2026-11-11 · GitHub AI 日报

统计窗口2026-11-10 10:30 ~ 2026-11-11 10:30（北京时间），对应 GitHub `committer-date >= 2026-11-10T02:30:00Z` 之后的默认分支提交。本期共收录 **195** 个条目，其中精选 5 条，四类分布为 ① 62 / ② 32 / ③ 28 / ④ 73。

覆盖方向AI Coding / 编程智能体与技能生态 · 医学 / 临床研究 AI 与 Agent 技能库 · 健康管理 / 营养与食品 AI · 模型底座与 Agent 基础设施

去重说明已与 10-16 至 11-10 二十二期推送记录逐条比对（用 `grep -ho "https://github.com/[^/]*/[^/]*"` 抽出**累计 1360 个历史仓库全集**再逐一比对，且比对时统一转小写以免大小写变体漏网），本期 195 个条目**全部为首次出现，零重复**；5 条精选亦均为首次入榜。本轮检索共采集 **11762** 个 `search/repositories` 唯一候选（160 条单词查询配合 `pushed:>2026-11-09` 日期限定，并合并上一轮同窗口采集结果），另单独抓取 GitHub trending 日榜 / 周榜 / 月榜共 **33** 个仓库逐个核实；剔除与历史重复、SEO 灌水与明显跑题项后得 **9535** 个非历史候选，再**逐仓用 core 端点 `GET /repos/{o}/{r}/commits?since=&per_page=1` 核实默认分支提交**（多线程并发，读取 `Link` 头 `rel="last"` 页码作为精确提交数），**最终保留 2828 个在窗口内有真实提交的仓库**（其中 33 个是 trending 页仓库，trending 榜本身不等于「今天有动作」，一律单独核实后才决定收录），人工挑选与信息密度筛选后定稿 195 个。**凡 `pushed_at` 显示更新但逐仓核实后落在窗口之外的仓库一律剔除，不凭 `pushed_at` 收录**。

核验口径所有"今天有动作"条目均以 GitHub 提交记录核实，**精确提交数一律以 core 端点 `Link` 头 `rel="last"` 的页码为准**（`per_page=1` 时页码即总条数）。**星标与许可证取自 GitHub API 实时快照**（截至 2026-11-11 10:30 北京时间）。**许可证以 core 端点 `GET /repos/{o}/{r}/license` 读取的 LICENSE 文件实文为准，API 的 `spdx_id` 只作线索**——本期实测 `ranxianglei/billion-context` 与 `heymrun/heym` 的 API 报 `NOASSERTION`，实文是 **MIT**；`Tencent/LoopForge` 实文是**腾讯自定义开源授权声明**；`ralforion/orionbelt-semantic-layer` 实文是 **Business Source License 1.1（源码可见，非开源）**；`Arize-ai/phoenix` 实文是 **Elastic License 2.0**；`TrenTorch/TrenTorch` 实文是 **PolyForm Noncommercial 1.0.0（研究与非商用）**；`pinellolab/CRISPResso2` 实文是**学术非商用限定授权**；`camlab-bioml/rakaia` 实文是**保留所有权利的专有声明**；`GiovanniDelGaudio1994/forma-ai` 与 `AA-0324/meliusme` 实文是**专有软件许可**；`Dicklesworthstone/coding_agent_session_search` 实文是 **MIT + OpenAI/Anthropic 附加条款**；`tuist/tuist`、`SourceWeft/SourceWeft`、`HL7/fhir`、`crbnos/carbon`、`OneUptime/oneuptime`、`ExplosiveCoderflome/AI-Novel-Writing-Assistant` 实文均为**多许可或双轨授权**。另有 **18** 个条目经 core 端点逐一确认**仓库根目录没有 LICENSE 文件**（含本期精选的 `Prabh54/groundedmed` 与 `shubhampawar05/nutrikit`），表中标注为"未声明"——**"未声明"不等于无授权条款，但选型阶段必须回查**。本期 195 个条目中 MIT 与 Apache-2.0 合计 128 个（66%），加上 BSD / ISC / MPL / CC0 等宽松型授权后共 146 个（75%）；AGPL-3.0 与 GPL / LGPL 合计 17 个属于传染性 copyleft，只能读不能抄；ELv2 / BUSL-1.1 / PolyForm NC 共 3 个属"源码可见但非开源"；多许可与自定义复合授权 6 个；专有授权与学术非商用 4 个；另有 18 个无 LICENSE 文件。选型时按三档分流。

一、今日精选 5 条

1. Prabh54/groundedmed：让生物医学 RAG 自己证明自己没编——每条结论都带引用校验，批评者模型还能打回重写

地址：https://github.com/Prabh54/groundedmed ｜ ★0 ｜未声明｜ Research demo: agentic RAG that gives citation-checked answers to biomedical yes/no/maybe questions, refuses when evidence is weak, and publishes its results from a frozen 200-question PubMedQA test. Not medical advice.

事件内容

v1.1.0 于 2026-11-11 02:24 发布，距 v1.0.0 公开版只隔一天多，主打「critic 引导重写 + 第二套锁定测试集」；同期还在 README 与 /eval 页面补上了生产冒烟测试结果。
架构是四段式 agentic RAG：planner（规划）→ retriever（检索）→ answer writer（作答）→ critic（批评），回答以 yes / no / maybe 三态给出，每条发现都必须指向它所依据的那篇摘要。
引用校验是代码干的，不是模型自己说的：程序逐条核对引用，critic 再把每条发现与被引摘要原文对照。critic 判定该拒答时，writer 获得一次「有引导的重写」机会，随后 critic 再审一遍。
评测用的是哈希锁定的 PubMedQA 测试集：第二套 400 题的判定规则在跑之前就写死落盘，README 与 /eval 页面同步公布 baseline 与 agent 两轮的分数。

值得关注的原因

这正是我公众号内容缺的那把尺子。 11-09 我给自己列的待办是「每次抽取附原文定位区间，拿不到就不许写进报告」——groundedmed 的做法比我的设想更严：不是「写清出处」就完，而是先做机器校验，再让一个不知道答案的批评者模型逐条对原文。它跟我 11-10 记的 bpf-benchmark 是同一个思路的两面：把「我自己说自己对」换成「一个不知情的第三方来判」。
「证据不足就明说」这条是可复制的。 我 GRPM 的干预报告一直担心模型在证据薄弱时会硬凑一个结论，groundedmed 直接把 not enough evidence 写进产品行为里——这是产品级设计，不是提示词技巧。
哈希锁定的测试集值得整套搬。 第二套 400 题「判定规则先于运行落盘」，这一条能治我公众号 A/B 测试「先看结果再定标准」的通病。下一步我可以把 GRPM 的 SNP 提取规则按同样的方式冻结一版。

2. Muratovnik/assay：给 coding agent 装上「读过的文件必须列全」的门禁——它把「我读过代码」这句话变成可自动核对的声明

地址：https://github.com/Muratovnik/assay ｜ ★1 ｜ MIT ｜ Evidence-grounded methods for coding agents: skills and agent profiles for Claude Code, Codex, Cursor and Gemini CLI

