1.优化隔离工作目录2.新增图像生成和编辑功能3.生成模型替换为本地flux2 klein

src/core/config.py

@@ -1,3 +1,5 @@
+from pathlib import Path
+
 from pydantic_settings import BaseSettings, SettingsConfigDict
 from pydantic import Field
 
@@ -44,3 +46,7 @@ class Settings(BaseSettings):
 
 settings = Settings()
 MONGO_URI = f"mongodb://{settings.MONGODB_USERNAME}:{settings.MONGODB_PASSWORD}@{settings.MONGODB_HOST}:{settings.MONGODB_PORT}"
+
+TOOL_DIR = Path(__file__).resolve().parent
+PROJECT_ROOT = TOOL_DIR.parent
+print(f"PROJECT_ROOT : {PROJECT_ROOT}")

src/routers/deep_agent_chat.py

@@ -1,4 +1,5 @@
 import logging
+import os
 import random
 import uuid
 import json
@@ -6,6 +7,8 @@ from typing import AsyncGenerator
 
 from fastapi import APIRouter
 from fastapi.responses import StreamingResponse
+
+from src.core.config import PROJECT_ROOT
 from src.schemas.deep_agent_chat import DeepAgentChatRequest, HistoryResponse, HistoryItem
 from langchain_core.messages import HumanMessage, SystemMessage, AIMessageChunk, ToolMessage, AIMessage, ToolMessageChunk
 
@@ -77,13 +80,15 @@ async def chat_stream(request: DeepAgentChatRequest):
     source_thread_id = request.thread_id
     checkpoint_id = request.checkpoint_id
 
-    # 构建主agent
-    main_agent = build_main_agent(request.use_report)
-
     # 1. 確定目標 thread_id
     is_branching = source_thread_id and checkpoint_id
     target_thread_id = str(uuid.uuid4())[:8] if is_branching else (source_thread_id or str(uuid.uuid4())[:8])
 
+    # 构建主agent
+    workspace_dir = os.path.join(PROJECT_ROOT, f"agent_workspace/{target_thread_id}")
+    print(f"target_thread_id : workspace_dir: {workspace_dir}")
+    main_agent = build_main_agent(request.use_report, workspace_dir)
+
     # 2. 配置參數
     temp = request.config_params.temperature if request.config_params else 0.7
 

src/schemas/deep_agent_chat.py

@@ -11,6 +11,7 @@ class AgentConfig(BaseModel):
 
 class DeepAgentChatRequest(BaseModel):
     message: str = Field(..., description="用户的输入指令")
+    # image_url: Optional[str] = Field(None, description="图片地址")  # ✅ 新增
     thread_id: Optional[str] = Field(None, description="会话线程ID，不传则开启新会话")
     checkpoint_id: Optional[str] = Field(None, description="回溯点的ID，用于从历史点开启新对话")
     config_params: Optional[AgentConfig] = None

src/server/deep_agent/agents/main_agent.py

@@ -1,21 +1,22 @@
 from pathlib import Path
 
+from daytona import Daytona
 from deepagents import create_deep_agent
 from deepagents.backends import FilesystemBackend
 from langchain.agents.middleware import SummarizationMiddleware
 from langgraph.checkpoint.mongodb import MongoDBSaver
 from langgraph.checkpoint.serde.jsonplus import JsonPlusSerializer
 from pymongo import MongoClient
+from daytona import CreateSandboxFromSnapshotParams, Daytona
+from langchain_daytona import DaytonaSandbox
 
 from src.core.config import MONGO_URI
-from src.server.deep_agent.agents.painter import painter_subagent
-from src.server.deep_agent.agents.researcher import research_subagent
+from src.server.deep_agent.agents.painter import build_painter_subagent
+from src.server.deep_agent.agents.researcher import build_researcher_subagent
 from src.server.deep_agent.agents.user_profile import user_profile_subagent
 from src.server.deep_agent.init_llm import main_llm
 from src.server.deep_agent.init_prompt import build_system_prompt
 
-TOOL_DIR = Path(__file__).resolve().parent
-PROJECT_ROOT = TOOL_DIR.parent
 client = MongoClient(MONGO_URI)
 checkpointer = MongoDBSaver(
     client=client["furniture_agent_db"],
@@ -23,22 +24,42 @@ checkpointer = MongoDBSaver(
     collection_name="fida_agent_collection",
     serde=JsonPlusSerializer(pickle_fallback=True),  # ← 關鍵這一行
 )
-subagents = [
-    painter_subagent,
-    research_subagent,
-    user_profile_subagent
-]
 
 
-def build_main_agent(use_report):
+class CanvasMiddleware:
+    def before_agent(self, state, agent_input, **kwargs):
+        canvas = state.get("canvas", {})
+
+        info = f"""
+        当前画布状态：
+        - image_path: {canvas.get("image_path")}
+        - 是否已有图片: {bool(canvas.get("image_path"))}
+        """
+
+        agent_input["messages"].append({
+            "role": "system",
+            "content": info
+        })
+
+        return state, agent_input
+
+
+def build_main_agent(use_report, workspace_dir):
+    research_subagent = build_researcher_subagent(workspace_dir)
+    painter_subagent = build_painter_subagent(workspace_dir)
+    subagents = [
+        painter_subagent,
+        research_subagent,
+        user_profile_subagent
+    ]
     main_agent = create_deep_agent(
         model=main_llm,
         system_prompt=build_system_prompt(use_report=use_report),
         subagents=subagents,
         checkpointer=checkpointer,
         backend=FilesystemBackend(
-            root_dir=str(PROJECT_ROOT / "agent_workspace"),
-            virtual_mode=False,  # 重要：關掉虛擬模式 → 真的寫硬碟
+            root_dir=workspace_dir,
+            virtual_mode=True,  # 重要：關掉虛擬模式 → 真的寫硬碟
         ),
         middleware=[
             SummarizationMiddleware(

src/server/deep_agent/agents/painter.py

@@ -2,7 +2,7 @@ from langchain.agents.middleware import wrap_tool_call
 
 from src.server.deep_agent.init_llm import llm
 from src.server.deep_agent.init_prompt import build_painter_prompt
-from src.server.deep_agent.tools.generate_furniture_sketch import generate_furniture
+from src.server.deep_agent.tools.generate_furniture_sketch import create_generate_furniture_tool, create_edit_furniture_tool
 
