gateway-config

A DSH plugin for configuring any OpenAI-compatible LLM gateway (any endpoint that serves GET /models) as a provider — one provider route per gateway, multiple gateways side by side.

What it provides

• Host tool: gateway_configure
• Web settings panel: LLM 网关配置
• Gateway model discovery via GET /models with non-chat model filtering
• Per-model reasoning-effort mappings
• Per-model text/image input capability mappings
• One managed credential per route (<ROUTE>_API_KEY, e.g. CINDY_API_KEY)
• Native ChatGPT OAuth for the openai-codex provider, including automatic access-token refresh through pi-ai
• Localized settings UI in English and Simplified Chinese, following the DSH interface locale

Providers are fixed presets; no ad-hoc routes are created. Anything that speaks the OpenAI GET /models shape (the de-facto standard DSH discovery reads) works:

• cindy (Cindy), sub2api (Sub2API) — private gateways; addresses never live in this repository. The host tool reads them from CINDY_BASE_URL / SUB2API_BASE_URL; the web panel remembers the last address per route.
• generic (通用) — any other OpenAI-compatible endpoint; supply its base URL (GATEWAY_BASE_URL env or the panel's Base URL field).
• Public vendors with OpenAI-compatible APIs carry their documented base URLs out of the box: deepseek, moonshot (Kimi), zhipu (GLM), qwen, openrouter, siliconflow, xai (Grok).
• openai-codex (OpenAI Codex / ChatGPT) uses the native pi-ai provider and OAuth transport. It does not require or accept a custom Base URL or API key.

Install into an existing DSH profile

This repository is a plugin package. It does not replace or contain a complete DSH home.

dsh plugin --profile web add https://github.com/iceprosurface/dsh-gateway-config.git

If your DSH version expects a Git URL package, use:

dsh plugin --profile web add git+https://github.com/iceprosurface/dsh-gateway-config.git

If the plugin entry is not reconciled automatically, add this to the existing $DSH_HOME/profiles/web/cordis.patch.yml without deleting other entries:

- insert:
    - id: gateway-config
      name: gateway-config

Then restart the existing DSH web process.

Configure a gateway

Open LLM Gateway / LLM 网关配置. Pick a gateway from the provider dropdown (Cindy, Sub2API, Generic, or a public vendor), fill in the Base URL if not pre-filled (the gateway's OpenAI-compatible root, e.g. ending in /v1), then click 拉取模型: the chat-capable models are listed with checkboxes; tick the ones to keep and mark one radio button as the default model, then click 写入配置. The key is stored in DSH local managed credentials and never written to this repository. Later, leave the key blank; the plugin reuses <ROUTE>_API_KEY.

For OpenAI Codex, select openai-codex and click Sign in to ChatGPT. The authorization page opens in a separate browser window. After the callback completes, DSH stores the grant as llm-pi-ai/openai-codex; tokens are never returned to the settings UI, and pi-ai refreshes them when needed. Then discover and import models as usual.

The host-side equivalent:

gateway_configure(apiKey?, baseURL?, route, displayName, defaultModel?)

apiKey and baseURL are omitted for route: "openai-codex" after OAuth sign-in has completed in the settings page.

Development

• lib/index.js: host-side tool
• lib/client.js: web-side settings panel

node --check lib/index.js
node --check lib/oauth-remote.js
node --check lib/client.js

Security

Never commit API keys, gateway addresses you consider private, credential files, personal DSH settings, sessions, attachments, generated dependencies, or runtime databases.

License

MIT