From b251c92573f1cd263ba199477e22af90358d078a Mon Sep 17 00:00:00 2001 From: arno Date: Sat, 9 May 2026 15:35:24 +0800 Subject: [PATCH] =?UTF-8?q?=E6=96=87=E6=A1=A3(=E5=B7=A5=E7=A8=8B):=20?= =?UTF-8?q?=E6=9B=B4=E6=96=B0=E9=A1=B9=E7=9B=AE=20Star=20=E6=95=B0?= =?UTF-8?q?=E5=B9=B6=E7=A7=BB=E9=99=A4=E5=AE=9E=E8=B7=B5=E4=B8=8E=20specki?= =?UTF-8?q?t=20=E9=85=8D=E7=BD=AE=E6=AE=B5=E8=90=BD?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit - 更新 AI Agent、RAG/LLM 框架等分类下项目的 Star 数 - 移除底部的实践经验总结和 speckit.constitution 配置 --- 工程/README.md | 120 +++++++++++++++++-------------------------------- 1 file changed, 42 insertions(+), 78 deletions(-) diff --git a/工程/README.md b/工程/README.md index c474efc..dc210f8 100644 --- a/工程/README.md +++ b/工程/README.md @@ -16,56 +16,56 @@ ### Claude Code / AI Agent -- [OpenClaw](https://github.com/openclaw/openclaw) ⭐ 352.6k - 你的个人 AI 助手,支持任何操作系统和平台 -- [n8n](https://github.com/n8n-io/n8n) ⭐ 183.1k - 具有原生 AI 能力的公平代码工作流自动化平台 -- [Claw Code](https://github.com/ultraworkers/claw-code) ⭐ 178.8k - GitHub 历史上最快突破 100K Stars 的仓库,Rust 实现的 CLI Agent Harness -- [Everything Claude Code](https://github.com/affaan-m/everything-claude-code) ⭐ 147.5k - Agent Harness 性能优化系统,提供技能、记忆、安全和研发优先开发 -- [Superpowers](https://github.com/obra/superpowers) ⭐ 142.3k - 有效的 Agent 技能框架和软件开发方法论 -- [Dify](https://github.com/langgenius/dify) ⭐ 136.9k - 生产就绪的 Agent 工作流开发平台 -- [LangChain](https://github.com/langchain-ai/langchain) ⭐ 132.9k - Agent 工程平台 -- [Open WebUI](https://github.com/open-webui/open-webui) ⭐ 130.8k - 用户友好的 AI 界面,支持 Ollama、OpenAI API 等 -- [Claude Code](https://github.com/anthropics/claude-code) ⭐ 111.4k - Anthropic 官方终端原生 AI 编程工具 -- [Spec Kit](https://github.com/github/spec-kit) ⭐ 86.4k - GitHub 规范工具包,帮助开始规范驱动开发 -- [nanobot](https://github.com/HKUDS/nanobot) ⭐ 38.6k - 超轻量级个人 AI Agent -- [OpenSpec](https://github.com/Fission-AI/OpenSpec) ⭐ 38.5k - AI 编码助手的规范驱动开发工具 -- [RAGFlow](https://github.com/infiniflow/ragflow) ⭐ 77.5k - 融合 RAG 与 Agent 能力的开源检索增强生成引擎 -- [Page Agent](https://github.com/alibaba/page-agent) ⭐ 16.5k - 纯 JavaScript 实现的页面 GUI Agent,使用自然语言操作 Web 界面 -- [Agent Skills for Context Engineering](https://github.com/muratcankoylan/Agent-Skills-for-Context-Engineering) ⭐ 14.9k - 上下文工程、多 Agent 架构和生产 Agent 系统的综合技能集合 -- [Claude Plugins Official](https://github.com/anthropics/claude-plugins-official) ⭐ 16.4k - Anthropic 官方管理的高质量 Claude Code 插件目录 -- [Compound Engineering Plugin](https://github.com/EveryInc/compound-engineering-plugin) ⭐ 13.8k - Claude Code、Codex 等的官方复合工程插件 -- [Claude Code Router](https://github.com/musistudio/claude-code-router) ⭐ 31.8k - 使用 Claude Code 作为编码基础设施基础,允许你决定如何与模型交互 -- [Happy Coder](https://github.com/slopus/happy) ⭐ 17.5k - Codex 和 Claude Code 的移动端和 Web 客户端,支持实时语音、加密 -- [VoltAgent](https://github.com/VoltAgent/voltagent) ⭐ 7.7k - 基于开源 TypeScript AI Agent 框架构建的 AI Agent 工程平台 -- [MasteringRAG](https://github.com/Steven-Luo/MasteringRAG) ⭐ 665 - 企业级 RAG 系统从入门到精通 +- [OpenClaw](https://github.com/openclaw/openclaw) ⭐ 369.9k - 你的个人 AI 助手,支持任何操作系统和平台 +- [n8n](https://github.com/n8n-io/n8n) ⭐ 187.1k - 具有原生 AI 能力的公平代码工作流自动化平台 +- [Claw Code](https://github.com/ultraworkers/claw-code) ⭐ 190.7k - GitHub 历史上最快突破 100K Stars 的仓库,Rust 实现的 CLI Agent Harness +- [Everything Claude Code](https://github.com/affaan-m/everything-claude-code) ⭐ 176.1k - Agent Harness 性能优化系统,提供技能、记忆、安全和研发优先开发 +- [Superpowers](https://github.com/obra/superpowers) ⭐ 183.6k - 有效的 Agent 技能框架和软件开发方法论 +- [Dify](https://github.com/langgenius/dify) ⭐ 140.6k - 生产就绪的 Agent 工作流开发平台 +- [LangChain](https://github.com/langchain-ai/langchain) ⭐ 136.1k - Agent 工程平台 +- [Open WebUI](https://github.com/open-webui/open-webui) ⭐ 136.2k - 用户友好的 AI 界面,支持 Ollama、OpenAI API 等 +- [Claude Code](https://github.com/anthropics/claude-code) ⭐ 121.8k - Anthropic 官方终端原生 AI 编程工具 +- [Spec Kit](https://github.com/github/spec-kit) ⭐ 93.8k - GitHub 规范工具包,帮助开始规范驱动开发 +- [nanobot](https://github.com/HKUDS/nanobot) ⭐ 42.0k - 超轻量级个人 AI Agent +- [OpenSpec](https://github.com/Fission-AI/OpenSpec) ⭐ 46.4k - AI 编码助手的规范驱动开发工具 +- [RAGFlow](https://github.com/infiniflow/ragflow) ⭐ 80.0k - 融合 RAG 与 Agent 能力的开源检索增强生成引擎 +- [Page Agent](https://github.com/alibaba/page-agent) ⭐ 17.6k - 纯 JavaScript 实现的页面 GUI Agent,使用自然语言操作 Web 界面 +- [Agent Skills for Context Engineering](https://github.com/muratcankoylan/Agent-Skills-for-Context-Engineering) ⭐ 15.5k - 上下文工程、多 Agent 架构和生产 Agent 系统的综合技能集合 +- [Claude Plugins Official](https://github.com/anthropics/claude-plugins-official) ⭐ 18.9k - Anthropic 官方管理的高质量 Claude