Cnyes Financial News Search

Purpose

Search Cnyes (鉅亨網, https://www.cnyes.com/) — Taiwan's largest Chinese-language financial news portal — for articles about any stock symbol (TW, US, HK), company name, ETF, fund, market theme, macro keyword, or news event. Returns a ranked list of recent articles with title, publish timestamp (Taipei time), byline / source, category, keyword tags, related tickers, cover image, snippet, and canonical article URL; can also dereference any article ID to its full body for fact extraction. Read-only — never logs in, never posts comments, never clicks paywalled "訂閱" subscription content. Output is summarised in Traditional Chinese (zh-TW) to match the source.

When to Use

Pull the latest Cnyes articles about a specific Taiwan stock (e.g. 台積電 / 2330), US stock (NVDA, AAPL, TSM), ETF, mutual fund, or HK ticker.
Track a market theme (AI, 半導體, 電動車, 加密貨幣, 美債, 匯率, 黃金, 油價) over recent days.
Build a Cnyes-backed RAG / research pipeline that needs structured fields (newsId, publishAt, category, keyword[], byline) rather than scraped HTML.
Monitor company news events: 股東會 (shareholder meetings), 法說會 (earnings calls), 公告 (regulatory announcements), 財報 (earnings reports), M&A / IPO news.
Pull the latest items in a category (頭條 / 台股 / 美股 / 科技 / 加密幣 / 外匯) without a keyword filter.

Workflow

Cnyes ships a public JSON news API at https://api.cnyes.com/media/api/v1/ — no auth, no API key, no cookies, no anti-bot stealth, no residential proxy. The web UI is a thin Next.js client over the same data. Lead with the API. Browser navigation is a ~50× cost premium fallback for the rare cases when the API is unreachable or you need pixel-level extraction (chart screenshots, paywalled-content detection).

1. Keyword search — preferred path

GET https://api.cnyes.com/media/api/v1/search/news?q={URL-encoded keyword}&page={1..last_page}

Works for Traditional Chinese (台積電, 半導體, 美債), Simplified Chinese (台积电 — auto-mapped to TC results), English (NVIDIA, AAPL, OpenAI), and ticker codes (2330, TSM, 0050).

Response shape (verified 2026-05-22):

{
  "items": {
    "total": 2817,
    "per_page": 20,            // hardcoded — see Gotchas
    "current_page": 1,
    "last_page": 141,
    "next_page_url": "/media/api/v1/search/news?q=...&page=2",
    "prev_page_url": null,
    "from": 1, "to": 20,
    "data": [
      {
        "newsId": 6466886,
        "title": "算力軍備到來！馬斯克砸千億美元建TeraFab ASML訂單排到2029年",
        "signature": "鉅亨網新聞中心",
        "content": "EUV 訂單，已排給<mark>台積電</mark> (TSM-US)(2330-TW)、三星電子",
        "keyword": ["ASML", "TeraFab", "特斯拉", "晶片", "算力"],
        "publishAt": 1779413406,
        "category": [{"name": "美股雷達", "categoryId": 831, "slug": "us_stock"}],
        "coverSrc": { "l": {"src": "https://cimg.cnyes.cool/.../l/....jpg", ...}, ... },
        "hasCoverPhoto": 0,
        "titleWithKwd": 0,
        "contentWithKwd": 1
      }
    ]
  }
}

Sort order is recency-relevance hybrid — first ~5 results are usually the most recent strong-match articles; deeper pages skew to older / weaker matches.
Pagination: increment page=; the limit parameter is silently ignored (see Gotchas). The next_page_url is a relative path — prepend https://api.cnyes.com to use it directly.

2. Decode the search response

For each item in items.data[]:

Field	Source	Notes
`url`	`https://news.cnyes.com/news/id/{newsId}`	Canonical article URL.
`title`	`title` with `<mark>…</mark>` stripped	Highlight tags wrap the query terms — strip before display.
`published_at` (UTC)	`new Date(publishAt * 1000)`	`publishAt` is unix-epoch seconds, UTC.
`published_at_taipei`	UTC + 8h, format `YYYY-MM-DD HH:mm`	Site displays Taipei time (Asia/Taipei, no DST).
`category_name` / `category_slug`	`category[0].name` / `category[0].slug`	`category[]` is missing for raw 公告 / 鏈文 reprints — fall back to `(未分類)`.
`byline` / `source`	`signature`	e.g. `鉅亨網新聞中心`, `鉅亨網編譯許家華`, `金色財經`, `優分析`, `Knowing 新聞`.
`snippet`	`content` with `<mark>…</mark>` and `<p>`-style HTML entities decoded	Often contains inline ticker refs like `(2330-TW)`, `(TSM-US)`.
`keywords`	`keyword[]` array	Used by Cnyes for tag pages (`https://news.cnyes.com/tag/{kw}`).
`cover_image`	`coverSrc.l.src` (640×360) or `coverSrc.xl.src` (960×539)	`null` when `hasCoverPhoto === 0`.
`related_tickers`	Regex `(\d{4,5}-TW)` and `([A-Z]{1,5}-US)` over `content` + (optional) full body	TW format is the 4–5-digit code + `-TW`; US format is the ticker + `-US`.

