@streamapp/stream 0.0.1

package/resources/packages/manmanbuy/manmanbuy-search.recipe.json

@@ -0,0 +1,115 @@
+{
+  "version": 1,
+  "kind": "html",
+  "sourceId": "manmanbuy-search",
+  "_why_discount_tab": "慢慢买有两个搜索 tab：`c=discount`（优惠流，SSR 直出，免签名免登录）和 `c=search`（商品比价，客户端渲染、走 prom2Extra 签名 XHR）。比价 tab 要慢慢买登录态、否则 7700『需登录验证』，是浏览器 recipe（贵一级）。优惠流 tab 每一行本身就是『某平台对某商品的报价』，搜一个商品名即得多平台价格——正是比价意图，且是 Tier-2 裸 HTTP，所以走这条。要精确到同一 SKU 分组再考虑升级比价 tab。",
+  "request": {
+    "url": "https://s.manmanbuy.com/pc/search/result?keyword={keyword}&c=discount&pageId={pageId}",
+    "method": "GET"
+  },
+  "list": {
+    "_why_class_prefix": "class 名带 CSS-module 哈希后缀（DiscountItemPC_box__m9G3M），站点一次重构就换，所以一律用 `[class*=\"前缀__\"]` 属性包含选择器钉住语义前缀，不写死哈希。",
+    "selector": "[class*=\"DiscountItemPC_box__\"]",
+    "limit": 40,
+    "fields": {
+      "title": {
+        "selector": "[class*=\"DiscountItemPC_itemTitle__\"] a",
+        "text": true
+      },
+      "link": {
+        "selector": "[class*=\"DiscountItemPC_itemTitle__\"] a",
+        "attr": "href",
+        "resolve": true
+      },
+      "guid": {
+        "selector": "[class*=\"DiscountItemPC_itemTitle__\"] a",
+        "attr": "href",
+        "resolve": true
+      },
+      "description": {
+        "_why": "itemSubTitle 是价格行（含促销与国补口径），如 `7619.02元+179.98元淘金币(含国补)`。这是这条比价条目的核心值，落成正文。",
+        "selector": "[class*=\"DiscountItemPC_itemSubTitle__\"] a",
+        "text": true
+      },
+      "author": {
+        "_why": "平台/店铺名（京东自营 / 天猫旗舰店 / 淘宝 …）当作这条报价的来源署名——多行同一商品不同平台，author 一栏就把『谁家的价』摆出来了。",
+        "selector": "[class*=\"DiscountItemPC_itemMall__\"]",
+        "text": true
+      },
+      "image": {
+        "_why": "顶层 image 字段：defaultNormalizer 读的是 raw.image（不是 cover），配价格文本渲染成 gallery 卡（商品图 + 价格）。",
+        "selector": "[class*=\"DiscountItemPC_itemCover__\"] img",
+        "attr": "src",
+        "resolve": true
+      }
+    }
+  },
+  "pagination": {
+    "_why": "优惠流按 pageId 翻页（URL 里 {pageId} 占位由 runner 每页替换，不是自动追加）。maxPages:2 = 约 70 条，够一次比价搜索；两页间新增的优惠会让少量条目重叠，按 guid=优惠链接 dedup 兜掉。",
+    "mode": "increment",
+    "param": "pageId",
+    "start": 1,
+    "step": 1,
+    "maxPages": 2
+  },
+  "assert": [
+    {
+      "selector": "[class*=\"DiscountItemPC_box__\"]",
+      "desc": "慢慢买优惠搜索列表行存在"
+    }
+  ],
+  "meta": {
+    "normalizer": "rsshub",
+    "type": "post",
+    "description": "慢慢买 · 全网比价关键词搜索 —— 搜一个商品名，返回各电商平台（京东/天猫/淘宝/拼多多…）对它的报价与促销，一行一个平台报价",
+    "topics": [
+      "manmanbuy",
+      "慢慢买",
+      "比价",
+      "价格",
+      "优惠",
+      "促销",
+      "电商",
+      "购物",
+      "shopping",
+      "price",
+      "搜索",
+      "search"
+    ],
+    "categories": [
+      "shopping"
+    ],
+    "facility": {
+      "key": "manmanbuy",
+      "label": "慢慢买"
+    },
+    "capabilities": [
+      "search"
+    ],
+    "provides": [
+      "search-price"
+    ],
+    "key_param": "keyword",
+    "priority": 80,
+    "discoverable": true,
+    "cadence_hint_seconds": 3600,
+    "example_queries": [
+      "慢慢买搜 iPhone 17 Pro 256G",
+      "比价 索尼 WH-1000XM5"
+    ],
+    "auth": {
+      "type": "none"
+    },
+    "homepage": "manmanbuy.com",
+    "params_schema": {
+      "keyword": {
+        "type": "string",
+        "required": true,
+        "description": "商品搜索关键词（越具体越能对上同一商品，如带型号/容量/颜色）"
+      }
+    },
+    "radar": [
+      "s.manmanbuy.com/pc/search/result"
+    ]
+  }
+}

