[gpt-oss] Small bug fixes for frontend (#22512)

Signed-off-by: Chen Zhang <zhangch99@outlook.com>

vllm/entrypoints/context.py

@@ -1,15 +1,20 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import json
 import logging
 from abc import ABC, abstractmethod
+from typing import TYPE_CHECKING, Union
 
-from openai_harmony import Message, Role, StreamState
+from openai_harmony import Author, Message, Role, StreamState, TextContent
 
 from vllm.entrypoints.harmony_utils import (
     get_encoding, get_streamable_parser_for_assistant, render_for_completion)
 from vllm.entrypoints.tool import Tool
 from vllm.outputs import RequestOutput
 
+if TYPE_CHECKING:
+    from mcp.client import ClientSession
+
 logger = logging.getLogger(__name__)
 
 
@@ -71,6 +76,7 @@ class HarmonyContext(ConversationContext):
     def append_output(self, output) -> None:
         if isinstance(output, RequestOutput):
             output_token_ids = output.outputs[0].token_ids
+            self.parser = get_streamable_parser_for_assistant()
             for token_id in output_token_ids:
                 self.parser.process(token_id)
             output_msgs = self.parser.messages
@@ -106,19 +112,41 @@ class HarmonyContext(ConversationContext):
     def render_for_completion(self) -> list[int]:
         return render_for_completion(self.messages)
 
-    async def call_search_tool(
-        self,
-        tool_session: Tool,
-        last_msg: Message,
-    ) -> list[Message]:
-        return await tool_session.get_result(self)
+    async def call_search_tool(self, tool_session: Union["ClientSession",
+                                                         Tool],
+                               last_msg: Message) -> list[Message]:
+        if isinstance(tool_session, Tool):
+            return await tool_session.get_result(self)
+        tool_name = last_msg.recipient.split(".")[1]
+        args = json.loads(last_msg.content[0].text)
+        result = await tool_session.call_tool(tool_name, args)
+        result_str = result.content[0].text
+        content = TextContent(text=result_str)
+        author = Author(role=Role.TOOL, name=last_msg.recipient)
+        return [
+            Message(author=author, content=[content], recipient=Role.ASSISTANT)
+        ]
 
-    async def call_python_tool(
-        self,
-        tool_session: Tool,
-        last_msg: Message,
-    ) -> list[Message]:
-        return await tool_session.get_result(self)
+    async def call_python_tool(self, tool_session: Union["ClientSession",
+                                                         Tool],
+                               last_msg: Message) -> list[Message]:
+        if isinstance(tool_session, Tool):
+            return await tool_session.get_result(self)
+        param = {
+            "code": last_msg.content[0].text,
+        }
+        result = await tool_session.call_tool("python", param)
+        result_str = result.content[0].text
+
+        content = TextContent(text=result_str)
+        author = Author(role=Role.TOOL, name="python")
+
+        return [
+            Message(author=author,
+                    content=[content],
+                    channel=last_msg.channel,
+                    recipient=Role.ASSISTANT)
+        ]
 
 
 class StreamingHarmonyContext(HarmonyContext):

vllm/entrypoints/openai/protocol.py

@@ -19,8 +19,8 @@ from openai.types.chat.chat_completion_message import (
 # yapf: enable
 from openai.types.responses import (ResponseFunctionToolCall,
                                     ResponseInputItemParam, ResponseOutputItem,
-                                    ResponsePrompt, ResponseStatus,
-                                    ResponseTextConfig)
+                                    ResponsePrompt, ResponseReasoningItem,
+                                    ResponseStatus, ResponseTextConfig)
 from openai.types.responses.response import ToolChoice
 from openai.types.responses.tool import Tool
 from openai.types.shared import Metadata, Reasoning
@@ -239,6 +239,7 @@ def get_logits_processors(processors: Optional[LogitsProcessors],
 
 
 ResponseInputOutputItem: TypeAlias = Union[ResponseInputItemParam,
+                                           ResponseReasoningItem,
                                            ResponseFunctionToolCall]
 
