dsh-websearch

@240xu/dsh-websearch

Aggregated web search provider for DeepSeek Harness — one plugin id unified that fans out to eleven backends concurrently, merges + URL-dedups results, and stays usable even when some backends go down.

A pure-Cordis drop-in: registers ONE provider at ctx.web so the dsh-web selection rule never fires WEB_PROVIDER_AMBIGUOUS. Zero-config search out of the box (Exa + Parallel + DuckDuckGo + SearXNG are keyless); API-key backends (DeepSeek / Anthropic / OpenAI / Brave / Tavily / Serper / Mojeek) auto-activate once their key is supplied via the credentials service or env.

DSH（DeepSeek Harness）原生插件：向 ctx.web 注册唯一的聚合搜索 provider unified，内置十一后端并发 fan-out，合并 + URL 去重后返回——即使部分后端宕机仍可用。

纯 Cordis 直插：只注册一个 provider，所以 dsh-web 的选择规则永远不会触发 WEB_PROVIDER_AMBIGUOUS。零配置即可搜索——Exa + Parallel + DuckDuckGo + SearXNG 四者皆无 key、开箱即用；DeepSeek / Anthropic / OpenAI / Brave / Tavily / Serper / Mojeek 在提供 API key 后自动激活。

Backends | 后端

可用性说明：available() 只做本地配置检查（不联网探测可达性）；密钥在每次搜索时按操作解析，缺失时该后端以明确错误码软失败（不阻塞其他后端）。全部缺 Key 时错误会指向设置页。

SearXNG 实例 | Public instances

SearXNG 默认指向 https://searx.be；公共实例可用性随网络环境波动，失败时错误信息会提示换源。 在设置面板修改 searxngBaseURL 即可切换，常见候选：

Install | 安装

方式一：作为 profile 依赖安装（推荐，规范做法）

cd ~/.dsh/profiles/web
# 在 package.json 的 dependencies 里加入：
#   "@240xu/dsh-websearch": "file:/path/to/@240xu/dsh-websearch"
pnpm install

插件自带 dsh.bundle 元数据（package.json 声明 "dsh": {"bundle": {"patch": "./cordis.patch.yml"}}），把它加入 profile package.json 的 dsh.profile.bundles 列表后即自动注册，无需在 cordis.patch.yml 手写条目。

只需在 ~/.dsh/profiles/web/cordis.patch.yml 里把内置 web_search 工具指到 unified provider：

- id: web
  name: '@deepseek-ai/dsh-web'
  config:
    searchProvider: unified
- id: web-search-deepseek
  name: '@deepseek-ai/dsh-web-search-deepseek'
  disabled: true

重启 dsh web 即生效。

Settings | 配置

v2.0.3 起，配置界面位于 Settings → 左侧栏「搜索 / Web Search」（顶级分区，与通用/模型/插件同级）。

每个 API Key 字段旁有 「获取 Key ↗」 直达对应平台控制台；保存即写入凭证库并立即生效。

插件向 DSH 设置面板注册 unified-search namespace（Settings → Unified Search），全部字段扁平化、开箱可编辑：

全局

v2.1 结果策略：设置面板新增「结果策略」分区。时间/语言过滤按各后端能力自动映射——Brave freshness/country/search_lang、Tavily time_range、Serper tbs/hl/gl（Google 日期语法）、SearXNG time_range/language、DDG df；不支持的后端自动忽略对应维度。内部超时以真实原因失败（backend "" timed out after Nms），不再被静默降级为取消；Mojeek 的 API Key 改走 Authorization 头，不再出现在 URL 中。

结果健康遥测（默认开）：每次搜索的 content 尾部附一行 [websearch backends] exa ✓5ms/3 · ddg ✗30000ms， 模型可据此自诊断并建议用户调整设置；设置面板「结果策略」可关闭（resultTelemetry）。

11 个后端开关：enableExa / enableParallel / enableDdg / enableSearxng（默认开）；enableBrave / enableTavily / enableSerper / enableMojeek / enableDeepseek / enableAnthropic / enableOpenai（默认关）