事件内容

v0.44.0 与 v0.44.1 在 2026-11-10 一天内连发两版，v0.44.1 直接针对 skill 文件签出的可见性。
核心是一组 hooks：把「读声明」与它所命名的文件绑定。 提交信息明确写着「keep a read claim's file list going past a name with no Read」「tie a read claim to every file its sentence lists」「compare a full-read claim with the file its own clause names」。
另一条是技能加载的条件化：只有 planning skill 被加载时，才向 plan 索取其 existing-solutions 段落；「skill-evaluation 签名」只对非插件仓库里的 skill 生效。
自评为「Evidence-grounded methods for coding agents: skills and agent profiles」，定位不是提示词库，而是给 agent 装上可执行的证据纪律。

值得关注的原因

这条直接命中我专家包质检最大的漏洞。 我现在的做法是「让模型自评」——它说读了某文件，我信。assay 的做法是让「我读过」这句话必须附带完整文件清单，并且清单里每个文件都有对应的 Read 记录，缺一个就红。 这正是我 11-09 给自己的第二件待办（「每次抽取附原文定位区间」）的可执行版本——我的版本是「要求模型写」，它是「让机器查」。
「read claim 挂在没有 Read 记录的名字后面」这种失败模式，我几乎肯定在自己专家包里也犯了。 报告里写「依据该文件的方法学」，但模型其实只看到了文件名——这两者在人眼审阅里完全一样，在 assay 的门禁下会立刻暴露。这条应该进我的专家包 checklist 第一行。
「只有技能加载了才问对应问题」这条对按需检索同样成立。 我的 GRPM 流程里，SNP 提取规则只在遇到营养相关变异时才该加载；如果无条件全量喂给模型，token 成本和幻觉面都会双涨。

3. vxxqv/openomicsbench：把「组学软件怎么验」做成能离线跑的包——v13 刚把 RNA-seq 输入包与校验和一起发出来

地址：https://github.com/vxxqv/openomicsbench ｜ ★2 ｜ Apache-2.0 ｜ Offline RNA, DNA, protein, single-cell and genomic interval benchmarks with a five-assay demo and traceable reports.

事件内容

v13.0.0 于 2026-11-10 17:30 发布，距 v12.0.0（11-08）两天一个版本，节奏很紧。
v13 的新增内容是 RNA-seq 输入包：既可按单个研究打包，也可为全部 12 个生物对象一起打包；每个包把 count matrix、sample sheet 与声明的 contrast 放在一起，并带校验和与设计预检（design preflight）。
结果表可原生对接 DESeq2、edgeR、limma 的风格并直接打分或转换；仓库同时保留参考结果（reference results），因此它明确声明这不是盲测挑战。
离线可运行是硬承诺：装完之后无需网络，序列文件留在本机；内置工具只覆盖检查与轻量预处理，不宣称替代比对器、变异检测器或分类器。

值得关注的原因

「校验和 + 设计预检 + 参考结果与盲测的区别写在明面上」这三件事合起来，就是我 GRPM 报告最缺的东西。 我 11-10 记的「数据归档 → 判定规则 → 可复现制品」三段路径，openomicsbench 已经把它做成了可 pip 安装的包——尤其是「这不是盲测挑战」这句自我限定，比多数论文都诚实。
它把「方法学可复现」的标准从「我说了算」降到了「我公开我的判定口径」。 12 个生物对象、7 组确定性序列与分类夹具、10x 风格单细胞基准、peak calling 基准——这些是可以直接对照着看的参照物，而不是各家口径不同的排行榜。
对我最实际的用法是「拿来当反例库」。 我的 GRPM 流程里，40 个营养相关 SNP 的提取与解读逻辑各不相同，如果每次换一批 SNP 就重新定义一遍「这条建议的证据强度」，那报告之间的横向比较就是假的。openomicsbench 的做法提醒我：判定规则要冻结成版本，而不是随报告走。

4. shubhampawar05/nutrikit：把营养数据库做成能自托管的 API + MCP——还顺手修了「印度菜配不上海数据库」这种真问题

地址：https://github.com/shubhampawar05/nutrikit ｜ ★0 ｜未声明｜ Self-hostable nutrition API: food search, portion-aware nutrients, meal logging from text, voice, photos, barcodes and labels, meal plans and an AI nutrition advisor. JS SDK on npm (@nutrikit/sdk).

事件内容

窗口内 3 次提交，最新的是 #21「fix/regional-food-matching」，同时补上了拼写纠正：拼错的食物名会给出纠错建议（"Suggest respellings for misspelled foods"）。
具体修的案例：把 dal makhani 匹配到 CoFID 的 black gram（黑豆）与 kidney bean curry——这是一个真实的跨数据库命名断层，不是造出来的演示。
形态是完整的可自托管栈：营养 API + 官网 + 开发者门户 + JavaScript SDK + MCP server，npm 上有 @nutrikit/sdk，Render 上有可试用的免费 API key 实例。
数据面覆盖 13,000+ 食物、每种最多 33 项营养素，且支持「2 medium」「1.5 cups」这类份量表达；另含文字/语音/照片三种记录入口与条码扫描。

值得关注的原因

这是本期健康线里唯一一个「数据库语义」而非「拍照算热量」的仓库，而后者占了 GitHub 健康线的大半。 拍照识别食物的项目很多，但把食物名对齐到营养数据库标识符这件事只有真正被生产使用过才会碰到——dal makhani 这个案例说明作者已经撞上了同一堵墙。
它同时提供 MCP server，这直接对上我 GLP-1 营养伴侣和 4P 平台的接入需求。 我现在的营养数据来源散在各处，统一到一个可自托管、可编程调用、带版本的数据源，正是我要的方向——而且自托管意味着患者的饮食数据不必出本机。
「纠错建议」这个功能看着小，实际是数据入口的第一道防线。 用户/患者输入食物名时打错字，如果系统静默匹配不上就返回空，用户就会放弃记录；给出候选纠错，记录成功率才立得住。我营养伴侣的输入层应该抄这个交互，而不是直接返回空结果。

5. chainloop-dev/chainloop：把「AI 写了什么」变成可留痕的 trace——今天把 Copilot、Pi skill 加载都纳进了记录范围

地址：https://github.com/chainloop-dev/chainloop ｜ ★611 ｜ Apache-2.0 ｜ Agentic SDLC infrastructure. Define guardrails. Collect signals. Enforce continuously. Signed records of what your AI coding agents and CI/CD pipelines did, checked against policies as code, with the loop closed in CI/CD. AI sessions, SBOMs

