Skip to main content
A

dsh-qwen-websearch

archerkattri/dsh-qwen-websearch

Qwen Token Plan web search provider that runs web_search on Token Plan credits through the Responses API web_search tool, summarizing with Qwen3.8 Flash and returning source URLs.

Install

dsh plugin --profile web add github:archerkattri/dsh-qwen-websearch

README

dsh-qwen-websearch

CI License: MIT

Web search for DeepSeek Harness (dsh) backed by Qwen Token Plan.

It registers a DSH WebSearchProvider that runs Qwen's native web search (Responses API + built-in web_search tool) on your Token Plan credits and summarizes results with Qwen3.8 Flash, returning the search call's own source URLs as citations. No DeepSeek per-search endpoint is used.

Why this provider

  • Native search, real sources. Retrieval and summarization happen in one Qwen request; the web search call reports the URLs it used, so web_search lists sources the way the tool expects.
  • No DeepSeek per-search fee. Searches are billed as Token Plan tokens.
  • No scraping. The original community plugin scraped DuckDuckGo, which now responds to server-side fetches with a bot-challenge page. This provider uses Qwen's server-side search instead.

Install

From GitHub:

dsh plugin --profile <your-profile> add github:archerkattri/dsh-qwen-websearch

Or open Plugins → Add plugin and choose the checked-out directory.

Configure

  1. Store a Qwen Token Plan API key (sk-sp-...).
    • Easiest: Settings → Models → Qwen4 (Token Plan) (or any Qwen Token Plan provider) and save the key. It lands in the DSH credential store.
    • Or export QWEN_TOKEN_PLAN_API_KEY, or QWEN_TOKEN_PLAN_INDIVIDUAL_API_KEY.
    • The plugin tries apiKeyEnv first, then QWEN_TOKEN_PLAN_API_KEY, then QWEN_TOKEN_PLAN_INDIVIDUAL_API_KEY, then DASHSCOPE_API_KEY.
  2. Nothing else is required: the bundle pins web.searchProvider to qwen-token-plan, so the existing web_search tool uses it immediately.
  3. Optionally set the summarizer model in the provider row (model:, default qwen3.8-flash).

Check Settings → Web search: the provider appears as Qwen Token Plan.

Configuration reference

Bundle patch:

- id: web
  name: '@deepseek-ai/dsh-web'
  config:
    searchProvider: qwen-token-plan
    fetchProvider: http
- insert:
    - id: web-search-qwen
      name: 'dsh-qwen-websearch'
      config:
        baseURL: https://token-plan.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1
        model: qwen3.8-flash

Provider options (row config):

FieldDefaultMeaning
apiKeyEnvQWEN_TOKEN_PLAN_API_KEYFirst credential reference to try
apiKey—Literal key (prefer apiKeyEnv)
baseURLhttps://token-plan.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1Qwen Token Plan endpoint
modelqwen3.8-flashSummarizer model
effortlowThinking effort for the search request (low/medium/xhigh)
maxToolCalls2Cap on server-side search calls per query
systemPromptbuilt inSearch-answer instructions

Calling Qwen: POST {baseURL}/responses with tools: [{ "type": "web_search" }], reasoning_effort, and max_tool_calls.

Speed and timeouts

Measured against the Singapore Token Plan endpoint: a forced web search takes ~35 s at the model's default thinking level, ~22 s at low, and ~11 s with max_tool_calls: 2 — the shipped defaults. Disabling thinking entirely (enable_thinking: false) makes search queries hang, so the provider keeps thinking on at low. Each request also has a hard 110 s client-side ceiling that fails with WEB_SEARCH_TIMEOUT instead of stalling.

DSH's web_search tool has its own timeout budget (tool-web searchTimeoutMs, default 30 s). If searches time out, raise it, e.g. in your profile patch:

- id: tool-web
  config:
    searchTimeoutMs: 120000

Notes and limitations

  • The bundle overrides the web row to pin the search provider and keep fetchProvider: http. If you use a different fetch provider, delete that override or change the value; otherwise web_fetch becomes ambiguous.
  • If another search provider is also usable and you remove the pin, DSH reports WEB_PROVIDER_AMBIGUOUS until you pin one in Settings → Web search.
  • Each search is a full Qwen request (search results included in the prompt), so it consumes Token Plan tokens; there is no separate per-search fee.
  • Requires a Token Plan key; pay-as-you-go keys are not interchangeable.

Credits

Provider structure based on umicorn07/dsh-web-search-qwen (MIT), with DuckDuckGo retrieval replaced by Qwen's native Responses web_search tool and Qwen3.8 Flash summarization.

  • dsh-qwen4-effort — official Qwen4 effort levels for Qwen4 models on an existing Qwen Token Plan route.

License

MIT

Related plugins