Marcel Hofer
b87d17e34d
Add native OpenAI Decisions and OpenRouter Decisions API routing to the image and text Decision Router nodes while keeping local in-process llama.cpp inference as the default. Discover OpenRouter decision models by input modality and decision output capability, then enforce server profile image policy and optional model allowlists both in the UI model catalog and before inference. Keep credentials and endpoints in server-side profiles; apply profile input, image, and timeout limits; never fall back to chat-generated JSON. Preserve saved workflow widget order, release any retained local model when switching to API mode, and normalize typed route choices, probabilities, and confidence into decision_json. Add API setup documentation and adapter/router/security regression coverage. Validation: 240 tests passed across six test modules run individually in bundled ComfyUI Python 3.12; Python and JavaScript syntax checks, API catalog validation, version metadata checks, and git diff --check passed. Provider inference is covered with mocked responses; no authenticated live inference was run.
3.8 KiB
Decision Router API mode
The Image and Text Decision Router nodes support both local in-process inference and native API inference. Local remains the default. API mode sends each item in a batch as a separate decision request and routes the original input item according to the provider's typed choice response.
Supported native providers
| Profile | Protocol | Endpoint | Initial model selection | Image input |
|---|---|---|---|---|
| OpenAI | openai_decisions |
https://api.openai.com/v1/decisions |
gpt-6-luna |
Supported |
| OpenRouter | openrouter_decisions |
https://openrouter.ai/api/alpha/decisions |
Decision models returned by its Models API | Only models whose input modalities include image |
Other providers in the OpenAI-compatible chat node are not exposed as native Decision APIs by this feature. The router does not fall back to a chat completion that generates JSON.
Credentials and profile setup
The built-in OpenAI and OpenRouter profiles use OPENAI_API_KEY and OPENROUTER_API_KEY. Set the key in the environment of the process that launches ComfyUI, then restart ComfyUI. Restart after adding or changing a custom profile as well, so the node's profile list is refreshed. The key and endpoint do not enter the workflow.
Custom profiles can opt into a native protocol using the server-side profile file:
{
"profiles": {
"OpenRouter Decision Production": {
"provider": "OpenRouter",
"base_url": "https://openrouter.ai/api/v1",
"decision_protocol": "openrouter_decisions",
"decision_endpoint": "https://openrouter.ai/api/alpha/decisions",
"auth": "bearer_env",
"api_key_env": "OPENROUTER_PRODUCTION_API_KEY",
"allowed_models": ["cloudflare/clef-flash", "typesafe/jev-1.14"],
"allow_images": true,
"max_image_pixels": 16777216,
"max_image_bytes": 20000000,
"max_timeout_seconds": 300
}
}
}
decision_protocol and decision_endpoint must appear together. The endpoint is trusted server configuration, validated at profile load, and is not a node input. For shared ComfyUI servers, use an exact allowed_models list.
Model and image capability handling
The node's API model list is filtered by decision output, required input modality, the profile's allow_images policy, and its optional allowed_models list. OpenRouter metadata is cached for up to one hour. If a stale cached list exists during a metadata outage, it can still be displayed; the selected model is checked against that list before inference. A model absent from the capability catalog is rejected rather than guessed to support images.
Image tensors are converted to bounded PNG data URLs using the selected server profile's allow_images, max_image_pixels, and max_image_bytes settings. OpenAI receives an input_image part; OpenRouter receives an image_url part in state. The profile's max_input_chars also bounds the system prompt, input text, and configured route names/descriptions. Provider-specific image limits can still apply and are reported as inference errors.
Output and limits
The APIs return typed route selections and probabilities, which are normalized into the router's existing decision_json output. Provider confidence is included when returned. Route descriptions become the choice criteria, and the system prompt becomes the decision instruction. The existing 52-route cap and dynamic output sockets continue to apply.
Native Decisions APIs do not generate a rationale. The local-only include_reason control is hidden in API mode; if it remains enabled in a saved workflow, the result explicitly marks the reason as unavailable. keep_model_loaded has no effect on remote inference, and switching to API mode releases a local model previously retained by the node.