# -*- coding: utf-8 -*-
"""
AI 聊天接口
功能：对话管理 + 消息存库 + function calling（工具调用）
"""

from fastapi import APIRouter, HTTPException, Depends
from pydantic import BaseModel
from typing import List, Optional
import json, base64, re, logging
from sqlalchemy.orm import Session as DBSession
from app.core.config import settings
from app.core.security import get_current_user
from app.db import get_db
from app.models.user import User
from app.models.chat import ChatSession, ChatMessage
from app.api.v1.ai_tools import PS_TOOLS, TOOL_DISPLAY_NAMES

log = logging.getLogger(__name__)
# 防止 uvicorn --reload 导致 handler 重复
if not log.handlers:
    log.setLevel(logging.INFO)
    log.propagate = False
    _h = logging.StreamHandler()
    _h.setFormatter(logging.Formatter('[%(asctime)s][%(name)s][%(levelname)s] %(message)s'))
    log.addHandler(_h)

router = APIRouter()

# ==================== 数据模型 ====================

class ChatRequest(BaseModel):
    message: str
    session_id: Optional[int] = None
    image_base64: Optional[str] = None
    model: Optional[str] = None           # 前端指定对话模型（覆盖默认）
    vision_model: Optional[str] = None    # 前端指定视觉模型
    image_edit_model: Optional[str] = None  # 前端指定图片编辑模型

class ToolResultRequest(BaseModel):
    """前端执行完工具后回传结果"""
    session_id: int
    tool_call_id: str
    tool_name: str
    result: str
    model: Optional[str] = None           # 前端指定对话模型
    vision_model: Optional[str] = None
    image_edit_model: Optional[str] = None

class GeneratePatternRequest(BaseModel):
    """从成衣图片生成面料花样"""
    image_base64: str          # 成衣图片 base64
    canvas_base64: Optional[str] = None  # 裁片轮廓截图 base64
    prompt: Optional[str] = None  # 可选自定义提示词
    image_edit_model: Optional[str] = None

class VerifyResultRequest(BaseModel):
    """验证套图结果"""
    garment_base64: str       # 原始成衣图片
    canvas_base64: str        # 套图后的画布截图
    prompt: Optional[str] = None
    vision_model: Optional[str] = None

# ==================== System Prompt ====================

SYSTEM_PROMPT = """你是 DesignerCEP 的 AI 助手，运行在 Adobe Photoshop CEP 插件中。

你的能力：
1. 回答关于 Photoshop 操作和插件使用的问题
2. 通过工具直接操作 Photoshop（创建图层、对齐、查看文档信息等）
3. 帮用户排查操作中遇到的错误
4. **AI 智能套图**：两阶段流程 — 先生成预览确认，再提取套到裁片上

## AI 智能套图流程（两阶段）

当用户上传成衣图片并要求套图时：

### 阶段 1 — 识别裁片 + 生成预览（需用户确认）
1. 调用 **identify_pieces** — 截取画布并识别每个图层是什么裁片部位（前片、后片、袖子等）
2. 告诉用户识别结果（如 "M-1=前片, M-2=后片, M-5=左袖..."）
3. 调用 **generate_garment_preview** — 会自动在裁片下方标注名称标签（如 "M-1 前片"），然后截取带标签的画布 + 成衣照片一起发给 AI 生成预览
4. 预览图显示在聊天中，每个裁片都有标签，**等用户确认 OK 后**进入阶段 2

### 阶段 2 — 提取花样 + 正式套图（用户确认后）
5. 根据用户反馈决定每个裁片的处理方式：
   - 需要花样的裁片 → extract_and_apply_all_pieces 中不设 color 字段
   - 纯色的裁片（如后幅） → 设置 color 字段为对应颜色值（如 "#F5E6D0"），直接 PS 填充不调 AI
6. 调用 **extract_and_apply_all_pieces** 执行套图
7. 可选：调用 **verify_pattern_result** 验证效果

重要：
- 阶段 1 完成后必须**等用户说"可以"/"OK"**才执行阶段 2
- 用户可能会说"袖子纯色"、"后幅不要花样"等，要根据 identify_pieces 的结果对应到正确的图层名
- 纯色裁片用 color 字段直接填充，不要浪费 AI 提取调用

重要规则：
- 当用户要求执行 PS 操作时，使用工具完成
- 执行操作前可以先了解当前文档和图层状态
- 用简洁的中文回答，适合在小面板中阅读
- 如果工具执行失败，向用户解释原因并建议解决方案
- 套图时应一次性完成「生成 → 套图 → 验证」全流程，不要中途停下等待用户指令
"""

VISION_PROMPT = """你是一位资深的服装设计分析师，同时也是 DesignerCEP 的 AI 助手。

当用户发送服装/成衣图片时，请从以下维度进行专业分析：

1. **服装类别** — 上衣/裤子/裙子/连衣裙/外套/配饰等，细分款式
2. **面料分析** — 根据视觉特征推测面料类型（棉、涤纶、丝绸、针织、牛仔、雪纺等），分析面料质感
3. **颜色与印花** — 主色调、配色方案、印花/图案类型及工艺（数码印花、丝网印刷、提花等）
4. **版型特点** — 修身/宽松/A字/H型等，分析领口、袖型、肩线、腰线、下摆处理
5. **工艺细节** — 缝线工艺、拉链/纽扣/暗扣、口袋设计、装饰细节、包边/锁边
6. **设计评价** — 设计亮点、风格定位（休闲/正装/运动/时尚等）、目标消费群体
7. **改进建议** — 如有可改进之处，给出专业建议

规则：
- 如果图片不是服装相关，也请尽力分析图片内容并给出有价值的反馈
- 如果用户同时提了文字问题，请结合图片和问题一起回答
- 用清晰、结构化的中文回答，适合在设计工作中参考
- 回答要专业但不啰嗦，突出重点信息
"""

# ==================== 对话管理接口 ====================

@router.get("/ai/models")
async def get_models(current_username: str = Depends(get_current_user)):
    """获取可用模型列表和当前默认值"""
    return {
        "code": 200,
        "data": {
            "chat_model": settings.AI_MODEL,
            "vision_model": settings.AI_VISION_MODEL,
            "image_edit_model": settings.AI_IMAGE_EDIT_MODEL,
            "chat_models": [
                {"id": "qwen3-max-2026-01-23", "name": "Qwen3 Max"},
                {"id": "gemini-2.5-pro-thinking", "name": "Gemini 2.5 Pro"},
            ],
            "vision_models": [
                {"id": "qwen-vl-max-latest", "name": "Qwen VL Max"},
                {"id": "gemini-2.5-flash-image", "name": "Gemini 2.5 Flash"},
            ],
            "image_edit_models": [
                {"id": "qwen-image-edit-max-2026-01-16", "name": "Qwen 图片编辑"},
                {"id": "gemini-3-pro-image-preview", "name": "Gemini 3 Pro 生图"},
            ],
        }
    }


