diff --git a/CLAUDE.md b/CLAUDE.md
new file mode 100644
index 0000000..3566beb
--- /dev/null
+++ b/CLAUDE.md
@@ -0,0 +1,50 @@
+# PhoneWork — Development Guide
+
+## Quick Reference
+
+| Need | Where to look |
+|------|---------------|
+| Project architecture, deployment, bot commands | `README.md` |
+| Claude Agent SDK usage patterns and tested examples | `docs/claude/` |
+| SDK migration plan (subprocess → ClaudeSDKClient) | `.claude/plans/toasty-pondering-nova.md` |
+| Feishu card / markdown formatting | `docs/feishu/` |
+
+## Claude Agent SDK
+
+When writing code that uses `claude-agent-sdk`, **first read `docs/claude/`**:
+
+- `_sdk_test_common.py` — .env auth loading pattern (`setup_auth()`, `auth_env()`, `make_tmpdir()`)
+- `test_query_read_edit.py` — `query()` one-shot: Read + Edit with `allowed_tools`
+- `test_client_write_resume.py` — `ClaudeSDKClient`: Write + session resume via `resume=session_id`
+- `test_hooks_audit_deny.py` — Hooks: `PostToolUse` audit + `PreToolUse` deny
+
+### Key rules (verified by testing)
+
+- Use `allowed_tools` for permission auto-approval. **Do not pass custom lists to `tools=`** — it causes CLI exit code 1.
+- Auth: set `ANTHROPIC_BASE_URL` + `ANTHROPIC_AUTH_TOKEN` in `.env`; pass via `ClaudeAgentOptions(env=auth_env())`. Clear `ANTHROPIC_API_KEY` to avoid auth conflicts.
+- SDK does not strip ANSI escape codes from `ResultMessage.result` — handle in application layer if needed.
+- On Windows, use manual `make_tmpdir()` / `remove_tmpdir()` instead of `tempfile.TemporaryDirectory()` context manager to avoid cleanup races with CLI subprocesses.
+
+## Project Structure
+
+```
+bot/           Feishu event handling, commands, message sending
+orchestrator/  LangChain agent + tools (session management, shell, files, web)
+agent/         Claude Code runner, session manager, task runner, scheduler, audit
+router/        Multi-host routing (public VPS side)
+host_client/   Host client (behind NAT, connects to router)
+shared/        Wire protocol for router ↔ host communication
+docs/claude/   Claude Agent SDK examples and reference tests
+docs/feishu/   Feishu API reference
+```
+
+## Running Tests
+
+```bash
+cd docs/claude
+../../.venv/Scripts/python test_query_read_edit.py
+../../.venv/Scripts/python test_client_write_resume.py
+../../.venv/Scripts/python test_hooks_audit_deny.py
+```
+
+Requires `.env` at project root with `ANTHROPIC_BASE_URL` and `ANTHROPIC_AUTH_TOKEN`.
diff --git a/docs/claude/_sdk_test_common.py b/docs/claude/_sdk_test_common.py
new file mode 100644
index 0000000..d2dc6db
--- /dev/null
+++ b/docs/claude/_sdk_test_common.py
@@ -0,0 +1,75 @@
+"""Shared helpers for Claude Agent SDK tests.
+
+Loads ANTHROPIC_BASE_URL / ANTHROPIC_AUTH_TOKEN from the project root .env
+and injects them into the process environment and ClaudeAgentOptions.env.
+"""
+
+from __future__ import annotations
+
+import os
+import shutil
+import tempfile
+from pathlib import Path
+
+
+PROJECT_ROOT = Path(__file__).resolve().parents[2]
+DOTENV_PATH = PROJECT_ROOT / ".env"
+
+
+def load_dotenv(dotenv_path: Path = DOTENV_PATH) -> dict[str, str]:
+    """Parse KEY=VALUE lines from a .env file (no third-party deps)."""
+    values: dict[str, str] = {}
+    if not dotenv_path.exists():
+        return values
+    for raw_line in dotenv_path.read_text(encoding="utf-8").splitlines():
+        line = raw_line.strip()
+        if not line or line.startswith("#") or "=" not in line:
+            continue
+        key, value = line.split("=", 1)
+        values[key.strip()] = value.strip().strip('"').strip("'")
+    return values
+
+
+def setup_auth() -> tuple[bool, str]:
+    """Inject .env auth vars into os.environ; clear ANTHROPIC_API_KEY to avoid conflicts."""
