diff --git a/ai-coding-assistant/packages/core/ai_coding_assistant/__init__.py b/ai-coding-assistant/packages/core/ai_coding_assistant/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/ai-coding-assistant/packages/core/ai_coding_assistant/agent/__init__.py b/ai-coding-assistant/packages/core/ai_coding_assistant/agent/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/ai-coding-assistant/packages/core/ai_coding_assistant/agent/messages.py b/ai-coding-assistant/packages/core/ai_coding_assistant/agent/messages.py
new file mode 100644
index 0000000..e671578
--- /dev/null
+++ b/ai-coding-assistant/packages/core/ai_coding_assistant/agent/messages.py
@@ -0,0 +1,121 @@
+"""Agent-internal message types.
+
+All message types share a common `to_dict()` / `from_dict()` interface for
+serialization.  `AgentMessage.from_dict()` is a factory that dispatches to the
+correct concrete class based on the ``role`` field.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from typing import Any
+
+
+@dataclass(slots=True)
+class ToolCall:
+    """A single tool call requested by the assistant."""
+
+    id: str
+    name: str
+    arguments: dict[str, Any] = field(default_factory=dict)
+
+    def to_dict(self) -> dict[str, Any]:
+        return {"id": self.id, "name": self.name, "arguments": self.arguments}
+
+    @classmethod
+    def from_dict(cls, data: dict[str, Any]) -> ToolCall:
+        return cls(
+            id=data["id"],
+            name=data["name"],
+            arguments=data.get("arguments", {}),
+        )
+
+
+@dataclass(slots=True)
+class SystemMessage:
+    role: str = field(default="system", init=False)
+    content: str
+
+    def to_dict(self) -> dict[str, Any]:
+        return {"role": self.role, "content": self.content}
+
+    @classmethod
+    def from_dict(cls, data: dict[str, Any]) -> SystemMessage:
+        return cls(content=data["content"])
+
+
+@dataclass(slots=True)
+class UserMessage:
+    role: str = field(default="user", init=False)
+    content: str
+
+    def to_dict(self) -> dict[str, Any]:
+        return {"role": self.role, "content": self.content}
+
+    @classmethod
+    def from_dict(cls, data: dict[str, Any]) -> UserMessage:
+        return cls(content=data["content"])
+
+
+@dataclass(slots=True)
+class AssistantMessage:
+    role: str = field(default="assistant", init=False)
+    content: str
+    tool_calls: list[ToolCall] = field(default_factory=list)
+
+    def to_dict(self) -> dict[str, Any]:
+        return {
+            "role": self.role,
+            "content": self.content,
+            "tool_calls": [tc.to_dict() for tc in self.tool_calls],
+        }
+
+    @classmethod
+    def from_dict(cls, data: dict[str, Any]) -> AssistantMessage:
+        tool_calls = [
+            ToolCall.from_dict(tc) for tc in data.get("tool_calls", [])
+        ]
+        return cls(content=data.get("content", ""), tool_calls=tool_calls)
+
+
+@dataclass(slots=True)
+class ToolResultMessage:
+    role: str = field(default="tool", init=False)
+    tool_call_id: str
+    content: str
+    is_error: bool = False
+
+    def to_dict(self) -> dict[str, Any]:
+        return {
+            "role": self.role,
+            "tool_call_id": self.tool_call_id,
+            "content": self.content,
+            "is_error": self.is_error,
+        }
+
+    @classmethod
+    def from_dict(cls, data: dict[str, Any]) -> ToolResultMessage:
+        return cls(
+            tool_call_id=data["tool_call_id"],
+            content=data["content"],
+            is_error=data.get("is_error", False),
+        )
+
+
+# Union type for type annotations
+AgentMessage = SystemMessage | UserMessage | AssistantMessage | ToolResultMessage
+
+
+def agent_message_from_dict(data: dict[str, Any]) -> AgentMessage:
+    """Factory: create the correct AgentMessage subclass from a dict."""
