| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Integration tests for MCP tool support in the Responses API.""" |
| 4 | |
| 5 | from __future__ import annotations |
| 6 | |
| 7 | from types import SimpleNamespace |
| 8 | from unittest.mock import AsyncMock, MagicMock |
| 9 | |
| 10 | import pytest |
| 11 | import pytest_asyncio |
| 12 | from openai import OpenAI |
| 13 | from openai_harmony import Message, ToolDescription, ToolNamespaceConfig |
| 14 | |
| 15 | import vllm.entrypoints.mcp.tool_server as tool_server_module |
| 16 | from tests.utils import RemoteOpenAIServer |
| 17 | from vllm.entrypoints.mcp.tool_server import ( |
| 18 | MCPToolServer, |
| 19 | init_tool_server, |
| 20 | post_process_tools_description, |
| 21 | ) |
| 22 | |
| 23 | from .conftest import ( |
| 24 | BASE_TEST_ENV, |
| 25 | events_contain_type, |
| 26 | log_response_diagnostics, |
| 27 | retry_for_tool_call, |
| 28 | retry_streaming_for, |
| 29 | validate_streaming_event_stack, |
| 30 | ) |
| 31 | |
| 32 | MODEL_NAME = "openai/gpt-oss-20b" |
| 33 | |
| 34 | _BASE_SERVER_ARGS = [ |
| 35 | "--enforce-eager", |
| 36 | "--tool-server", |
| 37 | "demo", |
| 38 | "--max_model_len", |
| 39 | "5000", |
| 40 | ] |
| 41 | |
| 42 | _PYTHON_TOOL_INSTRUCTION = ( |
| 43 | "You must use the Python tool to execute code. Never simulate execution." |
| 44 | ) |
| 45 | |
| 46 | |
| 47 | @pytest.mark.asyncio |
| 48 | async def test_init_tool_server_without_configuration(): |
| 49 | args = SimpleNamespace(tool_server=None) |
| 50 | |
| 51 | assert await init_tool_server(args) is None |
| 52 | |
| 53 | |
| 54 | @pytest.mark.asyncio |
| 55 | async def test_init_tool_server_with_mcp_url(monkeypatch: pytest.MonkeyPatch): |
| 56 | server = MagicMock() |
| 57 | server.add_tool_server = AsyncMock() |
| 58 | monkeypatch.setattr(tool_server_module, "MCPToolServer", lambda: server) |
| 59 | |
| 60 | result = await init_tool_server(SimpleNamespace(tool_server="http://mcp.test")) |
| 61 | |
| 62 | assert result is server |
| 63 | server.add_tool_server.assert_awaited_once_with("http://mcp.test") |
| 64 | |
| 65 | |
| 66 | @pytest.mark.skip_global_cleanup |
| 67 | class TestMCPToolServerUnit: |
| 68 | """Test MCP tool description processing and filtering. |
| 69 | |
| 70 | Note: The wildcard "*" is normalized to None by |
| 71 | _extract_allowed_tools_from_mcp_requests before reaching this layer, |
| 72 | so we only test None and specific tool filtering here. |
| 73 | See responses/test_serving_responses.py for "*" normalization tests. |
| 74 | """ |
| 75 | |
| 76 | def test_get_tool_description(self): |
| 77 | pytest.importorskip("mcp") |
| 78 | |
| 79 | server = MCPToolServer() |
| 80 | tool1 = ToolDescription.new( |
| 81 | name="tool1", description="First", parameters={"type": "object"} |
| 82 | ) |
| 83 | tool2 = ToolDescription.new( |
| 84 | name="tool2", description="Second", parameters={"type": "object"} |
| 85 | ) |
| 86 | tool3 = ToolDescription.new( |
| 87 | name="tool3", description="Third", parameters={"type": "object"} |
| 88 | ) |
| 89 | |
| 90 | server.harmony_tool_descriptions = { |
| 91 | "test_server": ToolNamespaceConfig( |
| 92 | name="test_server", |
| 93 | description="test", |
| 94 | tools=[tool1, tool2, tool3], |
| 95 | ) |
| 96 | } |
| 97 | |
| 98 | # Nonexistent server |
| 99 | assert server.get_tool_description("nonexistent") is None |
| 100 | |
| 101 | # None (no filter) - returns all tools |
| 102 | result = server.get_tool_description("test_server", allowed_tools=None) |
| 103 | assert len(result.tools) == 3 |
| 104 | |
| 105 | # Filter to specific tools |
| 106 | result = server.get_tool_description( |
| 107 | "test_server", allowed_tools=["tool1", "tool3"] |
| 108 | ) |
| 109 | assert len(result.tools) == 2 |
| 110 | assert result.tools[0].name == "tool1" |
