2026-09-28 GitHub 趋势研究简报
证据边界: 项目名称、星标、tags、description、license、size、forks 来自 2026-09-28 GitHub Search API 公开元数据(created_at / pushed_at / stargazers_count / forks_count / language / description / topics / license / size)+ 各项目 README 公开摘录(API readme 字段 base64 解码)。"X 天 Y⭐"基于 created_at→2026-09-28 总星数除以经过天数(粗略下限估计),stargazers REST endpoint 需要认证故未做精确单日增量统计。今日观察窗为 2026-09-25 ~ 2026-09-27 创建且截至 2026-09-28 仍在快速增长的 5 个高质量工程型项目;09-28 当日新创仓库在 GitHub API 暴露窗口 < 12 小时,本批重点放在 09-25 ~ 09-27 创建且仍持续增长的项目(04 日延续跟踪:Contrastive-LM/CLM / mikehasa/golive-skill / yetone/magpie / anishfn/shapeshift / samyost1/3dicon 五项目均处于 3 ~ 4 天增长窗口,今日新增 dzhng/jevgrep / feitangyuan/onetake / supermemoryai/company-brain / amitshekhariitbhu/ai-system-design 四项目加入构成「Jev 决策模型 CLI 工具层 + Agent Skill 视觉生成严肃工程化 + 商用 Slack 副驾开源 + AI 系统设计教程」严肃工程化三层)。今日最大事件是
dzhng/jevgrep2 天 699⭐ / fork 45 / fork/star 6.4% — 把 Jev / TypeSafe AI System One 决策模型从昨日 09-25 「对比学习 states↔actions + 缓存独立 + TypeSafe 兼容 API + 9× Jev 速度 + Terminal-Bench 87.6% + DeepSWE 81.6% SOTA verifier」推到「CLI 工具层 + coding agent 上下文检索 + Vercel AI Gateway / TypeSafe / OpenRouter / OpenCode Zen 四 provider + Claude Code / Codex / OpenCode 多 harness 自动检测 + Python / TypeScript declaration 解析 +npx skills add dzhng/jevgrep --skill jevgrep二次分发」严肃工程化形态——coding agent 在陌生任务上花时间找对文件,jevgrep 给它们一个起点:问仓库问题,jg返回相关文件 + 阅读线索 + 原代码摘录在一次 stdout 响应里。它用 Jev 在 folders / files / declarations 间判断相关性。coding agent 然后实现并测试变更。npm install -g @dzhng/jevgrep+jg auth+jg skill+jg "How are telemetry events recorded and sent?" ./my-project一行装 + 一行调用。需要 Node.js 22+ + macOS 或 Linux + Vercel AI Gateway / TypeSafe / OpenRouter / OpenCode Zen 四 provider 之一的 key。与昨日「Jev 决策模型多后端并存 + Agent Skill 严肃工程化跨工作流 + 跨 Agent 模型统一网关」三主线不同,今日主线转向「Jev / System One 决策模型推到 CLI / 工具层 + 多 provider + 多 harness 严肃工程化」——把 09-22 Rizzo-AI-Academy/rizzo-flow「Spark-X2.5-4B Apache-2.0 + Jev 兼容 API + 0 generated tokens」+ 09-20 logan-markewich/jeff「GLiFormer 400M Apache-2.0 + Jev 兼容 API + 自托管 drop-in」+ 09-23 deepopen-com/deepopen「Non-Auto Decision Engine Laya 改进 + 三检查点 + 7.8× Jev 速度」+ 09-25 Contrastive-LM/CLM「CLM-8B 对比学习 + 9× Jev + SOTA verifier」+ 09-26 anishfn/shapeshift「Jev 应用层 UI 端 18 类卡片」+ 09-22 fstandhartinger/chat-seek-vscode「VS Code 跨 CLI 聊天本地检索 + Laya reranking」六条 Jev 决策模型路径同步并存推到「CLI 工具层 + Vercel AI Gateway / TypeSafe / OpenRouter / OpenCode Zen 四 provider + Claude Code / Codex / OpenCode 多 harness 自动检测 + Python / TypeScript declaration 解析 + Node.js 22+ 一行装 +npx skills add二次分发 + summary + file locations + reading leads + verbatim source excerpts」严肃工程化形态——意味着 Jev / System One 决策模型生态从「决策 API(09-15 ~ 09-20)+ 应用层 UI 端具身(09-22 ~ 09-26)+ 可执行技能合集(09-21)+ 跨 CLI 检索(09-22)+ Android 聊天副驾(09-22 ~ 09-23)→ CLI 工具层 + 多 provider + 多 harness 自动检测 + declaration 解析 + 二次分发(09-28)」演化到全栈严肃工程化形态——CLI 工具层补齐了「coding agent 上下文检索」的最后一块,jevgrep 给「Jev 决策模型」补上「Agent 实际拿上下文」的最后一环。
今日重点趋势(按排名)
1. dzhng/jevgrep — Jev / System One 决策模型推到 CLI / 工具层 + 多 provider + 多 harness 严肃工程化(趋势分 92)
- dzhng/jevgrep(2 天 699⭐ / fork 45 / fork/star 6.4% / TypeScript / MIT / 4143 KB)—— 「Find code by asking what it does. A CLI for coding agents that uses Jev to discover relevant files and source context.」;定位 「CLI for coding agents + Vercel AI Gateway / TypeSafe / OpenRouter / OpenCode Zen 四 provider + Claude Code / Codex / OpenCode 多 harness 自动检测 + Python / TypeScript / JavaScript declaration 解析 + Node.js 22+ 一行装 +
npx skills add二次分发 + summary + file locations + reading leads + verbatim source excerpts」;架构npm install -g @dzhng/jevgrep+jg auth+jg skill+jg "How are telemetry events recorded and sent?" ./my-project+ Node.js 22+ + macOS / Linux + Vercel AI Gateway / TypeSafe / OpenRouter / OpenCode Zen 四 provider + Claude Code / Codex / OpenCode 多 harness 自动检测 +jg skill二次分发(npx skills add dzhng/jevgrep --skill jevgrep)+ Python / TypeScript / JavaScript declaration 解析 + 其它文本 fallback + summary + file locations + reading leads + verbatim source with line references + evidence not generated answer + 4-point quality criteria;核心机制——Find code by asking what it does——coding agent 在陌生任务上花时间找对文件,jevgrep 给它们一个起点:问仓库问题,jg返回相关文件 + 阅读线索 + 原代码摘录在一次 stdout 响应里;Uses Jev to judge relevance——用 Jev 在 folders / files / declarations 间判断相关性;Provider selection 0.3.0+——四 provider 任选:Vercel AI Gateway / TypeSafe / OpenRouter / OpenCode Zen;Install the agent skill——装 CLI 不会教 agent 用,必须jg skill也装;skill 教 agent 何时调jg+ 怎么用返回的 context + 何时用 normal tools 填空;skips redundant retrieval when context already