The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted code and documentation from external GitHub repositories for deep analysis. This creates a surface for indirect prompt injection attacks where an attacker-controlled repository could influence the agent's behavior. \n
Ingestion points: Cloned repository source files and READMEs located in the github-research-output directory structure. \n
Boundary markers: None identified; the skill instructions direct the agent to 'Actually Read Code' and trace execution flows directly. \n
Capability inventory: The skill executes git and gh via subprocess.run and performs extensive file system operations. \n
Sanitization: No specific mechanisms for filtering or sanitizing instructions embedded in the external code were identified. \n- [COMMAND_EXECUTION]: Several Python scripts (e.g., clone_repo.py, repo_metadata.py, search_github.py) use subprocess.run to call the git and gh command-line tools. These commands use arguments derived from user-provided topic slugs and automated search results. \n- [EXTERNAL_DOWNLOADS]: The skill clones repositories from GitHub (github.com) and fetches data from the Papers With Code API (paperswithcode.com). These are established, well-known services appropriate for the skill's stated purpose of software research.

github-research