The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill manages a local library of third-party repositories and documentation at $RESEARCH_HOME. As this content is cloned from external sources, it is inherently untrusted and could contain indirect prompt injection vectors.\n
Ingestion points: $RESEARCH_HOME/repos/, $RESEARCH_HOME/docs/, and $RESEARCH_HOME/pages/.\n
Boundary markers: The skill's primary design goal is to isolate these materials from the main project directory to prevent them from being treated as trusted context by default.\n
Capability inventory: Includes file management via git and metadata retrieval through gh.\n
Sanitization: Content is maintained as raw source for research purposes, but isolation is enforced by policy.\n- [COMMAND_EXECUTION]: The skill uses Python scripts (scripts/library.py, scripts/package_health.py) to execute system commands via subprocess.run.\n
Evidence: Interaction with git (cloning, sparse-checkout, maintenance), gh (GitHub API queries), node (version checking), and apt-cache (distro package queries).\n- [EXTERNAL_DOWNLOADS]: The skill performs network operations to fetch metadata and source code from external services.\n
Evidence: Communicates with PyPI, npm, crates.io, GitHub, Debian madison, and Launchpad APIs. These operations are essential to its function and target well-known, trusted package registries.

research-library