3. Latest-in-category — no keyword needed

GET https://api.cnyes.com/media/api/v1/newslist/category/{slug}?page=1&limit=20

Slugs observed in production (others may exist — discover via https://news.cnyes.com/news/cat/{slug}):

slug	Chinese name
`headline`	頭條
`tw_stock`	台股
`tw_premarket`	台股盤前 (categoryId 908)
`wd_stock`	美股
`us_stock`	美股雷達 (categoryId 831)
`tech`	科技
`cn_stock`	陸港股
`fund`	基金
`forex`	外匯
`future`	期貨
`bc`	區塊鏈 / 加密幣
`cnyeshouse`	房產
`tw_money`	理財
`celebrity_area`	新視界
`mag`	雜誌
`anue_live`	速報

The limit parameter is honoured on this endpoint (unlike /search/news).

4. Dereference article body — when snippets aren't enough

There is no public JSON article-detail endpoint — https://api.cnyes.com/media/api/v1/news/{id} returns 422 參數無效 and the /news/list?newsId=… / /news/detail/… variants all 404 (confirmed during discovery). Get the full body from the SSR'd HTML page:

browse cloud fetch "https://news.cnyes.com/news/id/{newsId}"

This returns 200 over the plain Browserbase fetch path — no --proxies, no browser session. Parse:

Title: <title>{title} | 鉅亨網 - {category_name}</title> or the page's <h1>.
Byline + publish time line: first paragraph after <h1> matches {signature}{YYYY-MM-DD HH:mm} (Taipei time) with no separator between byline and timestamp — e.g. 鉅亨網新聞中心2026-05-22 09:30. Split on the first 4-digit-year run.
Category breadcrumb: <a href="/news/cat/{parent_slug}">{parent_name}</a><a href="/news/cat/{child_slug}">{child_name}</a> directly above the <h1>.
Body: paragraphs inside <article> / the main content <div>. Strip image captions (<figcaption>) and the "本網站各類資訊報價由路孚特 REFINITIV 提供…" disclaimer footer.
Inline ticker links: <a href="https://invest.cnyes.com/usstock/detail/{TICKER}">…</a> for US, <a href="https://www.cnyes.com/twstock/{CODE}">…</a> for Taiwan. Use these as the authoritative ticker list for the article (more reliable than parsing (TSM-US) substrings, which can appear inside non-ticker prose).
Tags: bottom of the article — <a href="/tag/{tag}">{tag}</a> blocks. Same set as the search response's keyword[] in most cases.

5. Browser fallback (only when steps 1–4 are blocked)

If a CDN edge ever blocks the API path, drive the search UI:

sid=$(browse cloud sessions create --keep-alive | node -e "let s='';process.stdin.on('data',c=>s+=c).on('end',()=>process.stdout.write(JSON.parse(s).id))")
export BROWSE_SESSION="$sid"
# Three search URL variants — pick the most useful for the task:
# (a) richest news listing (cards with category + snippet + tags)
browse open "https://news.cnyes.com/search?q={URL-enc keyword}" --remote
# (b) news-only list on www subdomain (compact, same data, fewer ads)
browse open "https://www.cnyes.com/search/news?keyword={URL-enc keyword}" --remote
# (c) combined search — also surfaces quote / ticker matches in a top block
browse open "https://www.cnyes.com/search/all?keyword={URL-enc keyword}" --remote
browse get markdown body --remote   # extract article links + headlines from markdown
browse cloud sessions update "$sid" --status REQUEST_RELEASE

A bare session (no --verified, no --proxies) is sufficient — no Akamai or captcha was observed during a 1-iter probe. Cnyes does serve a cookie-consent overlay ("為優化網站服務…"), but it doesn't block content rendering or markdown extraction. If you do need to dismiss it, click button: 繼續使用 once.