事件内容

窗口内 6 次提交，主线是 GitHub Copilot 接入：#3801「feat(trace): GitHub Copilot support (spec, plan, and M1)」，#3804「Copilot skills, spec pointers, and pasted images (M2)」，#3805 启用 Copilot 的 Chainloop Trace hooks。
同期还把 Pi 的 skill 加载纳入记录（#3759「feat(trace): record Pi skill loads」）——即「加载了哪个技能」本身成为可审计事件。
自我定位是 agentic SDLC 基础设施：定义护栏、收集信号、执行一致性检查（"Define guardrails. Collect signals. Enforce contracts"）。
Apache-2.0，v1.134.0 于 2026-11-09 发布，v1.133.0 于 11-06；另有一次「让 lab compose stack 端到端跑通」的修复（#3797）。

值得关注的原因

「加载了哪个技能」成为可审计事件，这一条比 trace 本身更重要。 我 11-09 的第三件待办是「每条规则加『最后验证时间 + 指向目标是否仍存在』两个字段」——但如果连「这条结论当时是用哪套规则算出来的」都没记，我连判断结论是否过期都无从下手。 chainloop 记的是这一步。
Copilot / Claude Code / Pi / Codex 多宿主并存已是常态，单靠某个宿主的日志不够用。 我本机就同时装着 dsh、WorkBuddy、Ollama，任何跨宿主的合规要求都需要一个宿主无关的记录层——这正是 chainloop 的位置。
对 4P 平台最直接的启发是「契约」这个概念。 干预报告的输入数据快照、判定规则版本、引用清单，如果三者都带哈希并串成一条 trace，报告的可审计性就从一个愿望变成了一条流水线的一环。这和我 11-10 记的 bpf-benchmark「哈希绑定证据」是同一条路的两种实现。

