The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements a logging mechanism that ingests untrusted data from request headers and metadata objects to be stored in a database.
Ingestion points: The logAudit() function in assets/audit-log.ts processes input.redactedMetadata and req headers (specifically user-agent and x-forwarded-for).
Boundary markers: The implementation does not use prompt delimiters, but it isolates untrusted data by mapping it to specific database columns or including it in a cryptographic SHA-256 hash chain.
Capability inventory: The skill is capable of performing database write operations via the Prisma ORM.
Sanitization: The extractAuditClientIp function uses isIP from node:net to validate that the extracted IP address is a valid network address before processing or storage.
[COMMAND_EXECUTION]: The TypeScript implementation uses a raw database execution method to manage concurrency.
Evidence: In assets/audit-log.ts, db.$executeRawUnsafe is called to execute SELECT pg_advisory_xact_lock(hashtext('lgpd_audit_log_chain')).
Context: The query is a hardcoded static string used for acquiring a PostgreSQL advisory lock, which is a recommended practice for maintaining a serialized hash chain in high-concurrency environments. No user-controlled input is passed to the raw query.

lgpd-audit-logging