The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes the databricks CLI and jq utility via the Bash tool to interact with the Databricks workspace. These operations are scoped to specific binaries and are used exclusively for auditing workspace configuration and billing data.\n- [COMMAND_EXECUTION]: A local Python script, scripts/rank-and-report.py, is executed to perform data aggregation and generate the final report. This ensures that all financial calculations are deterministic and auditable, mitigating risks associated with model-based arithmetic errors.\n- [DATA_EXFILTRATION]: The skill accesses sensitive information from Databricks system tables, including system.billing.usage and system.billing.list_prices. This access is governed by Unity Catalog permissions and is a core requirement for the skill's functionality. The data is processed locally and formatted into a Markdown report for the user, with no evidence of unauthorized external transmission.\n- [PROMPT_INJECTION]: The skill identifies an indirect prompt injection surface by ingesting and reporting on workspace metadata (e.g., cluster names, job IDs).\n
Ingestion points: Data is retrieved from Databricks system tables and workspace API tools such as clusters_get and pipelines_get.\n
Boundary markers: No explicit delimiters are used in the final report template to separate ingested data from agent instructions.\n
Capability inventory: The toolset includes Bash (scoped), Write, Edit, and Read capabilities.\n
Sanitization: The scripts/rank-and-report.py script coerces numeric values but does not perform sanitization on text-based metadata fields for potentially malicious instructions.

databricks-cost-leak-hunter