二、分类速览

① AI Coding / 编程智能体与技能生态

项目	星标	一句话
morluto/rea	★450,372	Reverse engineer anything with agents, from app behavior down to native binaries.
thedotmack/claude-mem	★115,891	Persistent Context Across Sessions for Every Agent – Captures everything your agent does during sessions, compresses it with AI, and injects relevant context back into future sessions. Works with Claude Code, OpenClaw, Codex, Gemini, Hermes, Copilot, OpenCode + More
pbakaus/impeccable	★107,687	The design vocabulary for your coding agent. Steer with polish and bolder, review with critique and audit, catch AI slop before it ships.
heygen-com/hyperframes	★82,380	Write HTML. Render video. Built for agents.
pingdotgg/t3code	★55,913	（仓库未提供描述）
meta/astryx	★23,233	An open source design system that's fully customizable and agent ready
The-PR-Agent/pr-agent	★15,889	🚀 PR Agent: The Original Open-Source PR Reviewer. This project is not the Qodo free tier.
langchain-ai/open-swe	★11,197	An Open-Source Asynchronous Coding Agent
nexu-io/html-anything	★9,495	✨ The agentic HTML editor — your local AI agent writes the HTML, you ship it. 🚀 75 Skills × 9 Surfaces (magazine · deck · poster · XHS / tweet · prototype · data report · Hyperframes) 🛡️ Sandboxed preview · 📤 1-click to WeChat / X / Zhihu / HTML / PNG 🔑 Zero API key — Claude Code / Cursor / Codex /
l0ng-ai/openidea	★8,427	The IDE you know, rebuilt in Rust. IntelliJ IDEA–style keymap, tool windows and search-everywhere in a fast native app: editor, code insight, debugger, Git, terminal, database and HTTP clients for 60+ languages.
logancyang/obsidian-copilot	★7,978	Run agents in Obsidian - OpenCode, Codex, Claude Code etc.
beenuar/AiSOC	★7,199	Open-source AI Security Operations Center: alert fusion, LLM-agent triage, MITRE ATT&CK investigation, and a replayable decision ledger for every agent step. Self-hostable, runs with no API keys, MIT licensed. Ships an MCP server for Claude, Cursor and Continue.
ranxianglei/billion-context	★6,411	A context-compression plugin for month-long single sessions (billions of tokens) ,small context windows (a 100K context is enough) and token savings (5x fewer tokens) .上下文压缩插件，月级长会话会话，兼顾小窗口(100k上下文足矣)省token(省5倍token) . billion-context is all you need
tuist/tuist	★5,864	Tuist builds infrastructure for faster, more reliable build and test workflows across developers, CI, and coding agents. Shared cache, tests, and insights that sit alongside your existing CI.
callstack/agent-device	★5,327	Open source device layer for AI coding agents: a CLI, MCP server, and typed Node.js API to control mobile, web, TV, and desktop apps on devices and simulators, locally or in the cloud.
michael-denyer/pstack-claude	★4,216	Antigravity, Claude Code, Codex, Copilot, Gemini, Muse, OpenCode, Pi, and Prime Agent versions of Poteto's pstack. Rigorous agent workflows with Cursor primitives translated for other harnesses.
evalstate/fast-agent	★3,964	Code, Build and Evaluate agents - excellent Model and Skills/MCP/ACP/A2A Support
yvgude/lean-ctx	★3,941	LeanCTX is the Context Gateway for AI Systems. Select task-relevant context, apply supported controls, connect local knowledge and inspect recorded evidence. Open-source Engine: CLI, MCP, hooks and supported proxy paths.
Javis603/token-monitor	★3,629	Local-first desktop widget for tracking token usage, costs, limits, and session activity across 54+ AI coding tools—including Claude Code, Codex, Cursor, OpenCode, OpenClaw, and more—with multi-device sync.
taylorwilsdon/google_workspace_mcp	★3,501	Control Gmail, Google Calendar, Docs, Sheets, Slides, Chat, Forms, Tasks, Search & Drive with AI - Comprehensive Google Workspace MCP Server & CLI Tool
ExplosiveCoderflome/AI-Novel-Writing-Assistant	★3,424	面向长篇小说创作的 AI Native 开源系统，用 Agent、世界观、写法引擎、RAG 和整本生产工作流，帮助新手从一句灵感走到完整小说。AI-native engine for end-to-end novel creation — from idea to full chapters, with structured planning, worldbuilding, and agent-driven workflows.
xiufengsun/TokenTracker	★2,224	Local-first AI token usage & cost tracker for 31 coding tools incl. Claude Code, Codex, Cursor, Gemini & DeepSeek Harness—with native apps. Never reads prompts.
nimbalyst/nimbalyst	★2,058	Nimbalyst - The open-source visual workspace for Claude Code, Codex, and OpenCode. Run multiple coding agents in parallel, edit their work visually in markdown, mockups, and diagrams, and track tasks. Free, MIT-licensed desktop app for macOS, Windows, Linux, with mobile companion for iOS and Android
sweetpad-dev/sweetpad	★1,927	xcodebuild for humans and agents: build, run, debug, and test iOS, macOS, tvOS, watchOS, and visionOS apps from your terminal.
open-mercato/open-mercato	★1,919	The AI-Engineering Foundation Framework for business apps: open-source TypeScript, with multi-tenancy, RBAC, events and domain modules already decided as conventions and specs, so Cursor, Claude Code and Codex build features instead of re-deciding architecture. Start with 80% done.
heymrun/heym	★1,603	Build agentic systems. Run them with confidence. Orchestrate agents, automate business processes, inspect every execution, and keep humans in control. Deploy Heym on your own infrastructure.
davidondrej/cloudroom	★1,601	Cloud agents for any harness. Claude Code, Codex, Pi, and OpenCode run in isolated sandboxes, and live sessions move between your Mac and the cloud.
tddworks/ClaudeBar	★1,577	A macOS menu bar application that monitors AI coding assistant usage quotas. Keep track of your Claude, Codex, Antigravity ,and Gemini usage at a glance.
gastownhall/gascity	★1,511	Orchestration-builder SDK for multi-agent coding workflows
limecloud/lime	★1,494	Full-stack AI agent for coding, files, terminals, tools, research, content, multimodal work, and multi-agent workflows.
Dicklesworthstone/coding_agent_session_search	★1,203	Unified TUI and CLI to index and search your local coding agent session history across 30+ providers (Codex, Claude Code, Gemini, Cursor, Aider, Copilot, and more) into one searchable timeline. Hybrid search plus robot/JSON output for agents.
agentic-community/mcp-gateway-registry	★1,045	Enterprise-ready MCP Gateway & Registry that centralizes AI development tools with secure OAuth authentication, dynamic tool discovery, and unified access for both autonomous AI agents and AI coding assistants. Transform scattered MCP server chaos into governed, auditable tool access with Keycloak/E
CopilotKit/aimock	★1,009	Mock everything your AI app talks to — LLM APIs, MCP, A2A, AG-UI, vector DBs, search. One package, one port, zero dependencies.
pulseaiclub/phi	★1,002	A fast coding agent, rpc plugin, sub-agents, sloppy edits, and mcp
MichaelKinsy/PiG	★839	PiG (Pi in Go) is a faithful Go port of upstream Pi, the TypeScript codebase behind the Pi coding agent. It is a parity-bound translation, not a rewrite: upstream behavior is the contract, and Go is the implementation language.
Ark0N/Codeman	★838	Self-hosted mission control for AI coding agents: run Claude Code, OpenCode, Pi, Codex, and Antigravity & Gemini CLI 24/7, from any device, watch every subagent live
termio-sh/termio	★788	A terminal-first agentic development environment for agentic coding. Build for CLI/TUI agent. Runtime for Coding Agent, Tmux alternative
nanocodexyz/nanocode	★702	SDK for Frontier Embedded Agents.
avibe-bot/avibe	★639	Make your computer a home for your AI partner. Avibe is the local-first Agent OS: Vibey, its built-in AI partner, works out of the box, and Claude Code, Codex & OpenCode run unchanged beside it — reach them from your browser, phone, or any chat app.
pixtuoid/pixtuoid	★556	Terminal pixel-art office for AI coding agents
ansible/vscode-ansible	★492	Ansible IDE extension: auto-completion and integrating quality assurance tools like ansible-lint, ansible syntax check, yamllint, molecule and ansible-test.
SourceWeft/SourceWeft	★450	Open-source, self-hostable alternative to ChatGPT and NotebookLM — agents, skills, MCP, and cited knowledge in one AI workstation. Your models, your infrastructure. Free to use GPT 6: https://SourceWeft.com/
marcusquinn/aidevops	★448	Vibe-Coding is easy. DevOps is hard. OpenCode & Git token-efficient AI agent automation for your app, business, and personal development. Opinionated tools, services, CLI & API stack for speed, security, and 24/7 results. Open-source first. SOTA everything. The best AI harness for GitHub & teamwork.
Tencent/LoopForge	★240	Self-hosted workflows for coding-agent delivery. Orchestrate CodeBuddy, Codex and Claude Code with scripts and human approval gates. Includes Local workflows for coding assistants.
dev-toolings/superpowers-symfony	★233	Claude Code plugin for Symfony 7.4 LTS & 8.x — 44 skills, 7 AI subagents & 13 commands for API Platform v4, Doctrine ORM 3, TDD (Pest/PHPUnit), Messenger, security & DDD.
boadij/pi-herdsman	★222	Asynchronous Pi subagents and agent fleet orchestration for parallel coding agents with nested delegation, background work, and supervision in herdr.
NeverSight/learn-skills.dev	★221	Curated high-quality AI Agent Skills. Search, install, copy and share. Works with Claude Code, Cursor, OpenClaw, and other AI coding tools.
zoefix/agent-speed	★203	Agent Speed — A lightweight, real-time token speed monitor for Codex and Claude Code. Track tokens per second (TPS), detect slowdowns, and monitor AI agent performance at a glance.
AIpRoBuilder/FlowX	★203	MCP server and Python SDK that generates, runs and refines AI agent workflows from natural language (Generative Workflow, A2A, FastAPI).
skaft-software/octet	★200	fast rust coding agent, extensible in any language, multimodal inputs, deep provider integrations.
shashankswe2020-ux/whoop-mcp	★181	Read-only WHOOP MCP server for recovery, sleep, HRV, strain, workouts, and personal health analytics in Claude, Codex, and GitHub Copilot.
EternalNight996/memory-eternal	★165	Your AI forgets every session. Give it a memory you can audit: local-first, review-gated agent memory — one SQLite vault for DSH, Claude Code, Codex, Cursor, OpenClaw, Hermes and MengBot (any MCP client), with an immutable audit log and a token ledger. Agents propose, humans approve. MIT, self-hoste
dmae97/omk	★152	Open-source multi-agent coding CLI for the terminal (TypeScript, MIT). Switch models mid-session; finish goals only when your test passes.
vibex-ai/vibex	★138	Your agents. Your workspace. Your control—from prompt to commit, desktop to mobile. Works with Codex, Claude Code, OpenCode, Pi, Grok, DeepSeek Harness, Antigravity, and many more through ACP.
mikecann/threadbear	★131	A Mac app for running your Claude Code and Codex threads, on the subscriptions you already have. Comes with an iPhone app, and you get all the source.
makifbaysal/tasktrooper	★122	Local-first agent platform: board + role agents + agent CLI runs (Claude Code, Cursor, Antigravity, OpenCode) or local and API models (Ollama, LM Studio), all on your own Pc
autopus-ai/autopus-adk	★109	Autopus-ADK is of the agents, by the agents. for the agents. Multi-model orchestration (consensus/pipeline/debate/fastest). Architecture-as-Code, Lore decision tracking, SPEC/EARS engine.
ralforion/orionbelt-semantic-layer	★104	Source-available semantic and context layer, rule engine, and semantic sidecar for agentic AI and governed analytics. Compiles declarative YAML models into optimized SQL, KPIs, and semantic context across 8 SQL dialects.
amafjarkasi/electron-mcp-server	★102	MCP server for debugging Electron apps via Chrome DevTools Protocol. Start/attach processes, capture screenshots, read console errors & network logs, inspect DOM, evaluate JS, and automate UI — built for Cursor and other MCP clients.
wenlan-app/wenlan	★95	A living wiki your AI writes from your notes, code, AI chats (memories), and docs, checked against sources. It flags what it can’t back up and updates pages when sources change. Supports Claude, Codex, Obsidian...
cdeust/zetetic-team-subagents	★9	One epistemic standard on three hosts. Claude Code: 15 problem-shaped skills, 97 sourced reasoning patterns, 120 agents, commit-time gates. Codex and Gemini CLI: portable evidence-synthesis, design and goal-loop skills. Shared gates block unsourced constants, vacuous tests and AI-writing patterns. I
Muratovnik/assay	★1	Evidence-grounded methods for coding agents: skills and agent profiles for Claude Code, Codex, Cursor and Gemini CLI