@router.get("/ai/sessions")
async def list_sessions(
    db: DBSession = Depends(get_db),
    current_username: str = Depends(get_current_user)
):
    """获取当前用户的对话列表"""
    user = db.query(User).filter(User.username == current_username).first()
    if not user:
        raise HTTPException(status_code=404, detail="用户不存在")

    sessions = db.query(ChatSession)\
        .filter(ChatSession.user_id == user.id)\
        .order_by(ChatSession.updated_at.desc())\
        .limit(50).all()

    return {
        "code": 200,
        "data": [{
            "id": s.id,
            "title": s.title,
            "created_at": s.created_at.isoformat() if s.created_at else None,
            "updated_at": s.updated_at.isoformat() if s.updated_at else None,
        } for s in sessions]
    }


@router.get("/ai/sessions/{session_id}/messages")
async def get_session_messages(
    session_id: int,
    db: DBSession = Depends(get_db),
    current_username: str = Depends(get_current_user)
):
    """获取某个对话的全部消息"""
    user = db.query(User).filter(User.username == current_username).first()
    if not user:
        raise HTTPException(status_code=404, detail="用户不存在")

    session = db.query(ChatSession).filter(
        ChatSession.id == session_id, ChatSession.user_id == user.id
    ).first()
    if not session:
        raise HTTPException(status_code=404, detail="对话不存在")

    messages = db.query(ChatMessage)\
        .filter(ChatMessage.session_id == session_id)\
        .order_by(ChatMessage.created_at).all()

    return {
        "code": 200,
        "data": {
            "session_id": session.id,
            "title": session.title,
            "messages": [{
                "id": m.id, "role": m.role, "content": m.content,
                "tool_calls": m.tool_calls,
                "created_at": m.created_at.isoformat() if m.created_at else None,
            } for m in messages]
        }
    }


@router.delete("/ai/sessions/{session_id}")
async def delete_session(
    session_id: int,
    db: DBSession = Depends(get_db),
    current_username: str = Depends(get_current_user)
):
    """删除对话"""
    user = db.query(User).filter(User.username == current_username).first()
    if not user:
        raise HTTPException(status_code=404, detail="用户不存在")

    session = db.query(ChatSession).filter(
        ChatSession.id == session_id, ChatSession.user_id == user.id
    ).first()
    if not session:
        raise HTTPException(status_code=404, detail="对话不存在")

    db.query(ChatMessage).filter(ChatMessage.session_id == session_id).delete()
    db.delete(session)
    db.commit()
    return {"code": 200, "message": "对话已删除"}


# ==================== 聊天接口 ====================

@router.post("/ai/chat")
async def chat(
    data: ChatRequest,
    db: DBSession = Depends(get_db),
    current_username: str = Depends(get_current_user)
):
    """发送消息 → AI 回复（可能包含工具调用指令）"""
    log.info(f"{'='*60}")
    log.info(f"[Chat] 收到消息: '{data.message[:80]}' session={data.session_id} 有图片={'是' if data.image_base64 else '否'}")
    user = db.query(User).filter(User.username == current_username).first()
    if not user:
        raise HTTPException(status_code=404, detail="用户不存在")

    # 1. 获取或创建会话
    session = None
    if data.session_id:
        session = db.query(ChatSession).filter(
            ChatSession.id == data.session_id, ChatSession.user_id == user.id
        ).first()

    if not session:
        title = data.message[:30] + ("..." if len(data.message) > 30 else "")
        session = ChatSession(user_id=user.id, username=current_username, title=title)
        db.add(session)
        db.flush()

    # 2. 保存用户消息（图片消息加标记，不存 base64）
    has_image = bool(data.image_base64)
    save_content = ("[图片分析] " + data.message) if has_image else data.message
    db.add(ChatMessage(session_id=session.id, role="user", content=save_content))
    db.commit()

    # 3. 加载历史
    history = db.query(ChatMessage)\
        .filter(ChatMessage.session_id == session.id)\
        .order_by(ChatMessage.created_at).all()
    # tool 角色消息转为 user（避免 LLM 验证 tool_calls 结构报错）
    history_list = [{"role": m.role if m.role != "tool" else "user", "content": m.content} for m in history[-20:]]

    # 4. 调用 LLM（统一走工具模型，图片信息通过文本标记传递）
    try:
        if not settings.AI_API_KEY and not (data.model and _is_gemini_model(data.model)):
            reply_content = _mock_reply(data.message, has_image)
            tool_calls_data = None
        else:
            call_history = history_list
            # 有图片时：在消息中标注，让工具模型知道图片已上传可供工具使用
            if has_image:
                call_history = history_list[:-1] + [{
                    "role": "user",
                    "content": f"[用户上传了一张图片，已保存可供工具使用] {data.message or '请处理这张图片'}"
                }]

            # 根据模型类型路由
            if data.model and _is_gemini_model(data.model):
                log.info(f"[Chat] 路由到 Gemini: {data.model}")
                reply_content, tool_calls_data = _call_gemini_with_tools(call_history, data.model)
            else:
                reply_content, tool_calls_data = _call_llm_with_tools(call_history, model_override=data.model)
    except Exception as e:
        reply_content = f"AI 请求出错: {str(e)}"
        tool_calls_data = None

    # 5. 保存 AI 回复
    tc_json = json.dumps(tool_calls_data, ensure_ascii=False) if tool_calls_data else None
    db.add(ChatMessage(
        session_id=session.id, role="assistant",
        content=reply_content or "", tool_calls=tc_json
    ))
    db.commit()

    return {
        "code": 200,
        "data": {
            "session_id": session.id,
            "content": reply_content or "",
            "tool_calls": tool_calls_data
        }
    }


@router.post("/ai/tool-result")
async def submit_tool_result(
    data: ToolResultRequest,
    db: DBSession = Depends(get_db),
    current_username: str = Depends(get_current_user)
):
    """前端执行完工具后回传结果，后端继续让 AI 总结"""
    log.info(f"{'='*60}")
    log.info(f"[ToolResult] 工具: {data.tool_name}, 结果: {data.result[:150]}")
    user = db.query(User).filter(User.username == current_username).first()
    if not user:
        raise HTTPException(status_code=404, detail="用户不存在")

    session = db.query(ChatSession).filter(
        ChatSession.id == data.session_id, ChatSession.user_id == user.id
    ).first()
    if not session:
        raise HTTPException(status_code=404, detail="对话不存在")

