"""Agent 图缓存行为测试。""" from contextlib import ExitStack from datetime import datetime, timedelta from types import SimpleNamespace from unittest.mock import AsyncMock, patch import pytest from langchain_core.messages import AIMessage, HumanMessage from app.agent import MoviePilotAgent, ReplyMode, _CompiledAgentBundle from app.agent.mcp import AgentMcpToolSpec from app.agent.middleware.policy import AgentPolicyMiddleware from app.agent.tools.catalog import ( ToolCatalogSnapshot, ToolIdentityAmbiguousError, ) from app.agent.tools.impl.mcp import create_external_mcp_tools from app.core.config import settings from app.schemas.agent import AgentMcpServerConfig @pytest.fixture def anyio_backend(): """使用 asyncio 后端运行 anyio 异步测试。""" return "asyncio" class _FakeGraphState: """提供 LangGraph get_state 测试替身。""" def __init__(self, messages): """保存测试消息状态。""" self.values = {"messages": messages} class _CapturingAgent: """捕获传入消息的非流式 Agent 测试替身。""" def __init__(self): """初始化捕获容器。""" self.payload = None async def ainvoke(self, payload, config=None): """记录 Agent 调用输入。""" self.payload = payload def get_state(self, _config): """返回包含最终 AI 回复的图状态。""" return _FakeGraphState([AIMessage(content="ok")]) @pytest.mark.anyio async def test_create_agent_reuses_cached_graph_when_signature_matches(): """构造签名一致时应直接复用已编译 Agent 图。""" cached_graph = object() catalog = ToolCatalogSnapshot.from_tools( [], plugin_revision=0, factory_revision="factory-v1" ) agent = MoviePilotAgent(session_id="cache-hit", user_id="user-1") agent._compiled_agent_bundle = _CompiledAgentBundle( signature=("sig",), agent=cached_graph, streaming=False, created_at=datetime.now(), tool_catalog=catalog, subagent_catalog=catalog, plugin_revision=0, mcp_config_signature="mcp-config", catalog_checked_at=datetime.now(), ) with patch.object( agent, "_agent_bundle_signature", new=AsyncMock(return_value=("sig",)), ), patch( "app.agent.PluginManager.get_plugin_agent_tools_revision", return_value=0, ), patch( "app.agent.agent_mcp_manager.config_signature", return_value="mcp-config", ), patch("app.agent.create_agent") as create_agent: graph = await agent._create_agent(streaming=False) assert graph is cached_graph assert agent._last_agent_cache_hit is True create_agent.assert_not_called() @pytest.mark.anyio async def test_fresh_catalog_cache_hit_skips_tool_and_mcp_discovery() -> None: """目录仍在 freshness 窗口内时，缓存命中不得重建工具或访问 MCP。""" cached_graph = object() catalog = ToolCatalogSnapshot.from_tools( [], plugin_revision=0, factory_revision="factory-v1" ) agent = MoviePilotAgent(session_id="catalog-cache-hit", user_id="user-1") agent._compiled_agent_bundle = _CompiledAgentBundle( signature=("sig",), agent=cached_graph, streaming=False, created_at=datetime.now(), tool_catalog=catalog, subagent_catalog=catalog, plugin_revision=0, mcp_config_signature="mcp-config", catalog_checked_at=datetime.now(), ) with patch.object( agent, "_agent_bundle_signature", new=AsyncMock(return_value=("sig",)), ), patch.object( agent, "_initialize_local_tool_catalogs", side_effect=AssertionError("tool catalog rebuilt"), ), patch( "app.agent.PluginManager.get_plugin_agent_tools_revision", return_value=0, ), patch( "app.agent.agent_mcp_manager.config_signature", return_value="mcp-config", ), patch( "app.agent.agent_mcp_manager.list_enabled_tool_specs", new=AsyncMock(side_effect=AssertionError("MCP discovery called")), ): graph = await agent._create_agent(streaming=False) assert graph is cached_graph assert agent._last_agent_cache_hit is True @pytest.mark.anyio async def test_expired_unchanged_catalog_renews_freshness() -> None: """过期目录复核后若签名未变，应续期缓存而不是每轮重复 discovery。""" cached_graph = object() catalog = ToolCatalogSnapshot.from_tools( [], plugin_revision=0, factory_revision="factory-v1" ) expired_at = datetime.now() - timedelta(minutes=5) agent = MoviePilotAgent(session_id="catalog-refresh", user_id="user-1") agent._compiled_agent_bundle = _CompiledAgentBundle( signature=("sig",), agent=cached_graph, streaming=False, created_at=datetime.now(), tool_catalog=catalog, subagent_catalog=catalog, plugin_revision=0, mcp_config_signature="mcp-config", catalog_checked_at=expired_at, ) fake_llm = SimpleNamespace( _llm_type="openai-chat", model="fake", profile={"max_input_tokens": 64000}, ) with