feat(handler): 添加消息去重功能防止飞书重复投递

实现基于(user_id, content)的消息去重机制，避免飞书在网络抖动时重复投递相同消息。使用10秒时间窗口判断重复消息，超过窗口的旧记录会被自动清理。

agent/manager.py

@@ -10,7 +10,7 @@ from dataclasses import dataclass, field, asdict
 from pathlib import Path
 from typing import Dict, List, Optional
 
-from agent.pty_process import run_claude
+from agent.pty_process import run_claude, DEFAULT_PERMISSION_MODE, VALID_PERMISSION_MODES
 from agent.audit import log_interaction
 
 logger = logging.getLogger(__name__)
@@ -30,6 +30,7 @@ class Session:
     started: bool = False
     idle_timeout: int = DEFAULT_IDLE_TIMEOUT
     cc_timeout: float = DEFAULT_CC_TIMEOUT
+    permission_mode: str = field(default_factory=lambda: DEFAULT_PERMISSION_MODE)
 
     def touch(self) -> None:
         self.last_activity = asyncio.get_event_loop().time()
@@ -70,6 +71,7 @@ class SessionManager:
         owner_id: str = "",
         idle_timeout: int = DEFAULT_IDLE_TIMEOUT,
         cc_timeout: float = DEFAULT_CC_TIMEOUT,
+        permission_mode: str = DEFAULT_PERMISSION_MODE,
     ) -> Session:
         async with self._lock:
             session = Session(
@@ -78,12 +80,14 @@ class SessionManager:
                 owner_id=owner_id,
                 idle_timeout=idle_timeout,
                 cc_timeout=cc_timeout,
+                permission_mode=permission_mode,
             )
             self._sessions[conv_id] = session
             self._save()
             logger.info(
-                "Created session %s (owner=...%s) in %s (idle=%ds, cc=%.0fs)",
+                "Created session %s (owner=...%s) in %s (idle=%ds, cc=%.0fs, perm=%s)",
-                conv_id, owner_id[-8:] if owner_id else "-", working_dir, idle_timeout, cc_timeout,
+                conv_id, owner_id[-8:] if owner_id else "-", working_dir,
+                idle_timeout, cc_timeout, permission_mode,
             )
             return session
 
@@ -98,6 +102,7 @@ class SessionManager:
             cwd = session.cwd
             cc_session_id = session.cc_session_id
             cc_timeout = session.cc_timeout
+            permission_mode = session.permission_mode
             first_message = not session.started
             if first_message:
                 session.started = True
@@ -116,6 +121,7 @@ class SessionManager:
                     cc_session_id=cc_session_id,
                     resume=not first_message,
                     timeout=cc_timeout,
+                    permission_mode=permission_mode,
                 )
                 log_interaction(
                     conv_id=conv_id,
@@ -158,6 +164,7 @@ class SessionManager:
             cc_session_id=cc_session_id,
             resume=not first_message,
             timeout=cc_timeout,
+            permission_mode=permission_mode,
         )
 
         log_interaction(
@@ -195,10 +202,24 @@ class SessionManager:
                 "started": s.started,
                 "idle_timeout": s.idle_timeout,
                 "cc_timeout": s.cc_timeout,
+                "permission_mode": s.permission_mode,
             }
             for s in sessions
         ]
 
+    def set_permission_mode(self, conv_id: str, mode: str, user_id: Optional[str] = None) -> None:
+        """Change the permission mode for an existing session."""
+        session = self._sessions.get(conv_id)
+        if session is None:
+            raise KeyError(f"No session for conv_id={conv_id!r}")
+        if session.owner_id and user_id and session.owner_id != user_id:
+            raise PermissionError(f"Session {conv_id} belongs to another user")
+        if mode not in VALID_PERMISSION_MODES:
+            raise ValueError(f"Invalid permission mode {mode!r}. Valid: {VALID_PERMISSION_MODES}")
+        session.permission_mode = mode
+        self._save()
+        logger.info("Set permission_mode=%s for session %s", mode, conv_id)
+
     def _save(self) -> None:
         try:
             data = {cid: s.to_dict() for cid, s in self._sessions.items()}

agent/pty_process.py

@@ -17,12 +17,22 @@ def strip_ansi(text: str) -> str:
     return ANSI_ESCAPE.sub("", text)
 
 
+PERMISSION_MODE_FLAGS: dict[str, list[str]] = {
+    "bypassPermissions": ["--dangerously-skip-permissions"],
+    "acceptEdits":       ["--permission-mode", "acceptEdits"],
+    "plan":              ["--permission-mode", "plan"],
+}
+VALID_PERMISSION_MODES = list(PERMISSION_MODE_FLAGS)
+DEFAULT_PERMISSION_MODE = "bypassPermissions"
+
+
 async def run_claude(
     prompt: str,
     cwd: str,
     cc_session_id: str | None = None,
     resume: bool = False,
     timeout: float = 300.0,
+    permission_mode: str = DEFAULT_PERMISSION_MODE,
 ) -> str:
     """
     Run `claude -p <prompt>` in the given directory and return the output.
@@ -35,11 +45,10 @@ async def run_claude(
             - Subsequent calls: passed as --resume so CC has full history.
         resume: If True, use --resume instead of --session-id.
         timeout: Maximum seconds to wait before giving up.
+        permission_mode: One of 'bypassPermissions', 'acceptEdits', 'plan'.
     """
-    base_args = [
+    perm_flags = PERMISSION_MODE_FLAGS.get(permission_mode, PERMISSION_MODE_FLAGS[DEFAULT_PERMISSION_MODE])
-        "--dangerously-skip-permissions",
+    base_args = perm_flags + ["-p", prompt]
-        "-p", prompt,
-    ]
 
     if cc_session_id:
         if resume:

bot/commands.py

@@ -12,11 +12,35 @@ from typing import Optional, Tuple
 from agent.manager import manager
 from agent.scheduler import scheduler
 from agent.task_runner import task_runner
+from agent.pty_process import VALID_PERMISSION_MODES, DEFAULT_PERMISSION_MODE
 from orchestrator.agent import agent
 from orchestrator.tools import set_current_user, get_current_chat
 
 logger = logging.getLogger(__name__)
 
+# Permission mode aliases (user-facing shorthand → internal CC mode)
+_PERM_ALIASES: dict[str, str] = {
+    "bypass": "bypassPermissions",
+    "skip":   "bypassPermissions",
+    "accept": "acceptEdits",
+    "plan":   "plan",
+}
+_PERM_LABELS: dict[str, str] = {
+    "bypassPermissions": "bypass",
+    "acceptEdits":       "accept",
+    "plan":              "plan",
+}
+
+
+def _resolve_perm(alias: str) -> str | None:
+    """Map user-facing alias to internal permission mode, or None if invalid."""
+    return _PERM_ALIASES.get(alias.lower().strip())
+
+
+def _perm_label(mode: str) -> str:
+    """Return short human-readable label for a permission mode."""
+    return _PERM_LABELS.get(mode, mode)
+
 
