model-writtenrun 5809 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Justin Pryzby reported that repeated pg_upgrade failures append misleading old errors to per-database logs and proposed unlinking those logs before dumping in #1.
- Bruce Momjian questioned discarding previous logs in #2, while Tom Lane favored gathering output in a timestamped directory to simplify cleanup in #4, an approach Bruce supported in #5.
- Andrew Dunstan required a predictable directory name for buildfarm collection in #6, Michael Paquier favored a configurable path in #8, and Daniel Gustafsson defended separate diagnostic files for existing wrappers in #10.
- Justin replaced log deletion with directory-prefixed output paths in #11, preserving relative command-line paths and leaving the old-cluster deletion script outside because it must survive successful cleanup.
- Michael requested a command-line option and documentation in #12, leading to #13; Tom then proposed separating less-visible logs from user-facing output in #15, and Justin added log and dump subdirectories in #16.
- Windows CI exposed duplicate directory creation during restricted-token re-execution, so Justin moved token acquisition before option parsing in #17.
- Michael recommended separating filesystem initialization from option parsing, quoting dump paths and simplifying cleanup with rmtree() in #18, which Justin implemented as a two-patch submission in #19.
- Michael took over revision work in #22, combining the patches, introducing --outputdir and explicit path validation, and storing separate log and dump paths; #23 fixed the documentation build, while #25 addressed Justin's validation, permissions and error-message findings from #24.
- The configurable path prompted disagreement over whether mkdir() failure already protected recursive cleanup in #26 and #27, until Peter Eisentraut proposed locating output inside the target cluster in #28.
- Michael adopted Peter's proposal in #30, removing the option and its validation in favor of pg_upgrade_output.d under the new cluster; Bruce's cleanup questions were resolved by confirming automatic removal on success unless --retain is used in #31, #32 and #33.
- Michael arranged buildfarm log-collection changes in #38, and Andrew reported merging them while warning that client deployment would take time in #39.
- Michael committed #30 as 38bfae3652, with a permissions-ordering fix so generated files inherit the new cluster's group-access settings, and reported successful CI in #41.
- After landing, Julien Rouhaud raised concerns about buildfarm client rollout in #42; Michael clarified in #43 that older clients would still run tests but miss failure logs, and Andrew acknowledged an unannounced release and criticized landing before widespread deployment in #48.
- Michael offered a temporary revert in #49, but Tom saw no need in #50, leaving the change landed; the thread ended with Andrew explaining the delayed release announcement in #52 and Julien discussing mailing-list moderation in #53, with no further patch or revert recorded.