    # 1. 保存工具结果为一条消息
    db.add(ChatMessage(
        session_id=session.id, role="tool",
        content=f"[工具 {data.tool_name} 执行结果]: {data.result}"
    ))
    db.commit()

    # 2. 重新加载历史，让 AI 根据工具结果继续回答
    history = db.query(ChatMessage)\
        .filter(ChatMessage.session_id == session.id)\
        .order_by(ChatMessage.created_at).all()
    history_list = [{"role": m.role if m.role != "tool" else "user", "content": m.content} for m in history[-20:]]

    # 3. 再次调用 LLM（这次不带 tools，让 AI 总结结果）
    try:
        if not settings.AI_API_KEY and not (data.model and _is_gemini_model(data.model)):
            reply_content = f"工具 {data.tool_name} 执行完成。"
            tool_calls_data = None
        else:
            if data.model and _is_gemini_model(data.model):
                log.info(f"[ToolResult] 路由到 Gemini: {data.model}")
                reply_content, tool_calls_data = _call_gemini_with_tools(history_list, data.model)
            else:
                reply_content, tool_calls_data = _call_llm_with_tools(history_list, model_override=data.model)
    except Exception as e:
        reply_content = f"AI 总结出错: {str(e)}"
        tool_calls_data = None

    # 4. 保存 AI 总结
    tc_json = json.dumps(tool_calls_data, ensure_ascii=False) if tool_calls_data else None
    db.add(ChatMessage(
        session_id=session.id, role="assistant",
        content=reply_content or "", tool_calls=tc_json
    ))
    db.commit()

    return {
        "code": 200,
        "data": {
            "session_id": session.id,
            "content": reply_content or "",
            "tool_calls": tool_calls_data
        }
    }


# ==================== LLM 调用 ====================

def _call_llm_with_tools(messages_history: List[dict], model_override: str = None):
    """调用 LLM，支持 function calling"""
    from openai import OpenAI

    use_model = model_override or settings.AI_MODEL

    client = OpenAI(
        api_key=settings.AI_API_KEY,
        base_url=settings.AI_BASE_URL or "https://api.openai.com/v1",
    )

    messages = [{"role": "system", "content": SYSTEM_PROMPT}] + messages_history

    # ---- 日志：请求详情 ----
    log.info(f"{'='*60}")
    log.info(f"[LLM] 调用工具模型: {use_model}{' (override)' if model_override else ''}")
    log.info(f"[LLM] 消息数量: {len(messages)} (system + {len(messages_history)} history)")
    for i, m in enumerate(messages_history[-5:]):  # 只打最近 5 条
        role = m['role']
        content = m['content'][:120] if m.get('content') else '(empty)'
        log.info(f"[LLM] history[-{len(messages_history)-i}] {role}: {content}")
    log.info(f"[LLM] 工具数量: {len(PS_TOOLS)}")

    completion = client.chat.completions.create(
        model=use_model,
        messages=messages,
        tools=PS_TOOLS,
        tool_choice="auto",
    )

    choice = completion.choices[0]
    message = choice.message

    # ---- 日志：响应详情 ----
    log.info(f"[LLM] 响应 finish_reason={choice.finish_reason}")
    if message.content:
        log.info(f"[LLM] 回复文本: {message.content[:150]}...")
    if message.tool_calls:
        for tc in message.tool_calls:
            log.info(f"[LLM] 工具调用: {tc.function.name}({tc.function.arguments[:200]})")
    else:
        log.info(f"[LLM] 无工具调用")
    log.info(f"{'='*60}")

    # 检查是否有工具调用
    if message.tool_calls and len(message.tool_calls) > 0:
        tool_calls_data = []
        for tc in message.tool_calls:
            args = {}
            if tc.function.arguments:
                try:
                    args = json.loads(tc.function.arguments)
                except json.JSONDecodeError:
                    args = {}

            tool_calls_data.append({
                "id": tc.id,
                "name": tc.function.name,
                "display_name": TOOL_DISPLAY_NAMES.get(tc.function.name, tc.function.name),
                "args": args,
                "status": "pending"
            })

        return message.content or "", tool_calls_data

    # 普通文本回复
    return message.content or "", None


# ==================== Gemini 调用 ====================

def _is_gemini_model(model_name: str) -> bool:
    """判断是否是 Gemini 模型"""
    return model_name and "gemini" in model_name.lower()


def _call_gemini(messages_history: List[dict], model: str, images_b64: List[str] = None) -> str:
    """
    调用 Gemini API（通过第三方代理，OpenAI 兼容格式）
    支持纯文本对话和图片输入（视觉分析）
    返回文本内容
    """
    from openai import OpenAI as _OpenAI

    if not settings.GEMINI_API_KEY or not settings.GEMINI_BASE_URL:
        raise ValueError("GEMINI_API_KEY 或 GEMINI_BASE_URL 未配置")

    client = _OpenAI(
        api_key=settings.GEMINI_API_KEY,
        base_url=f"{settings.GEMINI_BASE_URL}/v1",
    )

    # 构造 OpenAI 格式消息
    messages = []
    for msg in messages_history:
        role = msg.get("role", "user")
        content = msg.get("content", "")
        # OpenAI 格式支持 system / user / assistant
        if role not in ("system", "user", "assistant"):
            role = "user"
        messages.append({"role": role, "content": content})

    # 如果有图片，把最后一条 user 消息改成多模态格式
    if images_b64:
        # 找到最后一条 user 消息
        last_user_idx = None
        for i in range(len(messages) - 1, -1, -1):
            if messages[i]["role"] == "user":
                last_user_idx = i
                break

        if last_user_idx is not None:
            text_content = messages[last_user_idx]["content"]
            multimodal_content = []
            # 先放图片
            for img_b64 in images_b64:
                multimodal_content.append({
                    "type": "image_url",
                    "image_url": {"url": f"data:image/jpeg;base64,{img_b64}"}
                })
            # 再放文本
            multimodal_content.append({"type": "text", "text": text_content})
            messages[last_user_idx]["content"] = multimodal_content

    log.info(f"{'='*60}")
    log.info(f"[Gemini] 调用模型: {model} (OpenAI 兼容)")
    log.info(f"[Gemini] 消息数: {len(messages)}, 图片数: {len(images_b64) if images_b64 else 0}")

    completion = client.chat.completions.create(
        model=model,
        messages=messages,
    )