package/resources/packages/manmanbuy/package.json

@@ -0,0 +1,15 @@
+{
+  "name": "@streamapp/manmanbuy",
+  "version": "1.0.0",
+  "author": "stream",
+  "keywords": [
+    "stream-recipe"
+  ],
+  "stream": {
+    "type": "recipe",
+    "name": "慢慢买",
+    "tagline": "全网比价 · 关键词搜索",
+    "facility": "manmanbuy",
+    "schemaVersion": 1
+  }
+}

package/resources/packages/mineru/README.md

@@ -0,0 +1,38 @@
+# MinerU
+
+Optional document/image → markdown backend (no credentials, no source).
+
+A service-only plugin: it contributes a backend container (reached through the gateway at
+`/_p/mineru`, or a remote relay when `mineru_url` is configured) but registers no source
+adapter/normalizer. Self-built mirror of MinerU behind a thin FastAPI wrapper.
+
+Local container = "free but slow" (CPU default, GPU when present); cloud relay (`mineru_url`) = the
+paid fast lane. Both tiers run the SAME MinerU — cloud sells throughput, not a better model
+(frontier VLMs are *less* faithful on documents than MinerU's pipeline). Output quality on the
+2026-07-30 run (see `backend.gpu` below for the setup): Chinese text accurate, tables reconstructed
+as HTML, nothing invented.
+
+## backend
+
+`gpu: true` —— MinerU's deterministic-first pipeline (layout → table-structure → formula → reading
+order) runs best on a GPU; born-digital PDFs extract the text layer directly (no VLM rewrite → exact
+numbers), the VLM only OCRs scanned/image regions.
+
+Measured 2026-07-30 (RTX 4080 16GB, host 46GB RAM; input = one screenshot dense with Chinese text,
+via `POST /_p/mineru/parse`). These are readings from that setup, not universal constants —
+VRAM/RSS scale with input size and model backend:
+
+- image on disk 9.64GB; model cache volume `stream_mineru-cache` 1.2GB
+- idle RSS 335MiB (models load lazily — nothing resident until a request arrives)
+- one-image parse: peak RSS 3.3GB (under a third of the `mem: 10G` cap)
+- VRAM delta ~1.6GB (whole card 7066 → 8652 MiB)
+
+With standby it costs 0 while unused.
+
+- `env.MINERU_DEVICE` —— MinerU device autodetect; pin a backend/model here if needed.
+- `volumes` —— persist the model + layout/table/formula weights so they don't re-download on restart.
+- `mem: 10G` —— cap host RAM (model lives in VRAM; this bounds the python/pdf/image side). swap
+  disabled (memswap = mem) so a runaway OOM-kills fast rather than host swap-thrash.
+- `standby.startTimeoutSeconds: 120` —— MinerU's model load (layout/table/formula weights) is heavier
+  than the other three backends' cold start; give it more runway before the standby-manager gives up
+  waiting for health.

