@lemoncat7/dsh-web-search

Configurable web access for DeepSeek Harness. The plugin provides DSH's native web_search and web_fetch tools: administrators select the search backend, while direct public-page retrieval uses the same plugin-scoped proxy and security boundary. A constrained web_source tool is also available when an agent must verify raw HTML markers or script-embedded data under an explicit source procedure; normal page reading remains the job of web_fetch.

Compatibility

Release 0.2.1 is built and deployment-tested against DeepSeek Harness 0.1.2-rc.1 on Node.js 22.19+ or 24+. Its DSH host, settings, credentials, tools, locale, and client UI peer contracts target the 0.1.2-rc.1 API family.

Supported providers:

• SearXNG — self-hosted, keyless general web search
• Wikipedia — keyless encyclopedia search; Simplified Chinese by default
• Tavily — agent-oriented API search
• Brave Search — traditional web and news search
• Gemini Grounded Search — Google Search grounding and URL Context for queries containing a URL

Only the selected provider receives each query. The plugin does not silently fan out requests or spend quota on fallback services.

Install

dsh plugin --profile web add @lemoncat7/dsh-web-search

Open Settings → Plugins → Plugin configuration → Web search to choose a provider, store API credentials, test the draft configuration, and save it without restarting DSH.

The default backend is keyless Chinese Wikipedia so a new installation can be verified immediately. Use SearXNG, Tavily, Brave, or Gemini for general web coverage.

Local SearXNG

docker compose -f deploy/searxng/compose.yml up -d
curl -fsS -X POST http://127.0.0.1:8080/search -d 'q=DeepSeek&format=json'

The included deployment binds to loopback only. Do not expose it publicly without authentication, rate limiting, and the controls recommended by SearXNG.

The settings card can read general-purpose engines from the configured instance, probe them with bounded concurrency, sort successful engines by latency, and save a per-client engine selection. Failed or empty searches can be retried from zero to three times.

Security

• Search JSON and fetched text are capped at 2 MiB; search responses are schema-validated.
• web_fetch accepts credential-free public HTTP(S) URLs only and blocks loopback, private, link-local, and reserved destinations.
• Fetch redirects are capped at five and every hop is revalidated. Binary bodies are rejected.
• web_source applies the same fetch policy and supports only bounded range reads or a bounded set of exact source markers, with at most 120,000 returned characters.
• API keys are resolved through DSH Credential Store and sent only in provider-defined authorization headers.
• Browser settings routes enforce exact same-origin writes, cap request bodies, and never return secret values.
• Fetch and source inspection are constrained read-only public-web capabilities, not internal-network access or arbitrary command execution.

Proxy

Set DSH_WEB_SEARCH_PROXY to an absolute HTTP(S) proxy URL. Search and fetch share this dispatcher; it remains scoped to the plugin and does not change networking for model providers or other DSH plugins. Loopback and private IPv4 destinations bypass it for local search backends, while web_fetch independently rejects non-public destinations. Add custom direct hosts to the comma-separated DSH_WEB_SEARCH_NO_PROXY variable.

Development

npm install
npm run check

Attribution

Maintained as an MIT-licensed derivative of zmh2000829/dsh-web-search-multi. The original license notice is retained.