② 医学 / 临床研究 AI 与 Agent 技能库

项目	星标	一句话
scverse/scvi-tools	★1,724	Deep probabilistic analysis of single-cell and spatial omics data
hail-is/hail	★1,082	Cloud-native genomic dataframes and batch computing
cBioPortal/cbioportal	★1,065	cBioPortal for Cancer Genomics
FirelyTeam/firely-net-sdk	★942	The official Firely .NET SDK for HL7 FHIR
HL7/fhir	★764	Official source for the HL7 FHIR Specification
OpenChemistry/avogadrolibs	★704	Avogadro libraries provide 3D rendering, visualization, analysis and data processing useful in computational chemistry, molecular modeling, bioinformatics, materials science, and related areas.
chemosim-lab/ProLIF	★555	Interaction Fingerprints for protein-ligand complexes and more
pinellolab/CRISPResso2	★359	Analysis of deep sequencing data for rapid and intuitive interpretation of genome editing experiments
cqframework/clinical_quality_language	★339	Clinical Quality Language (CQL) is an HL7 specification for the expression of clinical knowledge that can be used within both the Clinical Decision Support (CDS) and Clinical Quality Measurement (CQM) domains. This repository contains complementary tooling in support of that specification.
DIAGNijmegen/rse-grand-challenge	★195	A platform for end-to-end development of machine learning solutions in biomedical imaging
Gaoyuan-0423/biomedical-paper-reader	★156	生物医学论文精读 Skill 与本地阅读 App：逐图解析、证据评估、研究设计借鉴，支持中文、英文及双语报告。／ Biomedical paper-reading skill & local app: figure analysis, evidence appraisal, and research design insights. Chinese, English & bilingual reports.
David-OConnor/molchanica	★152	Protein and molecule viewer, editor, simulator
saezlab/corneto	★109	Unified knowledge-driven network inference from omics data
CMSgov/beneficiary-fhir-data	★87	An API designed to serve Medicare beneficiaries' demographic, enrollment, and claims data using the HL7® FHIR® Standard format.
didymo/OnkoDICOM	★73	OnkoDICOM was created with Radiation Oncologists to allow Radiation Oncologists to do research on DICOM standard image sets (DICOM-RT, CT, MRI, PET) using open source technologies, such as pydicom, dicompyler-core, PySide6, PIL, and matplotlib. OnkoDICOM is cross platform, open source software, and
camlab-bioml/rakaia	★73	Scalable spatial omics in the browser for hundreds of samples at once ／ 10x, IMC, IF, H&E, CosMX, & more ／ https://rakaia.io/
GenomicsStandardsConsortium/mixs	★65	Minimum Information about any (X) Sequence” (MIxS) specification
fulcrumgenomics/fgumi	★63	High-performance UMI tools for NGS data analysis
stjude/proteinpaint	★57	Data visualization and analysis framework focused on phenotype-molecular data integration at cohort level.
StatFunGen/xqtl-protocol	★57	Molecular QTL analysis protocol developed by ADSP Functional Genomics Consortium
microbiomedata/nmdc-schema	★53	National Microbiome Data Collaborative (NMDC) unified data model
HL7/US-Core	★50	FHIR R4 version of US Core IG
bioFAM/mofaflex	★50	factor models for omics data
molecule-dev/molecule	★44	Open-source, composable full-stack packages an AI can wire correctly. Swap any provider in one line. Export anytime.
virtool/virtool	★44	Viral infection diagnostics using next-generation sequencing
Victor-Xu-1/X-Science	★41	X-Science — an open-source biomedical research workspace. Apache-2.0 with component-specific third-party terms.
xinyuran/RLRefine	★12	Schema-driven LLM extraction with source-grounded validation, deterministic repair, and SFT → DPO → GRPO recipes.
genokeeper/genokeeper	★2	Local, privacy-first analyser for your own whole genome sequencing (WGS) data: FASTQ to GRCh38 calls with DeepVariant on your Mac, graded against ClinVar, gnomAD and ClinGen, with pharmacogenomics (PharmCAT, CPIC), polygenic scores, haplogroups and repeat expansions. Nothing leaves your machine unle
vxxqv/openomicsbench	★2	Offline RNA, DNA, protein, single-cell and genomic interval benchmarks with a five-assay demo and traceable reports.
Prabh54/groundedmed	★0	Research demo: agentic RAG that gives citation-checked answers to biomedical yes/no/maybe questions, refuses when evidence is weak, and publishes its results from a frozen 200-question PubMedQA test. Not medical advice.
zaidwhy/assay	★0	Multi-agent research system that checks every claim against the exact passage it cites, and abstains when the evidence is thin.
WannaBeSolopreneur/openevidence-guideline-tracker	★0	Live tracker of public clinical guideline and drug label changes, mapped to ICD-10 and RxNorm