Last-resort fallback — if both API and browser are blocked, use Google site-restricted search: site:cnyes.com OR site:news.cnyes.com "{keyword}". Quality is markedly lower (Google's index lag is ~24h for breaking news), and you only get the URL + headline, not structured fields.

6. Stock / ETF / fund detail landing pages (not articles — context only)

These pages render prices and charts; they are useful when the user asks "what's the latest on $TICKER" and you want to surface the canonical Cnyes page in addition to news.

Asset	URL pattern	Example
Taiwan stock	`https://www.cnyes.com/twstock/{4-5digit code}`	`/twstock/2330` (台積電), `/twstock/0050` (元大台灣50 ETF)
US stock	`https://invest.cnyes.com/usstock/detail/{TICKER}`	`/usstock/detail/TSM`, `/usstock/detail/AAPL`
HK stock	`https://www.cnyes.com/hkstock/{code}`	(verify via search response if uncertain)
Mutual fund	`https://invest.cnyes.com/funds/detail/{name URL-enc}/{id}`	`/funds/detail/.../A13029`
Crypto	`https://crypto.cnyes.com/coin/{symbol}`	(subdomain)
Forex	`https://www.cnyes.com/forex/{pair}`	e.g. `USDTWD`

These pages are JS-heavy SPA pages — browse cloud fetch returns the empty Next.js shell. To extract price data you must drive a browser session and wait for the React app to hydrate (~2–3s). For most news-search tasks you do not need to open these — the article-side ticker references in step 4 are sufficient.

7. Honesty layer — required output discipline

Cnyes mixes three content classes in the same search response. The skill must tag each result with its evidence class so downstream agents don't conflate them:

Class	How to detect	Output tag
Fact / news reporting	`signature` matches `鉅亨網新聞中心`, `鉅亨網編譯…`, `路透社`, `中央社`, `MoneyDJ` etc. and `category.slug ∈ {headline, tw_stock, wd_stock, us_stock, cn_stock, tech, anue_live, tw_premarket}`	`evidence_class: "fact"`
Opinion / analyst commentary	`signature` matches 點股成金 / 摩爾投顧 / 江國中 / 優分析 / 一手情報, or `category.slug ∈ {mag, celebrity_area}`	`evidence_class: "opinion"` — must NOT be cited as a factual claim
Regulatory announcement / boilerplate / sponsored	Title starts with `〈台股公告〉` / `公告：` / `盤中速報`; or `signature` matches 金色財經 / 鏈文 reprints (`categoryId: 894`)	`evidence_class: "announcement"`

Every JSON output field that is not directly present in the API/HTML response must carry a confidence qualifier: fact (verbatim from Cnyes), inferred (derived — e.g. sentiment, market-impact label, related-tickers regex), or unknown (couldn't determine). Sentiment / market-impact labels are always inferred.

Investment-advice disclaimer: never recommend buying, selling, holding, or shorting any asset. Limit output to "what Cnyes reported / what analysts in opinion pieces claimed" — direct quotation form. If the user asks "should I buy 2330?", respond with a Traditional-Chinese refusal: 「本工具僅整理鉅亨網新聞與分析內容，無法提供投資建議。請洽合格理財顧問。」 and still return any factual / opinion citations the user might find useful.

