model-writtenrun 2672 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Tom Lane proposed converting five contrib extensions to SQL-standard function bodies in #1 to prevent runtime search_path capture, hoping to include the work in PostgreSQL 14; he initially omitted adminpack and lo because their existing functions appeared unconditionally secure.
- Noah Misch identified a remaining installation-time vulnerability: inexact function or operator argument matches could bind the new bodies to hostile objects, as explained in #2 and #4, and Tom acknowledged that runtime safety alone did not solve this problem in #6.
- Mark Dilger proposed extension-schema substitution for cross-extension references in #7, and Tom favored the @extschema:cube@ spelling in #9; Robert Haas discussed separating creation schemas and broader schema lookup in #8 and #11, while Tom also considered restricting references to system and extension-owned objects in #10.
- Vik Fearing and Alvaro Herrera explored alternative SQL syntax in #17 and #18, but no such implementation was attached; Tom concluded that the security work was too large for PostgreSQL 14 in #12 and later marked the contrib proposal returned with feedback, reporting that the independently safe initdb conversions were already complete in #21.
- Ronan Dunklau revived the series after more than three years in #22, using the now-available @extschema:name@ facility for earthdistance, updating extension versions and adding lo; #24 then repaired missing Meson registrations and the omitted xml2 upgrade script.
- Michael Paquier asked about backpatching in #25, Tom argued against a general backpatch in #26, and Ronan identified earthdistance as an exception because the conversion also repaired pg_upgrade failures in #28.
- Michael committed the unchanged pg_freespacemap component of #24 as 3f323eba89, then requested missing xml2 and lo tests and postponed pageinspect to avoid conflicts with another pending fix in #30.
- Ronan supplied the requested tests and remaining conversions in #31, dropping committed pg_freespacemap and temporarily setting pageinspect aside; Peter Eisentraut suggested using named parameters in #32, but Tom later kept that independent cleanup outside this series in #35.
- Michael landed the lo_oid() test unchanged as 3ef038fc4f, the xml2 tests with formatting and missing alternate-output corrections as 93f9b4a93f, and the lo conversion unchanged as 13e3796c90, all from #31, explaining the results in #34.
- Tom prioritized earthdistance after another live failure report and committed its component of #31 as 31daa10fac, backpatching through PostgreSQL 16; he added exact float8 constants, removed unnecessary pg_catalog qualifiers and hardened CREATE DOMAIN in the existing installation script, explaining in #36 why an upgrade script could not repair an already-subverted domain base type.
- Tom resent the unchanged citext and xml2 remainder in #37, then landed citext as 97a5a16849 with redundant type qualification removed and xml2 as 667368fd26 with explicit text casts, announcing both in #38.
- After the competing pageinspect fix had landed, Tom returned to the component last attached in #24 and committed 68ff25eef1, adding an oid cast and preserving PARALLEL RESTRICTED because tuple_data_split() required it, as explained in #39.
- Tom proposed separate hardening for hstore and ltree transform modules in #40, qualifying references to their underlying extension types and removing obsolete same-schema documentation warnings; he again edited installation scripts directly because already-subverted definitions could not be repaired by an update script.
- Noah demonstrated in #41 that the landed pg_freespacemap conversion still allowed persistent capture of generate_series(int, bigint), prompting Michael's explicit-cast repair in #42.
- Noah reviewed the repair in #43 and recommended quoted numeric literals so parsing would produce constants instead of runtime cast calls; Michael incorporated that change in e0c3d5122e and confirmed completion in the thread's final message, #44.
- At #44, all six extension conversions, the added tests and the pg_freespacemap repair had landed, while the transform patch remained unannounced in the thread; it subsequently landed unchanged as bebe904038, leaving the broader named-parameter cleanup and proposed trusted status for earthdistance outside the completed work.