The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted CSV files to drive its workflow, creating a potential attack surface for indirect prompt injection.
Ingestion points: External CSV data enters the agent context through the utils/get_csvschema tool (introspection) and the csv-parse library in the staged import script.
Boundary markers: The skill uses mapping review tables and structured question gates (AskUserQuestion) to handle data-driven decisions, though it does not explicitly define delimiters for raw CSV content in all prose outputs.
Capability inventory: The skill can perform catalog writes (create_module, create_entity, create_field) and business data writes (postgrestRequest) via the semantius CLI, and execute local staged scripts using Bun.
Sanitization: The import script (import.ts) performs type-based coercion (integer, number, boolean, date) and validation against the mapping schema before data is sent to the API.
[COMMAND_EXECUTION]: The skill executes shell commands and local TypeScript scripts to perform data processing.
Evidence: Uses bun run to execute staged helpers (import.ts, create-fields.ts, render-plan.ts) and semantius call to interact with the platform CLI.
Context: These executions are limited to the skill's own template files and the vendor's CLI, serving the primary purpose of database automation.
[EXTERNAL_DOWNLOADS]: The skill downloads and installs a standard third-party dependency at runtime.
Evidence: bun add --cwd <run-folder> csv-parse is called during the workspace setup phase.
Context: The download targets a well-known, legitimate package (csv-parse) from a standard registry.
[DYNAMIC_EXECUTION]: The skill stages and runs local script templates based on runtime configurations.
Evidence: Script templates (import.template.ts) are copied to a temporary work folder and executed with bun. All execution parameters are derived from a local mapping.json file generated during the review phase.
Context: This design separates the data processing logic from the agent's direct reasoning, providing a more stable and observable execution environment.

semantius-importer