The Agent Skills Directory

[DATA_EXFILTRATION]: The local preview server (serve_preview.py) implements proactive security measures to prevent data exposure. It explicitly blocks access to hidden files (e.g., .env), prevents directory traversal via symbolic links, and restricts the server to the local loopback interface (127.0.0.1).
[DYNAMIC_EXECUTION]: The skill uses a custom-built, recursive-descent parser (manifest_parser.py) to handle user-provided configuration manifests. This parser is restricted to a data-only subset of JavaScript, effectively preventing code injection or prototype pollution attacks by strictly forbidding executable statements, function calls, and reserved keys like __proto__ or constructor.
[COMMAND_EXECUTION]: The Python scripts provided for scaffolding (scaffold_template.py) and image measurement (measure_preview.py) perform standard local development tasks. The scaffolding script includes safety checks to prevent overwriting existing directories and validates that no symbolic links are present in the source or target paths to avoid potential symlink-based attacks.
[EXTERNAL_DOWNLOADS]: The project requires standard, well-known scientific and image processing libraries (numpy, Pillow). These are established dependencies with no suspicious or unverified third-party code observed.
[INDIRECT_PROMPT_INJECTION]: While the skill processes user-defined theme manifests, the attack surface is minimized through strict schema validation and the use of the aforementioned non-executable parser, ensuring that instructions cannot be smuggled through data fields.

liuguang-banlan-ui