diff --git a/CHECKLIST.md b/CHECKLIST.md
new file mode 100644
index 0000000..ab8c1da
--- /dev/null
+++ b/CHECKLIST.md
@@ -0,0 +1,13 @@
+# Demo1 CHECKLIST
+
+## 过关前自检
+
+- [ ] `CLIAgent()` 可以正常创建
+- [ ] 初始状态是 `IDLE`
+- [ ] `process()` 对普通输入返回字符串
+- [ ] `process()` 对空输入不会崩
+- [ ] 数学输入能触发计算逻辑
+- [ ] `reset()` 后状态恢复为 `IDLE`
+- [ ] `python demo1_cli_agent/main.py` 能启动
+- [ ] `pytest demo1_cli_agent/tests/test_agent.py -q` 通过
+- [ ] push 到 `demo1-starter` 后，Actions 通过
diff --git a/FAQ.md b/FAQ.md
new file mode 100644
index 0000000..ad37b63
--- /dev/null
+++ b/FAQ.md
@@ -0,0 +1,27 @@
+# Demo1 FAQ
+
+## 1. 为什么 `process()` 一运行就报错？
+
+先确认你是不是还保留了 `NotImplementedError`。  
+这一关最常见的问题就是只改了一部分逻辑，但忘了去掉占位异常。
+
+## 2. 数学输入一定要做很完整吗？
+
+不用。  
+先满足测试里最基本的表达式场景，再逐步扩展。
+
+## 3. 为什么测试强调 `process()` 返回字符串？
+
+因为这一关先考察“接口稳定性”。  
+哪怕还不够智能，也要保证外部调用方始终拿到字符串结果。
+
+## 4. `reset()` 到底要做什么？
+
+至少两件事：
+
+- 清空上下文
+- 把状态恢复为 `IDLE`
+
+## 5. 怎么排查多轮对话相关问题？
+
+先打印或观察你保存的 history 结构是否真的在追加，而不是每次都被覆盖。
diff --git a/HINTS.md b/HINTS.md
new file mode 100644
index 0000000..b793eec
--- /dev/null
+++ b/HINTS.md
@@ -0,0 +1,22 @@
+# Demo1 HINTS
+
+只给思路，不给答案。
+
+## 你可以先想清楚的点
+
+- `process()` 的最低要求不是“聪明”，而是“稳定返回字符串”
+- 测试最关心的是接口行为，不是类设计得多复杂
+- 问候、数学、空输入，这三类先处理好，其他输入再统一兜底
+
+## 容易卡住的地方
+
+- 计算逻辑不需要一开始就支持很复杂的表达式
+- `reset()` 不只是清空历史，还要让 `state` 回到 `IDLE`
+- 不要让 `process()` 在空字符串时抛异常
+
+## 实现顺序建议
+
+1. 先让类能实例化
+2. 再让 `process()` 返回固定字符串
+3. 再逐步加意图分支
+4. 最后补 CLI 循环
diff --git a/README.md b/README.md
index 77b381d..1a7d1f9 100644
--- a/README.md
+++ b/README.md
@@ -21,6 +21,32 @@
   <strong>Agent</strong> · <strong>ML</strong> · <strong>FastAPI</strong> · <strong>RAG</strong> · <strong>SSE</strong> · <strong>CI Unlock Flow</strong>
 </p>
 
+## 当前分支
+
+你现在位于 `demo1-starter`。
+
+- 这个分支是第一关学习骨架，不是完整答案
+- 目标是自己实现 `demo1_cli_agent/` 里的核心逻辑
+- 当前关卡通过后，push 到 `demo1-starter` 会自动解锁下一关
+- 完整参考实现保留在 `main` 分支
+
+## 学习入口
+
+- 先读 [TODO.md](TODO.md)
+- 卡住时看 [HINTS.md](HINTS.md)
+- 易错点和排查看 [FAQ.md](FAQ.md)
+- 提交前对照 [CHECKLIST.md](CHECKLIST.md)
+- 完成后回看 [REFLECTION.md](REFLECTION.md)
+- 再看 `demo1_cli_agent/tests/test_agent.py`
+- 然后补 `demo1_cli_agent/agent.py` 和 `demo1_cli_agent/main.py`
+
+## 建议实现步骤
+
+1. 先让 `CLIAgent` 能被正常实例化。
+2. 让 `process()` 对任何输入都返回字符串。
+3. 加上简单问候和计算逻辑。
+4. 最后补 `reset()` 和 CLI 循环。
+
 ## 快速导航
 
 - [在线演示](#在线演示)
diff --git a/REFLECTION.md b/REFLECTION.md
new file mode 100644
index 0000000..a7b9859
--- /dev/null
+++ b/REFLECTION.md
@@ -0,0 +1,13 @@
+# Demo1 REFLECTION
+
+学完这一关，你应该能说清楚这些事：
