mirror of
https://github.com/stack-auth/stack.git
synced 2026-07-20 21:29:36 +08:00
Review of the structured-output change found that requiring improvementSuggestions reintroduced the failure-row problem: for good answers with nothing to suggest, the model omits the field, which would throw NoObjectGeneratedError and store a false 'needs human review' verdict. Verified live: the model omits it ~5/6 of the time on clean answers. Default it to "" so omission is tolerated, matching the old code's lenient handling.