dsh-web-search-tokenrhythm

A web-search provider for DeepSeek Harness that turns the built-in web_search tool into real searches through the Token Rhythm (基元律动) gateway's OpenAI Responses API (web_search tool). No extra account, no new API key — it reuses the key you already store for chat.

Quick start (3 steps)

Run these inside your dsh profile directory (e.g. %USERPROFILE%\.dsh\profiles\web):

# 1. Install the package
pnpm add dsh-web-search-tokenrhythm

// 2. Register the bundle in profiles/web/package.json
"dsh": {
  "profile": {
    "bundles": [
      "@deepseek-ai/dsh-base",
      "@deepseek-ai/dsh-web-app",
      "dsh-web-search-tokenrhythm"   // ← add this line
    ]
  }
}

# 3. Restart dsh web

That's it. The bundle mounts the provider and switches web_search to it automatically — no cordis.patch.yml editing required.

Prerequisites

• Your gateway must support the OpenAI Responses API with the native web_search tool. Token Rhythm does (deepseek-v4-flash-0731, which declares webSearch: true). Other gateways: check their model metadata before switching.
• Store an API key so the provider can authenticate. The default credential reference is DEEPSEEK_API_KEY — save it through the web UI (Settings → Models) or your credentials file. Point the provider at another reference with apiKeyEnv if you prefer.

Verify it works

In any chat, ask: “帮我搜一下 XXX” or “search the web for XXX”. The response will include real source links.

Configuration

Key	Default	Meaning
apiKeyEnv	DEEPSEEK_API_KEY	Credential reference resolved for each search.
baseURL	https://tokenrhythm.studio/v1	OpenAI-compatible endpoint base; /responses is appended.
model	deepseek-v4-flash-0731	Search model; must declare webSearch: true on your gateway.
maxTokens	2048	Upper bound on generated answer tokens.
maxRetries	2	Retries for transient gateway 503 responses.

Switching to another gateway

Override the row from your own profile cordis.patch.yml (later layers win):

- id: web-search-tokenrhythm
  name: dsh-web-search-tokenrhythm
  config:
    apiKeyEnv: MY_GATEWAY_API_KEY
    baseURL: https://my-gateway.example/v1
    model: my-web-search-model

Or edit the web-search-tokenrhythm section in the GUI Settings page — it takes effect on the next search without a restart.

Not using a gateway with web_search?

This provider cannot help. Install a different search backend instead (e.g. @deepseek-ai/dsh-web-search-exa or @deepseek-ai/dsh-web-search-perplexity) and set searchProvider to its id.

Troubleshooting

Symptom	Cause / Fix
WEB_PROVIDER_CREDENTIAL_MISSING	No key for apiKeyEnv. Save it in Settings → Models or set apiKeyEnv to a reference you have.
tools.0.input_schema 类型错误 / invalid tool type	The gateway does not support the web_search tool. Switch gateway or provider.
SERVICE_BUSY (HTTP 503)	Gateway overloaded; retried automatically up to maxRetries. Retry the search later if it still fails.
Search works but sources have no titles	Expected: this gateway returns opened-page URLs plus a generated answer; no structured title/snippet.

How it works

One search issues a full Responses model call with tools: [{ type: 'web_search' }]. The gateway runs server-side retrieval; the model may open several pages (web_search_call items). The provider:

• collects opened-page URLs as citeable sources[] (deduplicated, #ws_call_id= tracking fragments stripped);
• takes the model's final_answer text as content;
• retries transient 503 responses.

Failures surface as WEB_PROVIDER_ERROR; caller cancellation as WEB_ABORTED. MIT licensed.

Development

# Build (TypeScript → lib/)
pnpm exec tsc -b .

# Test (Vitest, 16 cases)
pnpm exec vitest run tests/provider.spec.ts

# Package
npm pack

The package follows the DeepSeek Harness plugin conventions: a Cordis plugin exporting name / inject / apply / Config, plus a bundle patch (cordis.patch.yml) that mounts itself when the package is listed in a profile's dsh.profile.bundles.

License

MIT