tower-data

Tower Data Analysis

Tower manages Apache Iceberg lakehouses. This skill answers questions about the user's business data with tower catalogs query, which vends short-lived read-only credentials and runs SQL through a sandboxed DuckDB in a single command — no credential handling on your side.

Core loop: read saved knowledge → discover schemas → query → answer → save what you learned back as knowledge.

Fast path (use this)

# 1. What does the team already know about this data? (semantics you can't infer)
tower catalogs knowledge list default -j

# 2. Every namespace, table, and column with types — one call, no credentials needed
tower catalogs show default --full -j

# 3. Query. Always -j.
tower catalogs query default -j --sql 'SELECT * FROM "default".bronze.users LIMIT 10'

# Longer SQL via stdin; non-default environment via -e
tower catalogs query default -j -e production < analysis.sql

Reference tables as <catalog>.<namespace>.<table>. Quote the catalog name — default is a SQL keyword: "default".gold.revenue.

Each invocation vends fresh short-lived credentials internally (~1s overhead), so token expiry and shell-state loss between commands are non-issues. Credentials never appear in output — nothing to redact.

Hard constraints of tower catalogs query

These are enforced by the command and will bite you if you write SQL the normal way. Read all five.

1. One statement per invocation. Multi-statement batches are rejected outright (Only a single SQL statement can be run at a time). There is no session between invocations, so CREATE TEMP TABLE then query is impossible — and it's blocked anyway (rule 2). Use CTEs (WITH) to build multi-step logic inside one statement.
2. SELECT only. Any DDL/DML is rejected unless you pass --write. SELECT, WITH ... SELECT, DESCRIBE, and SHOW ALL TABLES all work.
3. -j silently truncates at 1000 rows. The text output prints a "result truncated" notice; JSON does not — you just get 1000 rows with no signal. Never aggregate client-side over a -j result you didn't bound. Aggregate in SQL, or pass --max-rows 0 (no limit; can exhaust memory) or an explicit --max-rows N when you genuinely need the raw rows.
4. Errors are not JSON. On failure the command prints plain text (Error: Query failed: ...) and exits 1, even with -j. Check the exit code / non-[ first byte before parsing.
5. Always write an explicit AS for aliases. An implicit alias that is a date keyword (SELECT 1 day) misparses in the read-only statement checker and gets rejected as a write. SELECT 1 AS day is fine.