[{"_1":2,"_3":-5,"_4":-5},"loaderData",{"_5":6,"_7":8,"_9":10},"actionData","errors","root",{"_165":166,"_167":168},"routes/_app.x.$handle.$name",{"_11":12,"_13":14,"_15":-7,"_18":19,"_17":-7,"_125":-5,"_20":126,"_22":127,"_128":129,"_130":131,"_26":132,"_24":133,"_134":-5,"_135":-5,"_136":63,"_137":138,"_139":63,"_140":-5,"_141":-5,"_142":143},"routes/_app.x.$handle.$name.code.$",{"_11":12,"_13":14,"_15":16,"_17":-7,"_18":19,"_20":21,"_22":23,"_24":25,"_26":27,"_28":-5,"_29":-5,"_30":-5,"_31":-5,"_32":-5},"handle","yuxiaoy","name","biw-tech-rss","branchName","main","version","path","README.md","project",{"_33":70,"_13":14,"_44":71,"_72":-5,"_73":63,"_74":75,"_76":75,"_77":78,"_79":-5,"_80":81,"_82":-5,"_83":63,"_84":-5,"_85":-5,"_86":87,"_88":89,"_90":-5,"_91":-5,"_92":-5,"_93":-5,"_94":95,"_96":95,"_97":98,"_99":98,"_100":98,"_101":95,"_102":95,"_103":104,"_105":106,"_107":108,"_109":63},"tree",[50,51],"file",{"_33":34,"_35":36,"_37":38,"_13":19,"_39":40,"_41":24,"_17":42,"_43":-5,"_44":45,"_46":47,"_48":-5,"_18":49},"contents",[],"upstream","pull","ahead","behind","envVars","id","384b0e80-83af-4904-b84a-2fcf15c883c8","branchId","853a422d-e3ad-4c64-85f7-14098d3fdec9","fileId","4ff7b86e-8ef6-11f1-8ac6-1607ee4eb77e","content","# 车身开发前沿情报机器人（Val Town）\n\n面向**车身开发工程师**，自动抓取国内外关于**新材料 / 新工艺 /\n新技术**的实时资讯，去重、按主题归类后，通过邮件**每天一封（北京时间早上\n9:00）**定时推送。\n\n覆盖 10 个方向：\n\n| 方向 | 说明 |\n| ----------------------- | -------------------------------------- |\n| 新型高强钢 | 第三代高强钢、热成形钢、超高强钢 |\n| 压铸铝合金 / 一体化压铸 | 一体化压铸、免热处理合金、超大型压铸 |\n| 高强压铸铝合金 | 高强韧压铸铝、强度与塑性提升 |\n| 钢铝复合压铸 | 钢铝混合压铸、钢嵌件嵌铸 |\n| 镁合金 | 镁合金车身、镁压铸轻量化 |\n| 镁合金防腐技术 | 防腐涂层、表面处理、微弧氧化 |\n| 镁合金连接技术 | 镁合金焊接铆接、异种材料连接与电偶腐蚀 |\n| 低成本复合材料 | 低成本碳纤维、玻纤、热塑性复合材料 |\n| 高吸能材料 / 碰撞结构 | 吸能盒材料、碰撞吸能结构、泡沫铝 |\n| 连接工艺 | 自冲铆接、激光焊、异种材料胶接 |\n\n数据来源：Google News RSS（中文源 + 英文源），可在 `main.tsx` 的 `TOPICS`\n里自由增删。\n\n---\n\n## 一、它是怎么工作的\n\n1. 每个主题的多个关键词合并成一条 `OR` 检索，并附加 Google News 的 `when:Nd`\n 时间限定符，只拉时间窗内的结果。\n2. 抓取与 blob 读取并发，抓取内部按 `concurrency` 限流。\n3. 用 Val Town 的 `blob` 存储做**跨次去重**，同一条链接不会重复推送。\n4. **LLM 提炼**：给每条内容生成一句话干货，并剔除与车身开发无关的噪音。\n5. 按主题归类，生成中文排版的 HTML 邮件并发送。\n\n### 关于 LLM 提炼（`CONFIG.enableSummary`）\n\n用 Val Town 自带的免费 OpenAI 接口（`std/openai`），无需自备 API Key。\n\n它做两件事：\n\n- **剔除噪音**：股价涨跌、券商研报、招商引资、营销软文、供应商排行榜，以及虽然命中\n 关键词但主体与车身无关的内容（动力电池电芯、智驾、座舱等），都会被判为无关剔除。\n- **提炼干货**：一句话（≤45 字），优先保留材料/工艺名称、企业机构、关键参数、\n 进展阶段（研发/验证/量产/装车），去掉营销修饰语。\n\n> **重要限制**：Google News 的跳转链接是不透明 token，**拿不到原文正文**，\n> 所以摘要是基于**标题**提炼的。提示词里已严格禁止模型补充标题以外的信息，\n> 宁可返回空摘要也不编造。因此它的作用是**快速筛选**，细节仍需点开原文核对。\n>\n> 免费额度限 10 请求/分钟。本程序每次运行**只发 1 次请求**（所有条目批量处理），\n> 完全够用。若接口调用失败，会自动降级为\"只有标题、不过滤\"，不会导致任务失败。\n\n### 性能说明\n\n- 请求数：**58 → 20**（关键词按主题+语种合并成 OR 检索）\n- 依赖：移除 `npm:rss-parser`，改用内置正则解析，**零第三方依赖**、冷启动更快\n- 响应体积：`when:Nd` 让 Google 只返回时间窗内的条目，体积大幅下降\n- 并发：按 `CONFIG.concurrency`（默认 3）限流，配合 503/429 自动重试退避\n\n> **为什么不全并发？** 12 个请求同时打到 Google News，会因共享出口 IP 的突发流量\n> 被限流，全部返回 `HTTP 503`。实测并发 3 是速度与成功率的平衡点。\n> 若日志里仍出现大量 503，把 `concurrency` 调到 2 或 1，并适当调大\n> `retryBaseMs`。\n\n运行日志会打印「N 个请求，M 个有结果，耗时 Xms」，方便你判断是否被限流。\n\n---\n\n## 二、部署到 Val Town（5 分钟）\n\n### 1. 创建 Val\n\n1. 登录，点 **New → Val**。\n2. 把仓库里的 `main.tsx` 全部内容粘贴进去。\n3. 右上角 `+` → 选择 **CRON**，把这个文件设为 Cron 触发。\n\n### 2. 设置定时（每天早上 9 点）\n\nCron 时间按 **UTC** 计算。