 
 @wrap_tool_call
@@ -12,11 +12,16 @@ async def log_tool_calls(request, handler):
     return handler(request)
 
 
-painter_subagent = {
-    "name": "painter_subagent",
-    "description": "理解用户意图，使用prompt,调用generate_furniture工具生成家具sketch草图.",
-    "system_prompt": build_painter_prompt(),
-    "tools": [generate_furniture],
-    "model": llm,
-    # "middleware": [log_tool_calls],
-}
+def build_painter_subagent(workspace_dir):
+    generate_furniture = create_generate_furniture_tool(workspace_dir)
+    edit_furniture = create_edit_furniture_tool(workspace_dir)
+
+    painter_subagent = {
+        "name": "painter_subagent",
+        "description": "理解用户意图,利用prompt编辑或生成家具sketch图像",
+        "system_prompt": build_painter_prompt(),
+        "tools": [generate_furniture, edit_furniture],
+        "model": llm,
+        # "middleware": [log_tool_calls],
+    }
+    return painter_subagent

src/server/deep_agent/agents/researcher.py

@@ -1,21 +1,27 @@
 from src.server.deep_agent.init_llm import llm
 from src.server.deep_agent.init_prompt import build_researcher_prompt
-from src.server.deep_agent.tools.crawl_tool import crawl4ai_batch
-from src.server.deep_agent.tools.report_generator_tool import report_generator
+from src.server.deep_agent.tools.crawl_tool import create_crawl4ai_batch_tool
+from src.server.deep_agent.tools.report_generator_tool import create_report_generator_tool
 from src.server.deep_agent.tools.research_tool import topic_research
-from src.server.deep_agent.tools.structured_retrieval_tool import structured_retrieval
+from src.server.deep_agent.tools.structured_retrieval_tool import create_structured_retrieval_tool
 from src.server.deep_agent.tools.user_persona_tool import query_report_profile
 
-research_subagent = {
-    "name": "research_subagent",
-    "description": "通过网络搜索对家具设计开展深度研究并整合结论",
-    "system_prompt": build_researcher_prompt(),
-    "tools": [
-        query_report_profile,
-        topic_research,
-        crawl4ai_batch,
-        structured_retrieval,
-        report_generator
-    ],
-    "model": llm
-}
+
+def build_researcher_subagent(workspace_dir):
+    crawl4ai_batch = create_crawl4ai_batch_tool(workspace_dir)
+    structured_retrieval = create_structured_retrieval_tool(workspace_dir)
+    report_generator = create_report_generator_tool(workspace_dir)
+    research_subagent = {
+        "name": "research_subagent",
+        "description": "通过网络搜索对家具设计开展深度研究并整合结论",
+        "system_prompt": build_researcher_prompt(),
+        "tools": [
+            query_report_profile,
+            topic_research,
+            crawl4ai_batch,
+            structured_retrieval,
+            report_generator
+        ],
+        "model": llm
+    }
+    return research_subagent

src/server/deep_agent/agents/vision_subagent.py

@@ -0,0 +1,77 @@
+import json
+
+from src.server.deep_agent.init_llm import vision_llm
+from src.server.deep_agent.tools.vision_analyze_tool import vision_analyze_tool
+
+vision_subagent = {
+    "name": "vision_subagent",
+    "description": "分析用户上传的图片，提取家具、风格、颜色、材质等信息",
+    "system_prompt": """
+        你是一个专业的视觉分析助手（家具设计方向）。
+        
+        你的任务：
+        1. 理解用户提供的图片（路径或URL）
+        2. 分析家具内容
+        3. 输出结构化JSON（不要解释）
+        
+        格式：
+        {
+          "objects": [],
+          "style": "",
+          "color": [],
+          "material": [],
+          "room_type": "",
+          "description": ""
+        }
+        """,
+    "tools": [],  # ❗这里不用tool，直接用多模态模型
+    "model": vision_llm,
+}
+
+
+def vision_execute(state):
+    image = state.get("image")
+
+    if image is None:
+        return {
+            "error": "NO_IMAGE"
+        }
+
+    prompt = """
+你是一个家具视觉分析模型。
+
+任务：分析图片并输出JSON：
+
+{
+  "objects": [],
+  "style": "",
+  "color": [],
+  "material": [],
+  "room_type": "",
+  "description": ""
+}
+
+规则：
+- 只基于图像内容
+- 不允许编造
+- objects 最多5个
+- color 最多3个
+- 只输出JSON
+"""
+
+    result = vision_llm.generate(
+        image=image,  # ⭐ 关键：真正喂图
+        prompt=prompt
+    )
+
+    return safe_parse_json(result)
+
+
+def safe_parse_json(text):
+    try:
+        return json.loads(text)
+    except:
+        return {
+            "error": "INVALID_JSON",
+            "raw": text
+        }

src/server/deep_agent/init_llm.py

@@ -49,3 +49,12 @@ repoer_llm = ChatQwen(
     timeout=None,
     max_retries=2,
     api_key=settings.QWEN_API_KEY)
+
+vision_llm = ChatQwen(
+    enable_thinking=False,
+    model="qwen3-vl-plus",
+    temperature=0.2,
+    max_tokens=3_000,
+    timeout=None,
+    max_retries=2,
+    api_key=settings.QWEN_API_KEY)

src/server/deep_agent/init_prompt.py

@@ -15,7 +15,10 @@ def build_system_prompt(use_report):
        负责生成完整报告、调研、总结、分析。
     
