Bash script to submit, poll, and download a Sume video

A shell script can run the whole Sume video flow if it does four things: send an Idempotency-Key, poll the job's polling URL, stop on a terminal status, and return an exit code that tells a scheduler what happened. The script below does that with curl and jq in 22 lines. It exits 0 when the file is saved, 1 for a failed job, 2 for a canceled one, and 3 when its own deadline runs out.

The deadline is the part people forget. A client-side timeout does not cancel a Sume job. The job keeps running and keeps billing, so a script that gives up silently leaves paid work behind. This one prints the polling URL when it gives up, so you can read the job later.

The script

Run it as SUME_API_KEY=... bash video.sh order-8823-mug-v1. The argument becomes the idempotency key, so running the same command twice does not create a second job. The loop runs 40 times at 30 seconds, which is a 20-minute deadline.

#!/usr/bin/env bash
set -euo pipefail
API=https://api.sume.com
AUTH="Authorization: Bearer ${SUME_API_KEY:?set SUME_API_KEY}"
KEY="${1:?usage: video.sh <idempotency-key>}"
JOB=$(curl -fsS "$API/v1/videos" -H "$AUTH" \
  -H "Content-Type: application/json" -H "Idempotency-Key: $KEY" \
  -d '{"model":"seedance-2.5","prompt":"Slow push-in on a ceramic mug","duration":5}')
URL=$(jq -r .polling_url <<<"$JOB")
for _ in $(seq 1 40); do
  sleep 30
  S=$(curl -fsS "$URL" -H "$AUTH")
  case "$(jq -r .status <<<"$S")" in
    completed)
      curl -fsS "$(jq -r '.unsigned_urls[0]' <<<"$S")" -H "$AUTH" -o out.mp4
      exit 0 ;;
    failed) jq -c .error <<<"$S" >&2; exit 1 ;;
    cancelled) echo "job cancelled" >&2; exit 2 ;;
  esac
done
echo "deadline: job still running at $URL" >&2
exit 3

Exit codes

The /v1/videos route spells the canceled status with two l's, cancelled. The generic /v1/jobs routes use canceled. If you switch routes, change the case label.

Why curl -f and set -e

curl -f turns an HTTP error status into a non-zero exit, and set -euo pipefail stops the script on it. A 429 on submit then stops the script instead of trying to parse an error body as a job. If you want to retry, wrap the submit in a function that reads the retry-after header, and keep the same key.

The ${SUME_API_KEY:?...} form stops the script with a message if the key is missing, which is better than sending an empty bearer token and reading a 401.

Where this stops being enough

This script holds one job in one process. For a batch, the plan decides how many jobs are accepted at a time: Free accepts 6, Pro 24, Startup 48, and Scale 120, counting processing and queued together. Past that, a submit returns 429 queue_full. Use webhooks or a small worker with a job table once you move past a handful of jobs, and keep the polling URL stored next to each job id.

Running it from a scheduler

Cron, a CI job, and a Makefile target all read exit codes the same way. Map 0 to success, treat 1 and 2 as final, and treat 3 as a signal to look again rather than to resubmit. Because the idempotency key is the script argument, rerunning the same command after a 3 does not create a second job; it returns the original job and keeps polling.

Keep the key meaningful. A key built from an order number and a version, such as the one in the example, makes a retry safe and makes a deliberate change obvious, since a new version string is a new job. A key reused with a different body returns 409 idempotency_conflict, which is the right signal that the request changed.

Log the polling URL and the final JSON to a file. When a job fails, the request id in the error object is the fastest thing to quote to support.