import json
from openai import OpenAI

client = OpenAI()

# 1. 定義 Agent 可用的工具庫 (Tools)
def get_weather(location: str):
    """模擬查詢天氣的 API"""
    if "台北" in location:
        return json.dumps({"location": "台北", "temperature": "25°C", "condition": "晴天"})
    return json.dumps({"location": location, "temperature": "20°C", "condition": "陰天"})

def calculate(expression: str):
    """模擬數學計算器"""
    try:
        result = eval(expression)
        return json.dumps({"result": result})
    except Exception as e:
        return json.dumps({"error": str(e)})

# 將函數對映至名稱
available_functions = {
    "get_weather": get_weather,
    "calculate": calculate
}

# 2. 定義提供給 LLM 的工具聲明 (Tool Definitions)
tools = [
    {
        "type": "function",
        "function": {
            "name": "get_weather",
            "description": "取得特定城市的目前天氣狀況",
            "parameters": {
                "type": "object",
                "properties": {
                    "location": {"type": "string", "description": "城市名稱，例如：台北"}
                },
                "required": ["location"],
            },
        },
    },
    {
        "type": "function",
        "function": {
            "name": "calculate",
            "description": "執行基礎數學算式計算",
            "parameters": {
                "type": "object",
                "properties": {
                    "expression": {"type": "string", "description": "數學算式，例如：'25 * 1.2'"}
                },
                "required": ["expression"],
            },
        },
    }
]

# 3. Agent 核心執行循環 (The Agent Loop)
def run_agent(prompt: str):
    messages = [
        {"role": "system", "content": "你是一個自主 Agent。你可以思考並選擇呼叫工具來完成使用者的任務。"},
        {"role": "user", "content": prompt}
    ]
    
    print(f"🎯 任務目標: {prompt}\n" + "="*40)

    # Agent Loop 開始
    while True:
        # LLM 思考 (Reasoning)
        response = client.chat.completions.create(
            model="gpt-4o-mini",
            messages=messages,
            tools=tools,
            tool_choice="auto"
        )
        
        response_message = response.choices[0].message
        messages.append(response_message) # 將 LLM 的回應記錄回 Message 歷史

        # 判斷是否需要呼叫工具 (Action)
        tool_calls = response_message.tool_calls
        if not tool_calls:
            # 如果 LLM 沒有要求呼叫任何工具，代表任務完成，輸出最終回答
            print(f"🤖 Agent 最終回答:\n{response_message.content}")
            break

        # 執行工具 (Acting) 並取得觀察結果 (Observation)
        for tool_call in tool_calls:
            function_name = tool_call.function.name
            function_args = json.loads(tool_call.function.arguments)
            
            print(f"🛠️ [Action] 呼叫工具: {function_name}({function_args})")
            
            # 找到對應函數並執行
            function_to_call = available_functions[function_name]
            function_response = function_to_call(**function_args)
            
            print(f"👀 [Observation] 工具傳回結果: {function_response}")

            # 將工具執行的結果 append 回訊息列表，讓 LLM 在下一輪循環看到
            messages.append({
                "tool_call_id": tool_call.id,
                "role": "tool",
                "name": function_name,
                "content": function_response,
            })
            print("-" * 40)

# 執行範例：一個需要跨工具多步驟推理的任務
run_agent("請幫我查台北的氣溫，然後算出如果氣溫上升 15% 會是多少度？")