A8多套打包+M4归档+R5多图:存量功能备份

A8 多套交付包(packaging_task.py):
- 修复交付包只打第1条混乱note的bug,按ImageCandidate.strategy分A/B/C组
- 每组生独立note_0N夹(6图+文案.txt),同seq留最新去重,老数据兼容
- task74端到端验:3套各6图,独立agent7项交叉验证全过

M4 归档(tasks.py/exports.py/前端):
- list_tasks加date_from/date_to/product_id筛选+product_name批量填(防N+1)
- 新增exports.py:产品JSON导出+标杆CSV导出(UTF-8 BOM)
- 前端HistoryFilters日期/产品筛选+产品列+打回原因红banner
- response.py加raise_param_error;独立agent验A1/A2/A9通过

R5 产品多图(product_images.py/020迁移/前端):
- product_images表+5端点(上传/列/改场景/设主图/删图)
- 生图按ROLE_SCENE_PREFERENCE选对应场景图,回落primary
- 前端ProductImageManager多图画廊

R6 账号config拆页(settings/):
- 配置页按角色拆/settings(运营+组长+admin)+/config(仅admin)
- Key只显末4位不显余额(守红线)

核销表对齐真实代码状态:D1改稿框/M7裂变/E12评图分纠偏为已完成(曾漏回写)

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

backend/alembic/versions/020_product_images_table.py

@@ -0,0 +1,47 @@
+"""create product_images table
+
+Revision ID: 020
+Revises: 019
+Create Date: 2026-06-18
+
+R5 产品多图：一产品多张参考图，每张标 scene 场景类型（primary/scene/
+texture/ingredient/model），生图按分镜 role 选用对应图，取不到回落主图。
+product.image_path 仍保留当主图（向后兼容，零破坏）。
+"""
+from alembic import op
+import sqlalchemy as sa
+
+revision = "020"
+down_revision = "019"
+branch_labels = None
+depends_on = None
+
+_SCENES = ("primary", "scene", "texture", "ingredient", "model")
+
+
+def upgrade():
+    op.create_table(
+        "product_images",
+        sa.Column("id", sa.BigInteger, primary_key=True, autoincrement=True),
+        sa.Column(
+            "product_id", sa.BigInteger,
+            sa.ForeignKey("products.id", ondelete="CASCADE"),
+            nullable=False,
+        ),
+        sa.Column("path", sa.String(512), nullable=False, comment="绝对路径，worker直接open"),
+        sa.Column(
+            "scene", sa.Enum(*_SCENES, name="productimagescene"),
+            nullable=False, server_default="primary",
+            comment="场景类型：白底主图/使用场景/质地/成分/上脸",
+        ),
+        sa.Column("is_primary", sa.Boolean, nullable=False, server_default=sa.false(),
+                  comment="主图（同步 product.image_path）"),
+        sa.Column("sort_order", sa.Integer, nullable=False, server_default="0"),
+        sa.Column("created_at", sa.DateTime, server_default=sa.func.now(), nullable=False),
+    )
+    op.create_index("idx_product_images_product_id", "product_images", ["product_id"])
+
+
+def downgrade():
+    op.drop_index("idx_product_images_product_id", table_name="product_images")
+    op.drop_table("product_images")

backend/app/api/v1/exports.py

@@ -0,0 +1,100 @@
+"""
+app/api/v1/exports.py — 客户原始输入物导出（数据归属红线）
+产品/标杆属客户 client_data，可导出（JSON/CSV）；飞轮偏好属平台不在此列。
+require_write_permission：仅 workspace 成员可导出本工作区数据。
+"""
+
+import csv
+import io
+import json
+import logging
+from typing import Annotated
+
+from fastapi import APIRouter, Depends, Query
+from fastapi.responses import StreamingResponse
+from sqlalchemy.orm import Session
+
+from app.core.database import get_db
+from app.middleware.workspace_guard import CurrentUser, require_write_permission
+from app.models.product import BenchmarkNote, Product
+
+logger = logging.getLogger(__name__)
+router = APIRouter(tags=["exports"])
+
+
+def _csv_response(rows: list[dict], fieldnames: list[str], filename: str) -> StreamingResponse:
+    """通用 CSV 流式响应。加 UTF-8 BOM 防 Windows Excel 中文乱码。"""
+    buf = io.StringIO()
+    buf.write("")  # BOM
+    writer = csv.DictWriter(buf, fieldnames=fieldnames, extrasaction="ignore")
+    writer.writeheader()
+    for r in rows:
+        writer.writerow(r)
+    buf.seek(0)
+    return StreamingResponse(
+        iter([buf.getvalue()]),
+        media_type="text/csv; charset=utf-8",
+        headers={"Content-Disposition": f'attachment; filename="{filename}"'},
+    )
+
+
+@router.get("/products/export")
+def export_products(
+    current_user: Annotated[CurrentUser, Depends(require_write_permission)] = None,
+    db: Session = Depends(get_db),
+):
+    """产品档案导出（JSON）。selling_points 还原为数组，非 JSON 字符串。"""
+    rows = (
+        db.query(Product)
+        .filter(Product.workspace_id == current_user.workspace_id, Product.is_active == True)
+        .order_by(Product.id)
+        .all()
+    )
+    data = [
+        {
+            "id": p.id,
+            "name": p.name,
+            "category": p.category,
+            "selling_points": json.loads(p.selling_points) if p.selling_points else [],
+            "brand_keyword": p.brand_keyword,
+            "target_audience": p.target_audience,
+            "style_tone": p.style_tone,
+            "created_at": p.created_at.isoformat(),
+        }
+        for p in rows
+    ]
+    # 直接返数组(非包 ok 信封)：导出供下载/二次处理，前端按 blob 存盘。
+    return data
+
+
+@router.get("/benchmarks/export")
+def export_benchmarks(
+    format: str = Query(default="json", pattern="^(json|csv)$"),
+    current_user: Annotated[CurrentUser, Depends(require_write_permission)] = None,
+    db: Session = Depends(get_db),
+):
+    """标杆笔记导出（JSON/CSV）。JOIN Product 填 product_name 便于客户阅读。"""
+    rows = (
+        db.query(BenchmarkNote, Product.name)
+        .outerjoin(Product, BenchmarkNote.product_id == Product.id)
+        .filter(BenchmarkNote.workspace_id == current_user.workspace_id)
+        .order_by(BenchmarkNote.id)
+        .all()
+    )
+    data = [
+        {
+            "id": b.id,
+            "product_id": b.product_id,
+            "product_name": pname,
+            "highlights": b.highlights or "",
+            "link_url": b.link_url or "",
+            "analyze_status": b.analyze_status,
+            "created_at": b.created_at.isoformat(),
+        }
+        for b, pname in rows
+    ]
+    if format == "csv":
+        fields = ["id", "product_id", "product_name", "highlights",
+                  "link_url", "analyze_status", "created_at"]
+        return _csv_response(data, fields, "benchmarks.csv")
+    return data

backend/app/api/v1/product_images.py

@@ -0,0 +1,121 @@
+"""
+app/api/v1/product_images.py — R5 产品多图管理（列/删/改场景/设主图）
+
+上传走 products.py 的 /products/{id}/upload-image（带 scene）。
+本文件管已上传图的增删改。product.image_path 始终同步当前主图（向后兼容）。
+"""
+import logging
+from typing import Annotated
+
+from fastapi import APIRouter, Depends
+from pydantic import BaseModel
+from sqlalchemy.orm import Session
+
+from app.constants.enums import ProductImageScene
+from app.core.database import get_db
+from app.core.response import ok, raise_not_found, raise_business
+from app.middleware.workspace_guard import CurrentUser, require_write_permission
+from app.models.product import Product, ProductImage
+
+logger = logging.getLogger(__name__)
+router = APIRouter(tags=["product-images"])
+
+
+class SceneUpdate(BaseModel):
+    scene: str
+
+
+def _get_product(db: Session, product_id: int, workspace_id: int) -> Product:
+    p = db.query(Product).filter(
+        Product.id == product_id, Product.workspace_id == workspace_id
+    ).first()
+    if not p:
+        raise_not_found("产品不存在")
+    return p
+
+
+def _get_image(db: Session, product_id: int, image_id: int) -> ProductImage:
+    img = db.query(ProductImage).filter(
+        ProductImage.id == image_id, ProductImage.product_id == product_id
+    ).first()
+    if not img:
+        raise_not_found("产品图不存在")
+    return img
+
+
+@router.get("/products/{product_id}/images")
+def list_images(
+    product_id: int,
+    current_user: Annotated[CurrentUser, Depends(require_write_permission)] = None,
+    db: Session = Depends(get_db),
+):
+    _get_product(db, product_id, current_user.workspace_id)
+    imgs = db.query(ProductImage).filter(
+        ProductImage.product_id == product_id
+    ).order_by(ProductImage.sort_order).all()
+    return ok([
+        {"id": im.id, "path": im.path, "scene": im.scene,
+         "is_primary": im.is_primary, "sort_order": im.sort_order}
+        for im in imgs
+    ])
+
+
+@router.put("/products/{product_id}/images/{image_id}/scene")
+def update_scene(
+    product_id: int, image_id: int, body: SceneUpdate,
+    current_user: Annotated[CurrentUser, Depends(require_write_permission)] = None,
+    db: Session = Depends(get_db),
+):
+    _get_product(db, product_id, current_user.workspace_id)
+    if body.scene not in {s.value for s in ProductImageScene}:
+        raise_business(f"非法场景类型 {body.scene}")
+    img = _get_image(db, product_id, image_id)
+    img.scene = body.scene
+    db.commit()
+    logger.info("product image scene updated: id=%s scene=%s", image_id, body.scene)
+    return ok({"id": image_id, "scene": body.scene})
+
+
+@router.put("/products/{product_id}/images/{image_id}/primary")
+def set_primary(
+    product_id: int, image_id: int,
+    current_user: Annotated[CurrentUser, Depends(require_write_permission)] = None,
+    db: Session = Depends(get_db),
+):
+    p = _get_product(db, product_id, current_user.workspace_id)
+    img = _get_image(db, product_id, image_id)
+    db.query(ProductImage).filter(
+        ProductImage.product_id == product_id, ProductImage.is_primary == True
+    ).update({"is_primary": False})
+    img.is_primary = True
+    p.image_path = img.path   # 同步主图字段（管道/校验读此字段）
+    db.commit()
+    logger.info("product primary image set: product_id=%s image_id=%s", product_id, image_id)
+    return ok({"product_id": product_id, "primary_image_id": image_id})
+
+
+@router.delete("/products/{product_id}/images/{image_id}")
+def delete_image(
+    product_id: int, image_id: int,
+    current_user: Annotated[CurrentUser, Depends(require_write_permission)] = None,
+    db: Session = Depends(get_db),
+):
+    p = _get_product(db, product_id, current_user.workspace_id)
+    img = _get_image(db, product_id, image_id)
+    was_primary = img.is_primary
+    db.delete(img)
+    db.flush()
+    # 删的是主图：把剩余 sort_order 最小的一张顶上来当主图，保证 image_path 不悬空
+    if was_primary:
+        nxt = db.query(ProductImage).filter(
+            ProductImage.product_id == product_id
+        ).order_by(ProductImage.sort_order).first()
+        if nxt:
+            nxt.is_primary = True
+            p.image_path = nxt.path
+        else:
+            p.image_path = None
+    db.commit()
+    logger.info("product image deleted: product_id=%s image_id=%s was_primary=%s",
+                product_id, image_id, was_primary)
+    return ok({"deleted": image_id})

backend/app/api/v1/products.py

@@ -7,14 +7,14 @@ category 是纯数据字段，不在代码里做枚举（基石A）。
 import logging
 from typing import Annotated
 
-from fastapi import APIRouter, Depends, UploadFile, File
+from fastapi import APIRouter, Depends, UploadFile, File, Form
 from pydantic import BaseModel
 from sqlalchemy.orm import Session
 
 from app.core.database import get_db
 from app.core.response import ok, paginate, raise_not_found
 from app.middleware.workspace_guard import CurrentUser, require_admin, require_write_permission
-from app.models.product import BannedWord, BenchmarkNote, Product
+from app.models.product import BannedWord, BenchmarkNote, Product, ProductImage
 
 logger = logging.getLogger(__name__)
 router = APIRouter(tags=["products"])
@@ -50,6 +50,7 @@ class BannedWordCreate(BaseModel):
 
 def _fmt_product(p: Product) -> dict:
     import json
+    imgs = sorted(getattr(p, "images", None) or [], key=lambda im: im.sort_order)
     return {
         "id": p.id, "name": p.name, "category": p.category,
         "source": p.source, "is_active": p.is_active,
@@ -58,6 +59,11 @@ def _fmt_product(p: Product) -> dict:
         "text_angles": json.loads(p.text_angles) if p.text_angles else [],
         "custom_prompt": p.custom_prompt,
         "image_path": p.image_path,
+        "images": [
+            {"id": im.id, "path": im.path, "scene": im.scene,
+             "is_primary": im.is_primary, "sort_order": im.sort_order}
+            for im in imgs
+        ],
         "brand_keyword": p.brand_keyword,       # 012: 套2字段暴露
         "target_audience": p.target_audience,   # 012: 套2字段暴露
         "created_at": p.created_at.isoformat(),
@@ -82,7 +88,7 @@ def list_products(
 @router.post("/products")
 def create_product(
     body: ProductCreate,
-    current_user: Annotated[CurrentUser, Depends(require_admin)] = None,
+    current_user: Annotated[CurrentUser, Depends(require_write_permission)] = None,
     db: Session = Depends(get_db),
 ):
     import json
@@ -118,7 +124,7 @@ def get_product(
 @router.put("/products/{product_id}")
 def update_product(
     product_id: int, body: ProductCreate,
-    current_user: Annotated[CurrentUser, Depends(require_admin)] = None,
+    current_user: Annotated[CurrentUser, Depends(require_write_permission)] = None,
     db: Session = Depends(get_db),
 ):
     import json
@@ -155,10 +161,28 @@ _ALLOWED_CONTENT_TYPES = {"image/jpeg", "image/png", "image/webp"}
 _MAX_SIZE_BYTES = 10 * 1024 * 1024   # 10 MB
 
 
+def _check_image_magic(data: bytes) -> bool:
+    """
+    校验文件头 magic number，防 .exe 改名 .jpg 伪造 content_type 绕过。
+    content_type 来自客户端可伪造，magic number 是真实字节，二者都要过。
+    JPEG: FF D8 FF / PNG: 89 50 4E 47 0D 0A 1A 0A / WebP: RIFF....WEBP
+    """
+    if len(data) < 12:
+        return False
+    if data[:3] == b"\xff\xd8\xff":                       # JPEG
+        return True
+    if data[:8] == b"\x89PNG\r\n\x1a\n":                  # PNG
+        return True
+    if data[:4] == b"RIFF" and data[8:12] == b"WEBP":     # WebP
+        return True
+    return False
+
+
 @router.post("/products/{product_id}/upload-image")
 async def upload_product_image(
     product_id: int,
     file: UploadFile = File(...),
+    scene: str = Form("primary"),
     current_user: Annotated[CurrentUser, Depends(require_write_permission)] = None,
     db: Session = Depends(get_db),
 ):
@@ -166,10 +190,12 @@ async def upload_product_image(
     上传产品参考图（铁律：生图必须带产品图）。
     文件类型：JPEG/PNG/WebP；大小上限 10 MB。
     存储路径：uploads/products/{workspace_id}/{product_id}/{filename}
-    写回 product.image_path，生图管道读此字段构建 reference_images。
+    R5多图：每次上传落一条 product_images（带 scene 场景类型）。
+    首张或 scene=primary 时同步写 product.image_path 当主图（向后兼容）。
     """
     from app.core.response import raise_business
     from app.core.config import get_settings
+    from app.constants.enums import ProductImageScene
     import os, uuid
 
     p = db.query(Product).filter(
@@ -179,12 +205,18 @@ async def upload_product_image(
     if not p:
         raise_not_found("产品不存在")
 