| 111 | assert result.tools[1].name == "tool3" |
| 112 | |
| 113 | # Single tool |
| 114 | result = server.get_tool_description("test_server", allowed_tools=["tool2"]) |
| 115 | assert len(result.tools) == 1 |
| 116 | assert result.tools[0].name == "tool2" |
| 117 | |
| 118 | # No matching tools - returns None |
| 119 | result = server.get_tool_description( |
| 120 | "test_server", allowed_tools=["nonexistent"] |
| 121 | ) |
| 122 | assert result is None |
| 123 | |
| 124 | # Empty list - returns None |
| 125 | assert server.get_tool_description("test_server", allowed_tools=[]) is None |
| 126 | |
| 127 | def test_post_process_tools_description_uses_trim_schema_result(self, monkeypatch): |
| 128 | mcp_types = pytest.importorskip("mcp.types") |
| 129 | tool = mcp_types.Tool( |
| 130 | name="tool", |
| 131 | input_schema={"type": "object"}, |
| 132 | ) |
| 133 | result = mcp_types.ListToolsResult(tools=[tool]) |
| 134 | trimmed_schema = {"type": "object", "properties": {}} |
| 135 | |
| 136 | monkeypatch.setattr( |
| 137 | "vllm.entrypoints.mcp.tool_server.trim_schema", |
| 138 | lambda _schema: trimmed_schema, |
| 139 | ) |
| 140 | |
| 141 | post_process_tools_description(result) |
| 142 | |
| 143 | assert tool.input_schema is trimmed_schema |
| 144 | |
| 145 | def test_post_process_tools_description_honors_meta_opt_out(self): |
| 146 | mcp_types = pytest.importorskip("mcp.types") |
| 147 | included = mcp_types.Tool( |
| 148 | name="included", |
| 149 | input_schema={"type": "object"}, |
| 150 | ) |
| 151 | excluded = mcp_types.Tool( |
| 152 | name="excluded", |
| 153 | input_schema={"type": "object"}, |
| 154 | meta={"include_in_prompt": False}, |
| 155 | ) |
| 156 | result = mcp_types.ListToolsResult(tools=[included, excluded]) |
| 157 | |
| 158 | post_process_tools_description(result) |
| 159 | |
| 160 | assert result.tools == [included] |
| 161 | |
| 162 | def test_builtin_tools_consistency(self): |
| 163 | """MCP_BUILTIN_TOOLS must match BUILTIN_TOOL_TO_MCP_SERVER_LABEL values.""" |
| 164 | from vllm.entrypoints.openai.parser.harmony_utils import ( |
| 165 | BUILTIN_TOOL_TO_MCP_SERVER_LABEL, |
| 166 | MCP_BUILTIN_TOOLS, |
| 167 | ) |
| 168 | |
| 169 | assert set(BUILTIN_TOOL_TO_MCP_SERVER_LABEL.values()) == MCP_BUILTIN_TOOLS, ( |
| 170 | f"MCP_BUILTIN_TOOLS {MCP_BUILTIN_TOOLS} does not match " |
| 171 | f"BUILTIN_TOOL_TO_MCP_SERVER_LABEL values " |
| 172 | f"{set(BUILTIN_TOOL_TO_MCP_SERVER_LABEL.values())}" |
| 173 | ) |
| 174 | |
| 175 | |
| 176 | class TestMCPEnabled: |
| 177 | """Tests that require MCP tools to be enabled via environment variable.""" |
| 178 | |
| 179 | @pytest.fixture(scope="class") |
| 180 | def mcp_enabled_server(self): |
| 181 | env_dict = { |
| 182 | **BASE_TEST_ENV, |
| 183 | "VLLM_ENABLE_RESPONSES_API_STORE": "1", |
| 184 | "PYTHON_EXECUTION_BACKEND": "dangerously_use_uv", |
| 185 | "VLLM_GPT_OSS_SYSTEM_TOOL_MCP_LABELS": ("code_interpreter,container"), |
| 186 | "VLLM_GPT_OSS_HARMONY_SYSTEM_INSTRUCTIONS": "1", |
| 187 | } |
| 188 | with RemoteOpenAIServer( |
| 189 | MODEL_NAME, list(_BASE_SERVER_ARGS), env_dict=env_dict |
| 190 | ) as remote_server: |
| 191 | yield remote_server |
| 192 | |
| 193 | @pytest_asyncio.fixture |
| 194 | async def client(self, mcp_enabled_server): |
| 195 | async with mcp_enabled_server.get_async_client() as async_client: |
| 196 | yield async_client |
| 197 | |
| 198 | @staticmethod |
| 199 | def _mcp_tools_payload(*, allowed_tools: list[str] | None = None) -> list[dict]: |
| 200 | tool: dict = { |
| 201 | "type": "mcp", |
| 202 | "server_label": "code_interpreter", |
| 203 | "server_url": "http://localhost:8888", |
| 204 | } |
| 205 | if allowed_tools is not None: |
| 206 | tool["allowed_tools"] = allowed_tools |
| 207 | return [tool] |
| 208 | |
| 209 | @staticmethod |