known;Multi-harness auto-detect——installer 检测 Claude Code / Codex / OpenCode 等 coding agents 并问装哪里;--global用户级 /--yes无提示;Two install paths——jg skill委托给 skills CLI;或直接npx skills add dzhng/jevgrep --skill jevgrep;Upgrade——npm install -g @dzhng/jevgrep@latest升级 CLI;skill 需重跑jg skill;Start with a question——jg "Where is authentication checked before a request reaches a handler?" ./jg "How are database connections created, pooled, and closed?" ./src/jg "Which tests cover retry behavior when a request times out?" .;Output——summary 先来 + file locations + reading leads + selected source with line references;Python + TypeScript / JavaScript declaration 解析;其它文本 fallback;keeps qualifying file locations when excerpt 不自信;No separate runtime——No separate Python / Bun / ripgrep 安装 to usejg;Quality——output 是给 agent 用的 evidence,不是 generated answer 或 guaranteed complete search;topics——ai-sdk,claude-code,cli,code-search,codex,coding-agents,context-retrieval,developer-tools,jev,semantic-search,typescript,vercel-ai-gateway;4143 KB——TypeScript 中等 CLI;MIT;fork/star 6.4%——较健康 CLI 严肃工程化持续关注信号;699⭐ / 45 forks(09-28 当日较 09-26 推送持续)反映 CLI 工具层严肃工程化持续关注。 - 核心价值定位: 当前 coding agent 上下文检索赛道的痛点是 「agent 写陌生任务靠 grep + 读文件猜 + 不查 declarations + 在 README 里搜索 + 不知道哪 file 真相关 + 上下文质量不稳 + 一查半仓库 + 多 harness 配置分散」 ——jevgrep 用「CLI for coding agents + Jev 判断多 file 相关性 + 四 provider + 多 harness auto-detect + declaration 解析 + evidence not generated answer + npx skills 二次分发」是「Jev 决策模型 CLI 工具层 + 严肃工程化 + coding agent 上下文检索」的具体路径;Find code by asking what it does——agent 问仓库问题,
jg返回 file locations + reading leads + verbatim source with line references 一次 stdout;Jev 在 folders / files / declarations 间判断相关性——把 Jev / System One 决策模型从「决策 API」推到「CLI 工具层」严肃工程化;四 provider——Vercel AI Gateway / TypeSafe / OpenRouter / OpenCode Zen 任选;多 harness auto-detect——Claude Code / Codex / OpenCode 自动检测 +--global/--yes标志;二次分发——npx skills add dzhng/jevgrep --skill jevgrep二次分发;declaration 解析——Python + TypeScript / JavaScript;evidence not generated answer——output 是给 agent 用的 evidence;Node.js 22+——一行装;MIT 4143 KB——TypeScript 中等 CLI;fork/star 6.4%——较健康 CLI 严肃工程化信号。 - 架构亮点:
npm install -g @dzhng/jevgrep——一行装;jg auth——provider 认证;jg skill——agent skill 安装;jg "..." ./my-project——一行调用;Node.js 22+ + macOS / Linux——环境;Vercel AI Gateway / TypeSafe / OpenRouter / OpenCode Zen 四 provider——任选;Provider selection 0.3.0+——版本门槛;Claude Code / Codex / OpenCode multi-harness auto-detect——检测后问装哪里;--global/--yes——用户级 / 无提示;npx skills add dzhng/jevgrep --skill jevgrep——二次分发;Python / TypeScript / JavaScript declaration 解析——AST 级;其它文本 fallback——不抛错;summary 先 + file locations + reading leads + verbatim source with line references——output 结构;keeps qualifying file locations when excerpt 不自信——不强制 top-2;evidence not generated answer——不做 LLM 写答案;Quality 4 criteria——evidence-based / skip redundant / declaration-aware / not generated answer;topics 12 个——ai-sdk,claude-code,cli,code-search,codex,coding-agents,context-retrieval,developer-tools,jev,semantic-search,typescript,vercel-ai-gateway;MIT 4143 KB——TypeScript 中等 CLI。 - 价值判断: 它解决的是「coding agent 上下文检索 + Jev 决策模型 CLI 工具层 + 多 provider + 多 harness auto-detect + declaration 解析 + evidence not generated answer + npx skills 二次分发」的真痛点;TypeScript + MIT + 4143 KB + Node.js 22+ + Vercel AI Gateway / TypeSafe / OpenRouter / OpenCode Zen 四 provider + Claude Code / Codex / OpenCode 多 harness 自动检测 +
npx skills add二次分发 + Python / TS / JS declaration 解析 + summary + file locations + reading leads + verbatim source with line references + evidence not generated answer 是「Jev 决策模型 CLI 工具层 + 严肃工程化」形态;699⭐ / fork 45 / fork/star 6.4% / 2 天反映 Jev 决策模型 CLI 工具层严肃工程化持续关注(fork/star 6.4% 与昨日 09-25/26 的 shapeshift 9.4% / magpie 4.8% 同步,反映「CLI 工具层 + 严肃工程化 + coding agent 上下文检索」的典型严肃工程化 fork 率特征;45 forks 中包含「CLI 工具层严肃工程化 fork + coding agent 用户 fork + Jev 决策模型挑战者 fork + Vercel AI Gateway 用户 fork + 多 harness 用户 fork」五类);决定这条主线长期价值的是「Jev 在 folders / files / declarations 间判断相关性在多任务的稳定性 + 四 provider(Vercel AI Gateway / TypeSafe / OpenRouter / OpenCode Zen)的兼容 + 多 harness(Claude Code / Codex / OpenCode / Cursor / Copilot)auto-detect 的覆盖率 + declaration 解析在 Python / TS / JS 的准确率 + 其它文本 fallback 的稳定性 + summary + file locations + reading leads + verbatim source output 结构的实用性 + evidence not generated answer 边界的清晰度 +jg skill在npx skills add的二次分发兼容性 + Node.js 22+ 在多平台的稳定性 + 4-point quality criteria 在实际使用的边界 + topics 12 个覆盖(ai-sdk / claude-code / cli / code-search / codex / coding-agents / context-retrieval / developer-tools / jev / semantic-search / typescript / vercel-ai-gateway)的清晰度」——Jev 多任务稳定性 + 四 provider 兼容 + 多 harness auto-detect + declaration 解析 + 其它文本 fallback + output 实用性 + evidence 边界 + 二次分发兼容 + Node.js 22+ 兼容性 + 4-point quality + topics 覆盖是关键;对 coding agent 用户:可一行jg "..." ./my-project拿上下文检索;对 Jev / TypeSafe 生态:是「Jev 决策模型 CLI 工具层 + 严肃工程化」的具体路径;对 Vercel AI Gateway / OpenRouter 用户:是「多 provider + coding agent 上下文检索 + Jev 决策模型」的具体路径;对组织 / 企业:是 MIT 开源 + 多 provider + 多 harness + declaration 解析严肃工程化参考。