+    valid_scenes = {s.value for s in ProductImageScene}
+    if scene not in valid_scenes:
+        raise_business(f"非法场景类型 {scene}，仅支持 {sorted(valid_scenes)}")
+
     if file.content_type not in _ALLOWED_CONTENT_TYPES:
         raise_business(f"不支持的文件类型 {file.content_type}，仅支持 JPEG/PNG/WebP")
 
     data = await file.read()
     if len(data) > _MAX_SIZE_BYTES:
         raise_business("文件超过 10 MB 限制")
+    if not _check_image_magic(data):
+        raise_business("文件内容与扩展名不符（非真实 JPEG/PNG/WebP 图片）")
 
     settings = get_settings()
     ext = os.path.splitext(file.filename or "img.jpg")[1] or ".jpg"
@@ -197,10 +229,29 @@ async def upload_product_image(
     with open(save_path, "wb") as f:
         f.write(data)
 
-    p.image_path = save_path  # 绝对路径，worker 直接 open 可读
+    # 是否首张：决定要不要兜底设主图
+    existing = db.query(ProductImage).filter(
+        ProductImage.product_id == product_id
+    ).count()
+    is_primary = (existing == 0) or (scene == ProductImageScene.PRIMARY.value)
+    if is_primary:
+        # 同产品其它主图标记降级，保证唯一主图
+        db.query(ProductImage).filter(
+            ProductImage.product_id == product_id,
+            ProductImage.is_primary == True,
+        ).update({"is_primary": False})
+
+    img = ProductImage(
+        product_id=product_id, path=save_path, scene=scene,
+        is_primary=is_primary, sort_order=existing,
+    )
+    db.add(img)
+    if is_primary:
+        p.image_path = save_path  # 向后兼容：管道/校验仍读此字段当主图
     db.commit()
     db.refresh(p)
-    logger.info("product image uploaded: product_id=%s path=%s", product_id, save_path)
+    logger.info("product image uploaded: product_id=%s scene=%s primary=%s path=%s",
+                product_id, scene, is_primary, save_path)
     return ok(_fmt_product(p))
 
 
@@ -273,6 +324,8 @@ async def analyze_product_image(
     data = await file.read()
     if len(data) > _MAX_SIZE_BYTES:
         raise_business("文件超过 10 MB 限制")
+    if not _check_image_magic(data):
+        raise_business("文件内容与扩展名不符（非真实 JPEG/PNG/WebP 图片）")
 
     # key 解密（照抄 pipeline_steps.decrypt_user_key，基石B）
     api_key_row = db.query(UserApiKey).filter(

backend/app/api/v1/task_actions.py

@@ -58,7 +58,9 @@ def select_image(
     db: Session = Depends(get_db),
 ):
     """选图（飞轮信号 image_select +3）。"""
+    import json
     from app.services.flywheel_service import record_signal
+    from app.constants.enums import IMAGE_STRATEGY_ANGLE
     task = _check_task_ownership(
         db.query(GenerationTask).filter(GenerationTask.id == task_id).first(),
         current_user.workspace_id,
@@ -70,7 +72,14 @@ def select_image(
         raise_not_found("图片候选不存在")
     ic.is_selected = True
     db.commit()
-    record_signal(db, current_user, task, "image_select", candidate_id=ic.id)
+    # R7 断点2：选图带 strategy → 映射叙事角度标签进飞轮，与文案 angle_label 并轨；
+    # strategy 原值另存 signal_meta，便于后续按套别复盘。angle_label=None 不再空转。
+    angle = IMAGE_STRATEGY_ANGLE.get(ic.strategy or "")
+    meta = json.dumps({"strategy": ic.strategy, "role": ic.role}, ensure_ascii=False) if ic.strategy else None
+    record_signal(
+        db, current_user, task, "image_select",
+        candidate_id=ic.id, angle_label=angle, signal_meta=meta,
+    )
     return ok({"selected": body.candidate_id})
 
 
@@ -159,7 +168,14 @@ def get_preference_context(
         current_user.workspace_id,
     )
     from app.services.flywheel_service import get_preference_context
-    ctx = get_preference_context(db, current_user.workspace_id, task.product_id)
+    # 取产品档案供冷启动基线（与生产链同口径）
+    from app.models.product import Product
+    _p = db.query(Product).filter(Product.id == task.product_id).first()
+    product_dict = {
+        "custom_prompt": getattr(_p, "custom_prompt", "") or "",
+        "style_tone": getattr(_p, "style_tone", "") or "",
+    } if _p else {}
+    ctx = get_preference_context(db, current_user.workspace_id, task.product_id, product_dict)
     return ok(ctx)
 
 

backend/app/api/v1/tasks.py

@@ -170,16 +170,45 @@ def create_task(
 def list_tasks(
     page: int = 1, page_size: int = 20,
     status: list[str] | None = Query(default=None),
+    date_from: str | None = Query(default=None, description="起始日(YYYY-MM-DD)"),
+    date_to: str | None = Query(default=None, description="结束日(YYYY-MM-DD,含当日)"),
+    product_id: int | None = Query(default=None),
     current_user: Annotated[CurrentUser, Depends(require_write_permission)] = None,
     db: Session = Depends(get_db),
 ):
+    from datetime import datetime, timedelta
+    from app.models.product import Product
     q = db.query(GenerationTask).filter(GenerationTask.workspace_id == current_user.workspace_id)
     if status:
         # 支持多状态(?status=approved&status=rejected)，单值也兼容
         q = q.filter(GenerationTask.status.in_(status))
+    if product_id is not None:
+        q = q.filter(GenerationTask.product_id == product_id)
+    # 日期筛选：date 形参非法 → 40001(不静默吞成全量,防运营误判)
+    try:
+        if date_from:
+            q = q.filter(GenerationTask.created_at >= datetime.fromisoformat(date_from))
+        if date_to:
+            # date_to 含当日：取次日 0 点为开区间上界，覆盖当日全部时刻
+            _end = datetime.fromisoformat(date_to) + timedelta(days=1)
+            q = q.filter(GenerationTask.created_at < _end)
+    except ValueError:
+        from app.core.response import raise_param_error
+        raise_param_error("date_from/date_to 需为 YYYY-MM-DD 格式")
     total = q.count()
     items = q.order_by(GenerationTask.created_at.desc()).offset((page - 1) * page_size).limit(page_size).all()
-    return ok(paginate([_fmt_task(t) for t in items], total, page, page_size))
+    # product_name 一次性批量查(防 N+1)：收集本页 product_id → id→name dict
+    pids = {t.product_id for t in items if t.product_id}
+    name_map: dict[int, str] = {}
+    if pids:
+        for p in db.query(Product.id, Product.name).filter(Product.id.in_(pids)).all():
+            name_map[p.id] = p.name
+    rows = []
+    for t in items:
+        d = _fmt_task(t)
+        d["product_name"] = name_map.get(t.product_id)
+        rows.append(d)
+    return ok(paginate(rows, total, page, page_size))
 
 
 @router.get("/{task_id}")

backend/app/constants/enums.py

@@ -75,6 +75,16 @@ SIGNAL_WEIGHTS: dict[str, int] = {
 }
 
 
+# ── 生图3套正交叙事策略 → 飞轮角度标签 ───────────────────
+# 选图信号 angle_label 用此映射,与文案 angle_label 体系并轨进同一偏好聚合。
+# A痛点先行/B场景先行/C成分背书先行(倩倩姐2026-06-15起草,北哥过目版)。
+IMAGE_STRATEGY_ANGLE: dict[str, str] = {
+    "A": "痛点先行",
+    "B": "场景先行",
+    "C": "成分背书先行",
+}
+
+
 # ── 数据归属 ──────────────────────────────────────────
 class DataOwnership(str, Enum):
     CLIENT_DATA = "client_data"       # 原始输入产出，客户可导出
@@ -110,6 +120,15 @@ class ImageRole(str, Enum):
     MAIN = "main"
 
 
+# ── 产品参考图场景类型（R5多图：标注每张产品图拍的是什么，供生图按分镜role选用）──
+class ProductImageScene(str, Enum):
+    PRIMARY = "primary"        # 白底/正面主图（默认主图，必有）
+    SCENE = "scene"            # 使用场景图（梳妆台/生活环境）
+    TEXTURE = "texture"        # 质地特写（膏体/涂抹）
+    INGREDIENT = "ingredient"  # 成分/包装细节
+    MODEL = "model"            # 上脸/使用中
+
+
 # ── AI 图片提供商 ─────────────────────────────────────
 class ImageProvider(str, Enum):
     GPT = "gpt"

backend/app/core/config.py

@@ -42,7 +42,8 @@ class Settings(BaseSettings):
     CELERY_RESULT_BACKEND: str = ""
 
     # ── 并发限制（可配置，不写死）─────────────────────
-    MAX_CONCURRENT_TASKS_PER_USER: int = 2
+    # 倩倩姐红线：每用户并发上限 5 个任务（客户侧实际跑几个再观察调整）。
+    MAX_CONCURRENT_TASKS_PER_USER: int = 5
 
     # ── 文件存储路径 ──────────────────────────────────
     UPLOAD_BASE_PATH: str = "uploads/packages"

backend/app/core/response.py

@@ -59,5 +59,9 @@ def raise_business(message: str) -> None:
     raise CloverHTTPException(422, ErrorCode.BUSINESS_ERROR, message)
 
 
+def raise_param_error(message: str = "参数非法") -> None:
+    raise CloverHTTPException(400, ErrorCode.PARAM_INVALID, message)
+
+
 def raise_state_invalid(message: str = "状态机非法流转") -> None:
     raise CloverHTTPException(409, ErrorCode.STATE_INVALID, message)

backend/app/models/product.py

@@ -13,7 +13,7 @@ from sqlalchemy import (
 )
 from sqlalchemy.orm import Mapped, mapped_column, relationship
 
-from app.constants.enums import BannedWordLevel, ProductSource
+from app.constants.enums import BannedWordLevel, ProductImageScene, ProductSource
 from app.core.database import Base
 
 
@@ -52,12 +52,45 @@ class Product(Base):
     benchmark_notes: Mapped[list["BenchmarkNote"]] = relationship(
         back_populates="product", lazy="noload"
     )
+    images: Mapped[list["ProductImage"]] = relationship(
+        back_populates="product", lazy="selectin",
+        cascade="all, delete-orphan", order_by="ProductImage.sort_order",
+    )
 
     __table_args__ = (
         Index("idx_products_workspace_id", "workspace_id"),
     )
 
 
+class ProductImage(Base):
+    """产品参考图（R5多图）：一产品多张，每张标 scene 场景类型，生图按分镜role选用。
+
+    product.image_path 仍保留当主图（向后兼容，零破坏）；本表是其超集。
+    """
+    __tablename__ = "product_images"
+
+    id: Mapped[int] = mapped_column(BigInteger, primary_key=True, autoincrement=True)
+    product_id: Mapped[int] = mapped_column(
+        BigInteger, ForeignKey("products.id", ondelete="CASCADE"), nullable=False
+    )
+    path: Mapped[str] = mapped_column(String(512), nullable=False)  # 绝对路径，worker 直接 open
+    scene: Mapped[str] = mapped_column(
+        Enum(ProductImageScene, values_callable=lambda x: [e.value for e in x]),
+        default=ProductImageScene.PRIMARY, nullable=False,
+    )
+    is_primary: Mapped[bool] = mapped_column(default=False, nullable=False)  # 主图(同步 product.image_path)
+    sort_order: Mapped[int] = mapped_column(Integer, default=0, nullable=False)
+    created_at: Mapped[datetime] = mapped_column(
+        DateTime, server_default=func.now(), nullable=False
+    )
+
+    product: Mapped["Product"] = relationship(back_populates="images")
+
+    __table_args__ = (
+        Index("idx_product_images_product_id", "product_id"),
+    )
+
+
 class BenchmarkNote(Base):
     """标杆笔记（截图+手填亮点为主通道）"""
     __tablename__ = "benchmark_notes"

backend/app/services/ai_engine/_text_prompt.py

@@ -131,6 +131,38 @@ angle（本条角度标签）/ coverTitle（封面大字≤10字）/ imageBrief
 硬性格式：只输出JSON，不要markdown代码块，字符串内用中文引号「」。"""
 
 
+# ── 第2环标杆爆款配方 → 文案 prompt（借方法层结构，禁抄竞品原话）──────────────
+_BM_DIM_LABELS: list[tuple[str, str]] = [
+    ("title_formula", "标题公式"),
+    ("opening_hook", "开篇钩子"),
+    ("content_structure", "内容结构"),
+    ("selling_point_style", "卖点表达风格"),
+    ("emotion_tone", "情绪基调"),
+    ("topic_tags", "话题标签思路"),
+]
+
+
+def _build_benchmark_block(refs: list[dict]) -> str:
+    """把标杆 8维配方渲染成 prompt 块。只借'方法结构'，硬约束禁抄竞品品牌/功效原话。"""
+    if not refs:
+        return ""
+    lines: list[str] = []
+    for i, f in enumerate(refs[:3], 1):  # 最多取3条标杆，避免prompt膨胀
+        dims = [f"{label}=「{str(f.get(key, '')).strip()}」"
+                for key, label in _BM_DIM_LABELS if str(f.get(key, "")).strip()]
+        if dims:
+            lines.append(f"  标杆{i}：" + "；".join(dims))
+    if not lines:
+        return ""
+    body = "\n".join(lines)
+    return (
+        "\n【对标爆款配方参考（只借方法层结构，绝不照抄）】\n"
+        f"{body}\n"
+        "硬性约束：仅参考上面的标题套路/开篇方式/结构节奏/情绪基调来组织本产品文案；"
+        "禁止照搬竞品品牌名、产品名、功效原话或具体数字；产品信息一律以本产品为准。"
+    )
+
+
 def build_prompt(product: dict, count: int, extra_rules: str = "") -> str:
     """
     组装文案生成 user_prompt。
@@ -153,6 +185,7 @@ def build_prompt(product: dict, count: int, extra_rules: str = "") -> str:
 
     angle_hint = f"文案角度要覆盖：{'、'.join(angles)}（每条用不同角度）。" if angles else ""
     brand_rule = f"每条正文和标题中植入品牌词「{brand_kw}」一次（自然融入，不生硬）。" if brand_kw else ""
+    benchmark_block = _build_benchmark_block(product.get("benchmark_refs") or [])
 
     lines = [
         f"产品：{name}",
@@ -161,6 +194,7 @@ def build_prompt(product: dict, count: int, extra_rules: str = "") -> str:
         angle_hint,
         brand_rule,
         custom,
+        benchmark_block,
         f"\n【Q1随机变量池·每条身份/起因/小缺点各不相同，严格按下方分配使用】",
         combos_text,
         extra_rules,

backend/app/services/ai_engine/constants.py

@@ -70,6 +70,23 @@ PAGE_ROLES = [
 ]
 PAGE_ROLE_MAP = {r["role"]: r for r in PAGE_ROLES}
 
