Same as usual converter, but removes Javascript links and escapes URIs.

Same as usual, but be sure to start with a new line

Same as usual converter, but removes data URIs

Convert various types of resources (doc, ppt, pdf, excel, csv, zip file etc.) described by an file: or data: URI to markdown. Args: u

Convert MCP tool definitions to OpenAI function call format. Transforms the internal tool definition format used by MCP servers into

Create a linux sandbox. Args: timeout: Time in seconds before the sandbox is automatically shutdown. The default is 600 seconds. Ret

Remove one or more tasks from your plan. Use this to adjust your plan when tasks become irrelevant, duplicated, or no longer needed. Thi

Download a file from the internet to the `/home/user` dir of the sandbox (avoid large or slow URLs). Args: sandbox_id: The id of the sand

Download a file from the sandbox to local system. Files in sandbox cannot be processed by tools from other servers - only local files and internet URL

Check if current message_history + summary_prompt will exceed context If it will exceed, remove the last assistant-user pair and retu

Execute a single tool call. Args: tool_manager: The tool manager to use server_name: Name of the MCP server

Extract tool call information from LLM response

Format token usage statistics, return summary_lines for format_final_summary and log string

Format tool result for feeding back to LLM. Args: tool_result: The tool execution result Returns: F

Create a TaskLog instance from a dictionary. Args: d: Dictionary containing TaskLog field values. Returns:

Generate a minimal system prompt without MCP tool definitions. Used when no tools are available or when running in tool-less mode. Args

Get basic information

Return the custom identity prompt string.

Get execution flow

Get execution summary

Get performance summary

Get raw message data for debugging

Get spans summary

Get step logs summary

Tool to perform web searches via Serper API and retrieve rich results. It is able to retrieve organic search results, people also ask, r

Perform google searches via Serper API and retrieve rich results. It is able to retrieve organic search results, people also ask, related searches

Tool to perform web searches via Serper API and retrieve rich results. It is able to retrieve organic search results, people also ask, r

Main page

Check if a query has been executed before. Args: cache_name: Name of the cache (e.g., "main_google_search")

Calculate judge accuracy percentage

Calculate Level 1 accuracy percentage

Calculate overall Level 1 accuracy percentage

Calculate Level 2 accuracy percentage

Calculate overall Level 2 accuracy percentage

Calculate Level 3 accuracy percentage

Calculate overall Level 3 accuracy percentage

List available JSON files

Display the complete task plan with all items and their status. Use this to review your overall progress, see what's done and what remains,

Load trace file

Send message event with streaming content. Args: message_id: Unique identifier for the message delta_content

Patched version that doesn't return FORMAT_ERROR_MESSAGE. Instead of checking for \boxed{} content, we use the entire answer

Patched version that uses user-friendly prompt for main agent. For main agent in demo mode, uses a Markdown-friendly prompt instead

Patched version that prepends custom identity prompt.

Patched version that removes boxed format requirement.

Process LLM response

Put data safely from any thread

Use this tool to execute STATELESS Python code in your chain of thought. The code will not be shown to the user. This tool should be used for internal

You can use this tool use solve hard math problem, puzzle, riddle and IQ test question that requires a lot of chain of thought efforts. DO NOT use

You can use this tool to solve hard math problem, puzzle, riddle and IQ test question that requires a lot of chain of thought efforts. DO NOT use

Record that a query has been executed. Args: cache_name: Name of the cache query_str: The query string to re

Run the complete analysis and return summary statistics

Execute a lightweight shell command in the linux sandbox (no long-running, blocking, or resource-heavy processes). Args: command: The com

Run short, safe python code in a sandbox and return the execution result (avoid long loops or heavy tasks; must finish quickly). Args: co

Scrape content from a URL, including web pages, PDFs, code files, and other supported resources, and extract meaningful information using an LLM.

This tool is used to scrape a website for its content. Search engines are not supported by this tool. This tool can also be used to get YouTube video

This tool is used to scrape a website for its content. Search engines are not supported by this tool. This tool can also be used to get YouTube video

Search the Wayback Machine (archive.org) for archived versions of a webpage, optionally for a specific date. Args: url: The URL to search

Search for an entity in Wikipedia and return the revision history for a specific month. Args: entity: The entity to search for in Wikiped

Tool to perform web searches via Tencent Cloud SearchPro API (Sogou search engine). Sogou search offers superior results for Chinese-languag

Performs web searches using the Tencent Cloud SearchPro API to retrieve comprehensive information, with Sogou search offering superior results for Chi

Simulate vLLM's strftime_now function.

Load the chat template.

add_generation_prompt should add <|im_start|>assistant at the end.

Assistant with empty string content should work.

Assistant message should be wrapped correctly with <think> tags.

enable_thinking=false should output empty think tags.

enable_thinking=true should not output empty think tags.

Test a complete tool use flow.

Multi-turn conversation should maintain correct order.

Multiple tool calls should be separated by newlines.

Multiple consecutive tool responses should be merged into one user message.

Newlines in content should be preserved.

Only system message should work.

reasoning_content field should be wrapped in <think> tags.

Test reasoning content combined with tool use.

Special characters should be preserved.

System message should be wrapped correctly.

Custom system message should be prepended when tools are present.

Think tags should be preserved in historical messages, not removed.