+    values = load_dotenv()
+    base_url = values.get("ANTHROPIC_BASE_URL")
+    auth_token = values.get("ANTHROPIC_AUTH_TOKEN")
+    oauth_token = values.get("CLAUDE_CODE_OAUTH_TOKEN")
+
+    if not base_url:
+        return False, "Missing ANTHROPIC_BASE_URL in .env"
+    if not auth_token and not oauth_token:
+        return False, "Missing ANTHROPIC_AUTH_TOKEN or CLAUDE_CODE_OAUTH_TOKEN in .env"
+
+    os.environ["ANTHROPIC_BASE_URL"] = base_url
+    if auth_token:
+        os.environ["ANTHROPIC_AUTH_TOKEN"] = auth_token
+    if oauth_token:
+        os.environ["CLAUDE_CODE_OAUTH_TOKEN"] = oauth_token
+    os.environ.pop("ANTHROPIC_API_KEY", None)
+    return True, f"Auth loaded from {DOTENV_PATH}"
+
+
+def auth_env() -> dict[str, str]:
+    """Return env dict suitable for ClaudeAgentOptions(env=...)."""
+    return {
+        "ANTHROPIC_BASE_URL": os.environ["ANTHROPIC_BASE_URL"],
+        "ANTHROPIC_AUTH_TOKEN": os.environ.get("ANTHROPIC_AUTH_TOKEN", ""),
+        "CLAUDE_CODE_OAUTH_TOKEN": os.environ.get("CLAUDE_CODE_OAUTH_TOKEN", ""),
+    }
+
+
+def make_tmpdir(prefix: str = "sdk-test-") -> Path:
+    """Create a temporary directory manually (caller is responsible for cleanup)."""
+    return Path(tempfile.mkdtemp(prefix=prefix))
+
+
+def remove_tmpdir(path: Path) -> None:
+    """Remove a temporary directory, retrying once on Windows lock errors."""
+    try:
+        shutil.rmtree(path)
+    except OSError:
+        import time
+        time.sleep(1)
+        shutil.rmtree(path, ignore_errors=True)
diff --git a/docs/claude-agent-sdk.md b/docs/claude/claude-agent-sdk.md
similarity index 100%
rename from docs/claude-agent-sdk.md
rename to docs/claude/claude-agent-sdk.md
diff --git a/docs/claude/test_client_write_resume.py b/docs/claude/test_client_write_resume.py
new file mode 100644
index 0000000..e422145
--- /dev/null
+++ b/docs/claude/test_client_write_resume.py
@@ -0,0 +1,116 @@
+"""Example: ClaudeSDKClient — Write file + resume session.
+
+Demonstrates:
+- ClaudeSDKClient (stateful, interactive mode)
+- Creating a file via Write tool
+- Capturing session_id from SystemMessage
+- Resuming a session with options.resume
+- receive_response() for streaming messages
+- Manual tmpdir management
+
+Usage:
+    cd docs/claude
+    ../../.venv/Scripts/python test_client_write_resume.py
+"""
+
+from __future__ import annotations
+
+import asyncio
+from pathlib import Path
+
+from _sdk_test_common import auth_env, make_tmpdir, remove_tmpdir, setup_auth
+
+
+async def drain(client):
+    """Collect text from receive_response(), return (session_id, joined_text)."""
+    from claude_agent_sdk import AssistantMessage, ResultMessage, SystemMessage, TextBlock
+
+    texts: list[str] = []
+    session_id = None
+    async for msg in client.receive_response():
+        if isinstance(msg, SystemMessage) and msg.subtype == "init":
+            session_id = msg.data.get("session_id")
+        elif isinstance(msg, AssistantMessage):
+            for block in msg.content:
+                if isinstance(block, TextBlock):
+                    texts.append(block.text)
+        elif isinstance(msg, ResultMessage):
+            session_id = session_id or msg.session_id
+            if msg.result:
+                texts.append(msg.result)
+    return session_id, "\n".join(texts)
+
+
+async def main() -> int:
+    ok, msg = setup_auth()
+    print(msg)
+    if not ok:
+        return 1
+
+    from claude_agent_sdk import ClaudeAgentOptions, ClaudeSDKClient
+
+    tmpdir = make_tmpdir("write-resume-")
+    try:
+        target = tmpdir / "session_test.txt"
+
+        # --- Turn 1: create file ---
+        print("--- Turn 1: Write ---")
+        async with ClaudeSDKClient(
+            ClaudeAgentOptions(
+                cwd=str(tmpdir),
+                allowed_tools=["Write", "Read"],
+                permission_mode="acceptEdits",
+                max_turns=3,
+                env=auth_env(),
+            )
+        ) as client:
+            await client.query(
+                f"Use the Write tool to create session_test.txt "
+                f"in {tmpdir} with exactly the content 'Session 1'."
