按 topics 检索 GitHub 最近 N 天新建的仓库，硬性质量过滤与热度评分精选 Top N，并发拉取 README 自动生成中文摘要，输出暗黑卡片风 HTML 与 Markdown 双格式报告，可选推送飞书 / 钉钉群。 It searches GitHub for repos created within the last N days by topics, applies hard quality filters and a weighted trending score to pick the Top N, fetches READMEs concurrently to build automatic Chinese summaries, and outputs a dark-themed HTML + Markdown report — with optional Feishu / DingTalk group push.
摘自 2026-10-24 由计划任务自动生成的真实早报，数据未做修改 From the real report auto-generated by the scheduled task on 2026-10-24 — unedited
九个环节全部内置在一个 main.py 里（Config / GitHubClient / RepoFilter / Scorer / ReadmeSummarizer / Translator / ReportRenderer / Notifier / Pipeline），装好依赖即可运行。 All nine stages live inside a single main.py (Config / GitHubClient / RepoFilter / Scorer / ReadmeSummarizer / Translator / ReportRenderer / Notifier / Pipeline) — install the dependencies and you are ready to go.
调用 GitHub Search API，topic:{t} + created:>{N天前} + language:{lang} 多组合逐个查询、合并去重，只盯最近 N 天新冒出来的仓库。Queries the GitHub Search API with topic:{t} + created:>{N-days-ago} + language:{lang} combinations, merged and deduplicated — only repos created within the last N days.
多 PAT 自动轮换，实时监控 x-ratelimit-remaining；全部耗尽按 x-ratelimit-reset 等待（封顶 120 秒）；网络异常 / 5xx 指数退避重试，超时 connect 10s / read 30s。Rotates multiple PATs automatically while watching x-ratelimit-remaining; when exhausted it waits per x-ratelimit-reset (capped at 120 s). Network errors and 5xx get exponential-backoff retries; timeouts are 10 s connect / 30 s read.
剔除 fork 仓库、Star 低于 min_stars、无描述、README 拉取失败或为空的条目，全程打印各阶段过滤统计。Drops forks, repos below min_stars, entries without a description, and repos whose README fetch failed or came back empty — with per-stage filter stats printed along the way.
日均增星 + 对数压缩的 Star / Fork 体量三项加权求和，权重 w1/w2/w3 全部可在配置里调整，按分数排序取 Top N。Combines stars-per-day with log-scaled star and fork volume; the three weights w1/w2/w3 are fully configurable, and the list is ranked by score to keep the Top N.
线程池并发拉取 README（raw 优先、Base64 回退，截前 2000 字符），LLM 提炼【核心功能】【技术栈】【适用场景】与一句话亮点；未配置 LLM 时本地规则生成同等结构，任何字段不会留空。Fetches READMEs concurrently (raw first, Base64 fallback, first 2,000 chars) and distills core features, tech stack, use cases and a one-line highlight via any OpenAI-compatible LLM. Without an LLM key, local rules produce the same structure — no field is ever left empty.
MyMemory 免费引擎（无需 API Key，匿名额度约 5000 字符/天），翻译结果缓存到本地，重复内容零消耗；超额度内容次日自动补翻，失败只跳过该条，绝不中断报告。The MyMemory engine works without an API key (roughly 5,000 chars/day anonymously). Results are cached locally so repeated content costs nothing; anything over the daily budget is translated the next day, and a failed item is simply skipped — report generation never breaks.
Jinja2 内嵌模板渲染暗黑扁平风、响应式卡片 HTML（作者头像 / Star·Fork / 语言徽章 / 热度分 / 三个摘要小节 / 直达链接），同步生成 Markdown 报告。A built-in Jinja2 template renders a dark, flat, responsive card-style HTML report (author avatar, stars & forks, language badge, trending score, three summary sections, direct links), with a matching Markdown report generated side by side.