+# ── R5多图：分镜role → 优先产品图场景(scene) 偏好表 ──────────
+# 生图时按分镜 role 选该场景的产品图当参考；取不到回落主图(primary)。
+# scene 取值见 enums.ProductImageScene：primary/scene/texture/ingredient/model
+ROLE_SCENE_PREFERENCE = {
+    "hook":            ["scene", "primary"],       # 封面：真实生活场景优先
+    "pain_scene":      ["scene", "primary"],       # 痛点共鸣：使用前情境
+    "product_closeup": ["primary"],                # 单品特写：白底主图
+    "ingredient":      ["ingredient", "primary"],  # 成分拆解：成分/包装细节
+    "texture":         ["texture", "primary"],     # 质地展示：质地特写
+    "applied_proof":   ["model", "scene", "primary"],  # 上脸：上脸图/场景
+    "social_proof":    ["scene", "primary"],       # 社交背书：场景
+    "closer":          ["primary"],                # 促单收尾：主图
+    "scenario":        ["scene", "primary"],
+    "tutorial":        ["model", "scene", "primary"],
+}
+
+
 # ── 生图风格预设（扒 image.js STYLE_PROMPTS:26-29）──────────
 # 按 style 参数选小红书风格调性，注入 base_prompt 的"视觉风格"行
 STYLE_PROMPTS = {

backend/app/services/ai_engine/image_gen.py

@@ -14,13 +14,36 @@ import logging
 import os
 from typing import Any, Protocol
 
-from .constants import IMAGE_RETRY_ATTEMPTS, IMAGE_RETRY_BACKOFF_BASE, IMAGE_SIZE_DEFAULT
+from .constants import IMAGE_RETRY_ATTEMPTS, IMAGE_RETRY_BACKOFF_BASE, IMAGE_SIZE_DEFAULT, ROLE_SCENE_PREFERENCE
 from .image_scorer import score_image
 from .storyboard import plan_image_set, sanitize_text
 
 logger = logging.getLogger(__name__)
 
 
+def _pick_reference_for_role(
+    role: str,
+    images_by_scene: dict[str, list[bytes]] | None,
+    fallback: list[bytes] | None,
+) -> tuple[list[bytes] | None, str]:
+    """R5多图：按分镜 role 选该场景的产品图。取不到回落主图。
+
+    返回 (参考图bytes列表, 命中scene标签用于日志)。
+    """
+    if images_by_scene:
+        for scene in ROLE_SCENE_PREFERENCE.get(role, ["primary"]):
+            imgs = images_by_scene.get(scene)
+            if imgs:
+                return imgs, scene
+        # 偏好全落空：用任意可用图兜底（仍优先 primary）
+        if images_by_scene.get("primary"):
+            return images_by_scene["primary"], "primary"
+        for scene, imgs in images_by_scene.items():
+            if imgs:
+                return imgs, f"{scene}(兜底)"
+    return fallback, "fallback"
+
+
 class ImageClient(Protocol):
     """worker 注入的图片生成客户端协议（隔离 key 细节）"""
     async def gpt_edits(
@@ -129,12 +152,18 @@ async def generate_storyboard_images(
     strategy: str | None = None,
     target_role: str | None = None,
     custom_prompt: str | None = None,
+    images_by_scene: dict[str, list[bytes]] | None = None,
+    flywheel_fragment: str | None = None,
 ) -> list[dict]:
     """
     按 storyboard 逐张生图（asyncio.gather 并发），返回每张结果列表。
     strategy: None=默认叙事，'A'/'B'/'C'=三套正交叙事策略
     target_role: 非空时只生成该 role 那一张（R2 单张重生）
     custom_prompt: 非空时追加到每张 per_prompt 末尾（R2 人工提示词）
+    images_by_scene: R5多图，{scene: [bytes]}，按分镜 role 选对应场景图；
+                     为空则全分镜共用 reference_images（向后兼容）。
+    flywheel_fragment: R7 飞轮偏好片段（最近选图/打回真实信号聚合），注入图片
+                       排版偏好；仅影响文字角度/版式取向，绝不改瓶身（合规红线）。
     每项：{role, name, image_bytes, error}
     """
     plan = plan_image_set(note, product, image_count, analysis, strategy=strategy)
@@ -169,8 +198,20 @@ async def generate_storyboard_images(
         # R2 人工提示词：追加到末尾权重最高，但不覆盖前面合规/真实约束
         if custom_prompt:
             per_prompt += f"\n运营补充要求（在不违反上述合规与真实约束前提下尽量满足）：{sanitize_text(custom_prompt, 200)}。"
+        # R7 飞轮偏好：仅作用于文字角度/版式取向参考，绝不改瓶身（合规+真实红线）
+        if flywheel_fragment:
+            per_prompt += (
+                f"\n历史偏好参考（仅影响标题文字角度与排版取向，不得据此改动产品瓶身）："
+                f"{sanitize_text(flywheel_fragment, 300)}。"
+            )
         try:
-            img_bytes = await generate_one_image(client, per_prompt, reference_images)
+            # R5多图：按本张分镜 role 选对应场景产品图；无多图则共用 reference_images
+            ref_for_item, _scene_hit = _pick_reference_for_role(
+                item["role"], images_by_scene, reference_images
+            )
+            if images_by_scene:
+                logger.info("分镜 %s 选用产品图场景=%s", item["role"], _scene_hit)
+            img_bytes = await generate_one_image(client, per_prompt, ref_for_item)
             # 注：gpt-image-2 渲染中文偶发错别字(约1/6)。vision/OCR 文字校验闸门实测
             # 不可靠(漏报形近字+幻觉误伤品牌词),倩倩姐2026-06-16拍板先撤,纯生图,
             # 错别字作已知问题记录,后续迭代再处理。详见记忆 clover-image-text-check-shelved。

backend/app/services/ai_engine/preference_aggregator.py

@@ -52,7 +52,8 @@ def aggregate_preference_context(
         weight   = int(e.get("signal_weight", 1))
 
         # text_edit(改稿)是最强真实信号,角度按权重计入(倩倩姐2026-06-16拍板)
-        if sig_type in ("text_select", "approve", "text_edit") and angle:
+        # image_select(选图)按套别叙事角度计入,让选图偏好真正闭环回生图(R7断点2)
+        if sig_type in ("text_select", "approve", "text_edit", "image_select") and angle:
             angle_counts[angle] += weight
         elif sig_type == "reject_with_reason":
             reason = str(e.get("reason") or "").strip()

backend/app/services/fission_image_retry.py

@@ -60,6 +60,19 @@ def retry_fission_note_images(
         except Exception as e:  # noqa: BLE001
             logger.warning("裂变补图参考图读取失败，无参考图模式: %s", e)
 
+    # R5多图：补图也按场景分组，重生失败分镜时选对应场景图
+    images_by_scene: dict[str, list[bytes]] = {}
+    for _im in (product.get("images") or []):
+        _p = _resolve_image_path(_im.get("path", ""))
+        if _p and os.path.isfile(_p):
+            try:
+                with open(_p, "rb") as _f:
+                    images_by_scene.setdefault(_im.get("scene") or "primary", []).append(_f.read())
+            except Exception:  # noqa: BLE001
+                pass
+    if reference_images and not images_by_scene.get("primary"):
+        images_by_scene.setdefault("primary", []).extend(reference_images)
+
     upload_base = get_settings().UPLOAD_ABS_ROOT
     img_dir = os.path.join(upload_base, str(ft.workspace_id), f"fission_{ft.id}", str(fn.seq))
     os.makedirs(img_dir, exist_ok=True)
@@ -73,6 +86,7 @@ def retry_fission_note_images(
                 client=clients, note=note, product=product,
                 image_count=image_count, reference_images=reference_images,
                 target_role=role,
+                images_by_scene=images_by_scene or None,
             ))
         except Exception as exc:  # noqa: BLE001
             logger.error("裂变补图 seq=%s role=%s 仍失败: %s", fn.seq, role, exc)

backend/app/services/fission_images.py

@@ -39,6 +39,19 @@ def generate_fission_images(
         except Exception as e:  # noqa: BLE001
             logger.warning("裂变参考图读取失败，无参考图模式: %s", e)
 
+    # R5多图：裂变同样按场景分组，生图按分镜 role 选对应图
+    images_by_scene: dict[str, list[bytes]] = {}
+    for _im in (product.get("images") or []):
+        _p = _resolve_image_path(_im.get("path", ""))
+        if _p and os.path.isfile(_p):
+            try:
+                with open(_p, "rb") as _f:
+                    images_by_scene.setdefault(_im.get("scene") or "primary", []).append(_f.read())
+            except Exception:  # noqa: BLE001
+                pass
+    if reference_images and not images_by_scene.get("primary"):
+        images_by_scene.setdefault("primary", []).extend(reference_images)
+
     upload_base = get_settings().UPLOAD_ABS_ROOT
     for nid in note_ids:
         fn = db.query(FissionNote).filter(FissionNote.id == nid).first()
@@ -52,6 +65,7 @@ def generate_fission_images(
             results = asyncio.run(generate_storyboard_images(
                 client=clients, note=note, product=product,
                 image_count=image_count, reference_images=reference_images,
+                images_by_scene=images_by_scene or None,
             ))
         except Exception as exc:  # noqa: BLE001
             logger.error("裂变套 seq=%s 生图失败: %s", fn.seq, exc)

backend/app/services/flywheel_service.py

@@ -8,10 +8,10 @@ preference_aggregator：查最近50条 → 最常选角度 + 打回原因近3条
 import logging
 from typing import Any
 
-from sqlalchemy import desc, func
+from sqlalchemy import desc
 from sqlalchemy.orm import Session
 
-from app.constants.enums import SIGNAL_WEIGHTS, DataOwnership, SignalType
+from app.constants.enums import SIGNAL_WEIGHTS, DataOwnership
 from app.middleware.workspace_guard import CurrentUser
 from app.models.flywheel import PreferenceEvent
 from app.models.task import GenerationTask
@@ -20,8 +20,6 @@ logger = logging.getLogger(__name__)
 
 # 实时聚合窗口：最近50条事件
 _AGGREGATION_WINDOW = 50
-# 冷启动阈值：不足5条信号用产品档案冷启动
-_COLD_START_THRESHOLD = 5
 
 
 def record_signal(
@@ -32,12 +30,14 @@ def record_signal(
     candidate_id: int | None = None,
     angle_label: str | None = None,
     reason: str | None = None,
+    signal_meta: str | None = None,
 ) -> None:
     """
     写入飞轮信号。
     workspace_id + product_id 都必须有（基石C + 按产品分开学）。
     signal_weight 用枚举默认值，北哥可校准。
     data_ownership 默认 client_data（选择行为归客户）。
+    signal_meta：JSON 扩展（如选图存 strategy），不参与角度聚合主逻辑。
     """
     weight = SIGNAL_WEIGHTS.get(signal_type, 0)
     event = PreferenceEvent(
@@ -50,6 +50,7 @@ def record_signal(
         candidate_id=candidate_id,
         angle_label=angle_label,
         reason=reason,
+        signal_meta=signal_meta,
         data_ownership=DataOwnership.CLIENT_DATA,
     )
     try:
@@ -69,14 +70,17 @@ def record_signal(
 
 
 def get_preference_context(
-    db: Session, workspace_id: int, product_id: int
+    db: Session, workspace_id: int, product_id: int,
+    product_dict: dict[str, Any] | None = None,
 ) -> dict[str, Any]:
     """
-    实时聚合偏好上下文（最近50条 events）。
+    实时聚合偏好上下文（最近50条 events）供前端展示。
-    返回：recent_preference摘要 + reject_reasons近3条 + injected_count。
+    R7断点3：统一口径——委托给生产链同一个 aggregate_preference_context（权重口径），
-    不足5条 → 冷启动提示（产品档案兜底，由 AIE prompt 层读 products.custom_prompt）。
+    消除"前端展示按次数/生成按权重"的口径分裂（说一套做一套）。
     按 workspace_id + product_id 严格过滤（不串数据，基石C）。
     """
+    from app.services.ai_engine.preference_aggregator import aggregate_preference_context
+
     recent = (
         db.query(PreferenceEvent)
         .filter(
@@ -87,35 +91,18 @@ def get_preference_context(
         .limit(_AGGREGATION_WINDOW)
         .all()
     )
-
+    events_dicts = [
-    if len(recent) < _COLD_START_THRESHOLD:
+        {"signal_type": e.signal_type, "workspace_id": e.workspace_id,
+         "product_id": e.product_id, "angle_label": e.angle_label or "",
+         "signal_weight": e.signal_weight, "reason": e.reason or ""}
+        for e in recent
+    ]
+    ctx = aggregate_preference_context(
+        events_dicts, product_dict or {}, workspace_id, product_id
+    )
+    # 前端展示不需要 prompt_fragment（注入用），剥掉只回摘要+原因+计数
     return {
-            "recent_preference": "信号不足，使用产品档案基线（冷启动）",
+        "recent_preference": ctx.get("recent_preference", ""),
-            "reject_reasons": [],
+        "reject_reasons": ctx.get("reject_reasons", []),
-            "injected_count": len(recent),
+        "injected_count": ctx.get("injected_count", 0),
-        }
-
-    # 统计最常被选中的角度（text_edit 改稿=最强真实信号,按权重计入,倩倩姐2026-06-16）
-    angle_counts: dict[str, int] = {}
-    for ev in recent:
-        if ev.signal_type in (SignalType.TEXT_SELECT, SignalType.APPROVE, SignalType.TEXT_EDIT) and ev.angle_label:
-            angle_counts[ev.angle_label] = angle_counts.get(ev.angle_label, 0) + 1
-
-    top_angles = sorted(angle_counts.items(), key=lambda x: x[1], reverse=True)[:3]
-    if top_angles:
-        pref_desc = "；".join(f"{a}（已选{c}次）" for a, c in top_angles)
-        preference_summary = f"最近偏好：{pref_desc}"
-    else:
-        preference_summary = "暂无明显角度偏好"
-
-    # 取最近3条打回原因原文（不做 AI 归纳，契约§3）
-    reject_reasons = [
-        ev.reason for ev in recent
-        if ev.signal_type == SignalType.REJECT_WITH_REASON and ev.reason
-    ][:3]
-
-    return {
-        "recent_preference": preference_summary,
-        "reject_reasons": reject_reasons,
-        "injected_count": len(recent),
     }

backend/app/services/task_service.py

@@ -30,6 +30,28 @@ def _check_user_has_key(db: Session, user_id: int, workspace_id: int) -> None:
         raise_business("尚未配置 API Key，请先在设置中录入")
 
 
+def _check_concurrency_limit(db: Session, user_id: int, workspace_id: int) -> None:
+    """
+    校验该用户未完成任务数未超并发上限（红线=5,可配置)。
+    只算 pending/generating(真占 worker),挑选/审核态不计。超限引导稍后再试。
+    """
+    from app.core.config import get_settings
+    from app.constants.enums import TaskStatus
+
+    limit = get_settings().MAX_CONCURRENT_TASKS_PER_USER
+    running = (
+        db.query(GenerationTask)
+        .filter(
+            GenerationTask.operator_id == user_id,
+            GenerationTask.workspace_id == workspace_id,
+            GenerationTask.status.in_([TaskStatus.PENDING.value, TaskStatus.GENERATING.value]),
+        )
+        .count()
+    )
+    if running >= limit:
+        raise_business(f"您有 {running} 个任务正在生成，已达并发上限 {limit} 个，请等待完成后再发起")
+
+
 def create_generation_task(
     db: Session,
     current_user: CurrentUser,
@@ -42,6 +64,9 @@ def create_generation_task(
     if body.track == "ai":
         # 轨A：先检查有没有 key
         _check_user_has_key(db, current_user.user_id, current_user.workspace_id)
+        # 并发上限：只算正在消耗生成资源的任务(pending/generating)，
+        # 已生成完等挑选/审核的不占 worker。红线=每用户5个(可配置)。
+        _check_concurrency_limit(db, current_user.user_id, current_user.workspace_id)
 
     # 禁降级铁律：本次产品入镜(need_product_image=True)时，产品必须已上传参考图，
     # 否则拒绝建任务（不允许降级纯文生图，防产品包装跑偏/过抽检失败）。
@@ -57,6 +82,11 @@ def create_generation_task(
         if not (product.image_path or "").strip():
             raise_business("该产品未上传参考图，无法生成产品入镜内容；请先到产品库上传产品图，或关闭「产品入镜」开关")
 
