The Agent Skills Directory

[COMMAND_EXECUTION]: Uses the GitHub CLI (gh) and standard utilities like jq and file to interact with repositories and validate files.
[INDIRECT_PROMPT_INJECTION]: Processes Pull Request bodies as untrusted input. Mitigates risk by using a 'refuse-by-default' markdown parser (splice.mjs) that stops execution if it encounters unexpected or potentially malicious markdown/HTML constructs.
[DATA_EXPOSURE]: Implements strict controls on local file access. It refuses symlinks, verifies MIME types to ensure only images are uploaded, and requires an absolute root directory for scoping file operations.
[SAFE]: Implements industry best practices for security, including input normalization, output escaping, and verification of remote asset URLs against GitHub's official attachment infrastructure.

attaching-pr-screenshots