飞书 interactive 卡片与钉钉 markdown 消息两种格式独立开关；推送失败只打印告警，不影响本地 HTML / Markdown 输出；未配置 Webhook 时保存后自动用浏览器打开。Feishu interactive cards and DingTalk markdown messages, each with an independent switch. A failed push only prints a warning and never affects local output; with no webhook configured, the HTML opens in your browser right after saving.
双击 install_task.bat 注册每天 20:00 的计划任务；启动器部署在 %LOCALAPPDATA%，凭标记文件定位工具目录——本文件夹改名、移动都不会让任务失效。Double-click install_task.bat to register the daily 20:00 task. The launcher lives under %LOCALAPPDATA% and locates the tool by a marker file — renaming or moving the tool folder never breaks the task.
需要 Python 3.9+（建议 3.10+）。全部实现集中在 main.py 一个文件，没有 Web 服务、没有数据库，clone 下来就能用。 Requires Python 3.9+ (3.10+ recommended). Everything lives in the single main.py — no web server, no database. Clone it and go.
只有 5 个第三方依赖，一条命令装完：Just five third-party dependencies, installed in one command:
run 是默认子命令，可省略；schedule 启动后立即跑第一轮，Ctrl+C 退出：run is the default sub-command and can be omitted; schedule runs the first round immediately and stops with Ctrl+C:
运行时终端输出扫描进度条、过滤统计表与 Top N 摘要表格；结束后报告落盘，未配置 Webhook 且 auto_open: true 时自动在浏览器打开 HTML。 The terminal shows a scan progress bar, filter statistics and a Top N summary table; reports are written on finish, and with no webhook plus auto_open: true the HTML opens in your browser automatically.
双击 install_task.bat，注册名为 GitHubTrendingDaily 的计划任务，安装一次长期无人值守：Double-click install_task.bat to register a scheduled task named GitHubTrendingDaily — install once and forget about it:
报告按时间戳归档，同时维护固定文件名的「最新报告」与可翻阅的历史索引： Each report is archived by timestamp, alongside a fixed-name "latest" copy and a browsable history index:
所有字段均可省略，省略时使用内置默认值；工具附带完整注释版配置文件，改哪一行都有说明。 Every field is optional — omit it and the built-in default applies. A fully commented config file ships with the tool.
| 配置项Option | 默认值Default | 说明Description |
|---|---|---|
| topics | unity · game-development · llm · agent | 要监控的 GitHub topics，逐个参与组合查询。GitHub topics to watch; each is queried in combination. |
| languages | []（不限语言）[] (any language) | 只关注这些编程语言，与 topics 组合查询；留空表示不限。Restrict searches to these languages; empty means any language. |
| days | 7 | 只检索最近 N 天内创建的仓库。Only repos created within the last N days. |
| top_n | 12 | 最终榜单数量（建议 10~15）。Final list size (10–15 recommended). |
| min_stars | 20 | 硬性过滤：低于该 Star 数的仓库直接剔除。Hard filter: repos below this star count are dropped. |
| per_page | 50 | 单次搜索请求返回的仓库数上限（1~100）。Max results per search request (1–100). |
| weights | w1 2.0 / w2 1.0 / w3 0.5 | 热度评分公式三项权重，见下方公式。The three weights of the trending formula, see below. |
| tokens | []（匿名访问）[] (anonymous access) | GitHub PAT 列表，多 Token 自动轮换；留空为匿名限流 60 次/小时。GitHub PAT pool, rotated automatically; empty means anonymous access at 60 requests/hour. |
| max_workers | 6 | 拉取 README / 调用 LLM 的并发线程数。Worker threads for README fetching and LLM calls. |
| llm.enabled | false | 启用后由任意 OpenAI 兼容接口提炼中文摘要；关闭时走本地规则。When on, any OpenAI-compatible API writes the Chinese summaries; when off, local rules apply. |
| webhooks.feishu / dingtalk | false / "" | 飞书、钉钉群机器人的独立开关与 Webhook 地址。Independent Feishu / DingTalk bot toggles and webhook URLs. |
| translate.enabled | true | 免费中文翻译（MyMemory 引擎，结果缓存，超额度次日补翻）。Free Chinese translation (MyMemory engine, cached; over-budget content is caught up the next day). |
| output_dir | reports | 报告输出目录（相对路径基于脚本所在目录）。Report output directory (relative paths resolve against the script). |
| auto_open | true | 未配置任何 Webhook 时，保存报告后自动用浏览器打开 HTML。With no webhook configured, auto-open the HTML report in your browser after saving. |
更看重「刚创建就快速涨星」，同时保留体量参考： It favors repos that gain stars fast right after creation, while keeping volume in play:
仓库年龄下限 1.0 天，避免除零与超新仓库失真；想更偏体量就调大 w2 / w3，更偏增速就调大 w1。 Repo age is clamped to a 1.0-day minimum to avoid division blow-ups; raise w2 / w3 to favor volume, raise w1 to favor growth speed.
