The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill extracts sensitive session cookies from a running browser via the Chrome DevTools Protocol (CDP). Scripts such as scripts/resolve_pdf.py and scripts/refresh-auth-from-dia.sh retrieve cookies for domains including google.com, drive.google.com, virginia.edu, nyu.edu, and ssrn.com, and save them as plaintext JSON files in the ~/.claude-work/skills/paperpile/cookies/ directory.
[PERSISTENCE]: The skill establishes persistence by instructing the user to configure system-level scheduled tasks. scripts/warmup.sh and references/proxy_urls.md describe setting up a crontab entry to run session-refresh scripts every 25 minutes and a launchd agent (com.paperpile.warmup.plist) to ensure the automated browser instance remains active across sessions.
[COMMAND_EXECUTION]: The skill performs extensive shell command execution using Python's subprocess module and bash scripts. It invokes external binaries such as curl, bun, node, paperpile, and scholar to manage network traffic and library files.
[DYNAMIC_EXECUTION]: The skill uses browser automation (CDP) to interact with websites dynamically. It utilizes Runtime.evaluate to inject and execute JavaScript within the browser context for tasks such as scraping PDF links from the DOM and automating authentication button clicks.
[INDIRECT_PROMPT_INJECTION]: The skill has an attack surface for indirect prompt injection because it scrapes data from external publisher websites to influence browser navigation and tool behavior.
Ingestion points: Data is ingested via DOM scraping in _extract_publisher_pdf_url inside scripts/resolve_pdf.py.
Boundary markers: None identified.
Capability inventory: Subprocess execution, filesystem writing, and browser navigation.
Sanitization: None identified for the scraped URLs before they are passed to navigation commands.

paperpile