The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill provides a setup utility that fetches the Bun runtime installation script from the official bun.sh domain. Although the script is piped directly to a shell for execution, bun.sh is a well-known service in the development community, and the process is transparently described in the documentation.
[REMOTE_CODE_EXECUTION]: Includes js and eval commands that allow the agent to execute JavaScript within the headless browser context. The skill implements path validation in its execution logic to ensure that scripts loaded via the eval command are restricted to safe directories like /tmp or the current working directory, preventing arbitrary file reads.
[COMMAND_EXECUTION]: Utilizes the Bash tool and system utilities such as security (on macOS) for administrative tasks. These tasks include checking for updates, managing session state in ~/.gstack/, and retrieving decryption keys for local browser databases with user interaction.
[CREDENTIALS_UNSAFE]: Features a specialized browser cookie import tool that migrates session data from established browsers (Chrome, Edge, Arc, etc.) to the automated testing environment. To ensure safety, the skill redacts sensitive information such as 'Authorization' headers, 'X-API-Key' values, and HTML password fields from its logs and command outputs.
[DATA_EXFILTRATION]: Implements a telemetry system for tracking skill usage, outcomes, and performance. The system is designed to be privacy-conscious, sending only metadata and usage statistics while explicitly excluding code, file paths, or repository names. It includes a user consent workflow allowing users to opt into community mode or choose anonymous or disabled telemetry.
[INDIRECT_PROMPT_INJECTION]:
Ingestion points: Untrusted data enters the agent context through web content extraction commands like text, html, and snapshot in src/read-commands.ts and src/snapshot.ts.
Boundary markers: The skill does not currently implement specific delimiters or warnings to instruct the agent to ignore instructions found within external web content.
Capability inventory: The skill possesses significant capabilities, including shell command execution via Bash and JavaScript execution through the js and eval commands.
Sanitization: The skill performs basic sanitization by stripping script, style, noscript, and svg tags from page content before it is presented to the agent.
[DYNAMIC_EXECUTION]: Programmatically assembles and executes JavaScript code at runtime by wrapping input in asynchronous IIFEs. This mechanism is secured by internal validation logic that checks for path traversal and enforces location restrictions for any file-based execution.

browse