2. feitangyuan/onetake — Claude Agent Skill「一镜到底」连贯动效做产品发布片(趋势分 88)
- feitangyuan/onetake(2 天 529⭐ / fork 40 / fork/star 7.6% / Python / PolyForm Noncommercial 1.0.0 / 77237 KB)—— 「Motion films that never cut to the next slide. A Claude Agent Skill that makes product launch films, teasers and feature demos where every beat grows out of the one before — one continuous take, not a stack of scenes.」;定位 「Claude Agent Skill 一镜到底连贯动效 + boundary 设计 carry 跨镜头 + oracle 自动测 carry score + ~38 移动纯函数 + 同一相机不切 + real UI HTML 重建 + PolyForm Noncommercial 商用限制」;架构 Python · PolyForm Noncommercial 1.0.0 · 77237 KB · Agent Skill SKILL.md · 10 cases · probe.py oracle 测每帧边界 carry score 0.00 / 0.40 / 0.75 / 0.83 · ~38 moves (springs / entrances / carries / contact / sims / camera / fluid grounds) 各自为时间纯函数 · 同一相机不切镜头 · real UI rebuilt in HTML from screenshots on camera's plane · screen recording 不用 · uniform cadence / no stillness / clipped audio / fast moves without motion blur / subject leaves frame 五项 oracle 自动失败;核心机制——The problem: motion that falls apart——Most motion videos AI makes today (and most template tools) come out loose: title card / fade / UI shot / fade / feature card / fade / logo; each shot can look fine and the whole still reads like a slideshow, because nothing connects one beat to the next: scenes replace each other;What onetake does differently: every beat is carried——onetake treats the boundary between two beats as the thing to design; at every boundary, something on screen survives and visibly becomes the next beat: prompt bar opens into the app window; window reflows into the phone; bar collapses into a line, the line shoots across the desk and becomes the first grid line of the next film; card grows into the page; stage opens from the character; camera pushes through a card until it is the next scene; at the end, the cast folds into the lockup — in the launch film, the block cursor writes the name in one stroke;One camera holds it all together——never cuts; follows the subject a beat ahead; whips to where the next thing will land; floats like a hand-held operator; shakes when something hits; then holds dead still, because rests are what make the moves land;Continuity is measured, not hoped for——every film is checked by an oracle before a human watches it;
probe.pyrecords what is on screen in every frame, finds each boundary, and asks what carried across it; a film whose beats replace each other fails — same oracle fails uniform cadence / no stillness / clipped audio / fast moves without motion blur / subject that leaves the frame;What makes it look finished——Real UI, rebuilt——the product's interface is rebuilt in HTML from screenshots and sits on the camera's plane, so the camera can fly into it at any zoom; no screen recording;Measured moves——~38 moves (springs / entrances / carries / contact / sims / camera / fluid grounds), each a pure function of time;Cases——10 films covering launch films / teasers / feature demos;Verify——scripts/verify_promo.pyper-boundary carry score;PolyForm Noncommercial 1.0.0——any noncommercial use OK; commercial use needs separate license。 - 核心价值定位: 当前 Agent Skill 视觉生成赛道的痛点是 「一个 prompt 想要产品发布片但没有连贯性 + 镜头切得碎像 PPT + 边界不设计 carry 跨镜头 + 没有 oracle 自动测连续性 + 用 screen recording 而非 HTML 重建 UI + uniform cadence / no stillness / clipped audio / fast moves without motion blur / subject leaves frame 五项 quality issues 没人测 + 商用限制不清晰」 ——onetake 用「Claude Agent Skill + 一镜到底 + boundary 设计 carry 跨镜头 + oracle 自动测 carry score + ~38 移动纯函数 + 同一相机不切 + real UI HTML 重建 + PolyForm Noncommercial 商用限制」是「Agent Skill 视觉生成 + 一镜到底 + 严肃工程化 + 商用清晰」的具体路径;The problem: motion that falls apart——Most motion videos come out loose; scenes replace each other;every beat is carried——boundary between two beats as the thing to design; something on screen survives and visibly becomes the next beat;One camera——never cuts; follows the subject a beat ahead; whips to where the next thing will land; floats like a hand-held operator; shakes when something hits; holds dead still for rests;Continuity is measured, not hoped for——probe.py oracle per-boundary carry score 0.00 / 0.40 / 0.75 / 0.83;Real UI, rebuilt——HTML from screenshots on camera's plane so camera can fly into it at any zoom; no screen recording;Measured moves——~38 moves, each a pure function of time;Five auto-fail checks——uniform cadence / no stillness / clipped audio / fast moves without motion blur / subject leaves frame;PolyForm Noncommercial 1.0.0——any noncommercial use OK; commercial use needs separate license。
