fix(tests): resolve CI mock server issues and improve tool call handling

Several critical fixes were implemented to resolve test failures in CI environments:
1. Updated MockLLMServer to properly use respx_mock fixture instead of global respx,
   ensuring HTTP mocks work consistently across local and CI environments
2. Modified tool call argument handling to check tc_args is not None rather than truthy
3. Improved scenario matching logic to check all user messages instead of just the last
4. Added proper async handling for Google ADK integration tests
5. Fixed tool call argument formatting to use "{}" instead of empty strings

These changes resolve infinite loop issues with Google ADK tests and ensure
consistent behavior across different test environments.

fix(integration): 修复 CI mock 服务器问题并改进工具调用处理

实施了几项关键修复以解决 CI 环境中的测试失败：
1. 更新 MockLLMServer 以正确使用 respx_mock fixture 而不是全局 respx，
   确保 HTTP mocks 在本地和 CI 环境中一致工作
2. 修改工具调用参数处理以检查 tc_args 不为 None 而不是真值
3. 改进场景匹配逻辑以检查所有用户消息而不是仅最后一条
4. 为 Google ADK 集成测试添加了适当的异步处理
5. 修复工具调用参数格式为使用"{}"而不是空字符串

这些更改解决了 Google ADK 测试的无限循环问题并确保
跨不同测试环境的一致行为。

Change-Id: I46c8b4faedb357f8a4a2149756f4cab121c2fb8e
Signed-off-by: OhYee <oyohyee@oyohyee.com>

Author: OhYee

agentrun/integration/langgraph/agent_converter.py

@@ -495,7 +495,7 @@ def _convert_stream_updates_event(
                         if tc_id:
                             # 发送带有完整参数的 TOOL_CALL_CHUNK
                             args_str = ""
-                            if tc_args:
+                            if tc_args is not None:
                                 args_str = (
                                     AgentRunConverter._safe_json_dumps(tc_args)
                                     if isinstance(tc_args, dict)
@@ -570,7 +570,7 @@ def _convert_stream_values_event(
                 if tc_id:
                     # 发送带有完整参数的 TOOL_CALL_CHUNK
                     args_str = ""
-                    if tc_args:
+                    if tc_args is not None:
                         args_str = (
                             AgentRunConverter._safe_json_dumps(tc_args)
                             if isinstance(tc_args, dict)
@@ -694,7 +694,7 @@ def _convert_astream_events_event(
                             tool_name_to_call_ids[tc_name].append(tc_id)
                         # 第一个 chunk 包含 id 和 name
                         args_delta = ""
-                        if tc_args:
+                        if tc_args is not None:
                             args_delta = (
                                 AgentRunConverter._safe_json_dumps(tc_args)
                                 if isinstance(tc_args, (dict, list))
@@ -708,7 +708,7 @@ def _convert_astream_events_event(
                                 "args_delta": args_delta,
                             },
                         )
-                    elif tc_args:
+                    elif tc_args is not None:
                         # 后续 chunk 只有 args_delta
                         args_delta = (
                             AgentRunConverter._safe_json_dumps(tc_args)
@@ -765,7 +765,7 @@ def _convert_astream_events_event(
                                     ).append(tc_id)
 
                                 args_delta = ""
-                                if tc_args:
+                                if tc_args is not None:
                                     args_delta = (
                                         AgentRunConverter._safe_json_dumps(
                                             tc_args

tests/unittests/integration/conftest.py

@@ -320,9 +320,12 @@ def shared_mock_server(monkeypatch: Any, respx_mock: Any) -> MockLLMServer:
     """提供共享的 Mock LLM Server
 
     预配置了默认场景。
+
+    关键修复：传入 respx_mock fixture 给 MockLLMServer
+    - 确保 HTTP mock 在所有环境（本地/CI）中一致生效
     """
     server = MockLLMServer(expect_tools=True, validate_tools=False)
-    server.install(monkeypatch)
+    server.install(monkeypatch, respx_mock)
     server.add_default_scenarios()
     return server
 

tests/unittests/integration/langchain/test_agent_invoke_methods.py

@@ -400,7 +400,9 @@ def _normalize_agui_event(event: Dict[str, Any]) -> Dict[str, Any]:
         },
         {
             "type": "TOOL_CALL_ARGS",
-            "delta": "",
+            # 空参数在 LangGraph 中表现为 "{}" (Node.js SDK) 或 根据转换逻辑可能为空字符串
+            # 但当前 mock server 返回 "{}"，转换器保留了它
+            "delta": "{}",
             "hasToolCallId": True,
         },
         {"type": "TOOL_CALL_END", "hasToolCallId": True},
@@ -551,6 +553,15 @@ def _normalize_openai_stream(
             }],
             "finish_reason": None,
         },
+        {
+            "object": "chat.completion.chunk",
+            "tool_calls": [{
+                "name": None,
+                "arguments": "{}",
+                "has_id": False,
+            }],
+            "finish_reason": None,
+        },
         {
             "object": "chat.completion.chunk",
             "delta_role": "assistant",
@@ -612,7 +623,7 @@ def _normalize_openai_nonstream(resp: Dict[str, Any]) -> Dict[str, Any]:
         "content": "工具结果已收到: 2024-01-01 12:00:00",
         "tool_calls": [{
             "name": "get_time",
-            "arguments": "",
+            "arguments": "{}",
             "has_id": True,
         }],
         "finish_reason": "tool_calls",

tests/unittests/integration/mock_llm_server.py

@@ -43,7 +43,7 @@ class MockLLMServer:
     使用方式:
         # 基本用法
         server = MockLLMServer()
-        server.install(monkeypatch)
+        server.install(monkeypatch, respx_mock)  # 需要传入 respx_mock
 