**北京时间 9:00 = UTC 1:00**，所以填：\n\n```\n0 1 * * *\n```\n\n> 中国不实行夏令时，这个换算全年固定，不用随季节调整。\n>\n> 提示：Val Town 免费版 Cron 最快每 15 分钟一次，每天一次完全够用。\n> 需要换时间可用生成并自动做时区换算。\n\n### 3. 邮件发送\n\n代码按环境变量自动选择通道。\n\n#### 方式 A：不配环境变量（默认）\n\n使用 Val Town 内置 `std/email`，发到**你 Val Town\n账号绑定的邮箱**。免费版即可用。\n\n想转到其他邮箱可在账号邮箱里设 Gmail 过滤器（条件填发件人\n`notifications@val.town`，\n勾选「永不将其发送到垃圾邮件」+「转发给」+「归类为：主要」）。\n\n> 实践中 Gmail 转发坑较多：垃圾箱拦截、中文主题切词失败、转发地址绑定失效、\n> 标签页分类，且**转发后 SPF 校验会失败**，邮件常落在目标邮箱的垃圾箱里。\n> 若调不通，建议直接用方式 B。\n\n#### 方式 B：Brevo 直接投递到指定邮箱（免费，确定性更强）\n\n绕开 Gmail\n转发，程序直接把邮件发到目标地址；失败会立刻抛错并写进日志，便于排查。 Brevo\n免费版每天 300 封、永久免费、**不需要自有域名**。\n\n1. 注册（无需信用卡）\n2. **验证发件地址**：`Senders, Domains & Dedicated IPs` → `Senders` →\n `Add a sender`，填一个你能收信的邮箱，去点确认链接\n3. **拿 API Key**：右上角账号菜单 → `SMTP & API` → `API Keys` → 生成\n4. 在 Val 的 **Environment Variables** 里加三个变量（必须**同时**配齐才生效）：\n\n| 变量 | 值 |\n| ----------------- | --------------------------- |\n| `BREVO_API_KEY` | 上一步生成的 API Key |\n| `BREVO_SENDER` | 第 2 步**已验证**的发件邮箱 |\n| `RECIPIENT_EMAIL` | 你想接收情报的**任意**邮箱 |\n\n> 免费版邮件底部带 \"Sent with Brevo\" 标识，无法去除。\n\n### 4. 先跑一次验证\n\n在编辑器里点 **Run**，几秒后检查邮箱是否收到（首次运行时间窗内若无新内容，会打印\n“本次没有新内容”，可把 `CONFIG.lookbackDays` 临时调大再 Run 一次测试）。\n\n---\n\n## 三、常用自定义（改 `main.tsx` 顶部即可）\n\n| 想做的事 | 改哪里 |\n| ---------------------- | --------------------------------------------------------------------------------- |\n| 增删关注方向 / 关键词 | `TOPICS` 数组 |\n| 只要国内 or 只要国外 | `CONFIG.fetchZh` / `CONFIG.fetchEn` |\n| 时间窗、每主题条数 | `CONFIG.lookbackDays` / `CONFIG.maxItemsPerTopic` |\n| 某方向结果偏少 | 把 `CONFIG.combineQueries` 改成 `false`，退回一个关键词一个请求（更慢但召回更全） |\n| 关掉摘要 / 换模型 | `CONFIG.enableSummary` / `CONFIG.summaryModel` |\n| 觉得噪音剔太狠或不够狠 | 改 `main.tsx` 里 `SUMMARY_PROMPT` 的「keep 判定」规则 |\n| 推送频率 | Val Town 里的 Cron 表达式 |\n\n---\n\n## 四、环境变量清单\n\n全部可选。**不配** = 发到 Val Town 账号邮箱；**三个配齐** = 通过 Brevo\n直投指定邮箱。\n\n| 变量 | 说明 |\n| ----------------- | -------------------------------- |\n| `BREVO_API_KEY` | Brevo 的 API Key |\n| `BREVO_SENDER` | 在 Brevo 验证过的发件邮箱 |\n| `RECIPIENT_EMAIL` | 接收情报的邮箱（可以是任意地址） |\n\n---\n\n## 五、注意事项\n\n- Google News RSS 偶尔会对共享出口 IP\n 限流，代码已对**单个源失败做容错**，不影响其他源。\n- 若长期抓不到某方向，多半是关键词太窄，适当放宽 `TOPICS` 里的检索词。\n- 所有时间以 UTC 存储与比较，邮件里日期为发布日（YYYY-MM-DD）。","fileType",1,"parentId","createdAt",["D",1785732406555],"branchVersion",11,"label",[19],{"_33":34,"_13":19,"_37":38,"_43":-5,"_35":36,"_46":47,"_56":57,"_44":67,"_17":42,"_41":24,"_39":40,"_62":63,"_48":-5,"_64":68,"_18":69},{"_33":52,"_13":53,"_37":54,"_43":-5,"_35":36,"_46":55,"_56":57,"_44":58,"_17":59,"_41":60,"_39":61,"_62":63,"_48":-5,"_64":65,"_18":66},"67ba90d4-7663-4284-8e43-3086393249f1","main.ts","1b4351b4-86a9-11f1-8629-1607ee4eb77e",9,"latest",true,["D",1784951122461],8,"interval","// ============================================================================\n// 车身开发前沿情报机器人 (Car Body Material & Process Radar)\n// ----------------------------------------------------------------------------\n// 面向车身开发工程师，自动抓取国内外关于新材料 / 新工艺 / 新技术的实时资讯，\n// 去重、按主题归类后，通过邮件定时推送。\n//\n// 部署平台：Val Town (https://val.town) —— 这是一个 Cron 触发的 Val。\n// 运行时：Deno。除 Val Town 标准库外无第三方依赖，冷启动最快。\n//\n// 触发方式：把本文件设为 Cron 触发，每天一次，北京时间 9:00 → UTC `0 1 * * *`。