append(str(o))
+    kf_other = kf_other[:8] or ["无法判断"]
     return {
         "name": desc[:30],
         "brand": "无法判断",
@@ -621,7 +787,7 @@ def _assemble_v4(idx: int, fj: dict, ocr_texts: list[str]) -> dict[str, Any]:
         "appearance": desc[:200],
         "packaging": "无法判断",
         "text_on_package": text_on_package,
-        "key_features": [desc[:30]] if desc != "未识别" else ["无法判断"],
+        "key_features": kf_other,
         "scene": scene,
         "mood": mood,
         "portrait_prompt": f"{scene}，{mood}氛围，{desc}"[:200],

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

@@ -1,12 +1,13 @@
 # -*- coding: utf-8 -*-
-"""V2 兜底路径：image_analysis lite/fallback（默认 qwen3.7-plus / DashScope）单图调用。
+"""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。
 
 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
-- timeout=25s
+- 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
 """
 
@@ -22,7 +23,6 @@ from . import _prompt, assembler
 logger = logging.getLogger(__name__)
 
 _DEFAULT_TIMEOUT = 30
-_DEFAULT_MAX_TOKENS = 800
 
 
 def call_pro_vlm(
@@ -30,7 +30,9 @@ def call_pro_vlm(
     idx: int,
     *,
     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()
 
     try:
@@ -58,15 +60,17 @@ def call_pro_vlm(
     ]
 
     try:
-        raw = client.vision_completion(
-            messages=messages,
-            images=None,  # 图片已在 messages 中
-            temperature=0.3,
-            max_tokens=_DEFAULT_MAX_TOKENS,
-            timeout=timeout,
-            enable_thinking=False,
-            response_format={"type": "json_object"},
-        )
+        call_kwargs: dict[str, Any] = {
+            "messages": messages,
+            "images": None,  # 图片已在 messages 中
+            "temperature": 0.3,
+            "timeout": timeout,
+            "enable_thinking": False,
+            "response_format": {"type": "json_object"},
+        }
+        if max_tokens is not None:
+            call_kwargs["max_tokens"] = max_tokens
+        raw = client.vision_completion(**call_kwargs)
         elapsed = time.time() - t0
         if not raw:
             logger.warning("[vision.v2] pro 返回空 elapsed=%.1fs", elapsed)

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

@@ -1,5 +1,5 @@
 # -*- coding: utf-8 -*-
-"""V2 快速路径：image_analysis capability（默认 qwen3.8-flash / DashScope）强约束 JSON-only 调用。
+"""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。
 
 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：
@@ -7,8 +7,9 @@
 - 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=350、temperature=0.1（稳定输出 JSON）
-- timeout=12s（失败由外层走 pro 兜底）
+- 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）
 """
 
 from __future__ import annotations
@@ -23,7 +24,6 @@ from . import _prompt
 logger = logging.getLogger(__name__)
 
 _DEFAULT_TIMEOUT = 15
-_DEFAULT_MAX_TOKENS = 350
 
 
 def _strip_code_fence(s: str) -> str:
@@ -42,9 +42,13 @@ def call_fast_json(
     img_url: str,
     *,
     timeout: int = _DEFAULT_TIMEOUT,
-    max_tokens: int = _DEFAULT_MAX_TOKENS,
+    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。"""
+    """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:
@@ -72,15 +76,17 @@ def call_fast_json(
     ]
 
     try:
-        raw = client.vision_completion(
-            messages=messages,
-            images=None,  # 图片已在 messages 中
-            temperature=0.1,
-            max_tokens=max_tokens,
-            timeout=timeout,
-            enable_thinking=False,
-            response_format={"type": "json_object"},
-        )
+        call_kwargs: dict[str, Any] = {
+            "messages": messages,
+            "images": None,  # 图片已在 messages 中
+            "temperature": 0.1,
+            "timeout": timeout,
+            "enable_thinking": False,
+            "response_format": {"type": "json_object"},
+        }
+        if max_tokens is not None:
+            call_kwargs["max_tokens"] = max_tokens
+        raw = client.vision_completion(**call_kwargs)
         elapsed = time.time() - t0
         if not raw:
             logger.warning("[vision.v2] fast_json 返回空 elapsed=%.1fs", elapsed)

packages/application/cosyvoice_service.py

@@ -369,7 +369,7 @@ class CosyVoiceService:
         self._api_key = api_key or _router_key or settings.cosyvoice_api_key
         self._base_url = base_url or _router_url or settings.cosyvoice_base_url
         self._model = model or _router_model or settings.cosyvoice_model
-        self._clone_model = clone_model or getattr(settings, "cosyvoice_clone_model", "voice-enrollment")
+        self._clone_model = clone_model or getattr(settings, "cosyvoice_clone_model", "")
         self._audio_url_signer = audio_url_signer
 
