diff --git a/app/service/agent_service.py b/app/service/agent_service.py
index 0e607f8..48b9f28 100644
--- a/app/service/agent_service.py
+++ b/app/service/agent_service.py
@@ -1 +1,158 @@
-"""Agent 编排：LangGraph StateGraph + DeepSeek；Tool 节点；四 Agent 能力边界。"""
+"""Agent 编排（T-07）：LangGraph StateGraph 骨架 + DeepSeek 对话。
+
+链路（FLOW §2 最小闭环）：compose（角色 system + 历史窗口转 messages）
+→ llm（DeepSeek）→ guard（合规护栏）。四 Agent 能力边界（FLOW §4 与
+MEMORY 禁止项）由 system prompt 固化：客户无投资建议/收益承诺/自动下单；
+代理人草稿不外发；分析只读；风控辅助不自动处置。
+
+LLM 未配置（DEEPSEEK_API_KEY 为空）时降级固定提示回复（演示可跑且明确
+标注非模型生成）；单测经 FakeLLM 注入，不依赖外网。
+
+阶段 C（开发计划 C1/C2）将在 risk 分支接入 intent→tool→respond 与
+chat_tools（agent_tool_call 落库随 Tool 节点一起）；本骨架保证图编排
+可测、消息落库与 trace 贯通由 T-06 api/chat 承担。
+"""
+
+from __future__ import annotations
+
+import threading
+from typing import Annotated, Any, TypedDict
+
+from langchain_core.messages import AIMessage, BaseMessage, HumanMessage, SystemMessage
+from langgraph.graph import END, START, StateGraph
+from operator import add
+
+from app.config.settings import settings
+
+# 客户/对外口径的固定免责声明（随回复文本尾部输出；G-08 阻断响应另有两要素）
+CHAT_DISCLAIMER = "以上内容由 AI 生成，仅供业务参考，不构成投资建议。"
+
+# LLM 未配置时的降级回复（明确标注非模型生成，演示链路不断）
+_DEGRADED_PREFIX = "（LLM 未配置：请在 .env 设置 DEEPSEEK_API_KEY 后重启）"
+
+# 四 Agent 能力边界 system prompt（FLOW §4；MEMORY 禁止项固化）
+_SYSTEM_PROMPTS: dict[str, str] = {
+    "customer": (
+        "你是客户财富智能管家。仅服务当前登录客户本人；不得提供投资建议、"
+        "收益承诺或自动下单；事实性数据须注明来源，无法回答时如实告知。"
+    ),
+    "advisor": (
+        "你是代理人助手，服务理财顾问。仅可讨论名下客户（归属经平台校验）；"
+        "生成的草稿内容不会自动外发给客户；引用数据须可溯源。"
+    ),
+    "analyst": (
+        "你是数据分析助手。仅做只读查询与聚合分析，不写任何画像/预警/业务数据。"
+    ),
+    "risk": (
+        "你是风控监测助手。可查询预警台账与客户监测信息，输出仅供参考，"
+        "最终处置须经风控专员通过处置接口人工完成；不得自动处置预警。"
+    ),
+}
+
+
+class ChatState(TypedDict):
+    """对话图状态（messages 由各节点以 add 归并，避免覆盖历史）。"""
+
+    agent_type: str
+    history: list[dict]
+    user_message: str
+    messages: Annotated[list[BaseMessage], add]
+    reply: str
+    has_disclaimer: bool
+
+
+def _compose_messages(state: ChatState) -> list[BaseMessage]:
+    """system（角色边界）+ 历史窗口（role→message）+ 本轮用户消息。"""
+    msgs: list[BaseMessage] = [
+        SystemMessage(content=_SYSTEM_PROMPTS.get(state["agent_type"], _SYSTEM_PROMPTS["customer"]))
+    ]
+    for m in state["history"]:
+        role = m.get("role")
+        content = m.get("content", "")
+        if role == "user":
+            msgs.append(HumanMessage(content=content))
+        elif role == "assistant":
+            msgs.append(AIMessage(content=content))
+    msgs.append(HumanMessage(content=state["user_message"]))
+    return msgs
+
+
+def llm_node(state: ChatState) -> dict[str, Any]:
+    """组装 messages 后调 LLM；未配置 key 时降级（不抛异常，演示链路不断）。"""
+    messages = _compose_messages(state)
+    if not settings.deepseek_api_key:
+        reply = f"{_DEGRADED_PREFIX}已收到您的消息：{state['user_message']}"
+        return {"messages": messages + [AIMessage(content=reply)], "reply": reply}
+    llm = _get_llm()
+    result = llm.invoke(messages)
+    return {"messages": messages + [result], "reply": result.content}
+
+
+def guard_node(state: ChatState) -> dict[str, Any]:
+    """合规护栏：对外角色（customer/risk，未知类型 fail-safe 按最严口径）附免责声明。"""
+    external = state["agent_type"] in ("customer", "risk") or state["agent_type"] not in _SYSTEM_PROMPTS
+    if external:
+        return {"reply": f"{state['reply']}\n\n{CHAT_DISCLAIMER}", "has_disclaimer": True}
+    return {"has_disclaimer": False}
+
+
+def build_graph():
+    """StateGraph 骨架：START → llm → guard → END（阶段 C 扩展 tool 节点）。"""
+    graph = StateGraph(ChatState)
+    graph.add_node("llm", llm_node)
+    graph.add_node("guard", guard_node)
+    graph.add_edge(START, "llm")
+    graph.add_edge("llm", "guard")
+    graph.add_edge("guard", END)
+    return graph.compile()
+
+
+_graph = None
+_graph_lock = threading.Lock()
+_llm: Any | None = None
+
+
+def _get_graph():
+    global _graph
+    with _graph_lock:
+        if _graph is None:
+            _graph = build_graph()
+        return _graph
+
+
+def _get_llm() -> Any:
+    """DeepSeek 经 langchain-openai 兼容接口（懒构造；测试注入 _llm）。"""
+    global _llm