    result = completion.choices[0].message.content or ""
    log.info(f"[Gemini] 回复: {result[:200]}...")
    return result


def _call_gemini_with_tools(messages_history: List[dict], model: str) -> tuple:
    """
    用 Gemini 做对话（不支持 function calling，靠 prompt 引导调用工具）
    返回 (content, tool_calls_data) — tool_calls_data 始终为 None
    注意：Gemini 不原生支持 function calling，但文本对话正常
    """
    # 加入 system prompt
    full_history = [{"role": "system", "content": SYSTEM_PROMPT}] + messages_history

    log.info(f"{'='*60}")
    log.info(f"[Gemini] 调用对话模型: {model}")
    for i, m in enumerate(messages_history[-3:]):
        log.info(f"[Gemini] history[-{len(messages_history)-i}] {m['role']}: {str(m.get('content',''))[:120]}")

    result = _call_gemini(full_history, model)

    log.info(f"[Gemini] 回复文本: {result[:150]}...")
    log.info(f"[Gemini] 无工具调用（Gemini 不支持 function calling）")

    return result, None


def _call_vision_llm(user_message: str, image_base64: str, history: List[dict], model_override: str = None) -> str:
    """调用视觉模型分析图片（自动路由 Qwen / Gemini）"""

    use_model = model_override or settings.AI_VISION_MODEL

    # ---------- Gemini 路由 ----------
    if _is_gemini_model(use_model):
        log.info(f"[Vision] 使用 Gemini 视觉模型: {use_model}")
        msgs = [{"role": "system", "content": VISION_PROMPT}]
        for h in history[-10:]:
            role = h["role"] if h["role"] != "tool" else "user"
            msgs.append({"role": role, "content": h["content"]})
        msgs.append({"role": "user", "content": user_message})
        return _call_gemini(msgs, use_model, images_b64=[image_base64])

    # ---------- Qwen / OpenAI 路由 ----------
    from openai import OpenAI

    client = OpenAI(
        api_key=settings.AI_API_KEY,
        base_url=settings.AI_BASE_URL or "https://api.openai.com/v1",
    )

    # 构造多模态消息
    user_content = [
        {
            "type": "image_url",
            "image_url": {"url": f"data:image/jpeg;base64,{image_base64}"}
        },
        {
            "type": "text",
            "text": user_message
        }
    ]

    # 组装消息列表（历史 + 当前图片消息）
    messages = [
        {"role": "system", "content": VISION_PROMPT},
    ]
    # 添加近期历史（纯文本）
    for h in history[-10:]:
        role = h["role"] if h["role"] != "tool" else "user"
        messages.append({"role": role, "content": h["content"]})
    # 当前带图片的用户消息
    messages.append({"role": "user", "content": user_content})

    completion = client.chat.completions.create(
        model=use_model,
        messages=messages,
    )

    return completion.choices[0].message.content or ""


def _mock_reply(message: str, has_image: bool = False) -> str:
    """未配置 API Key 时的模拟回复"""
    if has_image:
        return "【模拟分析】收到图片。AI 分析功能需要配置 AI_API_KEY 和 AI_VISION_MODEL。\n\n配置后可以分析：服装类别、面料、颜色印花、版型、工艺细节等。"
    if "套图" in message:
        return "套图功能在「参数预设」页面。先选择花样组和裁片组，再添加规则，最后点击生成。"
    elif "对齐" in message:
        return "图层对齐功能在「参数预设」页面顶部。支持上下左右居中对齐、领口对齐等。"
    elif "裁片" in message or "PLT" in message:
        return "PLT 裁片处理在「PLT 裁片处理」页面。上传 PLT 文件，选择尺码，点击开始处理。"
    else:
        return "你好！我是 DesignerCEP AI 助手。你可以问我关于套图、裁片、对齐等功能的问题。"


# ==================== 图案生成 & 验证 ====================

@router.post("/ai/generate-preview")
async def generate_preview(
    data: GeneratePatternRequest,
    current_username: str = Depends(get_current_user)
):
    """阶段1：成衣照片 + 裁片截图 → AI 生成带轮廓的花样预览图"""
    log.info(f"{'='*60}")
    log.info(f"[Preview] 收到预览请求, 成衣图: {len(data.image_base64)//1024}KB, 裁片图: {len(data.canvas_base64)//1024 if data.canvas_base64 else 0}KB")
    if data.prompt:
        log.info(f"[Preview] 自定义提示词: {data.prompt[:150]}")
    if not settings.AI_API_KEY or not settings.AI_IMAGE_EDIT_MODEL:
        raise HTTPException(400, "AI_API_KEY 或 AI_IMAGE_EDIT_MODEL 未配置")

    try:
        # 构造图片列表：成衣 + 裁片截图（如果有）
        images = [data.image_base64]
        if data.canvas_base64:
            images.append(data.canvas_base64)

        default_prompt = (
            "Image 1 是一件成衣照片，Image 2 是裁片轮廓图。"
            "请将 Image 1 成衣上的面料花样提取出来，填充到 Image 2 的每个裁片轮廓内。"
            "要求：保持花样颜色、比例、方向一致；保留裁片轮廓线；不要改变裁片排列位置。"
        ) if data.canvas_base64 else (
            "请从这件成衣中提取面料花样，生成一张干净的花样平铺图。"
        )

        result_url, desc = _call_image_model(
            images_b64=images,
            prompt=data.prompt or default_prompt,
            model_override=data.image_edit_model,
        )
        return {"code": 200, "data": {"image_url": result_url, "description": desc}}
    except Exception as e:
        log.error(f"预览生成失败: {e}", exc_info=True)
        err_msg = str(e)
        if "data_inspection_failed" in err_msg or "inappropriate" in err_msg:
            raise HTTPException(400, "图片内容未通过平台安全审核，请更换图片后重试（避免使用含版权角色/敏感内容的图片）")
        raise HTTPException(500, f"预览生成失败: {err_msg}")


class CropPieceRequest(BaseModel):
    """从预览图中按坐标裁切指定裁片区域"""
    preview_base64: str    # 预览图 base64（从 URL 下载后传入）
    canvas_width: int      # PS 画布原始宽度
    canvas_height: int     # PS 画布原始高度
    piece_left: float      # 裁片在画布中的 left (px)
    piece_top: float       # 裁片在画布中的 top (px)
    piece_width: float     # 裁片宽度 (px)
    piece_height: float    # 裁片高度 (px)
    piece_name: str        # 裁片名称（用于日志）
    padding: float = 0.15  # 出血线比例（向外扩 15%）


