The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests and processes research papers from external URLs and local files, creating a potential surface for indirect prompt injection.
Ingestion points: Data enters the system via toolchains/fetch_paper.py (which fetches from arXiv, DOI, and generic URLs) and toolchains/read_any_file.py (which reads various local file formats).
Boundary markers: The skill uses triple-angle bracket delimiters (<<< >>>) in prompts like prompts/parse-paper.md to isolate external content from instructions.
Capability inventory: The agent context includes filesystem-write for managing a local paper cache and shell-exec for environment anchoring.
Sanitization: The toolchains perform structural cleaning (e.g., stripping HTML tags and LaTeX commands), but do not include semantic filtering for adversarial instructions embedded within the text.
[DYNAMIC_EXECUTION]: The universal file reader uses dynamic module loading to manage optional dependencies.
Evidence: The _opt function in toolchains/read_any_file.py utilizes __import__ to check for the presence of libraries such as pandas, pdfplumber, and beautifulsoup4.
Context: These modules are loaded using hardcoded static strings to provide functionality for specific file formats (like Excel or PDF) only when the respective libraries are installed by the user.
[COMMAND_EXECUTION]: The skill executes simple shell commands to determine system state for freshness checking.
Evidence: instructions/freshness.md specifies the use of date -u +%Y-%m-%d and stat -c %y.
Context: These are benign operations used to verify the publication recency of research papers and to anchor temporal searches.

read-research-paper