Agent-Skills-for-Context-Engineering：为你的AI智能体打造“情境工具箱” 🛠️🤖

Agent-Skills-for-Context-Engineering/
├── skills/
│   ├── context_compression/      # 情境压缩技能
│   ├── memory_management/        # 记忆管理技能
│   ├── state_tracking/           # 状态跟踪技能
│   ├── tool_context_integration/ # 工具与情境集成技能
│   └── ...                       # 更多技能分类
├── architectures/                # 多智能体架构示例
├── patterns/                     # 情境工程模式文档
└── examples/                     # 实战示例

# 概念性代码，展示思路
class DynamicConversationSummarizer:
    def __init__(self, llm_client):
        self.llm = llm_client
        self.full_history = []  # 原始记录
        self.summary_chain = [] # 摘要链（分层记忆）

    def add_interaction(self, user_input, agent_response):
        self.full_history.append((user_input, agent_response))
        if self._needs_summarization():  # 触发条件：轮次、长度、话题转折
            latest_chunk = self.full_history[-10:]  # 取最近片段
            new_summary = self.llm.generate(
                f"Summarize the key points and state changes from: {latest_chunk}"
            )
            # 将详细记录压缩，保留摘要
            self.summary_chain.append(new_summary)
            self.full_history = self.full_history[-5:]  # 只保留少量最新原始记录

    def get_context_for_prompt(self):
        # 组合：最新摘要 + 少量原始历史，构成高效上下文
        return "\n".join(self.summary_chain[-2:]) + "\n" + format_history(self.full_history)

class GoalOrientedStateTracker:
    STATES = ['GREETING', 'PROBLEM_IDENTIFICATION', 'SOLUTION_PROVIDING', 'CONFIRMATION', 'CLOSING']

    def __init__(self, initial_state='GREETING'):
        self.current_state = initial_state
        self.goal_stack = []  # 支持嵌套目标（如主目标：售后，子目标：查询物流）
        self.extracted_slots = {}  # 收集到的关键信息（如订单号、问题描述）

    def transition(self, user_utterance, llm_helper):
        # 使用一个小型LLM调用或规则，判断状态是否应迁移
        next_state = llm_helper.determine_state(
            current_state=self.current_state,
            user_input=user_utterance,
            filled_slots=self.extracted_slots
        )
        if next_state and next_state in self.STATES:
            self.current_state = next_state
        # 同时进行信息槽填充
        self._extract_slots(user_utterance)
        return self.current_state

    def get_guidance(self):
        # 返回当前状态下的系统指令片段，引导智能体行为
        state_guides = {
            'PROBLEM_IDENTIFICATION': "Focus on asking clarifying questions to understand the core issue.",
            'SOLUTION_PROVIDING': "Provide clear, step-by-step solutions based on the identified problem.",
        }
        return state_guides.get(self.current_state, "")