Blame for app.py

import json
import re
import os
from openai import OpenAI
from agent_build_sdk.builder import AgentBuilder
from agent_build_sdk.model.model import AgentResp, AgentReq, STATUS_DISTRIBUTION, STATUS_ROUND, STATUS_VOTE, \
    STATUS_START, STATUS_VOTE_RESULT, STATUS_RESULT
from agent_build_sdk.sdk.agent import BasicAgent
from agent_build_sdk.sdk.agent import format_prompt
from agent_build_sdk.utils.logger import logger

from prompts import DESC_PROMPT, VOTE_PROMPT

class SpyAgent(BasicAgent):
    def perceive(self, req=AgentReq):
        logger.info("spy perceive: {}".format(req))
        if req.status == STATUS_START:  
            self.memory.clear()
            self.memory.set_variable("name", req.message)
            self.memory.append_history(
                '主持人: 女士们先生们，欢迎来到《谁是卧底》游戏！我们有一个由6名玩家组成的小组，在其中有一名卧底。让我们开始吧！每个人都会收到一张纸。其中5人的纸上拥有相同的单词，而卧底则会收到含义上相似的单词。我们将大多数人拿到的单词称为"公共词"，将卧底拿到的单词称为"卧底词"。一旦你拿到了你的单词，首先需要根据其他人的发言判断自己是否拿到了卧底词。如果判断自己拿到了卧底词，请猜测公共词是什么，然后描述公共词来混淆视听，避免被投票淘汰。如果判断自己拿到了公共词，请思考如何巧妙地描述它而不泄露它，不能让卧底察觉，也要给同伴暗示。每人每轮用一句话描述自己拿到的词语，每个人的描述禁止重复，话中不能出现所持词语。每轮描述完毕，所有在场的人投票选出怀疑是卧底的那个人，得票数最多的人出局。卧底出局则游戏结束，若卧底未出局，游戏继续。现在游戏开始。')
        elif req.status == STATUS_DISTRIBUTION:  
            self.memory.set_variable("word", req.word)
            self.memory.append_history(
                '主持人: 你好，{}，你分配到的单词是:{}'.format(self.memory.load_variable("name"), req.word))
        elif req.status == STATUS_ROUND:  
            if req.name:
                self.memory.append_history(req.name + ': ' + req.message)
            else:
                self.memory.append_history('主持人: 现在进入第{}轮。'.format(str(req.round)))
                self.memory.append_history('主持人: 每个玩家描述自己分配到的单词。')
        elif req.status == STATUS_VOTE:  
            self.memory.append_history(req.name + ': ' + req.message)
        elif req.status == STATUS_VOTE_RESULT:  
            out_player = req.name if req.name else req.message
            if out_player:
                self.memory.append_history('主持人: 投票结果是：{}。'.format(out_player))
            else:
                self.memory.append_history('主持人: 无人出局。')
        elif req.status == STATUS_RESULT:
            self.memory.append_history(req.message)
        else:
            raise NotImplementedError

    def interact(self, req=AgentReq) -> AgentResp:
        logger.info("spy interact: {}".format(req))
        if req.status == STATUS_ROUND:
            prompt = format_prompt(DESC_PROMPT,
                                   {"name": self.memory.load_variable("name"),
                                    "word": self.memory.load_variable("word"),
                                    "history": "\n".join(self.memory.load_history())
                                    })
            logger.info("prompt:\n" + prompt)
            result = self.llm_caller(prompt)
            logger.info("spy interact raw result: {}".format(result))
            
            # 提取 JSON 中的发言内容，如果解析失败给个保守的保底回复
            speak_content = self.extract_json_field(result, "speak", fallback="这东西挺常见的，大家应该都熟悉。")
            # 强制保障不超过120个字符的限制，留一点裕量限制在110
            if len(speak_content) > 110:
                speak_content = speak_content[:110]
            
            logger.info("spy final speak: {}".format(speak_content))
            return AgentResp(success=True, result=speak_content, errMsg=None)

        elif req.status == STATUS_VOTE:
            self.memory.append_history('主持人: 到了投票的时候了。每个人，请指向你认为可能是卧底的人。')
            choices = [name for name in req.message.split(",") if name != self.memory.load_variable("name")]  
            self.memory.set_variable("choices", choices)
            prompt = format_prompt(VOTE_PROMPT, {"name": self.memory.load_variable("name"),
                                                 "choices": choices,
                                                 "history": "\n".join(self.memory.load_history())
                                                 })
            logger.info("prompt:\n" + prompt)
            result = self.llm_caller(prompt)
            logger.info("spy interact raw result: {}".format(result))
            
            # 提取 JSON 中的投票目标，如果解析失败默认投给列表第一个人
            fallback_vote = choices[0] if choices else ""
            vote_target = self.extract_json_field(result, "vote", fallback=fallback_vote)
            logger.info("spy final vote: {}".format(vote_target))
            return AgentResp(success=True, result=vote_target, errMsg=None)

        else:
            raise NotImplementedError

    def extract_json_field(self, text, field, fallback):
        """解析大模型返回的 JSON 并提取指定字段"""
        if not text:
            return fallback
        try:
            # 1. 尝试匹配 ```json ... ``` 包裹的内容
            match = re.search(r'```(?:json)?\s*(\{.*?\})\s*```', text, re.DOTALL)
            json_str = match.group(1) if match else text
            
            # 2. 解析 JSON
            data = json.loads(json_str)
            
            # 3. 提取字段并清理可能存在的多余空格
            result = str(data.get(field, fallback)).strip()
            return result
        except Exception as e:
            logger.error(f"JSON解析失败: {e}, 原始内容: {text}")
            return fallback

    def llm_caller(self, prompt):
        client = OpenAI(
            api_key=os.getenv('API_KEY'),
            base_url=os.getenv('BASE_URL')
        )
        
        # 增加 System Prompt 强化 JSON 输出要求，并将 temperature 调至 0.7 增加发散性
        completion = client.chat.completions.create(
            model=self.model_name,
            messages=[
                {'role': 'system', 'content': '你是一个顶尖的逻辑推理大师和心理博弈专家。你必须严格按照要求以 JSON 格式输出你的思考过程和行动。'},
                {'role': 'user', 'content': prompt}
            ],
            temperature=0.7 
        )
        try:
            return completion.choices[0].message.content
        except Exception as e:
            logger.error(f"LLM调用异常: {e}")
            return None


if __name__ == '__main__':
    name = 'spy'
    agent_builder = AgentBuilder(name, agent=SpyAgent(name, model_name=os.getenv('MODEL_NAME')))
    agent_builder.start()