The Agent Skills Directory

[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection due to its ingestion of untrusted data from external blueprints and URLs. Maliciously crafted instructions within these inputs could attempt to manipulate the agent's logic during the reconciliation process. Ingestion points: Processes blueprint files (Stage 1) and fetches raw bytes from user-provided URLs (Step 0). Boundary markers: Employs structured parsing of blueprint sections and specific header names to delimit data. Capability inventory: Utilizes bash for CLI interactions (semantius, yq, bun) and possesses file-writing capabilities for generating specifications. Sanitization: Implements a comprehensive Pre-save verification gate in Stage 11 that scans for DDL tokens, identifier leakage, and formatting violations to ensure the safety of the final output.
[COMMAND_EXECUTION]: The skill makes extensive use of shell commands to perform catalog lookups, update configurations, and run validation scripts. Evidence: Invokes semantius call crud for database operations, yq for managing customizations.yaml, and bun to execute a local consistency-check script.
[EXTERNAL_DOWNLOADS]: The skill includes functionality to download blueprint artifacts from remote sources. Evidence: Uses curl -s in Step 0 to retrieve raw blueprint data when a URL is provided by the user.
[SAFE]: The skill demonstrates a strong security design by incorporating mandatory, non-silent verification gates that must pass before any specification file is saved, effectively mitigating risks associated with malformed or malicious data processing.

semantius-analyst