 def parse_command(text: str) -> Optional[Tuple[str, str]]:
     """
@@ -67,6 +91,8 @@ async def handle_command(user_id: str, text: str) -> Optional[str]:
         return await _cmd_shell(args)
     elif cmd == "/remind":
         return await _cmd_remind(args)
+    elif cmd == "/perm":
+        return await _cmd_perm(user_id, args)
     elif cmd in ("/nodes", "/node"):
         return await _cmd_nodes(user_id, args)
     else:
@@ -76,61 +102,71 @@ async def handle_command(user_id: str, text: str) -> Optional[str]:
 async def _cmd_new(user_id: str, args: str) -> str:
     """Create a new session."""
     if not args:
-        return "Usage: /new <project_dir> [initial_message] [--timeout N]\nExample: /new todo_app fix the bug --timeout 600"
+        return "Usage: /new <project_dir> [initial_message] [--timeout N] [--perm MODE]\nModes: bypass (default), accept, plan"
 
     parser = argparse.ArgumentParser()
     parser.add_argument("working_dir", nargs="?", help="Project directory")
     parser.add_argument("rest", nargs="*", help="Initial message")
     parser.add_argument("--timeout", type=int, default=None, help="CC timeout in seconds")
     parser.add_argument("--idle", type=int, default=None, help="Idle timeout in seconds")
+    parser.add_argument("--perm", default=None, help="Permission mode: bypass, accept, plan")
 
     try:
         parsed = parser.parse_args(args.split())
     except SystemExit:
-        return "Usage: /new <project_dir> [initial_message] [--timeout N] [--idle N]"
+        return "Usage: /new <project_dir> [initial_message] [--timeout N] [--idle N] [--perm MODE]"
 
     if not parsed.working_dir:
         return "Error: project_dir is required"
 
+    permission_mode = _resolve_perm(parsed.perm) if parsed.perm else DEFAULT_PERMISSION_MODE
+    if permission_mode is None:
+        return f"Invalid --perm. Valid modes: bypass, accept, plan"
+
     working_dir = parsed.working_dir
     initial_msg = " ".join(parsed.rest) if parsed.rest else None
 
-    from orchestrator.tools import CreateConversationTool
+    from orchestrator.tools import _resolve_dir
 
-    tool = CreateConversationTool()
-    result = await tool._arun(
-        working_dir=working_dir,
-        initial_message=initial_msg,
-        cc_timeout=parsed.timeout,
-        idle_timeout=parsed.idle,
-    )
     try:
-        data = json.loads(result)
+        resolved = _resolve_dir(working_dir)
-        if "error" in data:
+    except ValueError as e:
-            return f"Error: {data['error']}"
+        return f"Error: {e}"
-        conv_id = data.get("conv_id", "")
+
+    import uuid as _uuid
+    conv_id = str(_uuid.uuid4())[:8]
+    await manager.create(
+        conv_id,
+        str(resolved),
+        owner_id=user_id,
+        idle_timeout=parsed.idle or 1800,
+        cc_timeout=float(parsed.timeout or 300),
+        permission_mode=permission_mode,
+    )
     agent._active_conv[user_id] = conv_id
-        cwd = data.get("working_dir", working_dir)
+
+    response = None
+    if initial_msg:
+        response = await manager.send(conv_id, initial_msg, user_id=user_id)
 
     chat_id = get_current_chat()
     if chat_id:
         from bot.feishu import send_card, send_text, build_sessions_card
         sessions = manager.list_sessions(user_id=user_id)
-            mode = "Direct 🟢" if agent.get_passthrough(user_id) else "Smart ⚪"
+        routing_mode = "Direct 🟢" if agent.get_passthrough(user_id) else "Smart ⚪"
-            card = build_sessions_card(sessions, conv_id, mode)
+        card = build_sessions_card(sessions, conv_id, routing_mode)
         await send_card(chat_id, "chat_id", card)
-            if initial_msg and data.get("response"):
+        if initial_msg and response:
-                await send_text(chat_id, "chat_id", data["response"])
+            await send_text(chat_id, "chat_id", response)
         return ""
 
-        reply = f"✓ Created session `{conv_id}` in `{cwd}`"
+    perm_label = _perm_label(permission_mode)
+    reply = f"✓ Created session `{conv_id}` in `{resolved}` [{perm_label}]"
     if parsed.timeout:
         reply += f" (timeout: {parsed.timeout}s)"
-        if initial_msg:
+    if initial_msg and response:
-            reply += f"\n\nSent: {initial_msg[:100]}..."
+        reply += f"\n\n{response}"
     return reply
-    except Exception:
-        return result
 
 
 async def _cmd_status(user_id: str) -> str:
@@ -152,7 +188,8 @@ async def _cmd_status(user_id: str) -> str:
     lines = ["**Your Sessions:**\n"]
     for i, s in enumerate(sessions, 1):
         marker = "→ " if s["conv_id"] == active else "  "
-        lines.append(f"{marker}{i}. `{s['conv_id']}` - `{s['cwd']}`")
+        perm = _perm_label(s.get("permission_mode", DEFAULT_PERMISSION_MODE))
+        lines.append(f"{marker}{i}. `{s['conv_id']}` - `{s['cwd']}` [{perm}]")
     lines.append(f"\n**Mode:** {mode}")
     lines.append("Use `/switch <n>` to activate a session.")
     lines.append("Use `/direct` or `/smart` to change mode.")
@@ -161,6 +198,28 @@ async def _cmd_status(user_id: str) -> str:
 
 async def _cmd_close(user_id: str, args: str) -> str:
     """Close a session."""
+    # If a specific conv_id is given by name (not a number), resolve it directly.
+    if args:
+        try:
+            int(args)
+            by_number = True
+        except ValueError:
+            by_number = False
+
+        if not by_number:
+            # Explicit conv_id given — look it up directly (may belong to another user).
+            conv_id = args.strip()
+            try:
+                success = await manager.close(conv_id, user_id=user_id)
+                if success:
+                    if agent.get_active_conv(user_id) == conv_id:
+                        agent._active_conv[user_id] = None
+                    return f"✓ Closed session `{conv_id}`"
+                else:
+                    return f"Session `{conv_id}` not found."
+            except PermissionError as e:
+                return str(e)
+
     sessions = manager.list_sessions(user_id=user_id)
     if not sessions:
         return "No sessions to close."
@@ -212,6 +271,38 @@ async def _cmd_switch(user_id: str, args: str) -> str:
         return f"Invalid number: {args}"
 
 
+async def _cmd_perm(user_id: str, args: str) -> str:
+    """Change the permission mode of the active (or specified) session."""
+    parts = args.split()
+    if not parts:
+        return (
+            "Usage: /perm <mode> [conv_id]\n"
+            "Modes: bypass (default), accept, plan\n"
+            "  bypass — skip all permission checks\n"
+            "  accept — auto-accept file edits, confirm shell commands\n"
+            "  plan   — plan only, no writes"
+        )
+
+    alias = parts[0]
+    conv_id = parts[1] if len(parts) > 1 else agent.get_active_conv(user_id)
+
+    permission_mode = _resolve_perm(alias)
+    if permission_mode is None:
+        return f"Unknown mode '{alias}'. Valid: bypass, accept, plan"
+
+    if not conv_id:
+        return "No active session. Use `/perm <mode> <conv_id>` or activate a session first."
+
+    try:
+        manager.set_permission_mode(conv_id, permission_mode, user_id=user_id)
+    except KeyError:
+        return f"Session `{conv_id}` not found."
+    except PermissionError as e:
+        return str(e)
+
+    return f"✓ Session `{conv_id}` permission mode set to **{_perm_label(permission_mode)}**"
+
+
 async def _cmd_retry(user_id: str) -> str:
     """Retry the last message (placeholder - needs history tracking)."""
     return "Retry not yet implemented. Just send your message again."
@@ -335,10 +426,11 @@ async def _cmd_nodes(user_id: str, args: str) -> str:
 def _cmd_help() -> str:
     """Show help."""
     return """**Commands:**
-/new <dir> [msg] [--timeout N] [--idle N] - Create session
+/new <dir> [msg] [--timeout N] [--idle N] [--perm MODE] - Create session
 /status - Show sessions and current mode
 /close [n] - Close session (active or by number)
 /switch <n> - Switch to session by number
+/perm <mode> [conv_id] - Set permission mode (bypass/accept/plan)
 /direct - Direct mode: messages → Claude Code (no LLM overhead)
 /smart - Smart mode: messages → LLM routing (default)
 /shell <cmd> - Run shell command (bypasses LLM)
@@ -347,4 +439,12 @@ def _cmd_help() -> str:
 /nodes - List connected host nodes
 /node <name> - Switch active node
 /retry - Retry last message
-/help - Show this help"""
+/help - Show this help
+
+**Permission modes** (used by /perm and /new --perm):
+  bypass — 跳过所有权限确认，CC 自动执行一切操作（默认）
+            适合：受信任的沙盒环境、自动化任务
+  accept — 自动接受文件编辑，但 shell 命令仍需手动确认
+            适合：日常开发，需要对命令执行保持控制
+  plan   — 只规划、不执行任何写操作
+            适合：先预览 CC 的操作计划再决定是否执行"""

