dsh-model-sync

Usage

1. Settings page

1. Start dsh web and open Settings.
2. Open the 模型同步 section in the left nav.
3. Click 刷新 if the list is empty.
4. Each provider card shows:
   • id and baseURL
   • discovered model chips (blue outline = new, not yet in settings)
   • 应用 N 个新模型 to write those ids into llm-pi-ai.providers..models
   • quota / balance text, or a short reason when the provider has no public window API
5. The top card toggles 模型同步 and optional 订阅制模型接入 (sub-model-access) by writing disabled on the matching profile row. That change is hot.

Catalog-only protocols (MiniMax, Kimi Coding, OpenAI Codex) do not speak OpenAI GET /models. The plugin falls back to known catalog ids instead of painting a red “this protocol has no /models listing” error.

2. Composer rings

1. Open any conversation and pick a model as usual.
2. A dual ring appears to the left of the model name in the composer.
3. Hover (or focus) it. The card title is provider / model for this session only.
4. Switching the session model updates the ring. Other providers stay on the settings page.

If the hover card says the current model is not recognized, the session provider is not in the discovered list yet — open Settings and refresh.

3. Quota sources

Provider	What you see	Endpoint
openai-codex	5h / 7d utilization	chatgpt.com/backend-api/wham/usage
kimi-coding	5h / 7d utilization	api.kimi.com/coding/v1/usages
zai	5h / 7d utilization	api.z.ai/api/monitor/usage/quota/limit
minimax-cn	5h / 7d remaining	api.minimaxi.com/v1/api/openplatform/coding_plan/remains
deepseek-official	balance vs last recharge	api.deepseek.com/user/balance
xai	reason only	grok.com gRPC billing is not ported yet
others	reason only	no public plan window

Plan percentages come from the same family of APIs as CC Switch. Remaining % = 100 − used. Zhipu (zai) must send the raw key in Authorization (no Bearer ). Codex also reads tokens.account_id from ~/.codex/auth.json when present.