\n// ============================================================================\n\nimport { email } from \"https://esm.town/v/std/email\";\nimport { blob } from \"https://esm.town/v/std/blob\";\nimport { OpenAI } from \"https://esm.town/v/std/openai\";\n\n// ----------------------------------------------------------------------------\n// 1) 关注的主题与检索关键词\n// 每个主题包含中文与英文关键词，分别去抓国内源和国外源。\n// 需要增删方向时，直接改这里即可。\n// ----------------------------------------------------------------------------\ninterface Topic {\n id: string;\n label: string; // 邮件里显示的中文分组名\n zh: string[]; // 中文检索词（Google News 中文）\n en: string[]; // 英文检索词（Google News 英文）\n}\n\nconst TOPICS: Topic[] = [\n {\n id: \"steel\",\n label: \"新型高强钢\",\n zh: [\"汽车第三代高强钢\", \"车身热成形钢\", \"汽车超高强钢应用\"],\n en: [\"automotive third generation AHSS\", \"car body press hardened steel\", \"gigapascal steel automotive\"],\n },\n {\n id: \"aldiecast\",\n label: \"压铸铝合金 / 一体化压铸\",\n zh: [\"一体化压铸车身\", \"免热处理压铸铝合金\", \"超大型压铸汽车\"],\n en: [\"gigacasting car body\", \"heat-free die casting aluminum alloy automotive\", \"megacasting vehicle structure\"],\n },\n {\n id: \"mg\",\n label: \"镁合金\",\n zh: [\"镁合金汽车车身\", \"镁合金压铸汽车轻量化\"],\n en: [\"magnesium alloy automotive body\", \"magnesium die casting vehicle lightweight\"],\n },\n {\n id: \"composite\",\n label: \"低成本复合材料\",\n zh: [\"汽车碳纤维复合材料低成本\", \"车身玻纤复合材料轻量化\", \"热塑性复合材料汽车\"],\n en: [\"low cost carbon fiber automotive\", \"thermoplastic composite car body\", \"glass fiber composite vehicle lightweight\"],\n },\n {\n id: \"energy\",\n label: \"高吸能材料 / 碰撞结构\",\n zh: [\"汽车吸能盒材料\", \"车身碰撞吸能结构新材料\", \"泡沫铝汽车吸能\"],\n en: [\"automotive crash energy absorption material\", \"crashworthiness structure new material vehicle\", \"aluminum foam energy absorbing car\"],\n },\n {\n id: \"joining\",\n label: \"连接工艺\",\n zh: [\"车身自冲铆接 SPR\", \"汽车激光焊接铝\", \"车身异种材料连接胶接\"],\n en: [\"self piercing rivet car body\", \"laser welding aluminum automotive\", \"dissimilar material joining car body adhesive\"],\n },\n];\n\n// ----------------------------------------------------------------------------\n// 2) 运行参数（可按需调整）\n// ----------------------------------------------------------------------------\nconst CONFIG = {\n // 只保留最近多少天内发布的内容（兜底窗口；配合去重使用）。\n // 每天推送时 2 天足够，留一天余量以防某次运行失败漏内容。\n lookbackDays: 2,\n // 每个主题最多展示多少条，避免邮件过长\n maxItemsPerTopic: 6,\n // 去重库里最多保留多少条历史链接，超出后自动清理最旧的\n maxSeenLinks: 4000,\n // 存储去重数据用的 blob key\n blobKey: \"car_body_radar_seen_links\",\n // 抓取语言：由用户选择“中英文都要”\n fetchZh: true,\n fetchEn: true,\n // 单个 RSS 请求超时（毫秒）\n fetchTimeoutMs: 10000,\n // 并发上限。Google News 会对共享出口 IP 的突发请求返回 503，\n // 全并发（12）必被限流，3 是实测下速度与成功率的平衡点。\n concurrency: 3,\n // 遇到 503 / 429 / 网络错误时的重试次数与退避基数\n maxRetries: 2,\n retryBaseMs: 1000,\n // 把同一主题的多个关键词合并成一条 OR 检索，请求数从 34 降到 12。\n // 若发现某方向结果变少，可改为 false 退回“一个关键词一个请求”。\n combineQueries: true,\n // 用 LLM 为每条内容提炼一句话干货，并剔除与车身开发无关的噪音。\n // 关掉后邮件只有标题，行为退回到之前的版本。\n enableSummary: true,\n // Val Town 免费版可用的低价模型（gpt-4o-mini / gpt-5-nano / gpt-4.1-nano）\n summaryModel: \"gpt-4o-mini\",\n // 送给模型的候选条数 = maxItemsPerTopic × 该倍数，为被剔除的噪音留出余量\n candidateMultiplier: 2,\n};\n\n// ----------------------------------------------------------------------------\n// 3) 工具函数\n// ----------------------------------------------------------------------------\ninterface Article {\n title: string;\n link: string;\n source: string;\n pubMs: number; // 发布时间戳，解析一次后复用，避免反复 new Date()\n lang: \"zh\" | \"en\";\n summary?: string; // LLM 提炼的一句话干货\n}\n\n// 预编译正则，避免每条 item 重新构造\nconst ITEM_RE = /]*>([\\s\\S]*?)