         # base_url 规范化：去掉末尾的路径残留（兼容旧版配置）

packages/domain/viral_video.py

@@ -225,8 +225,11 @@ class ViralVideoJob:
             self.error_msg = ""
             self.completed_at = None
             self.heartbeat_at = None
+        _now = datetime.now(timezone.utc)
+        self.started_at = _now
+        self.heartbeat_at = _now
         self.status = ViralVideoStatus.RUNNING
-        self.updated_at = datetime.now(timezone.utc)
+        self.updated_at = _now
 
     def resume_from_copy_generated(self, edited_copy: str | None = None) -> None:
         """阶段2->阶段3：用户确认/编辑口播文案，开始跑 TTS+单次Seedance渲染。"""
@@ -235,14 +238,20 @@ class ViralVideoJob:
         if edited_copy and isinstance(self.copy_result, dict):
             self.copy_result = {**self.copy_result, "voiceover_script": edited_copy}
             self.generated_copy_text = edited_copy
+        _now = datetime.now(timezone.utc)
+        self.started_at = _now
+        self.heartbeat_at = _now
         self.status = ViralVideoStatus.RUNNING
-        self.updated_at = datetime.now(timezone.utc)
+        self.updated_at = _now
 
     def resume_from_confirm(self) -> None:
         if self.status != ViralVideoStatus.WAIT_USER_CONFIRM:
             raise ValueError(f"Cannot resume from {self.status}")
+        _now = datetime.now(timezone.utc)
+        self.started_at = _now
+        self.heartbeat_at = _now
         self.status = ViralVideoStatus.RUNNING
-        self.updated_at = datetime.now(timezone.utc)
+        self.updated_at = _now
 
     def mark_completed(self, video_url: str) -> None:
         self.status = ViralVideoStatus.COMPLETED

packages/shared/ai_client.py

@@ -183,15 +183,20 @@ class DoubaoClient:
         provider: str = "volcengine",
     ) -> None:
         settings = get_shared_settings()
-        self.api_key: str = api_key or settings.doubao_api_key
-        self.model: str = model or settings.doubao_model
-        self.base_url: str = (base_url or settings.doubao_base_url).rstrip("/")
+        self.provider: str = provider
+        if provider == "dashscope":
+            self.api_key: str = api_key or getattr(settings, "dashscope_api_key", "")
+            self.model: str = model or getattr(settings, "dashscope_model", "")
+            self.base_url: str = (base_url or getattr(settings, "dashscope_base_url", "")).rstrip("/")
+        else:  # volcengine (default)
+            self.api_key = api_key or settings.doubao_api_key
+            self.model = model or settings.doubao_model
+            self.base_url = (base_url or settings.doubao_base_url).rstrip("/")
         self.timeout: int = timeout or settings.doubao_timeout
         self.max_retries: int = max_retries or settings.doubao_max_retries
         self.max_tokens: int | None = max_tokens
         self.temperature: float | None = temperature
         self.extra_params: dict = extra_params or {}
-        self.provider: str = provider
         self.vision_model: str = settings.doubao_vision_model
         self.vision_lite_model: str = settings.doubao_vision_lite_model
         self.fast_model: str = settings.doubao_fast_model
@@ -255,7 +260,7 @@ class DoubaoClient:
         self,
         messages: list[dict[str, str]],
         temperature: float = 0.7,
-        max_tokens: int = 1024,
+        max_tokens: int | None = None,
         model: str | None = None,
         timeout: int | None = None,
         **kwargs,
@@ -265,7 +270,7 @@ class DoubaoClient:
         Args:
             messages: 对话消息列表，[{"role": "user"/"system"/"assistant", "content": "..."}]
             temperature: 采样温度，0-2，默认0.7
-            max_tokens: 最大生成token数，默认1024
+            max_tokens: 最大生成token数，默认 None（使用实例 self.max_tokens DB 配置，兜底 1024）
 
         Returns:
             模型返回的文本内容，失败返回 None
@@ -278,11 +283,12 @@ class DoubaoClient:
             "Authorization": f"Bearer {self.api_key}",
             "Content-Type": "application/json",
         }
+        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": max_tokens,
+            "max_tokens": effective_max_tokens,
         }
         # 合并实例级额外参数和调用方传入的额外参数
         if self.extra_params:
@@ -303,6 +309,22 @@ class DoubaoClient:
                 )
                 response.raise_for_status()
                 data = response.json()
+                finish_reason = (data.get("choices") or [{}])[0].get("finish_reason", "")
+                if finish_reason == "length" and attempt < self.max_retries:
+                    # 输出被 max_tokens 截断：1.5x 扩容后重试（计入 max_retries，不额外增加）
+                    old_max = int(payload["max_tokens"])
+                    new_max = int(old_max * 1.5)
+                    payload["max_tokens"] = new_max
+                    wait = 0.5 * (2**attempt)
+                    logger.warning(
+                        "输出被max_tokens截断(%d)，扩容到%d后重试 (第%d/%d次)",
+                        old_max,
+                        new_max,
+                        attempt + 1,
+                        self.max_retries + 1,
+                    )
+                    time.sleep(wait)
+                    continue
                 content = data["choices"][0]["message"]["content"]
                 _elapsed = time.time() - _t0
                 logger.info(
@@ -336,7 +358,7 @@ class DoubaoClient:
         self,
         messages: list[dict],
         images: list[str] | None = None,
-        max_tokens: int = 2048,
+        max_tokens: int | None = None,
         temperature: float = 0.3,
         timeout: int | None = None,
         model: str | None = None,
@@ -345,12 +367,12 @@ class DoubaoClient:
         """调用豆包视觉理解 API（OpenAI 兼容多模态格式）.
 