patch.object( agent, "_agent_bundle_signature", new=AsyncMock(return_value=("sig",)), ), patch.object( agent, "_initialize_local_tool_catalogs", return_value=(catalog, catalog), ), patch.object( agent, "_initialize_mcp_tools", new=AsyncMock(return_value=[]), ), patch.object( agent, "_initialize_subagent_mcp_tools", new=AsyncMock(return_value=[]), ), patch.object( agent, "_initialize_llm", new=AsyncMock(return_value=fake_llm), ), patch.object( agent, "_sync_model_profile", ), patch( "app.agent.ServerToolRegistry.resolve_web_search", return_value=SimpleNamespace(use_local_web_search=True), ), patch( "app.agent.LLMHelper.get_server_tools", return_value=[], ), patch( "app.agent.prompt_manager.get_agent_prompt", return_value="prompt", ), patch( "app.agent.SkillsMiddleware", return_value=SimpleNamespace(name="skills", tools=[]), ), patch( "app.agent.create_subagent_middlewares", return_value=([], []), ), patch( "app.agent.PluginManager.get_plugin_agent_tools_revision", return_value=0, ), patch( "app.agent.agent_mcp_manager.config_signature", return_value="mcp-config", ), patch( "app.agent.agent_mcp_manager.list_enabled_tool_specs", new=AsyncMock(return_value=[]), ): graph = await agent._create_agent(streaming=False) assert graph is cached_graph assert agent._compiled_agent_bundle.catalog_checked_at > expired_at @pytest.mark.anyio async def test_agent_bundle_signature_changes_with_temperature(monkeypatch) -> None: """温度配置变化时应使会话内 Agent 图缓存失效。""" agent = MoviePilotAgent(session_id="temperature-change", user_id="user-1") runtime_config = { "provider": "openai", "model": "gpt-test", "api_key": "test-key", "base_url": "https://llm.example.com/v1", "base_url_preset": None, "user_agent": None, "use_proxy": False, "thinking_level": "off", "api_protocol": "auto", } with patch.object( agent, "_resolve_llm_runtime_config", new=AsyncMock(return_value=runtime_config), ): monkeypatch.setattr(settings, "LLM_TEMPERATURE", 0.3) initial_signature = await agent._agent_bundle_signature(streaming=False) monkeypatch.setattr(settings, "LLM_TEMPERATURE", 1.0) updated_signature = await agent._agent_bundle_signature(streaming=False) assert updated_signature != initial_signature @pytest.mark.anyio async def test_agent_bundle_signature_changes_with_tool_catalog() -> None: """工具目录 revision 必须参与会话内 Agent 图缓存签名。""" agent = MoviePilotAgent(session_id="tool-revision", user_id="user-1") runtime_config = {"provider": "openai", "model": "gpt-test"} first_catalog = ToolCatalogSnapshot.from_tools( [], plugin_revision=1, factory_revision="factory-v1" ) second_catalog = ToolCatalogSnapshot.from_tools( [], plugin_revision=2, factory_revision="factory-v1" ) with patch.object( agent, "_resolve_llm_runtime_config", new=AsyncMock(return_value=runtime_config), ): first_signature = await agent._agent_bundle_signature( streaming=False, tool_catalog=first_catalog, subagent_catalog=first_catalog, ) second_signature = await agent._agent_bundle_signature( streaming=False, tool_catalog=second_catalog, subagent_catalog=second_catalog, ) assert second_signature != first_signature @pytest.mark.anyio @pytest.mark.parametrize("max_tools", [0, 5]) async def test_graph_keeps_mcp_first_winner_and_catalogs_all_collisions( max_tools: int, ) -> None: """主图与子图保持 MCP first-wins，严格目录覆盖全部客户端工具。""" servers = [ AgentMcpServerConfig( id=server_id, name="Shared Name", transport="stdio", command=server_id, ) for server_id in ("one", "two") ] specs = [ AgentMcpToolSpec( server=server, name="echo", agent_tool_name="shared_echo", description="echo", input_schema={"type": "object", "properties": {}}, ) for server in servers ] main_tools = await create_external_mcp_tools( session_id="session", user_id="user", specs=specs, ) subagent_tools = await create_external_mcp_tools( session_id="session", user_id="user", specs=specs, ) empty_catalog = ToolCatalogSnapshot.from_tools( [], plugin_revision=0, factory_revision="factory-v1" ) fake_llm = SimpleNamespace( _llm_type="openai-chat", model="fake", profile={"max_input_tokens": 64000}, ) skill_tool = SimpleNamespace( name="shared_echo", description="skill collision", args_schema={"type": "object", "properties": {}}, _agent_tool_source="middleware:skills", ) activity_tool = SimpleNamespace( name="query_activity_log", description="activity log", args_schema={"type": "object", "properties": {}}, _agent_tool_source="middleware:activity_log", ) subagent_task_tool = SimpleNamespace( name="task", description="subagent task", args_schema={"type": "object", "properties": {}}, _agent_tool_source="middleware:subagents", ) captured = {} agent = MoviePilotAgent( session_id="mcp-collision", user_id="user", channel="web", source="test", ) def _capture_subagents(**kwargs): captured["subagent_tools"] = kwargs["tools"] captured["subagent_catalog"] = kwargs["catalog"] return [], [subagent_task_tool] def _capture_selector(**kwargs): captured["selection_tools"] = kwargs["selection_tools"] return