package/resources/packages/mineru/package.json

@@ -0,0 +1,28 @@
+{
+  "name": "@streamapp/mineru",
+  "version": "1.0.0",
+  "private": true,
+  "type": "module",
+  "stream": {
+    "id": "mineru",
+    "name": "MinerU",
+    "tagline": "文档 / 图片 → Markdown",
+    "description": "把 PDF、扫描件、图片转成 Markdown：版面还原、表格转成 HTML、公式转成 LaTeX。取正文那条链路上负责「这份东西不是文字，是一张纸」的那一档。本机容器免费但慢（吃 GPU 更快），配了 mineru_url 就走云端快车道——两边跑的是同一个模型，云端卖的是速度不是效果。",
+    "backend": {
+      "image": "ghcr.io/jaggerh/mineru-server:latest",
+      "service": "mineru",
+      "port": 80,
+      "health": "/health",
+      "gpu": true,
+      "env": {
+        "MINERU_DEVICE": "auto"
+      },
+      "volumes": ["mineru-cache:/root/.cache"],
+      "mem": "10G",
+      "standby": {
+        "idleMinutes": 10,
+        "startTimeoutSeconds": 120
+      }
+    }
+  }
+}

package/resources/packages/pansou/README.md

@@ -0,0 +1,14 @@
+# PanSou
+
+Search-only netdisk resource search (no credentials, no login).
+
+## backend
+
+- `env.CHANNELS` —— Default Telegram channel pool: the TG channels pansou searches when no
+  per-request `channels` param is given. Curated from the upstream reference config. Add/remove
+  channels here; they take effect on next `docker compose up -d` (compose regenerated from this
+  descriptor).
+- `env.ENABLED_PLUGINS` —— PanSou built-in indexer plugins to enable (non-TG web scrapers). Each is
+  a Go plugin in the pansou repo (`plugin/<name>/`).
+- `standby.idleMinutes: 30` —— 冷查询型：调用零散、单次搜索几秒内完事，没有播放/挂载那种"随时要来"
+  的时效压力——30 分钟闲置够覆盖同一会话内的连续几次搜索，又不会长期占着容器空转。

package/resources/packages/pansou/activate.ts

@@ -0,0 +1,11 @@
+import type { ActivateFn } from '../../src/packages/activate.ts'
+import { PansouAdapter } from './adapter.ts'
+import { pansouNormalizer } from '../../src/content/pansou.ts'
+
+/** 这个包贡献的东西：一个 adapter + 一个 normalizer。
+ *  `config.url` 是宿主解析好的显式覆盖（未配则 undefined，adapter 自己惰性回落到
+ *  PANSOU_URL env / plugin target）。 */
+export const activate: ActivateFn = (ctx) => ({
+  adapters: { pansou: new PansouAdapter(ctx.config.url as string | undefined) },
+  normalizers: { pansou: pansouNormalizer },
+})