     3. painter_subagent
-        负责根据用户描述,构造适用于生成家具sketch的prompt,使用prompt用工具生成图片.
+        负责根据用户描述,构造适用于 生成家具sketch的prompt或编辑家具sketch的prompt
+        1.利用prompt用工具生成图片.
+        2.利用prompt和图片路径用工具编辑图片.
+
         
     ========================
     执行规则
@@ -51,31 +54,26 @@ def build_system_prompt(use_report):
     - research-subagent 只负责 **报告生成**
     不要混用职责。
     ========================
-    严格输出规则
-    ========================
-    - 当生成图片时，绝对不要输出图片路径、file:// 地址、URL、本地链接
-    - 只输出文字描述，不输出任何图片链接或路径
     """
     return system_prompt
 
 
 def build_painter_prompt():
     prompt = """
-    你是一名专业的prompt优化专家，专注于家具设计草图生成。你的任务是：
-        1. 分析用户查询，理解核心意图，包括家具类型、风格、尺寸、颜色、材料等关键元素
-        2. 基于意图，优化并生成一个详细、精确的prompt，适合用于AI图片生成工具创建家具sketch草图（例如，线条简洁、手绘风格、焦点在设计细节上）
-        3. 使用优化的prompt调用图片生成工具，生成并返回草图图片
-        4. 如果需要，建议额外变体或改进
-
-        输出格式：
-        - 用户意图总结（1–2段）
-        - 优化后的prompt（完整文本）
-        - 生成的图片描述（如果工具返回）
-        - 建议改进（项目符号，可选）
-   【严格输出规则】
-    - 当生成图片时，**绝对不要输出图片路径、file:// 地址、URL、本地链接**。
-    - 只输出文字描述，不输出任何图片链接或路径。
+    你是 painter_subagent，专门生成或编辑 sketch 图。
+    1. 每次开始决策前，先调用工具 read_file("/current_sketch_path.txt") 获取当前路径。
+       - 如果文件不存在或返回空 → 当前没有历史图，使用 generate_sketch。
+       - 如果有路径 → 检查用户意图是否为「修改/编辑/改成/调整/优化/把...变成」，如果是则必须使用 edit_sketch，并传入 image_path = 读取到的路径。
+    2. 生成或编辑完成后，**必须立即**调用 write_file("/current_sketch_path.txt", content=本次生成的图片完整路径) 来更新状态。
+    3. 【对用户隐藏路径】：
+       - 永远不要在最终回复给用户的任何消息中出现路径、/tmp/、/current_sketch_path.txt 等字符串！
+       - 回复格式只能是：
+         "图片已成功生成！"
+         或
+         "已按你的要求把狗改成猫，图片更新完成！"
+       - 如果前端支持图片展示，你可以直接返回图片（但不要带路径文字）。
 
+    现在开始严格遵守以上规则。
     """
     return prompt
 

src/server/deep_agent/tools/crawl_tool.py

@@ -1,3 +1,4 @@
+import os
 import time
 import asyncio
 from typing import List, Dict, Any
@@ -8,19 +9,6 @@ import uuid
 from crawl4ai import AsyncWebCrawler, BrowserConfig, CrawlerRunConfig, CacheMode
 from langchain_core.tools import tool
 
-# ─────────────────────────────────────
-# 路径配置
-# ─────────────────────────────────────
-
-TOOL_DIR = Path(__file__).resolve().parent
-PROJECT_ROOT = TOOL_DIR.parent
-
-# DeepAgents 推荐目录
-SAVE_DIR = PROJECT_ROOT / "agent_workspace" / "raw_data"
-SAVE_DIR.mkdir(parents=True, exist_ok=True)
-
-print(f"tool save : {str(PROJECT_ROOT / "agent_workspace")}")
-
 # ─────────────────────────────────────
 # Browser 配置
 # ─────────────────────────────────────
@@ -65,7 +53,7 @@ def build_filename(url: str) -> str:
 # 单个 URL 抓取
 # ─────────────────────────────────────
 
-async def crawl_one(crawler, url: str, sem: asyncio.Semaphore) -> Dict[str, Any]:
+async def crawl_one(crawler, url: str, sem: asyncio.Semaphore, save_dir: str) -> Dict[str, Any]:
     async with sem:
         try:
             result = await crawler.arun(url=url, config=run_config)
@@ -87,7 +75,7 @@ async def crawl_one(crawler, url: str, sem: asyncio.Semaphore) -> Dict[str, Any]
                 }
 
             filename = build_filename(url)
-            filepath = SAVE_DIR / filename
+            filepath = os.path.join(save_dir, filename)
 
             header = (
                 f"<!-- Source: {url} -->\n"
@@ -115,7 +103,7 @@ async def crawl_one(crawler, url: str, sem: asyncio.Semaphore) -> Dict[str, Any]
 # Async 主逻辑
 # ─────────────────────────────────────
 
-async def _crawl4ai_batch(urls: List[str]) -> Dict[str, Any]:
+async def _crawl4ai_batch(urls: List[str], save_dir: str) -> Dict[str, Any]:
     urls = list(set(urls))  # 去重
 
     if not urls:
@@ -126,7 +114,7 @@ async def _crawl4ai_batch(urls: List[str]) -> Dict[str, Any]:
     async with AsyncWebCrawler(config=browser_config) as crawler:
 
         tasks = [
-            crawl_one(crawler, url, sem)
+            crawl_one(crawler, url, sem, save_dir)
             for url in urls
         ]
 
@@ -150,42 +138,46 @@ async def _crawl4ai_batch(urls: List[str]) -> Dict[str, Any]:
     }
 
 
-# ─────────────────────────────────────
-# Tool（同步）
-# ─────────────────────────────────────
-@tool
-def crawl4ai_batch(urls: List[str]) -> str:
-    """
-    Batch crawl webpages and save their content as markdown files.
+def create_crawl4ai_batch_tool(workspace_dir):
+    @tool
+    def crawl4ai_batch(urls: List[str]) -> str:
+        """
+        Batch crawl webpages and save their content as markdown files.
 