<\\/item>/gi;\nconst TITLE_RE = /]*>([\\s\\S]*?)<\\/title>/i;\nconst LINK_RE = /]*>([\\s\\S]*?)<\\/link>/i;\nconst PUBDATE_RE = /]*>([\\s\\S]*?)<\\/pubDate>/i;\nconst SOURCE_RE = /]*>([\\s\\S]*?)<\\/source>/i;\n\n// & 必须最后解，否则会把 < 这类二次转义错误展开\nfunction decodeXml(raw: string): string {\n return raw\n .replace(//g, \"$1\")\n .replace(/</g, \"<\")\n .replace(/>/g, \">\")\n .replace(/"/g, '\"')\n .replace(/'|�*39;/g, \"'\")\n .replace(/&#(\\d+);/g, (_, d) => String.fromCodePoint(Number(d)))\n .replace(/&/g, \"&\")\n .trim();\n}\n\nfunction pick(block: string, re: RegExp): string {\n const m = block.match(re);\n return m ? decodeXml(m[1]) : \"\";\n}\n\n// when:Nd 让 Google 只返回时间窗内的结果：响应体积大幅变小，\n// 且避免高相关度的旧闻挤占前 100 条名额、把新内容挤掉。\nfunction googleNewsUrl(query: string, lang: \"zh\" | \"en\", windowDays: number): string {\n const q = encodeURIComponent(`${query} when:${windowDays}d`);\n return lang === \"zh\"\n ? `https://news.google.com/rss/search?q=${q}&hl=zh-CN&gl=CN&ceid=CN:zh-Hans`\n : `https://news.google.com/rss/search?q=${q}&hl=en-US&gl=US&ceid=US:en`;\n}\n\n// 规范化链接，去掉跳转参数带来的重复。无 ?/# 的链接走快速路径\nfunction normalizeLink(link: string): string {\n if (!link.includes(\"?\") && !link.includes(\"#\")) return link;\n try {\n const u = new URL(link);\n u.hash = \"\";\n [\"utm_source\", \"utm_medium\", \"utm_campaign\", \"utm_term\", \"utm_content\", \"oc\"].forEach((p) =>\n u.searchParams.delete(p),\n );\n return u.toString();\n } catch {\n return link;\n }\n}\n\nfunction parseFeed(xml: string, lang: \"zh\" | \"en\"): Article[] {\n const out: Article[] = [];\n ITEM_RE.lastIndex = 0;\n let m: RegExpExecArray | null;\n while ((m = ITEM_RE.exec(xml)) !== null) {\n const block = m[1];\n const pubMs = new Date(pick(block, PUBDATE_RE)).getTime();\n if (isNaN(pubMs)) continue;\n\n // Google News 标题格式常为 \"标题 - 来源\"，标签里有更干净的来源名\n const rawTitle = pick(block, TITLE_RE);\n const source = pick(block, SOURCE_RE);\n const suffix = source ? ` - ${source}` : \"\";\n const title = suffix && rawTitle.endsWith(suffix) ? rawTitle.slice(0, -suffix.length).trim() : rawTitle;\n\n const link = normalizeLink(pick(block, LINK_RE));\n if (!title || !link) continue;\n out.push({ title, link, source, pubMs, lang });\n }\n return out;\n}\n\nconst sleep = (ms: number) => new Promise((r) => setTimeout(r, ms));\n\n// 真实浏览器 UA，自定义 UA 更容易被 Google 拦截\nconst UA =\n \"Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/126.0.0.0 Safari/537.36\";\n\nasync function fetchFeed(url: string, lang: \"zh\" | \"en\"): Promise {\n let lastErr = \"\";\n\n for (let attempt = 0; attempt <= CONFIG.maxRetries; attempt++) {\n const controller = new AbortController();\n const timer = setTimeout(() => controller.abort(), CONFIG.fetchTimeoutMs);\n try {\n const res = await fetch(url, {\n signal: controller.signal,\n headers: {\n \"User-Agent\": UA,\n Accept: \"application/rss+xml, application/xml;q=0.9, */*;q=0.8\",\n \"Accept-Language\": lang === \"zh\" ? \"zh-CN,zh;q=0.9\" : \"en-US,en;q=0.9\",\n },\n });\n if (res.ok) return parseFeed(await res.text(), lang);\n\n lastErr = `HTTP ${res.status}`;\n await res.body?.cancel(); // 不读走响应体会让连接一直挂着\n // 4xx（除 429）是稳定错误，重试没有意义\n if (res.status < 500 && res.status !