datarefresh

name: data/refresh description: Refresh a stable BigQuery-backed Parquet snapshot in the workspace and regenerate recurring dashboard or report artifacts from DuckDB. tools: - export_bigquery_query - data_agent_tools

data/refresh

Use this skill when the user wants a repeatable snapshot job rather than a one-off answer.

Workflow

1. Materialize the warehouse slice with materialize_bigquery_to_parquet.

   • Prefer datasets/<slug>/latest.parquet as the stable target path.
   • Use emit_csv=true only when a CSV mirror is explicitly useful.
   • Treat BigQuery as the source of truth and DuckDB as the local serving layer.

2. Validate the refreshed snapshot locally.

   • Call get_table_schema for the refreshed table.
   • Run one or two focused run_sql_query checks for freshness and row sanity.

3. Regenerate the recurring artifact.

   • Use generate_dashboard for an interactive HTML dashboard.
   • Use generate_summary for a canned HTML report.
   • Always pass a stable output_path so scheduled refreshes overwrite the same file.

4. Close with the refresh contract.

   • Tell the user which snapshot path was updated.
   • Tell the user which HTML path was regenerated.
   • If appropriate, suggest automating the exact prompt on a daily schedule in the user’s timezone.

Defaults

• Scheduled snapshots, not live warehouse HTML
• Canonical data artifact: Parquet + manifest
• CSV mirror is optional compatibility output, not the primary serving layer