The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill describes the evaluateJavascriptWithResponse method, which utilizes the new Function() constructor to execute string-based code in the webview context. This pattern provides a mechanism for remote code execution if the input script is derived from untrusted sources.\n- [COMMAND_EXECUTION]: Code examples in SKILL.md and reference/rpc-patterns.md demonstrate persistence through Bun.write, allowing the application to write JSON data to the local filesystem.\n- [PROMPT_INJECTION]: The RPC interface creates a surface for indirect prompt injection where data processed by the agent could trigger the aforementioned execution capabilities. \n
Ingestion points: RPC request parameters and the script parameter of evaluateJavascriptWithResponse in reference/rpc-patterns.md.\n
Boundary markers: There are no implementation-level boundary markers or delimiters suggested to isolate untrusted data from the RPC logic.\n
Capability inventory: The skill exposes code execution (new Function) and file system access (Bun.write).\n
Sanitization: No sanitization, validation, or escaping of incoming data is present in the provided documentation.

electrobun-rpc