The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads and installs an external software package from an unverified GitHub repository (https://github.com/ChicagoHAI/OpenAIReview.git) during its setup process. This is performed via pip install targeting the repository URL.\n- [COMMAND_EXECUTION]: The skill uses Python's subprocess module to execute shell commands for managing virtual environments, installing packages, and running CLI tools. Evidence is found in scripts/_engine.py and scripts/run_openaireview.py.\n- [DYNAMIC_EXECUTION]: The skill dynamically installs and then executes logic from the openaireview package at runtime by calling its CLI and scripts from a managed virtual environment.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external document data using multi-agent critique workflows, creating a surface where malicious instructions embedded in a paper could influence the agent's behavior. \n 1. Ingestion points: Ingests PDFs, DOCX files, and arXiv URLs into the workspace as full_text.md for analysis. \n 2. Boundary markers: None identified; the skill does not instruct subagents to ignore instructions contained within the analyzed paper text. \n 3. Capability inventory: The skill can write files to the local workspace and execute commands via subprocess.run. \n 4. Sanitization: No evidence of sanitization or escaping of extracted paper text before it is provided to LLM workers.

research-paper-review