直接用 HTTP 调用 {base_url}/chat/completions，不依赖 openai SDK；LLM 调用失败自动回退本地规则摘要： Calls {base_url}/chat/completions over plain HTTP with no openai SDK; a failed LLM call falls back to the local rule-based summarizer:
模型按 JSON 返回 core_features / tech_stack / use_cases / highlight 四个字段；终端会统计「LLM 提炼 N 个，规则兜底 M 个」。 The model returns four JSON fields — core_features / tech_stack / use_cases / highlight — and the terminal reports how many summaries came from the LLM versus the fallback rules.
以下内容摘自 reports/github_trending_20261024_2001.md——2026-10-24 20:01 由计划任务自动生成，数据未做任何修改。工具默认输出中文报告（免费翻译已开启）。 Excerpted from reports/github_trending_20261024_2001.md — generated automatically by the scheduled task at 20:01 on 2026-10-24, completely unedited. The tool produces reports in Chinese by default (free translation enabled).
| # | 项目 | Star | Fork | 语言 | 热度分 | 一句话亮点 |
|---|---|---|---|---|---|---|
| 1 | MemTensor/MemT-1.0 | 382+16 | 17 | Python | 145.9 | MemT-1.0：扩展和应用。研究记忆…… |
| 2 | h9-tec/ai-roadmap-2026 | 148+18 | 14 | 未知 | 101.9 | ai-roadmap-2026：逐层。选择一个图层，…… |
| 3 | liyupi/dafeiyu-cleaner | 55+26 | 1 | JavaScript | 83.5 | dafeiyu-cleaner: 一眼认出大肥鱼：评论一加载就鉴定… |
| 4 | QuickButlerPace65/AnythingLLM-Desktop🆕 新上榜 | 37 | 12 | 未知 | 76.1 | AnythingLLM-Desktop：🧳真正的便携性：无插件…… |
| 5 | C4rohan/Lampline | 171+7 | 30 | Python | 68.5 | Lampline：让您的代理商从Dialogflo过来…… |
| 6 | Umair/stalemap | 100+1 | 41 | TypeScript | 45.4 | stalemap：编码代理从tr学习库…… |
描述: MemT-1.0 的官方存储库
一句话亮点：MemT-1.0：扩展和应用。研究记忆……
【核心功能】
【技术栈】
【适用场景】
样例仅节选榜单前 6 名与 1 张详情卡；完整报告含 Top 12 全部条目，HTML 版为暗黑卡片风单文件。 This excerpt shows only the top 6 rows and one detail card; the full report covers all Top 12 entries as a dark-themed single-file HTML.
clone、装依赖、双击 install_task.bat——之后每天的早报都会安静地出现在 reports/ 目录里。 Clone it, install the dependencies, double-click install_task.bat — from then on, every day's report quietly lands in the reports/ directory.
回到安装步骤Back to the install steps