model-writtenrun 3721 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Quan Zongliang proposed implementing the PL/pgSQL TODO for %TYPE arrays, initially accepting only one empty bracket pair and deferring richer syntax because of implementation complexity #1 #6.
- Daniel Gustafsson welcomed the feature and questioned the limited syntax, while Pavel Stehule explained that PostgreSQL array dimensions and sizes depend on values rather than declarations; Quan then added explicit errors for unsupported cases #2 #3 #4.
- Pavel proposed a separate %ELEMENTTYPE facility, and Quan explored delegating richer array syntax to parse_datatype(), but neither proposal became an attached implementation in this thread #7 #8.
- After a month, Pavel supplied an alternative implementation that centralized array-type construction and supported %ROWTYPE, ARRAY, repeated brackets, and dimension specifications; his review also identified missing documentation and poor errors for RECORD references #10.
- Quan adopted Pavel's implementation, fixed its array-type validity check, and added documentation and tests, while proposing that RECORD-related %TYPE support be handled separately #11.
- Pavel and Quan refined documentation and regression coverage, corrected invalid %ROWTYPE examples, and added a missing-array-type error test; Pavel reported passing tests and declared the patch ready for committers #12 #13 #14.
- Tom Lane committed the array feature from #14 as 5e8674dc83, rewriting documentation and comments, making small code cleanups, and relocating the tests, as he reported in #15.
- Tom then supplied a separate patch allowing %TYPE to reference composite variables and removing undocumented bare type-name lookup; he left RECORD arrays unsupported and suggested better unknown-name errors as future work #15.
- Pavel endorsed the cleanup in the thread's final message #16, and Tom subsequently committed #15 as 43b46aae12 with an additional record-name qualification check and expanded tests; both submitted changes landed, while improved unknown-name diagnostics and RECORD-array support remained outside the committed work.