+
+- 一个最小可用的 Agent 至少需要输入、状态、上下文和输出
+- 状态机为什么适合描述 Agent 的执行过程
+- 为什么即使没有 LLM，也能先实现一个规则版 Agent
+- 为什么测试优先关注接口稳定性，而不是复杂能力
+
+如果你已经完成本关，说明你已经具备：
+
+- 实现一个最小 CLI Agent 的能力
+- 为后续任务型 Agent 打基础的能力
diff --git a/TODO.md b/TODO.md
new file mode 100644
index 0000000..bbf7047
--- /dev/null
+++ b/TODO.md
@@ -0,0 +1,30 @@
+# Demo1 TODO
+
+当前目标：完成一个最基础的 CLI Agent。
+
+## 你需要实现的文件
+
+- `demo1_cli_agent/agent.py`
+- `demo1_cli_agent/main.py`
+
+## 建议实现步骤
+
+1. 在 `CLIAgent.__init__` 中维护 `state` 和对话历史。
+2. 在 `process()` 中处理空输入，保证始终返回字符串。
+3. 加入简单意图识别：问候、计算、时间查询。
+4. 为数学表达式实现一个最小可用的计算逻辑。
+5. 在 `reset()` 中恢复 `IDLE` 状态并清空上下文。
+6. 在 `main.py` 中补一个最基本的 CLI 循环。
+
+## 完成标准
+
+- `pytest demo1_cli_agent/tests/test_agent.py -q` 通过
+- 推送到 `demo1-starter` 后，GitHub Actions 成功运行
+- Issues 页面出现 “Demo2 已解锁”
+
+## 卡住时看哪里
+
+- 当前分支的 `README.md`
+- `docs/demo_specs.md`
+- `docs/interview_qa.md`
+- 完整答案在 `main` 分支
diff --git a/demo1_cli_agent/agent.py b/demo1_cli_agent/agent.py
index 248b312..cd02857 100644
--- a/demo1_cli_agent/agent.py
+++ b/demo1_cli_agent/agent.py
@@ -1,98 +1,24 @@
-from __future__ import annotations
-
 from collections import deque
-from datetime import datetime
-from typing import Deque, Dict, List
-import ast
-import operator
-import re
-
-
-class SafeCalculator:
-    _bin_ops = {
-        ast.Add: operator.add,
-        ast.Sub: operator.sub,
-        ast.Mult: operator.mul,
-        ast.Div: operator.truediv,
-        ast.Pow: operator.pow,
-        ast.Mod: operator.mod,
-    }
-    _unary_ops = {
-        ast.UAdd: operator.pos,
-        ast.USub: operator.neg,
-    }
-
-    def evaluate(self, expression: str) -> float:
-        tree = ast.parse(expression, mode="eval")
-        return float(self._eval_node(tree.body))
-
-    def _eval_node(self, node: ast.AST) -> float:
-        if isinstance(node, ast.Constant) and isinstance(node.value, (int, float)):
-            return float(node.value)
-        if isinstance(node, ast.BinOp) and type(node.op) in self._bin_ops:
-            left = self._eval_node(node.left)
-            right = self._eval_node(node.right)
-            return self._bin_ops[type(node.op)](left, right)
-        if isinstance(node, ast.UnaryOp) and type(node.op) in self._unary_ops:
-            operand = self._eval_node(node.operand)
-            return self._unary_ops[type(node.op)](operand)
-        raise ValueError("Unsupported expression")
 
 
 class CLIAgent:
+    """Demo1 starter: implement the basic CLI agent yourself."""