package/resources/packages/pansou/adapter.ts

@@ -0,0 +1,94 @@
+import type { Adapter } from '../../src/adapters/types.ts'
+import type { SourceManifest } from '../../src/manifest/types.ts'
+import { pluginTarget } from '../../src/plugins/plugin-target.ts'
+import { withAwake } from '../../src/plugins/standby/hook.ts'
+
+/** server-side fetch base (NOT the client-facing gateway path): explicit ctor arg → PANSOU_URL
+ *  env (dev override) → the bootstrap-injected plugin target (compose mode: container DNS;
+ *  none: unset → ''). */
+export function resolvePansouUrl(explicit?: string): string {
+  return explicit ?? process.env.PANSOU_URL ?? pluginTarget('pansou') ?? ''
+}
+
+/** netdisk types we render well — the default `cloud_types` filter. */
+const DEFAULT_CLOUD_TYPES = ['baidu', 'quark', 'aliyun']
+/** bias toward complete resources: drop obvious non-full junk (overridable). */
+const DEFAULT_FILTER = { exclude: ['预告', '花絮', '试看', '样片', '片段', '主题曲'] }
+
+/** a param that may arrive as string[] (bootstrap) or a comma string (HTTP query). */
+function toArray(v: unknown): string[] | undefined {
+  if (Array.isArray(v)) return v.map(String).filter(Boolean)
+  if (typeof v === 'string' && v.trim()) return v.split(',').map((s) => s.trim()).filter(Boolean)
+  return undefined
+}
+/** a param that may arrive as an object (bootstrap) or a JSON string (HTTP query). */
+function toObject(v: unknown): Record<string, unknown> | undefined {
+  if (v && typeof v === 'object') return v as Record<string, unknown>
+  if (typeof v === 'string' && v.trim()) {
+    try {
+      return JSON.parse(v) as Record<string, unknown>
+    } catch {
+      return undefined
+    }
+  }
+  return undefined
+}
+
+/**
+ * PanSou adapter — a thin HTTP client over a (Stream-managed or external) PanSou
+ * netdisk search service. Search-only, no credentials. Returns PanSou's `results`
+ * (one SearchResult per source message: title + content + images + a links[] of
+ * netdisk shares); the presenter maps each to a normalized link Content.
+ *
+ * PanSou's search domain is per-request, not baked into the container: `channels`,
+ * `cloud_types`, `filter` (include/exclude), `ext` (plugin params, e.g. title_en)
+ * and `src` are all forwarded when the caller provides them — so the AI/MCP can
+ * customize a search at call time. Defaults bias toward complete resources.
+ */
+export class PansouAdapter implements Adapter {
+  readonly id = 'pansou'
+
+  constructor(private readonly explicitUrl?: string) {}
+
+  /** 惰性:host 档下 origin 是容器醒着时才存在的(standby Cell 缓存),构造期快照必得空串。
+   *  每次求值现解析;fetch 都在 withAwake 回调里,求值时容器已醒。compose 档恒定,无行为差。 */
+  private get baseUrl(): string {
+    return resolvePansouUrl(this.explicitUrl).replace(/\/$/, '')
+  }
+
+  async init(_env: Record<string, string>): Promise<void> {} // no credentials
+
+  async fetch(params: Record<string, unknown>, _manifest: SourceManifest): Promise<unknown[]> {
+    const kw = String(params.keyword ?? '')
+    if (!kw) throw new Error('[pansou] search needs `keyword`')
+    // res=results → the per-message SearchResult[] (title/content/images + links[]).
+    // Caller-supplied knobs win; otherwise default to render-able netdisks + a
+    // junk-excluding filter (lean toward 全集/合集 without nuking unlabeled packs).
+    const body: Record<string, unknown> = {
+      kw,
+      res: 'results',
+      cloud_types: toArray(params.cloud_types) ?? DEFAULT_CLOUD_TYPES,
+      filter: toObject(params.filter) ?? DEFAULT_FILTER,
+    }
+    const channels = toArray(params.channels)
+    if (channels) body.channels = channels
+    const ext = toObject(params.ext)
+    if (ext) body.ext = ext
+    if (params.src) body.src = String(params.src)
+
+    const r = await withAwake('pansou', () =>
+      fetch(`${this.baseUrl}/api/search`, {
+        method: 'POST',
+        headers: { 'content-type': 'application/json' },
+        body: JSON.stringify(body),
+      })
+    )
+    if (!r.ok) throw new Error(`[pansou] HTTP ${r.status}`)
+    // pansou nests the payload under `data`: { code, message, data: { total, results[] } }.
+    // (older builds returned a flat `results` — accept both so a version bump can't silently
+    // empty the source.)
+    const body2 = (await r.json()) as { results?: unknown[]; data?: { results?: unknown[] } }
+    const results = body2.data?.results ?? body2.results
+    return Array.isArray(results) ? results : []
+  }
+}

