The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes content extracted from external PDF files and interpolates it directly into the literature-reader-agent.md prompt, creating a surface for indirect injection.
Ingestion points: The markdown_content variable in agents/literature-reader-agent.md receives the full text of documents processed by the MinerU API.
Boundary markers: There are no delimiters or "ignore" instructions surrounding the untrusted content in the prompt template, allowing documents to potentially override agent behavior.
Capability inventory: The skill includes scripts that write files and directories to the local project structure (scripts/convert_pdfs_to_md_mineru_api.py).
Sanitization: No sanitization or filtering is performed on the Markdown content before it is read by the agent.
[EXTERNAL_DOWNLOADS]: The script scripts/convert_pdfs_to_md_mineru_api.py downloads conversion results from the MinerU API.
Evidence: The download_file function fetches ZIP archives from URLs provided by the mineru.net service.
Mitigation: The skill implements path traversal checks in the safe_extract_zip function to ensure files are only written within the designated output directory.
[COMMAND_EXECUTION]: The skill uses system commands to support its download functionality.
Evidence: scripts/convert_pdfs_to_md_mineru_api.py invokes curl via subprocess.run if the standard library's urllib fails.
Security: Commands are executed using a list of arguments without a shell, which prevents shell injection vulnerabilities.
[DYNAMIC_EXECUTION]: The skill performs local dynamic loading of its own scripts.
Evidence: scripts/convert-pdfs-to-md.py uses runpy.run_path to execute the main conversion logic in convert_pdfs_to_md_mineru_api.py. This is a standard method for delegating execution between bundled scripts.

academic-citation