+
     def __init__(self, max_context_len: int = 10):
         self.max_context_len = max_context_len
         self.state = "IDLE"
-        self._history: Deque[Dict[str, str]] = deque(maxlen=max_context_len)
-        self._calculator = SafeCalculator()
+        self.history = deque(maxlen=max_context_len)
 
     def process(self, user_input: str) -> str:
-        cleaned = (user_input or "").strip()
-        self.state = "THINKING"
-        self._history.append({"role": "user", "content": cleaned})
-
-        if not cleaned:
-            reply = "可以继续发我一个问题，或者让我帮你算一道题。"
-        elif self._looks_like_math(cleaned):
-            self.state = "ACTING"
-            reply = self._handle_math(cleaned)
-        elif any(token in cleaned.lower() for token in ["时间", "日期", "date", "time"]):
-            self.state = "ACTING"
-            reply = f"当前时间是 {datetime.now().strftime('%Y-%m-%d %H:%M:%S')}。"
-        elif any(token in cleaned.lower() for token in ["你好", "hi", "hello"]):
-            reply = "你好，我是这个学习项目里的 CLI Agent，可以陪你聊天，也能做简单计算。"
-        elif cleaned.lower() in {"exit", "quit", "bye"}:
-            reply = "本轮会话先到这里，随时可以继续。"
-        else:
-            reply = "我理解到这是一个普通问题。目前我支持基础对话、时间查询和简单数学计算。"
-
-        self.state = "DONE"
-        self._history.append({"role": "assistant", "content": reply})
-        return reply
+        """
+        TODO:
+        1. update agent state
+        2. record conversation history
+        3. route simple intents such as greeting / math / datetime
+        4. return a string reply instead of raising
+        """
+        raise NotImplementedError("Implement CLIAgent.process for Demo1")
 
     def reset(self) -> None:
-        self._history.clear()
-        self.state = "IDLE"
-
-    def _looks_like_math(self, text: str) -> bool:
-        if "计算" in text:
-            return True
-        return bool(re.search(r"\d+\s*[\+\-\*/%]\s*\d+", text))
-
-    def _handle_math(self, text: str) -> str:
-        match = re.search(r"([-+*/%().\d\s]+)", text)
-        expression = match.group(1).strip() if match else ""
-        try:
-            result = self._calculator.evaluate(expression)
-        except Exception:
-            return "这道题我没能正确解析，你可以换成更标准的表达式，例如 1 + 1。"
-
-        if result.is_integer():
-            result_text = str(int(result))
-        else:
-            result_text = f"{result:.4f}".rstrip("0").rstrip(".")
-        return f"计算结果是 {result_text}。"
-
-    @property
-    def history(self) -> List[Dict[str, str]]:
-        return list(self._history)
+        """TODO: clear context and set state back to IDLE."""
+        raise NotImplementedError("Implement CLIAgent.reset for Demo1")
diff --git a/demo1_cli_agent/main.py b/demo1_cli_agent/main.py
index 8ec3987..710f135 100644
--- a/demo1_cli_agent/main.py
+++ b/demo1_cli_agent/main.py
@@ -2,14 +2,14 @@
 
 
 def main() -> None:
+    """
+    Demo1 starter:
+    - instantiate CLIAgent
+    - build a simple REPL loop
+    - support exit / quit commands
+    """
     agent = CLIAgent()
-    print("CLI Agent 已启动，输入 exit 结束。")
-    while True:
-        user_input = input(">>> ").strip()
-        if user_input.lower() in {"exit", "quit"}:
-            print("Bye.")
-            break
-        print(agent.process(user_input))
+    print("Demo1 starter loaded. Please implement the CLI loop in main.py.")