SimpleNamespace(name="selector") def _capture_agent(**kwargs): captured["agent_tools"] = kwargs["tools"] captured["middlewares"] = kwargs["middleware"] return object() patchers = [ patch.object( agent, "_resolve_llm_runtime_config", new=AsyncMock(return_value={"provider": "openai", "model": "fake"}), ), patch.object( agent, "_initialize_local_tool_catalogs", return_value=(empty_catalog, empty_catalog), ), patch.object( agent, "_initialize_mcp_tools", new=AsyncMock(return_value=main_tools), ), patch.object( agent, "_initialize_subagent_mcp_tools", new=AsyncMock(return_value=subagent_tools), ), patch.object( agent, "_agent_bundle_signature", new=AsyncMock(return_value=("mcp-collision", max_tools)), ), patch.object( agent, "_initialize_llm", new=AsyncMock(return_value=fake_llm), ), patch.object(agent, "_sync_model_profile"), patch( "app.agent.PluginManager.get_plugin_agent_tools_revision", return_value=0, ), patch( "app.agent.agent_mcp_manager.config_signature", return_value="mcp-config", ), patch( "app.agent.agent_mcp_manager.list_enabled_tool_specs", new=AsyncMock(return_value=specs), ), patch( "app.agent.ServerToolRegistry.resolve_web_search", return_value=SimpleNamespace(use_local_web_search=True), ), patch("app.agent.LLMHelper.get_server_tools", return_value=[]), patch("app.agent.prompt_manager.get_agent_prompt", return_value="prompt"), patch( "app.agent.create_subagent_middlewares", side_effect=_capture_subagents, ), patch( "app.agent.MoviePilotToolFactory.get_tool_selector_always_include_names", return_value=[], ), patch( "app.agent.SkillsMiddleware", return_value=SimpleNamespace(name="skills", tools=[skill_tool]), ), patch( "app.agent.ActivityLogMiddleware", return_value=SimpleNamespace(name="activity", tools=[activity_tool]), ), patch( "app.agent.JobsMiddleware", return_value=SimpleNamespace(name="jobs"), ), patch( "app.agent.RuntimeConfigMiddleware", return_value=SimpleNamespace(name="runtime"), ), patch( "app.agent.MemoryMiddleware", return_value=SimpleNamespace(name="memory"), ), patch( "app.agent.SummarizationMiddleware", return_value=SimpleNamespace(name="summary"), ), patch( "app.agent.PatchToolCallsMiddleware", return_value=SimpleNamespace(name="patch"), ), patch( "app.agent.UsageMiddleware", return_value=SimpleNamespace(name="usage"), ), patch( "app.agent.ToolSelectorMiddleware", side_effect=_capture_selector, ), patch("app.agent.InMemorySaver", return_value=object()), patch("app.agent.create_agent", side_effect=_capture_agent), patch.object(settings, "LLM_MAX_TOOLS", max_tools), ] with ExitStack() as stack: for patcher in patchers: stack.enter_context(patcher) await agent._create_agent(streaming=False) assert captured["agent_tools"] == [main_tools[0], skill_tool, activity_tool] assert captured["subagent_tools"] == [subagent_tools[0]] if max_tools: assert captured["selection_tools"] == [ main_tools[0], skill_tool, activity_tool, subagent_task_tool, ] else: assert "selection_tools" not in captured policy_middleware = next( middleware for middleware in captured["middlewares"] if isinstance(middleware, AgentPolicyMiddleware) ) assert [ entry.source for entry in policy_middleware.catalog.collisions["shared_echo"] ] == ["mcp:one", "mcp:two", "middleware:skills"] assert ( policy_middleware.catalog.resolve_unique("query_activity_log").tool is activity_tool ) assert policy_middleware.catalog.resolve_unique("task").tool is subagent_task_tool with pytest.raises(ToolIdentityAmbiguousError, match="TOOL_IDENTITY_AMBIGUOUS"): policy_middleware.catalog.resolve_unique("shared_echo") assert [ entry.source for entry in captured["subagent_catalog"].collisions["shared_echo"] ] == ["mcp:one", "mcp:two"] @pytest.mark.anyio async def test_execute_agent_sends_only_latest_message_on_cache_hit(): """缓存命中时只把本轮新消息交给 LangGraph，避免重复提交历史。""" fake_graph = _CapturingAgent() agent = MoviePilotAgent(session_id="cache-hit", user_id="user-1") agent.reply_mode = ReplyMode.CAPTURE_ONLY agent._tool_context = {"user_reply_sent": False} agent._streamed_output = "" agent._should_stream = lambda: False agent.stream_handler = SimpleNamespace( stop_streaming=AsyncMock(return_value=(False, "")) ) async def _create_agent(streaming=False): """模拟缓存命中后的 Agent 创建结果。""" agent._last_agent_cache_hit = True return fake_graph agent._create_agent = _create_agent messages = [HumanMessage(content="上一轮"), HumanMessage(content="本轮")] with patch("app.agent.eventmanager.send_event"): await agent._execute_agent(messages) assert agent._streamed_output == "ok" assert fake_graph.payload["messages"] == [messages[-1]]