+            )
+            session_id, response = await drain(client)
+            print(f"  Response: {response[:200]}")
+
+        if not target.exists():
+            print("FAIL: file not created")
+            return 2
+
+        print(f"  File content: {target.read_text(encoding='utf-8')!r}")
+        print(f"  Session ID:   {session_id}")
+
+        if not session_id:
+            print("FAIL: no session_id captured")
+            return 3
+
+        # --- Turn 2: resume and verify context ---
+        print("\n--- Turn 2: Resume ---")
+        async with ClaudeSDKClient(
+            ClaudeAgentOptions(
+                cwd=str(tmpdir),
+                resume=session_id,
+                allowed_tools=["Read"],
+                permission_mode="acceptEdits",
+                max_turns=2,
+                env=auth_env(),
+            )
+        ) as resumed:
+            await resumed.query(
+                "Without modifying files, tell me the exact filename and content you created."
+            )
+            _, resume_response = await drain(resumed)
+            print(f"  Response: {resume_response[:300]}")
+
+        content_ok = target.read_text(encoding="utf-8").strip() == "Session 1"
+        resume_ok = "session_test.txt" in resume_response and "Session 1" in resume_response
+
+        if content_ok and resume_ok:
+            print("\nPASS")
+            return 0
+        print(f"\nFAIL: content_ok={content_ok} resume_ok={resume_ok}")
+        return 4
+    finally:
+        remove_tmpdir(tmpdir)
+
+
+if __name__ == "__main__":
+    raise SystemExit(asyncio.run(main()))
diff --git a/docs/claude/test_hooks_audit_deny.py b/docs/claude/test_hooks_audit_deny.py
new file mode 100644
index 0000000..c573056
--- /dev/null
+++ b/docs/claude/test_hooks_audit_deny.py
@@ -0,0 +1,111 @@
+"""Example: Hooks — audit log + PreToolUse deny.
+
+Demonstrates:
+- PostToolUse hook for audit logging
+- PreToolUse hook to deny dangerous commands
+- HookMatcher pattern matching by tool name
+- Hook callback signature and return format
+
+Usage:
+    cd docs/claude
+    ../../.venv/Scripts/python test_hooks_audit_deny.py
+"""
+
+from __future__ import annotations
+
+import asyncio
+from pathlib import Path
+
+from _sdk_test_common import auth_env, make_tmpdir, remove_tmpdir, setup_auth
+
+audit_log: list[dict] = []
+
+
+async def audit_hook(input_data, tool_use_id, context):
+    """PostToolUse: record every tool invocation."""
+    audit_log.append({
+        "tool": input_data.get("tool_name"),
+        "input": input_data.get("tool_input"),
+    })
+    return {}
+
+
+async def deny_rm_hook(input_data, tool_use_id, context):
+    """PreToolUse: block 'rm' commands."""
+    if input_data.get("tool_name") != "Bash":
+        return {}
+    command = input_data.get("tool_input", {}).get("command", "")
+    if "rm " in command:
+        return {
+            "hookSpecificOutput": {
+                "hookEventName": "PreToolUse",
+                "permissionDecision": "deny",
+                "permissionDecisionReason": "rm commands are blocked by policy",
+            }
+        }
+    return {}
+
+
+async def main() -> int:
+    ok, msg = setup_auth()
+    print(msg)
+    if not ok:
+        return 1
+
+    from claude_agent_sdk import (
+        AssistantMessage,
+        ClaudeAgentOptions,
+        HookMatcher,
+        ResultMessage,
+        TextBlock,
+        query,
+    )
+
+    tmpdir = make_tmpdir("hooks-")
+    try:
+        (tmpdir / "data.txt").write_text("important data\n", encoding="utf-8")
+
+        print("--- Query with hooks ---")
+        opts = ClaudeAgentOptions(
+            cwd=str(tmpdir),
+            allowed_tools=["Bash", "Read"],
+            permission_mode="acceptEdits",
+            max_turns=4,
+            env=auth_env(),
+            hooks={
+                "PostToolUse": [HookMatcher(matcher="Bash|Read", hooks=[audit_hook])],
+                "PreToolUse": [HookMatcher(matcher="Bash", hooks=[deny_rm_hook])],
+            },
+        )
+
+        async for msg in query(
+            prompt=(
+                "Do these steps in order:\n"
+                "1. Read data.txt\n"
+                "2. Run 'echo hello'\n"
+                "3. Run 'rm data.txt'\n"
+                "Report what happened for each step."
