dsh-providers-live-models

English | 中文

Live model catalogs for DeepSeek Harness (DSH): OpenCode Go, OpenCode Zen, native OpenAI Codex and OpenDesign. The plugin refreshes each route at runtime, adds per-model switches, sign-in and subscription usage, and keeps OpenCode's gateway headers and prompt-cache retention. Nothing is bundled, so upstream changes need no release here.

It also adds compaction recovery: long histories are summarized in capacity-aware batches, and failures or cancellation preserve the original history.

Routes

Route	Catalog	Sign-in / credential	Usage
opencode-go	Pi + live Go roster	OPENCODE_GO_API_KEY	Go quota
opencode (Zen)	Pi + live Zen roster	API key or Console device login	Zen balance
openai-codex	Pi public catalog	native ChatGPT OAuth	OpenAI subscription
opendesign (prerelease)	account catalog	API key or device authorization	quota and wallet in account mode

Install

Desktop: Plugins → Add plugin, paste github:Luca4Don3/dsh-providers-live-models, choose Enable now, restart DSH.

CLI:

dsh plugin --profile desktop add github:Luca4Don3/dsh-providers-live-models

Requires DSH 0.1.5-rc.1 or newer. OpenDesign also needs DSH 0.2.1-rc.2 with pi-ai 1.1.0; a mismatch disables only OpenDesign. The host needs GitHub access for the github: spec.

Use

1. Configure the routes you use in DSH's llm-pi-ai providers. Codex uses native OAuth; Go and Zen read the key references below.
2. Open Plugins → dsh-providers-live-models. Sign in to ChatGPT, OpenCode Console or OpenDesign, switch models under Model display (模型展示), then Save selection.
3. Open Settings → Account & balance for usage. Refresh usage forces a query; cards are cached across restarts and freshMs defaults to five minutes.

All sign-ins write only the plugin's own credential records; model API keys and Codex OAuth are not overwritten.

Configuration

Set these on the providers-live-models entry:

config:
  freshMs: 300000
  timeoutMs: 15000
  goApiKeyRef: OPENCODE_GO_API_KEY
  zenApiKeyRef: OPENCODE_API_KEY
  openDesign:
    protocol: chat-completions # or responses
    stripTemperature: false

stripTemperature is only for OpenDesign gateways that reject temperature. quotaSource accepts auto, vela or control.

Limits

• A model without a complete descriptor, or with an unsupported protocol, is reported and skipped, not guessed.
• Catalogs refresh every five minutes. Endpoints, interval and cache age are fixed in the source; gateway base URLs are deliberately fixed, so a remote catalog cannot redirect requests.
• OpenCode session headers and DeepSeek long prompt-cache retention are the plugin's two request rewrites; explicit request or model headers always win.
• A 403 from the Zen billing endpoint means the endpoint denied access, not that the key lacks permission; Console login is preferred.
• The plugin uses llm-pi-ai internals. A DSH update that changes them logs a warning and keeps the installed catalog.
• With DSH's bundled pi-ai 0.87.1, the OpenCode Go route has no Google transport, so gemini-4-argon-contributor is unavailable there.

Development

npm test

Offline fixtures; no network or credentials needed. CI runs Node 18 and 24.

More: OpenDesign · provider contract · protocol probing · catalog injection

License

MIT