@router.post("/ai/crop-piece")
async def crop_piece(
    data: CropPieceRequest,
    current_username: str = Depends(get_current_user)
):
    """从预览图中按坐标裁切指定裁片区域（Pillow，像素级精准）"""
    log.info(f"{'='*60}")
    log.info(f"[CropPiece] 裁切: {data.piece_name}")
    log.info(f"[CropPiece] 画布: {data.canvas_width}x{data.canvas_height}, 裁片区域: ({data.piece_left},{data.piece_top}) {data.piece_width}x{data.piece_height}")

    try:
        result_b64 = _crop_piece_from_preview(data)
        return {"code": 200, "data": {"cropped_base64": result_b64, "piece_name": data.piece_name}}
    except Exception as e:
        log.error(f"裁切失败: {e}", exc_info=True)
        raise HTTPException(500, f"裁切失败: {str(e)}")


class RefinePieceRequest(BaseModel):
    """AI 细化裁切后的花样图"""
    cropped_base64: str    # 裁切后的图片 base64
    piece_name: str        # 裁片名称
    pattern_type: str = "fill_pattern"  # fill_pattern / theme_pattern / mixed_fill / mixed_theme
    aspect_ratio: Optional[str] = None  # 裁片宽高比，如 "3:4"
    piece_width: Optional[int] = None   # 裁片像素宽
    piece_height: Optional[int] = None  # 裁片像素高
    prompt: Optional[str] = None
    image_edit_model: Optional[str] = None


@router.post("/ai/refine-piece")
async def refine_piece(
    data: RefinePieceRequest,
    current_username: str = Depends(get_current_user)
):
    """AI 细化裁切图：根据图案类型生成不同提示词"""
    log.info(f"{'='*60}")
    log.info(f"[RefinePiece] 细化: {data.piece_name}, 类型: {data.pattern_type}, 比例: {data.aspect_ratio}")
    log.info(f"[RefinePiece] 裁片尺寸: {data.piece_width}x{data.piece_height}, 图片: {len(data.cropped_base64)//1024}KB")

    if not settings.AI_API_KEY or not settings.AI_IMAGE_EDIT_MODEL:
        raise HTTPException(400, "AI_API_KEY 或 AI_IMAGE_EDIT_MODEL 未配置")

    try:
        # 尺寸描述
        size_hint = ""
        if data.aspect_ratio:
            size_hint = f"输出图片宽高比必须为 {data.aspect_ratio}。"
        if data.piece_width and data.piece_height:
            size_hint += f"输出尺寸至少 {int(data.piece_width * 1.15)}x{int(data.piece_height * 1.15)} 像素。"

        # 根据类型生成提示词
        if data.pattern_type == "fill_pattern":
            prompt = data.prompt or (
                f"这是从服装裁片中裁切出来的花型图案区域。"
                f"请将它处理为一张干净的矩形花型图："
                f"去掉所有轮廓线、标签文字和边缘空白；"
                f"花型图案要完整铺满整个矩形，无空白边距；"
                f"向四周自然延展重复纹样，保持花型连续无接缝；"
                f"保持花样的颜色和细节清晰。{size_hint}"
            )
        elif data.pattern_type == "theme_pattern":
            prompt = data.prompt or (
                f"这是从服装裁片中裁切出来的主题图案区域（如卡通人物、Logo、印花）。"
                f"请提取其中的主题图案，输出为纯白色背景(#FFFFFF)上的主题图案："
                f"主题图案要完整清晰，保持原始颜色和细节；"
                f"背景必须是纯白色(#FFFFFF)，不要任何其他底色或纹理；"
                f"去掉轮廓线和标签文字；主题图案在画面中居中放置。{size_hint}"
            )
        elif data.pattern_type == "mixed_fill":
            prompt = data.prompt or (
                f"这是从服装裁片中裁切出来的区域，包含主题图案和花型底纹。"
                f"请只提取底部的花型/纹理图案，去掉上面的主题图案（人物/Logo）："
                f"用底纹自然填充主题图案原来的位置；"
                f"花型要完整铺满整个矩形，保持纹样连续；"
                f"去掉轮廓线和标签文字。{size_hint}"
            )
        elif data.pattern_type == "mixed_theme":
            prompt = data.prompt or (
                f"这是从服装裁片中裁切出来的区域，包含主题图案和花型底纹。"
                f"请只提取主题图案（人物/Logo/大面积印花），输出为纯白色背景(#FFFFFF)："
                f"主题图案要完整清晰，保持原始颜色和细节；"
                f"背景必须是纯白色(#FFFFFF)，去掉所有底纹；"
                f"去掉轮廓线和标签文字；主题图案在画面中居中放置。{size_hint}"
            )
        else:
            prompt = data.prompt or f"细化这张图片，去掉轮廓线和标签文字，保持清晰。{size_hint}"

        log.info(f"[RefinePiece] 提示词: {prompt[:200]}")

        result_url, desc = _call_image_model(
            images_b64=[data.cropped_base64],
            prompt=prompt,
            model_override=data.image_edit_model,
        )
        return {"code": 200, "data": {"refined_url": result_url, "piece_name": data.piece_name, "pattern_type": data.pattern_type}}
    except Exception as e:
        log.error(f"细化失败: {e}", exc_info=True)
        err_msg = str(e)
        if "data_inspection_failed" in err_msg:
            raise HTTPException(400, "图片内容未通过安全审核")
        raise HTTPException(500, f"细化失败: {err_msg}")


def _crop_piece_from_preview(data: CropPieceRequest) -> str:
    """用 Pillow 从预览图中按 PS 坐标裁切指定区域（带出血线）"""
    from PIL import Image
    import io

    # 解码预览图
    img_bytes = base64.b64decode(data.preview_base64)
    preview_img = Image.open(io.BytesIO(img_bytes))
    pw, ph = preview_img.size

