Use this skill whenever you run or debug local PySpark pytest suites from the agent shell. Invoke `scripts/run-pyspark-tests.py` instead of calling `pytest` raw, so the full output goes to a log file and only a bounded digest reaches your context. The digest carries the exit code, counts, runnable failing node ids, and failures deduplicated by signature, so a suite where many tests fail with the same error collapses to one short block instead of flooding the window. Match the run to the suspected blast radius: when one change likely broke many tests, run the wider scope without `-x` so the digest groups the failures and you fix the shared cause in one pass; add `-x` only to drill on a single suspected failure.
Use this skill whenever you need to query a Databricks workspace from the agent shell — schema discovery (DESCRIBE, SHOW), row sampling, aggregations, EXPLAIN, anything that reads but must not mutate. Invoke `scripts/dbx-ro-query.py` instead of calling `databricks experimental aitools tools query` raw, so destructive SQL cannot run by accident. The wrapper allow-lists SELECT / WITH / SHOW / DESCRIBE / DESC / EXPLAIN, blocks every destructive verb after stripping comments and quoted strings, and emits agent-friendly output (auto / scalar / lines / csv / tsv / json) with a shape-aware default that returns 1x1 results as scalars and Nx1 results as lines for token efficiency.