[Frontend] Add MCP type support infrastructure to Responses API

Signed-off-by: Daniel Salib <danielsalib@meta.com>

Author: daniel-salib

tests/entrypoints/test_harmony_utils.py

@@ -2,6 +2,7 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 from openai.types.responses import ResponseFunctionToolCall, ResponseReasoningItem
+from openai.types.responses.response_output_item import McpCall
 from openai_harmony import Author, Message, Role, TextContent
 
 from vllm.entrypoints.harmony_utils import (
@@ -451,3 +452,167 @@ def test_has_custom_tools() -> None:
     assert has_custom_tools(
         {"web_search_preview", "code_interpreter", "container", "others"}
     )
+
+
+def test_parse_mcp_call_basic() -> None:
+    """Test that MCP calls are parsed with correct type and server_label."""
+    message = Message.from_role_and_content(Role.ASSISTANT, '{"path": "/tmp"}')
+    message = message.with_recipient("filesystem")
+    message = message.with_channel("commentary")
+
+    output_items = parse_output_message(message)
+
+    assert len(output_items) == 1
+    assert isinstance(output_items[0], McpCall)
+    assert output_items[0].type == "mcp_call"
+    assert output_items[0].name == "filesystem"
+    assert output_items[0].server_label == "filesystem"
+    assert output_items[0].arguments == '{"path": "/tmp"}'
+    assert output_items[0].status == "completed"
+
+
+def test_parse_mcp_call_dotted_recipient() -> None:
+    """Test that dotted recipients extract the tool name correctly."""
+    message = Message.from_role_and_content(Role.ASSISTANT, '{"cmd": "ls"}')
+    message = message.with_recipient("repo_browser.list")
+    message = message.with_channel("commentary")
+
+    output_items = parse_output_message(message)
+
+    assert len(output_items) == 1
+    assert isinstance(output_items[0], McpCall)
+    assert output_items[0].name == "list"
+    assert output_items[0].server_label == "repo_browser"
+
+
+def test_mcp_vs_function_call() -> None:
+    """Test that function calls are not parsed as MCP calls."""
+    func_message = Message.from_role_and_content(Role.ASSISTANT, '{"arg": "value"}')
+    func_message = func_message.with_recipient("functions.my_tool")
+    func_message = func_message.with_channel("commentary")
+
+    func_items = parse_output_message(func_message)
+
+    assert len(func_items) == 1
+    assert not isinstance(func_items[0], McpCall)
+    assert func_items[0].type == "function_call"
+
+
+def test_mcp_vs_builtin_tools() -> None:
+    """Test that built-in tools (python, container) are not parsed as MCP calls."""
+    # Test python (built-in tool) - should be reasoning, not MCP
+    python_message = Message.from_role_and_content(Role.ASSISTANT, "print('hello')")
+    python_message = python_message.with_recipient("python")
+    python_message = python_message.with_channel("commentary")
+
+    python_items = parse_output_message(python_message)
+
+    assert len(python_items) == 1
+    assert not isinstance(python_items[0], McpCall)
+    assert python_items[0].type == "reasoning"
+
+
+def test_parse_remaining_state_commentary_channel() -> None:
+    """Test parse_remaining_state with commentary channel and various recipients."""
+    from unittest.mock import Mock
+
+    from vllm.entrypoints.harmony_utils import parse_remaining_state
+
+    # Test 1: functions.* recipient → should return function tool call
+    parser_func = Mock()
+    parser_func.current_content = '{"arg": "value"}'
+    parser_func.current_role = Role.ASSISTANT
+    parser_func.current_channel = "commentary"
+    parser_func.current_recipient = "functions.my_tool"
+
+    func_items = parse_remaining_state(parser_func)
+
+    assert len(func_items) == 1
+    assert not isinstance(func_items[0], McpCall)
+    assert func_items[0].type == "function_call"
+    assert func_items[0].name == "my_tool"
+    assert func_items[0].status == "in_progress"
+
+    # Test 2: MCP tool (not builtin) → should return MCP call
+    parser_mcp = Mock()
+    parser_mcp.current_content = '{"path": "/tmp"}'
+    parser_mcp.current_role = Role.ASSISTANT
+    parser_mcp.current_channel = "commentary"
+    parser_mcp.current_recipient = "filesystem"
+
+    mcp_items = parse_remaining_state(parser_mcp)
+
+    assert len(mcp_items) == 1
+    assert isinstance(mcp_items[0], McpCall)
+    assert mcp_items[0].type == "mcp_call"
+    assert mcp_items[0].name == "filesystem"
+    assert mcp_items[0].server_label == "filesystem"
+    assert mcp_items[0].status == "in_progress"
+
+    # Test 3: Built-in tool (python)
+    # should NOT return MCP call, falls through to reasoning
+    parser_builtin = Mock()
+    parser_builtin.current_content = "print('hello')"
+    parser_builtin.current_role = Role.ASSISTANT
+    parser_builtin.current_channel = "commentary"
+    parser_builtin.current_recipient = "python"
+
+    builtin_items = parse_remaining_state(parser_builtin)
+
+    # Should fall through to reasoning logic
+    assert len(builtin_items) == 1
+    assert not isinstance(builtin_items[0], McpCall)
+    assert builtin_items[0].type == "reasoning"
+
+
+def test_parse_remaining_state_analysis_channel() -> None:
+    """Test parse_remaining_state with analysis channel and various recipients."""
+    from unittest.mock import Mock
+
+    from vllm.entrypoints.harmony_utils import parse_remaining_state
+
+    # Test 1: functions.* recipient → should return function tool call
+    parser_func = Mock()
+    parser_func.current_content = '{"arg": "value"}'
+    parser_func.current_role = Role.ASSISTANT
+    parser_func.current_channel = "analysis"
+    parser_func.current_recipient = "functions.my_tool"
+
+    func_items = parse_remaining_state(parser_func)
+
+    assert len(func_items) == 1
+    assert not isinstance(func_items[0], McpCall)
+    assert func_items[0].type == "function_call"
+    assert func_items[0].name == "my_tool"
+    assert func_items[0].status == "in_progress"
+
+    # Test 2: MCP tool (not builtin) → should return MCP call
+    parser_mcp = Mock()
+    parser_mcp.current_content = '{"query": "test"}'
+    parser_mcp.current_role = Role.ASSISTANT
+    parser_mcp.current_channel = "analysis"
+    parser_mcp.current_recipient = "database"
+
+    mcp_items = parse_remaining_state(parser_mcp)
+
+    assert len(mcp_items) == 1
+    assert isinstance(mcp_items[0], McpCall)
+    assert mcp_items[0].type == "mcp_call"
+    assert mcp_items[0].name == "database"
+    assert mcp_items[0].server_label == "database"
+    assert mcp_items[0].status == "in_progress"
+
+    # Test 3: Built-in tool (container)
+    # should NOT return MCP call, falls through to reasoning
+    parser_builtin = Mock()
+    parser_builtin.current_content = "docker run"
+    parser_builtin.current_role = Role.ASSISTANT
+    parser_builtin.current_channel = "analysis"
+    parser_builtin.current_recipient = "container"
+
+    builtin_items = parse_remaining_state(parser_builtin)
+
+    # Should fall through to reasoning logic
+    assert len(builtin_items) == 1
+    assert not isinstance(builtin_items[0], McpCall)
+    assert builtin_items[0].type == "reasoning"

