The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to execute arbitrary build and test commands inferred from a project's manifest files (e.g., 'npm run build', 'python -m build', 'pytest', 'cmake'). Additionally, the 'scripts/setup_env.ps1' script executes 'conda' and 'pip' commands to provision environments.
[INDIRECT_PROMPT_INJECTION]: The skill has a significant attack surface for indirect prompt injection as it is designed to process untrusted project repositories. A malicious repository could include harmful instructions or malicious dependencies in files like 'requirements.txt' or 'package.json', which the skill then processes and executes.

Ingestion points: Local repository files such as README, pyproject.toml, requirements.txt, and package.json are read in SKILL.md.
Boundary markers: Not present; the agent is instructed to read the project files directly to determine environment and build steps.
Capability inventory: The agent can create directories, install packages, and execute arbitrary build/test scripts as defined in the project files.
Sanitization: No sanitization or validation of the content within the manifest files is performed before they are used to determine and execute commands.

[EXTERNAL_DOWNLOADS]: The skill triggers the download and installation of dependencies from external registries like NPM and PyPI as part of its environment setup and build workflow.

github-project-builder