== 429) break;\n } catch (err) {\n lastErr = (err as Error).message;\n } finally {\n clearTimeout(timer);\n }\n\n // 指数退避 + 随机抖动，避免多个请求同时重试再次撞上限流\n if (attempt < CONFIG.maxRetries) {\n await sleep(CONFIG.retryBaseMs * 2 ** attempt + Math.random() * 500);\n }\n }\n\n console.warn(`抓取失败 [${lang}] ${lastErr}: ${url}`);\n return [];\n}\n\n// 限制同时在飞的请求数\nasync function mapLimit(items: T[], limit: number, fn: (item: T) => Promise): Promise {\n const results = new Array(items.length);\n let next = 0;\n const workers = Array.from({ length: Math.min(limit, items.length) }, async () => {\n while (true) {\n const i = next++;\n if (i >= items.length) break;\n results[i] = await fn(items[i]);\n }\n });\n await Promise.all(workers);\n return results;\n}\n\nfunction escapeHtml(s: string): string {\n return s\n .replace(/&/g, \"&\")\n .replace(//g, \">\")\n .replace(/\"/g, \""\");\n}\n\nfunction fmtDate(ms: number): string {\n return new Date(ms).toISOString().slice(0, 10);\n}\n\n// ----------------------------------------------------------------------------\n// 4) LLM 提炼：一句话干货 + 剔除无关噪音\n// 只有标题可用（Google News 的跳转链接是不透明 token，拿不到原文正文），\n// 因此必须严格约束模型不得补充标题以外的信息。\n// ----------------------------------------------------------------------------\nconst SUMMARY_PROMPT = `你是资深汽车车身开发工程师，负责筛选并提炼技术情报。\n\n输入是一个 JSON 数组，每个元素含 i(编号)、topic(所属方向)、title(新闻标题)、source(来源)。\n请对每一条做两件事：判断是否值得保留，以及提炼一句话干货。\n\n【keep 判定】剔除与车身材料/工艺/结构技术无关的条目，令 keep=false：\n- 股价涨跌、券商研报评级、市值与财报解读\n- 招商引资、产业园签约、企业营销软文、供应商/工厂排行榜\n- 虽含关键词但主体与车身无关的内容（如动力电池电芯、智驾、座舱、充电）\n- 纯行业规模预测、\"前景广阔\"类无技术信息的综述\n其余与材料牌号、工艺方法、结构设计、性能指标、装车应用相关的，keep=true。\n\n【summary 撰写】\n1. 只能使用 title 中已有的信息，严禁推测、补充或编造任何标题里没有的内容。\n2. 一句中文，不超过 45 字，优先包含：材料/工艺名称、企业或机构、关键参数、\n 进展阶段（研发/验证/量产/装车）。\n3. 去掉营销修饰语和\"据报道\"\"本文介绍\"这类废话，只留技术事实。\n4. 若标题信息太少、提炼不出有效技术信息，summary 返回空字符串。\n\n只输出 JSON，格式：{\"results\":[{\"i\":0,\"keep\":true,\"summary\":\"...\"}]}`;\n\ninterface Verdict {\n keep: boolean;\n summary: string;\n}\n\nasync function enrich(groups: { topic: Topic; items: Article[] }[]): Promise> {\n const out = new Map();\n if (!CONFIG.enableSummary) return out;\n\n const flat = groups.flatMap(({ topic, items }) => items.map((a) => ({ topicLabel: topic.label, a })));\n if (flat.length === 0) return out;\n\n const payload = flat.map((x, i) => ({\n i,\n topic: x.topicLabel,\n title: x.a.title,\n source: x.a.source,\n }));\n\n try {\n const openai = new OpenAI();\n const completion = await openai.chat.completions.create({\n model: CONFIG.summaryModel,\n response_format: { type: \"json_object\" },\n messages: [\n { role: \"system\", content: SUMMARY_PROMPT },\n { role: \"user\", content: JSON.stringify(payload) },\n ],\n });\n\n const parsed = JSON.parse(completion.choices[0]?.message?.content || \"{}\") as {\n results?: { i: number; keep?: boolean; summary?: string }[];\n };\n\n let dropped = 0;\n for (const r of parsed.results || []) {\n const item = flat[r.i];\n if (!item) continue;\n const keep = r.keep !== false;\n if (!keep) dropped++;\n out.set(item.a.link, { keep, summary: (r.summary || \"\").trim() });\n }\n console.log(`LLM 提炼完成：评估 ${flat.length} 条，剔除 ${dropped} 条噪音`);\n } catch (err) {\n // 提炼失败不应让整个任务失败，降级成\"只有标题、不过滤\"\n console.warn(`LLM 提炼失败，降级为无总结：${(err as Error).message}`);\n return new Map();\n }\n\n return out;\n}\n\n// ----------------------------------------------------------------------------\n// 5) 邮件发送：使用 Val Town 内置 std/email，发到你的 Val Town 账号邮箱。