vllm/entrypoints/harmony_utils.py

@@ -19,6 +19,7 @@
     ActionSearch,
     ResponseFunctionWebSearch,
 )
+from openai.types.responses.response_output_item import McpCall
 from openai.types.responses.response_reasoning_item import (
     Content as ResponseReasoningTextContent,
 )
@@ -155,11 +156,7 @@ def get_developer_message(
                 "web_search_preview",
                 "code_interpreter",
                 "container",
-                "mcp",
             ):
-                # These are built-in tools that are added to the system message.
-                # Adding in MCP for now until we support MCP tools executed
-                # server side
                 pass
 
             elif tool.type == "function":
@@ -427,6 +424,44 @@ def _parse_final_message(message: Message) -> ResponseOutputItem:
     )
 
 
+def _parse_mcp_recipient(recipient: str) -> tuple[str, str]:
+    """
+    Parse MCP recipient into (server_label, tool_name).
+
+    For dotted recipients like "repo_browser.list":
+        - server_label: "repo_browser" (namespace/server)
+        - tool_name: "list" (specific tool)
+
+    For simple recipients like "filesystem":
+        - server_label: "filesystem"
+        - tool_name: "filesystem"
+    """
+    if "." in recipient:
+        server_label = recipient.split(".")[0]
+        tool_name = recipient.split(".")[-1]
+    else:
+        server_label = recipient
+        tool_name = recipient
+    return server_label, tool_name
+
+
+def _parse_mcp_call(message: Message, recipient: str) -> list[ResponseOutputItem]:
+    """Parse MCP calls into MCP call items."""
+    server_label, tool_name = _parse_mcp_recipient(recipient)
+    output_items = []
+    for content in message.content:
+        response_item = McpCall(
+            arguments=content.text,
+            type="mcp_call",
+            name=tool_name,
+            server_label=server_label,
+            id=f"mcp_{random_uuid()}",
+            status="completed",
+        )
+        output_items.append(response_item)
+    return output_items
+
+
 def parse_output_message(message: Message) -> list[ResponseOutputItem]:
     """
     Parse a Harmony message into a list of output response items.
@@ -440,18 +475,13 @@ def parse_output_message(message: Message) -> list[ResponseOutputItem]:
     output_items: list[ResponseOutputItem] = []
     recipient = message.recipient
 