- 架构亮点: Agent Skill SKILL.md——Claude Code skill;10 cases——launch films / teasers / feature demos;probe.py oracle——per-frame boundary carry score;carry score 0.00 / 0.40 / 0.75 / 0.83——launch film v1 / v2 / v3 / onetake launch film;~38 moves——springs / entrances / carries / contact / sims / camera / fluid grounds;Each a pure function of time——纯函数化动画;One camera——never cuts; follows / whips / floats / shakes / holds;Real UI rebuilt in HTML——from screenshots on camera's plane;No screen recording——HTML rebuild;5 auto-fail checks——uniform cadence / no stillness / clipped audio / fast moves without motion blur / subject leaves frame;PolyForm Noncommercial 1.0.0——noncommercial OK; commercial separate license;scripts/verify_promo.py——per-boundary verify;77237 KB——Python 中大型项目,含 assets / cases / scripts;fork/star 7.6%——较健康 Agent Skill 严肃工程化持续关注信号;529⭐ / 40 forks(09-28 当日较 09-26 推送持续)反映 Agent Skill 视觉生成 + 一镜到底严肃工程化持续关注。
- 价值判断: 它解决的是「Agent Skill 视觉生成 + 一镜到底 + boundary 设计 carry 跨镜头 + oracle 自动测 carry score + ~38 移动纯函数 + 同一相机不切 + real UI HTML 重建 + PolyForm Noncommercial 商用限制」的真痛点;Python + PolyForm Noncommercial 1.0.0 + 77237 KB + Agent Skill SKILL.md + 10 cases + probe.py oracle + ~38 moves + real UI HTML 重建 + 5 auto-fail checks 是「Agent Skill 视觉生成 + 一镜到底严肃工程化」形态;529⭐ / fork 40 / fork/star 7.6% / 2 天反映 Agent Skill 视觉生成 + 一镜到底严肃工程化持续关注(fork/star 7.6% 与昨日 09-26 golive-skill 7.1% 同步,反映「Agent Skill + 视觉生成 + 一镜到底严肃工程化」的典型严肃工程化 fork 率特征;40 forks 中包含「Agent Skill 视觉生成 fork + 一镜到底 motion graphics fork + Claude Agent Skill 用户 fork + PolyForm Noncommercial 商用清晰 fork + oracle carry score fork」五类);决定这条主线长期价值的是「boundary 设计 carry 跨镜头在多 motion styles 的稳定性 + probe.py oracle 在 per-frame boundary carry score 的准确率 + ~38 移动纯函数在多视觉风格的扩展 + 同一相机不切在长视频的实用性 + real UI HTML 重建在多 zoom 的兼容性 + 5 auto-fail checks(uniform cadence / no stillness / clipped audio / fast moves without motion blur / subject leaves frame)在多视频的覆盖率 + PolyForm Noncommercial 1.0.0 商用清晰的边界 + Claude Code / Codex / Cursor 多 harness 的兼容性 + 10 cases 在多场景的覆盖度 + scripts/verify_promo.py 在 CI 的可用度」——boundary carry 稳定性 + oracle 准确率 + ~38 移动扩展 + 一相机实用性 + HTML 重建兼容性 + 5 auto-fail 覆盖率 + PolyForm 商用清晰 + 多 harness 兼容 + 10 cases 覆盖度 + CI 可用度是关键;对 Agent Skill 视觉生成用户:可一行 prompt 一镜到底连贯动效做产品发布片;对 Agent Skill 严肃工程化:是「视觉生成 + 一镜到底 + oracle carry score + ~38 移动纯函数 + 同一相机 + HTML 重建 + PolyForm Noncommercial」的具体路径;对组织 / 企业:是 PolyForm Noncommercial 商用清晰严肃工程化参考。
3. supermemoryai/company-brain — 商用 Slack 记忆副驾开源 + Cloudflare Workers 部署(趋势分 85)
- supermemoryai/company-brain(3 天 664⭐ / fork 93 / fork/star 14.0% / TypeScript / Apache-2.0 / 13452 KB)—— 「Open-sourcing our company brain - A teammate in your Slack that remembers everything your team says, and can go do the work.」;定位 「商用 Slack 记忆副驾开源 + Cloudflare Workers 部署 + Slack 原生 teammate + permissions graph + 主动回答 + 主动执行 + Skills + workspace prompt + 七个能力(Remembers / Answers from your team / Acts in your tools / Speaks up / Runs code / Works while you sleep / Learns your way)」;架构 TypeScript · Apache-2.0 · 13452 KB · Runs on Cloudflare Workers · memory by supermemory · Apache-2.0 fully open source · One-click deploy button · permissions graph · 七个能力 · Skills teach it your processes / formats / voice · workspace prompt sets how it behaves everywhere · 一键 deploy 到 own Cloudflare account · GitHub / Linear / Notion / Google Workspace + MCP hundreds more;核心机制——A teammate in your Slack that truly knows and understands your company, and can do anything——Your team's knowledge is scattered across Slack threads / docs / tickets / people's heads; Company Brain sits in Slack / remembers what your team says / answers from it; Then it goes and does the work: opens the issue / reads the PR / digs into the repo; And when a conversation needs something it knows, it speaks up on its own;Previously paid, now free and open source——Used to be a paid