The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze untrusted codebase data, which is an inherent vector for indirect prompt injection via malicious comments or documentation within the analyzed system.
Ingestion points: The skill reads project manifests (Setup Step 3), performs dependency graph analysis (Phase 1), and extracts logic/schema from source code (Phase 2).
Boundary markers: Employs a dedicated 'Assayer' (Skeptic) agent to challenge all deliverables for accuracy and completeness before advancing phases.
Capability inventory: Orchestrates sub-agents using TeamCreate and Agent tools; writes documentation artifacts to the local docs/ directory.
Sanitization: Includes a non-negotiable principle (Rule 4) forbidding secrets in context, though it does not explicitly describe sanitization of instruction-like content in code comments.
[CREDENTIALS_UNSAFE]: The skill includes high-value security principles, specifically Rule 4 ('No secrets in context'), which prohibits agents from including API keys, tokens, or PII in prompts, messages, or checkpoint files. It instructs agents to flag discovered secrets by file path/line number rather than value.

unearth-specification