Un script qui envoie ton workflow à ComfyUI

Write a script that drives ComfyUI from this folder, based on my workflow.

Workflow exported in API format: [Le fichier du workflow exporté au format API]
ComfyUI address: [L'adresse de ComfyUI sur ta machine]
Image folder: [Où enregistrer les images]
Language: [Le langage du script, ou « choisis pour moi »]

1. First of all
- Read ComfyUI's official documentation on the workflow API format (https://docs.comfy.org/development/api-development/workflow-api-format) and on the server routes (https://docs.comfy.org/development/comfyui-server/comms_routes), as well as the official example `script_examples/websockets_api_example.py` in the ComfyUI repository. Do not rely on your memory.
- Check that ComfyUI responds at the given address (`GET /system_stats`). If it doesn't, stop and tell me what to check.
- Open the workflow and list its nodes (number, `class_type`, `_meta.title`). If the file is not in API format, tell me to export it with "File → Export Workflow (API)".

2. Find nodes by their title
Do not rely on node numbers, which change from one workflow to another. Find the nodes to modify by their `_meta.title` (for example "PROMPT", "NEGATIF", "SEED", "TAILLE"). If they don't have a clear title yet, tell me which ones to rename in ComfyUI before exporting again. If a title is missing at run time, the script stops with a clear message.

3. The script
- It reads the workflow and replaces the description, the seed (random if I don't give one) and the size if I give one, without modifying the original file.
- It sends `{"prompt": <workflow>, "client_id": <identifier>}` as a POST to `/prompt` and reads the `prompt_id` from the response. If the response contains `error` or `node_errors`, it prints the error in a readable way and stops.
- It waits for completion by polling `/history/<prompt_id>` (once per second, with an adjustable maximum delay), or through the `/ws` WebSocket if that is simpler in this language.
- It downloads each image listed in `outputs` with `/view` (parameters `filename`, `subfolder`, `type`) and saves it in the image folder, with a name that contains the date and the seed.
- Next to each image, a small JSON file with the description, the seed, the size and the workflow name, so it can be regenerated.
- Command-line options: the description (required), the seed, the size, the number of images. Help (`--help`) and messages in French.
- No dependency if the standard library is enough; otherwise a single one, which you explain to me.

4. Try it
Run a real generation with a simple description, show me the path of the image and of its JSON file, then check that a wrong address or a missing node title gives a clear message.

5. Optional: a skill
Offer to turn it into an `/image` skill (`.claude/skills/image/SKILL.md`) that runs this script with the description I type (`$ARGUMENTS`), checking the skill format on https://code.claude.com/docs/en/skills. Only create it with my approval.

Talk to me in French.

Done when a real image has been generated and saved by the script, with its JSON file, and you have given me the command to generate another one.