    # 坐标映射：PS 画布坐标 → 预览图像素坐标
    scale_x = pw / data.canvas_width
    scale_y = ph / data.canvas_height

    # 出血线（按裁片尺寸的百分比向外扩）
    bleed_x = data.piece_width * data.padding
    bleed_y = data.piece_height * data.padding

    # PS 坐标（含出血）→ 预览图像素
    left = max(0, (data.piece_left - bleed_x) * scale_x)
    top = max(0, (data.piece_top - bleed_y) * scale_y)
    right = min(pw, (data.piece_left + data.piece_width + bleed_x) * scale_x)
    bottom = min(ph, (data.piece_top + data.piece_height + bleed_y) * scale_y)

    log.info(f"[CropPiece] PS 画布: {data.canvas_width}x{data.canvas_height}")
    log.info(f"[CropPiece] 预览图: {pw}x{ph}")
    log.info(f"[CropPiece] 缩放比: x={scale_x:.4f}, y={scale_y:.4f}")
    log.info(f"[CropPiece] PS 裁片区域: ({data.piece_left},{data.piece_top}) {data.piece_width}x{data.piece_height}")
    log.info(f"[CropPiece] 出血线: {data.padding*100:.0f}% → x±{bleed_x:.0f}px, y±{bleed_y:.0f}px")
    log.info(f"[CropPiece] 预览图裁切: ({left:.0f},{top:.0f})-({right:.0f},{bottom:.0f}) = {right-left:.0f}x{bottom-top:.0f}px")

    cropped = preview_img.crop((int(left), int(top), int(right), int(bottom)))

    # 调试：保存裁切结果
    import os, time
    debug_dir = os.path.join(os.path.dirname(__file__), '..', '..', '..', 'debug_images')
    os.makedirs(debug_dir, exist_ok=True)
    ts = int(time.time())
    debug_path = os.path.join(debug_dir, f'{ts}_crop_{data.piece_name}.png')
    cropped.save(debug_path)
    log.info(f"[CropPiece] 裁切结果已保存: {debug_path}")

    # 转 base64
    buf = io.BytesIO()
    cropped.save(buf, format='PNG')
    return base64.b64encode(buf.getvalue()).decode()


class ExtractPieceRequest(BaseModel):
    """从预览图中提取单个裁片花样（旧版 AI 提取，保留兼容）"""
    preview_base64: str
    piece_name: str
    piece_description: str = ""
    prompt: Optional[str] = None


@router.post("/ai/extract-piece-pattern")
async def extract_piece_pattern(
    data: ExtractPieceRequest,
    current_username: str = Depends(get_current_user)
):
    """阶段2：从预览图中提取指定裁片区域 → 输出矩形花样图"""
    log.info(f"{'='*60}")
    log.info(f"[ExtractPiece] 裁片: {data.piece_name}, 描述: '{data.piece_description}', 预览图: {len(data.preview_base64)//1024}KB")
    if not settings.AI_API_KEY or not settings.AI_IMAGE_EDIT_MODEL:
        raise HTTPException(400, "AI_API_KEY 或 AI_IMAGE_EDIT_MODEL 未配置")

    try:
        prompt = data.prompt or (
            f"这是一张服装裁片花样预览图，包含多个裁片。"
            f"请提取其中【{data.piece_name}】的花样区域"
            f"{'（' + data.piece_description + '）' if data.piece_description else ''}，"
            f"将该区域的花样输出为一张完整的矩形图片。"
            f"要求：只保留花样内容，去掉轮廓线，填满整个矩形，保持清晰。"
        )
        result_url, desc = _call_image_model(
            images_b64=[data.preview_base64],
            prompt=prompt,
            model_override=getattr(data, 'image_edit_model', None),
        )
        return {"code": 200, "data": {"pattern_url": result_url, "piece_name": data.piece_name, "description": desc}}
    except Exception as e:
        log.error(f"裁片提取失败: {e}", exc_info=True)
        raise HTTPException(500, f"裁片提取失败: {str(e)}")


class IdentifyPiecesRequest(BaseModel):
    """识别裁片部位 + 分析颜色花样"""
    canvas_base64: str                    # 画布截图（裁片轮廓）
    garment_base64: Optional[str] = None  # 成衣照片（用于分析颜色花样）
    layers_info: str                      # 图层信息
    vision_model: Optional[str] = None


@router.post("/ai/identify-pieces")
async def identify_pieces(
    data: IdentifyPiecesRequest,
    current_username: str = Depends(get_current_user)
):
    """用视觉模型识别裁片部位 + 分析成衣各部位的颜色花样"""
    if not settings.AI_API_KEY:
        raise HTTPException(400, "AI_API_KEY 未配置")

    try:
        result = _identify_garment_pieces(data.canvas_base64, data.layers_info, data.garment_base64, vision_model=data.vision_model)
        return {"code": 200, "data": result}
    except Exception as e:
        log.error(f"裁片识别失败: {e}", exc_info=True)
        raise HTTPException(500, f"裁片识别失败: {str(e)}")


def _identify_garment_pieces(canvas_b64: str, layers_info: str, garment_b64: str = None, vision_model: str = None) -> dict:
    """用视觉模型分析画布+成衣，识别裁片部位并判断颜色/花样（自动路由 Qwen/Gemini）"""

    use_model = vision_model or settings.AI_VISION_MODEL

    log.info(f"{'='*60}")
    log.info(f"[IdentifyPieces] 调用视觉模型识别裁片 + 分析花样")
    log.info(f"[IdentifyPieces] 模型: {use_model}")
    log.info(f"[IdentifyPieces] 画布: {len(canvas_b64)//1024}KB, 成衣: {len(garment_b64)//1024 if garment_b64 else 0}KB")
    log.info(f"[IdentifyPieces] 图层:\n{layers_info[:500]}")

    # 构造提示词
    if garment_b64:
        prompt = f"""你需要完成两个任务：

**任务1：识别裁片部位**
Image 2 是 PS 文档中的裁片轮廓图。以下是图层信息：
{layers_info}

请根据形状、大小识别每个图层是什么服装部位（前片、后片、袖子、领口等）。

**任务2：分析成衣颜色花样**
Image 1 是成衣照片。请分析这件衣服各个部位的视觉特征：
- 哪些部位有印花/花样/图案？描述花样内容
- 哪些部位是纯色/素色？给出颜色的 hex 值
- 衣服整体的底色是什么？

用 JSON 回答，格式：
```json
{{
  "mapping": {{
    "M-1": "前片",
    "M-2": "后片"
  }},
  "analysis": [
    {{"layer": "M-1", "piece": "前片", "type": "theme_pattern", "color": "#F5E6D0", "description": "卡通女孩图案，米白色纯色底"}},
    {{"layer": "M-2", "piece": "后片", "type": "solid", "color": "#F5E6D0", "description": "米白色纯色"}},
    {{"layer": "M-3", "piece": "左袖", "type": "fill_pattern", "description": "蓝白条纹重复图案"}}
  ],
  "base_color": "#F5E6D0"
}}
```

type 只有四种：
- solid — 纯色，必须给 color（hex值）
- fill_pattern — 重复性花型（碎花/条纹/格子/波点），整片铺满相同纹样
- theme_pattern — 主题图案（卡通/Logo/大印花）在纯色底上，必须给 color（底色 hex值）
- mixed_pattern — 主题图案叠在花型底纹上（最复杂），需描述主题和底纹
只返回 JSON，不要其他文字。"""
    else:
        prompt = f"""这是一张 PS 文档裁片截图。图层信息：
{layers_info}

