model-writtenrun 7453 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- The thread began with Michael Paquier's announcement that tablespace directory recreation had moved from makefiles into pg_regress to prevent parallel pg_upgrade and regression tests from interfering while preserving Windows administrator-account support #1.
- Christoph Berg reported that recreating the directory broke tests against a server running as another user; Michael suggested a cleanup-control option, while Christoph preferred server-side directory creation #2 #3 #4.
- Robert Haas and Michael questioned Windows shell compatibility, and Noah Misch identified shell-metacharacter problems and proposed a C mkdir helper in regress.c #5 #6 #7.
- Michael implemented the SQL-callable helper and tested it on Linux and MSVC, but Noah explained that switching between client and server users could still cause directory-removal failures; Michael subsequently reported successful Windows CI testing #8 #9 #10.
- Christoph supported the helper, but Noah emphasized its limited coverage of different-user scenarios, and Michael shelved the proposal because Debian had a workaround and no further action seemed necessary #11 #12 #13.
- A month later, Tom Lane found unwanted testtablespace directories while preparing beta tarballs and posted a separate patch adding --make-testtablespace-dir, with directory preparation moved out of source-file conversion and explicitly enabled by the core regression launchers #14 #15.
- Michael tested the Windows changes and questioned the new creation timing and duplicated path construction; Tom defended separating creation from input-directory detection, and Michael accepted the explanation, leaving the patch awaiting the release tag at the thread's close #16 #17 #18.
- Tom subsequently committed #15 as 413c1ef98e, changing only the help-option ordering; the earlier SQL-helper proposal remained uncommitted.