dsh-web-search-scrape

Since DSH 0.1.5, the model-facing web_search tool was moved to the agent preset plane, which shadow-plugin registrations at the profile level—the tier parameter never arrives and results are clipped to 8 entries, making deep research nearly impossible. The official search path, meanwhile, bills per token, an unnecessary cost for routine queries. dsh-web-search-scrape takes a direct approach: Node's built-in fetch scrapes public result pages from search engines (DuckDuckGo, Bing, Baidu, Google, Yandex) and social platforms (WeChat, Bilibili, Weibo, X, Zhihu, Douyin, Reddit), bypassing any API key entirely—scraping is always free. Six depth tiers range from T1 Ultra-fast (~4 s, 1 engine, 6 results) to T6 Research (~280 s, all engines plus seven social sites, 50+80 results, sub-agent multi-pass reading). The model can set tier explicitly; when omitted, an auto heuristic picks a tier, capped by maxAutoTier so common keywords do not silently escalate to T5 and incur model charges. Tiers T4 and above dispatch an LLM sub-agent for triage; T5/T6 can additionally open pages via curl to verify upstream sources and emit an evidence level. The plugin also registers a distinct web_search_deep tool that is never shadowed by presets, so tier and result ceiling (default 130) always follow this plugin's configuration. It suits users who need everyday source retrieval, comparative analysis, or trend tracking in DSH without per-query billing; research workflows that demand T5/T6 deep dives with multi-source cross-validation; and developers building or debugging DSH plugins who want to understand provider registration and preset shadowing mechanics.