③ 健康管理 / 营养与食品 AI

项目	星标	一句话
mealie-recipes/mealie	★13,875	Mealie is a self hosted recipe manager and meal planner with a RestAPI backend and a reactive frontend application built in Vue for a pleasant user experience for the whole family. Easily add recipes into your database by providing the url and mealie will automatically import the relevant data or ad
MichaIng/DietPi	★6,401	Lightweight justice for your single-board computer!
crbnos/carbon	★2,815	Open-source manufacturing ERP, MES and QMS. Quoting, MRP, inventory, shop floor, quality and lot/serial traceability on one Postgres schema, with a REST API and MCP server. Self-host or use Carbon Cloud.
vllm-project/recipes	★1,085	Common recipes to run vLLM
brandonp2412/Flex	★453	Offline-first workout tracker for strength and cardio, with progress graphs, rest timers, and custom themes. Built with Flutter.
1c7/Fitness	★410	健身教学视频。帮你减脂，增肌（以及拉伸放松，热身，体态，康复）（有大量 B 站和抖音视频）
MichaIng/DietPi-Docs	★242	Source repository for the DietPi OS documentation
slimeglitch/gryffin-calorai-ventus	★114	Top AI Calorie Tracker GitHub 2026
Ajhar424/Senior-Fit-Flow	★56	Senior Fitness Tracker App 2026: Guided Home Workouts for Elderly Health
tsubotti63/garmin-running-data-normalizer	★11	Local-first Python CLI and local browser GUI that turn Garmin Account Data Export into reproducible running, FIT, sleep, HRV, and training datasets with QA and provenance.
cbukenya/healthier	★9	A self-hostable personal health dashboard for your Google Health data: sleep, HRV, activity and evidence-based insights.
MMacLean305/Vivacity	★3	Vivacity is a full-stack health and fitness application featuring workout tracking, nutrition logging, progress monitoring, and AI-powered exercise recommendations.
BoraErenErdem/pulse-coach	★2	A multi-agent AI health & fitness coach built with LangChain, FastAPI, and local LLMs (Ollama), featuring RAG-powered nutrition/exercise guidance and proactive check-ins.
GiovanniDelGaudio1994/forma-ai	★0	AI-powered fitness and nutrition app featuring personalized workout plans, adaptive coaching, progress tracking, exercise video analysis, and allergy-aware meal planning.
SyberKraken/morrowkeep	★0	Local-first daily planner for Windows: tasks, chores, meals, workouts, calories and weight. Rust + egui, lives in the tray.
AA-0324/meliusme	★0	A minimalist, privacy-focused nutrition and meal-tracker. Local-only data storage and zero friction. Built for accurate tracking and user ownership.
victoriacarbonell/nutri-app	★0	Calorie and nutrition tracker PWA with AI photo analysis · work in progress
mzaiger/FastFoodDiet	★0	An application that builds you a fast food meal diet plan for 7 days with a calorie input. This is to help you stay within the calories you have for a goal each week.
ShamsahAld/Microbiome_Fermented	★0	Analyse metagenomic data from populations consuming traditional fermented foods (kefir, kombucha, Arabian laban) and construct microbial co-occurrence networks. Identify which microbial modules are enriched and link them to host metabolic pathways using network topology.
mariogiordano96/plico	★0	Open source Android app for workout tracking, meal planning and nutrition. Italian CREA food data, guided workouts, optional AI coach. No account, no ads.
sajedayad/Forge	★0	Forge: a mobile-first gym and nutrition tracker built with Claude. Workout logging, calorie and protein tracking, photo label scanning, sleep insights and weekly check-ins.
shubhampawar05/nutrikit	★0	Self-hostable nutrition API: food search, portion-aware nutrients, meal logging from text, voice, photos, barcodes and labels, meal plans and an AI nutrition advisor. JS SDK on npm (@nutrikit/sdk).
Maheshdasi/FitPlate	★0	AI Fitness & Nutrition Platform: explainable calorie targets, ML meal and workout recommendations, progress tracking and weekly insights. React, FastAPI, PostgreSQL, scikit-learn.
binDebug3/Nutrition_Planner	★0	An app for dietetics students to automatically build meals based on nutrient needs
sumaisa-mou/calorie-whisperer	★0	AI-powered fitness and nutrition tracker: snap a meal photo, type what you ate, or chat to log food and workouts.
chetan-tomar03/NutriScan-AI	★0	AI-powered nutrition analysis app that identifies food from images and estimates calories, protein, carbs, fats, and fiber
gemivnet/fitsheet	★0	Personal fitness companion — calories, weight, workouts, walks, recipes, notes, analytics. Web PWA + API in one Docker image.
tebl22/don-bolon	★0	🧀 Snap your plate, get the truth. AI nutrition coach for Android: photo calorie tracking, science-based goals, fasting and a grumpy cheese-ball mascot.