+    # 第2环：关联标杆笔记ID存库（JSON list）。pipeline 据此读 features_json 注入文案 prompt。
+    import json
+    _bids = getattr(body, "benchmark_ids", None) or []
+    benchmark_ids_json = json.dumps([int(i) for i in _bids]) if _bids else None
+
     task = GenerationTask(
         workspace_id=current_user.workspace_id,
         product_id=body.product_id,
@@ -66,6 +96,7 @@ def create_generation_task(
         image_count=body.image_count,
         track=body.track,
         need_product_image=need_img,
+        benchmark_ids=benchmark_ids_json,
         status="pending",
     )
     db.add(task)

backend/app/workers/packaging_task.py

@@ -42,47 +42,62 @@ def build_delivery_package(self, package_id: int) -> dict:
         settings = get_settings()
         upload_base = settings.UPLOAD_BASE_PATH.rstrip("/")
 
-        selected_text = db.query(TextCandidate).filter(
+        # A8 多套打包：按 strategy(A/B/C 三套正交叙事)分组，每套成一篇独立 note。
+        # 北哥拿到的交付包含完整 3 套（note_01/02/03），不再只打第 1 条文案。
+        selected_texts = db.query(TextCandidate).filter(
             TextCandidate.task_id == task_id, TextCandidate.is_selected == True,
-        ).first()
+        ).order_by(TextCandidate.id).all()
-        # 整套全打（倩倩姐2026-06-08拍板）：一条笔记的全部图按 seq 排序进包，
+        if not selected_texts:
-        # 不再只打 is_selected 的封面。北哥6张标准套 seq=1 是 hook 封面，天然排第一。
-        selected_images = db.query(ImageCandidate).filter(
-            ImageCandidate.task_id == task_id,
-        ).order_by(ImageCandidate.seq).all()
-
-        if not selected_text:
             raise ValueError("无已选文案，请先选择文案")
 
-        text_data = json.loads(selected_text.content or "{}")
+        # 整套全打：一套内全部图按 seq 排序进包，不只打封面。重生场景同 (strategy,seq)
-        images_data = []
+        # 可能多条(新增不删旧)，去重取最新(id最大)，避免包内重复图。
-        for ic in selected_images:
+        all_images = db.query(ImageCandidate).filter(
+            ImageCandidate.task_id == task_id,
+        ).order_by(ImageCandidate.strategy, ImageCandidate.seq).all()
+
+        def _read_image(ic) -> dict:
             img_bytes = b""
             if ic.url:
-                # url 形如 /uploads/ws/task/file.jpg，本身已含 uploads 前缀。
+                # url 已含 uploads 前缀；工作目录 /app，lstrip 当相对路径读，勿再拼 base(防 uploads/uploads)
-                # 工作目录是 /app，直接 lstrip("/") 当相对路径读，不能再拼 upload_base(会重复 uploads/uploads)。
-                rel = ic.url.lstrip("/")
-                abs_path = rel
                 try:
-                    with open(abs_path, "rb") as f:
+                    with open(ic.url.lstrip("/"), "rb") as f:
                         img_bytes = f.read()
                 except OSError as e:
-                    logger.warning("图片读取失败，跳过：%s %s", abs_path, e)
+                    logger.warning("图片读取失败，跳过：%s %s", ic.url, e)
-            images_data.append({
+            return {
                 "seq": ic.seq,
                 "role": ic.role.value if hasattr(ic.role, "value") else str(ic.role),
                 "data": img_bytes,
-            })
+            }
 
-        notes = [{
+        # 按 strategy 分组（A/B/C；老数据 strategy=None 归一套，向后兼容）
+        from collections import OrderedDict
+        groups: "OrderedDict[str, dict]" = OrderedDict()
+        for ic in all_images:
+            slot = groups.setdefault(ic.strategy or "_", {})
+            prev = slot.get(ic.seq)
+            if prev is None or ic.id > prev.id:
+                slot[ic.seq] = ic  # 同 seq 留最新
+
+        notes = []
+        for idx, (_strategy, slot) in enumerate(groups.items()):
+            images_data = [_read_image(slot[k]) for k in sorted(slot)]
+            # 文案配对：选中文案数≥套数则一套一条；否则各套共用第 1 条
+            # (图均以第 1 条文案为语境生成，共用合理；多选则尊重运营按套选的文案)
+            tc = selected_texts[idx] if idx < len(selected_texts) else selected_texts[0]
+            text_data = json.loads(tc.content or "{}")
+            notes.append({
                 "title": text_data.get("title", ""),
                 "content": text_data.get("content", ""),
                 "tags": text_data.get("tags", []),
                 "images": images_data,
-            "banned_word_status": (selected_text.banned_word_status.value
+                "banned_word_status": (tc.banned_word_status.value
-                                   if hasattr(selected_text.banned_word_status, "value")
+                                       if hasattr(tc.banned_word_status, "value")
-                                   else str(selected_text.banned_word_status)),
+                                       else str(tc.banned_word_status)),
-        }]
+            })
+        if not notes:
+            raise ValueError("无图片候选，无法打包")
 
         from app.services.ai_engine.package_exporter import build_delivery_package as do_build
         # 打包产物放专用目录 uploads/packages/，与图片目录 uploads/{ws}/{task}/ 分开

backend/app/workers/pipeline_io.py

@@ -148,7 +148,8 @@ def run_image_generation(db, clients, task, product_dict: dict,
                          first_copy: dict, upload_base_path: str,
                          regen_strategy: str | None = None,
                          regen_role: str | None = None,
-                         custom_prompt: str | None = None) -> int:
+                         custom_prompt: str | None = None,
+                         flywheel_fragment: str | None = None) -> int:
     """
     Step6+7+8(image): 调 generate_storyboard_images → 后处理 → 存 ImageCandidate → 推 SSE。
     返回 next_seq。
@@ -202,6 +203,23 @@ def run_image_generation(db, clients, task, product_dict: dict,
             "拒绝降级纯文生图。请确认产品已上传参考图。"
         )
 
+    # R5多图：按场景分组加载产品图，生图按分镜 role 选对应场景图
+    images_by_scene: dict[str, list[bytes]] = {}
+    for _im in (product_dict.get("images") or []):
+        _p = _resolve_image_path(_im.get("path", ""))
+        _scene = _im.get("scene") or "primary"
+        if _p and os.path.isfile(_p):
+            try:
+                with open(_p, "rb") as _f:
+                    images_by_scene.setdefault(_scene, []).append(_f.read())
+            except Exception as _e:
+                logger.warning("产品图(scene=%s)读取失败，跳过：%s %s", _scene, _p, _e)
+    # 主图始终保底进 primary（多图表为空或主图未入表时仍可用）
+    if reference_images and not images_by_scene.get("primary"):
+        images_by_scene.setdefault("primary", []).extend(reference_images)
+    if images_by_scene:
+        logger.info("R5多图已加载：%s", {k: len(v) for k, v in images_by_scene.items()})
+
     seq = seq_start
     # R2: 限定重生套别(regen_strategy)则只跑该套，否则全量 A/B/C 三套正交叙事
     _strategies = (regen_strategy,) if regen_strategy else ("A", "B", "C")
@@ -227,6 +245,8 @@ def run_image_generation(db, clients, task, product_dict: dict,
                 strategy=strategy,
                 target_role=regen_role,
                 custom_prompt=custom_prompt,
+                images_by_scene=images_by_scene or None,
+                flywheel_fragment=flywheel_fragment,
             ))
         except Exception as exc:
             img_success = False

backend/app/workers/pipeline_steps.py

@@ -58,6 +58,7 @@ def build_clients_and_clear_key(plain_key: str):
 def build_product_dict(product) -> dict:
     """把 ORM product 转成 AI 引擎所需的 dict（不含任何 key）。"""
     return {
+        "id": product.id,
         "name": product.name,
         "category": product.category or "通用好物",
         "selling_points": json.loads(product.selling_points or "[]"),
@@ -66,10 +67,52 @@ def build_product_dict(product) -> dict:
         "custom_prompt": product.custom_prompt or "",
         "brand_keyword": product.brand_keyword or "",       # S3: 品牌词透传进生成prompt(每条植入)
         "target_audience": product.target_audience or "",   # 012: 人群透传进storyboard/文案prompt
-        "image_path": product.image_path or "",             # 产品参考图路径（前端上传后填入）
+        "image_path": product.image_path or "",             # 产品参考图路径（主图，向后兼容）
+        # R5多图：每张产品图 {path, scene}，生图按分镜role选对应场景图
+        "images": [
+            {"path": im.path, "scene": im.scene}
+            for im in (getattr(product, "images", None) or [])
+        ],
+        # 第2环标杆配方，默认空；走 AI 主链时由 load_benchmark_features 覆盖填充
+        "benchmark_refs": [],
     }
 
 
+def load_benchmark_features(db, task, workspace_id: int) -> list[dict]:
+    """
+    第2环→第5环接线：读 task.benchmark_ids → 查 analyze_status=done 的标杆 features_json。
+    返回 8维配方 dict 列表（供 build_prompt 借方法层结构，禁抄竞品品牌/功效原话）。
+    未选/未分析完/解析失败都安全返空，绝不阻断生成。
+    """
+    from app.models.product import BenchmarkNote
+
+    raw_ids = getattr(task, "benchmark_ids", None)
+    if not raw_ids:
+        return []
+    try:
+        ids = [int(i) for i in json.loads(raw_ids)]
+    except Exception:
+        return []
+    if not ids:
+        return []
+
+    rows = db.query(BenchmarkNote).filter(
+        BenchmarkNote.id.in_(ids),
+        BenchmarkNote.workspace_id == workspace_id,
+        BenchmarkNote.analyze_status == "done",
+    ).all()
+
+    feats: list[dict] = []
+    for b in rows:
+        if not b.features_json:
+            continue
+        try:
+            feats.append(json.loads(b.features_json))
+        except Exception:
+            logger.warning("标杆 features_json 解析失败 id=%s", b.id)
+    return feats
+
+
 def load_flywheel_context(db, workspace_id: int, product_id: int, product_dict: dict) -> tuple[str, dict]:
     """
     查最近50条飞轮事件，聚合偏好上下文。

backend/app/workers/tasks.py

@@ -106,6 +106,7 @@ def run_generation_pipeline(self, task_id: int, regen_strategy: str | None = Non
             build_clients_and_clear_key,
             build_product_dict,
             load_flywheel_context,
+            load_benchmark_features,
         )
         from app.workers.pipeline_io import run_text_generation, run_image_generation
         from app.core.config import get_settings
@@ -135,6 +136,8 @@ def run_generation_pipeline(self, task_id: int, regen_strategy: str | None = Non
         }, seq)
 
         product_dict = build_product_dict(product)
+        # 第2环爆款配方接进文案链：选中并分析完的标杆 8维配方注入 build_prompt（借方法层结构）
+        product_dict["benchmark_refs"] = load_benchmark_features(db, task, workspace_id)
         flywheel_fragment, flywheel_ctx = load_flywheel_context(db, workspace_id, task.product_id, product_dict)
 
         if flywheel_fragment:
@@ -192,6 +195,7 @@ def run_generation_pipeline(self, task_id: int, regen_strategy: str | None = Non
             first_copy, settings.UPLOAD_BASE_PATH,
             regen_strategy=regen_strategy, regen_role=regen_role,
             custom_prompt=custom_prompt,
+            flywheel_fragment=flywheel_fragment,
         )
 
         # 最终状态 + task_done

backend/main.py

@@ -71,6 +71,7 @@ def _register_routers(app: FastAPI) -> None:
     from app.api.v1.auth import router as auth_router
     from app.api.v1.api_keys import router as api_keys_router
     from app.api.v1.products import router as products_router
+    from app.api.v1.product_images import router as product_images_router
     from app.api.v1.benchmarks import router as benchmarks_router
     from app.api.v1.tasks import router as tasks_router
     from app.api.v1.task_actions import router as task_actions_router
@@ -79,11 +80,15 @@ def _register_routers(app: FastAPI) -> None:
     from app.api.v1.stream import router as stream_router
     from app.api.v1.workspaces import router as workspaces_router
     from app.api.v1.fission import router as fission_router
+    from app.api.v1.exports import router as exports_router
 
     prefix = "/api/v1"
     app.include_router(auth_router, prefix=prefix)
     app.include_router(api_keys_router, prefix=prefix)
+    # 导出 router 必须在 products_router 之前：否则 /products/export 被 /products/{product_id} 吞掉
+    app.include_router(exports_router, prefix=prefix)
     app.include_router(products_router, prefix=prefix)
+    app.include_router(product_images_router, prefix=prefix)
     app.include_router(benchmarks_router, prefix=prefix)
     app.include_router(tasks_router, prefix=prefix)
     app.include_router(task_actions_router, prefix=prefix)

backend/tests/test_flywheel_wiring_p34.py

@@ -156,7 +156,10 @@ class TestPoint4SseFlywheelEvent:
         mq.limit.return_value = mq
         mq.all.return_value = orm_events
 
-        result = get_preference_context(mock_db, workspace_id=1, product_id=10)
+        result = get_preference_context(
+            mock_db, workspace_id=1, product_id=10,
+            product_dict={"custom_prompt": "", "style_tone": ""},
+        )
 
         fe_dto_keys = {"recent_preference", "reject_reasons", "injected_count"}
         missing = fe_dto_keys - result.keys()

frontend/src/app/config/page.tsx

@@ -1,30 +1,26 @@
 'use client';
 /**
- * 屏1 配置台（管理员 /config）
+ * 系统管理 /config（仅管理员）
- * 产品库 / 标杆笔记 / 违禁词 / API Key 录入
+ * 标杆笔记 / 违禁词库 —— 平台级配置，写操作后端锁 admin
- * 🔴 红线：key 只录不显余额（不做 Token 用量展示）
+ * 拆分自原 /config（B方案）：运营日常用的 API Key/产品档案已移到 /settings「工作配置」
  */
 import { useState } from 'react';
-import { ProductsTab } from '@/components/config/ProductsTab';
 import { BenchmarksTab } from '@/components/config/BenchmarksTab';
 import { BannedWordsTab } from '@/components/config/BannedWordsTab';
-import { ApiKeysTab } from '@/components/config/ApiKeysTab';
 
-type TabKey = 'products' | 'benchmarks' | 'banned_words' | 'api_keys';
+type TabKey = 'benchmarks' | 'banned_words';
 
 const TABS: { key: TabKey; label: string }[] = [
-  { key: 'products', label: '产品档案' },
   { key: 'benchmarks', label: '标杆笔记' },
   { key: 'banned_words', label: '违禁词库' },
-  { key: 'api_keys', label: 'API Key' },
 ];
 
 export default function ConfigPage() {
-  const [activeTab, setActiveTab] = useState<TabKey>('products');
+  const [activeTab, setActiveTab] = useState<TabKey>('benchmarks');
 
   return (
     <div className="p-6 max-w-5xl">
-      <h2 className="text-xl font-bold text-text-primary mb-6">配置台</h2>
+      <h2 className="text-xl font-bold text-text-primary mb-6">系统管理</h2>
 
       {/* Tab 导航 */}
       <div className="flex gap-1 border-b border-border-default mb-6" role="tablist">
@@ -48,10 +44,8 @@ export default function ConfigPage() {
 
