The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/resolve_exhibits.py utilizes subprocess.run to execute multiple shell commands, including rclone, ssh, scp, and tar. The ssh command specifically uses string formatting to construct a command string executed on the remote host, which includes parameters like --wrds-scratch that could be influenced by input.
[DYNAMIC_EXECUTION]: The skill employs JavaScript injection (01-install-hook.js and 02-run-pager.js) within the browser environment. These scripts monkey-patch XMLHttpRequest to intercept network data and automate user actions like paging through results.
[INDIRECT_PROMPT_INJECTION]: The workflow processes search results from Bloomberg Law and raw filing text from SEC EDGAR, creating an attack surface for instructions embedded in external documents.
Ingestion points: Network responses captured from the Bloomberg Law search API and document text downloaded from the SEC archives via WRDS.
Boundary markers: None identified in the prompt templates or processing scripts.
Capability inventory: Execution of shell commands (rclone, ssh, scp, tar) in resolve_exhibits.py, database queries via psycopg, and local file system write operations.
Sanitization: Uses BeautifulSoup for HTML parsing and regex for phrase verification in the extracted text.
[CREDENTIALS_UNSAFE]: The database connection string in scripts/resolve_exhibits.py contains a hardcoded username for the WRDS PostgreSQL service, which appears to be a vendor-specific resource associated with the skill author.

blaw-edgar-search