package/resources/packages/pansou/manifests.yaml

@@ -0,0 +1,50 @@
+- id: pansou-search
+  adapter: pansou
+  normalizer: pansou
+  title: 网盘搜索
+  description: 输入关键词搜索全网网盘（百度/阿里/夸克/天翼/迅雷等）的共享资源链接。支持网盘类型过滤与排除规则。
+  topics: [pansou, 网盘, 资源, 搜索, 百度网盘, 阿里云盘, 夸克网盘, 天翼云盘, netdisk, search, 影视, 剧集]
+  categories: [search, resource]
+  example_queries: [搜网盘资源, 海绵宝宝 网盘, 某电影网盘资源, 某剧全集 夸克]
+  capabilities: [search]
+  auth: { type: none }
+  cadence_hint_seconds: 3600
+  params_schema:
+    keyword:
+      type: string
+      required: true
+      description: 搜索关键词（必填），如电影名、剧名、资源名称。
+    cloud_types:
+      type: string
+      required: false
+      description: >-
+        限定网盘类型，逗号分隔。可选值: baidu, aliyun, quark, tianyi, uc, 115, 123,
+        mobile, xunlei, pikpak, magnet, ed2k。默认 baidu,quark,aliyun。
+    filter:
+      type: string
+      required: false
+      description: >-
+        JSON 格式过滤规则。exclude 排除含指定关键词的结果，如 {"exclude":["预告","花絮"]}。
+        默认排除预告/花絮/试看/样片/片段/主题曲。
+    channels:
+      type: string
+      required: false
+      description: >-
+        指定搜索的 Telegram 频道列表，逗号分隔（如 yunpanx,Aliyun_4K_Movies）。
+        不传则使用容器 CHANNELS 环境变量配置的默认频道池。
+    ext:
+      type: string
+      required: false
+      description: JSON 格式扩展参数，传给 PanSou 插件系统（如 title_en 供插件使用）。
+    src:
+      type: string
+      required: false
+      description: >-
+        限定搜索来源，三选一：all（默认，两者都搜）、tg（只搜 Telegram 频道，走
+        t.me 公开网页预览）、plugin（只搜 ENABLED_PLUGINS 里的网站爬虫）。
+  fan_out:
+    dimension: channels
+    strategy: batch
+    batch_key: [cloud_types, filter, src]
+  provides: [search-download]
+  key_param: keyword

package/resources/packages/pansou/package.json

@@ -0,0 +1,27 @@
+{
+  "name": "@streamapp/pansou",
+  "version": "1.0.0",
+  "private": true,
+  "type": "module",
+  "stream": {
+    "id": "pansou",
+    "name": "PanSou",
+    "tagline": "网盘资源聚合搜索",
+    "description": "Telegram 频道 + 异步插件的网盘资源聚合搜索 API。支持百度网盘、阿里云盘、夸克网盘、 天翼云盘、UC 网盘、115 网盘、123 网盘、移动云盘、迅雷云盘、PikPak、磁力、ed2k 等 多种网盘类型，按类型分类、智能排序、二级缓存。搜索无需登录态。",
+    "repository": "https://github.com/fish2018/pansou",
+    "backend": {
+      "image": "ghcr.io/jaggerh/pansou:latest",
+      "port": 8888,
+      "health": "/api/health",
+      "env": {
+        "CHANNELS": "Aliyun_4K_Movies,BaiduCloudDisk,Oscar_4Kmovies,Quark_Movies, alyp_1,bdbdndn11,bsbdbfjfjff,cloudtianyi,dianyingshare, gotopan,hdhhd21,jdjdn1111,leoziyuan,oneonefivewpfx, peccxinpd,pikpakpan,qixingzhenren,sbsbsnsqq,shareAliyun, taoxgzy,tianyifc,tianyirigeng,tgsearchers3,txtyzy, tyypzhpd,tyysypzypd,ucquark,ucwpzy,wp123zy,xx123pan, ydypzyfx,yggpan,yingshifenxiang123,yp123pan,yunpan139, yunpan189,yunpanuc,yunpanx,yunpanxunlei,zaihuayun,zyfb123",
+        "ENABLED_PLUGINS": "labi,zhizhen,shandian,duoduo,muou"
+      },
+      "standby": {
+        "idleMinutes": 30
+      }
+    },
+    "normalizer": "pansou",
+    "code": { "entry": "./activate.ts", "adapters": ["pansou"], "normalizers": ["pansou"] }
+  }
+}

package/resources/packages/qq/package.json

@@ -0,0 +1,15 @@
+{
+  "name": "@streamapp/qq",
+  "version": "1.0.0",
+  "author": "stream",
+  "keywords": [
+    "stream-recipe"
+  ],
+  "stream": {
+    "type": "recipe",
+    "name": "QQ",
+    "tagline": "给指定联系人发一条 QQ 消息",
+    "facility": "qq",
+    "schemaVersion": 1
+  }
+}