       {/* Tab 内容 */}
       <div id={`panel-${activeTab}`} role="tabpanel">
-        {activeTab === 'products' && <ProductsTab />}
         {activeTab === 'benchmarks' && <BenchmarksTab />}
         {activeTab === 'banned_words' && <BannedWordsTab />}
-        {activeTab === 'api_keys' && <ApiKeysTab />}
       </div>
     </div>
   );

frontend/src/app/history/HistoryFilters.tsx

@@ -0,0 +1,77 @@
+'use client';
+/**
+ * HistoryFilters — 历史归档页筛选条（日期区间 + 产品下拉 + 导出按钮）
+ * 按文件约束从 page.tsx 拆出，避免 page 超行。
+ */
+import { useEffect, useState } from 'react';
+import { api } from '@/lib/api';
+
+interface ProductOpt { id: number; name: string }
+
+interface Props {
+  dateFrom: string;
+  dateTo: string;
+  productId: string;
+  onChange: (patch: { dateFrom?: string; dateTo?: string; productId?: string }) => void;
+}
+
+// 通用导出：getBlob 下载，带 Authorization header（复用 DownloadButton 同款 blob 方案）
+async function exportData(path: string, filename: string) {
+  const blob = await api.getBlob(path);
+  const url = URL.createObjectURL(blob);
+  const a = document.createElement('a');
+  a.href = url;
+  a.download = filename;
+  document.body.appendChild(a);
+  a.click();
+  document.body.removeChild(a);
+  URL.revokeObjectURL(url);
+}
+
+export function HistoryFilters({ dateFrom, dateTo, productId, onChange }: Props) {
+  const [products, setProducts] = useState<ProductOpt[]>([]);
+  const [exporting, setExporting] = useState(false);
+
+  useEffect(() => {
+    api.get<{ items: ProductOpt[] }>('/api/v1/products?page_size=100')
+      .then((d) => setProducts(d.items || []))
+      .catch(() => setProducts([]));
+  }, []);
+
+  async function handleExport() {
+    setExporting(true);
+    try {
+      await exportData('/api/v1/products/export', 'products.json');
+    } catch { /* 失败静默，按钮恢复可重试 */ }
+    finally { setExporting(false); }
+  }
+
+  const inputCls = 'rounded-lg border border-border-default px-3 py-1.5 text-sm text-text-primary focus:border-brand-orange outline-none';
+
+  return (
+    <div className="flex flex-wrap items-center gap-3">
+      <label className="flex items-center gap-1.5 text-sm text-text-secondary">
+        起
+        <input type="date" value={dateFrom} max={dateTo || undefined}
+          onChange={(e) => onChange({ dateFrom: e.target.value })} className={inputCls} />
+      </label>
+      <label className="flex items-center gap-1.5 text-sm text-text-secondary">
+        止
+        <input type="date" value={dateTo} min={dateFrom || undefined}
+          onChange={(e) => onChange({ dateTo: e.target.value })} className={inputCls} />
+      </label>
+      <select value={productId} onChange={(e) => onChange({ productId: e.target.value })} className={inputCls}>
+        <option value="">全部产品</option>
+        {products.map((p) => <option key={p.id} value={String(p.id)}>{p.name}</option>)}
+      </select>
+      {(dateFrom || dateTo || productId) && (
+        <button onClick={() => onChange({ dateFrom: '', dateTo: '', productId: '' })}
+          className="text-xs text-text-tertiary hover:text-brand-orange underline">清除</button>
+      )}
+      <button onClick={handleExport} disabled={exporting}
+        className="ml-auto rounded-lg border border-border-default px-3 py-1.5 text-sm text-text-secondary hover:border-brand-orange disabled:opacity-50">
+        {exporting ? '导出中…' : '导出产品数据'}
+      </button>
+    </div>
+  );
+}

frontend/src/app/history/components.tsx

@@ -7,7 +7,7 @@
  * window.open 无法附带 Authorization header，会 401。
  * 改法：api.getBlob 拿二进制流 → URL.createObjectURL → a.click → revoke。
  */
-import { useState } from 'react';
+import { useState, Fragment } from 'react';
 import Link from 'next/link';
 import { clsx } from 'clsx';
 import { TaskListItem } from '@/types';
@@ -76,6 +76,7 @@ export function HistoryTable({ tasks }: { tasks: TaskListItem[] }) {
         <thead>
           <tr className="bg-surface-secondary text-text-secondary">
             <th className="px-4 py-3 text-left font-medium">任务</th>
+            <th className="px-4 py-3 text-left font-medium">产品</th>
             <th className="px-4 py-3 text-left font-medium">状态</th>
             <th className="px-4 py-3 text-left font-medium">文案</th>
             <th className="px-4 py-3 text-left font-medium">图片</th>
@@ -85,8 +86,8 @@ export function HistoryTable({ tasks }: { tasks: TaskListItem[] }) {
         </thead>
         <tbody>
           {tasks.map((task, i) => (
+            <Fragment key={task.id}>
             <tr
-              key={task.id}
               className={clsx(
                 'border-t border-border-default hover:bg-surface-tertiary transition-colors',
                 i % 2 === 0 ? 'bg-white' : 'bg-surface-primary'
@@ -95,6 +96,9 @@ export function HistoryTable({ tasks }: { tasks: TaskListItem[] }) {
               <td className="px-4 py-3 text-text-primary font-medium max-w-xs truncate">
                 {task.theme}
               </td>
+              <td className="px-4 py-3 text-text-secondary">
+                {task.product_name ?? '-'}
+              </td>
               <td className="px-4 py-3">
                 <span className={clsx(
                   'px-2 py-0.5 rounded-full text-xs font-medium',
@@ -128,6 +132,17 @@ export function HistoryTable({ tasks }: { tasks: TaskListItem[] }) {
                 </div>
               </td>
             </tr>
+            {/* 被打回任务：原因独占一行展示,运营见"为何被打回"可重做(R8历史复用) */}
+            {task.status === 'rejected' && task.reject_reason && (
+              <tr className={i % 2 === 0 ? 'bg-white' : 'bg-surface-primary'}>
+                <td colSpan={7} className="px-4 pb-3 pt-0">
+                  <div className="rounded-md bg-red-50 border border-red-200 px-3 py-2 text-xs text-red-600">
+                    <span className="font-medium">打回原因：</span>{task.reject_reason}
+                  </div>
+                </td>
+              </tr>
+            )}
+            </Fragment>
           ))}
         </tbody>
       </table>
@@ -142,6 +157,7 @@ export function HistorySkeleton() {
       {Array.from({ length: 6 }).map((_, i) => (
         <div key={i} className="flex gap-4 px-4 py-3 border-t border-border-default first:border-t-0">
           <div className="h-4 bg-gray-200 rounded flex-1" />
+          <div className="h-4 bg-gray-200 rounded w-20" />
           <div className="h-4 bg-gray-200 rounded w-16" />
           <div className="h-4 bg-gray-200 rounded w-12" />
           <div className="h-4 bg-gray-200 rounded w-12" />

frontend/src/app/history/page.tsx

@@ -10,6 +10,7 @@ import { TaskListItem, PaginatedResponse } from '@/types';
 import { getErrorAction } from '@/types/errors';
 import { clsx } from 'clsx';
 import { HistoryTable, HistorySkeleton, Pagination } from './components';
+import { HistoryFilters } from './HistoryFilters';
 
 const HISTORY_STATUSES = ['approved', 'archived', 'rejected', 'failed'];
 const STATUS_LABELS: Record<string, string> = {
@@ -24,14 +25,21 @@ export default function HistoryPage() {
   const [loading, setLoading] = useState(true);
   const [error, setError] = useState<string | null>(null);
   const [statusFilter, setStatusFilter] = useState<string>('all');
+  // M4: 日期区间 + 产品筛选
+  const [filters, setFilters] = useState({ dateFrom: '', dateTo: '', productId: '' });
 
-  const fetchHistory = useCallback(async (p: number, status: string) => {
+  const fetchHistory = useCallback(async (
+    p: number, status: string, f: { dateFrom: string; dateTo: string; productId: string },
+  ) => {
     setLoading(true);
     setError(null);
     try {
       const params = new URLSearchParams({ page: String(p), page_size: String(PAGE_SIZE) });
       if (status !== 'all') params.append('status', status);
       else HISTORY_STATUSES.forEach((s) => params.append('status', s));
+      if (f.dateFrom) params.append('date_from', f.dateFrom);
+      if (f.dateTo) params.append('date_to', f.dateTo);
+      if (f.productId) params.append('product_id', f.productId);
       const res = await api.get<PaginatedResponse<TaskListItem>>(`/api/v1/tasks?${params}`);
       setTasks(res.items);
       setTotal(res.total);
@@ -43,7 +51,7 @@ export default function HistoryPage() {
     }
   }, []);
 
-  useEffect(() => { fetchHistory(page, statusFilter); }, [page, statusFilter, fetchHistory]);
+  useEffect(() => { fetchHistory(page, statusFilter, filters); }, [page, statusFilter, filters, fetchHistory]);
 
   const totalPages = Math.ceil(total / PAGE_SIZE);
 
@@ -74,10 +82,18 @@ export default function HistoryPage() {
         ))}
       </div>
 
+      {/* M4: 日期区间 + 产品筛选 + 导出 */}
+      <HistoryFilters
+        dateFrom={filters.dateFrom}
+        dateTo={filters.dateTo}
+        productId={filters.productId}
+        onChange={(patch) => { setFilters((f) => ({ ...f, ...patch })); setPage(1); }}
+      />
+
       {error && (
         <div className="rounded-lg bg-red-50 border border-red-200 p-4 text-sm text-red-600">
           {error}
-          <button className="ml-4 underline" onClick={() => fetchHistory(page, statusFilter)}>
+          <button className="ml-4 underline" onClick={() => fetchHistory(page, statusFilter, filters)}>
             重试
           </button>
         </div>

frontend/src/app/settings/page.tsx

@@ -0,0 +1,53 @@
+'use client';
+/**
+ * 工作配置 /settings（运营+组长+管理员）
+ * API Key 录入 / 产品档案 —— 各角色日常自助配置
+ * 🔴 红线：key 各人自录自用各算各的；只录不显余额
+ * 拆分自原 /config（B方案）：admin 专属的标杆/违禁词留在 /config「系统管理」
+ */
+import { useState } from 'react';
+import { ProductsTab } from '@/components/config/ProductsTab';
+import { ApiKeysTab } from '@/components/config/ApiKeysTab';
+
+type TabKey = 'api_keys' | 'products';
+
+const TABS: { key: TabKey; label: string }[] = [
+  { key: 'api_keys', label: 'API Key' },
+  { key: 'products', label: '产品档案' },
+];
+
+export default function SettingsPage() {
+  const [activeTab, setActiveTab] = useState<TabKey>('api_keys');
+
+  return (
+    <div className="p-6 max-w-5xl">
+      <h2 className="text-xl font-bold text-text-primary mb-6">工作配置</h2>
+
+      {/* Tab 导航 */}
+      <div className="flex gap-1 border-b border-border-default mb-6" role="tablist">
+        {TABS.map((tab) => (
+          <button
+            key={tab.key}
+            role="tab"
+            aria-selected={activeTab === tab.key}
+            aria-controls={`panel-${tab.key}`}
+            onClick={() => setActiveTab(tab.key)}
+            className={`px-4 py-2 text-sm font-medium rounded-t-lg transition-colors ${
+              activeTab === tab.key
+                ? 'bg-surface-primary border border-border-default border-b-white text-brand-orange'
+                : 'text-text-secondary hover:text-text-primary'
+            }`}
+          >
+            {tab.label}
+          </button>
+        ))}
+      </div>
+
+      {/* Tab 内容 */}
+      <div id={`panel-${activeTab}`} role="tabpanel">
+        {activeTab === 'api_keys' && <ApiKeysTab />}
+        {activeTab === 'products' && <ProductsTab />}
+      </div>
+    </div>
+  );
+}

frontend/src/app/tasks/new/page.tsx

@@ -12,7 +12,7 @@ import { RecentTasksBar } from '@/components/tasks/RecentTasksBar';
 import { NewTaskForm } from '@/components/tasks/NewTaskForm';
 import { ImportTextModal } from '@/components/tasks/ImportTextModal';
 import { api } from '@/lib/api';
-import { Product, CreateTaskRequest } from '@/types/index';
+import { Product, CreateTaskRequest, Benchmark } from '@/types/index';
 import { PaginatedResponse } from '@/types/index';
 import { usePreferenceStore } from '@/stores/preferenceStore';
 import { getErrorAction } from '@/types/errors';
@@ -23,6 +23,7 @@ export default function NewTaskPage() {
   const { context } = usePreferenceStore();
   const [products, setProducts] = useState<Product[]>([]);
   const [selectedProduct, setSelectedProduct] = useState<Product | null>(null);
+  const [benchmarks, setBenchmarks] = useState<Benchmark[]>([]);
   const [form, setForm] = useState<Omit<CreateTaskRequest, 'product_id'>>({
     benchmark_ids: [],
     theme: '',
@@ -37,6 +38,15 @@ export default function NewTaskPage() {
 
   useEffect(() => { loadProducts(); }, []);
 
+  // 选品变化：拉该产品标杆笔记；切产品时清空已选标杆（避免跨产品串用）
+  useEffect(() => {
+    if (!selectedProduct) { setBenchmarks([]); return; }
+    setForm((f) => ({ ...f, benchmark_ids: [] }));
+    api.get<{ items: Benchmark[] }>(`/api/v1/products/${selectedProduct.id}/benchmarks`)
+      .then((d) => setBenchmarks(d.items || []))
+      .catch(() => setBenchmarks([]));
+  }, [selectedProduct]);
+
   async function loadProducts() {
     try {
       const data = await api.get<PaginatedResponse<Product>>('/api/v1/products');
@@ -125,6 +135,7 @@ export default function NewTaskPage() {
             products={products}
             selectedProduct={selectedProduct}
             onSelectProduct={setSelectedProduct}
+            benchmarks={benchmarks}
             form={form}
             onFormChange={(patch) => setForm((f) => ({ ...f, ...patch }))}
             submitting={submitting}

frontend/src/components/AuthGuard.tsx

@@ -2,7 +2,7 @@
 /**
  * AuthGuard — 路由守卫
  * - 未登录 → /login
- * - 越权（组长访问 /config，非管理员等）→ /dashboard
+ * - 越权（运营访问 /config 系统管理等）→ /dashboard
  * 扒 banana AuthGuard 范式，全新重建
  */
 import { useEffect } from 'react';
@@ -13,6 +13,7 @@ import { UserRole } from '@/types/index';
 // 路由 → 最低所需角色
 const ROUTE_ROLES: Record<string, UserRole[]> = {
   '/config': ['admin'],
+  '/settings': ['admin', 'supervisor', 'operator'],
   '/review': ['admin', 'supervisor'],
   '/history': ['admin', 'supervisor', 'operator'],
   '/tasks/new': ['admin', 'supervisor', 'operator'],

