The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes external processes via subprocess.run to perform its core functions securely. In scripts/new_report_skill.py, it executes a validator script to verify the report contract before generation. In assets/run_report.py, it utilizes the databricks CLI to run SQL statements against a warehouse. To prevent command injection in generated reports, the tool includes a safe_token function that rejects inputs containing backticks, dollar signs, or backslashes.
[EXTERNAL_DOWNLOADS]: The generated report template (assets/report-template.html) references the Apache ECharts library from unpkg.com, which is a well-known CDN for JavaScript packages. This allows reports to render interactive visualizations while adhering to a strict Content Security Policy that blocks unauthorized network connections.
[INDIRECT_PROMPT_INJECTION]: The skill architecture creates an ingestion point for external data from databases which is then summarized by the agent. To mitigate risks, the system implements mandatory HTML escaping in assets/render_report.py for all database content and AI-generated summaries, and provides specific instructions to the model to ground its summaries in the provided data.
[SAFE]: The skill implements a robust integrity protection mechanism using a contract.manifest.json file. The verify_manifest function in the report runner calculates SHA-256 hashes of the SQL queries and refuses to execute if the files have drifted from the generated state, protecting against unauthorized modifications to the report logic.

report-skill-builder