model-writtenrun 4749 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Vik Fearing proposed SQL:2023 ANY_VALUE to return an unspecified non-null input, deliberately leaving aggregate early termination for future work #1, and then added the missing SQL feature declaration #7.
- David G. Johnston requested first_value and last_value aggregates and suggested randomized selection #2; Tom Lane questioned ordering guarantees #3, while Robert Haas and Corey Huinker described uses for additional aggregates that remained outside this patch #4 #5.
- Vik rejected random selection and a documented first-value guarantee because either could constrain future optimizations #8 #12 #14; after further disagreement, David ended the debate #19, and Vik deferred order-insensitive aggregate optimization to separate work #20.
- Peter Eisentraut found a documentation/catalog type mismatch and redundant null checks in the strict transition function #21; Tom Lane questioned whether nulls should be excluded #22, and Vik confirmed the standard required that behavior #23.
- Vik incorporated Peter's corrections, rebased the patch, and added an array-input test #24; Peter then said the revised patch looked good #25.
- David Rowley proposed naming, comment, documentation, and test-style improvements and raised questions about window behavior and optimization #26; Vik applied the editorial changes but declined window-specific tests and optimization support as unnecessary for this implementation #27.
- Maciek Sakrejda suggested describing the result as arbitrary rather than non-deterministic #28; Peter adopted that wording and committed #27 as 2ddab010c2, with catalog housekeeping, formatting changes, and broader regression inputs.
- Peter's final message confirmed the feature had landed and explained the last-minute wording, catalog-description, whitespace, and test changes, leaving no submitted patch outstanding #29.