 
 
 if __name__ == "__main__":
diff --git a/demo2_task_agent/priority_queue.py b/demo2_task_agent/priority_queue.py
deleted file mode 100644
index 0a0e411..0000000
--- a/demo2_task_agent/priority_queue.py
+++ /dev/null
@@ -1,23 +0,0 @@
-from __future__ import annotations
-
-import heapq
-from typing import List, Tuple
-
-
-class PriorityQueue:
-    _priority_rank = {"high": 0, "mid": 1, "medium": 1, "low": 2}
-
-    def __init__(self) -> None:
-        self._heap: List[Tuple[int, int, dict]] = []
-        self._counter = 0
-
-    def push(self, task: dict) -> None:
-        rank = self._priority_rank.get(task.get("priority", "low"), 2)
-        heapq.heappush(self._heap, (rank, self._counter, task))
-        self._counter += 1
-
-    def pop(self) -> dict:
-        return heapq.heappop(self._heap)[2]
-
-    def peek(self) -> dict:
-        return self._heap[0][2]
diff --git a/demo2_task_agent/task_store.py b/demo2_task_agent/task_store.py
deleted file mode 100644
index 4cc42d6..0000000
--- a/demo2_task_agent/task_store.py
+++ /dev/null
@@ -1,40 +0,0 @@
-from __future__ import annotations
-
-from dataclasses import dataclass, asdict
-from typing import Dict, List, Optional
-from uuid import uuid4
-
-
-@dataclass
-class Task:
-    id: str
-    title: str
-    priority: str
-    due_date: str
-
-
-class TaskStore:
-    def __init__(self) -> None:
-        self._tasks: Dict[str, Task] = {}
-
-    def add(self, title: str, priority: str, due_date: str) -> dict:
-        task = Task(id=str(uuid4()), title=title, priority=priority, due_date=due_date)
-        self._tasks[task.id] = task
-        return asdict(task)
-
-    def get_all(self) -> List[dict]:
-        return [asdict(task) for task in self._tasks.values()]
-
-    def get(self, task_id: str) -> Optional[dict]:
-        task = self._tasks.get(task_id)
-        return asdict(task) if task else None
-
-    def update(self, task_id: str, **kwargs) -> dict:
-        task = self._tasks[task_id]
-        for field in ("title", "priority", "due_date"):
-            if field in kwargs and kwargs[field] is not None:
-                setattr(task, field, kwargs[field])
-        return asdict(task)
-
-    def delete(self, task_id: str) -> bool:
-        return self._tasks.pop(task_id, None) is not None
diff --git a/demo2_task_agent/tool_registry.py b/demo2_task_agent/tool_registry.py
deleted file mode 100644
index 41f6da7..0000000
--- a/demo2_task_agent/tool_registry.py
+++ /dev/null
@@ -1,40 +0,0 @@
-from __future__ import annotations
-
-from dataclasses import dataclass
-from typing import Any, Callable, Dict
-
-
-class ToolRegistryError(Exception):
-    pass
-
-
-class PermissionDeniedError(ToolRegistryError):
-    pass
-
-
-class ToolNotFoundError(ToolRegistryError):
-    pass
-
-
-@dataclass
-class ToolSpec:
-    func: Callable[..., Any]
-    permission: str
-
-
-class ToolRegistry:
-    _permission_order = {"read": 1, "write": 2, "admin": 3}
-
-    def __init__(self) -> None:
-        self._tools: Dict[str, ToolSpec] = {}
-
-    def register(self, name: str, func: Callable[..., Any], permission: str) -> None:
-        self._tools[name] = ToolSpec(func=func, permission=permission)
-
-    def call(self, name: str, args: Dict[str, Any], ctx_permission: str) -> Any:
-        if name not in self._tools:
-            raise ToolNotFoundError(name)
-        spec = self._tools[name]
-        if self._permission_order.get(ctx_permission, 0) < self._permission_order.get(spec.permission, 0):
-            raise PermissionDeniedError(f"{ctx_permission} cannot call {name}")