-    Args:
-        urls: List of webpage URLs to crawl.
+        Args:
+            urls: List of webpage URLs to crawl.
 
-    Returns:
-        A summary of crawling results and saved file paths.
-    """
+        Returns:
+            A summary of crawling results and saved file paths.
+        """
 
-    try:
-        result = asyncio.run(_crawl4ai_batch(urls))
+        try:
+            save_dir = os.path.join(workspace_dir, "raw_data")
+            if not os.path.exists(save_dir):
+                os.makedirs(save_dir, exist_ok=True)
 
-        if "error" in result:
-            return f"❌ Error: {result['error']}"
+            result = asyncio.run(_crawl4ai_batch(urls, save_dir))
 
-        output = [
-            "### 批量抓取完成 ###",
-            f"成功保存文件: {result['count']}",
-            f"保存目录: {SAVE_DIR}",
-            "",
-            "抓取详情:"
-        ]
+            if "error" in result:
+                return f"❌ Error: {result['error']}"
 
-        output.extend(result["summary"])
+            output = [
+                "### 批量抓取完成 ###",
+                f"成功保存文件: {result['count']}",
+                f"保存目录: {workspace_dir}",
+                "",
+                "抓取详情:"
+            ]
 
-        if result["saved_files"]:
-            output.append("\n可读取文件:")
-            output.extend(result["saved_files"])
+            output.extend(result["summary"])
 
-        return "\n".join(output)
+            if result["saved_files"]:
+                output.append("\n可读取文件:")
+                output.extend(result["saved_files"])
 
-    except Exception as e:
-        return f"🚨 爬虫系统异常: {str(e)}"
+            return "\n".join(output)
+
+        except Exception as e:
+            return f"🚨 爬虫系统异常: {str(e)}"
+
+    return crawl4ai_batch

src/server/deep_agent/tools/generate_furniture_sketch.py

@@ -1,15 +1,21 @@
 import json
 import logging
+import os
 import uuid
+from pathlib import Path
+from typing import Annotated
+
+import httpx
 from google.oauth2 import service_account
 from langchain_core.tools import tool
 from google import genai
 from google.genai.types import GenerateContentConfig, Modality
+from langgraph.prebuilt import ToolRuntime
 
 from minio import Minio
 
 from src.core.config import settings
-from src.server.utils.new_oss_client import oss_upload_image
+from src.server.utils.new_oss_client import oss_upload_image, oss_get_image, is_minio_file_exist, oss_upload_image_file
 
 logger = logging.getLogger(__name__)
 # 初始化全局凭证和客户端
@@ -27,47 +33,187 @@ client = genai.Client(
 )
 
 
-@tool
-async def generate_furniture(prompt: str) -> str:
-    """
-    使用 Gemini 图像生成模型根据详细的英文提示词生成家具设计草图。
-    """
-    print(f"\n[系统日志] 正在调用 Nano Banana (Gemini Image Gen) ...")
-
+def is_image_path_exist(image_path):
     try:
-        response = client.models.generate_content(
-            model="gemini-2.5-flash-image",
-            contents=(f"Generate a professional furniture design sketch: {prompt}"),
-            config=GenerateContentConfig(
-                response_modalities=[Modality.TEXT, Modality.IMAGE],
-            ),
-        )
+        return Path(image_path).exists()
+    except:
+        return False
 
-        image_bytes = None
-        for part in response.candidates[0].content.parts:
-            if part.inline_data:
-                image_bytes = part.inline_data.data
-                break
 
-        if not image_bytes:
-            return "未能生成图像数据。"
-        object_name = f"furniture/sketches/{uuid.uuid4()}.png"
-        bucket = "fida-test"  # 替换为你的 bucket 名称
-        # 3. 调用你的上传函数
-        upload_res = oss_upload_image(
-            oss_client=minio_client,
-            bucket=bucket,
-            object_name=object_name,
-            image_bytes=image_bytes
-        )
+def create_generate_furniture_tool(workspace_dir, width: int = 1024, height: int = 1024):
+    @tool
+    async def generate_furniture(prompt: str, runtime: ToolRuntime) -> str:
+        """
+        使用 Gemini 图像生成模型根据详细的英文提示词生成家具设计草图。
+        """
+        logger.info(f"\n[系统日志] 正在调用 generate_furniture ...")
+        try:
+            # 1. 生成图像 - local flux2-klein
+            object_name = f"furniture/sketches/{uuid.uuid4()}.png"
+            bucket_name = "fida-test"  # 替换为你的 bucket 名称
+            request_data = {
+                "prompt": prompt,
+                "bucket_name": bucket_name,
+                "object_name": object_name,
+                "width": width,
+                "height": height
+            }
+            async with httpx.AsyncClient(timeout=120) as client:
+                resp = await client.post(
+                    f"http://{settings.FLUX2_GEN_IMG_MODEL_URL}/predict",
+                    json=request_data,
+                )
+            result = resp.json()
+            image_url = result.get("output_path", None)
 
