mirror of
https://github.com/simstudioai/sim.git
synced 2026-09-19 09:30:49 +08:00
WaleedandClaude Opus 4.8
fd77bb4069
copilot_messages had no column preserving message order: created_at (set from each message's timestamp) ties at millisecond granularity in 58% of chats, and some chats have out-of-order timestamps within their array. The only other tiebreaker, id, is a random UUID — so ORDER BY created_at, id renders same-timestamp user/assistant pairs swapped. This blocks the R+1 read cutover. Add an integer seq = the message's 0-based index within the chat's JSONB array (ground-truth order), backfilled inline in migration 0219 (no script for self-hosters or us). Reads will use ORDER BY seq NULLS LAST, created_at, id at cutover; reads still come from JSONB after this PR. Design: - seq is a tiebreaker, not the sole sort key (concurrent-append/NULL safety). - Nullable now; defer NOT NULL so rolling-deploy old pods don't fail inserts. - replace (update-messages snapshot) overwrites seq = array index (re-densifies after a mid-conversation delete); append preserves existing seq via COALESCE and assigns base+idx from a single MAX(seq) read (never MAX+i in SQL — multi-row batches would collide). The non-atomic read-then-insert window is documented and bounded by the read tiebreak + snapshot re-densify. - Dedupe message ids before insert (87 prod chats carry dup ids; a repeated id in one INSERT...ON CONFLICT would otherwise throw). - Backfill picks first-occurrence per (chat,id), gap-free via ROW_NUMBER; validated on staging data (0-based, contiguous, 0 bad ranges). Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>