-        return spec.func(**args)
diff --git a/demo3_ml_visual/data_loader.py b/demo3_ml_visual/data_loader.py
deleted file mode 100644
index 4fb71a0..0000000
--- a/demo3_ml_visual/data_loader.py
+++ /dev/null
@@ -1,26 +0,0 @@
-from __future__ import annotations
-
-from pathlib import Path
-from typing import Tuple
-
-import pandas as pd
-from sklearn.datasets import load_iris, load_wine, load_breast_cancer
-
-
-class DataLoader:
-    def load_csv(self, path: str) -> Tuple[pd.DataFrame, pd.Series]:
-        df = pd.read_csv(path)
-        if df.shape[1] < 2:
-            raise ValueError("CSV must contain features and a target column")
-        return df.iloc[:, :-1], df.iloc[:, -1]
-
-    def load_builtin(self, name: str = "iris"):
-        mapping = {
-            "iris": load_iris,
-            "wine": load_wine,
-            "breast_cancer": load_breast_cancer,
-        }
-        if name not in mapping:
-            raise ValueError(f"Unsupported builtin dataset: {name}")
-        data = mapping[name](return_X_y=True)
-        return data
diff --git a/demo3_ml_visual/feature_engineer.py b/demo3_ml_visual/feature_engineer.py
deleted file mode 100644
index 1df1846..0000000
--- a/demo3_ml_visual/feature_engineer.py
+++ /dev/null
@@ -1,24 +0,0 @@
-from __future__ import annotations
-
-import numpy as np
-from sklearn.impute import SimpleImputer
-from sklearn.pipeline import Pipeline
-from sklearn.preprocessing import StandardScaler
-
-
-class FeatureEngineer:
-    def __init__(self) -> None:
-        self.pipeline = Pipeline(
-            steps=[
-                ("imputer", SimpleImputer(strategy="median")),
-                ("scaler", StandardScaler()),
-            ]
-        )
-
-    def fit_transform(self, X):
-        transformed = self.pipeline.fit_transform(X)
-        return np.asarray(transformed)
-
-    def transform(self, X):
-        transformed = self.pipeline.transform(X)
-        return np.asarray(transformed)
diff --git a/demo3_ml_visual/main.py b/demo3_ml_visual/main.py
deleted file mode 100644
index eaa504f..0000000
--- a/demo3_ml_visual/main.py
+++ /dev/null
@@ -1,20 +0,0 @@
-from sklearn.model_selection import train_test_split
-
-from data_loader import DataLoader
-from feature_engineer import FeatureEngineer
-from model_trainer import ModelTrainer
-
-
-def main() -> None:
-    X, y = DataLoader().load_builtin("iris")
-    fe = FeatureEngineer()
-    X_t = fe.fit_transform(X)
-    X_train, X_test, y_train, y_test = train_test_split(X_t, y, test_size=0.2, random_state=42)
-    trainer = ModelTrainer()
-    models = trainer.train(X_train, y_train)
-    metrics = trainer.evaluate(models, X_test, y_test)
-    print(metrics)
-
-
-if __name__ == "__main__":
-    main()
diff --git a/demo3_ml_visual/model_io.py b/demo3_ml_visual/model_io.py
deleted file mode 100644
index f59ec06..0000000
--- a/demo3_ml_visual/model_io.py
+++ /dev/null
@@ -1,11 +0,0 @@
-from __future__ import annotations
-
-import joblib
-
-
-def save_model(model, path: str) -> None:
-    joblib.dump(model, path)
-
-
-def load_model(path: str):
-    return joblib.load(path)
diff --git a/demo3_ml_visual/model_trainer.py b/demo3_ml_visual/model_trainer.py
deleted file mode 100644
index 8ec1b8d..0000000
--- a/demo3_ml_visual/model_trainer.py
+++ /dev/null
@@ -1,42 +0,0 @@
-from __future__ import annotations
-
-from typing import Dict
-
-from sklearn.ensemble import RandomForestClassifier
-from sklearn.linear_model import LogisticRegression
-from sklearn.metrics import accuracy_score, f1_score
-
-try:
-    from xgboost import XGBClassifier  # type: ignore
-except Exception:  # pragma: no cover
-    XGBClassifier = None
-
-
-class ModelTrainer:
-    def train(self, X_train, y_train) -> Dict[str, object]:
-        models: Dict[str, object] = {
-            "logistic_regression": LogisticRegression(max_iter=500),
-            "random_forest": RandomForestClassifier(n_estimators=100, random_state=42),
-        }
-        if XGBClassifier is not None:
-            models["xgboost"] = XGBClassifier(
-                n_estimators=50,
-                max_depth=3,
-                learning_rate=0.1,
-                eval_metric="mlogloss",
-                random_state=42,
-            )
-
-        for model in models.values():
-            model.fit(X_train, y_train)
-        return models
-
-    def evaluate(self, models, X_test, y_test):
-        metrics = {}
-        for name, model in models.items():
-            preds = model.predict(X_test)
-            metrics[name] = {
-                "accuracy": float(accuracy_score(y_test, preds)),
-                "f1": float(f1_score(y_test, preds, average="weighted")),
-            }
-        return metrics
diff --git "a/demo4_\347\273\274\345\220\210\351\241\271\347\233\256/main.py" "b/demo4_\347\273\274\345\220\210\351\241\271\347\233\256/main.py"
deleted file mode 100644
index 8612b7a..0000000
--- "a/demo4_\347\273\274\345\220\210\351\241\271\347\233\256/main.py"
+++ /dev/null
@@ -1,75 +0,0 @@
-from __future__ import annotations
-
-import json
-from typing import Iterator, List
-from uuid import uuid4
-
-from fastapi import FastAPI, HTTPException
-from pydantic import BaseModel
-from fastapi.responses import StreamingResponse
-
-
-app = FastAPI(title="Demo4 Integrated App")
-TASKS = {}
-
-
-class ChatRequest(BaseModel):
-    message: str
-
-
-class TaskCreateRequest(BaseModel):
-    title: str
-    priority: str
-    due_date: str
-
-
-@app.get("/health")
-def health():
-    return {"status": "ok", "version": "0.4.0"}
-
-
-@app.post("/chat")
-def chat(payload: ChatRequest):
-    message = payload.message.strip()
-    tools_used: List[str] = []
-    if "任务" in message and any(token in message for token in ["列", "list", "所有"]):
-        tools_used.append("list_tasks")
-        reply = f"当前共有 {len(TASKS)} 个任务。"
-    elif any(token in message for token in ["你好", "hi", "hello"]):
-        reply = "你好，这里是 Demo4 综合项目接口。"
-    else:
-        reply = "消息已收到，我可以处理聊天和任务相关请求。"
-    return {"reply": reply, "tools_used": tools_used}
-
-
-@app.get("/tasks")
-def get_tasks():
-    return {"tasks": list(TASKS.values())}
-
-
-@app.post("/tasks", status_code=201)
-def create_task(payload: TaskCreateRequest):
-    task = payload.model_dump()
-    task["id"] = str(uuid4())
-    TASKS[task["id"]] = task
-    return task
-
-
-@app.delete("/tasks/{task_id}")
-def delete_task(task_id: str):
-    if task_id not in TASKS:
-        raise HTTPException(status_code=404, detail="Task not found")
-    del TASKS[task_id]
-    return {"deleted": True}
-
-
-def _sse_event_stream(message: str) -> Iterator[str]:
-    for chunk in ["demo4", "stream", message]:
-        payload = json.dumps({"type": "token", "content": chunk}, ensure_ascii=False)
-        yield f"data: {payload}\n\n"
-    yield "data: {\"type\": \"done\"}\n\n"
-
-
-@app.get("/stream")
-def stream(message: str):
-    return StreamingResponse(_sse_event_stream(message), media_type="text/event-stream")
diff --git a/demo5_full_project/app/__init__.py b/demo5_full_project/app/__init__.py
deleted file mode 100644
index 8b13789..0000000
