Record execution result to quality manager for quality tracking.

Based on GroundingConfig.enabled_backends, register Provider instances to self._registry. Here only do *instantiation*, not a

Register the agent class in the registry upon instantiation.

Register SystemProvider separately because it requires GroundingClient instance. SystemProvider provides meta-level tools for queryin

Register a temporary file for later cleanup

Render agents section

Render grounding operations section

Render header section

Render logs section

Reorder servers dict according to _server_order.

Resize screenshot to display size for Computer Use API. For computer-use-2025-01-24, the screenshot must be resized to the

Restore MCP SDK logging to normal.

Run a shell command string (used for conda-wrapped scripts).

Sanitize raw cache and save to sanitized cache file.

Sanitize tool parameter schema to comply with Claude API requirements. Fixes common issues: - Empty object schemas (no properties, no req

Save tools metadata to cache file.

Scale coordinates from display size to actual screen size. The API returns coordinates in display_size (e.g., 1024x768).

Scan workspace directory and collect file information. Args: workspace_path: Path to workspace directory

Convert ToolSchema to OpenAI ChatCompletion tool format

Select key screenshots if there are too many.

Serialize AT-SPI element to dictionary Args: element: AT-SPI accessible element depth: Current depth

Serialize macOS accessibility element to dictionary Args: element: AX element depth: Current depth

Serialize Windows UIA element to dictionary Args: element: UIA element depth: Current depth

Store embedding in structured cache.

Show cursor

Use LLM to summarize large tool results.

Suppress MCP SDK's JSON parsing error logs. The MCP SDK logs errors when it receives non-JSON messages from servers. These ar

Convert a Pydantic BaseModel to a JSON-serialisable dict (by_alias=True). Leave all other types unchanged.

Convert ToolResult to LLMClient usable message format with LLM summarization for large results. Args: result: Tool execution result

Test JSON-RPC HTTP connection by sending an initialize request.

Wrap bash script with conda activation if needed.

Activate window (macOS uses AppleScript) Args: window_name: Window name or application name strict:

Add agent activity

Add execution record and update stats.

Add grounding operation

Adjust tool ranking using penalty-based approach. Args: tools_with_scores: List of (tool, semantic_score) tup

Attach grounding client Args: grounding_client: GroundingClient instance

Attach LLM client Args: llm_client: LLMClient instance

Build complete LLM configuration with auto-detection and environment variables. Auto-detects: - API key from environment variables (

Call an MCP tool with automatic reconnection handling and retry logic. Args: name: The name of the tool to call. argu

Call an MCP tool.

Capture screenshot with cursor using macOS native screencapture command Args: output_path: Output file path

Check all features with functional testing Args: test_endpoints: Whether to test endpoints (False only checks fe

Check for tool changes (new/updated/unchanged). Returns dict: {tool_key: "new"|"updated"|"unchanged"}

Clean up all temporary test files

Clear embeddings from cache. Args: backend: If provided, only clear this backend. If None, clear all. server:

Close HTTP session.

Close the HTTP session.

Provider shutdown cleanup

Close session

Close a session. Args: server_name: The name of the server to close the session for. Raises: ValueError: If

Close window or application (macOS uses AppleScript) Args: window_name: Window name or application name

Compute adaptive quality weight based on data confidence. Returns higher weight when we have more reliable quality data,

Connect to GUI desktop environment

Convert an MCP Tool to a BaseTool (RemoteTool) instance. This function extracts the tool schema from an MCP tool object and creates

Create a connector based on server configuration. Args: server_config: The server configuration section server_name: Name of

Create and initialize Session, return "session_name" (external visible) name is auto generated when it's None: <backend>-<index>

Create a session for the specified server with retry logic. Args: server_name: The name of the server to create a session for.

Create GUI session. Args: session_config: Session configuration Returns: GUISession

Create Web session

Create AnyTool UI instance Args: enable_live: Whether to enable live display updates compact: Use compact layout for sma

Close the session/connection and reset state. Ensures proper cleanup of all resources including aiohttp sessions.

Close the session/connection and reset state.

Ensure the dependencies of the MCP server are installed. This method checks if the dependencies are installed, and if not, asks the u

Evaluate tool description quality using LLM.

Run self-evolution cycle on given tools. This method: 1. Detects tool changes 2. Re-evaluates descriptions w

Run quality self-evolution cycle. This triggers: - Tool change detection - Description re-evaluation for upd

Execute a task with AnyTool. Args: task: Task instruction context: Additional context wo

Direct action execution (bypass LLM planning). Args: action_type: Action type from ACTION_SPACE para

Execute a desktop action (same logic as GUIConnector).

Execute a desktop action. This is the main method for action space execution. Args: action_type: Action

Build prompt for generating final summary across all iterations.