model-writtenrun 5361 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Haiying Tang proposed making query-based psql completion accept uppercase input in #1, initially downcasing inputs and adjusting returned names using keyword-case rules.
- Tom Lane objected to indiscriminate case-insensitive behavior in #2, and Kyotaro Horiguchi supplied a SET/RESET-specific alternative in #3 that Tang extended to SHOW in #4.
- Peter Eisentraut argued that the problem affected query-based completion generally and requested analysis of all callers in #6, leading Tang to return to the original approach and extend it in #7.
- David Zhang's testing exposed incomplete uppercase-table completion in #8, and Peter Eisentraut requested quoted-identifier coverage in #10; Tang added referenced-name handling, quoting fixes, and tests in #9 and #11.
- Peter Smith found intermediate-string leaks and cleanup issues in #12, which Tang addressed in #13 while adding tests for empty input.
- Kyotaro Horiguchi challenged applying COMP_KEYWORD_CASE to identifiers and reported broken schema-qualified quoting in #14 and #16, leading #21 to preserve identifier spelling independently of keyword-case settings.
- Discussion of locale-sensitive case conversion in #15, #18, #19, and #20 led Tang to restrict conversion to selected ASCII inputs in #22.
- Peter Eisentraut requested a simpler conversion helper and warned about case-sensitive enum labels and extension versions in #23; #24 refactored the helper and added an enum test, while leaving the enum-completion concern unresolved.
- After a several-month gap, platform-dependent enum ordering reported in #25 was avoided by testing one value in #27, and Japin Li's comment correction in #28 was incorporated in #29.
- Tom Lane argued in #31 that completion should follow backend identifier folding and handle valid quoted spellings, while Peter Eisentraut exposed encoding and enum-sensitivity errors in #34; Tang revised those areas in #33 and #35.
- Tom Lane took over the implementation in #37, replacing backend quote_ident() processing with client-side identifier parsing and requoting, and dropping the earlier tests because their terminal-control expectations were not portable.
- The four-patch series in #39 separated keywords from query results, extended SchemaQuery to handle referenced names, and added replacement tests; ambiguous multiword type-name completions were omitted because readline lost the necessary context.
- Tang questioned surprising inserted quotes and mixed-case matches in #40, and Tom clarified in #41 that supporting legal variant spellings was the primary goal rather than automatically supplying every missing quote.
- Tom's revision in #42 restricted quote-requiring matches for nonempty unquoted input, improved multibyte and reserved-word handling, and added LIKE prefix matching so catalog searches could use indexes; #43 rebased the series.
- Tang accepted and tested the revised behavior in #44 and #45, and Tom landed 02b8048ba5 from #43 with additional text-search schema support, naming cleanup, documentation, and a keyword test.
- Dagfinn Ilmari Mannsåker proposed keyword-case preservation and a separate NULL-check cleanup in #47, but Tom objected in #48 to losing visually distinct keywords in mixed completion menus and declined the isolated cleanup of longstanding NULL-check conventions.
- Tom proposed applying keyword-case preferences only when no object-name matches remained in #49, and Ilmari accepted that compromise while supplying uppercase keyword lists and SHOW SESSION AUTHORIZATION completion in #50.
- The combined changes from #49 and #50 landed as 020258fbd3, with a counter rename and a clarification of readline quoting limitations.
- Tom reported in #52 that mixed keyword and object-name casing prevented useful common-prefix completion, and f0cd9097cf subsequently restored Ilmari's unconditional keyword-case adjustment while also fixing qualified GUC names through verbatim completion.
- Months later, Tang reported leaked parsed names and supplied #53; Tom incorporated that patch and an additional patternToSQLRegex cleanup discovered with Valgrind in 90474c16a7.
- The thread ends with Tom confirming the leak fixes were pushed in #54; the completion feature and its follow-up fixes had landed, while the separate NULL-check cleanup from #47 remained unaccepted.