The Agent Skills Directory

[Indirect Prompt Injection Surface]: The skill processes untrusted contract text, which could potentially contain malicious instructions intended to influence the agent.
Ingestion points: Untrusted contract files (PDF, Word, Excel, etc.) are ingested into a local SQLite database (data.sqlite) via the corpus_prepare tool.
Boundary markers: The skill wraps document text in <DOCUMENT> tags and executes extraction passes using the --disallowed-tools "*" flag to isolate the agent from its capabilities while reading untrusted content.
Capability inventory: The skill utilizes tools for file system interaction, database operations (sql, write, find), and orchestrates subagents with scoped shell access (Bash).
Sanitization: The engine includes schema triggers and verification logic to ensure that citations and findings are contiguous substrings of the original document text before they are recorded.
[External Resource Downloads]: The skill relies on the @llamaindex/liteparse Node.js package and fetches OCR language data from a remote CDN during its initial setup. These resources are retrieved from well-known public registries and services.
[Local Command Execution]: The skill performs its analysis by executing local scripts and binaries, specifically the sweep.mjs utility and the claude CLI. This orchestration is a standard part of its design for parallel document processing and text extraction.
[Dynamic Orchestration of Subagents]: The skill dynamically spawns specialized subagents (healthcare:documents-reader-mcp, healthcare:documents-reader-cli) to handle complex tasks like visual OCR for scanned documents. These agents are restricted to specific tool allowlists to reduce the attack surface.

contracts