The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill manages a research library that ingests external, unvetted content including git repositories, PDFs, and documentation mirrors.
Ingestion points: Material is stored in $RESEARCH_HOME under repos/, docs/, and pages/ buckets.
Boundary markers: The skill explicitly mitigates risk by isolating this content outside the project's working directory, preventing agents from reading it 'ambiently'. Instructions require deliberate access to these paths only when task-relevant.
Capability inventory: Scripts use subprocess.run for git/GitHub operations and urllib for PyPI API requests.
Sanitization: Entry naming follows a strict <host>--<owner>--<repo> convention to ensure canonical mapping and avoid collision attacks.
[COMMAND_EXECUTION]: The included Python scripts interface with system tools to perform library maintenance and health checks.
Evidence: scripts/library.py executes git clone, git remote, and git rev-parse. scripts/package_health.py executes gh api.
Detail: Subprocess calls utilize argument lists rather than shell strings, effectively preventing command injection vulnerabilities.
[EXTERNAL_DOWNLOADS]: The skill performs network operations to fetch metadata and clone reference material.
Evidence: scripts/package_health.py fetches JSON metadata from pypi.org. scripts/library.py clones repositories from user-supplied URLs.
Detail: Interactions target well-known services (PyPI and GitHub). The skill correctly identifies itself via a User-Agent header pointing to the vendor's GitHub repository (theodoread/agent-skills).

research-library