The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes citation metadata from external research databases (CrossRef, PubMed, arXiv, OpenAlex) which could contain malicious strings.
Ingestion points: Metadata is ingested via API responses in extract_metadata.py, search_pubmed.py, search_openalex.py, and search_google_scholar.py.
Boundary markers: SKILL.md contains a prominent security section titled 'Treat extracted metadata as untrusted' that warns against shell injection.
Capability inventory: The skill uses requests for network access and Write/Edit tools for file management. It suggests using subprocess for execution.
Sanitization: scripts/_common.py includes a sanitize_key function and SKILL.md provides specific code examples for safe argument escaping and Python subprocess usage without shell=True.
[EXTERNAL_DOWNLOADS]: The skill fetches citation metadata from well-known academic APIs including api.openalex.org, api.crossref.org, eutils.ncbi.nlm.nih.gov, export.arxiv.org, and api.datacite.org. These are legitimate research infrastructure sources and no executable code is downloaded from these endpoints.
[CREDENTIALS_UNSAFE]: The scripts read optional environment variables (NCBI_API_KEY, NCBI_EMAIL, OPENALEX_EMAIL) to raise rate limits or identify callers to academic services. These are standard practices for these specific APIs and are documented in the skill's frontmatter.
[DYNAMIC_EXECUTION]: The bundled scripts use __import__("pathlib") for cross-platform path resolution during module setup. This is a benign use of dynamic loading for standard library components.

citation-management