feat: allow custom models

[AdirAmsalem]

Summary

Lets callers use models that aren't in the SDK's built-in registry (preview, experimental, private) by constructing a ModelDefinition directly. Mirrors the JS SDK's CustomModelDefinition.

Usage

Built-in helpers stay the same:

from decart import DecartClient, models

client = DecartClient(api_key=os.getenv("DECART_API_KEY"))

# Process (sync image)
await client.process({"model": models.image("lucy-image-2"), "prompt": "...", "data": b"..."})

# Queue (async video)
await client.queue.submit({"model": models.video("lucy-clip"), "prompt": "...", "data": b"..."})

For non-registry models, construct a ModelDefinition directly:

from decart import DecartClient, ModelDefinition, RealtimeClient, RealtimeConnectOptions

# Realtime
custom_realtime = ModelDefinition(
    name="lucy_2_rt_preview",
    url_path="/v1/stream",
    fps=20,
    width=1280,
    height=720,
)
await RealtimeClient.connect(..., options=RealtimeConnectOptions(model=custom_realtime, ...))

# Process (sync)
custom_image = ModelDefinition(
    name="lucy_image_preview",
    url_path="/v1/generate/lucy_image_preview",
    fps=25,
    width=1280,
    height=704,
)
await client.process({"model": custom_image, "prompt": "...", "data": b"..."})

# Queue (async)
custom_video = ModelDefinition(
    name="lucy_video_preview",
    url_path="/v1/jobs/lucy_video_preview",
    fps=20,
    width=1280,
    height=720,
)
await client.queue.submit({"model": custom_video, "prompt": "...", "data": b"..."})

If input_schema is omitted, inputs are forwarded as-is and the backend validates them.

Validation

• pytest -q — 128 passed (135 with decart[realtime] extras)
• black / ruff / mypy clean
• E2E smoke against the live API: built-in and custom ModelDefinition both succeed via client.process()

---

Note

Medium Risk
Expands supported model definitions beyond the SDK registry and changes request routing to rely on model.url_path, which could surface new runtime errors if callers supply incorrect paths or inputs.

Overview
Adds support for custom/non-registry models by letting callers construct ModelDefinition directly (exported as CustomModelDefinition) and making input_schema optional for passthrough validation.

Updates DecartClient.process() and QueueClient.submit() to accept any ModelDefinition[str], skip registry checks, and (when no schema is provided) forward non-None inputs as-is; queue submission now posts to base_url + model.url_path instead of /v1/jobs/{model.name}. Realtime connect typing is loosened to accept arbitrary model names, and tests are updated/added to cover custom models and backend error surfacing.

^{Reviewed by Cursor Bugbot for commit de8805b. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, have a team admin enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 985db94. Configure here.}