CGArtLab Search

Unified multi-backend internet search. Backends: Brave/Exa (web & deep research), Context7 (library/framework/API docs), Zhihu OpenAPI global search (中文全网), Tencent News (official CLI: 新闻搜索 + 热点榜), plus key-less web_fetch backends — Bing RSS (general web search), 腾讯新闻热榜 and 虎扑足球 (news/体育/足球), Stack Overflow, GitHub, Hacker News (developer sources). Use whenever the user needs information from the internet — searching the web, researching a topic, looking up docs, finding news or current events (including 足球比赛新闻), investigating anything — in Chinese or English, even when they don't explicitly say "search". Activate on 查一下/搜一下/调研/全面了解/最新动态/新闻/足球赛况/知乎上怎么看/look up/research/how do I use X/what's new. Do not activate for purely local tasks (reading/editing local files, running local commands) unless external sources are needed.

cgartlab-search — unified search

Route a query to the best backend(s), call each via the transport your environment supports, normalize the results, and answer the user with a cited Markdown summary.

When to use

Any external-information need: web search, deep research, library/API docs, news, sports news, Chinese knowledge. If the answer must come from the internet (not the user's local files or your training knowledge), use this skill.

Step 1 — route

Read references/routing.md for the full table (curl column vs DSH/web_fetch column). Quick guide:

Intent	backend(s)
general web / facts / "查一下"	bing (brave in curl harnesses)
行情 / 价格 / 走势 (market data)	bing (keyword-rewritten query) → direct-fetch data site (routing.md market two-step)
library/framework/API docs	context7
dev Q&A / errors	so (+ hn)
codebase / tool finding	github
latest / hot news	tencent (热点榜) or qqnews (+ bing)
中文新闻 / 中文全网	tencent (新闻搜索) + zhihu (全网中文)
足球 / 体育 news	hupu (+ bing, qqnews)
deep research / broad	bing + context7 + hn (add exa+brave with curl)
知乎 / Chinese knowledge	zhihu (official Global Search API — indexes the whole web, strong on 中文)

Step 2 — call

Open references/backends.md and follow the recipe for each chosen backend. Transport:

• Shell with network (Claude Code, Codex, terminal) → use the curl recipes (all backends), or run scripts/search.mjs <backends> "<query>" --limit 5 (Node 18+, zero deps, emits normalized JSON). scripts/search.sh is the curl+python variant (it probes $PYTHON/python3/python); prefer search.mjs wherever Node exists.
• No shell network (DSH) → use your harness's WebFetch/web_fetch tool on the GET URLs in backends.md. Check its host-reachability list first; skip unreachable hosts with an errors entry.

Keys only matter for brave/exa/tencent/zhihu: env vars (BRAVE_SEARCH_API_KEY, ...) or secrets.json next to SKILL.md. Key-less backends need nothing.

Step 3 — normalize + answer

Merge into the normalized shape (see backends.md), dedupe by URL, then write the answer:

markdown

## Docs (Context7)
- **<title>** — <one-line>. [source](url)
  <key excerpt>

## Web / Q&A (bing / so / github / hn)
- **<title>** — <snippet>. [link](url)

## News (qqnews / hupu / bing)
- **<title>** (<date>, <source>) — [link](url)

If a backend errored or was unreachable, add one line, e.g. "Brave/Exa 在本环境不可达，已用 bing 替代" or "知乎未配置，已用 bing site:zhihu.com 替代". Never fabricate results for a failed backend.

Degradation

Missing keys, anti-bot 403s, and blocked hosts are normal — record in errors and continue. Only if ALL chosen backends fail, say so plainly and suggest checking keys / network / backend configuration.

Backend status (verified 2026-10-20)

• Live & key-less (curl + web_fetch): bing, context7, qqnews, hupu, so, github, hn.
• curl-only (keyed): brave, exa — unreachable from DSH's sandboxed egress.
• Live (keyed): zhihu (official OpenAPI global search, HTTP) and tencent (official tencent-news-cli binary — needs a local install, not just a key).