-        if upload_res:
-            # 4. 构造访问链接 (如果是私有 bucket，需使用 presigned_get_object)
-            # 这里简单示例为直接访问地址
-            image_url = f"{bucket}/{object_name}"
-            return image_url
-        else:
-            return "图片生成成功，但上传至存储服务器失败。"
-    except Exception as e:
-        logger.warning(e)
-        return "绘图流程异常"
+            if image_url:
+                filename = os.path.join(workspace_dir, image_url)
+                # 2. 创建本地目录（确保目录存在）
+                local_dir = os.path.dirname(filename)
+                if not os.path.exists(local_dir):
+                    os.makedirs(local_dir, exist_ok=True)
+
+                img = oss_get_image(oss_client=minio_client, bucket=image_url.split('/')[0], object_name=image_url[image_url.find('/') + 1:])
+                img.save(filename)
+
+                return image_url
+            else:
+                return f"Image generation failed."
+
+        except Exception as e:
+            logger.warning(f"绘图流程异常：{e}")
+            return "绘图流程异常"
+
+    return generate_furniture
+
+
+def create_edit_furniture_tool(workspace_dir, width: int = 1024, height: int = 1024):
+    @tool
+    async def edit_furniture(prompt: str, input_image_path) -> str:
+        """
+        使用图像生成模型根据详细的英文提示词编辑家具设计草图。
+        """
+        logger.info(f"\n[系统日志] 正在调用 edit_furniture ...")
+
+        try:
+            # 0. 编辑前先检查工作环境和minio上是否存在该图像
+            input_image_path = input_image_path.lstrip('/')
+            filename = os.path.join(workspace_dir, input_image_path)
+            local_exist = is_image_path_exist(filename)
+            minio_exist = is_minio_file_exist(minio_client=minio_client, bucket_name=input_image_path.split('/')[0], object_name=input_image_path.split('/')[0])
+
+            if not local_exist and not minio_exist:
+                # 两个地方都不存在 直接报错
+                return f"Image generation failed."
+            elif local_exist and not minio_exist:
+                # 把本地的上传到minio
+                oss_upload_image_file(oss_client=minio_client, bucket=input_image_path.split('/')[0], object_name=input_image_path.split('/')[0], file_path=filename)
+            elif not local_exist and minio_exist:
+                # minio的下载到本地
+                img = oss_get_image(oss_client=minio_client, bucket=input_image_path.split('/')[0], object_name=input_image_path.split('/')[0], )
+                img.save(filename)
+            elif minio_exist and local_exist:
+                # 两个地方都存在 直接跳过
+                pass
+
+            # 1. 生成图像 - local flux2-klein
+            object_name = f"furniture/sketches/{uuid.uuid4()}.png"
+            bucket_name = "fida-test"  # 替换为你的 bucket 名称
+            request_data = {
+                "input_image_paths": [input_image_path],
+                "prompt": prompt,
+                "bucket_name": bucket_name,
+                "object_name": object_name,
+                "width": width,
+                "height": height
+            }
+            async with httpx.AsyncClient(timeout=120) as client:
+                resp = await client.post(
+                    f"http://{settings.FLUX2_GEN_IMG_MODEL_URL}/predict",
+                    json=request_data,
+                )
+            result = resp.json()
+            image_url = result.get("output_path", None)
+
+            if image_url:
+                filename = os.path.join(workspace_dir, image_url)
+                # 2. 创建本地目录（确保目录存在）
+                local_dir = os.path.dirname(filename)
+                if not os.path.exists(local_dir):
+                    os.makedirs(local_dir, exist_ok=True)
+
+                img = oss_get_image(oss_client=minio_client, bucket=image_url.split('/')[0], object_name=image_url[image_url.find('/') + 1:])
+                img.save(filename)
+                return image_url
+            else:
+                return f"Image generation failed."
+
+        except Exception as e:
+            logger.warning(f"edit_furniture error ：{e}")
+            return "edit_furniture error"
+
+    return edit_furniture
+
+# def create_generate_furniture_tool(workspace_dir):
+#     @tool
+#     async def generate_furniture(prompt: str) -> str:
+#         """
+#         使用 Gemini 图像生成模型根据详细的英文提示词生成家具设计草图。
+#         """
+#         print(f"\n[系统日志] 正在调用 Nano Banana (Gemini Image Gen) ...")
+#
+#         try:
+#             response = client.models.generate_content(
+#                 model="gemini-2.5-flash-image",
+#                 contents=(f"Generate a professional furniture design sketch: {prompt}"),
+#                 config=GenerateContentConfig(
+#                     response_modalities=[Modality.TEXT, Modality.IMAGE],
+#                 ),
+#             )
+#
+#             image_bytes = None
+#             for part in response.candidates[0].content.parts:
+#                 if part.inline_data:
+#                     image_bytes = part.inline_data.data
+#                     break
+#
+#             if not image_bytes:
+#                 return "未能生成图像数据。"
+#             # 1. 定义OSS存储路径和本地保存路径
+#             object_name = f"furniture/sketches/{uuid.uuid4()}.png"
+#             bucket = "fida-test"  # 替换为你的 bucket 名称
+#             filename = os.path.join(workspace_dir, f"{bucket}/{object_name}")
+#
+#             # 2. 创建本地目录（确保目录存在）
+#             local_dir = os.path.dirname(filename)
+#             if not os.path.exists(local_dir):
+#                 os.makedirs(local_dir, exist_ok=True)
+#
+#             # 3. 保存图片到本地文件（新增核心逻辑）
+#             try:
+#                 with open(filename, "wb") as f:
+#                     f.write(image_bytes)
+#                 print(f"[系统日志] 图片已保存到本地：{filename}")
+#             except Exception as save_e:
+#                 logger.warning(f"保存图片到本地失败：{save_e}")
+#                 # 本地保存失败不中断上传流程，仅记录日志
+#
+#             # 4. 上传图片到OSS（原有逻辑）
+#             upload_res = oss_upload_image(
+#                 oss_client=minio_client,
+#                 bucket=bucket,
+#                 object_name=object_name,
+#                 image_bytes=image_bytes
+#             )
+#
+#             if upload_res:
+#                 image_url = f"{bucket}/{object_name}"
+#                 return image_url
+#             else:
+#                 return f"图片生成成功（本地路径：{filename}），但上传至存储服务器失败。"
+#
+#         except Exception as e:
+#             logger.warning(f"绘图流程异常：{e}")
+#             return "绘图流程异常"
+#
+#     return generate_furniture

src/server/deep_agent/tools/report_generator_tool.py

@@ -32,104 +32,108 @@ class ReportInput(BaseModel):
 # LangGraph Tool
 # =========================
 