+            ),
+            options=opts,
+        ):
+            if isinstance(msg, AssistantMessage):
+                for block in msg.content:
+                    if isinstance(block, TextBlock):
+                        print(f"  Claude: {block.text[:300]}")
+            elif isinstance(msg, ResultMessage):
+                print(f"  Result: {msg.result!r:.300}")
+
+        print(f"\nAudit log ({len(audit_log)} entries):")
+        for entry in audit_log:
+            print(f"  {entry['tool']}: {entry['input']}")
+
+        file_exists = (tmpdir / "data.txt").exists()
+        print(f"\ndata.txt still exists: {file_exists}")
+        print("PASS" if file_exists else "FAIL: rm was not blocked")
+        return 0 if file_exists else 2
+    finally:
+        remove_tmpdir(tmpdir)
+
+
+if __name__ == "__main__":
+    raise SystemExit(asyncio.run(main()))
diff --git a/docs/claude/test_query_read_edit.py b/docs/claude/test_query_read_edit.py
new file mode 100644
index 0000000..8612159
--- /dev/null
+++ b/docs/claude/test_query_read_edit.py
@@ -0,0 +1,96 @@
+"""Example: query() one-shot — Read and Edit files.
+
+Demonstrates:
+- .env-based auth (ANTHROPIC_BASE_URL + ANTHROPIC_AUTH_TOKEN)
+- query() async iterator for simple one-shot tasks
+- allowed_tools for auto-approval
+- permission_mode="acceptEdits"
+- Reading and editing a file in cwd
+- Manual tmpdir management (avoids Windows cleanup races)
+
+Usage:
+    cd docs/claude
+    ../../.venv/Scripts/python test_query_read_edit.py
+"""
+
+from __future__ import annotations
+
+import asyncio
+from pathlib import Path
+
+from _sdk_test_common import auth_env, make_tmpdir, remove_tmpdir, setup_auth
+
+
+async def main() -> int:
+    ok, msg = setup_auth()
+    print(msg)
+    if not ok:
+        return 1
+
+    from claude_agent_sdk import (
+        AssistantMessage,
+        ClaudeAgentOptions,
+        ResultMessage,
+        TextBlock,
+        ToolUseBlock,
+        query,
+    )
+
+    tmpdir = make_tmpdir("read-edit-")
+    try:
+        test_file = tmpdir / "hello.txt"
+        test_file.write_text("hello world\n", encoding="utf-8")
+        print(f"cwd: {tmpdir}")
+
+        # --- Step 1: Read ---
+        print("\n--- Read ---")
+        opts = ClaudeAgentOptions(
+            cwd=str(tmpdir),
+            allowed_tools=["Read"],
+            permission_mode="acceptEdits",
+            max_turns=2,
+            env=auth_env(),
+        )
+        async for msg in query(prompt="Read hello.txt and show its content.", options=opts):
+            if isinstance(msg, AssistantMessage):
+                for block in msg.content:
+                    if isinstance(block, ToolUseBlock):
+                        print(f"  ToolUse: {block.name}({block.input})")
+                    elif isinstance(block, TextBlock):
+                        print(f"  Claude: {block.text[:200]}")
+            elif isinstance(msg, ResultMessage):
+                print(f"  Result: {msg.result!r:.200}")
+
+        # --- Step 2: Edit ---
+        print("\n--- Edit ---")
+        opts = ClaudeAgentOptions(
+            cwd=str(tmpdir),
+            allowed_tools=["Read", "Edit"],
+            permission_mode="acceptEdits",
+            max_turns=3,
+            env=auth_env(),
+        )
+        async for msg in query(
+            prompt=f"Edit hello.txt in {tmpdir}. Add '# edited' as the first line.",
+            options=opts,
+        ):
+            if isinstance(msg, AssistantMessage):
+                for block in msg.content:
+                    if isinstance(block, ToolUseBlock):
+                        print(f"  ToolUse: {block.name}({block.input})")
+                    elif isinstance(block, TextBlock):
+                        print(f"  Claude: {block.text[:200]}")
+            elif isinstance(msg, ResultMessage):
+                print(f"  Result: {msg.result!r:.200}")
+
+        content = test_file.read_text(encoding="utf-8")
+        print(f"\nFile after edit:\n{content}")
+        ok = "# edited" in content
+        print("PASS" if ok else "FAIL")
+        return 0 if ok else 2
+    finally:
+        remove_tmpdir(tmpdir)
+
+
+if __name__ == "__main__":
+    raise SystemExit(asyncio.run(main()))
diff --git a/requirements.txt b/requirements.txt
index 73e75e2..e13a338 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -7,3 +7,4 @@ langchain-community>=0.2.0
 pyyaml>=6.0.0
 rich>=13.0.0
 httpx>=0.27.0
+claude-agent-sdk