Site-Specific Gotchas

limit is silently ignored on /search/news — the response always paginates 20 items at a time regardless of limit=. Use page= to walk the result set. On /newslist/category/* the limit parameter IS honoured (verified per_page=3 with limit=3).
<mark>...</mark> highlight tags in title and content — the search API wraps the matched query terms with HTML <mark> tags. Strip them before display: re.sub(r'</?mark>', '', s).
content field is HTML-entity-encoded for some categories — e.g. category-list responses for tw_stock return <p></p><ul><li>… etc. Run an HTML-entity unescape (html.unescape in Python, etc.) before parsing. Search-API responses are usually plain text but may contain encoded characters in long snippets.
publishAt is unix-epoch SECONDS, UTC — not milliseconds. Format for display in Taipei time (UTC+8, no DST). The on-site human-readable timestamp is always Taipei-local.
category[] is OPTIONAL — some items (raw 公告, 鏈文 reprints from 金色財經, naked 盤中速報) ship without any category. Treat as category_slug = null and use signature + title-prefix heuristics to classify.
There is NO public article-detail JSON endpoint — verified during discovery: GET /media/api/v1/news/{id} returns {"items":[],"message":"參數無效","statusCode":422}; /news/?newsId={id}, /newslist?newsId={id}, /news/list, /news/detail/{id} all return 404. Don't waste turns probing for one — fetch the SSR'd HTML at https://news.cnyes.com/news/id/{newsId} instead. The HTML is fully server-rendered and parseable without a browser session.
api.cnyes.com requires network egress that resolves the cnyes domain — in a fresh Vercel sandbox without DNS to public DNS servers, plain curl can fail with "Could not resolve host". Always use browse cloud fetch <url> (or browse cloud fetch <url> --proxies if a corporate edge blocks the bare path). The fetch path resolves through Browserbase's outbound DNS.
Search ranking is recency-weighted but NOT strict — for a popular keyword (AI → 50k+ hits, 台積電 → 2.8k hits), the first page mixes the most-recent ~5 articles with older "hot" content. Always sort the returned items by publishAt descending if "latest N articles" is the user intent.
Simplified-Chinese query auto-maps to TC results — 台积电 returns the same set as 台積電. Useful for cross-region inbound traffic, but means a user query in mixed SC/TC still works.
Ticker references inside article body: TW stocks use the format (NNNN-TW) or (NNNNN-TW) (4 or 5 digits + suffix), US stocks (TICKER-US). ETFs use the same 4-5-digit + -TW pattern. However, plain text like (2330) (no suffix) also appears in some 公告 articles — these are less reliable. Always prefer the <a href=""> inline-ticker-link extraction over regex over prose.
Opinion vs news boundary is real and observable — Cnyes hosts content from 投顧 (investment-advisory firms: 摩爾投顧, 凱基投顧, etc.) under the same article infrastructure as its newsroom. The signature field is the reliable boundary marker. Mis-categorising 摩爾投顧's bullish/bearish stock pick as a Cnyes news factual claim is a silent integrity failure — the honesty layer in Step 7 prevents this.
The /search?q= route on news.cnyes.com (vs www.cnyes.com/search/news?keyword=) is the same data, different UI. Both render server-side. The news.cnyes.com/search page is richer (category badges, tag chips, longer snippets); the www.cnyes.com/search/news page is compact. For browser-fallback extraction prefer news.cnyes.com/search?q=.
news.cnyes.com/news/{id} (no /id/) redirects to /news/id/{id} — both URLs end at the same canonical page. Don't try to "normalise" the URL by stripping /id/; the inverse direction also works (Cnyes' homepage links use both forms inconsistently).
Cookie-consent overlay on first visit ("為優化網站服務，鉅亨網使用Cookie…"). It does not block markdown extraction or fetch — but if you do drive a browser session with screenshots in mind, click button: 繼續使用 to clear it before the user-visible screenshot.
Subscription-gated content (/anuestore, 鉅亨買 768, premium 雜誌 issues) — articles in the mag category or under https://www.cnyes.com/anuestore may show a paywall card on the HTML page. The search API still returns title + snippet for these, but the full body in the HTML page is truncated. Mark these as paywalled: true in output.

Expected Output

The skill emits one JSON envelope per query. All evidence_class and confidence qualifiers required by the honesty layer (Step 7) are present.

