The GitHub Actions job "CI" on rocketmq-dashboard.git/rz/ai-replay-overflow-live-frames has succeeded. Run started by GitHub user zjncs (triggered by zjncs).
Head commit for run: 0dedb9b1a2a37d10891816e5a6133a051fd92c35 / Zhao Jianing <[email protected]> fix(ai): keep live deltas when the replay buffer overflows deliver() promises that a frame with no persisted counterpart (seq 0, e.g. a text_delta or thinking fragment) is never dropped - it cannot be replayed from the database either, and losing a delta is worse than showing it twice - but the REPLAYING overflow branch dropped every further frame once 1024 were buffered, seq 0 included. The bound's justification only holds for frames the client can re-read from the persisted timeline when the run finishes, so the overflow branch now drops only frames with seq > 0; a seq 0 frame is always buffered, evicting the oldest replayable frame for room, or growing past the bound when every buffered frame is already unreplayable. Drain ordering and the dedup watermark are untouched, and the once-only overflow warning still fires whenever a replayable frame is dropped. Signed-off-by: zjncs <[email protected]> Report URL: https://github.com/apache/rocketmq-dashboard/actions/runs/39412845134 With regards, GitHub Actions via GitBox
