The Agent Skills Directory

[COMMAND_EXECUTION]: The main script scripts/fetch.py uses subprocess.run to execute a companion script scripts/cloak_pdf.py. It also runs a Python check to verify the presence of the cloakbrowser dependency. While these calls avoid shell=True and incorporate argument quoting via shlex, they represent a command execution surface that relies on the integrity of the environment (e.g., the CLOAKBROWSER_PYTHON variable).
[DYNAMIC_EXECUTION]: The helper script scripts/cloak_pdf.py executes dynamic JavaScript using page.evaluate within a browser environment. This mechanism is used to perform in-page fetches to bypass anti-bot protections like Cloudflare challenges.
[INDIRECT_PROMPT_INJECTION]: The skill processes data from multiple external and potentially untrusted sources.
Ingestion points: The skill fetches metadata such as paper titles and author names from various APIs (Crossref, Semantic Scholar, Unpaywall, arXiv) and scrapes content from Sci-Hub mirrors.
Boundary markers: None. The ingested strings are placed directly into JSON objects returned to the agent.
Capability inventory: The skill has the ability to write files to the local disk, perform network requests, and execute subprocesses.
Sanitization: While the skill includes excellent SSRF validation for URLs and binary validation for PDF files, it does not sanitize text-based metadata fields that are eventually returned to the AI agent.
[EXTERNAL_DOWNLOADS]: The skill fetches data and files from a variety of external research and preprint platforms. It mitigates SSRF risks by implementing a strict validation layer that checks all target hostnames and resolved IP addresses against private and restricted ranges before initiating any connection.

paper-fetch