package/resources/packages/qq/qq-send.recipe.json

@@ -0,0 +1,104 @@
+{
+  "version": 1,
+  "kind": "desktop",
+  "sourceId": "qq-send",
+  "app": {
+    "process": "QQ.exe",
+    "title": "QQ"
+  },
+  "_why_title": "聊天窗口可以被拖出来独立成窗，那时 QQ.exe 会有多个窗口，不限定就会被拒绝（枚举里列候选让人选）。主窗口的标题恒为 `QQ`，包含匹配够用。",
+  "_a11y_activation": "QQ 是 Electron，**a11y 树是懒建的**：窗口不在前台时整棵树是空的（find 一律返回 0 条），前台之后才逐步长出来。回执里那面 `unbuilt` 旗只说明「这次读不可采信」，不会替你把窗口提上来——focus 仍然是这条 recipe 的第一步。所以第一步的 focus 不只是为了打字——它是这条 recipe 能看见任何东西的前提。实测 2026-08-27：后台时 role=Edit / role=Button / nameContains=搜索 全部空手而归，同一个 agent 同一时刻读 explorer.exe 的任务栏拿到 21 个控件（证明 agent 没问题，是 QQ 不给）。",
+  "steps": [
+    {
+      "_why": "复位。上一轮跑完会把界面留在「搜索模式 + 上个联系人的搜索词」里，而 `type` 是往输入框里**追加**、不是替换——不复位，第二轮的搜索词会变成「上个联系人本次联系人」，然后每一步都'成功'地跑在错的界面上。点左栏「消息」会同时清空搜索框并回到会话列表（实测）。首次运行时它本来就在这个状态，点一下无副作用，所以不标 optional。",
+      "kind": "invoke",
+      "query": { "role": "Button", "name": "消息" }
+    },
+    {
+      "_why": "抢前台。下面两处 `type` 都没有 `query`（QQ 的输入框不认 ValuePattern，见 _why_type_steals_screen），走的是键盘、打给'此刻的焦点窗口'——不抢屏就会打到用户正在用的那个程序里去。",
+      "kind": "focus"
+    },
+    { "kind": "wait", "ms": 400 },
+    {
+      "kind": "invoke",
+      "query": { "role": "ComboBox", "name": "搜索" }
+    },
+    {
+      "_why_no_query": "给了 `query` 也没用：实测这个搜索框和下面的消息输入框都拒绝 setValue，runner 会退回键盘。写上反而假装它能后台跑。",
+      "kind": "type",
+      "text": "{contact}"
+    },
+    {
+      "_why": "等搜索结果。这里没有'就是它出现了'的新控件可指——结果列表本来就在，只是内容会换，所以只能钝等。",
+      "kind": "wait",
+      "ms": 1200
+    },
+    {
+      "_why_fragile": "搜索结果里**至少两行**的名字都含 `{contact}`：真正的联系人行（名字形如 `<联系人> <签名或描述>`）和兜底的「进入全网搜索<联系人> 查找用户、群聊等」。这里靠的是**真联系人行排在前面**（实测树序如此，invoke 取第一个匹配）。这条是本 recipe 最脆的一步：QQ 哪天调换了这两行的顺序，消息就会发进一个全网搜索页而不是会话——而每一步依然'成功'。要更稳就得能表达「排除 nameContains=进入全网搜索」，desktop recipe 的 A11yQuery 今天没有否定条件。",
+      "kind": "invoke",
+      "query": { "role": "ListItem", "nameContains": "{contact}" }
+    },
+    {
+      "_why": "等会话面板渲染出来。消息输入框是这一刻才出现的。",
+      "kind": "wait",
+      "ms": 1500
+    },
+    {
+      "_why_name_is_contact": "消息输入框的 a11y name 就是会话名（实测「我的手机」的输入框 name = `我的手机`），所以这一步天然带着'我确实在跟他说话'的校验——会话没切过去的话这个 query 就找不到，整条 recipe 当场停，而不是把消息发给上一个会话。**只在「我的手机」这一个会话上验过**，其他会话是否同样以会话名命名输入框，未验。",
+      "kind": "invoke",
+      "query": { "role": "Edit", "name": "{contact}" }
+    },
+    {
+      "kind": "type",