product with thousands of users; Now it's free and open source under Apache-2.0;Runs on Cloudflare Workers——on your own Cloudflare account;Private by design——memory isn't one big bucket; it's a permissions graph;seven abilities——🧠 Remembers decisions / projects / owners / context from the channels it's in, kept current as people talk; no one has to write anything down;💬 Answers from your team not from the internet, not from guesswork: from your own conversations and connected tools;🛠️ Acts in your tools GitHub / Linear / Notion / Google Workspace + hundreds more over MCP; it opens the issue instead of telling you to;📣 Speaks up chimes in when a conversation needs something it knows; you decide how chatty it is, org-wide and per channel;💻 Runs code its own sandbox: clones repos / runs scripts / hands back charts / CSVs / PDFs right in Slack;🌙 Works while you sleep scheduled digests to a channel or your DMs, plus research it kicks off on its own;🎓 Learns your way skills teach it your processes / formats / voice; a workspace prompt sets how it behaves everywhere;Use case——「Alex: is prod down? a couple of customers are pinging me」/「Company Brain: Not fully down. api/search is elevated, 42 errors in the last 15 minutes, and Kush is already on it. Probably this morning's deploy.」——Nobody mentioned the bot. It spoke up because it knew the answer。
- 核心价值定位: 当前商用 Slack 副驾赛道的痛点是 「团队知识散落在 Slack / docs / tickets / 人脑 + 谁来记 + 找时找不到 + 知道时不说 + 重复问 + 重复答 + 不主动 + 不会主动做 + 不接工具 + 没有 permissions graph + 没有 skills + 没有 workspace prompt + 没有自己的 sandbox + 不跑代码 + 没有 scheduled digest」 ——company-brain 用「商用产品停服 → Apache-2.0 完全开源 + Cloudflare Workers 部署 + Slack 原生 teammate + permissions graph + 主动回答 + 主动执行 + Skills + workspace prompt + 七个能力」是「商用副驾开源 + 严肃工程化 + Slack 原生 + 主动 + 接工具」的具体路径;Previously paid, now free and open source——thousands of users, now Apache-2.0;Runs on Cloudflare Workers——on your own Cloudflare account;A teammate in your Slack that truly knows and understands your company, and can do anything——sits in Slack / remembers / answers / goes and does the work / speaks up on its own;Private by design——memory isn't one big bucket; it's a permissions graph;seven abilities——Remembers / Answers from your team / Acts in your tools / Speaks up / Runs code / Works while you sleep / Learns your way;GitHub / Linear / Notion / Google Workspace + hundreds more over MCP——tool integrations;Use case——「prod down?」/「api/search elevated, 42 errors last 15min, Kush already on it, probably this morning's deploy」;Apache-2.0 13452 KB——TypeScript 大型项目;fork/star 14.0%——典型商用开源 + 高 fork 严肃工程化持续关注信号;664⭐ / 93 forks(09-28 当日较 09-25 创建持续)反映商用 Slack 副驾严肃工程化持续关注。
- 架构亮点: Apache-2.0——fully open source;Cloudflare Workers——runs on Cloudflare Workers;One-click deploy button——on your own Cloudflare account;permissions graph——memory is not one big bucket;seven abilities——Remembers / Answers from your team / Acts in your tools / Speaks up / Runs code / Works while you sleep / Learns your way;Skills——teach it your processes / formats / voice;workspace prompt——sets how it behaves everywhere;GitHub / Linear / Notion / Google Workspace + hundreds more over MCP——tool integrations;its own sandbox——clones repos / runs scripts / hands back charts / CSVs / PDFs in Slack;scheduled digests——to a channel or DMs;research it kicks off on its own——proactive;decide how chatty——org-wide and per channel;13452 KB——TypeScript 大型项目;Apache-2.0——商用清晰;memory by supermemory——already built;Discord community——supermemory.link/discord;fork/star 14.0%——典型商用开源 + 高 fork 严肃工程化持续关注信号。
- 价值判断: 它解决的是「商用 Slack 副驾开源 + Cloudflare Workers 部署 + Slack 原生 teammate + permissions graph + 主动回答 + 主动执行 + Skills + workspace prompt + 七个能力」的真痛点;TypeScript + Apache-2.0 + 13452 KB + Cloudflare Workers + Slack 原生 + permissions graph + 七个能力 + Skills + workspace prompt + MCP 集成 + 自己的 sandbox + scheduled digests 是「商用副驾开源 + 严肃工程化」形态;664⭐ / fork 93 / fork/star 14.0% / 3 天反映商用副驾开源 + 严肃工程化持续关注(fork/star 14.0% 与昨日 09-26 shapeshift 9.4% / 09-25 yetone/magpie 4.9% 形成对比,反映「商用开源 + 高 fork + 大型项目 + Apache-2.0 + Cloudflare Workers 部署」的典型严肃工程化 fork 率特征;93 forks 中包含「商用副驾用户 fork + Apache-2.0 严肃工程化 fork + Cloudflare Workers fork + permissions graph fork + Skills + workspace prompt fork + MCP 集成 fork」六类);决定这条主线长期价值的是「Slack 原生 teammate 在多 workspace 的部署稳定性 + permissions graph 在多用户的边界清晰度 + 主动回答在 prod down 等紧急情况的实用性 + 主动执行在 GitHub / Linear / Notion 等工具的兼容性 + Skills 在多 processes / formats / voice 的扩展 + workspace prompt 在多用户的可用度 + 自己的 sandbox 在多代码任务的稳定性 + scheduled digests 在多时间的实用性 + MCP 集成在 GitHub / Linear / Notion / Google Workspace 多工具的稳定性 + Cloudflare Workers 在多账户的部署兼容性 + Apache-2.0 在企业的商用清晰 + Discord 社区在多用户的活跃度」——Slack 原生部署稳定性 + permissions graph 边界 + 主动回答实用性 + 主动执行工具兼容性 + Skills 扩展 + workspace prompt 可用度 + sandbox 稳定性 + scheduled digests 实用性 + MCP 工具兼容性 + Cloudflare Workers 兼容性 + Apache-2.0 商用清晰 + Discord 社区活跃度是关键;对团队 / 企业:可一键 deploy 到自己 Cloudflare account + Apache-2.0 商用清晰 + 团队记忆 + 主动回答 + 主动执行;对商用副驾严肃工程化:是「商用产品停服 → Apache-2.0 完全开源 + Cloudflare Workers + 七个能力 + Skills + workspace prompt + MCP」的具体路径;对 Slack 重度用户:是「Slack 原生 teammate + permissions graph + 主动」的具体路径。
4. anishfn/shapeshift — Jev 应用层 UI 端具身形态持续 4 日增长(趋势分 82)
- anishfn/shapeshift(4 天 759⭐ / fork 71 / fork/star 9.4% / TypeScript / MIT / 12116 KB)—— 「An input that becomes what you mean. One text box that morphs into the right UI as you type — an event card, a checklist, a timer, a color picker, a bill splitter, a poll, a converter and more.」;定位 「Jev decides, code computes」——用 Jev 一次判断 14 个 typed questions 决定走哪张卡片 + 确定性代码算日期 / 金额 / 单位 / 数学 + 内置离线关键词分类器兜底;架构 Bun 1.2+ + Next.js + Three.js + Vite + Excalidraw 图表 + 18 类卡片 + 滞回状态机 challenger 赢两次才换 + URL 参数 ?debug=1 / ?demo=1&loop=1 / ?shape=event 强制卡片类型;持续 4 日增长:09-25 387⭐ / fork 43 → 09-26 566⭐ / fork 53 → 09-27 759⭐ / fork 71 持续 3 日增长(fork/star 9.4% 与昨日 09-26 9.4% 同步,典型高质量 UI 持续关注信号)。详细档案见
projects/shapeshift.md。 - 核心价值定位: 与昨日同构;关键新增:今日观察窗为 09-25 ~ 09-27 创建且截至 09-28 仍在快速增长,shapeshift 4 天 759⭐ / fork 71 是「Jev 决策模型应用层 UI 端具身」持续严肃工程化形态的具体路径。
- 价值判断: 与昨日 09-26 同;今日 759⭐ / fork 71 / fork/star 9.4% / 4 天反映 Jev 应用层 UI 端具身 + 严肃工程化 + 持续关注,fork/star 9.4% 较健康典型持续早期 fork 率特征。
5. amitshekhariitbhu/ai-system-design — AI 系统设计 step-by-step 教程合集持续 3 日增长(趋势分 76)
- amitshekhariitbhu/ai-system-design(3 天 448⭐ / fork 51 / fork/star 11.4% / Markdown / Apache-2.0 / 145 KB)—— 「AI System Design - Learn how to design AI systems built on LLMs, RAG, and AI Agents step by step.」;定位 「Markdown 教程合集 + step-by-step + Apache-2.0 商用清晰 + LLM + RAG + AI Agents 全栈设计」;架构 Markdown · Apache-2.0 · 145 KB · Topics: ai / ai-agents / ai-engineering / ai-system / ai-system-design / ai-systems / ai-systems-design / llm / system-design / system-design-interview;核心价值——Markdown 教程合集 + step-by-step——LLM + RAG + AI Agents 系统设计教程;topics 10 个——AI / ai-agents / ai-engineering / ai-system / ai-system-design / ai-systems / ai-systems-design / llm / system-design / system-design-interview;Apache-2.0 145 KB——纯 Markdown 极小项目;fork/star 11.4%——较健康 AI 系统设计教程严肃工程化持续关注信号;448⭐ / 51 forks(09-28 当日较 09-25 创建持续)反映 AI 系统设计教程严肃工程化持续关注。
- 核心价值定位: 当前 AI 系统设计学习赛道的痛点是 「AI 系统设计教程分散 + LLM / RAG / AI Agents 各自文档 + 没 step-by-step + 没 Apache-2.0 商用清晰 + 没 system-design-interview 视角 + 没 ai-engineering 工程化视角」 ——ai-system-design 用「Markdown 教程合集 + step-by-step + Apache-2.0 商用清晰 + LLM + RAG + AI Agents 全栈设计 + system-design-interview 视角 + ai-engineering 工程化视角」是「AI 系统设计教程 + 严肃工程化 + step-by-step」的具体路径。
- 架构亮点: Markdown——纯文档;Apache-2.0——商用清晰;145 KB——极小项目;topics 10 个——AI / ai-agents / ai-engineering / ai-system / ai-system-design / ai-systems / ai-systems-design / llm / system-design / system-design-interview;step-by-step——学习路径;LLM + RAG + AI Agents 全栈——覆盖广度;fork/star 11.4%——较健康 AI 系统设计教程严肃工程化持续关注信号。
- 价值判断: 它解决的是「AI 系统设计教程 + step-by-step + Apache-2.0 + LLM + RAG + AI Agents 全栈 + system-design-interview + ai-engineering」的真痛点;Markdown + Apache-2.0 + 145 KB + step-by-step + LLM + RAG + AI Agents 全栈是「AI 系统设计教程 + 严肃工程化」形态;448⭐ / fork 51 / fork/star 11.4% / 3 天反映 AI 系统设计教程严肃工程化持续关注;决定这条主线长期价值的是「step-by-step 教程在多主题(LLM / RAG / AI Agents)的覆盖广度 + Apache-2.0 商用清晰的边界 + topics 10 个覆盖的清晰度 + system-design-interview 视角在多问题的实用性 + ai-engineering 工程化视角在多场景的覆盖度 + Markdown 极小项目在多读者的可读性」——step-by-step 覆盖广度 + Apache-2.0 商用清晰 + topics 覆盖 + system-design-interview 实用性 + ai-engineering 覆盖度 + Markdown 可读性是关键。
最值得关注的方向
1. Jev / System One 决策模型「CLI 工具层 + 多 provider + 多 harness 严肃工程化」补齐最后一块
09-15 ~ 09-27 十三日内 Jev / System One 决策模型从「决策 API + SDK + 资源聚合 + 去 SaaS 化 + 具身智能 + 可执行技能合集 + 多后端并存 + 应用层 UI 端具身 + 跨 CLI 检索 + Android 聊天副驾」推到「CLI 工具层 + 多 provider + 多 harness 严肃工程化」:
- 09-15 ~ 09-20
logan-markewich/jeff(GLiFormer 400M Apache-2.0 + Jev 兼容 API + 自托管 drop-in replacement) - 09-22
Rizzo-AI-Academy/rizzo-flow(Spark-X2.5-4B / 1.7B Apache-2.0 + Jev 兼容 API + 0 generated tokens + Snake demo) - 09-22
fstandhartinger/chat-seek-vscode(VS Code 跨 CLI 聊天本地检索 + Laya reranking) - 09-23
