163 lines
6.6 KiB
Python
163 lines
6.6 KiB
Python
|
import socket
|
||
|
|
import json
|
||
|
|
import time
|
||
|
|
from openai import OpenAI
|
||
|
|
|
||
|
|
# 初始化 DeepSeek 客户端
|
||
|
|
client = OpenAI(
|
||
|
|
api_key="sk-420190f448fe41158c4e2ccff90e35ce",
|
||
|
|
base_url="https://api.deepseek.com"
|
||
|
|
)
|
||
|
|
MODEL_NAME = "deepseek-chat"
|
||
|
|
|
||
|
|
def fetch_and_translate_tools():
|
||
|
|
"""向 CAD 基座发送 tools/list，并转换为大模型认识的格式"""
|
||
|
|
try:
|
||
|
|
cad_socket = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
|
||
|
|
cad_socket.connect(("127.0.0.1", 8080))
|
||
|
|
|
||
|
|
# 1. 向 MCP Server 发送标准的 tools/list 请求
|
||
|
|
req = {
|
||
|
|
"jsonrpc": "2.0",
|
||
|
|
"id": 100,
|
||
|
|
"method": "tools/list"
|
||
|
|
}
|
||
|
|
cad_socket.sendall(json.dumps(req).encode('utf-8'))
|
||
|
|
|
||
|
|
response_data = cad_socket.recv(8192)
|
||
|
|
cad_socket.close()
|
||
|
|
|
||
|
|
mcp_response = json.loads(response_data.decode('utf-8'))
|
||
|
|
|
||
|
|
if "result" not in mcp_response or "tools" not in mcp_response["result"]:
|
||
|
|
print("[Error] 无法从 CAD 获取工具列表")
|
||
|
|
return []
|
||
|
|
|
||
|
|
llm_tools = []
|
||
|
|
# 2. 遍历 CAD 返回的工具，将其翻译为 DeepSeek/OpenAI 格式
|
||
|
|
for mcp_tool in mcp_response["result"]["tools"]:
|
||
|
|
llm_tools.append({
|
||
|
|
"type": "function",
|
||
|
|
"function": {
|
||
|
|
"name": mcp_tool["name"],
|
||
|
|
"description": mcp_tool["description"],
|
||
|
|
# 核心转换：MCP 的 inputSchema 等价于 LLM 的 parameters
|
||
|
|
"parameters": mcp_tool["inputSchema"]
|
||
|
|
}
|
||
|
|
})
|
||
|
|
|
||
|
|
print(f"[Init] 成功从 CAD 动态加载了 {len(llm_tools)} 个工具！")
|
||
|
|
return llm_tools
|
||
|
|
|
||
|
|
except Exception as e:
|
||
|
|
print(f"[Error] 获取工具列表失败: {e}")
|
||
|
|
return []
|
||
|
|
|
||
|
|
def call_cad_mcp_server(tool_name, arguments):
|
||
|
|
"""底层 TCP 通信保持不变"""
|
||
|
|
try:
|
||
|
|
cad_socket = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
|
||
|
|
cad_socket.connect(("127.0.0.1", 8080))
|
||
|
|
req = {
|
||
|
|
"jsonrpc": "2.0", "id": 1, "method": "tools/call",
|
||
|
|
"params": { "name": tool_name, "arguments": arguments }
|
||
|
|
}
|
||
|
|
cad_socket.sendall(json.dumps(req).encode('utf-8'))
|
||
|
|
response_data = b""
|
||
|
|
while True:
|
||
|
|
chunk = cad_socket.recv(8192)
|
||
|
|
response_data += chunk
|
||
|
|
if len(chunk) < 8192: break
|
||
|
|
cad_socket.close()
|
||
|
|
return json.loads(response_data.decode('utf-8'))
|
||
|
|
except Exception as e:
|
||
|
|
print(f"[Error] 连接 CAD 基座失败: {e}")
|
||
|
|
return None
|
||
|
|
|
||
|
|
def run_agent(user_prompt):
|
||
|
|
print("=== 🚀 CAD 具身智能 Agent 启动 ===")
|
||
|
|
|
||
|
|
# 动态向 CAD 请求可用工具，彻底解耦！
|
||
|
|
tools = fetch_and_translate_tools()
|
||
|
|
|
||
|
|
if not tools:
|
||
|
|
print("没有可用的工具，Agent 退出。")
