public/skills/arka-provider/SKILL.md

name, description

name	description
arka-provider	Manage the custom Arka-AI (arka-ai.ru) LLM provider in opencode config. Use when the user needs to add, refresh, or fix the arka-ai.ru provider, sync its model catalog, change the Arka-AI default model, or when arka models are missing from /models.

Arka-AI provider

arka-ai.ru is not in models.dev, so opencode ships no built-in config for it. It is wired up locally with a plugin + a static provider block in ~/.config/opencode/opencode.json.

How it works

• Plugin: ~/.config/opencode/plugins/arka-ai.ts
  • On startup it loads the published chat catalog https://git.arka-ai.ru/arka/public/raw/branch/main/catalog.json (no API key; refreshed daily at 05:00 Europe/Moscow).
  • If that fetch fails, it falls back to https://api.arka-ai.ru/v1/models and filters out non-chat models (TTS/ASR/image/video/OCR/embedding/reranker/…) via a heuristic regex so the /models picker stays clean.
  • Caches the result at ~/.cache/opencode/arka-ai-models.json (TTL 24h).
  • Injects the provider into config via the config hook, layered over the static block in opencode.json.
• Static block (opencode.json → provider["Arka-AI"]) holds name, api, env, and a small fallback models map that is used when the catalog is unreachable.
• Inference still uses ARKA_API_KEY against https://api.arka-ai.ru/v1.

Refresh the model list

Catalog ~216 raw models / ~116 chat; the public snapshot updates at 05:00 Europe/Moscow. Changes land in opencode only after cache expiry + restart (config is not hot-reloaded).

1. rm -f ~/.cache/opencode/arka-ai-models.json
2. Restart opencode (or run opencode again) — the plugin re-fetches on the next cold start.

Optionally inspect the live API (no key required for listing):

curl -sS https://api.arka-ai.ru/v1/models | python3 -m json.tool
curl -sS https://git.arka-ai.ru/arka/public/raw/branch/main/catalog.json | python3 -m json.tool

Change default model

Edit opencode.json → "model": "Arka-AI/<model-id>". Use the IDs exactly as returned by the catalog (e.g. Arka-AI/DeepSeek-V4-Pro, Arka-AI/qwen3.8-max, Arka-AI/kimi-k3).

Adjust filtering

Edit NON_CHAT (regex) or FORCED_MODELS in the plugin file and the same lists in scripts/sync-catalog.py in arka/public, then re-run the installer and restart. FORCED_MODELS keeps must-have chat models even if the heuristic would drop them.