Find and return all dangerous tokens in the command. Returns empty list if no dangerous tokens found.

Format agent actions for display.

Create SecurityPolicy from configuration dict. Supports two formats for blocked_commands: 1. List format (applies to

Deserialize from dict.

Generate a comprehensive summary of the recording session.

Get accessibility tree locally.

Get accessibility tree from desktop environment. Args: max_depth: Maximum depth of accessibility tree traversal

Get macOS accessibility tree Args: max_depth: Maximum depth Returns: Accessibil

Get the configuration of the specified agent

Get all sanitized cached tools, grouped by server.

Get statistics about the embedding cache. Returns: Dict with structure: { "total_embeddings": int,

Get current mouse cursor position. Useful for GUI debugging and relative positioning. Returns: (x, y) tu

Get cursor position Returns: (x, y)

Get the global installer manager instance.

Get the host URL for a specific port. Args: port: The port number to get the host for Returns:

Get debug info from the last search operation. Returns: Dict containing search debug info, or None if no search has been

Extract the model name from LangChain LLM using BaseChatModel standards.

Extract the model provider from LangChain LLM using BaseChatModel standards.

Auto-generate JSON-schema from _run() or _arun() signature. Returns empty dict for tools with no parameters. Priority: prefer

Get penalty factor for a tool (0.2-1.0).

Get a prompt by name.

Get a prompt by name.

Get the global quality manager instance.

Get screen size. Returns: Dict with 'width' and 'height', or None on failure

Capture screenshot locally using ScreenshotHelper.

Get screenshot from desktop environment. Returns: Screenshot image bytes (PNG/JPEG), or None on failure

Get the list of configured server names. Returns: List of server names.

Get simple feature dict (for banner display)

Get summary

Get macOS system information Returns: System information dictionary

Get system prompt based on platform. Args: platform: Platform type (Ubuntu, Windows, macOS, or Darwin) Returns:

Get terminal output (macOS Terminal.app or iTerm2) Returns: Terminal output content

Intelligent tool retrieval: automatically decides whether to return all tools or trigger search. Logic: - If tool_co

Get top N tools by quality score. Args: n: Number of tools to return backend: Filter by backend type

Check if sanitized cache exists and has data.

Initialize the MCP provider. If config["eager_sessions"] is True, creates sessions for all configured servers. Otherwise, ses

Initialize the MCP session and discover available tools. Returns: The session information returned by the MCP implementation.

Set cache breakpoints for the 3 most recent turns. One cache breakpoint is left for tools/system prompt, to be shared across sessions.

Dispatch by name — same routing as ShellConnector via AioHttpConnector.

Universal tool invocation method. Supports multiple calling patterns: 1. Using BaseTool instance with bound runtime

Build feedback message to pass iteration summary to next iteration.

List static tools for every registered backend.

List all available prompts.

List all available resources from the MCP implementation.

List tools from MCP sessions. Args: session_name: If provided, only list tools from that session.

List all available tools from the MCP implementation.

List all available tools from the MCP implementation.

List all windows Returns: Window list

Load all quality records and global execution count from disk. Returns: Tuple of (records_dict, global_execution_count)

Print startup banner with server information

Print check results

Read a resource by URI.

Read a resource by URI.

Record an agent action. Args: agent_name: Name of the agent performing the action action_type: Type

Record an agent's action and decision-making process. Args: agent_name: Name of the agent performing the action

Record tool execution result and increment global counter.

Record a single iteration's LLM conversation to conversations.jsonl (real-time). Args: iteration: Iteration numb

Record the tools retrieved for a task Args: task_instruction: The task instruction used for retrieval

Refresh MCP tool cache by starting servers one by one and saving tool metadata.

Convert Anthropic response to parameter list. Handles both text blocks, tool use blocks, and thinking blocks. Args: response

Execute a Bash script locally. Return format matches the server's ``/run_bash_script`` endpoint.

Asynchronous running health check

Run main function

Execute a Python script locally. Return format matches the server's ``/run_python`` endpoint.

Execute Python script on remote server Args: code: Python code string timeout: Execution timeout in

Save initial screenshot to screenshots dir and update metadata.

Save sanitized tool metadata to disk. Args: servers: Dict mapping server_name -> list of sanitized tool metadata

Save/update a single server's tools to cache (incremental append). Args: server_name: Name of the MCP server

Search tools from backend(s) or session. Args: task_description: Task description for searching relevant tools

Set the global quality manager instance.

Set expected server order (from config). Used when saving to disk.

Set desktop wallpaper Args: image_path: Image path Returns: Result dictionary