bot/handler.py

@@ -25,6 +25,26 @@ _ws_connected: bool = False
 _last_message_time: float = 0.0
 _reconnect_count: int = 0
 
+# Deduplication: drop Feishu re-deliveries by (user_id, content) within a short window.
+# Feishu retries on network hiccups within ~60s using the same payload.
+# We use a 10s window: identical content from the same user within 10s is a re-delivery,
+# not a deliberate repeat (user intentional repeats arrive after the bot has already replied).
+_recent_messages: dict[tuple[str, str], float] = {}  # key: (user_id, content) → timestamp
+_DEDUP_WINDOW = 10.0  # seconds
+
+
+def _is_duplicate(user_id: str, content: str) -> bool:
+    """Return True if this (user, content) pair arrived within the dedup window."""
+    now = time.time()
+    expired = [k for k, ts in _recent_messages.items() if now - ts > _DEDUP_WINDOW]
+    for k in expired:
+        del _recent_messages[k]
+    key = (user_id, content)
+    if key in _recent_messages:
+        return True
+    _recent_messages[key] = now
+    return False
+
 
 def get_ws_status() -> dict[str, Any]:
     """Return WebSocket connection status."""
@@ -79,10 +99,14 @@ def _handle_message(data: P2ImMessageReceiveV1) -> None:
             logger.info("Empty text after stripping, ignoring")
             return
 
-        logger.info("✉  ...%s → %r", open_id[-8:], text[:80])
-
         user_id = open_id or chat_id
 
+        if _is_duplicate(user_id, text):
+            logger.info("Dropping duplicate delivery: user=...%s text=%r", user_id[-8:], text[:60])
+            return
+
+        logger.info("✉  ...%s → %r", open_id[-8:], text[:80])
+
         if _main_loop is None:
             logger.error("Main event loop not set; cannot process message")
             return

conftest.py

@@ -0,0 +1,16 @@
+"""
+Root conftest — runs before pytest collects any test files or imports any
+production modules. Patches config._CONFIG_PATH to point at the test keyring
+so that `import config` never tries to open the real keyring.yaml.
+
+Must live at the repo root (not inside tests/) to fire before collection.
+"""
+from pathlib import Path
+import importlib
+
+_TEST_KEYRING = Path(__file__).parent / "tests" / "keyring_test.yaml"
+
+# Patch config before anything else imports it
+import config as _config_mod
+_config_mod._CONFIG_PATH = _TEST_KEYRING
+importlib.reload(_config_mod)

orchestrator/agent.py

@@ -40,28 +40,45 @@ Pass these names directly to `create_conversation` — the tool resolves them au
 
 {active_session_line}
 
-Your responsibilities:
+## Tools — two distinct categories
+
+### Bot control commands (use `run_command`)
+`run_command` executes PhoneWork slash commands. Use it when the user asks to:
+- Change permission mode: "切换到只读模式" → run_command("/perm plan")
+- Close/switch sessions: "关掉第一个" → run_command("/close 1")
+- Change routing mode: "切换到直连模式" → run_command("/direct")
+- Set a reminder: "10分钟后提醒我" → run_command("/remind 10m 提醒我")
+- Check status: "看看现在有哪些 session" → run_command("/status")
+- Any other /command the user would type manually
+
+Available bot commands (pass verbatim to run_command):
+  /new <dir> [msg] [--perm bypass|accept|plan]  — create session
+  /close [n|conv_id]   — close session
+  /switch <n>          — switch active session
+  /perm <mode> [conv_id] — permission mode: bypass (default), accept, plan
+  /direct              — direct mode (bypass LLM for CC messages)
+  /smart               — smart mode (LLM routing, default)
+  /status              — list sessions
+  /remind <Ns|Nm|Nh> <msg> — one-shot reminder
+  /tasks               — list background tasks
+
+### Host shell commands (use `run_shell`)
+`run_shell` executes shell commands on the host machine (git, ls, cat, pip, etc.).
+NEVER use `run_shell` for bot control. NEVER use `run_command` for shell commands.
+
+## Session responsibilities
 1. NEW session: call `create_conversation` with the project name/path. \
    If the user's message also contains a task, pass it as `initial_message` too.
 2. Follow-up to ACTIVE session: call `send_to_conversation` with the active conv_id shown above.
-3. List sessions: call `list_conversations`.
+3. BOT CONTROL: call `run_command` with the appropriate slash command.
-4. Close session: call `close_conversation`.
+4. GENERAL QUESTIONS: answer directly — do NOT create a session for simple Q&A.
-5. GENERAL QUESTIONS: If the user asks a general question (not about a specific project or file), \
+5. WEB / SEARCH: use `web` at most twice, then synthesize and reply.
-   answer directly using your own knowledge. Do NOT create a session for simple Q&A.
+6. BACKGROUND TASKS: when a task starts, reply immediately — do NOT poll `task_status`.
-6. WEB / SEARCH: Use the `web` tool when the user needs current information. \
-   Call it ONCE (or at most twice with a refined query). Then synthesize and reply — \
-   do NOT keep searching in a loop. If the first search returns results, use them.
-7. BACKGROUND TASKS: When `create_conversation` or `send_to_conversation` returns a \
-   "Task #... started" message, the task is running in the background. \
-   Immediately reply to the user that the task has started and they will be notified. \
-   Do NOT call `task_status` in a loop waiting for it — the system sends a notification when done.
 