Code 插件目录 +- [Compound Engineering Plugin](https://github.com/EveryInc/compound-engineering-plugin) ⭐ 16.3k - Claude Code、Codex 等的官方复合工程插件 +- [Claude Code Router](https://github.com/musistudio/claude-code-router) ⭐ 33.6k - 使用 Claude Code 作为编码基础设施基础,允许你决定如何与模型交互 +- [Happy Coder](https://github.com/slopus/happy) ⭐ 20.3k - Codex 和 Claude Code 的移动端和 Web 客户端,支持实时语音、加密 +- [VoltAgent](https://github.com/VoltAgent/voltagent) ⭐ 8.7k - 基于开源 TypeScript AI Agent 框架构建的 AI Agent 工程平台 +- [MasteringRAG](https://github.com/Steven-Luo/MasteringRAG) ⭐ 679 - 企业级 RAG 系统从入门到精通 ### RAG / LLM 框架 -- [AutoGen](https://github.com/microsoft/autogen) ⭐ 56.9k - Microsoft Agent AI 编程框架,用于创建可自主行动或与人类协作的多 Agent AI 应用 -- [Flowise](https://github.com/FlowiseAI/Flowise) ⭐ 51.7k - 可视化构建 AI Agents 的低代码平台 -- [LlamaIndex](https://github.com/run-llama/llama_index) ⭐ 48.4k - 领先的文档 Agent 和 OCR 平台 -- [crewAI](https://github.com/crewAIInc/crewai) ⭐ 48.4k - 协调角色扮演、自主 AI Agents 的框架,通过培养协作智能使 Agents 无缝协作 +- [AutoGen](https://github.com/microsoft/autogen) ⭐ 57.8k - Microsoft Agent AI 编程框架,用于创建可自主行动或与人类协作的多 Agent AI 应用 +- [Flowise](https://github.com/FlowiseAI/Flowise) ⭐ 52.6k - 可视化构建 AI Agents 的低代码平台 +- [LlamaIndex](https://github.com/run-llama/llama_index) ⭐ 49.2k - 领先的文档 Agent 和 OCR 平台 +- [crewAI](https://github.com/crewAIInc/crewai) ⭐ 50.9k - 协调角色扮演、自主 AI Agents 的框架,通过培养协作智能使 Agents 无缝协作 - [Quivr](https://github.com/QuivrHQ/quivr) ⭐ 39.1k - 为你的应用集成 GenAI 的 RAG 引擎,专注于产品而非 RAG -- [LangGraph](https://github.com/langchain-ai/langgraph) ⭐ 28.8k - 构建有弹性的语言 Agents 作为图 -- [FastGPT](https://github.com/labring/FastGPT) ⭐ 27.7k - 基于 LLMs 的知识库平台,提供数据处理、RAG 检索和可视化 AI 工作流编排 -- [SemanticKernel](https://github.com/microsoft/semantic-kernel) ⭐ 27.7k - 轻松快速地将尖端 LLM 技术集成到应用中 -- [Haystack](https://github.com/deepset-ai/haystack) ⭐ 24.8k - 开源 AI 编排框架,用于构建上下文工程、生产就绪的 LLM 应用 -- [Hindsight](https://github.com/vectorize-io/hindsight) ⭐ 8.6k - 能够学习的 Agent Memory +- [LangGraph](https://github.com/langchain-ai/langgraph) ⭐ 31.5k - 构建有弹性的语言 Agents 作为图 +- [FastGPT](https://github.com/labring/FastGPT) ⭐ 27.9k - 基于 LLMs 的知识库平台,提供数据处理、RAG 检索和可视化 AI 工作流编排 +- [SemanticKernel](https://github.com/microsoft/semantic-kernel) ⭐ 27.8k - 轻松快速地将尖端 LLM 技术集成到应用中 +- [Haystack](https://github.com/deepset-ai/haystack) ⭐ 