+    if _llm is None:
+        from langchain_openai import ChatOpenAI
+
+        _llm = ChatOpenAI(
+            model="deepseek-chat",
+            api_key=settings.deepseek_api_key,
+            base_url=settings.deepseek_base_url,
+            temperature=0.3,
+        )
+    return _llm
+
+
+def chat(agent_type: str, history: list[dict], user_message: str) -> dict[str, Any]:
+    """对话入口（T-06 api/chat 调用）：返回 {reply, has_disclaimer}。"""
+    final = _get_graph().invoke(
+        {
+            "agent_type": agent_type,
+            "history": history,
+            "user_message": user_message,
+            "messages": [],
+            "reply": "",
+            "has_disclaimer": False,
+        }
+    )
+    return {"reply": final["reply"], "has_disclaimer": final["has_disclaimer"]}
+
+
+def reset_cache() -> None:
+    """测试隔离出口：清空图与 LLM 单例缓存。"""
+    global _graph, _llm
+    with _graph_lock:
+        _graph = None
+        _llm = None
diff --git a/tests/test_agent_service.py b/tests/test_agent_service.py
new file mode 100644
index 0000000..da40f1c
--- /dev/null
+++ b/tests/test_agent_service.py
@@ -0,0 +1,92 @@
+"""T-07 agent_service StateGraph 骨架（mock LLM，不依赖外网/真 key）。
+
+FakeLLM 捕获收到的 messages 供断言（system 边界 + 历史窗口 + 本轮消息）；
+降级路径（无 key）与免责声明护栏单独覆盖。图/LLM 缓存经 reset_cache 隔离。
+"""
+
+from __future__ import annotations
+
+import pytest
+from langchain_core.messages import AIMessage
+
+from app.config.settings import settings
+from app.service import agent_service
+
+
+class FakeLLM:
+    def __init__(self, reply: str = "模拟回复"):
+        self.reply = reply
+        self.calls: list[list] = []
+
+    def invoke(self, messages):
+        self.calls.append(list(messages))
+        return AIMessage(content=self.reply)
+
+
+@pytest.fixture()
+def fake_llm(monkeypatch):
+    llm = FakeLLM()
+    monkeypatch.setattr(agent_service, "_llm", llm)
+    monkeypatch.setattr(settings, "deepseek_api_key", "test-key")
+    yield llm
+    agent_service.reset_cache()
+
+
+HISTORY = [
+    {"role": "user", "content": "你好"},
+    {"role": "assistant", "content": "您好，有什么可以帮您？"},
+]
+
+
+def test_chat_customer_appends_disclaimer(fake_llm):
+    out = agent_service.chat("customer", HISTORY, "查一下我的持仓")
+    assert out["has_disclaimer"] is True
+    assert out["reply"].endswith(agent_service.CHAT_DISCLAIMER)
+    assert "模拟回复" in out["reply"]
+
+
+def test_chat_risk_also_disclaimer(fake_llm):
+    out = agent_service.chat("risk", [], "今天有多少待审预警")
+    assert out["has_disclaimer"] is True
+
+
+def test_chat_advisor_no_disclaimer(fake_llm):
+    out = agent_service.chat("advisor", [], "总结客户需求")
+    assert out["has_disclaimer"] is False
+    assert agent_service.CHAT_DISCLAIMER not in out["reply"]
+
+
+def test_messages_compose_system_history_and_user(fake_llm):
+    agent_service.chat("customer", HISTORY, "最新问题")
+    msgs = fake_llm.calls[0]
+    assert msgs[0].__class__.__name__ == "SystemMessage"
+    assert "不得提供投资建议" in msgs[0].content  # customer 边界注入
+    assert [m.__class__.__name__ for m in msgs[1:]] == ["HumanMessage", "AIMessage", "HumanMessage"]
+    assert msgs[-1].content == "最新问题"
+
+
+def test_role_boundary_per_agent_type(fake_llm):
+    agent_service.chat("analyst", [], "聚合一下")
+    system = fake_llm.calls[0][0]
+    assert "只读" in system.content
+
+
+def test_degraded_reply_without_api_key(monkeypatch):
+    """key 缺失 → 降级固定提示（不抛异常，演示链路不断）。"""
+    monkeypatch.setattr(settings, "deepseek_api_key", "")
+    agent_service.reset_cache()
+    out = agent_service.chat("customer", [], "在吗")
+    assert "LLM 未配置" in out["reply"] and "在吗" in out["reply"]
+    assert out["has_disclaimer"] is True  # 降级回复同样经 guard 护栏
+
+
+def test_graph_skeleton_nodes():
+    graph = agent_service.build_graph()
+    assert {"llm", "guard"} <= set(graph.nodes)  # __start__ 为 LangGraph 内部节点
+
+
+def test_unknown_agent_type_falls_back_customer_boundary(fake_llm):
+    """未知 agent_type 兜底 customer 边界（fail-safe：最严声明口径）。"""
+    out = agent_service.chat("unknown", [], "hi")
+    assert out["has_disclaimer"] is True
+    assert "不得提供投资建议" in fake_llm.calls[0][0].content