SQLFluff Redshift GRANT: Copilot’s “safe to merge” fix underwent human rewrite

Primary record (September 29, 2026 reading): original Redshift bug report #7361, agent-authored PR #7563, release 4.1.0.

A Redshift user reported on January 14, 2026 that SQLFluff rejected a valid GRANT statement with multiple GROUP targets, supplying a reproducer and noting a possible analogous problem with ALTER DEFAULT PRIVILEGES. Maintainer peterbud dispatched Copilot March 2; Copilot’s first fix changed the parser and added fixtures in around ten minutes of visible agent session time. Greptile’s bot summary rated it 5/5 and safe to merge. Peterbud marked ready for review and initially approved. But after reviewer keraion’s March 10 review, peterbud pushed six commits including “Simplify and add more tests,” “Remove FunctionSegment,” and “Replicate ANSI GRANT statement with multiple TO parts” (other commits merged main into the branch). WittierDinosaur approved and merged the resulting nine-commit PR March 14; release 4.1.0 of March 26 lists it. The visible page does not expose the full text of keraion’s inline comment, so do not claim the precise request behind each human change or that all six commits altered behavior. The reporter’s acceptance, cost and later defects are not in the public issue/PR/release chain.

What it changes: the bot’s high confidence and the first human approval did not constitute a final implementation. Attribution based solely on PR author would obscure subsequent human work and additional review; the agent’s ten-minute session is not end-to-end elapsed engineering time. This is a separate case from the DuckDB MAP follow-up, useful when examining the same repository’s handoff under different reviewer decisions. Related: economics and review autonomy.