The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes the Python subprocess module to execute git and gh (GitHub CLI) commands for cloning repositories and interacting with the GitHub API. The implementation in scripts such as scripts/clone_repo.py, scripts/repo_metadata.py, and scripts/search_github.py follows security best practices by passing arguments as lists rather than using shell strings, which mitigates the risk of command injection from research metadata.- [EXTERNAL_DOWNLOADS]: The skill performs legitimate network operations to fetch repository data and research metadata from well-known services. Specifically, it clones repositories from github.com and queries the paperswithcode.com API to map papers to implementations. These downloads are intrinsic to the skill's function and target trusted domains.- [PROMPT_INJECTION]: The skill possesses a vulnerability surface for indirect prompt injection (Category 8). During Phase 4, the agent is directed to clone and deeply read source code from repositories identified through automated search. If an analyzed repository contains adversarial instructions hidden in code comments or documentation, the agent might inadvertently adopt those instructions.
Ingestion points: Repository content is cloned to the phase4_deep_dive/repos/ directory.
Boundary markers: There are no explicit instructions to use delimiters or defensive prompting when the agent reads the external code files.
Capability inventory: The skill environment allows for subprocess execution, network access via urllib, and file system modifications.
Sanitization: No content filtering is applied to the retrieved source files before they are presented to the agent for analysis.

github-research