Questions spanning thousands of tables force the model to hand-write 20k-character UNION ALL statements (19% of cost in one batch of sessions) #1489

Problem

A warehouse with one table per entity (2,753 tables) has no tool-supported way to query across a set of tables. Every route that would let the engine iterate over a table list is rejected or classified as a write, so the model generates the full UNION ALL text itself and re-types it into sql_execute.

Mechanism

• DuckDB query_table([...]) rejects subqueries and lateral arguments ("Table function cannot contain subqueries" 36 times, "does not support lateral join column" 34, "Invalid named parameter" 12). An engine limit.
• The statements that work around it (SET max_expression_depth which DuckDB's own error recommends, SET VARIABLE + query_table(getvariable(...)), EXECUTE IMMEDIATE, CREATE TEMP MACRO ... AS TABLE) are denied as writes (packages/opencode/src/altimate/tools/sql-classify.ts:27, :226-235; sql-execute.ts:113-121).
• The model builds the text with string_agg; the single huge result cell is clipped in the preview (sql-result-format.ts:40, :178, :342), so it reads the spill file via bash/read and re-types the SQL (output tokens).

Evidence

• 33 of 44 preview truncations are generated SQL.
• Median output tokens per session on the affected question: 42,355 (Claude Opus) and 23,545 (gpt-5.6-sol), against 362-3,404 on other questions on the same warehouse.
• 246-675 s and $5.98-$10.87 per Opus session, the 8 costliest of a 270-session batch; $58.2 of $307.9 (19%) for the two affected questions; context up to 144k tokens; one session used 675 of an 840 s budget; gpt-5.6-sol sessions ran 621-851 s and contributed to budget misses.
• Generations of 15-25k tokens took 100-160 s each.

Impact

Time and cost, with budget risk. No incorrect answers caused directly in Opus sessions.

Proposed fix

Pick one or more:

• A table-set primitive on sql_materialize / sql_execute (for example tables: "A*" or a list obtained from a scratch query) that expands UNION ALL BY NAME server-side.
• Allow session-local non-writing settings (SET max_expression_depth, threads, memory_limit) on read-only connections, and CREATE [TEMP] MACRO in scratch.
• A query_from: {warehouse, sql} option that classifies the generated text before running it.

Acceptance check

A question over 1,000+ tables completes without any single model output above 2k tokens; median output tokens and cost drop by at least 5x on the affected workload.