-@tool("report_generator", args_schema=ReportInput)
-async def report_generator(
-        report_topic: str,
-        structured_data: List[Dict],
-        language: str = "English"
-) -> dict:
-    """
-    Generate a professional design/market report
-    directly from structured retrieval results.
-    """
+def create_report_generator_tool(workspace_dir):
+    @tool("report_generator", args_schema=ReportInput)
+    async def report_generator(
+            report_topic: str,
+            structured_data: List[Dict],
+            language: str = "English"
+    ) -> dict:
+        """
+        Generate a professional design/market report
+        directly from structured retrieval results.
+        """
 
-    writer = get_stream_writer()
-    if not structured_data:
-        error_msg = "Error: No structured data provided."
-        writer({"type": "report_error", "message": error_msg})
-        return error_msg
+        writer = get_stream_writer()
+        if not structured_data:
+            error_msg = "Error: No structured data provided."
+            writer({"type": "report_error", "message": error_msg})
+            return error_msg
 
-    collected_data_str = json.dumps(
-        structured_data,
-        ensure_ascii=False,
-        indent=2
-    )
-
-    # =========================
-    # Prompt
-    # =========================
-
-    system_prompt = f"""
-    You are a professional design trend analyst.
-    
-    Generate a long, structured Markdown report.
-    
-    REQUIREMENTS:
-    
-    1. Follow MECE principle.
-    2. Embed images ONLY if they start with https://
-       using: ![alt](url)
-    3. Insert images inline.
-    4. Every key insight must cite source:
-       [Website Name](url)
-    5. Use Markdown headings.
-    6. Start directly with title.
-    7. Be detailed and analytical.
-    
-    Output Language: {language}
-    """
-
-    user_prompt = f"""
-    Topic: {report_topic}
-    
-    Input Data:
-    {collected_data_str}
-    """
-
-    # =========================
-    # 调用 LLM
-    # =========================
-    writer({"type": "report_start", "topic": report_topic, "language": language})
-
-    full_report = ""
-    try:
-        report_llm = repoer_llm.with_config(
-            callbacks=[]
+        collected_data_str = json.dumps(
+            structured_data,
+            ensure_ascii=False,
+            indent=2
         )
-        async for chunk in report_llm.astream(
-                [
-                    SystemMessage(content=system_prompt),
-                    HumanMessage(content=user_prompt)
-                ]
-        ):
-            if chunk.content:  # Gemini 返回的 chunk.content
-                delta = chunk.content
-                full_report += delta
-                # return {"type": "report_delta", "delta": delta}
-                writer({"type": "report_delta", "delta": delta})  # ← 实时推送给前端
-        writer({"type": "report_stop", "topic": report_topic, "language": language})
-    except Exception as e:
-        error_msg = f"LLM generation failed: {str(e)}"
-        writer({"type": "report_error", "message": error_msg})
-        return error_msg
 
-    report_content = full_report.strip()
+        # =========================
+        # Prompt
+        # =========================
 
-    # =========================
-    # 保存报告
-    # =========================
-    output_dir = "workspace/reports"
-    os.makedirs(output_dir, exist_ok=True)
+        system_prompt = f"""
+        You are a professional design trend analyst.
 
-    safe_topic = re.sub(r'[\\/*?:"<>|]', "", report_topic.replace(" ", "_"))
-    filename = f"{output_dir}/{safe_topic}.md"
+        Generate a long, structured Markdown report.
 
-    try:
-        with open(filename, "w", encoding="utf-8") as f:
-            f.write(report_content)
-        writer({"type": "report_complete", "file_path": filename})
-    except Exception as e:
-        writer({"type": "report_save_warning", "message": str(e)})
+        REQUIREMENTS:
 
-    # 返回完整内容（作为 tool result），同时正文已通过 delta 流式输出
-    return report_content + f"\n\n✅ Report saved to: {filename}"
+        1. Follow MECE principle.
+        2. Embed images ONLY if they start with https://
+           using: ![alt](url)
+        3. Insert images inline.
+        4. Every key insight must cite source:
+           [Website Name](url)
+        5. Use Markdown headings.
+        6. Start directly with title.
+        7. Be detailed and analytical.
+
+        Output Language: {language}
+        """
+
+        user_prompt = f"""
+        Topic: {report_topic}
+
+        Input Data:
+        {collected_data_str}
+        """
+
+        # =========================
+        # 调用 LLM
+        # =========================
+        writer({"type": "report_start", "topic": report_topic, "language": language})
+
+        full_report = ""
+        try:
+            report_llm = repoer_llm.with_config(
+                callbacks=[]
+            )
+            async for chunk in report_llm.astream(
+                    [
+                        SystemMessage(content=system_prompt),
+                        HumanMessage(content=user_prompt)
+                    ]
+            ):
+                if chunk.content:  # Gemini 返回的 chunk.content
+                    delta = chunk.content
+                    full_report += delta
+                    # return {"type": "report_delta", "delta": delta}
+                    writer({"type": "report_delta", "delta": delta})  # ← 实时推送给前端
+            writer({"type": "report_stop", "topic": report_topic, "language": language})
+        except Exception as e:
+            error_msg = f"LLM generation failed: {str(e)}"
+            writer({"type": "report_error", "message": error_msg})
+            return error_msg
+
+        report_content = full_report.strip()
+
+        # =========================
+        # 保存报告
+        # =========================
+        output_dir = os.path.join(workspace_dir, "reports")
+        if not os.path.exists(output_dir):
+            os.makedirs(output_dir, exist_ok=True)
+
+        safe_topic = re.sub(r'[\\/*?:"<>|]', "", report_topic.replace(" ", "_"))
+        filename = f"{output_dir}/{safe_topic}.md"
+
+        try:
+            with open(filename, "w", encoding="utf-8") as f:
+                f.write(report_content)
+            writer({"type": "report_complete", "file_path": filename})
+        except Exception as e:
+            writer({"type": "report_save_warning", "message": str(e)})
+
+        # 返回完整内容（作为 tool result），同时正文已通过 delta 流式输出
+        return report_content + f"\n\n✅ Report saved to: {filename}"
+
+    return report_generator

src/server/deep_agent/tools/structured_retrieval_tool.py

@@ -32,121 +32,6 @@ class StructuredRetrievalInput(BaseModel):
     source_url: Optional[str] = Field(None, description="Optional global source URL")
 
