model-writtenrun 904 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Jelte Fennema-Nio opened the thread with four improvements exposing regression diffs, failed-command output, Perl exceptions, and caller locations directly in test output, accompanied by deliberately failing demonstrations in #1.
- Zsolt Parragi identified stale command-helper documentation in #2, while Andres Freund argued for bounded output, less tempfile overhead, shared formatting, safer exception handling, and general croak-based caller reporting in #3.
- Jelte incorporated that feedback in #5, making regression diff output automatic but bounded, sharing truncated command diagnostics, fixing pg_ctl restart redirection, improving the die handler, and replacing localized caller tracking with croak throughout the helper modules.
- Nazir Bilal Yavuz approved the revised Perl patches but identified incorrect counting of long diff lines in #6; Jelte introduced continuation-aware diagnostic output in #7 and explained why the descriptor hang was specific to daemon-spawning commands.
- Andrew Dunstan gave the series an initial positive review in #8, and Zsolt reported that successful commands had lost their command-start log lines in #9.
- Jelte restored command logging, fixed Windows file-locking and line-count errors, protected done_testing() with eval, and added command_ok conversions for the pg_upgrade and streaming regression tests in #10.
- Corey Huinker tested intentional regression failures and approved the series in #11; Jelte made editorial adjustments in #12, and Corey confirmed clean application and passing tests in #13.
- Andrew committed the pg_regress diagnostics, failed-command diagnostics, and Perl die handler from #12 as 5720ae0143, 1402b8d2fc, and 76540fdedf, retaining their implementation apart from whitespace and rewriting their commit messages.
- Andrew also committed the test-script conversions from #12 as bb6ae9707c and the croak conversion as b8da9869b8, adding @CARP_NOT so errors traversing Cluster.pm wrappers identify the test-script caller, as explained in #14.
- Peter Eisentraut objected that wrapped diff lines overwhelmed terminals and displaced useful summaries in #16; after Jelte defended local diagnostics and suggested configurable limits in #17, Peter explained that crash-related diffs often show collateral failures and highlighted interleaved output in #18.
- Andrew proposed making diff extraction opt-in through PG_REGRESS_DIFF_EXTRACT and enabling it in Cirrus CI in #19, but this proposal was not shown landing in the thread.
- Alexander Lakhin reported that truncating command output hid the failing test's timing in #20, revealing a separate drawback of the command_ok conversions.
- Jelte offered an alternative follow-up in #21 that suppresses diffs after test-process failures, emits one summary at the end, introduces PG_REGRESS_DIFF_LINES, accounts for wrapping at 80 columns, and lowers the default limit to 50; he deferred broader backend-crash diagnostics to future work.
- Jelte acknowledged in #22 that the missing timings remained unresolved and considered optional command_ok truncation controls; Andrew questioned why TAP tests could not enable extraction themselves in #23, and Jelte preferred useful default behavior over blanket enabling in #24.
- The original five changes had landed, but the follow-up output policy remained under discussion when Christoph Berg requested a show-everything mode for Debian builds that preserve logs but not test artifacts in #25.