API Key 授权（credential-ref，填环境变量名，实际 key 存于 credentials 服务或环境变量）：

Base URL / Model：每个 key-gated 后端都有对应 ${id}BaseURL；DeepSeek/Anthropic/OpenAI 另有 ${id}Model；Tavily 有 tavilySearchDepth（basic/advanced）。

环境变量兜底：未在 credentials 配置时回退读同名环境变量；DSH_UNIFIED_SEARCH_BACKENDS 可逗号分隔强制指定启用集合。

Design | 设计

• One provider, no ambiguity: a single registeredSearchProvider({id:"unified"}) — the dsh-web seam's selection rule picks it unambiguously, and search() caps maxResults itself.
• Promise.allSettled fan-out: every enabled + available backend fires concurrently; a single backend failure is demoted to a soft null so the rest still contribute — only when ALL fail does the provider throw WEB_PROVIDER_ERROR.
• Per-backend abort demotion: a single backend aborting becomes soft-null; the provider only rethrows WEB_ABORTED when the caller's own AbortSignal fires.
• Dedup strategies: url keeps the historical URL-key merge with cross-backend field fill-in; url+title additionally collapses same-story syndicated mirrors (title Jaccard >= 0.9, CJK-aware tokenizer) while still enriching the kept entry.
• Deterministic rerank (optional): query-term overlap scoring (title x3 / snippet x2 / URL x1); stable ties preserve fan-out priority — same input always yields the same order.
• Honest timeouts: each backend runs under an internal abort controller merged into the caller signal; when only the internal timer fires, the failure is reclassified as WEB_PROVIDER_ERROR ("backend timed out") instead of being masked as a user cancellation.
• Concurrency & Timeout Control: concurrency (default 6) limits simultaneous calls; backendTimeoutMs (default 30s) caps each backend.
• Streamable-http MCP, custom handshake: lib/util/mcp-client.js implements initialize → notifications/initialized → tools/call with Mcp-Session-Id header caching against Exa and Parallel — no dependency on the full MCP SDK.
• Host-logger diagnostics: each backend routes request/outcome through lib/util/log.js to the host logger (ctx.logger), best-effort and never thrown - deliberately NOT session events: the session event vocabulary is a closed set third-party plugins must not extend (an unknown envelope type makes the whole session log unreadable).

Architecture | 架构

lib/
  index.js                 # name/inject/Config/apply — registers settings + provider
  provider.js              # UnifiedSearchProvider — fan-out, abort demotion, dedup, truncate, concurrency, timeout
  util/
    abort.js               # isAbortError / searchAborted / throwIfSearchAborted / maybeAbortError
    mcp-client.js          # streamable-http MCP client (initialize + tools/call + session cache)
    log.js                 # recordBackendRequest / recordBackendOutcome → session event log
  backends/
    exa.js                 # Exa (web_search_exa) via mcp-client
    parallel.js            # Parallel (web_search) via mcp-client
    ddg.js                 # DuckDuckGo HTML scrape (uddg redirect decode)
    searxng.js             # SearXNG REST JSON (keyless)
    brave.js               # Brave Search REST (key-gated)
    tavily.js              # Tavily REST (key-gated, AI answer + deep search)
    serper.js              # Serper.dev REST (key-gated, Google scrape)
    mojeek.js              # Mojeek REST (key-gated)
    anthropic-like.js      # shared: DeepSeek + Anthropic (Messages API + web_search_* server tool)
    openai.js              # OpenAI /responses + web_search tool (url_citation parsing)
    index.js               # unified registry + individual exports
tests/
  parse.test.js            # 28 unit tests for each backend's pure parse fn
  provider.test.js         # 5 fan-out tests (merge/dedup, abort demotion, all-fail, maxResults cap, concurrency)

Test | 测试

node --test tests/

33/33 pass.

License | 许可

MIT © 2026 240xu