The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: Fetches curated skill lists from the OpenAI repository (a trusted organization) and downloads skill archives from GitHub (a well-known service).
[COMMAND_EXECUTION]: Executes git clone and git sparse-checkout using subprocess.run with list arguments to avoid shell injection. All git operations are directed towards well-known GitHub domains.
[CREDENTIALS_UNSAFE]: Uses the GITHUB_TOKEN or GH_TOKEN from environment variables to authenticate with the GitHub API, facilitating the installation of skills from private repositories.
[REMOTE_CODE_EXECUTION]: Installs new agent skills from remote repositories, which introduces new code and instructions into the agent's environment. The skill enforces directory name validation and path traversal checks to mitigate risks.
[PROMPT_INJECTION]:
Ingestion points: scripts/list-curated-skills.py (reads GitHub API responses for skill directory listings).
Boundary markers: Absent.
Capability inventory: Includes the ability to clone repositories using git and write skill files to the local $CODEX_HOME/skills directory.
Sanitization: Implements _safe_extract_zip to prevent directory traversal attacks and validates that skill names are single path segments.

skill-installer