@dsh-next/dsh-next-oauth-providers
DeepSeek Harness 插件：从 Models 页面登录 Kimi、Grok、ChatGPT 和 Claude 编程订阅。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add @dsh-next/dsh-next-oauth-providers@0.1.0Next Oauth Providers 以 @dsh-next/dsh-next-oauth-providers 发布，当前版本为 0.1.0。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
English | 中文
This DeepSeek Harness plugin adds Subscriptions under Settings → Models. Sign in with a Kimi, Grok, ChatGPT, or Claude coding subscription using the same Add provider / Edit / Delete flow as API-key rows. Connected models appear in the existing model selector.
The editor card keeps the stock layout. Where API-key rows have an API key
field, this card has Sign in / Reconnect. Tokens are stored in
$DSH_HOME/.credentials.yaml under dsh-next-oauth-providers, not in
settings.yaml.
An omitted or empty models list uses the adapter defaults. A non-empty list replaces the catalog. Fetch available models requires sign-in and does not write until you add selected models and Apply. Kimi, ChatGPT, and Claude use the bundled catalogs. Grok first asks its coding endpoint, falling back to the bundled catalog on failure or after a 30-second request timeout.
Grok traffic uses the Grok coding proxy, not the ordinary xAI API endpoint.
dsh plugin --profile <name> add @dsh-next/dsh-next-oauth-providers
<name> is your DSH profile (for example web).
Needs DeepSeek Harness 0.1.3-alpha.2 or newer (official dsh-llm-pi-ai).
Catalog rows persist under dsh-next-oauth-providers.providers as a block
list keyed by official catalog ids (xai, kimi-coding, openai-codex,
anthropic):
dsh-next-oauth-providers:
providers:
- id: xai
displayName: Grok
models:
- id: grok-4.6
name: Grok 4.6
contextWindow: 500000
maxTokens: 500000
The model selector uses the same ids with an -oauth suffix (xai-oauth, …)
so an API-key anthropic row can coexist with a Claude subscription. They
cannot live under llm-pi-ai: — that section only accepts API-key
providers.
ChatGPT OAuth keeps the pi-ai model list but uses official API context /
max-output sizes as defaults (1050K / 128K for current GPT-5.4+ ids; 400K /
128K for gpt-5.4-mini). Customized settings still override a row.
One account per family in this version. Reconnect replaces that grant.
One unusable subscription does not take the others down. If a provider catalog cannot be prepared, only that provider's models fail to load — the model selector names it and offers Retry — while the remaining routes, and chat on them, keep working. The host logs one warning per provider.
Closing an editor, switching provider editors, or leaving the Models page cancels its unfinished sign-in. Cancelling a login queued for credential storage preserves the previously saved grant.
Subscription access is controlled by the provider. Signing in here does not grant a plan you do not already have.
Claude and ChatGPT callback servers bind loopback ports 53692 and 1455.
ChatGPT sign-in fails if something else (often VS Code Codex) already owns
1455.
Contributors: see CONTRIBUTING.md.