deepopen-com/deepopen(Non-Auto Decision Engine Laya 改进 + 三检查点 + 7.8× Jev 速度) - 09-25
Contrastive-LM/CLM(CLM-8B 对比学习 + 9× Jev + SOTA verifier + Apache-2.0) - 09-26
anishfn/shapeshift(Jev 应用层 UI 端 18 类卡片 + 滞回状态机) - 09-28
dzhng/jevgrep(CLI for coding agents + Jev 判断多 file 相关性 + 四 provider + 多 harness auto-detect + declaration 解析 + npx skills 二次分发)
今日 dzhng/jevgrep 把 13 日 Jev 决策模型生态从「决策 API + 应用层 + 跨 CLI + 移动端副驾」推到「CLI 工具层 + coding agent 上下文检索」严肃工程化形态——意味着任何 Jev 应用层(shapeshift / Astra-Ares / deepopen / jev-chat-JARVIS / jev-chat-windows / jeff / rizzo-flow / chat-seek-vscode)现在都有 coding agent 上下文检索 CLI 工具:npm install -g @dzhng/jevgrep + jg auth + jg skill + jg "How are telemetry events recorded and sent?" ./my-project。CLI 工具层补齐了「coding agent 上下文检索」的最后一块,jevgrep 给「Jev 决策模型」补上「Agent 实际拿上下文」的最后一环——这是 Jev / System One 决策模型生态从「决策 API + 应用层 UI + 跨 CLI + 移动端副驾 → CLI 工具层」全栈严肃工程化的关键演化信号。
2. Agent Skill 严肃工程化向「Agent Skill 工具自身 + 商用开源 + 教程合集」三层展开
09-19 ~ 09-28 十日内 Agent Skill 严肃工程化多线同时铺开:
- 09-19
wuyoscar/jev-skill(Awesome Jev Skills 9 技能 + 90 场景 + Action tests) - 09-21
Haleclipse/CometixCode(Rust Claude Code TUI 1:1 重写 + AGPL-3.0) - 09-22
fstandhartinger/chat-seek-vscode(VS Code 跨 CLI 聊天记录本地检索) - 09-23
freestylefly/WeChatBridge(macOS 微信 Share Extension + Developer ID + Apple 公证 + 双语界面) - 09-24
lhlGitHub/threejs-architecture-effects(Three.js 古建程序化动态组装) - 09-25
samyost1/3dicon(一个 prompt → 透明循环 3D 图标) - 09-25
mikehasa/golive-skill(Agent Skill 把 agent 产物上线六路 provider) - 09-25
yetone/magpie(跨 Agent 模型统一网关菜单栏 App) - 09-28
feitangyuan/onetake(Claude Agent Skill「一镜到底」连贯动效 + oracle carry score) - 09-28
supermemoryai/company-brain(商用 Slack 副驾开源 + Cloudflare Workers + Apache-2.0) - 09-28
amitshekhariitbhu/ai-system-design(AI 系统设计 step-by-step 教程 + Apache-2.0)
意味着 Agent Skill 严肃工程化从「Claude Code 周边生态 + 单点工具」推到「跨工作流(视觉 / 上线 / 检索 / 切换)+ 跨平台(macOS / Windows / Linux)+ 跨客户端(Codex / Claude Code / OpenCode / Cursor / Copilot / Gemini CLI / Pi / Goose)」展开,再推到「Agent Skill 工具自身(onetake 一镜到底 carry score)+ 商用开源(company-brain 停服 → Apache-2.0 Cloudflare Workers)+ 教程合集(ai-system-design step-by-step Apache-2.0)」三层严肃工程化形态——这是 Agent Skill 严肃工程化进入「跨场景 + 跨平台 + 跨客户端 + 工具自身 + 商用开源 + 教程合集」演化阶段的关键信号。
3. 严肃工程化个人开发者工具「隐私边界 + 离线优先 + 不联网 + Native OAuth + Developer ID + 公证 + 自我表达 / 创作工具」共识
09-22 ~ 09-28 七日内严肃工程化个人开发者工具多线同时铺开:
- 09-23
freestylefly/WeChatBridge(macOS 微信 Share Extension + Developer ID + Apple 公证) - 09-23
unreallabsai/unreal-agent(async-first Go harness 八组件) - 09-24
SewCabinSpout/cleanupper(macOS 终端清理 Trash-first + 保护路径黑名单 + Zero telemetry) - 09-24
edison-land/paragravity(Chromium user-data-dir 多账号并行 + Native Google OAuth + Spotlight) - 09-25
yetone/magpie(Wails 菜单栏 + 跨 Agent 网关 + 不读 shell env) - 09-28
dzhng/jevgrep(CLI 工具层 + evidence not generated answer + 多 harness auto-detect) - 09-28
supermemoryai/company-brain(Cloudflare Workers 部署 + permissions graph + Apache-2.0 商用清晰) - 09-28
feitangyuan/onetake(PolyForm Noncommercial 商用清晰 + 一镜到底 + oracle carry score)
意味着严肃工程化个人开发者工具从「macOS / Windows 单点 CLI / App」推到「Trash-first + Zero telemetry + 保护路径黑名单 + JSON 脚本化 + 不读 shell env + Atomic writes + Native OAuth + Developer ID + Apple 公证 + Cloudflare Workers 部署 + permissions graph + PolyForm / Apache-2.0 商用清晰 + oracle 测连续性 + evidence not generated answer + 多 harness auto-detect」共识——这是 个人开发者工具从「能跑就行」推到「严肃工程化 + 隐私边界 + 离线优先 + 不联网 + 商用清晰 + 自我表达 / 创作工具」演化。
风险与机遇
风险
- Jev / System One 决策模型 CLI 工具层的稳定性风险:dzhng/jevgrep 把 Jev 决策模型从「决策 API + 应用层」推到「CLI 工具层 + coding agent 上下文检索」,但 Jev 在 folders / files / declarations 间判断相关性在多任务的稳定性 + 四 provider(Vercel AI Gateway / TypeSafe / OpenRouter / OpenCode Zen)的兼容 + 多 harness(Claude Code / Codex / OpenCode / Cursor / Copilot)auto-detect 的覆盖率 + declaration 解析在 Python / TS / JS 之外的扩展 + 其它文本 fallback 的稳定性 均是「Jev CLI 工具层」严肃工程化的关键变量——任何 provider / harness / declaration 不稳,「CLI 工具层」价值会被削弱。
- Agent Skill 严肃工程化向「工具自身 + 商用开源 + 教程合集」三层展开的工程化深度风险:onetake 的 boundary 设计 carry + oracle carry score + ~38 移动纯函数 + real UI HTML 重建在多 motion styles 的稳定性 + company-brain 的 permissions graph + 主动回答 + 主动执行 + MCP 集成在多工具的兼容性 + ai-system-design 的 step-by-step 在多主题的覆盖广度,均是「Agent Skill 三层严肃工程化」的关键挑战。
- 商用副驾开源的可持续性风险:supermemoryai/company-brain 是商用产品停服后 Apache-2.0 完全开源,但 权限图 + 主动回答 + 主动执行 + MCP 集成在多用户的稳定性 + Apache-2.0 维护承诺 + Cloudflare Workers 部署兼容性 + Discord 社区活跃度 均是商用开源严肃工程化的关键变量——任何维护承诺断,「商用开源」可能成为短期热度。
- macOS / Windows 严肃工程化个人开发者工具商业化路径不清晰:dzhng/jevgrep / yetone/magpie / edison-land/paragravity / freestylefly/WeChatBridge / SewCabinSpout/cleanupper 等严肃工程化个人开发者工具多为 MIT + 个人开发者作品,商业化路径(SaaS / 付费版 / 赞助)不清晰,长期可持续性(个人开发者 burnout / 安全更新 / 多平台兼容性)存疑。
