dsh-deep-search
采用来源分级（A/B/C/D）和错误信息启发式判断的多轮迭代搜索，适用于 DeepSeek Harness——6 个免费引擎 + 14 个面向平台的检索器、研究状态、verify_claims，以及注入式 SKILL 协议。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:Xiao-wuwu/dsh-deep-search#310fcc75ce34744b7688d249ba506586d5ab5035Deep Search 以 dsh-deep-search 发布，当前版本为 0.1.6。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
让你的 DSH 不再一次搜索就下结论。 多轮迭代深度检索 · 14 平台定向 · 来源可信度分级 (A/B/C/D) · 交叉验证 · 虚假信息检测
一次 web_search 就回答"XX 发布了没有"？模型把聚合站的标题党当成事实、把"没搜到"说成"不存在"？
dsh-deep-search 把「查询规划 → 并行检索 → 解析 → 重写 → 验证分级」这套多轮迭代协议变成 DSH 会话的默认行为。智能由你的会话模型驱动，苦力由插件承担——去重、聚类、可信度分级、虚假信号检测全部自动完成。全程零强制付费依赖：不配任何 API Key 也能跑满核心平台，配额耗尽的引擎自动退避到链尾并自愈。
English README: README.en.md
dsh plugin --profile desktop add dsh-deep-search
# 重启 DSH（web profile 关闭了 HMR）
装好后直接问：「Qwen4-Max 到底发布了吗？」——会话模型会自动加载 deep-research 协议，多轮迭代直到拿到官方 A 级证据或如实报告"未检索到"。
| 说明 | |
|---|---|
| 🔁 迭代协议默认化 | 系统提示注入 + SKILL：时效/真伪问题自动走多轮（≤5 轮），无需用户催 |
| 🌐 14 平台定向检索 | GitHub · HuggingFace · ModelScope · 知乎 · Reddit · HN · StackOverflow · Wikipedia · V2EX · npm · B站 · 微博 · X(尽力而为) · + 免Key通用引擎(Bing/DDG/SearXNG/Mojeek/Firecrawl) |
| 🧮 可信度分级 A/B/C/D | 官方公告 > 官方互动 > 社区(需交叉验证) > 媒体/未知。数据驱动规则表可扩 |
| 🕵️ 虚假信息信号 | 内容农场 / 标题党 / 讽刺源 / 孤证 / 日期矛盾 / 旧闻 自动打标 |
| 🔗 交叉验证矩阵 | 同事件多源聚类 + 独立域名计数，杜绝"3 家转载当 3 个信源" |
| 🚫 诚实降级 | X/微博/B站 反爬失败 → gaps 明示 unavailable/degraded，web 替代绝不冒充 X 覆盖 |
| 📄 "未检索到"纪律 | 找不到时输出：覆盖范围 + 现有线索分级 + "这不等于不存在" |
| 💸 零强制付费 | Keyed 引擎(Tavily/Serper/博查/Exa/Firecrawl-Key)有免费额度，配置了才启用，缺失自动禁用并在 doctor 说明 |
// 主力：并行检索 + 跨轮去重 + 自动分级
deep_search({
question: "Qwen4-Max 发布了吗",
queries: [
"Qwen4-Max released open weights",
{ q: "\"Qwen4-Max\"", target: "huggingface", timeRange: "week" },
{ q: "from:Alibaba_Qwen Qwen4", target: "x" }
],
researchId: "rs-xxxx" // 续轮，自动只回新增
})
// → { results[trustRank/flags/extracted], clusters, gaps, officialConfirmation, hint, … }
// 交叉验证矩阵 + 轮次历史，写最终答案前先读
research_state({ researchId: "rs-xxxx", action: "summary" })
// 断言 ↔ 已收集证据核对（不再联网），给裁决建议
verify_claims({ researchId: "rs-xxxx", claims: [{ text: "Qwen4-Max 2026-10-10 开源" }] })
// 零配额自检：各引擎/平台就绪、缺哪些可选凭据、谁在冷却
deep_search({ doctor: true })
你(规划器) ──生成 10-20 组查询──▶ deep_search(执行器：并行/去重/聚类/分级/预算)
▲ │
└──解析+重写：读 extracted 实体 + hint◀────┘
│ 找到 A 级公告 | 轮次尽 | 无新词
▼
research_state(交叉验证矩阵) ──▶ 你(验证器)生成分级答案
| 插件 | 定位 |
|---|---|
| dsh-deep-search | 多轮迭代协议 + 平台定向 + A/B/C/D 分级 + 虚假信息检测；不替代内置 web_search，而是给它装上"查证"的肌肉记忆 |
| dsh-web-search-pro | 重型检索基础设施（缓存/证据包/浏览器渲染），单次检索很强，迭代协议需自配 |
| dsh-free-search | 19 引擎免Key兜底 + 图片墙，广度好，单轮 |
| dsh-zhihu-search | 知乎官方 API 单平台专家 |
| modsearch | 免Key网页/X 桥接（X 需本机 Grok），单次检索 |
它们共存不冲突：本插件自带引擎层，不 override ctx.web seam，工具名独立。
全部字段可经 DSH 设置页 → 插件 → dsh-deep-search 修改；API Key 一律走 DSH 凭据存储，配置里只存引用名：
{
"enabledEngines": ["bing", "ddg", "searxng", "mojeek", "firecrawl-keyless", "tavily"],
"enabledPlatforms": ["github", "huggingface", "modelscope", "reddit", "hackernews", "zhihu", "x"],
"proxy": "", "proxyByTarget": { "x": "http://127.0.0.1:7890" },
"concurrency": 8, "maxRounds": 5, "budgetChars": 9000,
"promptInjection": true
}
可选凭据（缺了不影响默认流程）：TAVILY_API_KEY · SERPER_API_KEY · BOCHA_API_KEY · EXA_API_KEY · FIRECRAWL_API_KEY · ZHIHU_ACCESS_SECRET（知乎升稳定档）· ROLLI_API_TOKEN（X 升稳定档）。
pnpm install
pnpm run typecheck && pnpm test # 98 项测试，无需网络
pnpm run build # tsc 类型检查 + tsdown 打包 lib/
MIT
这个项目从设计到测试都是一次次「被一次搜索坑了」的真实痛点堆出来的。如果它让你的 DSH 回答变得敢引用、可溯源：
给个 Star 就是最大的支持 —— 它让更多人（和未来的你）能找到它。发现问题或想要更多平台适配，欢迎开 Issue / PR。
如果 dsh-deep-search 帮你的 AI 回答做到了可溯源，欢迎 在 GitHub Sponsors 打赏 —— 一杯咖啡的钱，就是独立开发者持续维护的最大动力。也欢迎 PR 与 Issue。
