dsh-bing-search

简体中文

<p align="center"> <img src="docs/social-preview.png" width="100%" alt="dsh-bing-search: DuckDuckGo first, Bing by language, quality scores" /> </p>

Web search for DeepSeek Harness (DSH), implemented as a small MCP server and powered by curl_cffi.

search order:

1. Probe DuckDuckGo HTML (html.duckduckgo.com) and cache reachability for about 60 seconds. From mainland China this probe often fails unless a proxy is configured.
2. Use DDG when it is reachable.
3. Fall back to Bing when DDG is down, rate-limited (HTTP 202 / challenge), or the result set is quality_label=poor.
4. Route Bing by language: Chinese / zh-* markets go to cn.bing.com, otherwise www.bing.com.

Every search response includes quality_score (0–1) and quality_label (good / weak / poor). Treat poor as unusable (dictionary pages, first-token junk). Do not cite those titles.

It gives a DSH agent three browser-style tools:

• mcp__web__search — search the public web and return normalized organic results.
• mcp__web__open — open a public web page and extract readable text.
• mcp__web__find — find text inside a long page and return nearby context.

DSH agent
  -> @deepseek-ai/dsh-mcp-client
  -> dsh-bing-search (MCP/stdio)
  -> curl_cffi.AsyncSession(impersonate="chrome")
  -> html.duckduckgo.com          (if reachable)
  -> else cn.bing.com / www.bing.com

Mainland China: DuckDuckGo is often unreachable without a proxy or VPN. That is expected. The plugin then uses Bing and sets warnings to duckduckgo_unreachable. The MCP child does not inherit your shell HTTP_PROXY / HTTPS_PROXY (trust_env=False). To force a proxy, set DSH_WEB_PROXY on the plugin process (for example http://127.0.0.1:10808 in the cordis env: map). Do not assume DDG will work on a typical mainland home or campus network.

Community plugin: DeepSeek Harness asks third-party plugins to use the dsh-plugin GitHub topic for discovery.

Fastest install: give this repo to an agent

If your coding agent has terminal and filesystem access (Codex, Claude Code, Pi, OpenCode, etc.), paste this:

Install this DeepSeek Harness plugin into my current DSH setup:
https://github.com/Biogod2020/dsh-bing-search

Read the repository README and INSTALL.md first. Install it with uv, detect my active
DSH profile, add it through cordis.patch.yml using the required `insert` patch form,
preserve all unrelated config, use the absolute path of the installed dsh-bing-search
executable, then verify that mcp__web__search, mcp__web__open, and mcp__web__find are
registered. Finally run one real web search smoke test and report what changed.

That is the recommended path. INSTALL.md contains a deterministic install contract written for agents.

Manual install

1. Install the executable

Python 3.10+ is required. With uv:

uv tool install --force git+https://github.com/Biogod2020/dsh-bing-search.git

Find the tool bin directory:

uv tool dir --bin

Use the absolute path to dsh-bing-search (or dsh-bing-search.exe on Windows) in the DSH config below.

For development instead of a tool install:

git clone https://github.com/Biogod2020/dsh-bing-search.git
cd dsh-bing-search
uv sync --extra dev

The repository includes uv.lock for reproducible development installs.

2. Add it to DSH

DSH profiles combine a root cordis.yml with a patch layer cordis.patch.yml. When adding a new plugin through the patch layer, the entry must be wrapped in insert:

- insert:
    - id: mcp-web
      name: '@deepseek-ai/dsh-mcp-client'
      config:
        serverName: web
        transport: stdio
        command: /ABSOLUTE/PATH/TO/dsh-bing-search
        args: []
        toolCallTimeoutMs: 30000
        failOnStartupError: true
        reconnect:
          enabled: true
          initialDelayMs: 500
          maxDelayMs: 30000
          maxAttempts: 10

Do not add a bare - id: mcp-web entry to cordis.patch.yml: bare entries patch existing IDs and an unknown ID can be skipped. If you are editing the root cordis.yml directly, a normal bare plugin entry is correct. See cordis.example.yml.

3. Verify

After DSH reloads the profile, the model should see:

mcp__web__search
mcp__web__open
mcp__web__find

Then ask the agent to search for something current and open one result. A successful round trip verifies both search access and MCP registration. Restart DSH (or the MCP child) after changing plugin code; the stdio process does not hot-reload Python.

Tools

search

{
  "query": "DeepSeek Harness GitHub",
  "count": 8,
  "offset": 0,
  "market": "en-US",
  "safe_search": "Moderate"
}

Returns:

Field	Meaning
provider	duckduckgo or bing
title / url / snippet / rank	Organic result
source_id	Stable ID from the canonical URL
quality_score	0–1 overlap of the query with titles/snippets
quality_label	good / weak / poor
warnings	Fallback reason and quality notes

Use market=zh-CN for Chinese queries. If the query contains CJK, Bing fallback still uses cn.bing.com even when market is en-US.

