The Agent Skills Directory

[PROMPT_INJECTION]: The skill acts as an orchestration layer that ingests external project data (e.g., project_state.json, STATUS.md, PROJECT.md) to determine next steps. This ingestion of untrusted data while having access to tools like Bash and file writing represents a surface for indirect prompt injection.
Ingestion points: project_state.json, STATUS.md, CLAIMS.md, REVIEW_LOG.md, and user-provided medical data files (CSV/Excel).
Boundary markers: Delimiters for untrusted data are not explicitly enforced in the orchestration logic itself, though the skill routes to specialized tools.
Capability inventory: Access to Bash, Write, Edit, and local scripts (scripts/validate_skill_contracts.py).
Sanitization: The skill implements a 'PHI Safety Gate' to ensure clinical identifiers are removed before further processing, which serves as a pre-processing validation layer.
[SAFE]: The skill contains an explicit 'PHI Safety Gate' (Dialogue Node N6) that requires human confirmation or a de-identification step before any sensitive medical data is processed by LLM-based tools, effectively mitigating data exposure risks.
[SAFE]: A 'Tier-3 Worker Guard' is implemented to restrict high-risk autonomous operations (e.g., sending emails, external code pushes, or publishing to repositories) during end-to-end runs, ensuring human oversight for external communication.

orchestrate