English version: jump to English
Stop letting your DSH agent conclude from a single search. Multi-round iterative deep research · 14 platform-targeted retrievers · A/B/C/D source grading · cross-validation · misinformation heuristics.
A one-shot web_search answering "has model X shipped?" — with an aggregator's clickbait headline — is how hallucinated facts get into your context.
dsh-deep-search makes the plan → parallel-retrieve → parse → rewrite → verify loop the default behavior of a DSH session. The intelligence is your session model; the grunt work (dedup, clustering, trust grading, rumor flags) is the plugin. Zero forced payments: no API key needed for the core platforms; quota-exhausted engines fall to the back of the chain and self-heal.
中文 README：README.md
dsh plugin --profile desktop add dsh-deep-search
# restart DSH (the web profile disables HMR)
Then just ask: "Has Qwen4-Max actually been released?" — the model auto-loads the deep-research protocol and iterates until it finds A-grade official evidence or honestly reports not found.
| 🔁 Default iteration | System-prompt injection + SKILL: time-sensitive / contested questions iterate up to 5 rounds without being asked |
| 🌐 14 platforms | GitHub · HuggingFace · ModelScope · Zhihu · Reddit · HackerNews · StackOverflow · Wikipedia · V2EX · npm · bilibili · Weibo · X (best-effort) · free general engines (Bing/DDG/SearXNG/Mojeek/Firecrawl) |
| 🧮 Trust grades | A official announcement → B official interaction → C community (needs cross-validation) → D media/unknown. Data-driven, user-extensible |
| 🕵️ Rumor detection | content-farm / clickbait / satire / single-source / date-conflict / stale flags, automatic |
| 🔗 Cross-validation matrix | Same-event clustering + independent-domain counting kills "3 reprints ≠ 3 sources" |
| 🚫 Honest degradation | X/Weibo/bilibili anti-bot failures surface as gaps with unavailable/degraded — web substitutes are never presented as X coverage |
| 📄 "Not found" discipline | When nothing confirms: coverage + graded clues + explicitly "this does not mean it did not happen" |
| 💸 No paid walls | Optional keyed engines (Tavily/Serper/Bocha/Exa/Firecrawl) activate only when credentials resolve; free channels cover the default flow |
deep_search({ question, queries: [ "…", { q, target, timeRange } ], researchId, budgetChars })
// → results[trustRank, flags, extracted] · clusters · gaps · officialConfirmation · hint · stats
deep_search({ doctor: true }) // zero-quota readiness of every channel
research_state({ researchId, action }) // summary / close / reset — the citation matrix
verify_claims({ researchId, claims }) // evidence match + rule-suggested verdict
Four roles — the model plans/parses/rewrites/adjudicates; the plugin executes, dedups, clusters, grades, budgets, and reports honestly. See SKILL.md for the full protocol.
dsh-deep-search coexists with all of them: it never overrides the ctx.web seam.
pnpm install && pnpm run typecheck && pnpm test # 98 offline tests
pnpm run build # tsdown → lib/
MIT
Every feature here came from getting burned by a lazy one-shot answer. If it makes your agent citable and traceable, a star is the cheapest way to say so — and the best way to help the next person find it. Issues and PRs welcome.
If dsh-deep-search makes your agent's answers citable and traceable, consider sponsoring on GitHub — a coffee goes a long way for an indie maintainer. PRs and issues welcome.