机遇
- Jev / TypeSafe AI System One 决策模型生态从「决策 API → 应用层 UI → CLI 工具层 → Agent harness 自托管 → 跨 CLI 检索 → 移动端副驾」全栈严肃工程化是 Apache-2.0 严肃工程化的具体路径:jevgrep 给「Jev 决策模型」补上「Agent 实际拿上下文」的最后一环,意味着任何想在 Jev / TypeSafe AI 生态做 CLI 工具 / 应用层 / 集成层 / Agent 层 / Harness 层 / 移动端副驾的开发者的具体机会窗口——Jev 决策模型生态已进入「全栈严肃工程化」阶段。
- Agent Skill 严肃工程化向「跨工作流 + 跨平台 + 跨客户端 + 工具自身 + 商用开源 + 教程合集」三层展开是 Agent 能力「跨平台分发」的具体路径:onetake(一镜到底)/ company-brain(商用开源)/ ai-system-design(教程合集)/ jevgrep(CLI 工具层)+ 3dicon(视觉)/ golive-skill(上线)/ magpie(模型切换)/ shapeshift(UI 端)/ threejs-architecture-effects(3D 古建)/ CometixCode(TUI)/ WeChatBridge(Share Extension)/ jev-skill(场景合集)等多线同时铺开,意味着 Agent Skill 严肃工程化已从「单点工具」推到「跨工作流 + 跨平台 + 跨客户端 + 工具自身 + 商用开源 + 教程合集」六层展开——这是 Agent Skill 严肃工程化生态的关键演化窗口。
- 商用副驾严肃工程化开源 + Cloudflare Workers 部署 + Apache-2.0 商用清晰是反 SaaS 范式的具体路径:supermemoryai/company-brain 把「商用 Slack 副驾 → Apache-2.0 完全开源 + Cloudflare Workers 部署 + 七个能力 + Skills + workspace prompt + MCP 集成 + permissions graph」严肃工程化,意味着 任何商用副驾停服后 Apache-2.0 开源 + Cloudflare Workers 部署 + 自主托管是反 SaaS 范式的具体路径——这是商用副驾严肃工程化生态的关键演化窗口。
重点项目深度分析
- dzhng/jevgrep(2 天 699⭐ / fork 45 / fork/star 6.4% / TypeScript / MIT / 4143 KB / Score 92)—— Jev / System One 决策模型 CLI 工具层 + 多 provider + 多 harness 严肃工程化;Vercel AI Gateway / TypeSafe / OpenRouter / OpenCode Zen 四 provider;Claude Code / Codex / OpenCode 多 harness auto-detect;Python / TypeScript / JavaScript declaration 解析;
npm install -g @dzhng/jevgrep+jg auth+jg skill+jg "..." ./my-project一行装 + 一行调用;Node.js 22+;summary + file locations + reading leads + verbatim source with line references;evidence not generated answer;npx skills add dzhng/jevgrep --skill jevgrep二次分发;topics 12 个覆盖(ai-sdk / claude-code / cli / code-search / codex / coding-agents / context-retrieval / developer-tools / jev / semantic-search / typescript / vercel-ai-gateway);fork/star 6.4% 较健康 CLI 严肃工程化持续关注信号。 - feitangyuan/onetake(2 天 529⭐ / fork 40 / fork/star 7.6% / Python / PolyForm Noncommercial 1.0.0 / 77237 KB / Score 88)—— Claude Agent Skill「一镜到底」连贯动效做产品发布片;boundary 设计 carry 跨镜头;probe.py oracle 自动测 carry score 0.00 / 0.40 / 0.75 / 0.83;~38 moves (springs / entrances / carries / contact / sims / camera / fluid grounds) 各自为时间纯函数;同一相机不切镜头;real UI rebuilt in HTML from screenshots on camera's plane 让相机以任何 zoom 飞入;screen recording 不用;uniform cadence / no stillness / clipped audio / fast moves without motion blur / subject leaves frame 五项 oracle 自动失败;10 cases;PolyForm Noncommercial 1.0.0 商用清晰;fork/star 7.6% 较健康 Agent Skill 视觉生成严肃工程化持续关注信号。
- supermemoryai/company-brain(3 天 664⭐ / fork 93 / fork/star 14.0% / TypeScript / Apache-2.0 / 13452 KB / Score 85)—— 商用 Slack 记忆副驾开源 + Cloudflare Workers 部署;Apache-2.0 fully open source;runs on Cloudflare Workers;one-click deploy button;permissions graph;七个能力(Remembers / Answers from your team / Acts in your tools / Speaks up / Runs code / Works while you sleep / Learns your way);Skills teach it your processes / formats / voice;workspace prompt sets how it behaves everywhere;GitHub / Linear / Notion / Google Workspace + hundreds more over MCP;its own sandbox clones repos / runs scripts / hands back charts / CSVs / PDFs in Slack;scheduled digests;research it kicks off on its own;decide how chatty org-wide and per channel;fork/star 14.0% 典型商用开源 + 高 fork 严肃工程化持续关注信号。
- anishfn/shapeshift(4 天 759⭐ / fork 71 / fork/star 9.4% / TypeScript / MIT / 12116 KB / Score 82)—— Jev 应用层 UI 端具身形态持续 4 日增长;Jev 14 typed questions + 18 类卡片 + 滞回状态机 challenger + Next.js + Three.js + Excalidraw;内置离线关键词分类器兜底;URL 参数 ?debug=1 ?demo=1&loop=1;shapeshiftui.vercel.app 在线 demo;TypeSafe 兼容后端可切换 4 个 Apache-2.0 开源后端;09-25 387⭐ → 09-26 566⭐ → 09-27 759⭐ 持续 3 日增长 → 09-28 持续;fork/star 9.4% 较健康个人开发者严肃工程化 + Jev 生态持续关注信号。
- amitshekhariitbhu/ai-system-design(3 天 448⭐ / fork 51 / fork/star 11.4% / Markdown / Apache-2.0 / 145 KB / Score 76)—— AI 系统设计 step-by-step 教程合集;Markdown 极小项目;Apache-2.0 商用清晰;topics 10 个覆盖(ai / ai-agents / ai-engineering / ai-system / ai-system-design / ai-systems / ai-systems-design / llm / system-design / system-design-interview);LLM + RAG + AI Agents 全栈设计;step-by-step 教程路径;fork/star 11.4% 较健康 AI 系统设计教程严肃工程化持续关注信号。
数据来源: GitHub Search API (2026-09-28) | Stars: 699 / 529 / 664 / 759 / 448 | License: MIT / PolyForm Noncommercial 1.0.0 / Apache-2.0 / MIT / Apache-2.0 | Created: 2026-09-26 ~ 2026-09-27 | 观察窗: 2026-09-25 ~ 2026-09-27