The Agent Skills Directory

[Secure Input Processing]: The skill processes external infrastructure and catalog files through a security-hardened Python script (scripts/render_catalog_erd.py). This script implements multiple layers of protection to handle untrusted data safely.
Ingestion points: The agent reads user-supplied Terraform, Bicep, ARM, and DS_CATALOG_V1 markdown files to extract relationship data.
Boundary markers: SKILL.md defines clear success criteria and stop rules that prevent the agent from guessing or inferring relationships outside the confirmed source scope.
Capability inventory: The skill is limited to local file reading and executing a specific, well-defined rendering script; it does not perform arbitrary shell execution or network operations.
Sanitization: The renderer enforces strict limits on display text (maximum 120 characters) and utilizes a character blacklist (<, >, {, }, ", etc.) to prevent markup injection or diagram syntax manipulation in the generated Mermaid or ASCII output.
[Path Traversal Protection]: File handling logic in the rendering script includes explicit checks for .. segments and ensures that all resolved paths remain within authorized directory roots before any read operations occur.
[Safe YAML Parsing]: The catalog parser uses a custom loader inherited from yaml.SafeLoader that is further hardened to reject dangerous YAML features such as aliases, anchors, merge keys, and explicit tags, effectively mitigating risks of logic bypass or denial-of-service through YAML structure.
[Resource Constrainment]: A hard limit of 5MB is enforced on input files (MAX_INPUT_BYTES), providing a safeguard against resource exhaustion attacks when parsing large or malformed infrastructure configuration files.

architecture-diagrams