 Guidelines:
 - Relay Claude Code's output verbatim.
-- If no active session and the user sends a task without naming a directory, ask them which project.
+- If no active session and the user sends a task without naming a directory, ask which project.
-- For general knowledge questions (e.g., "what is a Python generator?", "explain async/await"), \
+- After using any tool, always produce a final text reply. Never end a turn on a tool call.
-  answer directly without creating a session.
-- After using any tool, always produce a final text reply to the user. Never end a turn on a tool call.
 - Keep your own words brief — let Claude Code's output speak.
 - Reply in the same language the user uses (Chinese or English).
 """

orchestrator/tools.py

@@ -716,12 +716,49 @@ class TaskStatusTool(BaseTool):
         }, ensure_ascii=False)
 
 
+class RunCommandInput(BaseModel):
+    command: str = Field(
+        ...,
+        description=(
+            "A bot slash command to execute (e.g. '/perm accept', '/close 1', '/switch 2'). "
+            "This runs bot control commands — NOT shell commands on the host machine. "
+            "Use run_shell for host shell commands (git, ls, etc.)."
+        ),
+    )
+
+
+class RunCommandTool(BaseTool):
+    name: str = "run_command"
+    description: str = (
+        "Execute a PhoneWork bot slash command on behalf of the user. "
+        "Use this to control sessions, switch modes, change permissions, etc. "
+        "Examples: '/perm accept', '/close 1', '/switch 2', '/direct', '/smart', '/status'. "
+        "Do NOT use this for shell commands — use run_shell for those."
+    )
+    args_schema: Type[BaseModel] = RunCommandInput
+
+    def _run(self, command: str) -> str:
+        raise NotImplementedError("Use async version")
+
+    async def _arun(self, command: str) -> str:
+        from bot.commands import handle_command
+        from orchestrator.tools import get_current_user
+        user_id = get_current_user()
+        if not user_id:
+            return "Error: no user context"
+        result = await handle_command(user_id, command.strip())
+        if result is None:
+            return f"Unknown command: {command!r}. Use /help to see available commands."
+        return result
+
+
 # Module-level tool list for easy import
 TOOLS = [
     CreateConversationTool(),
     SendToConversationTool(),
     ListConversationsTool(),
     CloseConversationTool(),
+    RunCommandTool(),
     ShellTool(),
     FileReadTool(),
     FileWriteTool(),

pyproject.toml

@@ -0,0 +1,3 @@
+[tool.pytest.ini_options]
+asyncio_mode = "auto"
+testpaths = ["tests"]

requirements-dev.txt

@@ -0,0 +1,5 @@
+pytest>=8.0.0
+pytest-asyncio>=0.24.0
+pytest-bdd>=7.0.0
+pytest-recording>=0.13.0
+pytest-mock>=3.12.0

tests/conftest.py

@@ -0,0 +1,167 @@
+"""
+Master test fixtures for PhoneWork BDD tests.
+"""
+from __future__ import annotations
+
+import asyncio
+import time
+from pathlib import Path
+from typing import Any
+from unittest.mock import AsyncMock, MagicMock, patch
+
+import pytest
+
+TESTS_DIR = Path(__file__).parent
+CASSETTES_DIR = TESTS_DIR / "cassettes"
+CASSETTES_DIR.mkdir(exist_ok=True)
+
+
+# ── Feishu send mock ─────────────────────────────────────────────────────────
+
+@pytest.fixture
+def feishu_calls():
+    """
+    Capture all calls to bot.feishu send functions.
+    Lazy imports inside commands.py pull from bot.feishu at call time,
+    so patching the module attributes is sufficient.
+    """
+    captured: dict[str, list] = {"texts": [], "cards": [], "files": []}
+
+    async def mock_send_text(receive_id, receive_id_type, text):
+        captured["texts"].append({"receive_id": receive_id, "text": text})
+
+    async def mock_send_card(receive_id, receive_id_type, card):
+        captured["cards"].append({"receive_id": receive_id, "card": card})
+
+    async def mock_send_file(receive_id, receive_id_type, file_path, file_type="stream"):
+        captured["files"].append({"receive_id": receive_id, "file_path": file_path})
+
+    with patch("bot.feishu.send_text", side_effect=mock_send_text), \
+         patch("bot.feishu.send_card", side_effect=mock_send_card), \
+         patch("bot.feishu.send_file", side_effect=mock_send_file):
+        yield captured
+
+
+# ── run_claude mock ──────────────────────────────────────────────────────────
+
+@pytest.fixture
+def mock_run_claude():
+    """
+    Replace run_claude in both its definition and its import site in manager.py.
+    Default return value is a short CC-style output string.
+    """
+    mock = AsyncMock(return_value="Claude Code: task complete.")
+    with patch("agent.pty_process.run_claude", mock), \
+         patch("agent.manager.run_claude", mock):
+        yield mock
+
+
+# ── Singleton state resets ───────────────────────────────────────────────────
+
+@pytest.fixture(autouse=True)
+def reset_manager():
+    from agent.manager import manager
+    manager._sessions.clear()
+    yield
+    manager._sessions.clear()
+
+
+@pytest.fixture(autouse=True)
+def reset_agent():
+    from orchestrator.agent import agent
+    agent._history.clear()
+    agent._active_conv.clear()
+    agent._passthrough.clear()
+    agent._user_locks.clear()
+    yield
+    agent._history.clear()
+    agent._active_conv.clear()
+    agent._passthrough.clear()
+    agent._user_locks.clear()
+
+
+@pytest.fixture(autouse=True)
+def reset_task_runner():
+    from agent.task_runner import task_runner
+    task_runner._tasks.clear()
+    yield
+    task_runner._tasks.clear()
+
+
+@pytest.fixture(autouse=True)
+def reset_scheduler():
+    from agent.scheduler import scheduler
+    for task in list(getattr(scheduler, "_tasks", {}).values()):
+        task.cancel()
+    scheduler._jobs.clear()
+    yield
+    for task in list(getattr(scheduler, "_tasks", {}).values()):
+        task.cancel()
+    scheduler._jobs.clear()
+
+
+@pytest.fixture(autouse=True)
+def reset_contextvars():
+    from orchestrator.tools import set_current_user, set_current_chat
+    set_current_user(None)
+    set_current_chat(None)
+    yield
+    set_current_user(None)
+    set_current_chat(None)
+
+
+@pytest.fixture(autouse=True)
+def reset_reply(pytestconfig):
+    """Clear _reply before each test so stale values don't leak between scenarios."""
+    pytestconfig._reply = None
+    yield
+
+
+# ── Working directory isolation ──────────────────────────────────────────────
+
+@pytest.fixture
+def tmp_working_dir(tmp_path, monkeypatch):
+    import config
+    import orchestrator.tools as tools_mod
+    monkeypatch.setattr(config, "WORKING_DIR", tmp_path)
+    monkeypatch.setattr(tools_mod, "WORKING_DIR", tmp_path)
+    (tmp_path / "myproject").mkdir()
+    return tmp_path
+
+
+# ── VCR cassette factory ─────────────────────────────────────────────────────
+
+def make_vcr_cassette(cassette_name: str):
+    """
+    Return a vcrpy context manager for the given cassette name.
+    Set VCR_RECORD_MODE=new_episodes locally to record; CI uses 'none'.
+    Authorization headers are stripped so cassettes are safe to commit.
+    If the cassette doesn't exist in 'none' mode, the test is skipped.
+    """
+    import os
+    try:
+        import vcr
+    except ImportError:
+        import pytest
+        pytest.skip("vcrpy not installed")
+
+    record_mode = os.environ.get("VCR_RECORD_MODE", "none")
+    cassette_path = CASSETTES_DIR / cassette_name
+    cassette_path.parent.mkdir(parents=True, exist_ok=True)
+
+    if record_mode == "none" and not cassette_path.exists():
+        import pytest
+        pytest.skip(f"No cassette recorded yet: {cassette_name}. Run with VCR_RECORD_MODE=new_episodes to record.")
+
+    my_vcr = vcr.VCR(
+        record_mode=record_mode,
+        match_on=["method", "scheme", "host", "port", "path", "body"],
+        filter_headers=["authorization", "x-api-key"],
+        decode_compressed_response=True,
+    )
+    return my_vcr.use_cassette(str(cassette_path))
+
+
+@pytest.fixture
+def vcr_cassette():
+    return make_vcr_cassette