 
-@tool("structured_retrieval", args_schema=StructuredRetrievalInput)
-def structured_retrieval(
-        file_paths: List[str],
-        query: str,
-        source_url: Optional[str] = None
-) -> Dict:
-    """
-    Batch structured extraction from markdown files.
-    - Performs vector search + re-ranking
-    - Saves extracted structured data as JSON file to disk
-    - Returns ONLY summary (status, count, file path)
-    """
-
-    # ── 1. 收集所有文件內容 ──────────────────────────────────────
-    all_docs_pool: List[Document] = []
-
-    for path in file_paths:
-        if not os.path.exists(path) or not path.endswith((".md", ".markdown")):
-            continue
-
-        file_name = os.path.basename(path)
-
-        with open(path, "r", encoding="utf-8") as f:
-            content = f.read()
-
-        current_source = source_url or _extract_source_from_md(content) or "unknown"
-
-        sections = _split_markdown_by_headers(content)
-
-        for sec in sections:
-            all_docs_pool.append(
-                Document(
-                    page_content=sec,
-                    metadata={"source_url": current_source, "file_name": file_name}
-                )
-            )
-
-    if not all_docs_pool:
-        return {"status": "no_documents_found", "items_count": 0, "json_path": None}
-
-    # ── 2. Vector search ────────────────────────────────────────────
-    vector_store = FAISS.from_documents(all_docs_pool, _EMBEDDING_MODEL)
-    retrieved = vector_store.similarity_search(query, k=200)
-
-    # ── 3. 提取結構化片段 ──────────────────────────────────────────
-    structured_items = []
-
-    for doc in retrieved:
-        text = doc.page_content.strip()
-        if len(text) < 30:
-            continue
-
-        images = list(set(re.findall(r"!\[.*?\]\((.*?)\)", text)))
-
-        structured_items.append(
-            {
-                "text": text,
-                "images": images,
-                "source_url": doc.metadata.get("source_url"),
-                "file_name": doc.metadata.get("file_name")
-            }
-        )
-
-    # ── 4. Re-rank ──────────────────────────────────────────────────
-    if structured_items:
-        unique_items = {item["text"]: item for item in structured_items}.values()
-        pairs = [[query, item["text"]] for item in unique_items]
-        scores = _RERANK_MODEL.predict(pairs)
-
-        sorted_items = sorted(
-            zip(scores, unique_items),
-            key=lambda x: x[0],
-            reverse=True
-        )
-        top_items = [item for _, item in sorted_items[:50]]
-    else:
-        top_items = []
-
-    # ── 5. 寫入 JSON 文件 ──────────────────────────────────────────
-    if not top_items:
-        return {"status": "no_relevant_content", "items_count": 0, "json_path": None}
-
-    # 產生有意義的檔名
-    safe_query = re.sub(r'[^a-zA-Z0-9\u4e00-\u9fa5]', '_', query)[:40]
-    timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
-    json_filename = f"extracted_{safe_query}_{timestamp}.json"
-
-    # 建議的儲存目錄（與 crawl4ai_batch 對齊）
-    output_dir = os.path.join(os.path.dirname(file_paths[0]), "..", "extracted")
-    os.makedirs(output_dir, exist_ok=True)
-
-    json_path = os.path.join(output_dir, json_filename)
-
-    with open(json_path, "w", encoding="utf-8") as f:
-        json.dump(
-            {
-                "query": query,
-                "extracted_at": timestamp,
-                "item_count": len(top_items),
-                "items": top_items
-            },
-            f,
-            ensure_ascii=False,
-            indent=2
-        )
-
-    # ── 6. 只回傳摘要 ──────────────────────────────────────────────
-    return {
-        "status": "success",
-        "items_count": len(top_items),
-        "json_path": json_path,
-        "summary": f"已提取 {len(top_items)} 個高相關片段，儲存於 {json_path}"
-    }
-
-
 def _extract_source_from_md(content: str) -> Optional[str]:
     match = re.search(r"<!--\s*Source:\s*(.*?)\s*-->", content)
     return match.group(1).strip() if match else None
@@ -223,3 +108,126 @@ def _chunk_text(
         start = max(0, end - overlap)
 