④ 模型底座与 Agent 基础设施

项目	星标	一句话
netdata/netdata	★81,359	The fastest path to AI-powered full stack observability, even for lean teams.
grafana/grafana	★77,825	The open and composable observability and data visualization platform. Visualize metrics, logs, and traces from multiple sources like Prometheus, Loki, Elasticsearch, InfluxDB, Postgres and many more.
hiyouga/LlamaFactory	★75,988	Unified Efficient Fine-Tuning of 100+ LLMs & VLMs (ACL 2024)
deepspeedai/DeepSpeed	★43,319	DeepSpeed is a deep learning optimization library that makes distributed training and inference easy, efficient, and effective.
facebookresearch/faiss	★41,375	A library for efficient similarity search and clustering of dense vectors.
servo/servo	★38,458	Servo aims to empower developers with a lightweight, high-performance alternative for embedding web technologies in applications.
seaweedfs/seaweedfs	★36,013	SeaweedFS is a distributed storage system for object storage (S3), file systems, Iceberg table buckets, and Lance vector buckets, designed to handle billions of files with O(1) disk access and effortless horizontal scaling.
apify/crawlee	★26,415	Crawlee—A web scraping and browser automation library for Node.js to build reliable crawlers. In JavaScript and TypeScript. Extract data for AI, LLMs, RAG, or GPTs. Download HTML, PDF, JPG, PNG, and other files from websites. Works with Puppeteer, Playwright, Cheerio, JSDOM, and raw HTTP. Both headf
longbridge/gpui-kit	★20,287	Rust GUI components for building fantastic cross-platform desktop application by using GPUI.
NVIDIA-NeMo/Speech	★18,727	A scalable generative AI framework built for researchers and developers working on Large Language Models, Multimodal, and Speech AI (Automatic Speech Recognition and Text-to-Speech)
NVIDIA/Megatron-LM	★18,373	Ongoing research training transformer models at scale
modelscope/ms-swift	★16,081	Use PEFT or Full-parameter to CPT/SFT/DPO/GRPO 600+ LLMs (Qwen3.8, DeepSeek-V4, GLM-5.1, InternLM3, Llama4, ...) and 300+ MLLMs (Qwen3-VL, Qwen3-Omni, InternVL3.5, Ovis2.5, GLM5.3, Gemma4, Llava, Phi4, ...) (AAAI 2025).
facebookresearch/sam3	★12,262	The repository provides code for running inference and finetuning with the Meta Segment Anything Model 3 (SAM 3), links for downloading the trained model checkpoints, and example notebooks that show how to use the model.
Arize-ai/phoenix	★12,238	AI Observability & Evaluation
lancedb/lancedb	★11,876	Developer-friendly OSS embedded retrieval library for multimodal AI. Search More; Manage Less.
storytold/vectorcraft	★11,602	VectorCraft: An open-source vector art tool written from scratch in pure Rust
ClickHouse/ClickStack	★10,020	Resolve production issues, fast. An open source observability platform unifying session replays, logs, metrics, traces and errors powered by ClickHouse and OpenTelemetry.
NVIDIA/cudf	★9,799	cuDF - GPU DataFrame Library
LearningCircuit/local-deep-research	★9,560	~95% on SimpleQA (e.g. Qwen3.6-27B on a 3090). Supports all local and cloud LLMs (llama.cpp, Ollama, Google, ...). 10+ search engines - arXiv, PubMed, your private documents. Everything Local & Encrypted.
flyteorg/flyte	★7,833	Type-safe, distributed orchestration of agents, ML pipelines, and real-time inference on your k8s — in pure Python with async/await, also other languages (rust, go and ts)
OneUptime/oneuptime	★7,810	Complete open-source monitoring and observability platform.
getopenscreen/openscreen	★6,879	Record your screen, ship a demo. Free and open-source, GPU-accelerated, no watermarks, no subscriptions. Windows, macOS, Linux. Actively maintained.
skyzh/tiny-llm	★5,712	learn LLM inference system on Apple Silicon for systems engineers: build a tiny vLLM + Qwen
ModelTC/LightLLM	★4,359	LightLLM is a Python-based LLM (Large Language Model) inference and serving framework, notable for its lightweight design, easy scalability, and high-speed performance.
uxlfoundation/oneDNN	★4,091	Open-source library of optimized deep learning operations (matmul, convolution, attention) for CPUs (x64, AArch64, RISC-V) and Intel GPUs. Used by PyTorch, TensorFlow, OpenVINO, and ONNX Runtime.
checkpoint-restore/criu	★4,066	Checkpoint/Restore tool
NVIDIA/cuda-rust	★4,023	NVIDIA's CUDA platform for Rust. Host runtime crates plus Tile (cutile-rs) and SIMT (cuda-oxide) kernel programming models in idiomatic Rust.
HakanSeven12/OpenCADStudio	★3,972	A CAD application built with Rust — 2D/3D drawing, DWG/DXF support, and GPU-accelerated rendering
pytorch/TensorRT	★3,021	PyTorch/TorchScript/FX compiler for NVIDIA GPUs using TensorRT
thu-pacman/chitu	★2,999	High-performance inference framework for large language models, focusing on efficiency, flexibility, and availability.
NVIDIA/gpu-operator	★2,980	NVIDIA GPU Operator creates, configures, and manages GPUs in Kubernetes
mw00/project-maya	★2,766	Run GLM-5.3 (744B), GLM-5.3-Flash (321B), DeepSeek-V4.1-Flash (748B) and MiniMax-M3.1-Flash (428B) on your own GPUs - one engine, one dashboard, OpenAI- and Anthropic-compatible API.
jolibrain/deepdetect	★2,551	Deep Learning Server and CLI for Torch and TensorRT
noonghunna/club-3090	★2,490	Community recipes for serving LLMs on RTX 3090/4090/5090 CUDA gpus. Multi-engine (vLLM, sglang, llama.cpp and other custom engines) and model-agnostic. Currently shipping Qwen3.8-27B Qwen3.6 35B Gemma 4 26B Gemma 4 31B configs for 1×/2×/4x/
lupinemachines/lupine	★2,463	LUPINE is a GPU over IP bridge allowing GPUs on remote machines to be attached to CPU-only machines.
ovg-project/kvcached	★2,038	Virtualized Elastic KV Cache for Dynamic GPU Sharing and Beyond
SemiAnalysisAI/InferenceX	★2,022	Open Source AI Accelerator Research Platform Standard / 开源推理研究平台
CrossGL/crosstl	★1,989	Translate GPU shaders and compute kernels between GLSL, HLSL, Metal, SPIR-V, Slang, CUDA and HIP, and to WGSL, through the CrossGL IR
datawhalechina/zero-to-sglang	★1,976	Official SGLang x Datawhale course on LLM inference: understand inference, build a mini-sglang from scratch, then read the real SGLang source and land your first PR. Available in English and Chinese.
AdamPlatin123/dsh-plugin-radar	★1,500	DSH Plugin Radar — open-source ecosystem radar for DeepSeek Harness plugins: continuous discovery (21k+ candidates), k8s runtime validation (13k+ tests), 15-min snapshots; the catalog is a generated artifact — 开源 DSH 插件生
alibaba/rtp-llm	★1,394	RTP-LLM: Alibaba's high-performance LLM inference engine for diverse applications.
ModelCloud/GPTQModel	★1,289	LLM model quantization (compression) toolkit with HW acceleration support for Nvidia (CUDA), AMD (ROCm), Intel GPU (XPU), Huawei Ascent (NPU), and Intel/AMD/Apple CPU via HF, vLLM, and SGLang.
sgl-project/SpecForge	★1,279	Train speculative decoding models effortlessly and port them smoothly to SGLang serving.
CurvineIO/curvine	★1,058	AI-native, cloud-native file system in Rust. Full POSIX on object storage with a multi-tier distributed cache, for AI training, inference, and agents. CNCF Sandbox project.
kividbio/kividb	★1,057	A fast, multi-threaded Redis® and Valkey alternative written in Rust: drop-in compatible, with full-text and vector search built in
NVIDIA/cudnn-frontend	★996	cuDNN Frontend is NVIDIA's modern, open-source entry point to the cuDNN library and a growing collection of high-performance open-source kernels.
livekit/agents-js	★962	Build realtime multimodal AI agents with Node.js
openvinotoolkit/model_server	★956	A scalable inference server for models optimized with OpenVINO™
synchopate/llama-modchip	★897	🐋 Run 284B DeepSeek-V4-Flash (<59GiB RAM, 44.4 tok/s pp, 9.1-14.7 tok/s) & 30B MoEs (25-111 tok/s) on a pure 64GB CPU at 256K ctx (0 GPUs). SwiGLU-H2, Stockfish-PVS L3 & 13ms Claude Code restore.
ai-dynamo/aiperf	★866	AIPerf is a comprehensive benchmarking tool that measures the performance of generative AI models served by your preferred inference solution.
TrenTorch/TrenTorch	★835	Don't memorize ML. Understand it from first principles including Inference, Cuda, Deep Learning, Data Science and Reinforcement Learning
usemoss/moss	★806	The retrieval layer for production AI systems. Lightning-fast (<10ms) search without vector databases. Built for browser, edge, on-device, and cloud.
Helldez/BigMoeOnEdge	★741	Run MoE models bigger than your RAM. Frontier-size MoE on a phone, laptop, edge. Lossless, on stock llama.cpp. An adaptive streaming optimizer. Any hardware, any OS, any model, any quantization, behind a local OpenAI-compatible server.
pegainfer-project/pegainfer	★739	Pure Rust + CUDA LLM inference engine — no PyTorch, OpenAI-compatible, serves Qwen3 to Kimi-K2
nace-ai/kapitsa	★732	Blazing fast Apple Silicon inference engine optimized for Qwen & Nemotron. OpenAI + Anthropic API compatible. Built by folks from nace.ai
Basekick-Labs/arc	★690	Open, SQL-native time-series database for telemetry you need to keep. 38M+ records/sec ingestion, 25M+ rows/sec queries. InfluxDB Line Protocol and Telegraf compatible. Open Parquet on your storage. Single binary. S3/Azu
zhongkaifu/TensorSharp	★671	A native .NET LLM inference engine and agent runtime for GGUF models. TensorSharp provides a console application, a web-based chatbot interface, iPhone App, and Ollama/OpenAI-compatible HTTP APIs for programmatic access. It supports Windows
chainloop-dev/chainloop	★611	Agentic SDLC infrastructure. Define guardrails. Collect signals. Enforce continuously. Signed records of what your AI coding agents and CI/CD pipelines did, checked against policies as code, with the loop closed in CI/CD. AI sessions, SBOMs
kacper-daftcode/vLLM-Moet	★572	Frontier MoE models on RTX PRO 6000 / RTX 5090 (sm_120): DeepSeek-V4.1-Flash on 4x RTX PRO 6000, as a GPU+CPU hybrid on 4x RTX 5090, and 2-bit experts on vLLM v0.24.0
alchemy-run/distilled	★522	Effect-native SDKs for cloud providers
bytedance/vArmor	★516	vArmor is a cloud-native container hardening system that leverages AppArmor/BPF/Seccomp and NetworkProxy technologies to enforce access control from system calls to application protocols — protecting workloads including AI Agents.
vllm-project/vime	★503	An LLM post-training framework with vLLM for RL Scaling
volcano-sh/kthena	★497	Lightweight, Modular, Kubernetes-native AI serving platform for scalable model serving.
openxla/tokamax	★329	Tokamax: A GPU and TPU kernel library.
ai-dynamo/aitune	★290	NVIDIA AITune is an inference toolkit designed for tuning and deploying Deep Learning models with a focus on NVIDIA GPUs.
OrbitInfra/OrbitKV	★281	OrbitKV is a KV cache for vLLM and SGLang.
dsx-ai-factory/nvcf	★265	Platform for deploying and routing GPU-accelerated inference, streaming, and batch workloads at scale.
adventists-ai/sophon-realtime	★227	Sophon-Realtime v1: open full-duplex voice models that hear your tone and speak with one. Tag and Flash run on one GPU with vLLM-Omni; E2E in preview.
sgl-project/sgl-kernel-npu	★200	SGLang kernel library for NPU
NVIDIA/kumo-engine	★75	Kumo Engine is an accelerator for context generation over structured data, designed for teams building and running predictive models in production at scale.
sgl-project/sgl-kernel-xpu	★38	SGLang kernel library for Intel XPU
intel/enterprise-ai-solutions	★11	Production-ready platform for deploying OpenAI-compatible LLM inference on Kubernetes. A one-command installer for enterprise generative AI on Intel® Xeon®, with vLLM and other serving engines, secure Keycloak authentication, and built-in o
intel/enterprise-inference	★8	Enterprise-grade, high-performance LLM serving on Kubernetes. OpenAI-compatible inference powered by vLLM and OpenVINO™ Model Server behind a unified AI gateway, with NUMA-aware CPU optimization for Intel® Xeon® - production model serving,

