Handle tool execution results.

Write the content of a file to a given path; raise a ToolError if an error occurs.

Execute a single tool and handle errors.

Format a single content item using the appropriate formatter. Uses the Strategy pattern to dispatch to the correct formatter based on content typ

Resolve a model ID to its config. Tries an exact match first, then the longest known key that the ID starts with, so undated aliases, newly d

Press a keyboard key or key combination.

Render an API response to a streamlit tab

Run argv under sandbox-exec, capping both wall-clock time and captured output. stdout and stderr are merged so the cap applies to the total.

Remove advisor server_tool_use and advisor_tool_result blocks in place. Required when dropping the advisor from `tools` mid-conversation; the API

Extract an image and return a file reference.

Format conversation messages into a readable transcript. Args: messages: List of message dictionaries from session state include_

Calculate scale factors for converting Claude coordinates to viewport coordinates. Args: viewport_width: Actual browser

Load data from a file in the storage directory.

Load a prompt template from the prompts directory.

ceil(px / px_per_token) using integer math.

Handle agent output - both text and tool use.

Scale a coordinate pair list/tuple. Args: coords: [x, y] coordinate pair viewport_width: Actual browser view

Scale coordinates from Claude's vision to actual viewport. Args: x: X coordinate from Claude y: Y coordinate

Test a single command against the security hook.

Return the server-hosted tool param. The server supplies the description and input_schema; we only declare the type and the display si

Validate chmod commands - only allow making files executable with +x. Returns: Tuple of (is_allowed, reason_if_blocked)

Validate init.sh script execution - only allow ./init.sh. Returns: Tuple of (is_allowed, reason_if_blocked)

Process user input and handle tool calls in a loop

Handle various click actions.

Helper to create a message with circular reference.

Create the read/write context based on connection type.

Perform a drag operation.

Make targeted changes to a file.

Launch browser and ensure page is ready.

Seed .streamlit/config.toml from the shared theme on first run only, so a user's local edits survive.

Execute JavaScript code in the page context. Returns the result of the last expression.

Find elements on the page matching the search query using AI.

Fill a form field with a value.

Format message content based on its type. This is the main entry point that handles different content structures.

Extract ALL text content from the current page. USE THIS to read articles, posts, or any text content. Returns the actual tex

Move the mouse cursor to a position without clicking. Useful for revealing tooltips, dropdown menus, or triggering hover states.

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

List files in a directory.

Convert an agent ToolResult to an API ToolResultBlockParam.

With the assumption that images are screenshots that are of diminishing value as the conversation progresses, remove all but the final `image

Press mouse button down.

Release mouse button.

Navigate to a URL.

Read a file from disk. Args: path: Path to the file to read max_lines: Maximum number of lines to read (0 mea

Extract the DOM tree with structured content and element references. USE THIS to analyze page structure and find specific elements.

Render assistant message content. Handles both single items and lists of content blocks, properly routing tool results to the TOOL se

Route message rendering based on role. Args: message: Message dictionary containing role and content

Render the actual message content based on its type. Args: message: The message content to render

Render a tool result that was stored in session state. Args: message: Dictionary containing the tool_use_id reference

Render a tool result with output, error, and optional image. Args: tool_result: The ToolResult object to render

Render a tool use message with coordinate scaling for browser tools. Args: message: Dictionary containing tool use information

Render user message content. Handles both single items and lists of content blocks, skipping image blocks in conversation history.

Same for Accessibility: AXIsProcessTrustedWithOptions with the prompt option pops the system dialog and adds <app> to the list.

Trigger the system "<app> would like to record this computer's screen" dialog, which also adds <app> to the Screen Recording list (unchecked) so

Resolve `path` inside the scratch root; raise if it would escape.

Replace inline base64 image sources with on-disk file paths.

Run a single test and track results. Args: test_name: Name of the test test_func: Test function to execute

Replace ``{name}`` placeholders without ``str.format``'s attribute/index traversal (which would let a user-supplied template reach ``__globals__``

Apply coordinate scaling to browser tool inputs for display. This ensures the displayed coordinates match what the browser tool will actually

Scroll the page or element.

Scroll to a specific element.

Only nudge after a lone click/type/key/scroll/wait, not after screenshots or batch calls (which is what the model is told to do anyway).