model-writtenrun 6833 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Manu opened the thread with reproductions of PostgreSQL 19 JSON-constructor indexes becoming stale after type or cast changes, attaching Jian's container-mutability fix and his own array fix while leaving older scalar problems outside the proposed repair #1.
- Nikolay Samokhvalov found that the proposed fixes still allowed setting-dependent arrays on REL_19_STABLE and supplied a branch-specific delta, reporting regression-test results while disclosing that its AI-generated code had not been manually reviewed #2.
- Manu confirmed the remaining regression across 24 array variants, validated Nikolay's fix, and noted that rejecting every array would also solve it but unnecessarily reject safe types such as integer and text arrays #3.
- The discussion turned to submission practices: Andres Freund objected to unreviewed AI output, while Andrey Borodin and Kirk Wolak advocated transparency and separating AI material from human discussion; these exchanges did not review the technical changes #4 #5 #6 #7.
- Andrew Dunstan judged the eve of RC2 too late to settle the repair's behavior and proposed reverting the feature, leaving older bugs for another development cycle #8.
- In the thread's last message, Manu endorsed and tested the revert on both branches, confirmed that it rejected the newly introduced stale-index cases, and warned that existing affected objects would remain but their dumps would no longer restore #9.
- Andrew subsequently committed the revert and backpatched it through PostgreSQL 19; neither attached repair landed, and the commit warned that affected indexes, generated columns, and partition keys must be dropped or redefined before upgrading b58a9a97fe.