frontend/src/components/config/ProductImageManager.tsx

@@ -0,0 +1,154 @@
+'use client';
+/**
+ * ProductImageManager — R5 产品多图管理
+ * 多图上传(带场景类型) + 设主图 + 改场景 + 删图。
+ * scene 场景类型决定生图时哪张图喂给哪个分镜 role。
+ */
+import { useRef, useState } from 'react';
+import { api } from '@/lib/api';
+import { Product, ProductImage, ProductImageScene } from '@/types/dto';
+
+const SCENE_OPTIONS: { value: ProductImageScene; label: string }[] = [
+  { value: 'primary', label: '主图/白底' },
+  { value: 'scene', label: '使用场景' },
+  { value: 'texture', label: '质地特写' },
+  { value: 'ingredient', label: '成分/包装' },
+  { value: 'model', label: '上脸/使用中' },
+];
+
+const sceneLabel = (s: string) =>
+  SCENE_OPTIONS.find((o) => o.value === s)?.label ?? s;
+
+function toUploadSrc(path: string) {
+  return `/uploads/${path.replace(/^.*\/uploads\//, '').replace(/^uploads\//, '')}`;
+}
+
+export function ProductImageManager({
+  product,
+  onChanged,
+}: {
+  product: Product;
+  onChanged: (updated: Product) => void;
+}) {
+  const inputRef = useRef<HTMLInputElement>(null);
+  const [scene, setScene] = useState<ProductImageScene>('primary');
+  const [uploading, setUploading] = useState(false);
+  const [busyId, setBusyId] = useState<number | null>(null);
+  const [error, setError] = useState('');
+
+  const images = product.images ?? [];
+
+  async function refresh() {
+    const updated = await api.get<Product>(`/api/v1/products/${product.id}`);
+    onChanged(updated);
+  }
+
+  async function handleFile(e: React.ChangeEvent<HTMLInputElement>) {
+    const file = e.target.files?.[0];
+    if (!file) return;
+    if (file.size > 10 * 1024 * 1024) { setError('文件超过 10 MB'); return; }
+    setError('');
+    setUploading(true);
+    try {
+      const form = new FormData();
+      form.append('file', file);
+      form.append('scene', scene);
+      await api.postForm<Product>(`/api/v1/products/${product.id}/upload-image`, form);
+      await refresh();
+    } catch {
+      setError('上传失败，请重试');
+    } finally {
+      setUploading(false);
+      if (inputRef.current) inputRef.current.value = '';
+    }
+  }
+
+  async function setPrimary(imageId: number) {
+    setBusyId(imageId);
+    try {
+      await api.put(`/api/v1/products/${product.id}/images/${imageId}/primary`, {});
+      await refresh();
+    } finally { setBusyId(null); }
+  }
+
+  async function changeScene(imageId: number, next: string) {
+    setBusyId(imageId);
+    try {
+      await api.put(`/api/v1/products/${product.id}/images/${imageId}/scene`, { scene: next });
+      await refresh();
+    } finally { setBusyId(null); }
+  }
+
+  async function removeImage(imageId: number) {
+    setBusyId(imageId);
+    try {
+      await api.delete(`/api/v1/products/${product.id}/images/${imageId}`);
+      await refresh();
+    } finally { setBusyId(null); }
+  }
+
+  return (
+    <div className="mt-2 space-y-2">
+      {images.length > 0 && (
+        <div className="grid grid-cols-2 gap-2">
+          {images.map((im: ProductImage) => (
+            <div key={im.id}
+              className="flex items-center gap-2 border border-border-default rounded-lg p-2">
+              {/* eslint-disable-next-line @next/next/no-img-element */}
+              <img src={toUploadSrc(im.path)} alt={`${sceneLabel(im.scene)}图`}
+                className="h-12 w-12 object-cover rounded border border-border-default shrink-0" />
+              <div className="flex-1 min-w-0 space-y-1">
+                <select value={im.scene} disabled={busyId === im.id}
+                  onChange={(e) => changeScene(im.id, e.target.value)}
+                  aria-label="图片场景类型"
+                  className="w-full text-xs border border-border-default rounded px-1 py-0.5">
+                  {SCENE_OPTIONS.map((o) => (
+                    <option key={o.value} value={o.value}>{o.label}</option>
+                  ))}
+                </select>
+                <div className="flex items-center gap-2 text-xs">
+                  {im.is_primary ? (
+                    <span className="text-brand-orange font-medium">★ 主图</span>
+                  ) : (
+                    <button type="button" disabled={busyId === im.id}
+                      onClick={() => setPrimary(im.id)}
+                      className="text-text-secondary hover:text-brand-orange">设主图</button>
+                  )}
+                  <button type="button" disabled={busyId === im.id}
+                    onClick={() => removeImage(im.id)}
+                    className="text-red-500 hover:underline ml-auto">删除</button>
+                </div>
+              </div>
+            </div>
+          ))}
+        </div>
+      )}
+
+      <div className="flex items-center gap-2">
+        <select value={scene} onChange={(e) => setScene(e.target.value as ProductImageScene)}
+          aria-label="新图场景类型"
+          className="text-xs border border-border-default rounded px-2 py-1">
+          {SCENE_OPTIONS.map((o) => (
+            <option key={o.value} value={o.value}>{o.label}</option>
+          ))}
+        </select>
+        <button type="button" onClick={() => inputRef.current?.click()} disabled={uploading}
+          className="text-xs border border-dashed border-border-default rounded px-3 py-1
+                     text-text-secondary hover:border-brand-orange hover:text-brand-orange
+                     disabled:opacity-50 transition-colors">
+          {uploading ? '上传中…' : '+ 上传产品图'}
+        </button>
+      </div>
+
+      {images.length === 0 && (
+        <p className="text-xs text-text-tertiary leading-relaxed">
+          首张必须含产品主体（瓶身/包装本体）当主图，生图以此为锚点保证产品一致。
+          质地图、上脸图等设对应场景类型，生图会按分镜自动选用。
+        </p>
+      )}
+      {error && <p className="text-xs text-red-500">{error}</p>}
+      <input ref={inputRef} type="file" accept="image/jpeg,image/png,image/webp"
+        className="hidden" aria-label="选择产品图" onChange={handleFile} />
+    </div>
+  );
+}

frontend/src/components/config/ProductSubComponents.tsx

@@ -1,92 +1,16 @@
 'use client';
 /**
- * ProductCard + ProductForm + ProductsSkeleton + ProductImageUpload — 产品档案子组件
+ * ProductCard + ProductForm + ProductsSkeleton — 产品档案子组件
+ * 产品多图管理拆到 ProductImageManager（R5）。
  */
-import { useRef, useState } from 'react';
+import { useState } from 'react';
 import { api } from '@/lib/api';
 import { Product, CreateProductRequest } from '@/types/index';
+import { ProductImageManager } from './ProductImageManager';
 
-// --- ProductImageUpload ---
+// ProductImageUpload 旧单图组件已被 ProductImageManager（多图+场景）取代。
-export function ProductImageUpload({
+// 保留导出名向后兼容引用方。
-  product,
+export { ProductImageManager as ProductImageUpload };
-  onUploaded,
-}: {
-  product: Product;
-  onUploaded: (updated: Product) => void;
-}) {
-  const inputRef = useRef<HTMLInputElement>(null);
-  const [uploading, setUploading] = useState(false);
-  const [error, setError] = useState('');
-
-  async function handleFile(e: React.ChangeEvent<HTMLInputElement>) {
-    const file = e.target.files?.[0];
-    if (!file) return;
-    if (file.size > 10 * 1024 * 1024) { setError('文件超过 10 MB'); return; }
-    setError('');
-    setUploading(true);
-    try {
-      const form = new FormData();
-      form.append('file', file);
-      const updated = await api.postForm<Product>(
-        `/api/v1/products/${product.id}/upload-image`, form
-      );
-      onUploaded(updated);
-    } catch {
-      setError('上传失败，请重试');
-    } finally {
-      setUploading(false);
-      if (inputRef.current) inputRef.current.value = '';
-    }
-  }
-
-  return (
-    <div className="mt-2">
-      {product.image_path ? (
-        <div className="flex items-center gap-2">
-          {/* eslint-disable-next-line @next/next/no-img-element */}
-          <img
-            src={`/uploads/${product.image_path.replace(/^.*\/uploads\//, '').replace(/^uploads\//, '')}`}
-            alt={`${product.name}参考图`}
-            className="h-12 w-12 object-cover rounded border border-border-default"
-          />
-          <button
-            type="button"
-            onClick={() => inputRef.current?.click()}
-            className="text-xs text-brand-orange hover:underline"
-          >
-            换图
-          </button>
-        </div>
-      ) : (
-        <div>
-          <button
-            type="button"
-            onClick={() => inputRef.current?.click()}
-            disabled={uploading}
-            className="text-xs border border-dashed border-border-default rounded px-3 py-1
-                       text-text-secondary hover:border-brand-orange hover:text-brand-orange
-                       disabled:opacity-50 transition-colors"
-          >
-            {uploading ? '上传中…' : '+ 上传产品参考图'}
-          </button>
-          <p className="text-xs text-text-tertiary mt-1 leading-relaxed">
-            必须含产品主体（瓶身/包装本体），生图以此为锚点保证产品一致。
-            质地图、手臂上脸图等可作补充，但不能替代主图单独上传。
-          </p>
-        </div>
-      )}
-      {error && <p className="text-xs text-red-500 mt-1">{error}</p>}
-      <input
-        ref={inputRef}
-        type="file"
-        accept="image/jpeg,image/png,image/webp"
-        className="hidden"
-        aria-label="选择产品参考图"
-        onChange={handleFile}
-      />
-    </div>
-  );
-}
 
 export function ProductCard({ product, onUpdated }: { product: Product; onUpdated: () => void }) {
   const [current, setCurrent] = useState<Product>(product);
@@ -109,9 +33,9 @@ export function ProductCard({ product, onUpdated }: { product: Product; onUpdate
           </span>
         ))}
       </div>
-      <ProductImageUpload
+      <ProductImageManager
         product={current}
-        onUploaded={(updated) => { setCurrent(updated); onUpdated(); }}
+        onChanged={(updated) => { setCurrent(updated); onUpdated(); }}
       />
     </div>
   );
@@ -124,14 +48,18 @@ export function ProductForm({ onSaved, onCancel }: { onSaved: () => void; onCanc
     custom_prompt: null, banned_word_ids: [],
   });
   const [saving, setSaving] = useState(false);
+  const [error, setError] = useState('');
 
   async function handleSave(e: React.FormEvent) {
     e.preventDefault();
-    if (!form.name) return;
+    if (!form.name.trim()) { setError('请填写产品名称'); return; }
-    setSaving(true);
+    setSaving(true); setError('');
     try {
       await api.post('/api/v1/products', form);
       onSaved();
+    } catch (err) {
+      const msg = (err as { message?: string })?.message;
+      setError(msg || '保存失败，请重试');
     } finally {
       setSaving(false);
     }
@@ -158,6 +86,7 @@ export function ProductForm({ onSaved, onCancel }: { onSaved: () => void; onCanc
           />
         </div>
       </div>
+      {error && <p role="alert" className="text-sm text-red-500">{error}</p>}
       <div className="flex gap-3 justify-end">
         <button type="button" onClick={onCancel}
           className="px-4 py-2 text-sm text-text-secondary border border-border-default rounded-lg hover:bg-surface-tertiary">

frontend/src/components/layout/Sidebar.tsx

@@ -2,7 +2,7 @@
 /**
  * Sidebar — 左侧8项导航
  * 按 PRD-前端§2 全局布局
- * 角色守卫：/review 只组长+管理员；/config 只管理员
+ * 角色守卫：/review 只组长+管理员；/settings 全员；/config 只管理员
  */
 import Link from 'next/link';
 import { usePathname } from 'next/navigation';
@@ -26,7 +26,8 @@ const NAV_ITEMS: NavItem[] = [
   { label: '裂变扩散', href: '/fission', icon: '🌱' },
   { label: '审核台', href: '/review', roles: ['supervisor', 'admin'], icon: '✅' },
   { label: '历史归档', href: '/history', icon: '🗂️' },
-  { label: '配置中心', href: '/config', roles: ['admin'], icon: '⚙️' },
+  { label: '工作配置', href: '/settings', icon: '🔧' },
+  { label: '系统管理', href: '/config', roles: ['admin'], icon: '⚙️' },
 ];
 
 export function Sidebar() {

frontend/src/components/tasks/NewTaskForm.tsx

@@ -3,13 +3,14 @@
  * NewTaskForm — 开任务表单（屏2主体）
  * 选产品 + 今天主题 + 数量 + 双轨入口
  */
-import { Product, CreateTaskRequest } from '@/types/index';
+import { Product, CreateTaskRequest, Benchmark } from '@/types/index';
 import { CountStepper } from './CountStepper';
 
 interface NewTaskFormProps {
   products: Product[];
   selectedProduct: Product | null;
   onSelectProduct: (p: Product | null) => void;
+  benchmarks: Benchmark[];
   form: Omit<CreateTaskRequest, 'product_id'>;
   onFormChange: (patch: Partial<Omit<CreateTaskRequest, 'product_id'>>) => void;
   submitting: boolean;
@@ -19,7 +20,7 @@ interface NewTaskFormProps {
 }
 
 export function NewTaskForm({
-  products, selectedProduct, onSelectProduct,
+  products, selectedProduct, onSelectProduct, benchmarks,
   form, onFormChange, submitting, error,
   onSubmitAi, onSubmitImport,
 }: NewTaskFormProps) {
@@ -28,6 +29,15 @@ export function NewTaskForm({
     onFormChange({ [field]: Math.max(1, Math.min(max, form[field] + delta)) });
   }
 
+  // 只允许选已分析完(analyze_status=done)的标杆——未分析的无 features 注入无意义
+  const usableBenchmarks = benchmarks.filter((b) => b.analyze_status === 'done');
+  function toggleBenchmark(id: number) {
+    const cur = form.benchmark_ids || [];
+    onFormChange({
+      benchmark_ids: cur.includes(id) ? cur.filter((x) => x !== id) : [...cur, id],
+    });
+  }
+
   // 产品参考图状态：image_path 绝对路径(/app/uploads/...)取末段映射成 /uploads/... 供前端访问
   const rawPath = selectedProduct?.image_path || '';
   const hasProductImage = !!rawPath;
@@ -93,6 +103,30 @@ export function NewTaskForm({
         <p className="text-xs text-text-tertiary mt-1">{form.theme.length}/100</p>
       </div>
 
+      {/* 对标爆款（可选，多选）：选中标杆把其8维配方注入文案生成 */}
+      {usableBenchmarks.length > 0 && (
+        <div className="bg-surface-secondary rounded-lg p-4 space-y-2">
+          <p className="text-sm font-medium text-text-primary">
+            对标爆款（可选）<span className="text-xs text-text-tertiary ml-1">借爆款方法结构，不抄竞品原话</span>
+          </p>
+          <div className="space-y-1.5 max-h-40 overflow-auto">
+            {usableBenchmarks.map((b) => (
+              <label key={b.id} className="flex items-start gap-2 cursor-pointer text-sm">
+                <input type="checkbox"
+                  checked={(form.benchmark_ids || []).includes(b.id)}
+                  onChange={() => toggleBenchmark(b.id)}
+                  className="w-4 h-4 mt-0.5 accent-brand-orange"
+                  aria-label={`选用标杆 ${b.id}`} />
+                <span className="text-text-secondary line-clamp-2">
+                  {b.highlights || `标杆 #${b.id}`}
+                </span>
+              </label>
+            ))}
+          </div>
+          <p className="text-xs text-text-tertiary">已选 {(form.benchmark_ids || []).length} 个，最多取前 3 个注入</p>
+        </div>
+      )}
+
       {/* 数量设定（不写死） */}
       <div className="flex gap-6">
         <CountStepper label="生几条文案" value={form.text_count}

frontend/src/types/dto.tasks.ts

@@ -8,6 +8,7 @@ export type TaskStatus =
 export interface TaskListItem {
   id: number;
   product_id: number;
+  product_name?: string | null;  // M4: 历史页"产品"列,list_tasks 批量 JOIN 填充
   theme: string;
   status: TaskStatus;
   created_at: string;

frontend/src/types/dto.ts

@@ -37,6 +37,16 @@ export interface ApiKey {
 export interface CreateApiKeyRequest { provider: string; api_key: string; }
 
