<think> tags in content should be extracted and reformatted.

Args section should be auto-generated from parameters.properties.

If description already has Args:, don't add another.

Tool call with dict arguments should be JSON serialized.

Tool calls should be formatted with <use_mcp_tool> tags.

Tool call with None content should work.

Tool description should be indented with 4 spaces.

Tool with empty description should not crash.

Input JSON schema should be included.

Tool should be formatted with ### Tool name: header.

Tool with None description should not crash.

Tool response should be embedded in a user message.

Tool responses should NOT be wrapped in <tool_response> tags.

Tool server should be my_mcp_server.

Tools can be passed without the function wrapper.

When tools are provided, a special system prompt should be generated.

Unicode content should be preserved.

User message should be wrapped in <|im_start|>user ... <|im_end|>.

Get today's date in YYYY-MM-DD format.

Calculate overall completion rate percentage

Calculate overall judge accuracy percentage

Update message history with tool calls data (llm client specific)

Upload a local file to the `/home/user` dir of the remote python interpreter. Args: sandbox_id: The id of the sandbox to run the code in.

Use GAIA-style judge to verify if the predicted answer is correct.

Ask question about an image or a video and get the answer with GPT-4o vision model. Args: media_path_or_url: The path of the image/video

Ask question about an image or a video and get the answer with a vision language model. Args: image_path_or_url: The path of the image fi

Get specific Wikipedia page content for the specific entity (people, places, concepts, events) and return structured information. This tool searc

Decorator: wraps any *async* function in asyncio.wait_for(). Usage: @with_timeout(20) async def create_message_foo(...): ...

Decorator: wraps any *async* function in asyncio.wait_for(). Usage: @with_timeout(20) async def create_message_foo(...): ...