The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes the browse CLI and Node.js scripts to perform automation. It includes a dedicated parseCommand utility in evaluate.mjs that manually validates and parses arguments before execution to prevent shell injection attacks.
[EXTERNAL_DOWNLOADS]: The framework manages its environment by downloading standard libraries such as playwright, zod, and @anthropic-ai/sdk from official registries via npm. These dependencies are consistent with its stated purpose of browser automation.
[PROMPT_INJECTION]: The skill has an indirect prompt injection surface as it ingests accessibility trees and text from external websites during browsing tasks. This data is used to iteratively update navigation strategies and generate automation code. This is a low-risk surface inherent to the tool's primary purpose.
[SAFE]: To ensure data privacy, the skill implements a lockDownTrace function that enforces strict file system permissions (chmod 0700/0600) on all trace artifacts, screenshots, and logs, ensuring that sensitive session data is only accessible to the process owner.
[COMMAND_EXECUTION]: The codegen.mjs script generates Playwright and Stagehand automation scripts derived from agent traces and subsequently executes them for verification using the tsx runner.

autobrowse