Set Bailian Coding Plan LLM Service Connector 🐑

SetBailianCodingPlanLLMServiceConnector is the connector for an 阿里云百炼 Coding Plan subscription - the sk-cp- key, the 编程订阅 tier that only exposes Qwen-Coder models. Same shape as its siblings: a key, a model, one llm_service_connector output you wire into any generating node in the pack. Print nothing, generate nothing, just carry the credential.

The thing to internalise is that Bailian's three tiers are three separate services, not three plans on one API. The Coding Plan has its own host (coding.dashscope.aliyuncs.com), its own key prefix, and a model list that stops at the Coder family. A PAYG sk- key here, or this key on the Token Plan endpoint, fails - and it fails as a generic auth or 400 error that reads like your key is broken rather than in the wrong slot.

The inputs

• api_token - the sk-cp-... key. Blank is fine and arguably better: the connector then reads the bailian_coding entry from mie_llm_keys.json in the pack folder. That file is gitignored on purpose, which matters because a key typed into the node travels inside your workflow JSON every time you share one.
• model_select - default qwen3-coder-plus, with qwen3-coder-flash, the older qwen-coder, and Custom. That's the whole list, and it's the whole list because the endpoint only serves that family.
• custom_model - used when model_select is Custom. This is where dated ids go.
• config_file, config_key (default bailian_coding), prefer_local_config - the usual plumbing, same semantics as every other connector in the pack: with the default true, a non-empty entry in the JSON file wins over whatever you typed.

Output: llm_service_connector. Wire it to CallLLMService, Translator, any of the prompt-generator nodes, or CheckLLMServiceConnectivity - the last one is worth the twenty seconds, since it reports HTTP status and the first 200 characters of the reply, which is how you discover "actually this tier hates my model id" without queuing a whole workflow.

Should this be your connector?

Here's the blunt version: not for the H3 pipeline. The storyboard and loop-plan nodes in this pack are doing narrative work - reading a cast list and a dialogue block, keeping spatial continuity, writing structured prose back. Qwen-Coder is a very good code model, and "very good code model" is not the same job. Alibaba's own docs are explicit that non-coder model ids (qwen3-max, qwen3-vl-plus, and so on) come back as an HTTP 400 here, so the tier can't quietly cover those tasks either.

Where it earns its place: text plumbing. Structured JSON reshaping, schema fiddling, regex for your own batch tooling, quick script help through CallLLMService while you build the workflow, and translation of long technical strings where you'd rather pay a coding-plan rate than a frontier one. If your graph is mostly code-adjacent text transformation rather than creative direction, this is the cheapest competent option in the pack's connector list. If it's image- or dialogue-shaped, buy the Token Plan connector instead - that one carries the multimodal lineup, VL models included.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/MieMieeeee/ComfyUI-MieNodes
cd ComfyUI-MieNodes && pip install -r requirements.txt

Manager search is ComfyUI_MieNodes if you'd rather not use a terminal. No weights, no GPU, no extra dependencies beyond the pack's regular requirements.txt. Preferred setup is one line in mie_llm_keys.json:

{
  "bailian_coding": "sk-cp-your-key-here"
}

Where people get burned

• HTTP 401 with a valid key. Wrong tier. Check whether the key starts sk-, sk-sp- or sk-cp- and match it to the node - this is by far the most common complaint about these connectors, and it's a config error every time.
• HTTP 400 on a model you "know" works. You're on the Coding Plan and pointed at a non-coder model. The endpoint rejects it server-side; no amount of custom_model guessing fixes it.
• Date-suffixed model ids. Family names in the dropdown are the stable ones. If the console shows you something longer, Custom plus the full id is the way.
• Expecting vision. There's no VL model on this tier, so anything that sends an image content part - captioning, image-driven prompt rewriting - isn't a fit. It doesn't fail gracefully; it fails server-side.