tests/features/agent/passthrough.feature

@@ -0,0 +1,19 @@
+Feature: Direct (passthrough) mode — bypass LLM for CC sessions
+
+  Background:
+    Given user "user_abc123" is sending commands
+    And run_claude returns "Done. Here is the result."
+
+  Scenario: Passthrough sends directly to CC without LLM
+    Given user has session "sess01" in "/tmp/proj1"
+    And active session is "sess01"
+    And direct mode is enabled for user "user_abc123"
+    When user sends agent message "run the tests"
+    Then run_claude was called
+    And reply contains "Done. Here is the result."
+
+  Scenario: Passthrough on missing session clears active conv
+    Given active session is "ghost_session_id" which does not exist
+    And direct mode is enabled for user "user_abc123"
+    When user sends agent message "hello"
+    Then active session for user "user_abc123" is None

tests/features/agent/routing.feature

@@ -0,0 +1,35 @@
+Feature: LLM smart routing — agent routes messages to correct tools
+
+  Background:
+    Given user "user_abc123" is in smart mode
+    And run_claude returns "I created the component for you."
+
+  @vcr
+  Scenario: Agent creates new session for project task
+    Given vcr cassette "agent/routing_new_session.yaml"
+    When user sends agent message "create a React app in todo_app folder"
+    Then agent created a session for user "user_abc123"
+    And reply is not empty
+
+  @vcr
+  Scenario: Agent answers general question without creating session
+    Given vcr cassette "agent/routing_general_qa.yaml"
+    When user sends agent message "what is a Python generator?"
+    Then no session is created for user "user_abc123"
+    And reply is not empty
+
+  @vcr
+  Scenario: Agent sends follow-up to existing session
+    Given user has active session "sess01" in "/tmp/proj1"
+    And vcr cassette "agent/routing_follow_up.yaml"
+    When user sends agent message "now add tests for that"
+    Then run_claude was called
+    And reply is not empty
+
+  @vcr
+  Scenario: Agent answers direct QA without tools when no active session
+    Given no active session for user "user_abc123"
+    And vcr cassette "agent/routing_direct_qa.yaml"
+    When user sends agent message "explain async/await in Python"
+    Then reply is not empty
+    And reply does not contain "Max iterations"

tests/features/commands/close.feature

@@ -0,0 +1,38 @@
+Feature: /close command — terminate a session
+
+  Background:
+    Given user "user_abc123" is sending commands
+
+  Scenario: No sessions returns error
+    When user sends "/close"
+    Then reply contains "No sessions to close"
+
+  Scenario: Close active session by default
+    Given user has session "sess01" in "/tmp/proj1"
+    And active session is "sess01"
+    When user sends "/close"
+    Then reply contains "Closed session"
+    And session manager has 0 sessions for user "user_abc123"
+
+  Scenario: Close session by number
+    Given user has session "sess01" in "/tmp/proj1"
+    And user has session "sess02" in "/tmp/proj2"
+    When user sends "/close 1"
+    Then reply contains "Closed session"
+    And session manager has 1 session for user "user_abc123"
+
+  Scenario: Invalid number returns error
+    Given user has session "sess01" in "/tmp/proj1"
+    When user sends "/close 9"
+    Then reply contains "Invalid session number"
+
+  Scenario: Cannot close another user's session
+    Given session "sess01" in "/tmp/proj1" belongs to user "other_user"
+    When user sends "/close sess01"
+    Then reply contains "belongs to another user"
+
+  Scenario: Closing active session clears active conv
+    Given user has session "sess01" in "/tmp/proj1"
+    And active session is "sess01"
+    When user sends "/close"
+    Then active session for user "user_abc123" is None

tests/features/commands/direct_smart.feature

@@ -0,0 +1,27 @@
+Feature: /direct and /smart mode toggle
+
+  Background:
+    Given user "user_abc123" is sending commands
+
+  Scenario: /direct requires active session
+    When user sends "/direct"
+    Then reply contains "No active session"
+
+  Scenario: /direct enables passthrough mode
+    Given user has session "sess01" in "/tmp/proj1"
+    And active session is "sess01"
+    When user sends "/direct"
+    Then reply contains "Direct mode ON"
+    And passthrough mode is enabled for user "user_abc123"
+
+  Scenario: /smart disables passthrough mode
+    Given user has session "sess01" in "/tmp/proj1"
+    And active session is "sess01"
+    And direct mode is enabled for user "user_abc123"
+    When user sends "/smart"
+    Then reply contains "Smart mode ON"
+    And passthrough mode is disabled for user "user_abc123"
+
+  Scenario: /smart always succeeds even without active session
+    When user sends "/smart"
+    Then reply contains "Smart mode ON"

tests/features/commands/help.feature

@@ -0,0 +1,25 @@
+Feature: /help command — show command reference
+
+  Background:
+    Given user "user_abc123" is sending commands
+
+  Scenario: /help lists all commands
+    When user sends "/help"
+    Then reply contains "/new"
+    And reply contains "/status"
+    And reply contains "/close"
+    And reply contains "/switch"
+    And reply contains "/direct"
+    And reply contains "/smart"
+    And reply contains "/shell"
+    And reply contains "/remind"
+    And reply contains "/tasks"
+    And reply contains "/nodes"
+
+  Scenario: /h alias works
+    When user sends "/h"
+    Then reply contains "/new"
+
+  Scenario: Unknown command is not handled
+    When user sends "/unknown_xyz_cmd"
+    Then command is not handled

tests/features/commands/new.feature

@@ -0,0 +1,37 @@
+Feature: /new command — create a Claude Code session
+
+  Background:
+    Given user "user_abc123" is sending commands
+
+  Scenario: No arguments shows usage
+    When user sends "/new"
+    Then reply contains "Usage: /new"
+
+  Scenario: Creates session with valid directory
+    Given run_claude returns "Session ready."
+    When user sends "/new myproject"
+    Then reply contains "myproject"
+    And session manager has 1 session for user "user_abc123"
+
+  Scenario: Creates session with initial message
+    Given run_claude returns "Fixed the bug."
+    When user sends "/new myproject fix the login bug"
+    Then reply contains "myproject"
+
+  Scenario: Path traversal attempt is blocked
+    When user sends "/new ../../etc"
+    Then reply contains "Error"
+    And session manager has 0 sessions for user "user_abc123"
+
+  Scenario: Custom timeout is accepted
+    Given run_claude returns "Done."
+    When user sends "/new myproject --timeout 60"
+    Then reply contains "myproject"
+    And reply contains "timeout: 60s"
+
+  Scenario: Creates session and sends card when chat_id is set
+    Given the current chat_id is "chat_xyz"
+    And run_claude returns "Ready."
+    When user sends "/new myproject"
+    Then a sessions card is sent to chat "chat_xyz"
+    And text reply is empty

tests/features/commands/nodes.feature

@@ -0,0 +1,13 @@
+Feature: /nodes and /node commands — multi-host node management
+
+  Background:
+    Given user "user_abc123" is sending commands
+    And ROUTER_MODE is disabled
+
+  Scenario: /nodes outside router mode returns explanation
+    When user sends "/nodes"
+    Then reply contains "Not in router mode"
+
+  Scenario: /node outside router mode returns explanation
+    When user sends "/node myhost"
+    Then reply contains "Not in router mode"