| 210 | def _python_exec_input(code: str = "") -> str: |
| 211 | if not code: |
| 212 | code = "import random; print(random.randint(1, 1000000))" |
| 213 | return f"Execute the following code: {code}" |
| 214 | |
| 215 | @pytest.mark.asyncio |
| 216 | @pytest.mark.parametrize("model_name", [MODEL_NAME]) |
| 217 | async def test_mcp_tool_env_flag_enabled(self, client: OpenAI, model_name: str): |
| 218 | response = await retry_for_tool_call( |
| 219 | client, |
| 220 | model=model_name, |
| 221 | expected_tool_type="mcp_call", |
| 222 | input=self._python_exec_input(), |
| 223 | instructions=_PYTHON_TOOL_INSTRUCTION, |
| 224 | tools=self._mcp_tools_payload(), |
| 225 | temperature=0.0, |
| 226 | extra_body={"enable_response_messages": True}, |
| 227 | ) |
| 228 | |
| 229 | assert response.status == "completed" |
| 230 | log_response_diagnostics(response, label="MCP Enabled") |
| 231 | |
| 232 | tool_call_found = False |
| 233 | tool_response_found = False |
| 234 | for message in response.output_messages: |
| 235 | recipient = message.get("recipient") |
| 236 | if recipient and recipient.startswith("python"): |
| 237 | tool_call_found = True |
| 238 | assert message.get("channel") == "commentary" |
| 239 | parsed_message = Message.from_dict(message) |
| 240 | if parsed_message.author.role == "tool" and ( |
| 241 | parsed_message.author.name or "" |
| 242 | ).startswith("python"): |
| 243 | tool_response_found = True |
| 244 | assert message.get("channel") == "commentary" |
| 245 | |
| 246 | assert tool_call_found, ( |
| 247 | f"No Python tool call found. " |
| 248 | f"Output types: " |
| 249 | f"{[getattr(o, 'type', None) for o in response.output]}" |
| 250 | ) |
| 251 | assert tool_response_found, "No Python tool response found" |
| 252 | |
| 253 | for message in response.input_messages: |
| 254 | assert Message.from_dict(message).author.role != "developer" |
| 255 | |
| 256 | @pytest.mark.asyncio |
| 257 | @pytest.mark.parametrize("model_name", [MODEL_NAME]) |
| 258 | async def test_mcp_tool_with_allowed_tools_star( |
| 259 | self, client: OpenAI, model_name: str |
| 260 | ): |
| 261 | response = await retry_for_tool_call( |
| 262 | client, |
| 263 | model=model_name, |
| 264 | expected_tool_type="mcp_call", |
| 265 | input=self._python_exec_input(), |
| 266 | instructions=_PYTHON_TOOL_INSTRUCTION, |
| 267 | tools=self._mcp_tools_payload(allowed_tools=["*"]), |
| 268 | temperature=0.0, |
| 269 | extra_body={"enable_response_messages": True}, |
| 270 | ) |
| 271 | |
| 272 | assert response.status == "completed" |
| 273 | log_response_diagnostics(response, label="MCP Allowed Tools *") |
| 274 | |
| 275 | tool_call_found = any( |
| 276 | (msg.get("recipient") or "").startswith("python") |
| 277 | for msg in response.output_messages |
| 278 | ) |
| 279 | assert tool_call_found, ( |
| 280 | f"No Python tool call with '*'. " |
| 281 | f"Output types: " |
| 282 | f"{[getattr(o, 'type', None) for o in response.output]}" |
| 283 | ) |
| 284 | |
| 285 | @pytest.mark.asyncio |
| 286 | @pytest.mark.parametrize("model_name", [MODEL_NAME]) |
| 287 | async def test_mcp_tool_calling_streaming_types( |
| 288 | self, |
| 289 | pairs_of_event_types: dict[str, str], |
| 290 | client: OpenAI, |
| 291 | model_name: str, |
| 292 | ): |
| 293 | def _has_mcp_events(events: list) -> bool: |
| 294 | return events_contain_type(events, "mcp_call") |
| 295 | |
| 296 | events = await retry_streaming_for( |
| 297 | client, |
| 298 | model=model_name, |
| 299 | validate_events=_has_mcp_events, |
| 300 | input=("What is 123 * 456? Use Python to calculate the result."), |
| 301 | tools=[{"type": "mcp", "server_label": "code_interpreter"}], |
| 302 | instructions=_PYTHON_TOOL_INSTRUCTION, |
| 303 | temperature=0.0, |
| 304 | ) |
| 305 | |
| 306 | validate_streaming_event_stack(events, pairs_of_event_types) |