The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external PDF files from arXiv using the scripts/extract_paper.py tool. This creates a surface for indirect prompt injection where a maliciously crafted paper could contain instructions intended to influence the agent's behavior.
Ingestion points: scripts/extract_paper.py reads and parses untrusted PDF files downloaded from the web.
Boundary markers: The SKILL.md template lacks explicit boundary markers or instructions telling the AI to ignore commands embedded within the extracted text.
Capability inventory: The skill has the ability to write to the file system, make network requests for market data, and execute internal scripts via subprocess.
Sanitization: Path sanitization is present for paper IDs, but the extracted content is passed directly to the agent's context.
[COMMAND_EXECUTION]: The scripts/run_research.py script orchestrates the workflow by using subprocess.run to execute other Python scripts in the package (search_arxiv.py, extract_paper.py, reproduce_paper.py). While this involves shell execution, the command arguments are passed as lists and targeted at internal package files, which is a standard method for script chaining.
[EXTERNAL_DOWNLOADS]: The skill performs automated downloads of research papers from arxiv.org and market data from sina.com.cn (via the akshare library). These target well-known, reputable services for the intended research purpose.

paper-replication