tests/features/commands/perm.feature

@@ -0,0 +1,70 @@
+Feature: /perm command — change session permission mode
+
+  Background:
+    Given user "user_abc123" is sending commands
+
+  Scenario: No args shows usage
+    When user sends "/perm"
+    Then reply contains "Usage"
+    And reply contains "bypass"
+    And reply contains "accept"
+    And reply contains "plan"
+
+  Scenario: Set active session to accept mode
+    Given user has session "sess01" in "/tmp/proj1"
+    And active session is "sess01"
+    When user sends "/perm accept"
+    Then reply contains "accept"
+    And reply contains "sess01"
+    And session "sess01" has permission mode "acceptEdits"
+
+  Scenario: Set active session to plan mode
+    Given user has session "sess01" in "/tmp/proj1"
+    And active session is "sess01"
+    When user sends "/perm plan"
+    Then reply contains "plan"
+    And session "sess01" has permission mode "plan"
+
+  Scenario: Set active session back to bypass
+    Given user has session "sess01" in "/tmp/proj1"
+    And active session is "sess01"
+    When user sends "/perm bypass"
+    Then reply contains "bypass"
+    And session "sess01" has permission mode "bypassPermissions"
+
+  Scenario: Unknown mode returns error
+    Given user has session "sess01" in "/tmp/proj1"
+    And active session is "sess01"
+    When user sends "/perm turbo"
+    Then reply contains "Unknown mode"
+
+  Scenario: No active session returns error
+    Given no active session for user "user_abc123"
+    When user sends "/perm accept"
+    Then reply contains "No active session"
+
+  Scenario: Set permission on specific conv_id
+    Given user has session "sess01" in "/tmp/proj1"
+    And user has session "sess02" in "/tmp/proj2"
+    And active session is "sess01"
+    When user sends "/perm plan sess02"
+    Then reply contains "sess02"
+    And session "sess02" has permission mode "plan"
+
+  Scenario: Cannot change permission of another user's session
+    Given session "sess01" in "/tmp/proj1" belongs to user "other_user"
+    When user sends "/perm accept sess01"
+    Then reply contains "another user"
+
+  Scenario: New session with --perm accept
+    When user sends "/new myproject --perm accept"
+    Then reply contains "accept"
+    And session manager has 1 session for user "user_abc123"
+
+  Scenario: New session with --perm plan
+    When user sends "/new myproject --perm plan"
+    Then reply contains "plan"
+
+  Scenario: New session with invalid --perm
+    When user sends "/new myproject --perm turbo"
+    Then reply contains "Invalid"

tests/features/commands/remind.feature

@@ -0,0 +1,33 @@
+Feature: /remind command — schedule a one-time reminder
+
+  Background:
+    Given user "user_abc123" is sending commands
+    And the current chat_id is "chat_xyz"
+
+  Scenario: No arguments shows usage
+    When user sends "/remind"
+    Then reply contains "Usage: /remind"
+
+  Scenario: Missing message part shows usage
+    When user sends "/remind 10m"
+    Then reply contains "Usage: /remind"
+
+  Scenario: Invalid time format returns error
+    When user sends "/remind badtime check build"
+    Then reply contains "Invalid time format"
+
+  Scenario: Valid reminder with seconds is scheduled
+    When user sends "/remind 30s check the build"
+    Then reply contains "Reminder #"
+    And reply contains "30s"
+    And scheduler has 1 pending job
+
+  Scenario: Valid reminder with minutes is scheduled
+    When user sends "/remind 5m deploy done"
+    Then reply contains "5m"
+    And scheduler has 1 pending job
+
+  Scenario: Valid reminder with hours is scheduled
+    When user sends "/remind 2h weekly report"
+    Then reply contains "2h"
+    And scheduler has 1 pending job

tests/features/commands/shell.feature

@@ -0,0 +1,22 @@
+Feature: /shell command — run host shell commands
+
+  Background:
+    Given user "user_abc123" is sending commands
+
+  Scenario: No arguments shows usage
+    When user sends "/shell"
+    Then reply contains "Usage: /shell"
+
+  Scenario: Runs echo and returns output
+    When user sends "/shell echo hello"
+    Then reply contains "hello"
+    And reply contains "exit code: 0"
+
+  Scenario: Blocked dangerous command is rejected
+    When user sends "/shell rm -rf /"
+    Then reply contains "Blocked"
+    And reply does not contain "exit code"
+
+  Scenario: Non-zero exit code is reported
+    When user sends "/shell exit 1"
+    Then reply contains "exit code"

tests/features/commands/status.feature

@@ -0,0 +1,40 @@
+Feature: /status command — list sessions and current mode
+
+  Background:
+    Given user "user_abc123" is sending commands
+
+  Scenario: No sessions returns empty message
+    When user sends "/status"
+    Then reply contains "No active sessions"
+
+  Scenario: Shows session list
+    Given user has session "sess01" in "/tmp/proj1"
+    And user has session "sess02" in "/tmp/proj2"
+    When user sends "/status"
+    Then reply contains "sess01"
+    And reply contains "sess02"
+
+  Scenario: Shows active marker on current session
+    Given user has session "sess01" in "/tmp/proj1"
+    And active session is "sess01"
+    When user sends "/status"
+    Then reply contains "→"
+
+  Scenario: Shows current mode as Smart by default
+    Given user has session "sess01" in "/tmp/proj1"
+    When user sends "/status"
+    Then reply contains "Smart"
+
+  Scenario: Shows Direct mode after /direct
+    Given user has session "sess01" in "/tmp/proj1"
+    And active session is "sess01"
+    And direct mode is enabled for user "user_abc123"
+    When user sends "/status"
+    Then reply contains "Direct"
+
+  Scenario: Sends card when chat_id is set
+    Given user has session "sess01" in "/tmp/proj1"
+    And the current chat_id is "chat_xyz"
+    When user sends "/status"
+    Then a sessions card is sent to chat "chat_xyz"
+    And text reply is empty

tests/features/commands/switch.feature

@@ -0,0 +1,30 @@
+Feature: /switch command — activate a different session
+
+  Background:
+    Given user "user_abc123" is sending commands
+
+  Scenario: No sessions returns error
+    When user sends "/switch 1"
+    Then reply contains "No sessions available"
+
+  Scenario: Valid switch updates active session
+    Given user has session "sess01" in "/tmp/proj1"
+    And user has session "sess02" in "/tmp/proj2"
+    When user sends "/switch 2"
+    Then reply contains "Switched to session"
+    And active session for user "user_abc123" is "sess02"
+
+  Scenario: Out of range number returns error
+    Given user has session "sess01" in "/tmp/proj1"
+    When user sends "/switch 5"
+    Then reply contains "Invalid session number"
+
+  Scenario: Non-numeric argument returns error
+    Given user has session "sess01" in "/tmp/proj1"
+    When user sends "/switch notanumber"
+    Then reply contains "Invalid number"
+
+  Scenario: Missing argument shows usage
+    Given user has session "sess01" in "/tmp/proj1"
+    When user sends "/switch"
+    Then reply contains "Usage: /switch"

