# L0 元规则示例
{
    "dont_delete_files": True,
    "confirm_before_execute": ["rm", "del", "format"],
    "max_retry_attempts": 3,
    "require_human_approval_for": ["payment", "send_email"]
}

# 任务执行时，记忆逐步积累
task_start:
    - 加载 L1（快速路由）
    - 加载 L2（全局事实）
    - 加载 L3（相关 Skills）
    - 加载 L4（历史类似任务）

task_progress:
    - 每步执行后更新 working_checkpoint
    - 遇到新模式时写入 L2

task_complete:
    - 将执行路径固化为 L3 Skill
    - 将对话记录归档为 L4 Session

# agent_loop.py 核心逻辑（约 100 行）

async def agent_loop():
    state = await load_context()          # 加载上下文（含记忆）
    
    while not task_completed(state):
        # 1. 感知环境状态
        observation = await perceive(state)
        
        # 2. 任务推理
        reasoning = await reason(state, observation)
        
        # 3. 执行工具
        result = await execute_tools(state, reasoning)
        
        # 4. 更新记忆
        await update_memory(state, result)
        
        # 5. 检查是否完成
        if is_task_complete(state):
            await crystallize_skill(state)  # 固化为 Skill
            break
    
    await archive_session(state)          # 归档会话

# 动态安装包
code_run("pip install mootdx")
    ↓
# 写脚本
code_run("""
import mootdx
# 选股逻辑
stocks = mo.TopList(field='总市值').query('现价>10')
""")
    ↓
# 保存为 Skill
# 从此以后一句话即可调用

// 打开网页
web_execute_js({
    action: "goto",
    url: "https://www.meituan.com"
})

// 点击元素
web_execute_js({
    action: "click",
    selector: ".search-input"
})

// 输入文本
web_execute_js({
    action: "input",
    selector: ".search-input",
    text: "珍珠奶茶"
})

// 截图（视觉反馈）
web_execute_js({
    action: "screenshot"
})

# Step 1: 安装依赖
code_run("pip install mootdx")

# Step 2: 编写选股脚本
code_run("""
import mootdx
mo = mootdx.Minute()

# 选股条件：EXPMA 金叉 + 换手率 > 5%
stocks = mo.Market(market='gem')
# 筛选逻辑...
""")

# Step 3: 配置定时任务
code_run("""
import schedule

def check_stocks():
    # 执行选股逻辑
    ...

schedule.every().day.at("09:30").do(check_stocks)
""")

# Step 4: 保存为 Skill
# 以后只需说"监控股票并提醒我"

# mykey.py
# 无需额外配置

# 启动
python frontends/wechatapp.py

# mykey.py
qq_app_id = "YOUR_APP_ID"
qq_app_secret = "YOUR_APP_SECRET"
qq_allowed_users = ["YOUR_USER_OPENID"]

python frontends/qqapp.py

# mykey.py
fs_app_id = "cli_xxx"
fs_app_secret = "xxx"
fs_allowed_users = ["ou_xxx"]

python frontends/fsapp.py

# 企业微信
wecom_bot_id = "your_bot_id"
wecom_secret = "your_secret"

# 钉钉
dingtalk_client_id = "your_app_key"
dingtalk_client_secret = "your_secret"

python launch.pyw
# 或
streamlit run frontends/stapp.py

# 1. 克隆仓库
git clone https://github.com/lsdefine/GenericAgent.git
cd GenericAgent

# 2. 安装最小依赖
pip install streamlit pywebview

# 3. 配置 API Key
cp mykey_template.py mykey.py
# 编辑 mykey.py，填入你的 LLM API Key

# 4. 启动
python launch.pyw

# mykey.py
# 支持 Claude / Gemini / Kimi / MiniMax / OpenAI

openai_api_key = "sk-xxx"  # OpenAI
anthropic_api_key = "sk-ant-xxx"  # Claude
gemini_api_key = "xxx"  # Gemini
kimi_api_key = "xxx"  # Kimi
minimax_api_key = "xxx"  # MiniMax

# 启动后，在输入框输入：
"帮我打开百度并搜索 'AI Agent'"

# Agent 会：
# 1. 打开浏览器
# 2. 导航到百度
# 3. 搜索 AI Agent
# 4. 保存这个技能到 L3

# agent_loop.py 完整伪代码

class GenericAgent:
    def __init__(self, model, memory_system):
        self.model = model
        self.memory = memory_system
        self.tools = ToolRegistry()
    
    async def run(self, task: str):
        # 1. 加载上下文
        context = await self.memory.load_context(task)
        
        # 2. 主循环
        for step in range(max_steps):
            # 感知
            observation = await self.perceive(context)
            
            # 推理
            plan = await self.model.think(context, observation)
            
            # 执行
            if plan.needs_tool:
                result = await self.tools.execute(plan.tool, plan.args)
                context.add_result(result)
            else:
                # 直接回复
                return plan.response
            
            # 更新记忆
            await self.memory.update(context, step)
            
            # 检查完成
            if self.is_complete(context):
                break
        
        # 3. 固化为 Skill
        await self.crystallize_skill(context)
        
        # 4. 归档会话
        await self.archive_session(context)

# memory/l3_task_skills.py 示例

class TaskSkills:
    def __init__(self, storage_path):
        self.skills = self.load_skills(storage_path)
    
    async def add_skill(self, task: str, execution_path: list):
        skill = {
            "name": self.generate_name(task),
            "trigger": self.generate_trigger(task),
            "steps": execution_path,
            "created_at": now()
        }
        self.skills.append(skill)
        await self.save()
    
    async def recall(self, user_input: str) -> Skill | None:
        # 使用 embedding 匹配最相似的 Skill
        embedding = self.embed(user_input)
        for skill in self.skills:
            similarity = cosine_similarity(embedding, skill.embedding)
            if similarity > threshold:
                return skill
        return None