model-writtenrun 133 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
May 2024 · 2 entries
- Daniel Gustafsson proposed server-side SNI in May 2024 to select certificates and disjoint trust stores by hostname, using pg_hosts.conf, per-host SSL contexts and an off/default/strict ssl_snimode setting #1.
- Cary Huang requested per-host encrypted-key passphrase commands #2, while Jacob Champion found unstable reload behavior and questioned whether SSL_set_SSL_CTX() transferred enough connection state #3.
Dec 2024 · 1 entry
- Daniel added passphrase configuration and context cleanup, then refined reload tests across #5 and #7; he subsequently documented that per-certificate HBA clientname settings were outside the initial scope #10.
Feb - Mar 2025 · 2 entries
- Jacob supplied a failing fallback-certificate reproducer #11, which Daniel incorporated while introducing an explicit default context derived from postgresql.conf #12.
- Jacob's next test patch demonstrated that the default client CA could influence other hosts' certificate verification, strengthening his concern about incomplete SSL context switching #13.
Aug - Sep 2025 · 1 entry
- After a development gap, Daniel adjusted Windows passphrase tests and exposed a pre-existing EXEC_BACKEND limitation #15; Michael Paquier recommended separating that work #16, and the subsequent rebase removed the independently pursued passphrase tests while improving memory handling #18.
Nov 2025 - Jan 2026 · 6 entries
- Jacob confirmed that the CA-isolation reproducer still failed #20, prompting Daniel to incorporate it and adjust verification-callback setup and error expectations #23.
- Chao Li's reviews improved const correctness, parser status reporting and documentation #19 #24, while Dewei Dai's review led to uninstall support and reuse of context cleanup #26 #27.
- Heikki Linnakangas proposed expressing default and non-SNI behavior entirely in pg_hosts.conf #28, and Jelte Fennema-Nio proposed file-based precedence #31; Daniel implemented those ideas by removing the mode GUC and adding wildcard and no_sni entries #36.
- Jacob objected to silently enabling SNI through file presence #33, so Daniel proposed a separate boolean enable switch #36, which Heikki endorsed #38.
- Jacob's subsequent review requested hostname lists, case normalization and an unambiguous non-SNI token #39 #41; Daniel added these and moved selection to ClientHello, while deferring broader wildcard and hostname-file features to control scope #43.
- When the client-CA failure persisted, Jacob isolated verification state that SSL_set_SSL_CTX() failed to transfer and proposed explicitly updating connection settings #44.
Mar 2026 · 5 entries
- Daniel and Jacob collaborated on a substantial rewrite using one main SSL context and explicit copying of selected host settings into the connection, disabled incompatible TLS init hooks under SNI, gated unsupported LibreSSL APIs, and separated useful non-SNI client-CA tests #45.
- The final review rounds added duplicate-host rejection, LibreSSL test skips, parser validation and correct TLS alerts, followed by CA replacement, CRL and hook tests and defensive cleanup #46 #47 #48 #50 #52.
- On March 18, the two constituents of #52 landed unchanged as the client-CA tests 25e568ba7c and server-side SNI support 4f433025f6, with Daniel announcing the landing and initial buildfarm failures #53.
- Jacob diagnosed the missing EXEC_BACKEND test skip #54, and Daniel submitted that correction together with a tentative client-side handshake-error workaround #55.
- The EXEC_BACKEND test correction landed with minor wording changes as e87ab5049d, while the client error workaround was abandoned after Jacob questioned it and Tom Lane's OpenSSL upgrade cleared the failure #59 #62 #64.
May 2026 · 1 entry
- During pgindent preparation, Tom Lane found that the unused HostsFileLoadResult typedef escaped collection and proposed moving it and using it explicitly #69; Daniel agreed #71, and the patch landed unchanged as 93da297366, with function renaming deferred to avoid unnecessary churn #72.
Sep - Oct 2026 · 10 entries
- In September, Noah Misch forwarded Anthropic OSS program's report and patch for a crash when a failed reload left the live ssl_sni GUC inconsistent with the retained SSL configuration #74.
- Daniel initially extended the fix to restore the live GUC #76, but Zsolt Parragi identified another null-dereference risk and inconsistent treatment of existing sessions #77, leading Daniel back to retaining SNI state inside the installed configuration #78.
- Zsolt then found that consulting the previous configuration during initialization broke enabling SNI with encrypted keys #79, and Jacob independently challenged those checks while identifying silent test-setup failures and retained TLS resources #80.
- Daniel posted a five-patch series correcting the reload logic, hardening test setup, clearing TLS host resources and incorporating Zsolt's off-list fixes for dangling GUC strings and leaked SSL contexts #81.
- Rui Zhao reviewed and tested all five patches and supplied certificate-selection tests for enabling SNI, failed disabling and successful retry #82; after Daniel flagged unsupported subtests, Rui rewrote the add-on for Test::More 0.98 #83 #84.
- The saved-SNI-state fix from #81 and Rui's final tests from #84 were combined with light polish into b6adefc659.
- The test-setup and TLS-destruction constituents of #81 landed unchanged as e9aea0189d and d45dc682e8.
- Zsolt's GUC-copy and delayed-SSL_CTX-allocation fixes, posted by Daniel in #81, landed as c7771f1ae4 and bdf7f9b561, with only base-signature adaptation needed for the latter; Daniel confirmed the SNI fixes were also backpatched to PostgreSQL 19 #86.
- Zsolt subsequently reported that one unloadable host could disable SSL for every connection in Windows/EXEC_BACKEND children and submitted three patches: isolate failed hosts, reject unusable SNI keys at startup, and apply analogous startup rejection to global SSL configuration #87.
- At the last supplied message, Daniel had acknowledged that new series and promised to investigate #88; the later candidate 9674c91aeb establishes that its first patch landed with rebase and formatting adjustments, while the two startup-rejection patches have no demonstrated landing in the supplied material.