model-writtenrun 8873 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Jan 2020 · 16 entries
- The thread began with Tom Lane announcing experimental TAP coverage for psql keyword and query-driven completion, using interactive sessions and skipping installations without IO::Pty #1; when Michael Paquier proposed making that module mandatory #3, Tom retained optional support because Windows and some older systems could not provide it #6.
- Christoph Berg reported a failing backslash-command test on Debian unstable #2, and Tom traced it to libedit prematurely removing backslashes #4; ddd87d5645 implemented his workaround of reading the original text from rl_line_buffer.
- Tom requested readable control-character diagnostics #5, and Dagfinn Ilmari Mannsåker supplied Data::Dumper-based output, visible failure diagnostics, and caller-line reporting #8; Tom committed that patch with a less restrictive backspacing assertion as 56a3921a2f, announcing the combined fix in #11.
- Discussion of Debian's runtime substitution of readline for libedit exposed a separate compatibility concern: Tom warned that forthcoming filename-completion code could select inappropriate hooks when compile-time headers and runtime libraries differed #9, while Christoph explained the packaging arrangement and Robert Haas opposed breaking it #10 #12.
- Peter Geoghegan then reported that multiple-choice completion still failed on his Ubuntu workstation because the output contained terminal escapes #14; Tom's first matching patch failed Peter's test #22 #23, but allowing mixed spaces and escapes passed #26 #27 and landed unchanged as fac1c04fec.
- Forcing TERM to xterm immediately caused new buildfarm failures #29, so Tom added temporary environment logging in 60ab7c80b4; its results confirmed that affected machines normally ran without TERM set #32.
- Tom proposed unsetting TERM and LS_COLORS and removing explicit escape matching #33, but Peter still reproduced the original failure #34; 7e42478186 therefore landed only the immediate removal of forced TERM while the broader replacement remained under investigation.
- An apparent cure from reinstalling terminal data proved misleading #35 #38: Peter rebased Tom's proposed replacement and reproduced the failure again #37, then traced it to colored-completion settings in his personal .inputrc #39.
- Tom proposed suppressing .inputrc with INPUTRC=/dev/null #40, Peter verified that this worked with the rebased patch #42, and 48e03583cd combined that isolation with the TERM-unsetting patch from #37, abandoning explicit escape-sequence matching.
- Ilmari supplied a separate cleanup using qr// patterns, like() for the banner, and failed-pattern diagnostics #46; 8c081a2f4e landed it with minor formatting edits and a portability correction to an accidentally tightened assertion, explained by Tom and acknowledged by Ilmari #47 #48.
- Tom reported that upstream libedit had fixed the premature-dequoting bug and hoped eventually to remove PostgreSQL's temporary workaround; he declined to weaken trailing-space and ordering checks merely to accommodate ancient broken libraries #49.
- Christoph reported success on several Debian releases but substantial failures on Ubuntu Xenial and an unrelated recovery-test failure on Bionic #50; he requested an explicit way to disable completion tests for supported distributions with problematic libraries #52.
- After reproducing the Xenial problem, Tom found that its libedit exposed the previous input line through rl_line_buffer, making the new workaround worsen an existing incompatibility even though reverting it would not restore full correctness #55.
- The Bionic report also opened a recovery-testing investigation: Tom connected buildfarm startup failures to an artificially low file-descriptor limit and descriptors inherited from cron #59, rejected forcibly closing inherited descriptors #61, and Amit Kapila cautioned that removing the test might conceal separately observed crashes #62.
- Robert Haas questioned the continuing portability cost #63, and discussion settled on an independent test-only environment switch rather than overloading with_readline #66 #67 #68; Tom implemented SKIP_READLINE_TESTS in e7ee433137 and announced it for affected package and buildfarm configurations #69.
- By the thread's end, Christoph confirmed that skipping completion tests resolved the Xenial build problem #70 and that reverting the separate recovery test eliminated the Bionic failure #71, which Amit acknowledged #72; the completion fixes had landed, while removal of the temporary libedit workaround remained a future intention.