|
||
|
|
return
|
||
|
|
|
||
|
|
# 核心心智设定：教它怎么形成视觉闭环
|
||
|
|
system_prompt = """你是一个高阶 AutoCAD 视觉检查专家。
|
||
|
|
请严格遵循以下工作流：
|
||
|
|
1. 第一步永远是调用 `get_viewport_screenshot` 观察当前画面。
|
||
|
|
2. 仔细评估图像。如果目标物体（如图元、文字）太小导致无法确信细节，请调用 `zoom_window_normalized` 放大该局部区域。
|
||
|
|
3. 【关键指令】：每次执行缩放操作（zoom）后，你必须在下一回合再次调用 `get_viewport_screenshot` 获取放大后的新截图！
|
||
|
|
4. 反复观察和放大，直到你 100% 看清细节，再用自然语言向用户输出最终结论。"""
|
||
|
|
|
||
|
|
messages = [
|
||
|
|
{"role": "system", "content": system_prompt},
|
||
|
|
{"role": "user", "content": user_prompt}
|
||
|
|
]
|
||
|
|
|
||
|
|
MAX_TURNS = 8 # 熔断机制：最多允许思考 8 个回合，防止无限套娃
|
||
|
|
|
||
|
|
for turn in range(MAX_TURNS):
|
||
|
|
print(f"\n--- [第 {turn + 1} 回合] 思考中 ---")
|
||
|
|
|
||
|
|
response = client.chat.completions.create(
|
||
|
|
model=MODEL_NAME,
|
||
|
|
messages=messages,
|
||
|
|
tools=tools
|
||
|
|
)
|
||
|
|
|
||
|
|
assistant_message = response.choices[0].message
|
||
|
|
messages.append(assistant_message) # 记录思维路径
|
||
|
|
|
||
|
|
# 1. 检查是否需要执行物理动作
|
||
|
|
if assistant_message.tool_calls:
|
||
|
|
for tool_call in assistant_message.tool_calls:
|
||
|
|
tool_name = tool_call.function.name
|
||
|
|
args = json.loads(tool_call.function.arguments)
|
||
|
|
|
||
|
|
print(f"[动作决定] ⚡ 调用工具: {tool_name}")
|
||
|
|
if args: print(f" 参数: {args}")
|
||
|
|
|
||
|
|
# 执行 CAD 通信
|
||
|
|
cad_res = call_cad_mcp_server(tool_name, args)
|
||
|
|
time.sleep(0.5) # 给予 CAD 渲染刷新窗口的时间缓冲
|
||
|
|
|
||
|
|
# 处理执行结果，并准备发回给 LLM 的观测报告
|
||
|
|
tool_result_content = []
|
||
|
|
|
||
|
|
if cad_res and "result" in cad_res:
|
||
|
|
for item in cad_res["result"]["content"]:
|
||
|
|
if item["type"] == "text":
|
||
|
|
tool_result_content.append({"type": "text", "text": item["text"]})
|
||
|
|
elif item["type"] == "image":
|
||
|
|
base64_img = item["data"]
|
||
|
|
mime = item.get("mimeType", "image/png")
|
||
|
|
tool_result_content.append({
|
||
|
|
"type": "image_url",
|
||
|
|
"image_url": {"url": f"data:{mime};base64,{base64_img}"}
|
||
|
|
})
|
||
|
|
print("[观测反馈] 📸 已将最新 CAD 屏幕画面传回视觉中枢。")
|
||
|
|
else:
|
||
|
|
tool_result_content.append({"type": "text", "text": "CAD 工具执行失败或无响应。"})
|
||
|
|
|
||
|
|
# 必须向大模型提交 Tool 返回结果
|
||
|
|
messages.append({
|
||
|
|
"role": "tool",
|
||
|
|
"tool_call_id": tool_call.id,
|
||
|
|
"content": tool_result_content
|
||
|
|
})
|
||
|
|
|
||
|
|
# 2. 如果不调工具，说明任务已完成，输出最终自然语言结论
|
||
|
|
else:
|
||
|
|
print("\n==================================")
|
||
|
|
print(f"🎯 [最终结论]:\n{assistant_message.content}")
|
||
|
|
print("==================================")
|
||
|
|
break
|
||
|
|
|
||
|
|
else:
|
||
|
|
print("\n[警告] 达到最大思考回合数限制，Agent 强行中止。")
|
||
|
|
|
||
|
|
if __name__ == "__main__":
|
||
|
|
run_agent("请仔细检查当前图纸，告诉我图纸中心那些小圆的内部，是否还包含了更小的同心圆？如果看不清，请务必放大确认。")
|