--- a/demo5_full_project/app/__init__.py
+++ /dev/null
@@ -1 +0,0 @@
-
diff --git a/demo5_full_project/app/core/__init__.py b/demo5_full_project/app/core/__init__.py
deleted file mode 100644
index 8b13789..0000000
--- a/demo5_full_project/app/core/__init__.py
+++ /dev/null
@@ -1 +0,0 @@
-
diff --git a/demo5_full_project/app/core/rag/__init__.py b/demo5_full_project/app/core/rag/__init__.py
deleted file mode 100644
index 8b13789..0000000
--- a/demo5_full_project/app/core/rag/__init__.py
+++ /dev/null
@@ -1 +0,0 @@
-
diff --git a/demo5_full_project/app/core/rag/embedder.py b/demo5_full_project/app/core/rag/embedder.py
deleted file mode 100644
index b4ee5dd..0000000
--- a/demo5_full_project/app/core/rag/embedder.py
+++ /dev/null
@@ -1,21 +0,0 @@
-from __future__ import annotations
-
-import hashlib
-
-import numpy as np
-
-
-class Embedder:
-    def __init__(self, dimension: int = 32):
-        self.dimension = dimension
-
-    def encode(self, text: str) -> np.ndarray:
-        text = text or ""
-        buckets = np.zeros(self.dimension, dtype=float)
-        for token in text.encode("utf-8"):
-            buckets[token % self.dimension] += 1.0
-        digest = hashlib.sha256(text.encode("utf-8")).digest()
-        for idx, byte in enumerate(digest[: self.dimension]):
-            buckets[idx] += byte / 255.0
-        norm = np.linalg.norm(buckets)
-        return buckets if norm == 0 else buckets / norm
diff --git a/demo5_full_project/app/core/rag/retriever.py b/demo5_full_project/app/core/rag/retriever.py
deleted file mode 100644
index 66146da..0000000
--- a/demo5_full_project/app/core/rag/retriever.py
+++ /dev/null
@@ -1,35 +0,0 @@
-from __future__ import annotations
-
-from dataclasses import dataclass
-from typing import List
-
-import numpy as np
-
-from app.core.rag.embedder import Embedder
-
-
-@dataclass
-class DocumentChunk:
-    text: str
-
-
-class Retriever:
-    def __init__(self) -> None:
-        self.embedder = Embedder()
-        self._documents = [
-            DocumentChunk("Agent 是一种能够感知、决策并执行动作的软件实体。"),
-            DocumentChunk("RAG 会先检索知识，再把检索结果拼到生成提示词中。"),
-            DocumentChunk("SSE 适合服务端向客户端单向推送流式文本。"),
-        ]
-
-    def retrieve(self, query: str, top_k: int = 3) -> List[str]:
-        if not self._documents:
-            return []
-        query_vec = self.embedder.encode(query)
-        scored = []
-        for doc in self._documents:
-            doc_vec = self.embedder.encode(doc.text)
-            score = float(np.dot(query_vec, doc_vec))
-            scored.append((score, doc.text))
-        scored.sort(reverse=True, key=lambda item: item[0])
-        return [text for _, text in scored[: max(0, min(top_k, 5))]]
diff --git a/demo5_full_project/app/ml/__init__.py b/demo5_full_project/app/ml/__init__.py
deleted file mode 100644
index 8b13789..0000000
--- a/demo5_full_project/app/ml/__init__.py
+++ /dev/null
@@ -1 +0,0 @@
-
diff --git a/demo5_full_project/app/ml/model_io.py b/demo5_full_project/app/ml/model_io.py
deleted file mode 100644
index 11b195b..0000000
--- a/demo5_full_project/app/ml/model_io.py
+++ /dev/null
@@ -1,24 +0,0 @@
-from __future__ import annotations
-
-import json
-from pathlib import Path
-
-import joblib
-
-
-def save_model(model, path: str, metadata=None) -> None:
-    path_obj = Path(path)
-    path_obj.parent.mkdir(parents=True, exist_ok=True)
-    joblib.dump(model, path_obj)
-    meta_path = path_obj.with_suffix(path_obj.suffix + ".meta.json")
-    meta_path.write_text(json.dumps(metadata or {}, ensure_ascii=False, indent=2), encoding="utf-8")