+    role = data.get("role")
+    dispatch: dict[str, type[AgentMessage]] = {
+        "system": SystemMessage,
+        "user": UserMessage,
+        "assistant": AssistantMessage,
+        "tool": ToolResultMessage,
+    }
+    cls = dispatch.get(role)
+    if cls is None:
+        raise ValueError(f"unknown role: {role!r}")
+    return cls.from_dict(data)
diff --git a/ai-coding-assistant/packages/core/ai_coding_assistant/context/__init__.py b/ai-coding-assistant/packages/core/ai_coding_assistant/context/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/ai-coding-assistant/packages/core/ai_coding_assistant/llm/__init__.py b/ai-coding-assistant/packages/core/ai_coding_assistant/llm/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/ai-coding-assistant/packages/core/ai_coding_assistant/llm/base.py b/ai-coding-assistant/packages/core/ai_coding_assistant/llm/base.py
new file mode 100644
index 0000000..0d2717f
--- /dev/null
+++ b/ai-coding-assistant/packages/core/ai_coding_assistant/llm/base.py
@@ -0,0 +1,109 @@
+"""LLM Provider abstraction — types and ABC.
+
+Defines the contract that every LLM provider adapter must implement,
+plus shared data types for messages, tool definitions, and stream chunks.
+"""
+
+from __future__ import annotations
+
+from abc import ABC, abstractmethod
+from collections.abc import AsyncIterator
+from dataclasses import dataclass, field
+from typing import Any
+
+
+# ── LLM API message format ───────────────────────────────────
+
+
+@dataclass(slots=True)
+class Message:
+    """A message in the LLM provider's native API format."""
+
+    role: str  # "system" | "user" | "assistant" | "tool"
+    content: str | None = None
+    tool_calls: list[dict[str, Any]] | None = None
+    tool_call_id: str | None = None
+    name: str | None = None
+
+
+# ── Tool definition types ────────────────────────────────────
+
+
+@dataclass(slots=True)
+class ToolParameter:
+    """Describes a single parameter of a tool."""
+
+    type: str
+    description: str = ""
+    required: bool = True
+    enum: list[str] | None = None
+
+
+@dataclass(slots=True)
+class ToolDef:
+    """Tool definition passed to LLM providers."""
+
+    name: str
+    description: str
+    parameters: dict[str, ToolParameter] = field(default_factory=dict)
+
+    def to_openai_format(self) -> dict[str, Any]:
+        """Convert to OpenAI function-calling tool format."""
+        properties: dict[str, Any] = {}
+        required: list[str] = []
+        for pname, param in self.parameters.items():
+            prop: dict[str, Any] = {"type": param.type, "description": param.description}
+            if param.enum is not None:
+                prop["enum"] = param.enum
+            properties[pname] = prop
+            if param.required:
+                required.append(pname)
+
+        return {
+            "type": "function",
+            "function": {
+                "name": self.name,
+                "description": self.description,
+                "parameters": {
+                    "type": "object",
+                    "properties": properties,
+                    "required": required,
+                },
+            },
+        }
+
+
+# ── Stream chunk ─────────────────────────────────────────────
+
+
+@dataclass(slots=True)
+class StreamChunk:
+    """A single chunk emitted during LLM streaming."""
+
+    type: str  # "content" | "tool_call" | "done" | "error"
+    content: str | None = None
+    tool_call: dict[str, Any] | None = None
+
+
+# ── LLM Provider ABC ─────────────────────────────────────────
+
+
+class LLMProvider(ABC):
+    """Abstract base class for LLM provider adapters."""
+
+    @abstractmethod
+    async def stream_chat(
+        self,
+        messages: list[Message],
+        tools: list[ToolDef] | None = None,
+        **kwargs: Any,
+    ) -> AsyncIterator[StreamChunk]:
+        """Stream a chat completion from the LLM."""
+
+    @abstractmethod
+    def convert_messages(self, agent_messages: list[Any]) -> list[Message]:
+        """Convert agent-internal messages to the provider's Message format."""
+
+    @abstractmethod
+    def convert_tools(self, tools: list[ToolDef]) -> list[dict[str, Any]]:
+        """Convert ToolDef list to the provider's native tool format."""
