The Agent Skills Directory

Local Data Processing: The skill is designed to operate on the user's local filesystem. It uses a local SQLite database and Node.js server to handle document extraction and storage, ensuring that sensitive contract data does not leave the local environment unless explicitly configured otherwise by the user.
Path and Access Control: The system implements strict controls over filesystem paths. Paths enter only through the corpus_register tool and are canonicalized to prevent traversal attacks. Worker agents are restricted by a tool allowlist that prevents them from accessing arbitrary files or performing unauthorized write operations.
Verification Mechanisms: The skill uses schema triggers to verify that citations match the original document content at the time of insertion, providing a structural guarantee against fabricated evidence.
Subagent Sandbox Design: The documents-reader subagents are purpose-built with limited capabilities, holding only the permissions necessary to read shards and file findings, which mitigates the risk of privilege escalation within the agentic flow.
User Oversight: The architecture includes a mandatory 'Plan' phase where the AI must present its intended scope and assumptions to the user for approval before performing any resource-intensive reading or processing.

contracts