The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes paper metadata (titles, abstracts) fetched from external academic APIs which could contain adversarial instructions.
Ingestion points: Paper data is retrieved via scripts/providers/openalex.py, scripts/providers/semantic_scholar.py, and scripts/providers/crossref.py.
Boundary markers: The skill outputs structured JSONL files (candidates_normalized.jsonl, candidates_deduped.jsonl), providing structural separation, though it does not explicitly wrap retrieved text in instruction-ignore delimiters.
Capability inventory: scripts/search_runner.py performs file system writes to a designated bundle directory and makes network GET requests to academic providers.
Sanitization: scripts/candidate_schema.py validates incoming data against a strict schema (rls.paper.v1), and scripts/search_runner.py includes a _safe_error function to redact credentials from logs.
[SAFE]: The skill demonstrates security best practices by implementing path traversal checks in scripts/search_runner.py (_safe_bundle_dir) and scripts/manifest.py (safe_relative), ensuring operations remain within the authorized scope.
[SAFE]: Network operations are limited to well-known, legitimate academic service providers (OpenAlex, Semantic Scholar, Crossref) using the standard requests library.

research-literature-search