DSH CLIProxyAPI Provider

An auto-discovery CLIProxyAPI provider for DeepSeek Harness (DSH).

It discovers a local CPA service, synchronizes the live model catalog into DSH's native model picker, and lets you choose a vision-capable model from that catalog—without typing model IDs by hand.

简体中文说明

Why this plugin?

Generic OpenAI-compatible routes usually require a manually maintained base URL and model list. This plugin treats CLIProxyAPI as a first-class, dynamic DSH provider:

• Zero-address local setup — probes 127.0.0.1:8317, localhost:8317, and [::1]:8317.
• Live model discovery — reads /v1/models?client_version=pi, with standard data[].id fallback.
• Native model switching — models appear under CLIProxyAPI (auto) in DSH's model picker.
• Newest versions first — versions are sorted newest-first inside each model family.
• Vision model dropdown — enter the CPA URL and key, then select from the returned catalog; model IDs are never entered manually.
• Capability sync — imports context/output limits, reasoning support, image input, and service tiers when CPA reports them.
• Automatic refresh — CPA account or model changes are picked up without reinstalling the plugin.
• Credential-safe — the API key is stored through DSH's credential service, not in settings.yaml.
• Standard DSH bundle — works with DSH Web, headless profiles, and DSH Desktop; it is not a Desktop-only extension.

Requirements

• DeepSeek Harness 0.1.0-rc.7 or a compatible release
• Node.js 22.19+ or 24+
• A running CLIProxyAPI instance

The plugin is currently built and tested against the public DSH 0.1.0-rc.7 contract.

Install

Install directly from GitHub into the profile you use:

dsh plugin --profile web add github:ChadCSong/dsh-cliproxyapi-provider

For DSH Desktop, use its active profile (normally desktop):

dsh plugin --profile desktop add github:ChadCSong/dsh-cliproxyapi-provider

Restart the corresponding DSH process once after installation. No manual profile patching is needed; the standard DSH bundle declaration lives in package.json and cordis.patch.yml.

Install from a local checkout

git clone https://github.com/ChadCSong/dsh-cliproxyapi-provider.git
cd dsh-cliproxyapi-provider
pnpm install
pnpm build
dsh plugin --profile web add "$PWD"

Replace web with the profile you actually use.

Setup and usage

1. Start CLIProxyAPI. Leave the URL empty when it listens on the default port, 8317.
2. Open Settings → Plugins → CLIProxyAPI (auto discovery) in DSH.
3. If CPA bearer authentication is enabled, enter the API key once.
4. After the URL and key are valid, choose the vision model from the automatically loaded dropdown.
5. Return to a conversation and select any model under CLIProxyAPI (auto) in DSH's native picker.

Useful commands:

/cpa-status
/cpa-refresh

/cpa-status reports the current endpoint and synchronized model count. /cpa-refresh immediately probes CPA and refreshes the DSH model catalog.

Configuration

The settings namespace intentionally remains dsh-cliproxyapi for upgrade compatibility:

dsh-cliproxyapi:
  enabled: true
  # Empty means local auto-discovery. Remote CPA is also supported.
  baseURL: ''
  apiKeyEnv: CLIPROXYAPI_API_KEY
  # Empty means the first catalog model declaring image input.
  visionModel: ''
  protocol: openai-completions
  refreshIntervalSeconds: 300
  probeTimeoutMs: 2000

Endpoint resolution

Candidates are deduplicated and tried in this order:

1. Explicit baseURL setting
2. CLIPROXYAPI_BASE_URL from the DSH launch environment
3. Local loopback candidates on port 8317

A non-loopback service is contacted only when you explicitly configure it.

API key storage

apiKeyEnv is a credential reference, not the secret itself. The settings UI writes the key through DSH's credential service. The default reference is CLIPROXYAPI_API_KEY.

Inference protocol

The default, openai-completions, uses CPA's /v1/chat/completions route. Select openai-responses only when your CPA deployment exposes the target models through the Responses API.

Vision model behavior

DSH checks a model's inputModalities before attaching images. The plugin first trusts CPA's input_modalities metadata. Selecting a vision model explicitly adds image capability to that model inside DSH when the catalog omits it. The selected model must still genuinely accept OpenAI-compatible image content.

How it works

The plugin manages one route in DSH's official @deepseek-ai/dsh-llm-pi-ai adapter through the public settings contract. It does not reimplement Harness model transport. Streaming, tool calls, attachments, reasoning, and errors therefore continue through the official adapter.

The host bundle imports no Electron or dsh-plugin-desktop APIs. Its settings card uses public DSH client services and slots, allowing the same package to run in CLI/Web and Desktop profiles.

Development

pnpm install
pnpm test
pnpm typecheck
pnpm build

When developing against a local Harness checkout:

DSH_CHECKOUT=/absolute/path/to/deepseek-harness pnpm test
DSH_CHECKOUT=/absolute/path/to/deepseek-harness pnpm build

Alternatively, point DSH_PACKAGES_ROOT at a DSH node_modules directory. Generated local path files are ignored by Git and never included in published packages.

Security

Automatic discovery only contacts loopback addresses. The plugin never writes a literal API key to its settings, model cache, or logs. See SECURITY.md to report a vulnerability privately.

Contributing

Issues and pull requests are welcome. For behavior changes, please include or update Vitest coverage and verify pnpm test, pnpm typecheck, and pnpm build before submitting.

License

MIT