{
  "query": "台積電",
  "method_used": "api",
  "endpoint": "https://api.cnyes.com/media/api/v1/search/news?q=%E5%8F%B0%E7%A9%8D%E9%9B%BB&page=1",
  "fetched_at_utc": "2026-05-22T02:15:16Z",
  "fetched_at_taipei": "2026-05-22 10:15 (Asia/Taipei, UTC+8)",
  "total_hits": 2817,
  "returned": 5,
  "page": 1,
  "last_page": 141,
  "results": [
    {
      "newsId": 6466886,
      "url": "https://news.cnyes.com/news/id/6466886",
      "title": "算力軍備到來！馬斯克砸千億美元建TeraFab ASML訂單排到2029年",
      "published_at_utc": "2026-05-22T01:30:06Z",
      "published_at_taipei": "2026-05-22 09:30",
      "byline": "鉅亨網新聞中心",
      "category_name": "美股雷達",
      "category_slug": "us_stock",
      "evidence_class": "fact",
      "snippet": "EUV 訂單，已排給台積電 (TSM-US)(2330-TW)、三星電子…",
      "keywords": ["ASML", "TeraFab", "特斯拉", "晶片", "算力"],
      "related_tickers": [
        {"ticker": "TSM",  "market": "US", "confidence": "fact"},
        {"ticker": "2330", "market": "TW", "confidence": "fact"},
        {"ticker": "ASML", "market": "US", "confidence": "fact"},
        {"ticker": "TSLA", "market": "US", "confidence": "fact"}
      ],
      "cover_image": "https://cimg.cnyes.cool/prod/news/6466886/l/e4b6c92c6404bac723b084d9e14d3067.jpg",
      "paywalled": false,
      "summary_zh_tw": "ASML 執行長證實已與馬斯克就 TeraFab 半導體計畫直接溝通；該項目由特斯拉、SpaceX、xAI 共同發起，落腳德州，初始投資 550 億美元、遠期最高 1,190 億美元，鎖定 2 奈米以下製程。ASML EUV 訂單已排至 2029 年，由台積電、三星、Intel 共享。",
      "summary_confidence": "fact",
      "market_impact": {
        "direction": "positive_for_ASML_TSM_supply_chain",
        "confidence": "inferred",
        "rationale": "Article reports a long-dated demand commitment from a new high-spending fab buyer; sentiment label is derived, not stated."
      }
    }
  ],
  "disclaimer_zh_tw": "本資料僅整理鉅亨網公開新聞與分析內容，不構成投資建議。市場影響欄位為自動推論，可能有誤；請以原始報導為準並洽合格理財顧問。"
}

Outcome shape — keyword has no hits

{
  "query": "this-is-definitely-not-a-real-keyword-zzz",
  "method_used": "api",
  "fetched_at_utc": "2026-05-22T02:15:16Z",
  "total_hits": 0,
  "returned": 0,
  "results": [],
  "fallback_suggested": "site:cnyes.com OR site:news.cnyes.com \"<keyword>\""
}

Outcome shape — single-article dereference (step 4 only)

{
  "newsId": 6466886,
  "url": "https://news.cnyes.com/news/id/6466886",
  "title": "算力軍備到來！馬斯克砸千億美元建TeraFab ASML訂單排到2029年",
  "published_at_taipei": "2026-05-22 09:30",
  "byline": "鉅亨網新聞中心",
  "category_breadcrumb": ["美股", "美股雷達"],
  "category_slug": "us_stock",
  "body_zh_tw": "在比利時安特衛普 5 月 15 日一場半導體活動上，ASML(ASML-US) 執行長 Christophe Fouquet 向《路透社》確認了一件事…(full Traditional-Chinese body)",
  "body_word_count_zh": 1834,
  "related_tickers_inline": [
    {"ticker": "ASML", "market": "US", "first_offset_chars": 47, "confidence": "fact"},
    {"ticker": "TSLA", "market": "US", "first_offset_chars": 312, "confidence": "fact"},
    {"ticker": "TSM",  "market": "US", "first_offset_chars": null, "confidence": "fact"},
    {"ticker": "2330", "market": "TW", "first_offset_chars": null, "confidence": "fact"}
  ],
  "tags": ["ASML", "TeraFab", "特斯拉", "晶片", "算力"],
  "evidence_class": "fact",
  "key_facts": [
    {"claim": "ASML 執行長 Christophe Fouquet 已與馬斯克就 TeraFab 進行直接溝通", "confidence": "fact"},
    {"claim": "SpaceX 提交的德州 Grimes 廠區備案文件揭露初始投資 550 億美元、遠期最高 1,190 億美元", "confidence": "fact"},
    {"claim": "目標製程為 2 奈米及以下", "confidence": "fact"}
  ],
  "key_inferences": [
    {"claim": "對台積電、三星、Intel 等已下單 EUV 客戶為中性偏正面（新訂單抬升 ASML 訂單能見度至 2029）", "confidence": "inferred"}
  ],
  "paywalled": false,
  "disclaimer_zh_tw": "本資料僅整理鉅亨網公開新聞內容，不構成投資建議。"
}

Outcome shape — paywalled / truncated article

{
  "newsId": 6466346,
  "url": "https://news.cnyes.com/news/id/6466346",
  "title": "520行情vs.最牛台股",
  "category_slug": "mag",
  "evidence_class": "opinion",
  "paywalled": true,
  "snippet_available": true,
  "body_available": false,
  "snippet_zh_tw": "(API snippet only — article body requires Cnyes 雜誌 訂閱 subscription)",
  "next_step": "Surface title + snippet + URL to user; do NOT attempt to bypass paywall.",
  "disclaimer_zh_tw": "此為訂閱內容，僅能提供標題與摘要。"
}