+      "text": "{message}"
+    },
+    {
+      "_why_button_not_enter": "点「发送」按钮，不按回车：回车键在 QQ 里可以被用户设成换行，那样消息永远发不出去而每一步都'成功'。按钮有句柄、走 invoke，是确定的。",
+      "kind": "invoke",
+      "query": { "role": "Button", "name": "发送" }
+    },
+    {
+      "_why": "等消息进入会话流。observer 立刻读会读到发送前的状态。",
+      "kind": "wait",
+      "ms": 1200
+    }
+  ],
+  "observer": {
+    "itemQuery": { "role": "Text", "nameContains": "{message}" },
+    "fields": {
+      "content": { "read": "name" }
+    },
+    "dedupeBy": "content"
+  },
+  "_why_observer_reads_back_what_it_sent": "这条 recipe 是**动作**不是采集，本来无物可读。但 desktop runner 对'一条都没读到'的处置是 `blocked('drift')`（`src/replay/desktop-runner.ts:183`），而 browser recipe 那边表达动作用的 `allowEmpty` 在 desktop schema 里**不存在**。与其伪造一个恒真的 observer 去骗过这道闸，不如让它读回**刚发出去的那条消息**：发成功 → 读到 1 条 → 通过；没发出去（会话错了、按钮没点着、输入框没收下文字）→ 读不到 → 当场 blocked。于是那道闸从障碍变成了**送达确认**。代价是它会产出一条 item；这个源 `pick_in: []`、`discoverable: false`，不进任何选择面。",
+  "read": {
+    "dedupeBy": "content",
+    "targetCount": 1
+  },
+  "meta": {
+    "description": "给指定 QQ 联系人发一条消息：复位回消息列表 → 搜联系人 → 点进会话 → 写进输入框 → 点发送 → 读回刚发出去的那条当送达确认。桌面 recipe（a11y 词汇，Windows UIA 后端）。",
+    "action": true,
+    "effects": ["write"],
+    "_why_effects": "会真的发消息、写用户账户——安装 preview（src/replay/recipe-install.ts）靠这一格亮副作用牌，不声明就是让装它的人看不见这条会写东西。",
+    "_why_action": "这是动作 recipe（发消息，副作用不可撤回），不是采集——必须显式 opt-in 才能被 MCP 的 run_action_recipe 工具跑，见 src/replay/recipe.ts 的 RecipeMeta.action 头注。",
+    "_why_type_steals_screen": "**这条 recipe 会抢屏，不能当定时任务跑。** Telegram 那条能后台静默跑，是因为它的输入框认 ValuePattern（setValue 直接写进元素，不经键盘）；QQ 的搜索框和消息输入框实测都不认，runner 只能退回键盘，而键盘打给'此刻的焦点窗口'——所以必须先 focus。半夜跑它 = 半夜把用户的屏幕抢过去打字。",
+    "discoverable": false,
+    "_why_pick_in": "两个选择面都不出现——它是动作，不是内容来源",
+    "pick_in": [],
+    "capabilities": [],
+    "member_timeout_ms": 25000,
+    "_why_member_timeout": "一趟要点 5 次、等 3 段（0.4+1.2+1.5+1.2s），加上 Electron 懒建 a11y 树的首次开销，实测约 8-10s。按流式路默认的 15s 会贴着边，照 telegram 那条的先例申报 25s。",
+    "params_schema": {
+      "contact": {
+        "type": "string",
+        "required": true,
+        "description": "发给谁。填搜索框里能命中他的名字（包含匹配即可），例如 `我的手机`。**别填得太短**：短词会同时命中一堆会话，而这一步取的是第一个匹配。"
+      },
+      "message": {
+        "type": "string",
+        "required": true,
+        "description": "消息正文。它同时被当作送达确认的判据（发完从会话里读回这段文字），所以别填一句会话里本来就常见的话——那样即使没发出去也能'读到'。"
+      }
+    }
+  }
+}

