vitas/dsh-web-search-gateway

dsh-web-search-gateway is a native plugin built specifically for DeepSeek Harness that re-routes the built-in web_search tool from DeepSeek's own billable endpoint to any OpenRouter-compatible gateway, so citations are paid for on the same account as the chat and at the gateway's price rather than as a full DeepSeek model turn. The package ships a bundle patch that the install line wires in automatically, and it is mounted with dsh plugin add dsh-web-search-gateway --profile web from the npm package, dsh plugin add github:vitas/dsh-web-search-gateway from a GitHub checkout, or dsh plugin add link:$PWD from a local clone built with npm install && npm run build. The API key is supplied via dsh credential set OPENROUTER_API_KEY (or exported before launching dsh) and is resolved per call through the credentials service first and then the process environment, so rotating it never requires a restart, but the patch and the linked host entry are read at boot so dsh web must be restarted once after install.

The typical workflow begins in Settings → Plugins → Web search (OpenRouter), where the user sets Endpoint, Search model, and pastes the API key, then runs Test search to see a source count, latency, and the first hits, with the chat model kept independent so a cheap chat model stays exactly as it is. Cost levers, cheapest first, are: pick the cheapest model that actually searches on the chosen gateway (verified on api.b.ai to be gpt-5.4-nano returning 2 citations), set dsh-tool-web's searchMaxQueries to 1 so one tool call becomes one search, keep maxResults low, and leave includeAnswer off so the search model's prose is not returned as content. The card exposes protocol (openai or openrouter), baseURL, model, apiKeyEnv, literal apiKey, maxResults (1–25), maxOutputTokens (1024 default), allowedDomains, excludedDomains, and OpenRouter-specific passthroughs such as engine, searchContextSize, maxUses, maxTotalResults, referer, and title. It targets DSH users who want lower search bills on api.b.ai-style gateways and anyone needing verified server-side web search through DeepSeek native clients. Dependencies, limits, and first-run caveats are concrete: Node ≥ 20 is required, the build uses esbuild + typescript with the committed lib/client.js enforced by CI, 35 tests run without network or credentials and one live smoke test is opt-in via DSH_WEB_SEARCH_LIVE=1 OPENROUTER_API_KEY=sk-... plus DSH_WEB_SEARCH_BASE_URL and DSH_WEB_SEARCH_MODEL, and there are no runtime imports of @deepseek-ai/* and no install-time script. The license is Apache-2.0; measured gateway limits show that glm-*, deepseek-*, qwen*, minimax-*, mimo-*, and gpt-5-nano accept the request and silently drop the search (HTTP 200 with 0 results), Gemini and Anthropic endpoints fail, and chat-completions-only gateways return 404 on /responses. The configuration entry is marked overridden and can be reset from the card; on first run, expect WEB_PROVIDER_CREDENTIAL_MISSING if no key is set, an explicit failure if the named model does not search on that gateway, and the requirement to restart dsh web once after install.