DuckDB, summarized in your digest

DuckDB v2.0 CLI adds agent mode, automatically detecting AI coding agents and producing compact Markdown tables, explicit truncation markers, JSON errors, compact query plans, cost estimates, progress updates, and early query termination. In a Claude Code test answering 22 TPC-H questions, all 132 runs were correct while output tokens fell 59% (123.6k to 50.8k), though overall cost and runtime barely changed because system prompts dominated input. Explicit format flags and CLI settings override the mode.

DuckDB can serve as a tiny, shareable catalog without storing any rows: a database file containing only view definitions can point to remote Parquet files, remaining a few hundred kilobytes regardless of dataset size. Publishers share one URL, while consumers attach it read-only and query stable named relations; DuckDB fetches only needed columns and row groups. Publishers can repartition, relocate, or rename source files by updating views, though access, network performance, and remote read-only limitations remain.