diff --git a/ai-coding-assistant/packages/core/ai_coding_assistant/skills/__init__.py b/ai-coding-assistant/packages/core/ai_coding_assistant/skills/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/ai-coding-assistant/packages/core/ai_coding_assistant/tools/__init__.py b/ai-coding-assistant/packages/core/ai_coding_assistant/tools/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/ai-coding-assistant/packages/core/ai_coding_assistant/tools/builtin/__init__.py b/ai-coding-assistant/packages/core/ai_coding_assistant/tools/builtin/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/ai-coding-assistant/packages/core/ai_coding_assistant/transport/__init__.py b/ai-coding-assistant/packages/core/ai_coding_assistant/transport/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/ai-coding-assistant/pyproject.toml b/ai-coding-assistant/pyproject.toml
new file mode 100644
index 0000000..eaa1585
--- /dev/null
+++ b/ai-coding-assistant/pyproject.toml
@@ -0,0 +1,30 @@
+[build-system]
+requires = ["hatchling"]
+build-backend = "hatchling.build"
+
+[project]
+name = "ai-coding-assistant"
+version = "0.1.0"
+description = "An extensible AI coding assistant with multi-LLM support"
+requires-python = ">=3.11"
+dependencies = [
+    "httpx>=0.27",
+    "pydantic>=2.0",
+]
+
+[project.optional-dependencies]
+dev = [
+    "pytest>=8.0",
+    "pytest-asyncio>=0.23",
+    "respx>=0.21",
+]
+
+[project.scripts]
+ai-ca = "ai_coding_assistant.cli:main"
+
+[tool.hatch.build.targets.wheel]
+packages = ["packages/core/ai_coding_assistant"]
+
+[tool.pytest.ini_options]
+asyncio_mode = "auto"
+testpaths = ["tests"]
diff --git a/ai-coding-assistant/tests/__init__.py b/ai-coding-assistant/tests/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/ai-coding-assistant/tests/core/__init__.py b/ai-coding-assistant/tests/core/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/ai-coding-assistant/tests/core/agent/__init__.py b/ai-coding-assistant/tests/core/agent/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/ai-coding-assistant/tests/core/agent/test_messages.py b/ai-coding-assistant/tests/core/agent/test_messages.py
new file mode 100644
index 0000000..dbce9b7
--- /dev/null
+++ b/ai-coding-assistant/tests/core/agent/test_messages.py
@@ -0,0 +1,156 @@
+"""Tests for agent message models — RED phase."""
+
+from __future__ import annotations
+
+import json
+
+import pytest
+
+from ai_coding_assistant.agent.messages import (
+    AgentMessage,
+    AssistantMessage,
+    SystemMessage,
+    ToolCall,
+    ToolResultMessage,
+    UserMessage,
+    agent_message_from_dict,
+)
+
+
+# ── Construction ──────────────────────────────────────────────
+
+class TestUserMessage:
+    def test_create(self):
+        msg = UserMessage(content="hello")
+        assert msg.role == "user"
+        assert msg.content == "hello"
+
+    def test_to_dict(self):
+        msg = UserMessage(content="hello")
+        d = msg.to_dict()
+        assert d == {"role": "user", "content": "hello"}
+
+    def test_from_dict(self):
+        msg = UserMessage.from_dict({"role": "user", "content": "hello"})
+        assert msg.role == "user"
+        assert msg.content == "hello"
+
+
+class TestSystemMessage:
+    def test_create(self):
+        msg = SystemMessage(content="you are helpful")
+        assert msg.role == "system"
+        assert msg.content == "you are helpful"
+
+    def test_to_dict_roundtrip(self):
+        original = SystemMessage(content="you are helpful")
+        restored = SystemMessage.from_dict(original.to_dict())
+        assert restored.content == original.content
+
+
+class TestAssistantMessage:
+    def test_create_text_only(self):
+        msg = AssistantMessage(content="hi there")
+        assert msg.role == "assistant"
+        assert msg.content == "hi there"
+        assert msg.tool_calls == []
+
+    def test_create_with_tool_calls(self):
+        tc = ToolCall(id="call_1", name="echo", arguments={"text": "hello"})
+        msg = AssistantMessage(content="", tool_calls=[tc])
+        assert len(msg.tool_calls) == 1
+        assert msg.tool_calls[0].name == "echo"
+
+    def test_to_dict_with_tool_calls(self):
+        tc = ToolCall(id="call_1", name="echo", arguments={"text": "hello"})
