Retry one scene of a Format run without paying for the whole run
Send previous_run_id to continue a Sume Format run as another turn: redo one scene, keep the rest. The refusals, 404, 400 and 409, and what each means.
To redo one part of a Format run, create a new run on the same Format with previous_run_id set to the earlier run. The next turn continues the same conversation: the agent is replayed what it produced, so it can redo one scene and leave the rest alone. The docs say this is how live-commerce integrations retry a single scene without paying for the whole show again.
The request
Name the old run, say what to redo in instruction, pass any data in input, bind the same output_schema as before (it is per run, not inherited) and give the new run its own cap.
{
"previous_run_id": "RUN_ID",
"instruction": "Retry the selected scene only. Keep every other scene unchanged.",
"input": {"scene_id": "sc_7"},
"generation_spend_cap_usd": 8
}Refusals
A continuation is a new run with a new id, receipt, cap and webhook. The original never changes.
| Response | Means |
|---|---|
404 previous_run_not_found | Unknown id, or another owner's run |
400 previous_run_format_mismatch | The run was created on a different Format |
409 previous_run_not_terminal | Still running; poll it, then call again |
400 previous_run_not_resumable | No thread_id, or neither completed nor any artifacts |
| Continuable if | thread_id is not null and it completed or has non-empty artifacts[] |
Practical notes
A failed run that left work behind can be continued; one that left nothing cannot, so start fresh. Never send a thread id: it is 400 unknown_parameter; name the run. On a continued run artifacts[] lists everything the whole conversation made, while usage stays per run, so budget each turn on its own receipt.
Keeping scene state
Your own record should hold the first run id, each continuation id, and which scene each one touched. The platform keeps the artifacts, but it does not know your scene numbering; the scene_id in input is yours. A per-scene retry count in your database prevents a loop of costly retries on a scene the brief simply cannot satisfy.
Limits
The retry spends its own cap and is not free; a narrow instruction keeps it cheap but the agent can still change more than you asked, so compare artifacts afterward. Send a fresh Idempotency-Key per continuation, derived from the scene and a version. A canceled run delivers no webhook, so for webhook-only integrations read the cancel receipt.
Related posts
More in Formats
- Send a video to a Format run: URLs in input, not attachments
Format and Agent Completion attachments take images only. For video, put the media.sume.com URL in input; Format runs count it toward 10 videos per run.
- Service-account key gets 403 on a Format run: use a user or team key
Sume service-account keys cannot create Format runs and fail with 403 insufficient_scope and a service_account_format_runs_unsupported reason. Here is the fix.
- Share a Sume Format with a partner using its docs call sheet URL
Every Format with a handle and slug has a call sheet at docs.sume.com/formats/{handle}/{slug}: curl, scopes and polling, without the Format body.
- Before/after ad from one packshot: how the Sume Format builds it
sume-before-after builds a matched first frame with ChatGPT Image 2, then animates a reveal with Seedance 2.5. What to brief, what to check, and the limits.
Written by Sume