DuckDuckGo /l/?uddg= and Bing /ck/a redirects are decoded where possible. Common tracking parameters are stripped and duplicate URLs are merged.

For people, papers, or illustrated blogs, search the author name or a short proper noun first. If quality_label is poor, do not keep lengthening the query. Chinese academic metadata belongs in a specialized corpus (for example CNKI), not this general web search.

open

{
  "url": "https://example.com/article",
  "max_chars": 24000
}

Fetches public HTTP(S) pages with curl_cffi, applies DNS/IP checks and safe redirects, limits response size, and extracts readable text without executing JavaScript.

open is built for article-like HTML. It is not a browser. Live DSH runs showed that weather and other widget-heavy sites (tianqi.com, weather.com.cn, and similar) often yield navigation chrome or near-empty text: Trafilatura finds no main article, then the fallback dumps the whole DOM. status can still be ok. For those pages, trust the search snippet, or open a simpler article URL. Do not expect live temperature, maps, or other JS-rendered UI.

find

{
  "url": "https://example.com/article",
  "pattern": "DeepSeek",
  "max_matches": 5,
  "context_chars": 700
}

Returns matching regions without injecting the entire page into the model context.

Why three tools instead of one giant search_and_summarize tool?

The plugin keeps retrieval deterministic and lets the DSH model control the research loop:

search -> inspect candidates -> open -> find / search again -> synthesize

The plugin handles HTTP, parsing, cleaning, caching, engine fallback, provenance, and a quality mark. The agent decides what to search, which sources to trust, when to reformulate the query, and when enough evidence has been collected. The agent must read quality_label and warnings.

Configuration

Environment variable	Default	Purpose
DSH_BING_SEARCH_URL	https://www.bing.com/search	Override Bing HTML endpoint only when set to a non-default value (tests). Otherwise the host is chosen by language
DSH_WEB_IMPERSONATE	chrome	curl_cffi browser fingerprint
DSH_WEB_PROXY	empty	HTTP/HTTPS/SOCKS proxy. The process uses trust_env=False and does not inherit HTTP_PROXY
DSH_WEB_TIMEOUT_SECONDS	20	Transfer timeout
DSH_WEB_CONNECT_TIMEOUT_SECONDS	8	Connect timeout
DSH_WEB_MAX_BODY_BYTES	5242880	Maximum body size for open
DSH_BING_MAX_BODY_BYTES	2097152	Maximum search-page body size
DSH_WEB_MAX_REDIRECTS	8	Maximum redirects
DSH_WEB_CONCURRENCY	8	Maximum in-process concurrent requests
DSH_BING_CACHE_TTL_SECONDS	90	Search cache TTL
DSH_WEB_CACHE_TTL_SECONDS	600	Page cache TTL

Tests

Offline tests (parsers, quality score, locale routing, DDG-first / Bing fallback):

uv run pytest -m "not live"

Live smoke test:

RUN_LIVE_BING=1 uv run pytest -m live -s

The marker name is still live / RUN_LIVE_BING. A live run hits DDG first and only uses Bing if DDG is unavailable.

CI covers Python 3.10, 3.12, 3.13 and 3.14.

Design and safety notes

This is an unofficial DuckDuckGo HTML + Bing HTML adapter. It does not use the retired Bing Search API.

• DDG markup lives in src/dsh_bing_search/providers/ddg.py.
• Bing markup lives in src/dsh_bing_search/providers/bing_parser.py.
• Quality scoring lives in src/dsh_bing_search/quality.py and is engine-agnostic.
• Requests use curl_cffi.AsyncSession with browser impersonation.
• User-supplied page URLs are restricted to public HTTP(S) targets and safe redirect handling is enabled.
• Response bodies are size-limited.
• CAPTCHA / challenge / HTTP 202 pages are reported as status="blocked"; the plugin does not attempt to bypass them.
• Headless Bing on www.bing.com often returns structurally valid but unrelated cards. cn.bing.com helps some hot Chinese queries; long-tail names and titles can still collapse to the first token. That is what the quality mark is for.
• open does not automatically retry slow target sites; increase the timeout environment variables if needed.

Community

DeepSeek Harness is currently in developer preview, so plugin interfaces may still evolve. For DSH-specific support and discovery:

• Browse the dsh-plugin topic.
• See the DeepSeek Harness repository.
• Join the DSH community channels linked from the official repository.

Contributions and parser fixes are welcome.

License

MIT