@gausszhou/dsh-web-search-local
面向 dsh Web seam 的无需密钥多引擎网页搜索和页面抓取提供商（不依赖 DeepSeek）。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add @gausszhou/dsh-web-search-local@0.2.1Web Search Local 以 @gausszhou/dsh-web-search-local 发布，当前版本为 0.2.1。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
English | 简体中文
Keyless, multi-engine web search + page fetch providers for the DeepSeek Harness (dsh) ctx.web seam. Works with any model backend — including fully local models — with no API keys and no dependency on DeepSeek's server-side search.
dsh's built-in web_search tool is model-agnostic: it only calls ctx.web.search(). What depends on DeepSeek is the default search provider (dsh-web-search-deepseek), which sends each query to DeepSeek's web_search_20250305 server tool with DEEPSEEK_API_KEY. Switch to a local model (Ollama etc.) and that provider has no key, so search dies.
This package registers two providers that do the HTTP themselves:
| provider id | capability | engines |
|---|---|---|
local-multi | web_search | three layers in order — SearXNG (when configured) → Google/DuckDuckGo/Mojeek (global) → Bing/Baidu/Sogou/360 (CN); same-layer engines are queried in parallel and their results merged round-robin; a layer with no results degrades to the next |
local-fetch | web_fetch | direct GET, charset-aware decoding (incl. gbk), html/text bodies |
A node process does not use the OS/browser proxy. If engines like DuckDuckGo are unreachable in your network, the provider resolves a proxy automatically:
proxyUrl config (explicit, or 'off' to force direct)HTTPS_PROXY / HTTP_PROXY / ALL_PROXY environment variables7890 Clash, 7897, 10809/10808 v2rayN, 1080, …)The proxy applies only to the global engines (google, duckduckgo, mojeek) — they are the ones that need a tunnel in CN-like networks. The CN engines (bing, baidu, sogou, 360) and a private SearXNG instance always connect directly (CONNECT tunnel is only used for the global engines; if the tunnel dies at transport level those requests fall back to direct too). This split matters: proxying CN engines from a foreign exit IP triggers Baidu's verification wall, Sogou's antispider captcha, and 360's 302 redirects; and routing your own SearXNG instance through a VPN node can serve stale or empty result sets.
Published on the npm registry.
npm install @gausszhou/dsh-web-search-local
Or register and activate it with the dsh CLI in one step:
dsh plugin add @gausszhou/dsh-web-search-local
The package ships a dsh.bundle patch, so dsh plugin add installs and
auto-activates it: the web profile switches to searchProvider: local-multi
/ fetchProvider: local-fetch and the built-in web-search-deepseek provider is
disabled. No manual cordis.patch.yml edit is required.
To tune the defaults the bundle applies, override them in your profile's
cordis.patch.yml ($DSH_HOME/profiles/web/cordis.patch.yml for the web
profile) after the bundle layer:
- id: web
config:
searchProvider: local-multi
fetchProvider: local-fetch
- id: web-search-deepseek
disabled: true
- insert:
- id: web-search-local
name: '@gausszhou/dsh-web-search-local'
config:
engines: [searxng, google, duckduckgo, mojeek, bing, baidu, sogou, 360]
Put this package anywhere the dsh process can read, e.g. $DSH_HOME/profiles/web/plugins/web-search-local/ (Windows: C:\Users\<you>\.dsh\profiles\web\plugins\web-search-local\). Add to your profile's cordis.patch.yml ($DSH_HOME/profiles/web/cordis.patch.yml for the web profile):
- id: web
config:
searchProvider: local-multi
fetchProvider: local-fetch
- id: web-search-deepseek
disabled: true
- insert:
- id: web-search-local
name: 'file:///C:/Users/<you>/.dsh/profiles/web/plugins/web-search-local/index.js'
config:
engines: [searxng, google, duckduckgo, mojeek, bing, baidu, sogou, 360]
web_search now returns plain source lists (no server-side summary) and works with any model.config:
engines: [searxng, google, duckduckgo, mojeek, bing, baidu, sogou, 360] # member list (execution is layered: searxng → global → cn, parallel within a layer)
skipWithoutProxy: [google, duckduckgo, mojeek] # engines NOT attempted when no proxy is available ([] = always attempt)
searxngBaseUrl: 'http://127.0.0.1:8080' # optional; runs first when set
proxyUrl: '' # '' auto | 'off' | 'http://host:port'
searchTimeoutMs: 12000
fetchTimeoutMs: 20000
maxFetchBytes: 1048576
maxSources: 12
cacheTtlMs: 300000 # in-memory result cache
engineMinIntervalMs: 1500 # min gap between engine calls (anti rate-limit)
engineCooldownMs: 600000 # skip engine after captcha/anomaly/verification-wall (0 = off)
engineRetryCooldownMs: 60000 # skip engine after generic failure (0 = off)
userAgent: '<browser-like UA>'
The default engine list executes in three layers, in order, with same-layer engines queried in parallel and merged:
searxngBaseUrl is set) is already a meta-search aggregation, so its results win immediately and lower layers are skippedskipWithoutProxy)A layer with no results (empty, blocked, or skipped engines) degrades to the next, so the global engines never break the directly reachable cn engines. The google engine is scrape-fragile (consent wall, sorry/ bot detection, JS-required enablejs wall); for reliable Google results prefer a SearXNG instance with the google engine enabled. On open networks where the global engines work directly, set skipWithoutProxy: [].