+        msg = AssistantMessage(content="", tool_calls=[tc])
+        d = msg.to_dict()
+        assert d["role"] == "assistant"
+        assert len(d["tool_calls"]) == 1
+        assert d["tool_calls"][0]["id"] == "call_1"
+
+    def test_from_dict_with_tool_calls(self):
+        data = {
+            "role": "assistant",
+            "content": "",
+            "tool_calls": [
+                {"id": "call_1", "name": "echo", "arguments": {"text": "hello"}}
+            ],
+        }
+        msg = AssistantMessage.from_dict(data)
+        assert msg.role == "assistant"
+        assert len(msg.tool_calls) == 1
+        assert msg.tool_calls[0].name == "echo"
+        assert msg.tool_calls[0].arguments == {"text": "hello"}
+
+    def test_roundtrip(self):
+        tc = ToolCall(id="call_1", name="echo", arguments={"text": "hello"})
+        original = AssistantMessage(content="thinking...", tool_calls=[tc])
+        restored = AssistantMessage.from_dict(original.to_dict())
+        assert restored.content == original.content
+        assert len(restored.tool_calls) == 1
+        assert restored.tool_calls[0].id == tc.id
+
+
+class TestToolResultMessage:
+    def test_create(self):
+        msg = ToolResultMessage(
+            tool_call_id="call_1", content="result text", is_error=False
+        )
+        assert msg.role == "tool"
+        assert msg.tool_call_id == "call_1"
+        assert msg.is_error is False
+
+    def test_create_error(self):
+        msg = ToolResultMessage(
+            tool_call_id="call_2", content="something failed", is_error=True
+        )
+        assert msg.is_error is True
+
+    def test_to_dict_roundtrip(self):
+        original = ToolResultMessage(
+            tool_call_id="call_1", content="result text", is_error=False
+        )
+        restored = ToolResultMessage.from_dict(original.to_dict())
+        assert restored.tool_call_id == original.tool_call_id
+        assert restored.content == original.content
+        assert restored.is_error == original.is_error
+
+
+# ── AgentMessage union helpers ────────────────────────────────
+
+class TestAgentMessageUnion:
+    def test_from_dict_user(self):
+        msg = agent_message_from_dict({"role": "user", "content": "hi"})
+        assert isinstance(msg, UserMessage)
+
+    def test_from_dict_system(self):
+        msg = agent_message_from_dict({"role": "system", "content": "sys"})
+        assert isinstance(msg, SystemMessage)
+
+    def test_from_dict_assistant(self):
+        msg = agent_message_from_dict({"role": "assistant", "content": "hi"})
+        assert isinstance(msg, AssistantMessage)
+
+    def test_from_dict_tool(self):
+        msg = agent_message_from_dict(
+            {"role": "tool", "tool_call_id": "c1", "content": "ok", "is_error": False}
+        )
+        assert isinstance(msg, ToolResultMessage)
+
+    def test_from_dict_unknown_role_raises(self):
+        with pytest.raises(ValueError, match="unknown role"):
+            agent_message_from_dict({"role": "alien", "content": "???"})
+
+    def test_json_serializable(self):
+        """All message types should be JSON-serializable via to_dict()."""
+        tc = ToolCall(id="call_1", name="echo", arguments={"text": "hi"})
+        msgs: list[AgentMessage] = [
+            SystemMessage(content="system"),
+            UserMessage(content="hello"),
+            AssistantMessage(content="", tool_calls=[tc]),
+            ToolResultMessage(tool_call_id="call_1", content="hi", is_error=False),
+        ]
+        for msg in msgs:
+            json_str = json.dumps(msg.to_dict())
+            assert json_str  # non-empty
diff --git a/ai-coding-assistant/tests/core/context/__init__.py b/ai-coding-assistant/tests/core/context/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/ai-coding-assistant/tests/core/llm/__init__.py b/ai-coding-assistant/tests/core/llm/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/ai-coding-assistant/tests/core/llm/test_base.py b/ai-coding-assistant/tests/core/llm/test_base.py
new file mode 100644
index 0000000..71b20a1
--- /dev/null
+++ b/ai-coding-assistant/tests/core/llm/test_base.py
@@ -0,0 +1,147 @@
+"""Tests for LLM Provider abstraction — RED phase."""