-    # Browser tool calls
-    if recipient is not None and recipient.startswith("browser."):
-        output_items.append(_parse_browser_tool_call(message, recipient))
-
-    # Analysis channel (reasoning/chain-of-thought)
-    elif message.channel == "analysis":
-        output_items.extend(_parse_reasoning_content(message))
+    if recipient is not None:
+        # Browser tool calls
+        if recipient.startswith("browser."):
+            output_items.append(_parse_browser_tool_call(message, recipient))
 
-    # Commentary channel
-    elif message.channel == "commentary":
-        # Function calls
-        if recipient is not None and recipient.startswith("functions."):
+        # Function calls (should only happen on commentary channel)
+        elif message.channel == "commentary" and recipient.startswith("functions."):
             output_items.extend(_parse_function_call(message, recipient))
 
         # Built-in tools on commentary channel are treated as reasoning for now
@@ -463,10 +493,19 @@ def parse_output_message(message: Message) -> list[ResponseOutputItem]:
             # multiple functions - explanatory text with no recipient. Built-in tool
             # recipients (python/browser/container) also generate reasoning output.
             output_items.extend(_parse_reasoning_content(message))
+
+        # All other recipients are MCP calls
         else:
-            raise ValueError(f"Unknown recipient: {recipient}")
+            output_items.extend(_parse_mcp_call(message, recipient))
+
+    # No recipient - handle based on channel for non-tool messages
+    elif message.channel == "analysis":
+        output_items.extend(_parse_reasoning_content(message))
+
+    elif message.channel == "commentary":
+        # Commentary channel without recipient shouldn't happen
+        raise ValueError(f"Commentary channel message without recipient: {message}")
 
-    # Final output message
     elif message.channel == "final":
         output_items.append(_parse_final_message(message))
 
@@ -485,20 +524,70 @@ def parse_remaining_state(parser: StreamableParser) -> list[ResponseOutputItem]:
     if current_recipient is not None and current_recipient.startswith("browser."):
         return []
 
-    if parser.current_channel == "analysis":
-        reasoning_item = ResponseReasoningItem(
-            id=f"rs_{random_uuid()}",
-            summary=[],
-            type="reasoning",
-            content=[
-                ResponseReasoningTextContent(
-                    text=parser.current_content, type="reasoning_text"
+    if current_recipient and parser.current_channel in ("commentary", "analysis"):
+        if current_recipient.startswith("functions."):
+            rid = random_uuid()
+            return [
+                ResponseFunctionToolCall(
+                    arguments=parser.current_content,
+                    call_id=f"call_{rid}",
+                    type="function_call",
+                    name=current_recipient.split(".")[-1],
+                    id=f"fc_{rid}",
+                    status="in_progress",
                 )
-            ],
-            status=None,
-        )
-        return [reasoning_item]
-    elif parser.current_channel == "final":
+            ]
+        # Built-in tools (python, browser, container) should be treated as reasoning
+        elif not (
+            current_recipient.startswith("python")
+            or current_recipient.startswith("browser")
+            or current_recipient.startswith("container")
+        ):
+            # All other recipients are MCP calls
+            rid = random_uuid()
+            server_label, tool_name = _parse_mcp_recipient(current_recipient)
+            return [
+                McpCall(
+                    arguments=parser.current_content,
+                    type="mcp_call",
+                    name=tool_name,
+                    server_label=server_label,
+                    id=f"mcp_{rid}",
+                    status="in_progress",
+                )
+            ]
+
+    if parser.current_channel == "commentary":
+        return [
+            ResponseReasoningItem(
+                id=f"rs_{random_uuid()}",
+                summary=[],
+                type="reasoning",
+                content=[
+                    ResponseReasoningTextContent(
+                        text=parser.current_content, type="reasoning_text"
+                    )
+                ],
+                status=None,
+            )
+        ]
+
+    if parser.current_channel == "analysis":
+        return [
+            ResponseReasoningItem(
+                id=f"rs_{random_uuid()}",
+                summary=[],
+                type="reasoning",
+                content=[
+                    ResponseReasoningTextContent(
+                        text=parser.current_content, type="reasoning_text"
+                    )
+                ],
+                status=None,
+            )
+        ]
+
+    if parser.current_channel == "final":
         output_text = ResponseOutputText(
             text=parser.current_content,
             annotations=[],  # TODO
@@ -515,6 +604,7 @@ def parse_remaining_state(parser: StreamableParser) -> list[ResponseOutputItem]:
             type="message",
         )
         return [text_item]
+
     return []