Gemini Omni multi-turn edits: previous_interaction_id vs Sume jobs

On Google's API, multi-turn editing is done by passing previous_interaction_id so the model continues from the last turn (read 2026-10-03). Sume's Video Router has no interaction id. Each edit on gemini-omni-flash-1.1 is its own job that starts from a video_url, so a second edit uses the first edit's output file as the new source.

Side by side

Chaining by hand

Read the first job's result, take the artifact URL, and submit it as the next edit's video_url. Keep each edit prompt narrow, as in the docs example: say what changes and say keep everything else the same.

Every pass re-encodes the video, so expect quality to slip over many turns. Make the final edits from the original with a combined prompt when you can. Output length follows the source clip, and Google's input cap of 10 seconds still applies.