识别每个图层的服装部位，用 JSON 回答：
```json
{{"mapping": {{"M-1": "前片", "M-2": "后片"}}}}
```
只返回 JSON。"""

    # ---------- Gemini 路由 ----------
    if _is_gemini_model(use_model):
        log.info(f"[IdentifyPieces] 使用 Gemini 视觉: {use_model}")
        images = []
        if garment_b64:
            images.append(garment_b64)
        images.append(canvas_b64)

        content = _call_gemini(
            [{"role": "user", "content": prompt}],
            use_model,
            images_b64=images
        )
        log.info(f"[IdentifyPieces] Gemini 回复: {content[:500]}")
    else:
        # ---------- Qwen / OpenAI 路由 ----------
        from openai import OpenAI

        client = OpenAI(
            api_key=settings.AI_API_KEY,
            base_url=settings.AI_BASE_URL or "https://api.openai.com/v1",
        )

        content_parts = []
        if garment_b64:
            content_parts.append({"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{garment_b64}"}})
        content_parts.append({"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{canvas_b64}"}})
        content_parts.append({"type": "text", "text": prompt})

        completion = client.chat.completions.create(
            model=use_model,
            messages=[{"role": "user", "content": content_parts}],
        )

        content = completion.choices[0].message.content or ""
        log.info(f"[IdentifyPieces] 视觉模型回复: {content[:500]}")

    # 提取 JSON
    try:
        clean = content.strip()
        if "```" in clean:
            clean = clean.split("```")[1]
            if clean.startswith("json"):
                clean = clean[4:]
            clean = clean.strip()
        parsed = json.loads(clean)

        # 兼容旧格式（纯 mapping）和新格式（mapping + analysis）
        if "mapping" in parsed:
            return parsed
        else:
            return {"mapping": parsed}
    except json.JSONDecodeError:
        log.warning(f"[IdentifyPieces] JSON 解析失败")
        return {"mapping": {}, "raw_response": content}


@router.post("/ai/verify-result")
async def verify_result(
    data: VerifyResultRequest,
    current_username: str = Depends(get_current_user)
):
    """对比原始成衣和套图结果，给出验证反馈"""
    if not settings.AI_API_KEY:
        raise HTTPException(400, "AI_API_KEY 未配置")

    try:
        feedback = _verify_pattern_result(data.garment_base64, data.canvas_base64, data.prompt, vision_model=data.vision_model)
        return {"code": 200, "data": {"feedback": feedback}}
    except Exception as e:
        log.error(f"验证失败: {e}", exc_info=True)
        raise HTTPException(500, f"验证失败: {str(e)}")


# ==================== 公共：图片模型调用 + 响应解析 ====================

def _call_image_model(images_b64: List[str], prompt: str, model_override: str = None) -> tuple:
    """
    调用图片编辑/生成模型（自动路由 DashScope / Gemini）
    返回 (result_url_or_base64, description)
    """
    import requests as http_requests

    use_model = model_override or settings.AI_IMAGE_EDIT_MODEL

    # ---------- Gemini 图片生成路由（OpenAI 兼容格式，走代理） ----------
    if _is_gemini_model(use_model):
        log.info(f"[ImageModel] 使用 Gemini 图片模型: {use_model}")

        if not settings.GEMINI_API_KEY or not settings.GEMINI_BASE_URL:
            raise ValueError("GEMINI_API_KEY 或 GEMINI_BASE_URL 未配置")

        from openai import OpenAI as _OpenAI

        # 第三方代理走 OpenAI 兼容接口（/v1/chat/completions）
        client = _OpenAI(
            api_key=settings.GEMINI_API_KEY,
            base_url=f"{settings.GEMINI_BASE_URL}/v1",
        )

        # 构造 OpenAI 格式的多模态消息（和正常调用一样）
        content_parts = [{"type": "text", "text": prompt}]
        for img_b64 in images_b64:
            content_parts.append({
                "type": "image_url",
                "image_url": {"url": f"data:image/jpeg;base64,{img_b64}"}
            })

        log.info(f"[ImageModel] Gemini OpenAI 兼容模式")
        log.info(f"[ImageModel] 模型: {use_model}")
        log.info(f"[ImageModel] 图片数: {len(images_b64)}, 各图大小: {[len(b)//1024 for b in images_b64]}KB")
        log.info(f"[ImageModel] 提示词: {prompt[:200]}")

        completion = client.chat.completions.create(
            model=use_model,
            messages=[{"role": "user", "content": content_parts}],
        )

        result_content = completion.choices[0].message.content or ""
        log.info(f"[ImageModel] Gemini 回复长度: {len(result_content)} chars")

        # 从 markdown 中提取 base64 图片：![...](data:image/...;base64,...)
        import re
        match = re.search(r'!\[.*?\]\((data:image/(\w+);base64,([^)]+))\)', result_content)

        if not match:
            # 也尝试直接匹配 data URI（有些响应不带 markdown 格式）
            match2 = re.search(r'(data:image/(\w+);base64,([A-Za-z0-9+/=]+))', result_content)
            if match2:
                match = match2

        if not match:
            log.warning(f"[ImageModel] Gemini 响应中无图片，前500字: {result_content[:500]}")
            raise ValueError("Gemini 未返回图片，请检查模型是否支持图片生成")

        data_uri = match.group(1)
        img_format = match.group(2)
        image_b64 = match.group(3)

        # base64 填充修正
        padding = 4 - len(image_b64) % 4
        if padding != 4:
            image_b64 += '=' * padding

        log.info(f"[ImageModel] Gemini 返回图片: image/{img_format}, {len(image_b64)//1024}KB")

        # 保存调试图片
        import os, time as _time
        debug_dir = os.path.join(os.path.dirname(__file__), '..', '..', '..', 'debug_images')
        os.makedirs(debug_dir, exist_ok=True)
        ts = int(_time.time())
        debug_path = os.path.join(debug_dir, f'{ts}_gemini_output.{img_format}')
        try:
            with open(debug_path, 'wb') as f:
                f.write(base64.b64decode(image_b64))
            log.info(f"[ImageModel] Gemini 输出图片已保存: {debug_path}")
        except Exception as e:
            log.warning(f"[ImageModel] 保存调试图片失败: {e}")

