Python argparse CLI that checks seconds per Sume model first

You can stop a bad Sume video request before it leaves your machine with an argparse CLI that holds each model's duration range, rejects a model name it does not know, and offers --dry-run to print the request body. The 25-line script below does that. sys.exit with a string prints it to stderr and exits with status 1, which is what a shell script or CI job expects.

The ranges come from the Sume video docs and are copied into the script. That is quick, and it can go stale. The last section shows how to replace the table with a live catalog read.

The documented limits

Other models have their own ranges, and the catalog lists them in supported_durations. Add a row to LIMITS only after you read that field.

The script

Run python3 clip.py wan-3.0 "A mug on a desk" --seconds 12 --dry-run and it prints the JSON body. Run it with minimax-h3 and --seconds 20 and it exits with minimax-h3 accepts 5 to 15 seconds, not 20. Without --dry-run it posts to /v1/videos, so set SUME_API_KEY and CLIP_KEY, the idempotency key for the job.

import argparse, json, os, sys, urllib.request

LIMITS = {"seedance-2.5": (4, 30), "wan-3.0": (2, 30), "minimax-h3": (5, 15),
          "gemini-omni-flash-1.1": (3, 10)}  # seconds, from the Sume video docs

p = argparse.ArgumentParser(prog="clip")
p.add_argument("model", choices=LIMITS)
p.add_argument("prompt")
p.add_argument("--seconds", type=int, default=8)
p.add_argument("--dry-run", action="store_true")
a = p.parse_args()
lo, hi = LIMITS[a.model]
if not lo <= a.seconds <= hi:
    sys.exit(f"{a.model} accepts {lo} to {hi} seconds, not {a.seconds}")
body = {"model": a.model, "prompt": a.prompt, "duration": a.seconds}
if a.dry_run:
    print(json.dumps(body))
    sys.exit(0)
req = urllib.request.Request(
    os.environ.get("SUME_BASE", "https://api.sume.com") + "/v1/videos", method="POST",
    data=json.dumps(body).encode(),
    headers={"Authorization": "Bearer " + os.environ["SUME_API_KEY"],
             "Content-Type": "application/json", "Idempotency-Key": os.environ["CLIP_KEY"]})
with urllib.request.urlopen(req, timeout=35) as r:
    print(r.status, json.load(r))

Choices as a first gate

• choices=LIMITS makes argparse reject a misspelled model with a list of valid ones, before any network call.
• The default of 8 seconds is valid for three of the four models and wrong for none; gemini-omni-flash-1.1 accepts up to 10.
• The dry run prints exactly what would be sent, so you can diff it against an earlier run.

Replacing the table with the catalog

GET /v1/videos/models returns supported_durations for each model. Fetch it once at start, build LIMITS from the list, and keep the hard-coded table only as a fallback for offline runs. Then a launch that changes a range changes the CLI with no edit. The related post on checking a new model shows the fields to read.

Keep the script's output small and stable. A deploy job that parses it will break the day you add a chatty line.

Exit codes and CI

Used in a pipeline, the script has three outcomes. A valid dry run exits 0 and prints the body. A duration outside the range, or a model that is not a choice, exits non-zero before any request. A live run that gets an HTTP error raises, and Python exits with status 1 and a traceback.

That is enough for a CI gate: run the dry run for every clip in a shot list, and fail the build if any line exits non-zero. No job is created and nothing is billed, so the gate can run on every commit.

If you later add polling, keep it out of this file. A small tool that does one thing is easier to test than one that submits, waits, and downloads.