Flow keeps 3 edit turns and History; build the audit trail on Sume

Google Flow lets you keep editing an Omni clip for up to 3 conversational turns without losing context, and its History panel lists all previous versions of the video. Sume has no conversation: every edit is a separate job with its own id, so the audit trail is whatever you record, and the cheap way to record it is the job id plus the idempotency key.

What exactly does Flow remember?

The Flow help page says you can continue for up to 3 conversational turns without losing the context, and that in the History panel you can find all previous versions of the video. It does not say what happens on turn four, so plan as if context ends there and re-describe the scene when you start a fresh chain.

The Gemini API describes the same idea as conversational editing through the Interactions API, where each instruction builds on the previous one. Sume's docs describe no such session, which is covered in one Sume job per edit.

How do I recreate the History panel on Sume?

Every Sume submit returns a job and a durable result. The Video Router edit takes a video_url source; the finished job's result carries the new video, which becomes the video_url of your next edit. Nothing else links version 2 to version 1, so write that link down yourself.

A simple ledger needs four fields per row: the Idempotency-Key you sent, the job id returned, the source URL, and the prompt. The Jobs and results page covers polling and the job status and result paths.

What does a three-edit chain look like?

Each call below uses its own key so a retry cannot create a duplicate. Paste the result URL from step one into step two.

Because an edit prompt has no memory of the earlier one, a later prompt that says only "make it darker" is ambiguous. Name the subject again: "Keep the same scene. Make the room darker." The Omni prompt-style notes in the Omni edit prompts post cover wording.

for n in 1 2 3; do
  curl -s -X POST https://api.sume.com/v1/video-router/generate \
    -H "Authorization: Bearer $SUME_API_KEY" \
    -H "Content-Type: application/json" \
    -H "Idempotency-Key: chain-demo-edit-$n" \
    -d "{\"model\":\"gemini-omni-flash-1.1\",\"prompt\":\"Edit $n: warm the lighting. Keep everything else the same.\",\"video_url\":\"$SRC_URL\",\"mode\":\"async\"}"
  echo
done

What is the catch?

The loop above sends the same source three times, so it makes three parallel versions, not a chain. To chain, wait for each job, read its result URL, and use it as SRC_URL for the next. That is the real difference from Flow: its turns are a conversation, yours is a loop you own.

Re-encoding also compounds across a chain. Each generation is a new file, so check the clip with video inspect after the last step rather than assuming it matches the first.