03 · 核心循环：让大模型学会"思考 → 行动 → 观察"

你 > 现在几点？顺便算下 365*24 等于多少

知了 > （内部实际发生了什么）
  1. [思考] 用户问两件事，都需要工具。并行调用：
     → get_current_time()
     → calculator("365*24")
  2. [观察] 时间工具返回 "2026-09-30 14:22 星期三"
     计算器返回 "365*24 = 8760"
  3. [思考] 拿到全部事实，可以回答了
  4. [回答] "现在是 2026-09-30 14:22（星期三）。一年有 8760 小时。"

[1] {"role":"system",  "content":"你是知了…可用工具：calculator…"}
[2] {"role":"user",    "content":"现在几点？顺便算下 365*24"}
[3] {"role":"assistant","tool_calls":[{"function":{"name":"get_current_time","arguments":"{}"}}]}
[4] {"role":"tool",    "tool_call_id":"call_1","content":"2026-09-30 14:22 星期三"}
[5] {"role":"assistant","tool_calls":[{"function":{"name":"calculator","arguments":"{\"expression\":\"365*24\"}"}}]}
[6] {"role":"tool",    "tool_call_id":"call_2","content":"365*24 = 8760"}
[7] {"role":"assistant","content":"现在是 2026-09-30 14:22（星期三）…"}

def chat(self, messages, tools=None, temperature=0.7):
    kwargs = dict(model=self.model, messages=messages, temperature=temperature)
    if tools:
        kwargs["tools"] = tools
        kwargs["tool_choice"] = "auto"     # 模型自己决定用不用工具

    resp = self.client.chat.completions.create(**kwargs)
    msg = resp.choices[0].message

    result = {"role": "assistant", "content": msg.content or ""}
    if getattr(msg, "tool_calls", None):
        result["tool_calls"] = [
            {"id": tc.id, "type": "function",
             "function": {"name": tc.function.name, "arguments": tc.function.arguments}}
            for tc in msg.tool_calls
        ]
    return result

def run(self, user_input: str) -> dict:
    # 1) 用户发言进入记忆
    self.memory.add("user", user_input)

    # 2) 系统提示词（人设+规矩+技能目录）+ 检索到的记忆/知识，注入本轮
    system_text = build_system_prompt(...)
    context_blocks = [...]                       # 长期记忆片段、知识库片段
    if context_blocks:
        system_text += "\n\n" + "\n\n".join(context_blocks)

    # 3) 本轮完整消息 = 系统提示 + 会话历史
    messages = [{"role": "system", "content": system_text}] + self.memory.view()

    # ★★★ 核心循环：思考 → 行动 → 观察 ★★★
    steps = []
    for _ in range(self.cfg.max_iterations):      # 上限 8 次，防失控
        assistant_msg = self.llm.chat(messages, tools=self.registry.schemas())
        messages.append(assistant_msg)            # 模型发言必须回填！

        tool_calls = assistant_msg.get("tool_calls")
        if not tool_calls:
            # 模型不再要工具 → 这就是最终回答
            reply = assistant_msg["content"].strip()
            self.memory.add("assistant", reply)
            self.memory.maybe_summarize(self.llm) # 历史太长就滚动摘要
            return {"reply": reply, "steps": steps}

        for call in tool_calls:                   # 可能并行要多个工具
            result = self.registry.call(call["function"]["name"],
                                        call["function"]["arguments"])
            steps.append({"tool": call["function"]["name"], "result": result[:300]})
            # 协议要求：结果以 role="tool" 回传，id 对应
            messages.append({"role": "tool",
                             "tool_call_id": call["id"], "content": result})

    # 4) 兜底：循环超限，去掉工具强制模型总结
    final = self.llm.chat(messages)
    ...

python main.py
你 > 现在几点？3000 美元换成人民币大概多少？（汇率自己查）