 // 产品档案
+export type ProductImageScene = 'primary' | 'scene' | 'texture' | 'ingredient' | 'model';
+
+export interface ProductImage {
+  id: number;
+  path: string;
+  scene: ProductImageScene;
+  is_primary: boolean;
+  sort_order: number;
+}
+
 export interface Product {
   id: number;
   name: string;
@@ -47,7 +57,8 @@ export interface Product {
   text_angles: string[];
   custom_prompt: string | null;
   banned_word_ids: number[];
-  image_path: string | null;   // 产品参考图路径（铁律：生图带产品图）
+  image_path: string | null;   // 产品参考图路径（铁律：生图带产品图，向后兼容=主图）
+  images?: ProductImage[];      // R5多图：每张带 scene 场景类型
 }
 
 export interface CreateProductRequest {
@@ -67,6 +78,8 @@ export interface Benchmark {
   screenshot_url: string;
   highlights: string;
   link_url: string | null;
+  analyze_status?: 'pending' | 'analyzing' | 'done' | 'failed';
+  features?: Record<string, unknown> | null;
 }
 
 export interface CreateBenchmarkRequest {

frontend/src/types/index.ts

@@ -9,6 +9,7 @@ export type {
   SwitchWorkspaceRequest, SwitchWorkspaceResponse,
   ApiKey, CreateApiKeyRequest,
   Product, CreateProductRequest,
+  ProductImage, ProductImageScene,
   Benchmark, CreateBenchmarkRequest,
   BannedWordLevel, BannedWord, CreateBannedWordRequest,
 } from './dto';

plans/R-裂变重写核销表.md

@@ -64,3 +64,25 @@
 ## 搁置/待北哥（攒批）
 - 裂变三档/参考度真实业务措辞
 - N套真差异化质量验收(北哥过目)
+
+## 补图机制（中转站波动兜底，2026-06-18 追加并实测✅）
+
+> 实测发现 codeproxy 偶发不稳，单套中某张图重试3次仍失败会记 error。倩倩姐拍板"不稳定的需要加再次重试的机制"。原则：只重生失败分镜，成功的图与 token 不浪费；生图引擎零改动(R2)。
+
+| # | 文件 | 改什么 | 状态 |
+|---|------|--------|------|
+| RT1 | alembic 019 + models/fission.py | fission_tasks 加 source_task_id(新架构不扇出GenerationTask,补图需复刻首轮上下文:产品/operator key/张数,故落库回查;原 source_fission_id 反查链在新架构下断裂) | ✅ 迁移019 head,列已加 |
+| RT2 | services/fission_service.py | create_fission 写入 source_task_id | ✅ |
+| RT3 | services/fission_image_retry.py | 只补标 error 的 role(引擎 target_role 单张重生,R2零改动),merge 回 images_json,全恢复则 status=image_done | ✅ 105行 |
+| RT4 | workers/tasks.py | retry_fission_images Celery task,key归属源任务operator(基石B不传明文key),幂等锁防重复补图 | ✅ 已注册(celery inspect registered) |
+| RT5 | api/v1/fission.py | POST /fission/{id}/notes/{seq}/retry-images,无失败分镜守卫 queued=false | ✅ |
+| RT6 | 前端 FissionNoteCard + notes/page.tsx | 失败提示改"重试补图"按钮+loading,父页轮询6次×8s刷新 | ✅ tsc过 |
+
+- 端到端实测✅: task5 seq1(hook图首轮codeproxy重试3次失败,note_id=2)触发补图→**只补hook**(1024×1536去AI化新落盘)、**applied_proof原图不动**(时间戳未变)、recovered=1 still_failed=0、status=image_done; apiports503回落codeproxy gpt-5.5强档兜底(R3); seq0无失败图守卫返回queued=false不浪费token
+- git: 285791c "裂变补图：中转站波动导致单张失败时只重生失败分镜"(8文件)
+
+## git 提交记录（2026-06-18）
+- 7f419f4 存量后端(生图素人感约束+评图分+幂等防重跑+审核回路)
+- d85dcd4 第11环裂变重写(一次LLM出N套完整笔记包)
+- 25f86b2 存量前端(生图分组/灯箱/重生控件等)
+- 285791c 裂变补图机制

plans/总核销表.md

@@ -24,15 +24,15 @@ KR1链路连通 + KR2质量达标 + KR3自主可验 + KR4引擎补全(第2环+
 
 | ID | 任务 | 模块核销表 | 细化 | 状态 |
 |----|------|-----------|------|------|
-| M3 | 第8环 审核(死链/打回回路) | plans/核销-第8环-审核.md | 🟢S1-S6/A1-A8 | □ |
+| M3 | 第8环 审核(死链/打回回路) | plans/核销-第8环-审核.md | 🟢S1-S6/A1-A8 | ✅ 对账核实(2026-06-18独立agent):review.py approve(:101)/reject(:127)端点真在,_fmt_task返review_status/reject_reason/reviewed_at(:64-66),RejectReasonBanner在text/images页均挂载,/review审核工作台页+ReviewCard+RejectModal全在,打回→status=pending_selection→运营见原因可重做闭环真通。曾误标□ |
-| M4 | 第12环 归档(日期品类筛选/导出) | plans/核销-第12环-归档.md | 🟢S1-S8/A1-A12 | □ |
+| M4 | 第12环 归档(日期品类筛选/导出) | plans/核销-第12环-归档.md | 🟢S1-S8/A1-A12 | ◐ 后端验过(2026-06-18独立agent):✅A1日期筛选(79→当日4/未来0/非法→40001)✅A2品类筛选(product_id=1命中78+product_name批量防N+1)✅A9产品导出(3条+selling_points真数组)✅A6/A7 zip结构+文案格式;前端代码就绪过typecheck(产品列/日期控件/打回原因banner/导出按钮)但A3/A4/A12目视项待Z阶段浏览器实跑;A10标杆导出逻辑已审ws3无数据未触发 |
-| M5 | 工程底座(5并发/magic/选标杆/置灰/lightbox) | plans/核销-工程底座.md | 🟢S1-S6/A1-A8 | □ |
+| M5 | 工程底座(5并发/magic/选标杆/置灰/lightbox) | plans/核销-工程底座.md | 🟢S1-S6/A1-A8 | ✅ 完成(2026-06-18):✅lightbox ✅置灰 ✅magic(文件头校验,products.py _check_image_magic,2端点接入,验证真图过.exe拒) ✅5并发(config值2→5红线,task_service._check_concurrency_limit算pending/generating,验证5拦4放) ✅选标杆(注入链4断点全打通:task_service存benchmark_ids→pipeline_steps.load_benchmark_features读done态→tasks塞product_dict→_text_prompt渲染8维配方+负向约束;前端NewTaskForm多选UI;DB回环验证benchmark_ids真存库+done过滤+ws隔离;独立agent重核5项真修好,不入图片prompt合规保持) |
 
 ## P2 引擎补全（第2环标杆 / 第11环裂变）
 
 | ID | 任务 | 模块核销表 | 细化 | 状态 |
 |----|------|-----------|------|------|
-| M6 | 第2环 标杆8特征分析 | plans/核销-第2环-标杆.md | 🟢S1-S12(S12拆a/b)/A1-A12 | ◐ **引擎双路径端到端真验完成**(2026-06-16 c1ac2a6):①文本路径(手填亮点)8维齐全conf0.9 ②截图vision路径(真喂4.7MB素颜霜图)source=vision真读图,识别三段式卖点/√图标/大字报模板/办公背景,看不到的诚实标"截图未体现"。features_json落库done。端点健壮(404/422/读图失败降级/key缺失提示)。**剩非引擎本体待办:①北哥过目8维草案标✅满分 ②第5/6环消费features_json对标(接线,归M6下游或随M1/M2)**。8维=banana不变量+俊达配方+3链接 |
+| M6 | 第2环 标杆8特征分析 | plans/核销-第2环-标杆.md | 🟢S1-S12(S12拆a/b)/A1-A12 | ◐ **引擎双路径+第5环消费接线完成**(2026-06-16~18)：①文本路径(手填亮点)8维齐全conf0.9 ②截图vision路径(真喂4.7MB素颜霜图)source=vision真读图,识别三段式卖点/√图标/大字报模板/办公背景,看不到的诚实标"截图未体现"。features_json落库done。端点健壮(404/422/读图失败降级/key缺失提示)。**✅第5环消费features_json已接通(2026-06-18)**:选标杆注入链4断点全打通(详见M5),用户选的标杆8维配方真进文案build_prompt(借方法层结构+禁抄竞品负向约束),独立agent重核真修好。**剩非引擎本体待办:北哥过目8维草案标✅满分**。8维=banana不变量+俊达配方+3链接 |
 | M7 | 第11环 裂变(工作台内) | plans/核销-第11环-裂变.md | 🟢S1-S13(含S10.5)/AC1-12 | ◐ **引擎端到端真跑通**(2026-06-16 cefdbaa):①fission_service扇出1源→N子任务(回填source_fission_id,复用主管道run_generation_pipeline),MAX_FANOUT=5红线 ②fission_prompt三档(high/mid/low)规则 ③tasks.py worker识别裂变子任务→注入对应档位进文案上下文(自抓关键缺口,**worker曾跑旧内存代码未热重载,restart后实测注入日志level=high真出现**) ④POST/fission触发+GET/fission/{id}进度聚合,main.py注册路由 ⑤实测fission_id=2扇出task70/71,子任务带source_fission_id+档位注入生效,进度端点聚合正确(total=2/done=0/generating)。**剩待办:①三档真实措辞待北哥定义(现占位规则) ②前端裂变入口+进度呈现(归前端攒批) ③子任务真出3套差异化文案+配图的质量验收(需北哥过目)** |
 
 ## P3 前端打通+飞轮（2026-06-16 九文档审核结论·按序逐条修）
@@ -44,7 +44,7 @@ KR1链路连通 + KR2质量达标 + KR3自主可验 + KR4引擎补全(第2环+
 | ID | 任务 | 根据 | 状态 |
 |----|------|------|------|
 | A1 | Sidebar 4个死链路由(/tasks /images /tasks/build /tasks/list 全不存在) | Sidebar.tsx确证 | ✅ 2026-06-16 删4死链+补回/review审核台(限supervisor+admin),tsc全量EXIT0,commit 77f52e7 |
-| D1 | 改稿进飞轮(SignalType.text_edit + TextCandidate.edited + 改稿端点走task_actions) | 最强真实信号 | ◐ 后端✅2026-06-16端到端真跑(edited False→True+飞轮信号落库weight=5,实测抓出ENUM未同步bug已修,commit e3fb6f2)。**前端改稿框入口攒批做(倩倩姐2026-06-16:后端先串完前端攒批)** |
+| D1 | 改稿进飞轮(SignalType.text_edit + TextCandidate.edited + 改稿端点走task_actions) | 最强真实信号 | ✅ 全链完整(2026-06-18独立agent核实):后端✅端到端真跑(edited False→True+飞轮信号weight=5,commit e3fb6f2)。**前端✅已做完整**:TextCandidateCard.tsx:29-47可编辑textarea+保存改稿/取消按钮,text/page.tsx:29-35真调POST /text-candidates/{id}/edit+同步store,保存后展示"🌿已学习你的改动"+"已改稿"角标。曾误标◐ |
 | C2 | 审核界面显分(读score_json总分,不读eval_score NULL列) | 审核可判断 | ✅ 2026-06-16 review.py复用_score_array_to_obj算总分+7维,ReviewCard接ScoreDimBars(passLine=80红线),端到端验task40返回total=83,新旧维度兼容,commit ea423e8 |
 | B7 | 事件处理补齐(SSE/轮询状态流转) | 前端诊断 | ✅ 2026-06-16 核实已做无缺口:SSE签ticket换JWT(?ticket=合规红线)+useSse消费progress/done/error全链+轮询兜底,后端事件发布到位,标绿 |
 | B6 | 错误态呈现(status=failed前端可见) | 前端诊断 | ✅ 2026-06-16 后端加TaskStatus.FAILED终态+异常区分重试中vs耗尽(耗尽落failed不retry)+014迁移ALTER status ENUM;前端fatalError state+useSse error不再静默丢+FatalErrorBanner提示+failed标签(列表/历史页可见,history纳入HISTORY_STATUSES),tsc EXIT0 |
@@ -52,16 +52,16 @@ KR1链路连通 + KR2质量达标 + KR3自主可验 + KR4引擎补全(第2环+
 ### 桶B 飞轮专项（产品最大亮点·做扎实+前端呈现）
 | ID | 任务 | 守的底线 | 状态 |
 |----|------|---------|------|
-| E12 | AI评图分进飞轮:仅"生成时筛选+前端展示X分已入飞轮" | 权重只认真实信号(选/改/审),eval_score留NULL铁律 | □ |
+| E12 | AI评图分进飞轮:仅"生成时筛选+前端展示X分已入飞轮" | 权重只认真实信号(选/改/审),eval_score留NULL铁律 | ✅ 全链完整(2026-06-18独立agent核实):ImageStrategyGroup.tsx:89-96 ai_visual_score徽章(≥80绿/≥60橙/<60灰),后端image_gen.py:228→pipeline_io.py:303→tasks.py:146落库透传;eval_score全程NULL(models/task.py:93/124注释+image_gen/pipeline_io/review.py均不碰),红线守住。曾误标□ |
-| FV | 飞轮可视化=轻量化关键节点反馈(让客户感到确实越用越好) | 不做独立飞轮主视图大页(砍E01) | □ |
+| FV | 飞轮可视化=轻量化关键节点反馈(让客户感到确实越用越好) | 不做独立飞轮主视图大页(砍E01) | ✅ 符合决策(2026-06-18核实):FlywheelBanner(new页/text页/images页SSE flywheel_injected事件)+TextCandidateCard"🌿已学习"+AI分徽章,轻量节点反馈嵌现有流程;无独立大看板=砍E01已拍板。曾误标□ |
 
 ### 桶C 接线（数据流薄，主链通后接）
 | ID | 任务 | 状态 |
 |----|------|------|
-| B1 | 图片按strategy分组(generationStore现一维平铺) | □ |
+| B1 | 图片按strategy分组(generationStore现一维平铺) | ✅ 同R1已做(对账核实):ImageStrategyGroup按strategy分A/B/C三组,images页:45-47分组+:96-108渲染,非空壳。曾重复挂账□ |
-| C1 | 任务看板 | □ |
+| C1 | 任务看板 | □ 真欠账(2026-06-18核实:仅/history列表+状态tab,无kanban按状态分列视图,零代码)。⚠️攒批问倩倩姐:/history已有状态筛选,是否还需独立kanban看板?可能伪需求 |
-| B2/B3/B4/B5 | 前端数据流补齐 | □ |
+| B2/B3/B4/B5 | 前端数据流补齐 | ✅ 主链完整(2026-06-18核实:generationStore/taskStore/preferenceStore全实现无空壳无TODO,hydrateFromTask历史回填已做);B2-B5核销表本身需求定义模糊,代码层无欠账 |
-| C3/C4 | 展示层补薄 | □ |
+| C3/C4 | 展示层补薄 | C3✅留口子(OVERLAY_TEXT_RENDER_ENABLED=False拍板缓做);C4□代码无对应实现且需求不明,⚠️攒批问倩倩姐C4具体指什么 |
 
 ### 桶D 踩刹车（砍/缓，防堆伪需求偏离纵切片）
 | ID | 任务 | 裁决 |
@@ -72,10 +72,10 @@ KR1链路连通 + KR2质量达标 + KR3自主可验 + KR4引擎补全(第2环+
 ### 桶E 需补充（缺口补全）
 | ID | 任务 | 状态 |
 |----|------|------|
-| 补1 | 上传主图校验(必含产品瓶身,防质地图/手臂图当主图) | □ |
+| 补1 | 上传主图校验(必含产品瓶身,防质地图/手臂图当主图) | ✅ 同R3已做(对账核实):后端2层硬拦截(task_service.py:46-58建任务校验+pipeline_io.py:199-204 worker兜底防绕过)+前端NewTaskForm:66-77红色警告+按钮置灰。曾重复挂账□ |
 | 补2 | status=failed 完整处理(后端+前端) | ✅ 同B6(2026-06-16),后端FAILED终态+前端可见,commit 1b48875/ffc213f |
 | 补3 | git版本兜底 | ✅2026-06-16已建Clover独立仓库,基线6a2632d |
-| D2 | 数据复盘(飞轮喂料可见) | □ |
+| D2 | 数据复盘(飞轮喂料可见) | ◐ 实时注入反馈已做(FlywheelBanner+"已学习"提示,见FV),但历史注入摘要/复盘页未做。⚠️攒批问倩倩姐:轻量节点反馈是否已够,还需独立历史复盘页? |
 | D3 | 爆款解析(并入M6标杆) | ⏸️随M6 |
 