tests/features/commands/tasks.feature

@@ -0,0 +1,27 @@
+Feature: /tasks command — list background tasks
+
+  Background:
+    Given user "user_abc123" is sending commands
+
+  Scenario: No tasks returns empty message
+    When user sends "/tasks"
+    Then reply contains "No background tasks"
+
+  Scenario: Shows running task with spinner emoji
+    Given there is a running task "task001" described as "CC session abc: fix bug"
+    When user sends "/tasks"
+    Then reply contains "task001"
+    And reply contains "fix bug"
+    And reply contains "⏳"
+
+  Scenario: Shows completed task with checkmark
+    Given there is a completed task "task002" described as "CC session xyz: deploy"
+    When user sends "/tasks"
+    Then reply contains "task002"
+    And reply contains "✅"
+
+  Scenario: Shows failed task with cross
+    Given there is a failed task "task003" described as "CC session err: bad cmd"
+    When user sends "/tasks"
+    Then reply contains "task003"
+    And reply contains "❌"

tests/keyring_test.yaml

@@ -0,0 +1,10 @@
+FEISHU_APP_ID: test_app_id
+FEISHU_APP_SECRET: test_app_secret
+OPENAI_BASE_URL: https://open.bigmodel.cn/api/paas/v4/
+OPENAI_API_KEY: test_api_key_for_vcr
+OPENAI_MODEL: glm-4.7
+WORKING_DIR: /tmp/phonework_test
+METASO_API_KEY: ""
+ROUTER_MODE: false
+ROUTER_SECRET: ""
+ALLOWED_OPEN_IDS: []

tests/step_defs/common_steps.py

@@ -0,0 +1,192 @@
+"""
+Shared Given/Then step definitions used across all feature files.
+"""
+from __future__ import annotations
+
+from pytest_bdd import given, then, parsers
+
+
+# ── Given: user identity ─────────────────────────────────────────────────────
+
+@given(parsers.parse('user "{user_id}" is sending commands'))
+def set_user(user_id, pytestconfig):
+    from orchestrator.tools import set_current_user
+    set_current_user(user_id)
+    pytestconfig._test_user_id = user_id
+
+
+@given(parsers.parse('the current chat_id is "{chat_id}"'))
+def set_chat(chat_id):
+    from orchestrator.tools import set_current_chat
+    set_current_chat(chat_id)
+
+
+# ── Given: session setup ─────────────────────────────────────────────────────
+
+@given(parsers.parse('user has session "{conv_id}" in "{cwd}"'))
+def add_session(conv_id, cwd, pytestconfig, tmp_path):
+    from agent.manager import manager, Session
+    user_id = getattr(pytestconfig, "_test_user_id", "user_abc123")
+    session = Session(conv_id=conv_id, cwd=str(tmp_path / conv_id), owner_id=user_id, cc_timeout=50.0)
+    (tmp_path / conv_id).mkdir(exist_ok=True)
+    manager._sessions[conv_id] = session
+
+
+@given(parsers.parse('session "{conv_id}" in "{cwd}" belongs to user "{owner}"'))
+def add_foreign_session(conv_id, cwd, owner, tmp_path):
+    from agent.manager import manager, Session
+    session = Session(conv_id=conv_id, cwd=str(tmp_path / conv_id), owner_id=owner, cc_timeout=50.0)
+    (tmp_path / conv_id).mkdir(exist_ok=True)
+    manager._sessions[conv_id] = session
+
+
+@given(parsers.parse('active session is "{conv_id}"'))
+def set_active_session(conv_id, pytestconfig):
+    from orchestrator.agent import agent
+    user_id = getattr(pytestconfig, "_test_user_id", "user_abc123")
+    agent._active_conv[user_id] = conv_id
+
+
+@given(parsers.parse('active session is "{conv_id}" which does not exist'))
+def set_ghost_active_session(conv_id, pytestconfig):
+    from orchestrator.agent import agent
+    user_id = getattr(pytestconfig, "_test_user_id", "user_abc123")
+    agent._active_conv[user_id] = conv_id
+    # intentionally NOT added to manager._sessions
+
+
+@given(parsers.parse('no active session for user "{user_id}"'))
+def ensure_no_active_session(user_id):
+    from orchestrator.agent import agent
+    agent._active_conv[user_id] = None
+
+
+# ── Given: mode toggles ──────────────────────────────────────────────────────
+
+@given(parsers.parse('direct mode is enabled for user "{user_id}"'))
+def enable_direct_mode(user_id):
+    from orchestrator.agent import agent
+    agent._passthrough[user_id] = True
+
+
+# ── Given: mocks ─────────────────────────────────────────────────────────────
+
+@given(parsers.parse('run_claude returns "{output}"'))
+def set_run_claude_return(output, mock_run_claude):
+    mock_run_claude.return_value = output
+
+
+# ── Given: config ────────────────────────────────────────────────────────────
+
+@given("ROUTER_MODE is disabled")
+def disable_router_mode(monkeypatch):
+    import config
+    monkeypatch.setattr(config, "ROUTER_MODE", False)
+
+
+# ── Then: reply assertions ───────────────────────────────────────────────────
+
+@then(parsers.parse('reply contains "{text}"'))
+def reply_contains(text, pytestconfig):
+    reply = getattr(pytestconfig, "_reply", None)
+    assert text in (reply or ""), \
+        f"Expected {text!r} in reply, got: {reply!r}"
+
+
+@then(parsers.parse('reply does not contain "{text}"'))
+def reply_not_contains(text, pytestconfig):
+    reply = getattr(pytestconfig, "_reply", None)
+    assert text not in (reply or ""), \
+        f"Expected {text!r} NOT in reply, got: {reply!r}"
+
+
+@then("reply is not empty")
+def reply_not_empty(pytestconfig):
+    reply = getattr(pytestconfig, "_reply", None)
+    assert reply and reply.strip(), \
+        f"Expected non-empty reply, got: {reply!r}"
+
+
+@then("text reply is empty")
+def reply_is_empty(pytestconfig):
+    reply = getattr(pytestconfig, "_reply", None)
+    assert reply == "", \
+        f"Expected empty reply, got: {reply!r}"
+
+
+@then("command is not handled")
+def command_not_handled(pytestconfig):
+    reply = getattr(pytestconfig, "_reply", None)
+    assert reply is None
+
+
+# ── Then: session state ──────────────────────────────────────────────────────
+
+@then(parsers.parse('session manager has {count:d} session for user "{user_id}"'))
+@then(parsers.parse('session manager has {count:d} sessions for user "{user_id}"'))
+def check_session_count(count, user_id):
+    from agent.manager import manager
+    sessions = manager.list_sessions(user_id=user_id)
+    assert len(sessions) == count, \
+        f"Expected {count} sessions, got {len(sessions)}: {sessions}"
+
+
+@then(parsers.parse('active session for user "{user_id}" is "{conv_id}"'))
+def check_active_session(user_id, conv_id):
+    from orchestrator.agent import agent
+    assert agent._active_conv.get(user_id) == conv_id
+
+
+@then(parsers.parse('active session for user "{user_id}" is None'))
+def check_no_active_session(user_id):
+    from orchestrator.agent import agent
+    assert agent._active_conv.get(user_id) is None
+
+
+@then(parsers.parse('session "{conv_id}" has permission mode "{mode}"'))
+def check_session_perm_mode(conv_id, mode):
+    from agent.manager import manager
+    session = manager._sessions.get(conv_id)
+    assert session is not None, f"Session {conv_id!r} not found"
+    assert session.permission_mode == mode, \
+        f"Expected permission_mode={mode!r}, got {session.permission_mode!r}"
+
+
+# ── Then: mode state ─────────────────────────────────────────────────────────
+
+@then(parsers.parse('passthrough mode is enabled for user "{user_id}"'))
+def check_passthrough_on(user_id):
+    from orchestrator.agent import agent
+    assert agent._passthrough.get(user_id) is True
+
+
+@then(parsers.parse('passthrough mode is disabled for user "{user_id}"'))
+def check_passthrough_off(user_id):
+    from orchestrator.agent import agent
+    assert agent._passthrough.get(user_id) is False
+
+
+# ── Then: Feishu output ──────────────────────────────────────────────────────
+
+@then(parsers.parse('a sessions card is sent to chat "{chat_id}"'))
+def check_card_sent(chat_id, feishu_calls):
+    cards = feishu_calls["cards"]
+    assert any(c["receive_id"] == chat_id for c in cards), \
+        f"No card sent to {chat_id!r}, captured: {cards}"
+
+
+# ── Then: scheduler ──────────────────────────────────────────────────────────
+
+@then(parsers.parse('scheduler has {count:d} pending job'))
+@then(parsers.parse('scheduler has {count:d} pending jobs'))
+def check_scheduler_jobs(count):
+    from agent.scheduler import scheduler
+    assert len(scheduler._jobs) == count, \
+        f"Expected {count} jobs, got {len(scheduler._jobs)}"
+
+
+# ── Then: run_claude ─────────────────────────────────────────────────────────
+
+@then("run_claude was called")
+def check_run_claude_called(mock_run_claude):
+    assert mock_run_claude.call_count >= 1, "Expected run_claude to be called"

