T8 LLM Provider Config

The three T8 enhancer nodes - MiniMax H3, Seedance 2.0 and Music 3 - each carry their own API mode, model picker, local GGUF settings and temperature strategy. If you run all three in one workflow, that's a lot of duplicated dropdowns to keep in sync. This node is the shared control: set the channel once, wire the single provider_config output into any or all of the enhancers' matching green socket, and they all run on that channel until you disconnect.

It's purely optional. The enhancers work completely standalone; nothing here changes their behavior until you connect the wire.

How it works

provider selects the shared channel - Seedance NZ (the author's gateway, default), Local Qwen, the AI Workshop, or an OpenAI-compatible endpoint. Pick cloud and the relevant model fields appear (ai_workshop_model, custom_model, openai_base_url). Pick Local Qwen and you get the full local stack: local_model and local_mmproj (scanned from ComfyUI/models/LLM), local_context_size, local_max_tokens, think mode, unload policy, and VRAM policy - all of which override whatever the enhancer nodes had set individually.

temperature_policy is where the shared config gets opinionated. The pack normally sends temperature with requests, matching its 1.2.0 behavior; the AUTO policy is what omits it for the one known Kimi Coding Plan endpoint that chokes on the field, and you can force send or omit explicitly. extra_parameters_json accepts only a whitelist of parameters - it cannot override model, messages, auth, media or streaming. credential_alias is the genuinely nice feature: you store a real key once in the ComfyUI user directory under a name, and workflows save only the alias. The real key never enters the graph JSON.

Credential priority when both exist: a key wired directly into the enhancer's api_key socket wins, then the credential alias, then the node's environment-variable fallback.

The trap

The green "共享 LLM 渠道配置" socket is not a STRING. It only accepts this node's provider_config output. The README calls this out because it's the first thing beginners hit - they try to run a plain text or API-key node into it and nothing happens. Add this node from the search list, connect its right-side output to the enhancer's same-named input, and disconnect it to instantly restore each enhancer's own widgets.

Install

It ships in the same pack, so the steps are identical: Manager search "MiniMax H3 / Seedance 2.0 / Music 3 Prompt Enhancer (T8)", or:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-prompt-enhancer-T8.git

Restart and hard-refresh with Ctrl+F5. There's an example workflow (basic_workflow_multi_task_connections.json) that wires one shared config into all three enhancers with only pack-internal nodes.

Where people get burned

• Forgetting it's a special socket type and trying to feed it text.
• Expecting the shared config to store the API key - it routes credentials via alias; the key still lives in the ComfyUI user dir, which is the point.
• Mixing local vision settings for H3/Seedance with Music 3: Music 3 ignores the projector entirely, so an AUTO mmproj match there is harmless but meaningless.