         将 images 附加到最后一条 user message 的 content 中，
-        使用 vision_model（默认 doubao-1-5-vision-pro-250328）。
+        使用构造函数传入的 self.model（DB capability 绑定的视觉模型，默认 qwen-vl-plus）。
 
         Args:
             messages: 对话消息列表。最后一条 user message 会被注入图片内容。
             images: 图片列表，支持 base64 data URI 或 HTTP(S) URL。
-            max_tokens: 最大生成 token 数，默认 2048。
+            max_tokens: 最大生成 token 数，默认 None（使用实例 self.max_tokens DB 配置，兜底 2048）。
             temperature: 采样温度，默认 0.3（视觉任务偏低更稳定）。
             timeout: 单次请求超时秒数，不传则使用默认 self.timeout。
 
@@ -390,11 +412,12 @@ class DoubaoClient:
             "Authorization": f"Bearer {self.api_key}",
             "Content-Type": "application/json",
         }
+        effective_max_tokens = max_tokens if max_tokens is not None else (self.max_tokens or 2048)
         payload: dict[str, Any] = {
-            "model": model or self.vision_model,
+            "model": model or self.model,
             "messages": vision_messages,
             "temperature": temperature,
-            "max_tokens": max_tokens,
+            "max_tokens": effective_max_tokens,
         }
         if self.extra_params:
             payload.update(self.extra_params)
@@ -414,6 +437,22 @@ class DoubaoClient:
                 )
                 response.raise_for_status()
                 data = response.json()
+                finish_reason = (data.get("choices") or [{}])[0].get("finish_reason", "")
+                if finish_reason == "length" and attempt < self.max_retries:
+                    # 视觉输出被 max_tokens 截断：1.5x 扩容后重试（计入 max_retries）
+                    old_max = int(payload["max_tokens"])
+                    new_max = int(old_max * 1.5)
+                    payload["max_tokens"] = new_max
+                    wait = 0.5 * (2**attempt)
+                    logger.warning(
+                        "视觉输出被max_tokens截断(%d)，扩容到%d后重试 (第%d/%d次)",
+                        old_max,
+                        new_max,
+                        attempt + 1,
+                        self.max_retries + 1,
+                    )
+                    time.sleep(wait)
+                    continue
                 content = data["choices"][0]["message"]["content"]
                 _elapsed = time.time() - _t0
                 logger.info(

packages/shared/ai_router.py

@@ -429,8 +429,8 @@ class AIRouter:
         api_key = getattr(settings, "cosyvoice_api_key", "")
         if not api_key:
             return None
-        base_url = getattr(settings, "cosyvoice_base_url", "https://dashscope.aliyuncs.com/api/v1")
-        model = getattr(settings, "cosyvoice_model", "cosyvoice-v3-flash")
+        base_url = getattr(settings, "cosyvoice_base_url", "")
+        model = getattr(settings, "cosyvoice_model", "")
 
         return TTSClient(provider="dashscope", api_key=api_key, base_url=base_url, model=model)
 
@@ -439,8 +439,8 @@ class AIRouter:
         api_key = getattr(settings, "doubao_api_key", "")
         if not api_key:
             return None
-        base_url = getattr(settings, "doubao_base_url", "https://ark.cn-beijing.volces.com/api/v3")
-        model = getattr(settings, "doubao_image_model", "doubao-seedream-5-0-flash-260915")
+        base_url = getattr(settings, "doubao_base_url", "")
+        model = getattr(settings, "doubao_image_model", "")
 
         return ImageGenClient(
             provider="volcengine",
@@ -455,8 +455,8 @@ class AIRouter:
         api_key = getattr(settings, "doubao_api_key", "")
         if not api_key:
             return None
-        base_url = getattr(settings, "doubao_base_url", "https://ark.cn-beijing.volces.com/api/v3")
-        model = getattr(settings, "doubao_video_model", "doubao-seedance-2-5-260628")
+        base_url = getattr(settings, "doubao_base_url", "")
+        model = getattr(settings, "doubao_video_model", "")
 
         return VideoGenClient(
             provider="volcengine",