The Agent Skills Directory

[COMMAND_EXECUTION]: The workflow involves several agents that generate and execute shell commands. Specifically, the solution_verifier_agent is instructed to solve lab assignments by executing code and following setup instructions in a README, while the starter_code_agent validates the scaffold by running environment setup and tests.- [REMOTE_CODE_EXECUTION]: The skill facilitates the installation of third-party software and dependencies. The starter_code_agent is responsible for creating environment files such as requirements.txt or package.json, and the solution_verifier_agent installs these dependencies in a fresh environment during the verification phase.- [DATA_EXFILTRATION]: The skill manages sensitive ground truth data, including seeds, planted properties, and reference solutions, which are intended for professor use only. While there are internal rules to isolate this data (such as ground_truth.md), the presence of this sensitive information in the agent context alongside capabilities to generate student-facing artifacts creates a potential data exposure surface.- [PROMPT_INJECTION]: The skill has an attack surface for indirect prompt injection. Ingestion points include external professor-supplied lab briefs and handout documents ingested by the lab_designer_agent. The agent team possesses broad capabilities including file writing and code execution, and there are no explicit boundary markers or sanitization processes mentioned to prevent instructions in the ingested data from influencing these actions.

lab-forge