tests/step_defs/test_agent.py

@@ -0,0 +1,76 @@
+"""
+Step definitions for agent routing and passthrough features.
+"""
+from __future__ import annotations
+
+from pytest_bdd import scenarios, given, when, then, parsers
+
+from tests.step_defs.common_steps import *  # noqa: F401,F403 — import shared steps
+
+scenarios(
+    "../features/agent/routing.feature",
+    "../features/agent/passthrough.feature",
+)
+
+
+# ── Given: agent-specific setup ──────────────────────────────────────────────
+
+@given(parsers.parse('user "{user_id}" is in smart mode'))
+def set_smart_mode(user_id, pytestconfig):
+    from orchestrator.agent import agent
+    from orchestrator.tools import set_current_user
+    set_current_user(user_id)
+    agent._passthrough[user_id] = False
+    pytestconfig._test_user_id = user_id
+
+
+@given(parsers.parse('user has active session "{conv_id}" in "{cwd}"'))
+def add_and_activate_session(conv_id, cwd, pytestconfig, tmp_path):
+    from agent.manager import manager, Session
+    from orchestrator.agent import agent
+    user_id = getattr(pytestconfig, "_test_user_id", "user_abc123")
+    session = Session(conv_id=conv_id, cwd=str(tmp_path / conv_id), owner_id=user_id, cc_timeout=50.0)
+    (tmp_path / conv_id).mkdir(exist_ok=True)
+    manager._sessions[conv_id] = session
+    agent._active_conv[user_id] = conv_id
+
+
+@given(parsers.parse('vcr cassette "{cassette_name}"'))
+def set_vcr_cassette(cassette_name, pytestconfig):
+    pytestconfig._vcr_cassette = cassette_name
+
+
+# ── When: send message through agent ─────────────────────────────────────────
+
+@when(parsers.parse('user sends agent message "{text}"'))
+def send_agent_message(text, pytestconfig, mock_run_claude, feishu_calls):
+    import asyncio
+    from orchestrator.agent import agent
+    from tests.conftest import make_vcr_cassette
+    user_id = getattr(pytestconfig, "_test_user_id", "user_abc123")
+    cassette_name = getattr(pytestconfig, "_vcr_cassette", None)
+    loop = asyncio.get_event_loop()
+
+    if cassette_name:
+        with make_vcr_cassette(cassette_name):
+            reply = loop.run_until_complete(agent.run(user_id, text))
+    else:
+        reply = loop.run_until_complete(agent.run(user_id, text))
+
+    pytestconfig._reply = reply
+
+
+# ── Then: agent-specific assertions ─────────────────────────────────────────
+
+@then(parsers.parse('agent created a session for user "{user_id}"'))
+def check_session_created(user_id):
+    from orchestrator.agent import agent
+    assert agent._active_conv.get(user_id) is not None, \
+        f"Expected active session to be set for {user_id}"
+
+
+@then(parsers.parse('no session is created for user "{user_id}"'))
+def check_no_session(user_id):
+    from orchestrator.agent import agent
+    assert agent._active_conv.get(user_id) is None, \
+        f"Expected no active session for {user_id}, got {agent._active_conv.get(user_id)}"

tests/step_defs/test_commands.py

@@ -0,0 +1,79 @@
+"""
+Step definitions for all slash command features.
+"""
+from __future__ import annotations
+
+import time
+
+from pytest_bdd import scenarios, given, when, then, parsers
+
+from tests.step_defs.common_steps import *  # noqa: F401,F403 — import shared steps
+
+scenarios(
+    "../features/commands/new.feature",
+    "../features/commands/status.feature",
+    "../features/commands/switch.feature",
+    "../features/commands/close.feature",
+    "../features/commands/direct_smart.feature",
+    "../features/commands/shell.feature",
+    "../features/commands/remind.feature",
+    "../features/commands/tasks.feature",
+    "../features/commands/nodes.feature",
+    "../features/commands/help.feature",
+    "../features/commands/perm.feature",
+)
+
+
+# ── When: send slash command ─────────────────────────────────────────────────
+
+@when(parsers.parse('user sends "{text}"'))
+def send_command(text, pytestconfig, feishu_calls, mock_run_claude):
+    import asyncio
+    from bot.commands import handle_command
+    user_id = getattr(pytestconfig, "_test_user_id", "user_abc123")
+    reply = asyncio.get_event_loop().run_until_complete(handle_command(user_id, text))
+    pytestconfig._reply = reply
+
+
+# ── Given: task runner state ─────────────────────────────────────────────────
+
+@given(parsers.parse('there is a running task "{task_id}" described as "{desc}"'))
+def add_running_task(task_id, desc):
+    from agent.task_runner import task_runner, BackgroundTask, TaskStatus
+    task = BackgroundTask(
+        task_id=task_id,
+        description=desc,
+        started_at=time.time(),
+        status=TaskStatus.RUNNING,
+    )
+    task_runner._tasks[task_id] = task
+
+
+@given(parsers.parse('there is a completed task "{task_id}" described as "{desc}"'))
+def add_completed_task(task_id, desc):
+    from agent.task_runner import task_runner, BackgroundTask, TaskStatus
+    now = time.time()
+    task = BackgroundTask(
+        task_id=task_id,
+        description=desc,
+        started_at=now - 5,
+        status=TaskStatus.COMPLETED,
+        completed_at=now,
+        result="success",
+    )
+    task_runner._tasks[task_id] = task
+
+
+@given(parsers.parse('there is a failed task "{task_id}" described as "{desc}"'))
+def add_failed_task(task_id, desc):
+    from agent.task_runner import task_runner, BackgroundTask, TaskStatus
+    now = time.time()
+    task = BackgroundTask(
+        task_id=task_id,
+        description=desc,
+        started_at=now - 3,
+        status=TaskStatus.FAILED,
+        completed_at=now,
+        error="subprocess failed",
+    )
+    task_runner._tasks[task_id] = task