+
+from __future__ import annotations
+
+import pytest
+
+from ai_coding_assistant.llm.base import (
+    LLMProvider,
+    Message,
+    StreamChunk,
+    ToolDef,
+    ToolParameter,
+)
+
+
+# ── Data model construction ──────────────────────────────────
+
+class TestMessage:
+    def test_user_message(self):
+        msg = Message(role="user", content="hello")
+        assert msg.role == "user"
+        assert msg.content == "hello"
+        assert msg.tool_calls is None
+
+    def test_assistant_with_tool_calls(self):
+        msg = Message(
+            role="assistant",
+            content="",
+            tool_calls=[{"id": "c1", "name": "echo", "arguments": {"text": "hi"}}],
+        )
+        assert len(msg.tool_calls) == 1
+        assert msg.tool_calls[0]["name"] == "echo"
+
+    def test_system_message(self):
+        msg = Message(role="system", content="you are helpful")
+        assert msg.role == "system"
+
+    def test_tool_message(self):
+        msg = Message(
+            role="tool",
+            content="result",
+            tool_call_id="c1",
+        )
+        assert msg.role == "tool"
+        assert msg.tool_call_id == "c1"
+
+
+class TestToolParameter:
+    def test_create(self):
+        param = ToolParameter(
+            type="string",
+            description="text to echo",
+        )
+        assert param.type == "string"
+
+    def test_with_enum(self):
+        param = ToolParameter(
+            type="string",
+            description="mode",
+            enum=["fast", "slow"],
+        )
+        assert param.enum == ["fast", "slow"]
+
+
+class TestToolDef:
+    def test_create(self):
+        td = ToolDef(
+            name="echo",
+            description="Echo back the input",
+            parameters={
+                "text": ToolParameter(type="string", description="text to echo"),
+            },
+        )
+        assert td.name == "echo"
+        assert "text" in td.parameters
+
+    def test_to_openai_format(self):
+        td = ToolDef(
+            name="echo",
+            description="Echo back the input",
+            parameters={
+                "text": ToolParameter(type="string", description="text to echo"),
+            },
+        )
+        result = td.to_openai_format()
+        assert result["type"] == "function"
+        assert result["function"]["name"] == "echo"
+        assert "text" in result["function"]["parameters"]["properties"]
+
+
+class TestStreamChunk:
+    def test_content_chunk(self):
+        chunk = StreamChunk(type="content", content="Hello")
+        assert chunk.type == "content"
+        assert chunk.content == "Hello"
+        assert chunk.tool_call is None
+
+    def test_tool_call_chunk(self):
+        chunk = StreamChunk(
+            type="tool_call",
+            tool_call={"id": "c1", "name": "echo", "arguments": "{}"},
+        )
+        assert chunk.type == "tool_call"
+        assert chunk.tool_call["name"] == "echo"
+
+    def test_done_chunk(self):
+        chunk = StreamChunk(type="done")
+        assert chunk.type == "done"
+
+    def test_error_chunk(self):
+        chunk = StreamChunk(type="error", content="rate limited")
+        assert chunk.type == "error"
+        assert chunk.content == "rate limited"
+
+
+# ── LLMProvider ABC ──────────────────────────────────────────
+
+class TestLLMProviderABC:
+    def test_cannot_instantiate(self):
+        """LLMProvider is abstract and cannot be instantiated directly."""
+        with pytest.raises(TypeError):
+            LLMProvider()  # type: ignore[abstract]
+
+    def test_subclass_must_implement_methods(self):
+        """A subclass that doesn't implement all abstract methods cannot be instantiated."""
+
+        class IncompleteProvider(LLMProvider):
+            pass
+
+        with pytest.raises(TypeError):
+            IncompleteProvider()  # type: ignore[abstract]
+
+    def test_subclass_can_be_created_with_all_methods(self):
+        """A subclass that implements all abstract methods can be instantiated."""
+
+        class DummyProvider(LLMProvider):
+            async def stream_chat(self, messages, tools=None, **kwargs):
+                yield StreamChunk(type="done")
+
+            def convert_messages(self, agent_messages):
+                return []
+
+            def convert_tools(self, tools):
+                return []
+
+        provider = DummyProvider()
+        assert isinstance(provider, LLMProvider)
diff --git a/ai-coding-assistant/tests/core/skills/__init__.py b/ai-coding-assistant/tests/core/skills/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/ai-coding-assistant/tests/core/tools/__init__.py b/ai-coding-assistant/tests/core/tools/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/ai-coding-assistant/tests/core/transport/__init__.py b/ai-coding-assistant/tests/core/transport/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/memory/2026-05-13.md b/memory/2026-05-13.md
new file mode 100644
index 0000000..f2c07fb
--- /dev/null
+++ b/memory/2026-05-13.md
@@ -0,0 +1,5 @@
+# 2026-05-13
+
+## 纪要
+
+- 工作区已初始化，可在此持续记录当天上下文。