\n// 不传 to，免费版即可正常发送（to 留空时默认发给账号所有者）。\n// 如需转投到其他邮箱，在账号邮箱里设一条自动转发规则即可。\n// ----------------------------------------------------------------------------\nasync function sendEmail(subject: string, html: string) {\n await email({ subject, html });\n console.log(\"已通过 std/email 发送到 Val Town 账号邮箱\");\n}\n\n// ----------------------------------------------------------------------------\n// 6) 主流程（Cron 入口）\n// ----------------------------------------------------------------------------\ninterface FetchTask {\n topicIndex: number;\n lang: \"zh\" | \"en\";\n query: string;\n}\n\nfunction buildTasks(): FetchTask[] {\n const tasks: FetchTask[] = [];\n TOPICS.forEach((topic, topicIndex) => {\n for (const lang of [\"zh\", \"en\"] as const) {\n if (lang === \"zh\" && !CONFIG.fetchZh) continue;\n if (lang === \"en\" && !CONFIG.fetchEn) continue;\n const keywords = lang === \"zh\" ? topic.zh : topic.en;\n if (CONFIG.combineQueries) {\n tasks.push({ topicIndex, lang, query: keywords.map((k) => `(${k})`).join(\" OR \") });\n } else {\n keywords.forEach((query) => tasks.push({ topicIndex, lang, query }));\n }\n }\n });\n return tasks;\n}\n\nasync function loadSeen(): Promise> {\n try {\n return (await blob.getJSON(CONFIG.blobKey)) || {};\n } catch {\n return {};\n }\n}\n\nexport default async function (interval: { lastRunAt?: Date } = {}) {\n const startedAt = Date.now();\n const now = startedAt;\n // 兜底时间窗：取“上次运行时间”与“lookbackDays”里更早的那个，避免漏内容\n const lookbackMs = now - CONFIG.lookbackDays * 24 * 3600 * 1000;\n const lastRunMs = interval?.lastRunAt ? new Date(interval.lastRunAt).getTime() : lookbackMs;\n const sinceMs = Math.min(lookbackMs, isNaN(lastRunMs) ? lookbackMs : lastRunMs);\n\n const tasks = buildTasks();\n // 多留 1 天余量，避开 Google 与本地时间窗的边界差\n const windowDays = Math.min(30, Math.ceil((now - sinceMs) / 86400000) + 1);\n\n // 抓取与 blob 读取并发；抓取内部按 concurrency 限流，避免触发 Google 503\n const [seen, feeds] = await Promise.all([\n loadSeen(),\n mapLimit(tasks, CONFIG.concurrency, (t) => fetchFeed(googleNewsUrl(t.query, t.lang, windowDays), t.lang)),\n ]);\n const okCount = feeds.filter((f) => f.length > 0).length;\n console.log(\n `抓取完成：${tasks.length} 个请求，${okCount} 个有结果，耗时 ${Date.now() - startedAt}ms`,\n );\n\n // 按主题归拢抓取结果\n const byTopic: Article[][] = TOPICS.map(() => []);\n feeds.forEach((articles, i) => byTopic[tasks[i].topicIndex].push(...articles));\n\n // 多取一些候选，给 LLM 剔除噪音后留出余量\n const candidateLimit = CONFIG.enableSummary\n ? CONFIG.maxItemsPerTopic * CONFIG.candidateMultiplier\n : CONFIG.maxItemsPerTopic;\n\n const candidates = TOPICS.map((topic, i) => {\n // 主题内去重（按链接），并过滤时间窗 + 全局已推送\n const dedup = new Map();\n for (const a of byTopic[i]) {\n if (a.pubMs < sinceMs) continue;\n if (seen[a.link]) continue;\n if (!dedup.has(a.link)) dedup.set(a.link, a);\n }\n // 按发布时间倒序\n const items = [...dedup.values()].sort((x, y) => y.pubMs - x.pubMs).slice(0, candidateLimit);\n return { topic, items };\n });\n\n const verdicts = await enrich(candidates);\n\n const grouped: { topic: Topic; items: Article[] }[] = [];\n let totalNew = 0;\n\n for (const { topic, items } of candidates) {\n const kept: Article[] = [];\n for (const a of items) {\n const v = verdicts.get(a.link);\n if (v && !v.keep) {\n seen[a.link] = now; // 判定无关的也记下，避免每天重复送进模型\n continue;\n }\n if (v?.summary) a.summary = v.summary;\n kept.push(a);\n }\n\n // 超出上限的留着不标记，下次运行还有机会出现\n const picked = kept.slice(0, CONFIG.maxItemsPerTopic);\n picked.forEach((a) => (seen[a.link] = now));\n totalNew += picked.length;\n grouped.push({ topic, items: picked });\n }\n\n // 清理去重库，避免无限膨胀\n let toStore = seen;\n const seenEntries = Object.entries(seen);\n if (seenEntries.length > CONFIG.maxSeenLinks) {\n seenEntries.sort((a, b) => b[1] - a[1]);\n toStore = Object.fromEntries(seenEntries.slice(0, CONFIG.maxSeenLinks));\n }\n\n if (totalNew === 0) {\n await blob.setJSON(CONFIG.blobKey, toStore);\n console.log(`本次没有新内容，跳过发送。总耗时 ${Date.now() - startedAt}ms`);\n return { totalNew: 0 };\n }\n\n // 写去重库与发信互不依赖，并发执行\n const html = renderEmail(grouped, totalNew, sinceMs, now);\n const dateStr = new Date(now).toISOString().slice(0, 10);\n await Promise.all([\n blob.setJSON(CONFIG.blobKey, toStore),\n sendEmail(`【车身前沿情报】${dateStr} · 新增 ${totalNew} 条`, html),\n ]);\n\n console.log(`完成：共推送 ${totalNew} 条，总耗时 ${Date.now() - startedAt}ms`);\n return { totalNew };\n}\n\n// ----------------------------------------------------------------------------\n// 7) 邮件模板\n// ----------------------------------------------------------------------------\nfunction renderEmail(\n grouped: { topic: Topic; items: Article[] }[],\n totalNew: number,\n sinceMs: number,\n now: number,\n): string {\n const sinceStr = new Date(sinceMs).toISOString().slice(0, 10);\n const nowStr = new Date(now).toISOString().slice(0, 10);\n\n const sections = grouped\n .filter((g) => g.items.length > 0)\n .map((g) => {\n const rows = g.items\n .map((a) => {\n const flag = a.lang === \"zh\" ? \"国内\" : \"国外\";\n const meta = [a.source, fmtDate(a.pubMs)].filter(Boolean).join(\" · \");\n const summary = a.summary\n ? `

${escapeHtml(a.summary)}

`\n : \"\";\n return `\n

\n \n ${escapeHtml(a.title)}\n \n ${summary}\n

\n ${flag}\n ${escapeHtml(meta)}\n

`;\n })\n .join(\"\");\n return `\n

\n ${escapeHtml(g.topic.label)}\n （${g.items.length}）\n

${rows}\n

`;\n })\n .join(\"\");\n\n return `\n

车身开发前沿情报

新材料 · 新工艺 · 新技术｜${sinceStr} ~ ${nowStr}｜共 ${totalNew} 条

\n ${sections}\n

\n 本邮件由部署在 Val Town 的自动化程序生成，数据来源：Google News（中/英）。\n 摘要由 LLM 基于标题提炼，仅供快速筛选，细节请点开原文核对。\n 如需调整关注方向或推送频率，请修改 Val 中的 TOPICS / Cron 配置。\n

`;\n}","deleted",false,"children",[],[53],["D",1785732406555],[],[19],"44f6b27a-dccf-41dc-ad02-50c2620e967c",["D",1784819628999],"deletedAt","legacyEnvVars","privacy","public","httpPrivacy","description","","imageUrl","userId","96c2793d-57c8-41d2-a49a-d1ee82f9dd24","pinnedFileId","pinPreview","pinMigrated","screenshotVersion","likeCount",0,"orgId","93515af8-c7b7-459c-9bf2-3bb3df310a75","tursoName","tursoHostname","tursoMeasuredAt","pinned","userPerm","r","valPerm","blobPerm","rw","emailPerm","sqlitePerm","projectPerm","telemetryPerm","user",{"_11":12,"_124":120},"org",{"_11":12,"_118":-5,"_119":120,"_80":81,"_121":122,"_123":63},"branch",{"_17":47,"_13":16,"_33":36,"_110":70,"_111":-5,"_112":-5,"_113":-5,"_114":-5,"_115":-5,"_116":14,"_117":12},"isOwned","projectId","forkedBranchId","forkedBranchVersion","forkedProjectPrivacy","forkedProjectId","forkedProjectOrgId","projectName","userHandle","displayName","avatar","https://img.clerk.com/eyJ0eXBlIjoiZGVmYXVsdCIsImlpZCI6Imluc18yTm10cGhQUklwU0tqZHNiOHNnNGd2ZjJURFUiLCJyaWQiOiJ1c2VyXzJObXd1QWdCZmVtaUh0bGU5TkswOG1sM0xheiIsImluaXRpYWxzIjoiRlkifQ","tier","free","canRestrictVals","profileImageUrl","adoptOrgId",{"_33":70,"_13":14,"_44":161,"_72":-5,"_73":63,"_74":75,"_76":75,"_77":78,"_79":-5,"_80":81,"_82":-5,"_83":63,"_84":-5,"_85":-5,"_86":87,"_88":89,"_90":-5,"_91":-5,"_92":-5,"_93":-5,"_94":95,"_96":95,"_97":98,"_99":98,"_100":98,"_101":95,"_102":95,"_103":162,"_105":163,"_107":164,"_109":63},[155,156],"files",[155,156],"counts",{"_146":87,"_147":42,"_148":87,"_32":149,"_150":151},[],{"_33":34,"_13":19,"_37":38,"_43":-5,"_35":36,"_46":47,"_56":57,"_44":144,"_17":42,"_41":24,"_39":40,"_62":63,"_48":-5,"_18":145},"upstreamProject","upstreamBranch","emptyEnvVars","flags",[],"hasHttpFiles","grants","orgs","customDomains",{},["D",1785732406555],[19],"pulls","branches","forks",{"_152":87,"_154":87},"connections",{"_152":87,"_153":87},"count","unfilledCount","emptyCount",{"_33":34,"_13":19,"_37":38,"_43":-5,"_35":36,"_46":47,"_56":57,"_44":67,"_17":42,"_41":24,"_39":40,"_62":63,"_48":-5,"_64":159,"_18":160},{"_33":52,"_13":53,"_37":54,"_43":-5,"_35":36,"_46":55,"_56":57,"_44":58,"_17":59,"_41":60,"_39":61,"_62":63,"_48":-5,"_64":157,"_18":158},[],[53],[],[19],["D",1784819628999],{"_11":12,"_124":120},{"_11":12,"_118":-5,"_119":120,"_80":81,"_121":122,"_123":63},{"_17":47,"_13":16,"_33":36,"_110":70,"_111":-5,"_112":-5,"_113":-5,"_114":-5,"_115":-5,"_116":14,"_117":12},"ctx",{"_103":-5,"_141":198,"_199":200,"_201":202,"_203":63,"_204":63,"_205":63,"_206":57,"_207":-5,"_208":-7,"_209":-7,"_210":87,"_211":63,"_212":63,"_213":-5,"_214":63,"_215":63,"_216":63,"_217":-5},"ENV",{"_169":170,"_171":172,"_173":16,"_174":175,"_176":177,"_178":179,"_180":181,"_182":183,"_184":185,"_186":187,"_188":189,"_190":191,"_192":193,"_194":195,"_196":197},"REACT_APP_SENTRY_DSN","https://586020d86f05489db63a554690acf1e9@o1398359.ingest.sentry.io/6724499","RENDER_GIT_COMMIT","e87eb04e707cf64e7e532c8e8ec395f209e17f15","RENDER_GIT_BRANCH","POSTHOG_API_TOKEN","phc_nmoyuAkStEZuUp69rYlg1DZgrIt19Gx0Y4Hwf8tGgJP","POSTHOG_API_HOST","https://wart.val.town","WEB_API_TEMPLATE","https://{val}.web.val.run","APP_URL","https://www.val.town","MODULE_URL","https://esm.town","VANITY_WEB_API_TEMPLATE","https://{label}.val.run/","DENOLS_URL","https://lsp.val.town","NODE_ENV","production","CLOUDFLARE_TURNSTILE_SITE_KEY","0x4AAAAAACtaIZJH7s9rBMZ6","GOOGLE_CLIENT_ID","875989696899-jtsqkhlbdv5jrlsk8631p7ipjg3c6o62.apps.googleusercontent.com","STRIPE_PUBLISHABLE_KEY","pk_live_51N2zvWEpZTTLipIkVEwXrUG9AT0yeOOEYxszarZT1jpsZV58Ly6wlg3SltZpHN26CUalOBmUSY7adQRQiSkCZz2500qSXNkf8Y","API_URL","https://api.val.town",[],"preferencesKeybindings","default","preferencesTheme","light","preferencesFixers","preferencesCopilot","preferencesSelectionAi","preferencesCloseBrackets","preferencesChatPanelState","preferencesDprint","preferencesDenofmt","notificationCount","flagDenols","isBusinessOrgMember","businessOrgHandle","isBusinessOrgAdmin","isHigherTownieTier","navCollapsed","currentOrgId"]