        # 提取文本描述（去掉图片部分）
        description = re.sub(r'!\[.*?\]\(data:image/[^)]+\)', '', result_content).strip()

        # 返回 data URI（前端可直接用于 <img src>）
        image_data_uri = f"data:image/{img_format};base64,{image_b64}"
        log.info(f"[ImageModel] Gemini 生成完成")
        return image_data_uri, description

    # ---------- DashScope 原生接口 ----------
    # DashScope 图片编辑模型用原生接口，不用 /compatible-mode
    api_url = "https://dashscope.aliyuncs.com/api/v1/services/aigc/multimodal-generation/generation"

    # 构造 DashScope 原生格式的消息
    content_parts = []
    for img_b64 in images_b64:
        content_parts.append({"image": f"data:image/jpeg;base64,{img_b64}"})
    content_parts.append({"text": prompt})

    payload = {
        "model": use_model,
        "input": {
            "messages": [{
                "role": "user",
                "content": content_parts
            }]
        },
        "parameters": {
            "n": 1,
            "watermark": False,
            "prompt_extend": True,
        }
    }

    headers = {
        "Content-Type": "application/json",
        "Authorization": f"Bearer {settings.AI_API_KEY}",
    }

    log.info(f"{'='*60}")
    log.info(f"[ImageModel] 调用 DashScope 原生 API")
    log.info(f"[ImageModel] 模型: {use_model}")
    log.info(f"[ImageModel] 图片数: {len(images_b64)}, 各图大小: {[len(b)//1024 for b in images_b64]}KB")
    log.info(f"[ImageModel] 提示词: {prompt[:200]}")
    log.info(f"[ImageModel] 端点: {api_url}")

    # 调试：把发给模型的图片保存到磁盘
    import os
    debug_dir = os.path.join(os.path.dirname(__file__), '..', '..', '..', 'debug_images')
    os.makedirs(debug_dir, exist_ok=True)
    import time
    ts = int(time.time())
    for idx, img_b64 in enumerate(images_b64):
        debug_path = os.path.join(debug_dir, f'{ts}_input_{idx}.jpg')
        try:
            with open(debug_path, 'wb') as f:
                f.write(base64.b64decode(img_b64))
            log.info(f"[ImageModel] 调试图片已保存: {debug_path}")
        except Exception as e:
            log.warning(f"[ImageModel] 保存调试图片失败: {e}")

    resp = http_requests.post(api_url, json=payload, headers=headers, timeout=120)

    if resp.status_code != 200:
        error_data = resp.json() if resp.headers.get("content-type", "").startswith("application/json") else {}
        err_msg = error_data.get("message", resp.text[:300])
        err_code = error_data.get("code", "")
        log.error(f"[ImageModel] API 错误: {resp.status_code} {err_code} {err_msg}")
        if "data_inspection_failed" in str(error_data):
            raise ValueError("图片内容未通过安全审核，请更换图片")
        raise ValueError(f"图片模型调用失败({resp.status_code}): {err_msg}")

    data = resp.json()
    log.info(f"[ImageModel] 响应 keys: {list(data.keys())}")

    # 解析响应：output.choices[0].message.content[].image
    output = data.get("output", {})
    choices = output.get("choices", [])

    if not choices:
        log.warning(f"[ImageModel] 无 choices: {str(data)[:500]}")
        raise ValueError("模型未返回结果")

    content_list = choices[0].get("message", {}).get("content", [])
    result_b64 = None
    description = ""

    image_url = None
    for item in content_list:
        if isinstance(item, dict):
            if "image" in item:
                image_url = item["image"]
                log.info(f"[ImageModel] 获取到图片 URL: {image_url[:100]}...")
            elif "text" in item:
                description += item["text"]

    if not image_url:
        log.warning(f"[ImageModel] content 中无图片: {str(content_list)[:500]}")
        raise ValueError("模型未返回图片")

    # 调试：保存输出图片 URL
    log.info(f"[ImageModel] 输出图片 URL: {image_url[:120]}...")
    try:
        import os, time
        debug_dir = os.path.join(os.path.dirname(__file__), '..', '..', '..', 'debug_images')
        ts = int(time.time())
        out_resp = http_requests.get(image_url, timeout=60)
        if out_resp.status_code == 200:
            debug_path = os.path.join(debug_dir, f'{ts}_output.png')
            with open(debug_path, 'wb') as f:
                f.write(out_resp.content)
            log.info(f"[ImageModel] 输出图片已保存: {debug_path}")
    except Exception as e:
        log.warning(f"[ImageModel] 保存输出图片失败: {e}")

    return image_url, description


def _verify_pattern_result(garment_b64: str, canvas_b64: str, extra_prompt: str = None, vision_model: str = None) -> str:
    """用视觉模型对比原始成衣和套图结果（自动路由 Qwen/Gemini）"""

    use_model = vision_model or settings.AI_VISION_MODEL

    log.info(f"{'='*60}")
    log.info(f"[Verify] 调用视觉模型验证套图效果")
    log.info(f"[Verify] 模型: {use_model}")
    log.info(f"[Verify] 成衣图: {len(garment_b64)//1024}KB, 画布图: {len(canvas_b64)//1024}KB")
    if extra_prompt:
        log.info(f"[Verify] 用户补充: {extra_prompt[:100]}")

    verify_prompt = (
        "请对比这两张图片：第一张是原始成衣照片，第二张是套图结果。\n"
        "验证：1. 花样还原度（颜色/比例/方向）2. 裁片覆盖完整度 "
        "3. 对齐质量 4. 整体效果。给出评分(1-10)和具体改进建议。"
    )
    if extra_prompt:
        verify_prompt += f"\n用户补充：{extra_prompt}"

    # ---------- Gemini 路由 ----------
    if _is_gemini_model(use_model):
        log.info(f"[Verify] 使用 Gemini 视觉: {use_model}")
        return _call_gemini(
            [
                {"role": "user", "content": "你是服装套图质量检验专家。"},
                {"role": "user", "content": verify_prompt},
            ],
            use_model,
            images_b64=[garment_b64, canvas_b64]
        )

    # ---------- Qwen / OpenAI 路由 ----------
    from openai import OpenAI

    client = OpenAI(
        api_key=settings.AI_API_KEY,
        base_url=settings.AI_BASE_URL or "https://api.openai.com/v1",
    )

    completion = client.chat.completions.create(
        model=use_model,
        messages=[
            {"role": "system", "content": "你是服装套图质量检验专家。"},
            {"role": "user", "content": [
                {"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{garment_b64}"}},
                {"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{canvas_b64}"}},
                {"type": "text", "text": verify_prompt},
            ]},
        ],
    )

    return completion.choices[0].message.content or ""