Single Google search -> title/url/snippet per result. Results are cached 24h, so repeating a query is free -- prefer re-querying over caching results yourself. For latest/today/breaking queries set SURF_CACHE_TTL_SEARCH_MS=0 to bypass the cache. Default limit 10 (max 20). First call ~4s (Chromium warmup), then ~2s. On CAPTCHA a visible Chrome opens for a human to solve (shared-IP protection); SURF_CLOUD_MODE=true makes it fail-fast instead.

Run 2-10 Google searches concurrently. Use to compare multiple angles in one call. Each query counts against the internal rate limit (~10/min) -- do not loop this for bulk scraping. First call adds 5-10s pool warmup. Per-query failures are isolated in the results array. Disabled in cloud mode.

Fetch one public URL -> clean article text. HTML via Mozilla Readability; academic PDFs (arxiv/biorxiv/Nature/OpenReview/NeurIPS/JMLR/PMLR/Springer/PubMed-via-PMC) auto-detected via Content-Type, %PDF magic, citation_pdf_url meta, and per-domain URL rules. Tiered depth: mode="abstract" returns ~1500 chars (PDF page 1 or HTML meta description) -- cheap survey to triage relevance before paying for full body. mode="full" (default) returns the whole article. Best-effort: failures return an errorInfo instead of throwing.

One-shot Google search + parallel extract of the top results. Default mode="abstract" returns SERP enriched with ~1500-char abstracts per result -- a cheap survey of what the top results actually contain, far fewer tokens than fetching all bodies. Switch to mode="full" only when you need the actual article texts (slower, much more tokens). Per-page extract failures are isolated. Disabled in cloud mode.

MCP server status: cascade mode + transitions, rate-limiter usage, cache size, config. Call this if searches start failing or returning empty -- check cascade.totalCaptchas and rateLimiter.queueSize, and reduce search volume if they are high.