-
-
-def load_model(path: str, return_metadata: bool = False):
-    path_obj = Path(path)
-    model = joblib.load(path_obj)
-    meta_path = path_obj.with_suffix(path_obj.suffix + ".meta.json")
-    metadata = json.loads(meta_path.read_text(encoding="utf-8")) if meta_path.exists() else {}
-    if return_metadata:
-        return model, metadata
-    return model
diff --git a/demo5_full_project/app/ml/tuner.py b/demo5_full_project/app/ml/tuner.py
deleted file mode 100644
index 3a71c66..0000000
--- a/demo5_full_project/app/ml/tuner.py
+++ /dev/null
@@ -1,26 +0,0 @@
-from __future__ import annotations
-
-from sklearn.ensemble import RandomForestClassifier
-from sklearn.model_selection import cross_val_score
-
-
-def tune_model(X_train, y_train, n_trials: int = 5):
-    candidate_depths = [2, 3, 4, 5, None]
-    candidate_estimators = [20, 50, 100, 150, 200]
-    best_model = None
-    best_score = -1.0
-
-    for idx in range(max(1, n_trials)):
-        model = RandomForestClassifier(
-            n_estimators=candidate_estimators[idx % len(candidate_estimators)],
-            max_depth=candidate_depths[idx % len(candidate_depths)],
-            random_state=42 + idx,
-        )
-        score = float(cross_val_score(model, X_train, y_train, cv=3).mean())
-        if score > best_score:
-            best_score = score
-            best_model = model
-
-    assert best_model is not None
-    best_model.fit(X_train, y_train)
-    return best_model, best_score
diff --git a/demo5_full_project/main.py b/demo5_full_project/main.py
deleted file mode 100644
index 1918571..0000000
--- a/demo5_full_project/main.py
+++ /dev/null
@@ -1,75 +0,0 @@
-from __future__ import annotations
-
-import json
-from typing import Iterator, List
-from uuid import uuid4
-
-from fastapi import FastAPI
-from fastapi.responses import StreamingResponse
-from pydantic import BaseModel
-
-
-app = FastAPI(title="Demo5 Full Project")
-TASKS = {}
-
-
-class ChatRequest(BaseModel):
-    message: str
-
-
-class TaskCreateRequest(BaseModel):
-    title: str
-    priority: str
-    due_date: str
-
-
-@app.get("/health")
-def health():
-    return {"status": "ok", "version": "1.0.0"}
-
-
-@app.post("/chat")
-def chat(payload: ChatRequest):
-    message = payload.message.strip()
-    tools_used: List[str] = []
-    if "任务" in message and any(token in message for token in ["列", "list", "所有"]):
-        tools_used.append("list_tasks")
-        reply = f"当前共有 {len(TASKS)} 个任务。"
-    else:
-        reply = "Demo5 已收到请求，支持任务、流式输出和检索模块演示。"
-    return {"reply": reply, "tools_used": tools_used}
-
-
-@app.get("/tasks")
-def get_tasks():
-    return {"tasks": list(TASKS.values())}
-
-
-@app.post("/tasks", status_code=201)
-def create_task(payload: TaskCreateRequest):
-    task = payload.model_dump()
-    task["id"] = str(uuid4())
-    TASKS[task["id"]] = task
-    return task
-
-
-@app.delete("/tasks/{task_id}")
-def delete_task(task_id: str):
-    if task_id not in TASKS:
-        from fastapi import HTTPException
-
-        raise HTTPException(status_code=404, detail="Task not found")
-    del TASKS[task_id]
-    return {"deleted": True}
-
-
-def _sse_event_stream(message: str) -> Iterator[str]:
-    for chunk in ["收到消息", "正在处理", message]:
-        payload = json.dumps({"type": "token", "content": chunk}, ensure_ascii=False)
-        yield f"data: {payload}\n\n"
-    yield "data: {\"type\": \"done\"}\n\n"
-
-
-@app.get("/stream")
-def stream(message: str):
-    return StreamingResponse(_sse_event_stream(message), media_type="text/event-stream")