     return chunks
+
+
+def create_structured_retrieval_tool(workspace_dir):
+    @tool("structured_retrieval", args_schema=StructuredRetrievalInput)
+    def structured_retrieval(
+            file_paths: List[str],
+            query: str,
+            source_url: Optional[str] = None
+    ) -> Dict:
+        """
+        Batch structured extraction from markdown files.
+        - Performs vector search + re-ranking
+        - Saves extracted structured data as JSON file to disk
+        - Returns ONLY summary (status, count, file path)
+        """
+
+        # ── 1. 收集所有文件內容 ──────────────────────────────────────
+        all_docs_pool: List[Document] = []
+
+        for path in file_paths:
+            if not os.path.exists(path) or not path.endswith((".md", ".markdown")):
+                continue
+
+            file_name = os.path.basename(path)
+
+            with open(path, "r", encoding="utf-8") as f:
+                content = f.read()
+
+            current_source = source_url or _extract_source_from_md(content) or "unknown"
+
+            sections = _split_markdown_by_headers(content)
+
+            for sec in sections:
+                all_docs_pool.append(
+                    Document(
+                        page_content=sec,
+                        metadata={"source_url": current_source, "file_name": file_name}
+                    )
+                )
+
+        if not all_docs_pool:
+            return {"status": "no_documents_found", "items_count": 0, "json_path": None}
+
+        # ── 2. Vector search ────────────────────────────────────────────
+        vector_store = FAISS.from_documents(all_docs_pool, _EMBEDDING_MODEL)
+        retrieved = vector_store.similarity_search(query, k=200)
+
+        # ── 3. 提取結構化片段 ──────────────────────────────────────────
+        structured_items = []
+
+        for doc in retrieved:
+            text = doc.page_content.strip()
+            if len(text) < 30:
+                continue
+
+            images = list(set(re.findall(r"!\[.*?\]\((.*?)\)", text)))
+
+            structured_items.append(
+                {
+                    "text": text,
+                    "images": images,
+                    "source_url": doc.metadata.get("source_url"),
+                    "file_name": doc.metadata.get("file_name")
+                }
+            )
+
+        # ── 4. Re-rank ──────────────────────────────────────────────────
+        if structured_items:
+            unique_items = {item["text"]: item for item in structured_items}.values()
+            pairs = [[query, item["text"]] for item in unique_items]
+            scores = _RERANK_MODEL.predict(pairs)
+
+            sorted_items = sorted(
+                zip(scores, unique_items),
+                key=lambda x: x[0],
+                reverse=True
+            )
+            top_items = [item for _, item in sorted_items[:50]]
+        else:
+            top_items = []
+
+        # ── 5. 寫入 JSON 文件 ──────────────────────────────────────────
+        if not top_items:
+            return {"status": "no_relevant_content", "items_count": 0, "json_path": None}
+
+        # 產生有意義的檔名
+        safe_query = re.sub(r'[^a-zA-Z0-9\u4e00-\u9fa5]', '_', query)[:40]
+        timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
+        json_filename = f"extracted_{safe_query}_{timestamp}.json"
+
+        # 建議的儲存目錄（與 crawl4ai_batch 對齊）
+        output_dir = os.path.join(workspace_dir, "extracted")
+        if not os.path.exists(output_dir):
+            os.makedirs(output_dir, exist_ok=True)
+
+        if not os.path.exists(output_dir):
+            # 2. 不存在则创建（makedirs 支持创建多级目录，mkdir 只能创建单级）
+            os.makedirs(output_dir, exist_ok=True)
+
+        json_path = os.path.join(output_dir, json_filename)
+
+        with open(json_path, "w", encoding="utf-8") as f:
+            json.dump(
+                {
+                    "query": query,
+                    "extracted_at": timestamp,
+                    "item_count": len(top_items),
+                    "items": top_items
+                },
+                f,
+                ensure_ascii=False,
+                indent=2
+            )
+
+        # ── 6. 只回傳摘要 ──────────────────────────────────────────────
+        return {
+            "status": "success",
+            "items_count": len(top_items),
+            "json_path": json_path,
+            "summary": f"已提取 {len(top_items)} 個高相關片段，儲存於 {json_path}"
+        }
+
+    return structured_retrieval

src/server/deep_agent/tools/vision_analyze_tool.py

@@ -0,0 +1,21 @@
+from langchain.tools import tool
+from langchain_core.messages import HumanMessage
+from PIL import Image
+import requests
+from io import BytesIO
+
+from src.server.deep_agent.init_llm import vision_llm
+
+
+@tool
+def analyze_image(image_url: str) -> str:
+    """分析给定URL的图像。输入图像URL，输出图像描述和关键观察。"""
+    response = requests.get(image_url)
+    image = Image.open(BytesIO(response.content))
+    # 这里使用模型直接分析图像（简化示例）
+    msg = HumanMessage(content=[
+        {"type": "text", "text": "详细描述这张图像，包括物体、颜色、场景和任何文本。"},
+        {"type": "image_url", "image_url": {"url": image_url}}
+    ])
+    result = vision_llm.invoke([msg])
+    return result.content

src/server/utils/new_oss_client.py

@@ -36,7 +36,7 @@ http_client = urllib3.PoolManager(
 
 
 #  获取图片
-def oss_get_image(oss_client, bucket, object_name, data_type):
+def oss_get_image(oss_client, bucket, object_name):
     # cv2 默认全通道读取
     image_object = None
     try:
@@ -57,9 +57,44 @@ def oss_upload_image(oss_client, bucket, object_name, image_bytes):
     return req
 
 
+def oss_upload_image_file(oss_client, bucket, object_name, file_path):
+    req = None
+    try:
+        req = oss_client.fput_object(
+            bucket_name=bucket,
+            object_name=object_name,
+            file_path=file_path
+        )
+    except Exception as e:
+        logger.warning(f" | 上传图片出现异常 ######: {e}")
+    return req
+
+
+def get_presigned_url(oss_client, bucket, object_name):
+    try:
+        presigned_url = oss_client.presigned_get_object(
+            bucket_name=bucket,
+            object_name=object_name,
+            expires=3600
+        )
+        return presigned_url
+    except Exception as e:
+        print(f"get_presigned_url exception :{e}")
+        return "object not found"
+
+
+def is_minio_file_exist(minio_client: Minio, bucket_name: str, object_name: str) -> bool:
+    try:
+        # 核心判断：检查MinIO中指定bucket+object是否存在
+        minio_client.stat_object(bucket_name, object_name)
+        return True
+    except Exception as e:
+        return False
+
+
 if __name__ == '__main__':
-    url = "fida-test/furniture/sketches/4449a66d-6267-43f7-86a2-1e42bd19ec61.png"
+    url = 'fida-test/furniture/sketches/9356e3d8-d56e-4478-adde-61b29119979b.png'
     read_type = "2"
-    img = oss_get_image(oss_client=minio_client, bucket=url.split('/')[0], object_name=url[url.find('/') + 1:], data_type=read_type)
+    img = oss_get_image(oss_client=minio_client, bucket=url.split('/')[0], object_name=url[url.find('/') + 1:])
     img.show()
     img.save("result.png")