 

vllm/entrypoints/openai/serving_responses.py

@@ -16,8 +16,7 @@ from fastapi import Request
 from openai import BaseModel
 # yapf conflicts with isort for this block
 # yapf: disable
-from openai.types.responses import (ResponseContentPartDoneEvent,
-                                    ResponseCreatedEvent,
+from openai.types.responses import (ResponseCreatedEvent,
                                     ResponseFunctionToolCall,
                                     ResponseInProgressEvent,
                                     ResponseOutputItem,
@@ -54,7 +53,7 @@ from vllm.entrypoints.openai.protocol import (ErrorResponse,
 # yapf: enable
 from vllm.entrypoints.openai.serving_engine import OpenAIServing
 from vllm.entrypoints.openai.serving_models import OpenAIServingModels
-from vllm.entrypoints.tool_server import ToolServer
+from vllm.entrypoints.tool_server import MCPToolServer, ToolServer
 from vllm.inputs.data import TokensPrompt as EngineTokensPrompt
 from vllm.logger import init_logger
 from vllm.outputs import CompletionOutput
@@ -238,6 +237,15 @@ class OpenAIServingResponses(OpenAIServing):
         if raw_request:
             raw_request.state.request_metadata = request_metadata
 
+        if self.tool_server is not None and isinstance(
+                self.tool_server, MCPToolServer
+        ) and (request.background or request.stream) and request.tools and any(
+                tool.type in ["web_search_preview", "code_interpreter"]
+                for tool in request.tools):
+            return self.create_error_response(
+                "MCP tool server is not supported in background mode and "
+                "streaming mode")
+
         # Schedule the request and get the result generator.
         generators: list[AsyncGenerator[ConversationContext, None]] = []
 
@@ -844,9 +852,13 @@ class OpenAIServingResponses(OpenAIServing):
                             type="reasoning",
                             content=[
                                 ResponseReasoningTextContent(
-                                    text=previous_item.content[0].text),
+                                    text=previous_item.content[0].text,
+                                    type="reasoning_text",
+                                ),
                             ],
                             status="completed",
+                            id=current_item_id,
+                            summary=[],
                         )
                         yield _send_event(
                             ResponseReasoningTextDoneEvent(
@@ -857,15 +869,6 @@ class OpenAIServingResponses(OpenAIServing):
                                 content_index=current_content_index,
                                 text=previous_item.content[0].text,
                             ))
-                        yield _send_event(
-                            ResponseContentPartDoneEvent(
-                                type="response.content_part.done",
-                                item_id=current_item_id,
-                                sequence_number=-1,
-                                output_index=current_output_index,
-                                content_index=current_content_index,
-                                part=reasoning_item,
-                            ))
                         yield _send_event(
                             ResponseOutputItemDoneEvent(
                                 type="response.output_item.done",

vllm/entrypoints/tool.py

@@ -2,7 +2,9 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 import os
 from abc import ABC, abstractmethod
-from typing import TYPE_CHECKING, Any
+from typing import TYPE_CHECKING, Any, Optional
+
+from openai_harmony import Message
 
 from vllm.logger import init_logger
 
@@ -70,7 +72,16 @@ class HarmonyPythonTool(Tool):
                 "gpt_oss is not installed, code interpreter is disabled")
             return
 
-        self.python_tool = PythonTool()
+        # NOTE (Chen): as of gpt-oss 0.0.2, there is a bug in _make_response
+        # and we do the following monkey patch to fix it.
+        class PatchedGptOssPythonTool(PythonTool):
+
+            def _make_response(self,
+                               output: str,
+                               channel: Optional[str] = None) -> Message:
+                return super()._make_response(output)
+
+        self.python_tool = PatchedGptOssPythonTool()
         logger.info_once("Code interpreter tool initialized")
 
     async def get_result(self, context: "ConversationContext") -> Any:

vllm/entrypoints/tool_server.py

@@ -4,7 +4,7 @@ from abc import ABC, abstractmethod
 from contextlib import AbstractAsyncContextManager, asynccontextmanager
 from typing import TYPE_CHECKING, Any, Optional
 
-from openai_harmony import ToolNamespaceConfig
+from openai_harmony import ToolDescription, ToolNamespaceConfig
 
 from vllm.entrypoints.tool import HarmonyBrowserTool, HarmonyPythonTool, Tool
 from vllm.logger import init_logger
@@ -105,7 +105,6 @@ class MCPToolServer(ToolServer):
         self.harmony_tool_descriptions = {}
 
     async def add_tool_server(self, server_url: str):
-        from mcp.types import ToolDescription
         tool_urls = server_url.split(",")
         self.harmony_tool_descriptions = {}
         self.urls: dict[str, str] = {}
@@ -133,6 +132,8 @@ class MCPToolServer(ToolServer):
                 logger.warning(
                     "Tool %s already exists. Ignoring duplicate tool server %s",
                     tool_from_mcp.name, url)
+        logger.info("MCPToolServer initialized with tools: %s",
+                    list(self.harmony_tool_descriptions.keys()))
 
     def has_tool(self, tool_name: str):
         return tool_name in self.harmony_tool_descriptions