 ---
@@ -131,33 +131,33 @@ KR1链路连通 + KR2质量达标 + KR3自主可验 + KR4引擎补全(第2环+
 ### R5 产品多图（P1）
 | 子项 | 去重映射 | 根因证据 | 状态 |
 |------|---------|---------|------|
-| 产品图多张(DB+上传端点+前端multiple) | 问题2/14新 | image_path单值字符串,无multiple,覆盖式 | □ |
+| 产品图多张(DB+上传端点+前端multiple) | 问题2/14新 | image_path单值字符串,无multiple,覆盖式 | ✅ product_images表(alembic020)+ProductImage模型(lazy=selectin)+5端点(upload带scene/列/改场景/设主图/删图,删主图自动提升下张防image_path悬空);前端ProductImageManager.tsx多图画廊(每张缩略图+场景下拉+设主图+删除);tsc EXIT0;API链路实测建品→传2图(primary+scene)→改场景→设主图→删图全code=0 |
-| 生图按场景选用对应图 | 问题2新 | pipeline_io只读单条image_path | □ |
+| 生图按场景选用对应图 | 问题2新 | pipeline_io只读单条image_path | ✅ ROLE_SCENE_PREFERENCE覆盖全10分镜role;_pick_reference_for_role按偏好链选图缺失回落primary再回落fallback;image_gen/pipeline_io/fission_images/fission_image_retry四路均构造并传images_by_scene;真实数据验证texture分镜命中texture图(不同字节确认)、其余回落primary、无多图回落原reference;独立agent交叉核验6链路点全真0实质bug |
-| operator选品/快速建品/产品入镜 | 问题1新 | POST/products限admin,operator不能建 | □ |
+| operator选品/快速建品/产品入镜 | 问题1新 | POST/products限admin,operator不能建 | ✅ create/update放开require_admin→require_write_permission;delete_product保留require_admin(破坏性不放开);operator账号实测建品id=6+传图成功code=0 |
 
 ### R6 账号体系+Key+权限（P1）
 | 子项 | 去重映射 | 根因证据 | 状态 |
 |------|---------|---------|------|
-| 后端users.py增/删/列用户(删=软删is_active) | 问题15新 | 无users CRUD,is_active登录已校验:37 | □ |
+| 后端users.py增/删/列用户(删=软删is_active) | 问题15新 | 无users CRUD,is_active登录已校验:37 | ⏸️暂停(倩倩姐2026-06-18拍板:现有4种子账号够北哥测,加人能力留后;先推R7飞轮) |
-| 前端配置台账号管理Tab | 问题15新 | 无账号管理入口 | □ |
+| 前端配置台账号管理Tab | 问题15新 | 无账号管理入口 | ⏸️暂停(同上,随后端users CRUD一起做) |
-| Key绑定账号显式展示 | 问题15新 | UserApiKey已按user隔离,只缺前端展示 | □ |
+| Key绑定账号显式展示 | 问题15新 | UserApiKey已按user隔离,只缺前端展示 | ✅ B方案拆页:运营进/settings「工作配置」自录key;实测operator录key(id=4 last4=0001)归user_id=3、admin的key(id=2)归user_id=1,DB核对user_id严格隔离各算各的;只显末4位不显余额(守红线) |
-| operator配置编辑权(放开全部产品) | 问题13a新 | /config硬限admin,products CRUD require_admin | □ |
+| operator配置编辑权(放开全部产品) | 问题13a新 | /config硬限admin,products CRUD require_admin | ✅ B方案:配置页按角色拆/settings(运营+组长+admin,含APIKey+产品档案)+/config(仅admin,标杆+违禁词);Sidebar+AuthGuard双层放开/settings给三角色;前缀匹配无串扰;角色矩阵静态核查无死链;独立agent交叉核验0穿透漏洞;后端require_write_permission已对齐 |
-| 保存按钮补错误提示(治点不动) | 问题13b新 | ProductForm校验静默return无提示 | □ |
+| 保存按钮补错误提示(治点不动) | 问题13b新 | ProductForm校验静默return无提示 | ✅ ProductForm加error state:空名提示"请填写产品名称"+catch后端报错展示(role=alert);tsc EXIT0 |
 
 ### R7 飞轮闭环+可视化（P2）
 | 子项 | 去重映射 | 根因证据 | 状态 |
 |------|---------|---------|------|
-| 飞轮片段注入生图链路(补断点1) | 新 | run_image_generation没传flywheel_fragment | □ |
+| 飞轮片段注入生图链路(补断点1) | 新 | run_image_generation没传flywheel_fragment | ✅ generate_storyboard_images+run_image_generation加flywheel_fragment参数,tasks.py传load_flywheel_context真实聚合结果;拼per_prompt措辞"仅影响文字角度/排版不改瓶身"守红线;独立agent复核"真修好" |
-| 选图信号带strategy进飞轮(补断点2) | 新 | image_select的angle_label=None空转 | □ |
+| 选图信号带strategy进飞轮(补断点2) | 新 | image_select的angle_label=None空转 | ✅ IMAGE_STRATEGY_ANGLE映射A痛点/B场景/C成分背书,image_select取ic.strategy映射angle_label+原值存signal_meta;record_signal加signal_meta参;strategy空安全降级;复核"真修好" |
-| 统一前端查询与注入的计数口径 | 新 | service用次数/aggregator用权重不一致 | □ |
+| 统一前端查询与注入的计数口径 | 新 | service用次数/aggregator用权重不一致 | ✅ get_preference_context改委托aggregate_preference_context(权重口径),image_select纳入角度聚合;展示端点与生产链同源,交叉验证输出字符串完全一致;复核"真修好" |
-| AI评图分仅展示+生成筛选 | =既有E12 | 守eval_score NULL红线 | □ |
+| AI评图分仅展示+生成筛选 | =既有E12 | 守eval_score NULL红线 | ✅ 既有已达标:ai_visual_score只展示+筛选,eval_score全程NULL,本轮改动复核确认未碰;ImageStrategyGroup徽章展示AI分 |
-| 飞轮可视化轻量节点反馈 | =既有FV/D2 | 让客户感到越用越好 | □ |
+| 飞轮可视化轻量节点反馈 | =既有FV/D2 | 让客户感到越用越好 | ◐ 已有雏形:生成时FlywheelBanner"本次已注入偏好角度"+改稿TextCandidateCard"🌿已学习你的改动"+AI分徽章;无独立大看板(符合砍E01红线) |
 
 ### R8 归档（P2）
 | 子项 | 去重映射 | 根因证据 | 状态 |
 |------|---------|---------|------|
-| 历史打回显示原因 | =问题10,既有M4沾边 | _fmt_task()没返reject_reason,前端没读 | □ |
+| 历史打回显示原因 | =问题10,既有M4沾边 | ✅后端_fmt_task返review_status/reject_reason/reviewed_at(tasks.py:58),前端HistoryTable加rejected行红banner展示打回原因(代码就绪,Z阶段目视) | ◐ |
-| 日期/品类筛选+导出 | =既有M4 | M4未开工 | □ |
+| 日期/品类筛选+导出 | =既有M4 | ✅后端筛选链+导出端点独立agent验过(A1/A2/A9),前端HistoryFilters就绪待目视 | ◐ |
 
 ---
 

plans/核销-第12环-归档.md

@@ -14,24 +14,24 @@
 | S3 | history/page.tsx 加 dateFrom/dateTo state + 两个 `<input type=date>`(提 DateRangeFilter 子组件防超100行) | history/page.tsx:28-44 + JSX 50-94 |
 | S4 | HistoryTable 表头/行加"产品"列(显 product_name ?? '-') | history/components.tsx:70-127 |
 | S5 | TaskListItem 加 `product_name?:string` | dto.tasks.ts:8-16 |
-| S6 | ⚠️多套打包：查所有 is_selected=True TextCandidate 按 angle_label 分组每组取第一；notes 列表长度=选中套数(改 line 77-85 为 for loop)；images_data 循环外查一次，每套传 `list(images_data)` 副本(退化:图共享复制进各套夹) | packaging_task.py:45-47,77-85 |
+| S6 | ✅多套打包(2026-06-18实做修正分组依据)：实测真相=图按 `ImageCandidate.strategy`(A/B/C三套正交叙事)分组,三套图均以first_copy第1条文案为语境生成→**按strategy分组比angle_label更贴数据实际,且无需加set_index列**。改 packaging_task.py:查全部图按strategy分组(OrderedDict)+同(strategy,seq)取最新id去重(防重生重复)+老数据strategy=None归"_"单套兼容;文案配对:选中文案数≥套数则一套一条否则共用第1条;每套成独立note进exporter | packaging_task.py:45-100 |
 | S7 | 新建 GET /products/export + /benchmarks/export(format=json/csv)；require_write_permission；⚠️benchmarks/export 需 Product JOIN 填 product_name(复用 _fmt_benchmark 不够,只有 product_id) | products.py + benchmarks.py |
 | S8 | 配置页产品 tab + 历史归档页各加"导出数据"按钮，通用 ExportButton(url,filename) 调 getBlob 下载 | history/components.tsx 或 config/ExportTab.tsx |
 
 ## 验收清单（做完一个勾一个）
 
-- [ ] **A1**〔auto〕日期筛选：传 date_from/date_to → items 全落区间；传 date_from=2099 → 空数组 total=0。⚠️审计补:再验 `date_from=not-a-date` → code=40001(非500)。
+- [x] **A1**〔auto〕日期筛选：✅2026-06-18 ws3实测:全量79条→当日(2026-06-17)区间过滤=4条,未来(2099+)=0条;非法日期"not-a-date"→raise_param_error返400/40001(非500)。date_to含当日(+1天开区间)。
-- [ ] **A2**〔auto〕product_id 筛选：传 product_id=1 → 每条 product_id=1 且 product_name 非空。
+- [x] **A2**〔auto〕product_id 筛选：✅2026-06-18 ws3实测:product_id=1过滤命中78条,product_name批量查到'素颜霜'(防N+1:set收集→一次in_查询建name_map)。
-- [ ] **A3**〔human〕前端产品列：/history 表头有"产品"列，每行显产品名(非数字)，与 GET /products/{id}.name 一致。
+- [ ] **A3**〔human〕前端产品列：/history 表头有"产品"列，每行显产品名(非数字)，与 GET /products/{id}.name 一致。【代码已就绪:components.tsx加产品列+dto加product_name,typecheck过+独立审核过;待Z阶段浏览器目视】
-- [ ] **A4**〔human〕日期控件：有开始/结束 date input，填后表格刷新，XHR URL 含 date_from/date_to，结果≤全量。
+- [ ] **A4**〔human〕日期控件：有开始/结束 date input，填后表格刷新，XHR URL 含 date_from/date_to，结果≤全量。【代码已就绪:HistoryFilters.tsx日期input+产品下拉+清除,page接入筛选变化重置页码;待Z阶段浏览器目视】
 - [ ] **A5**〔human〕下载交付包：approved 任务点"下载交付包"，≤60s 触发下载 delivery-{id}.zip(>1KB)；`unzip -l` 含 note_01/文案.txt/发布清单/合规说明。
-- [ ] **A6**〔auto〕zip 结构完整：`unzip -l` grep 文案.txt/发布清单/合规说明/note_01//.jpg 全非空。⚠️审计补:若 .jpg 空先查 ImageCandidate.url 文件容器内可达(packaging_task.py:64-70 读失败静默跳过)。
+- [x] **A6**〔auto〕zip 结构完整：✅2026-06-18 task74端到端验:zip解出note_01/02/03各7项(6图+文案.txt),发布清单+合规说明在root,18张图全非空真写入(磁盘文件真读到)。
-- [ ] **A7**〔auto〕文案.txt 格式：`unzip -p note_01/文案.txt` 含"【标题】"、不以 `{` 开头(JSON 已解析)。
+- [x] **A7**〔auto〕文案.txt 格式：✅2026-06-18 task74验:note_02/文案.txt以"【标题】懒人三分钟搞定,倍分子上脸气色回来了"开头(非{,JSON已解析),品牌词"倍分子"已植入。
-- [ ] **A8**〔human+auto〕多套打包：⚠️审计改 human(需先 seed 3条 is_selected TextCandidate)→构造后 `unzip -l|grep 'note_0[1-3]/'|wc -l`≥3，三套标题互不相同。
+- [x] **A8**〔human+auto〕多套打包：✅2026-06-18 task74(6图×3套A/B/C)端到端真跑:选1条文案→建package→build_delivery_package→`grep note_0[1-3]`=3套各6图,每套独立note_0N夹+文案.txt。独立agent交叉验证7条全过(分组/去重/老数据兼容/越界/异常/契约/红线)。⚠️攒批问倩倩姐:当前只生成选1条文案,三套图共用同文案→交付包"3套"仅图风格(A/B/C)不同、文案全同,品牌词"倍分子"已正确植入。是否要三套各配不同角度文案?
-- [ ] **A9**〔auto〕产品 JSON 导出：GET /products/export → 200,application/json,数组,每元素含 id/name/selling_points,且 selling_points 是数组(非JSON字符串)。
+- [x] **A9**〔auto〕产品 JSON 导出：✅2026-06-18 ws3实测:export_products返3条,selling_points类型=list(['烟酰胺改善暗沉提亮','水解珍珠锁水匀肤']非JSON字符串);端点真注册路由表+顺序在/products/{product_id}之前(idx9<13不被吞)。
-- [ ] **A10**〔auto〕标杆 CSV 导出：GET /benchmarks/export?format=csv → 200,text/csv,首行表头含 product_id/highlights/created_at,行数≥已录标杆数。
+- [ ] **A10**〔auto〕标杆 CSV 导出：⏸️逻辑已审(CSV带UTF-8 BOM+JOIN Product填product_name+json/csv双格式),但ws3无标杆数据未触发真导出;待录标杆后验或Z阶段一条龙带验。
-- [ ] **A11**〔auto〕导出权限：⚠️审计注——require_write_permission 当前只查 workspace_members 不分角色(workspace_guard.py:65-88)；未登录调 → code=40101，workspace 成员 → 200。
+- [ ] **A11**〔auto〕导出权限：⚠️审计注——require_write_permission 当前只查 workspace_members 不分角色(workspace_guard.py:65-88)；未登录调 → code=40101，workspace 成员 → 200。【两端点均挂require_write_permission,独立审核确认】
-- [ ] **A12**〔human〕不显 token 余额(红线)：⚠️审计改 human(curl 拿不到Next.js水合后DOM)——浏览器 Ctrl+F 搜"余额/用量/tokens/balance"无结果。
+- [ ] **A12**〔human〕不显 token 余额(红线)：⚠️审计改 human(curl 拿不到Next.js水合后DOM)——浏览器 Ctrl+F 搜"余额/用量/tokens/balance"无结果。【HistoryFilters/page/components均无余额字段,独立审核确认】
 
 ## 待北哥定义（不阻断代码）
 - 多套打包图片套间对应：ImageCandidate 无 set_index，是否加列关联文案？还是退化"图共享复制进每套夹"(本期默认退化)。