Preflight a Sume image request against capability descriptors

Sume publishes what each image model accepts as typed capability descriptors, and rejects anything else with 400 unsupported_parameter. A preflight function can check your request body against those descriptors first, so you find a bad aspect_ratio or an unsupported reference count in your own code. The check is local; the descriptors come from one catalog call.

Descriptor types

From the Image API docs.

The preflight

It reports problems as a list. Free-text fields such as prompt are boolean descriptors, so only presence is checked. input_references is checked by length.

def preflight(body, params):
    problems = []
    for key, val in body.items():
        if key in ("model", "metadata", "mode", "webhook_url", "wait_timeout_seconds"):
            continue
        d = params.get(key)
        if d is None:
            problems.append(f"{key}: not supported")
        elif d["type"] == "enum" and val not in d["values"]:
            problems.append(f"{key}: {val!r} not in {d['values']}")
        elif d["type"] == "range":
            size = len(val) if isinstance(val, list) else val
            if not d["min"] <= size <= d["max"]:
                problems.append(f"{key}: {size} outside {d['min']}-{d['max']}")
    return problems

Wire it up

Fetch supported_parameters for the model from GET /v1/images/models, call preflight(body, params), and send the request only if the list is empty. Cache the catalog for a short time, not forever.

Limits

The check cannot know about token-priced quality reservations or content policy, and it skips routing fields. Treat it as a first filter; the API stays the authority.