三、今日趋势观察

⭐⭐⭐ 「读了没有」正在变成一个可以被机器判红的问题。 11-10 我记的 astryx 是「测试失败报具名场景」，11-09 我记的 squad 是「规范/文档/矩阵三者对齐代码 + 回归守卫」，今天 assay 直接把这件事做到了 agent 身上：tie a read claim to every file its sentence lists——模型说「我读了某文件」，必须给出完整清单，清单里每个名字都得有对应的 Read 记录。我专家包质检目前还是「让模型自评」，这条是可以整套抄的。
⭐⭐⭐ 「把结论钉在原始材料上」出现了三种不同实现。 groundedmed 让批评者模型逐条对照摘要原文；openomicsbench 让判定规则连同校验和一起冻结成可安装的包；chainloop 让「加载了哪个技能」本身成为可审计事件。技术栈毫无交集，但都在回答同一个问题——三个月后我怎么知道当时那份报告是怎么得出来的。 我 GRPM 的「规则加验证时间字段」只解决了一半：字段加了，但产出这份结论时用的到底是哪套规则、哪份原始数据，仍然没有留痕。
⭐⭐ 健康线的分水岭：数据库语义 vs 拍照算热量。 本期健康线 28 个条目里绝大多数仍是「拍照识别食物 + 算卡路里」，而 nutrikit 做的是把食物名对齐到营养数据库标识符——它修的那个 dal makhani → black gram 案例，是只有真正把这件事用在生产上才会撞到的墙。这一条的分量在于：它证明营养 AI 的难点不在识别，在于对齐。 我做补剂配方与功能医学干预评估时，卡住我的也正是这一层。自托管 + MCP 组合还直接对上了我 GLP-1 营养伴侣的接入需求。
⭐⭐ dsh 生态继续在榜。 AdamPlatin123/dsh-plugin-radar（★1,500，窗口内 238 次提交）是一个开源生态雷达，持续快照 + 渲染 dsh 插件生态的变化——我本机装着 dsh，这条线应该持续优先跟踪。上一期还有 dsh-Go、dsh-blender-plugin、dsh-mnemon、dsh-memory 四个不同团队的 dsh 扩展同时在榜，说明这不是单人项目，而是已经长出生态。
⭐⭐ 授权形态这一期出现了一个新的三档样本。 真开源（MIT 与 Apache-2.0 合计 128 个，66%；加上 BSD / ISC / MPL 等宽松型授权后共 146 个，75%）、源码可见但非开源（orionbelt-semantic-layer 的 BUSL-1.1、phoenix 的 ELv2、TrenTorch 的 PolyForm NC 各 1 个）、完全不开放（forma-ai 与 meliusme 的专有许可、rakaia 的保留所有权利、CRISPResso2 的学术非商用 4 个）。对一人公司的含义仍是三件不同的事：第三档连读都不行、第二档可读可学设计但绝不能进产品依赖、只有第一档是选型候选。 而本期最刺眼的一条是：两个精选仓库（groundedmed 与 nutrikit）根本没有 LICENSE 文件——它们的方法论值得抄，但代码不能直接依赖。

四、GitHub 地址速查

GitHub AI 日报 · 2026-11-11（周三）

一、今日精选 5 条

1. Prabh54/groundedmed：让生物医学 RAG 自己证明自己没编——每条结论都带引用校验，批评者模型还能打回重写

2. Muratovnik/assay：给 coding agent 装上「读过的文件必须列全」的门禁——它把「我读过代码」这句话变成可自动核对的声明

3. vxxqv/openomicsbench：把「组学软件怎么验」做成能离线跑的包——v13 刚把 RNA-seq 输入包与校验和一起发出来

4. shubhampawar05/nutrikit：把营养数据库做成能自托管的 API + MCP——还顺手修了「印度菜配不上海数据库」这种真问题

5. chainloop-dev/chainloop：把「AI 写了什么」变成可留痕的 trace——今天把 Copilot、Pi skill 加载都纳进了记录范围

二、分类速览

① AI Coding / 编程智能体与技能生态

② 医学 / 临床研究 AI 与 Agent 技能库

③ 健康管理 / 营养与食品 AI

④ 模型底座与 Agent 基础设施

三、今日趋势观察

四、GitHub 地址速查