         # 添加自定义场景
         server.add_scenario(Scenarios.simple_chat("你好", "你好！"))
@@ -67,15 +67,22 @@ class MockLLMServer:
     validate_tools: bool = True
     """是否验证工具格式（默认 True）"""
 
-    def install(self, monkeypatch: Any) -> "MockLLMServer":
+    _respx_router: Any = field(default=None, init=False, repr=False)
+    """内部使用的 respx router 实例"""
+
+    def install(
+        self, monkeypatch: Any, respx_mock: Any = None
+    ) -> "MockLLMServer":
         """安装所有 mock
 
         Args:
             monkeypatch: pytest monkeypatch fixture
+            respx_mock: pytest respx_mock fixture（必须传入以确保 mock 生效）
 
         Returns:
             self: 返回自身以便链式调用
         """
+        self._respx_router = respx_mock
         self._patch_model_info(monkeypatch)
         self._patch_litellm(monkeypatch)
         self._setup_respx()
@@ -240,7 +247,20 @@ async def fake_acompletion(*args: Any, **kwargs: Any) -> ModelResponse:
             pass  # google.adk not installed
 
     def _setup_respx(self):
-        """设置 respx HTTP mock"""
+        """设置 respx HTTP mock
+
+        关键修复：使用 pytest-respx fixture 提供的 router 而不是全局 respx
+
+        问题背景：
+        - 之前直接使用全局 respx.route() 在 CI 环境中不生效
+        - 全局 respx router 在某些环境中可能没有正确初始化
+        - 导致 HTTP 请求没有被拦截，Google ADK 发送真实请求
+
+        解决方案：
+        - 使用 pytest-respx 提供的 respx_mock fixture
+        - 通过 install() 方法传入 respx_mock
+        - 确保 mock 在所有环境中一致生效
+        """
 
         def extract_payload(request: Any) -> Dict[str, Any]:
             try:
@@ -274,7 +294,10 @@ def build_response(request: Any, route: Any) -> respx.MockResponse:
                 )
             return respx.MockResponse(status_code=200, json=response_json)
 
-        respx.route(url__startswith=self.base_url).mock(
+        # 关键修复：使用传入的 respx_router 而不是全局 respx
+        # 如果没有传入 respx_router，回退到全局 respx（向后兼容）
+        router = self._respx_router if self._respx_router is not None else respx
+        router.route(url__startswith=self.base_url).mock(
             side_effect=build_response
         )
 
@@ -304,6 +327,27 @@ def _build_response(
             tools_payload is not None,
         )
 
+        # 添加详细的消息日志，帮助调试框架的消息格式
+        for i, msg in enumerate(messages):
+            role = msg.get("role", "unknown")
+            content_preview = str(msg.get("content", ""))[:100]
+            logger.debug(
+                "Message[%d] role=%s, content_preview=%s",
+                i,
+                role,
+                content_preview,
+            )
+            if "tool_calls" in msg:
+                logger.debug(
+                    "Message[%d] has tool_calls: %s", i, msg.get("tool_calls")
+                )
+            if "tool_call_id" in msg:
+                logger.debug(
+                    "Message[%d] has tool_call_id: %s",
+                    i,
+                    msg.get("tool_call_id"),
+                )
+
         # 验证工具格式
         if self.validate_tools and self.expect_tools and tools_payload:
             self._assert_tools(tools_payload)
@@ -319,16 +363,19 @@ def _build_response(
             turn = scenario.get_response(messages)
             return turn.to_response()
 
-        # 默认逻辑：根据最后一条消息决定响应
+        # 默认逻辑：未匹配场景时使用
         return self._build_default_response(messages, tools_payload)
 
     def _build_default_response(
         self, messages: List[Dict], tools_payload: Optional[List]
     ) -> Dict[str, Any]:
         """构建默认响应（无场景匹配时使用）"""
-        last_role = messages[-1].get("role")
+        # 检查消息历史中是否已经有 tool 结果
+        # 这是关键修复：不只检查最后一条消息，而是检查整个历史
+        has_tool_results = any(msg.get("role") == "tool" for msg in messages)
 
-        if last_role == "tool":
+        if has_tool_results:
+            # 已经有 tool 结果，应该返回最终答案而不是再次调用工具
             return {
                 "id": "chatcmpl-mock-final",
                 "object": "chat.completion",
@@ -349,7 +396,7 @@ def _build_default_response(
                 },
             }
 
