The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The documentation and examples (SKILL.md, EXAMPLES.md) instruct the agent or user to download an OpenAPI specification using curl -fsS http://127.0.0.1:8090/openapi.json -o openapi.json. While this targets the local loopback address, which is common in development environments for fetching metadata from a running service, it represents an external download operation.
[INDIRECT_PROMPT_INJECTION]: The skill processes external data (OpenAPI JSON files and Firestore rules) which could potentially contain malicious instructions or confusing data if the source is compromised.
Ingestion points: The scripts scripts/sync_openapi.py and scripts/fetch_firestore_schema.py read openapi.json, firestore.rules, and firestore.schema.json from the local file system.
Boundary markers: The instructions do not define explicit boundary markers to prevent the agent from following instructions potentially embedded within these JSON/rules files.
Capability inventory: The skill has the capability to write files (lib/models/api_models.dart) based on the content of the ingested files.
Sanitization: The skill performs basic sanitization by using regex to clean class names (re.sub(r"[^a-zA-Z0-9_]", "", name)) and standard JSON parsing, which mitigates simple code injection but does not prevent logic-based indirect injection.
[COMMAND_EXECUTION]: The skill relies on the agent executing shell commands (python scripts/sync_openapi.py) and provides example commands involving curl. The Python implementation correctly implements path validation using .resolve() and relative_to() to ensure the scripts only access files within the intended project root, preventing path traversal.

dart-type-bridger