package/resources/packages/quark/package.json

@@ -0,0 +1,16 @@
+{
+  "name": "@streamapp/quark",
+  "version": "1.0.0",
+  "author": "stream",
+  "keywords": [
+    "stream-recipe"
+  ],
+  "stream": {
+    "type": "recipe",
+    "name": "夸克",
+    "tagline": "验分享链是否存活 + 列文件",
+    "facility": "quark",
+    "schemaVersion": 1,
+    "cookieDomain": "quark.cn"
+  }
+}

package/resources/packages/quark/quark-share.recipe.json

@@ -0,0 +1,40 @@
+{
+  "version": 3,
+  "kind": "http",
+  "sourceId": "quark-share",
+  "acceptNonOk": true,
+  "output": "object",
+  "request": {
+    "url": "https://drive-h.quark.cn/1/clouddrive/share/sharepage/detail?pr=ucpro&fr=pc&uc_param_str=&ver=2&pwd_id={pwd_id}&pdir_fid=0&_page=1&_size=50",
+    "method": "GET",
+    "headers": { "referer": "https://pan.quark.cn/" }
+  },
+  "compute": {
+    "capabilities": [],
+    "prefetch": [
+      {
+        "as": "tok",
+        "request": {
+          "url": "https://drive-h.quark.cn/1/clouddrive/share/sharepage/token?pr=ucpro&fr=pc&uc_param_str=",
+          "method": "POST",
+          "headers": { "content-type": "application/json", "referer": "https://pan.quark.cn/" },
+          "body": "{\"pwd_id\":\"{pwd_id}\",\"passcode\":\"{passcode}\",\"support_visit_limit_private_share\":true}"
+        }
+      }
+    ],
+    "sign": "(() => { const t = input.pre.tok; const st = (t && t.data && t.data.stoken) || ''; return { url: 'https://drive-h.quark.cn/1/clouddrive/share/sharepage/detail?pr=ucpro&fr=pc&uc_param_str=&ver=2&pwd_id=' + encodeURIComponent(input.params.pwd_id) + '&stoken=' + encodeURIComponent(st) + '&pdir_fid=0&force=0&_page=1&_size=50&_fetch_total=1&_sort=file_type:asc,file_name:asc' }; })()",
+    "decode": "(() => { const t = input.pre.tok; if (!t || typeof t.code !== 'number') throw new Error('quark token: unreadable body'); if (t.code !== 0) { const s = t.status; if (typeof s === 'number' && s >= 400 && s < 500 && s !== 429) return { validity: 'not-usable', files: [], reason: String(t.message || ('code ' + t.code)) }; throw new Error('quark token status ' + s + ' code ' + t.code + ': ' + (t.message || '?')); } if (!t.data || !t.data.stoken) throw new Error('quark token: code 0 without stoken'); const b = input.body; const list = (b && b.data && b.data.list) || []; const files = list.map((f) => ({ name: String(f.file_name || ''), is_dir: !!f.dir, size: Number(f.size || 0) })); return files.length ? { validity: 'alive', files: files } : { validity: 'not-usable', files: [], reason: '分享里没有文件' }; })()"
+  },
+  "assert": [{ "path": "validity", "desc": "quark 探针判决必须带 validity" }],
+  "meta": {
+    "description": "夸克分享验活 + 列文件（probe recipe，只读，判决对象输出；转存见 quark-save）",
+    "discoverable": false,
+    "_why_pick_in": "只在 Provider 成员面挑得到，不在频道订阅面——netdisk-verify-quark 行的成员",
+    "pick_in": ["provider"],
+    "key_param": "pwd_id",
+    "params_schema": {
+      "pwd_id": { "type": "string", "required": true, "description": "夸克分享 id（pan.quark.cn/s/<pwd_id>）" },
+      "passcode": { "type": "string", "required": false, "default": "", "description": "分享提取码（如有）" }
+    }
+  }
+}

package/resources/packages/reddit/package.json

@@ -0,0 +1,15 @@
+{
+  "name": "@streamapp/reddit",
+  "version": "1.0.0",
+  "author": "stream",
+  "keywords": [
+    "stream-recipe"
+  ],
+  "stream": {
+    "type": "recipe",
+    "name": "Reddit",
+    "tagline": "subreddit 热门帖流",
+    "facility": "reddit",
+    "schemaVersion": 1
+  }
+}