-        # 如果有工具，返回工具调用
+        # 如果有工具且未调用过，返回工具调用
         if tools_payload:
             return {
                 "id": "chatcmpl-mock-tools",

tests/unittests/integration/scenarios.py

@@ -113,12 +113,31 @@ def get_response(self, messages: List[Dict]) -> MockTurn:
         - 如果最后一条消息是 tool 类型，说明工具已执行，进入下一轮
         - 否则返回当前轮次
         """
+        import logging
+
+        logger = logging.getLogger(__name__)
+
         # 计算当前应该返回哪一轮
         tool_rounds = sum(1 for msg in messages if msg.get("role") == "tool")
 
+        logger.debug(
+            "Scenario '%s': Found %d tool messages, total turns: %d",
+            self.name,
+            tool_rounds,
+            len(self.turns),
+        )
+
         # 根据工具消息数量确定当前轮次
         # 每个工具响应对应一个轮次的推进
         current_idx = min(tool_rounds, len(self.turns) - 1)
+
+        logger.debug(
+            "Scenario '%s': Returning turn %d, has_tool_calls=%s",
+            self.name,
+            current_idx,
+            self.turns[current_idx].has_tool_calls(),
+        )
+
         return self.turns[current_idx]
 
     def reset(self):
@@ -145,12 +164,14 @@ def simple_chat(trigger: str, response: str) -> MockScenario:
         """
 
         def trigger_fn(messages: List[Dict]) -> bool:
-            # 查找最后一条用户消息
-            for msg in reversed(messages):
+            # 检查所有用户消息（任意一条包含trigger即匹配）
+            # 修复：不只检查最后一条，避免框架插入的额外消息干扰匹配
+            for msg in messages:
                 if msg.get("role") == "user":
                     content = msg.get("content", "")
                     if isinstance(content, str):
-                        return trigger in content
+                        if trigger in content:
+                            return True
                     elif isinstance(content, list):
                         # 处理 content 是列表的情况
                         for item in content:
@@ -188,11 +209,13 @@ def single_tool_call(
         """
 
         def trigger_fn(messages: List[Dict]) -> bool:
-            for msg in reversed(messages):
+            # 检查所有用户消息（任意一条包含trigger即匹配）
+            # 修复：避免框架插入的额外消息干扰匹配
+            for msg in messages:
                 if msg.get("role") == "user":
                     content = msg.get("content", "")
-                    if isinstance(content, str):
-                        return trigger in content
+                    if isinstance(content, str) and trigger in content:
+                        return True
             return False
 
         return MockScenario(
@@ -230,11 +253,13 @@ def multi_tool_calls(
         """
 
         def trigger_fn(messages: List[Dict]) -> bool:
-            for msg in reversed(messages):
+            # 检查所有用户消息（任意一条包含trigger即匹配）
+            # 修复：避免框架插入的额外消息干扰匹配
+            for msg in messages:
                 if msg.get("role") == "user":
                     content = msg.get("content", "")
-                    if isinstance(content, str):
-                        return trigger in content
+                    if isinstance(content, str) and trigger in content:
+                        return True
             return False
 
         return MockScenario(
@@ -273,11 +298,13 @@ def multi_round_tools(
         """
 
         def trigger_fn(messages: List[Dict]) -> bool:
-            for msg in reversed(messages):
+            # 检查所有用户消息（任意一条包含trigger即匹配）
+            # 修复：避免框架插入的额外消息干扰匹配
+            for msg in messages:
                 if msg.get("role") == "user":
                     content = msg.get("content", "")
-                    if isinstance(content, str):
-                        return trigger in content
+                    if isinstance(content, str) and trigger in content:
+                        return True
             return False
 
         turns = []

tests/unittests/integration/test_agentscope.py

@@ -124,9 +124,13 @@ class TestAgentScopeIntegration(AgentScopeTestMixin):
 
     @pytest.fixture
     def mock_server(self, monkeypatch: Any, respx_mock: Any) -> MockLLMServer:
-        """创建并安装 Mock LLM Server"""
+        """创建并安装 Mock LLM Server
+
+        关键修复：传入 respx_mock fixture 给 MockLLMServer
+        - 确保 HTTP mock 在所有环境（本地/CI）中一致生效
+        """
         server = MockLLMServer(expect_tools=True, validate_tools=False)
-        server.install(monkeypatch)
+        server.install(monkeypatch, respx_mock)
         server.add_default_scenarios()
         return server
 

tests/unittests/integration/test_crewai.py

@@ -123,9 +123,13 @@ class TestCrewAIIntegration(CrewAITestMixin):
 
     @pytest.fixture
     def mock_server(self, monkeypatch: Any, respx_mock: Any) -> MockLLMServer:
-        """创建并安装 Mock LLM Server"""
+        """创建并安装 Mock LLM Server
+
+        关键修复：传入 respx_mock fixture 给 MockLLMServer
+        - 确保 HTTP mock 在所有环境（本地/CI）中一致生效
+        """
         server = MockLLMServer(expect_tools=True, validate_tools=False)
-        server.install(monkeypatch)
+        server.install(monkeypatch, respx_mock)
         server.add_default_scenarios()
         return server