A private SearXNG instance (Docker: docker run -p 8080:8080 searxng/searxng) is the most robust engine of all: meta-search aggregation, a JSON API, no per-engine scraping.
Search engines (especially DuckDuckGo) throttle scripts. Three mechanisms keep a single-engine setup usable:
engineMinIntervalMs (anti rate-limit), while different engines in one layer start together.blocked by captcha / anomaly check / Baidu's verification wall, or HTTP 403/429), it is skipped for engineCooldownMs (default 10 min); generic failures (transport, HTTP errors) only trip the shorter engineRetryCooldownMs (default 60 s). While cooling down the engine is skipped and the failure is reported in the aggregated error.html.duckduckgo.com endpoint is bot-walled, the same query is retried once against lite.duckduckgo.com/lite/, which is more tolerant of scripts. If the lite endpoint is walled too, the engine reports blocked by anomaly check (html and lite) and trips the long engineCooldownMs breaker instead of hammering both endpoints on every search.A blocked engine never makes the whole search fail if other engines remain; with a single engine it fails fast with a "cooling down" reason instead of hammering the walled endpoint.
The model can steer which engine a search uses, per call, two ways:
web_search, this plugin registers web_search_engine with two optional arguments:
engine: one engine — searxng, google, duckduckgo, mojeek, bing, baidu, sogou, 360engines: an ordered priority list of engines
When neither is given, the call degrades to the configured default three-layer chain, exactly like web_search.ctx.web.search({ query, engine }) or ctx.web.search({ query, engines }) gets the same override; unknown engine names fail with WEB_PROVIDER_ERROR listing the valid ids.An explicit override replaces the configured chain entirely (including the SearXNG auto-prepend) — the model's explicit choice wins. The requested engines are grouped into the same three layers (searxng / global / cn) and parallel-merged within a layer, exactly like the default chain; a single engine simply runs alone. Pacing, circuit breaking, and skipWithoutProxy still apply to the requested engines, so a pinned-but-unreachable engine fails fast instead of breaking the search.
Remove the web override, the web-search-deepseek disable, and the inserted row from cordis.patch.yml.
/link?url= wrappers are resolved server-side (the stub page embeds the real target); 360's wrappers expose the real URL in the anchor's data-mdurl attribute, which the parser reads directly.google engine scrapes the HTML SERP with a dual-layout parser (basic gbv=1 markup and the modern JS-era markup) and sends a CONSENT/SOCS cookie to bypass the EU consent interstitial. Google often serves scripts a JS-required wall (/httpservice/retry/enablejs) or a sorry/ captcha instead of results — both are detected and trip the long circuit-breaker cooldown with a clear reason, and the global layer degrades to the cn layer. For reliable Google results, use a SearXNG instance with the google engine enabled.web_search returns { sources: [{ url, title?, snippet?, publishedAt? }], truncated }. Within a layer the engines run in parallel and their sources are merged round-robin, deduped, and capped at maxSources (truncated is set when the merged list exceeds the cap); a layer with no results degrades to the next. publishedAt is a best-effort YYYY-MM-DD filled when the engine renders a date (SearXNG's publishedDate, or date text in Bing/Baidu/Sogou/360 result blocks) and omitted otherwise — the same optional semantics as the official provider's page_age field.fetch + node:http/https/net/tls only, plus the dsh-provided @deepseek-ai/dsh-web (declared as a peerDependency; every dsh profile already ships it).WebError with WEB_PROVIDER_ERROR (engine/transport/timeout, engine errors aggregated) or WEB_ABORTED (caller cancellation) — the same vocabulary the official providers use.web_fetch needs tool-web's fetch: true; the shipped standard agent preset ships with fetch: false — copy the preset to $DSH_HOME/.agent-presets/ and flip it there.The plugin declares a schemastery Config schema (one field per entry of defaultConfig()) and registers a settings namespace (web-search-local) through dsh's settings service — the same mechanism the built-in web-search-deepseek, shell, and agent-loop plugins use. This gives you:
Config and SETTINGS_NAMESPACE are also named exports.)config passed to apply(ctx, config) from cordis.patch.yml still wins.The visible card in the "Plugin configuration" panel is a client-side React component — the built-in cards (Terminal / Agent loop / Web search) are hand-written inside the bundled
dsh-client-ui-settings-pluginsclient package. For this external plugin to gain an editable card in that panel it must additionally ship a./clienthalf that registers asettings.plugin.itemcard. So "the configuration UI" for this plugin = the server-side schema integration above plus a client card.
MIT