25.1k - 开源 AI 编排框架,用于构建上下文工程、生产就绪的 LLM 应用 +- [Hindsight](https://github.com/vectorize-io/hindsight) ⭐ 12.7k - 能够学习的 Agent Memory - [FlashRAG](https://github.com/RUC-NLPIR/FlashRAG) ⭐ 3.4k - 高效 RAG 研究的 Python 工具包 ### 推理部署 -- [Milvus](https://github.com/milvus-io/milvus) ⭐ 43.7k - 高性能、云原生向量数据库,用于可扩展的向量 ANN 搜索 +- [Milvus](https://github.com/milvus-io/milvus) ⭐ 44.1k - 高性能、云原生向量数据库,用于可扩展的向量 ANN 搜索 - [FastChat](https://github.com/lm-sys/FastChat) ⭐ 39.4k - 用于训练、服务和评估大语言模型的开放平台,Vicuna 和 Chatbot Arena 的发布仓库 -- [SGLang](https://github.com/sgl-project/sglang) ⭐ 25.6k - 大语言模型和多模态模型的高性能服务框架 -- [TensorRT-LLM](https://github.com/NVIDIA/TensorRT-LLM) ⭐ 13.3k - NVIDIA 提供的大语言模型高性能推理优化框架 +- [SGLang](https://github.com/sgl-project/sglang) ⭐ 27.5k - 大语言模型和多模态模型的高性能服务框架 +- [TensorRT-LLM](https://github.com/NVIDIA/TensorRT-LLM) ⭐ 13.5k - NVIDIA 提供的大语言模型高性能推理优化框架 - [Xorbits Inference](https://github.com/xorbitsai/inference) ⭐ 9.2k - 通过更改一行代码即可替换 GPT 为任何 LLM 的推理 API -- [CooperBench](https://github.com/cooperbench/CooperBench) ⭐ 11 - AI 模型性能基准测试 +- [CooperBench](https://github.com/cooperbench/CooperBench) ⭐ 12 - AI 模型性能基准测试 ### SDK / 开发工具 -- [Remotion](https://github.com/remotion-dev/remotion) ⭐ 42.4k - 使用 React 以编程方式制作视频 -- [docker-credential-pass](https://github.com/docker/docker-credential-helpers) ⭐ 1.3k - 通过存储在平台密钥库中来保护 Docker 登录凭据 -- [Anthropic SDK Java](https://github.com/anthropics/anthropic-sdk-java) ⭐ 281 - Anthropic Java SDK +- [Remotion](https://github.com/remotion-dev/remotion) ⭐ 46.2k - 使用 React 以编程方式制作视频 +- [docker-credential-pass](https://github.com/docker/docker-credential-helpers) ⭐ 1.2k - 通过存储在平台密钥库中来保护 Docker 登录凭据 +- [Anthropic SDK Java](https://github.com/anthropics/anthropic-sdk-java) ⭐ 305 - Anthropic Java SDK ### 设计工具 @@ -73,15 +73,15 @@ ### 资源集合 -- [Jujutsu](https://github.com/jj-vcs/jj) ⭐ 27.7k - 既简单又强大的 Git 兼容版本控制系统 -- [gitbucket](https://github.com/gitbucket/gitbucket) ⭐ 9.4k - 由 Scala 驱动的 Git 平台,易于安装、高可扩展性且兼容 GitHub API -- [top-rss-list](https://github.com/weekend-project-space/top-rss-list) ⭐ 6.1k - 订阅人数最多的中文优质 RSS 源列表 +- [Jujutsu](https://github.com/jj-vcs/jj) ⭐ 28.6k - 既简单又强大的 Git 兼容版本控制系统 +- [gitbucket](https://github.com/gitbucket/gitbucket) ⭐ 9.3k - 由 Scala 驱动的 Git 平台,易于安装、高可扩展性且兼容 GitHub API +- [top-rss-list](https://github.com/weekend-project-space/top-rss-list) ⭐ 6.2k - 订阅人数最多的中文优质 RSS 源列表 - [awesome-n8n](https://github.com/restyler/awesome-n8n) ⭐ 2.8k - 有用的 n8n 资源:社区节点和教程列表 ### AllenAI -- [OLMo](https://github.com/allenai/OLMo-core) ⭐ 1.1k - OLMo 生态系统的 PyTorch 构建块 -- [olmocr](https://github.com/allenai/olmocr) ⭐ 17.1k - 用于线性化 PDF 以进行 LLM 数据集/训练的工具包 +- [OLMo](https://github.com/allenai/OLMo-core) ⭐ 1.2k - OLMo 生态系统的 PyTorch 构建块 +- [olmocr](https://github.com/allenai/olmocr) ⭐ 17.2k - 用于线性化 PDF 以进行 LLM 数据集/训练的工具包 ### 原型 @@ -95,39 +95,3 @@ - llama.cpp 解决了"如何在普通硬件上跑得飞快" - KTransformers 解决了"如何用有限显存跑大模型" - [LMSYS Projects](https://lmsys.org/projects/) - -## 实践 - -- 提出需求,让 AI 来角色扮演,让它列出 N 个功能点,然后 AI 就会根据它的理解生成一个需求列表,接下来你再和 AI 对这些需求逐个打磨,这其实是一个高效冷启动的方法。 -- 在需求提出后,在规划阶段(Planning)不要给 AI 太具体的方案,而是让 AI 来生成方案,你只需要关注最终你想要的结果。 -- 提前告诉 AI 有哪些基础设施和环境的问题,让它少走弯路。 -- 提出需求的第一阶段要求 Agent 做的一些关键动作。 - - 比如无论接下来做什么,都要把计划和 todo 列表放在一个 work.md 或 todo.md 这类文件里。 - - 每完成一个阶段的工作,就把上一阶段的经验教训更新到 agents.md 里。 - - 当一个计划完成并且代码合并后,把这个工作的设计文档添加到项目的知识库中。 -- 在 Research 的过程中,告诉模型它拥有无限的预算和时间,尽可能充分地进行调研。 -- 如果模型有推理深度的参数,在这个阶段把它们全部调到最高级别。在这个阶段多烧一些 token、做好更好的规划、了解更多上下文,对后续的实现阶段会更有帮助。 -- 实现阶段要注意,要么你就让 AI 完全去做,要么你就完全自己做,千万别混着来,目前是倾向于完全零人工干预的模式效果更好。 -- 在测试和验收结果的阶段。你只要知道如何评估和测试你要的东西,AI 就一定能把东西给你做出来。 -- AI 不擅长集成测试、端到端测试。 -- 在完成大目标前,先和 AI 一起做一个方便的集成测试框架,并提前准备好测试的基础设施,收集和生成一些现成集成测试的用例,尽量一键能运行那种,这样在开发阶段就能事半功倍。 -- 关于如何使用这些测试的基础设施的信息,在正式开始前就固化在 agents.md 里。 -- 让 AI 帮你生成测试,但一定要告诉它一些生成的逻辑,标准和目的,另外就是千万不要把生成测试的 Context 和实际进行开发工作的 Agent 的 Context 混在一起。 -- 当前的 Coding Agent 在面对单一模块复杂度超过大约 5 万行代码之后,就开始很难把问题一次性解决掉,只要任务复杂度控制在这个阈值之下,在一个足够好的提示词驱动下,很多事情确实可以做到一次过。 - -/speckit.constitution -1. @docs/req.md @docs/tech.md @docs/ui.md 这3个文件是用户需求来源。 -2. 遵循敏捷开发方式。 -3. 项目的开发环境运行在 VS Code 的开发容器中,容器名是 `ai193`,对应的镜像是`arno:devcontainer-arno`,该镜像由 @.devcontainer/Dockerfile 构建。 -4. 项目根目录: `/home/arno/workspace/ai193`。 -5. 无论接下来做什么,都要把计划和 todo 列表放在 @docs/work.md 或者 @docs/todo.md 这类文件里。 -6. 每完成一个阶段的工作,就把上一阶段的经验教训更新到 @AGENTS.md 里。**注意:** @CLAUDE.md 是链接到 @AGENTS.md 文件的。 -7. 当进行 Research 的过程中,你拥有无限的预算和时间,尽可能充分地进行调研。 -8. 当单一模块复杂度超过大约 5 万行代码之后,需要拆分模块来实现。 - -/speckit.constitution -1. 当需要简化代码的时候,可以使用 `@"code-simplifier:code-simplifier (agent)"` -2. 当需要绘制架构图、甘特图、UML图、部署图、流程图、时序图、序列图等编码相关的图时,使用 'Mermaid' 格式 -3. 文档统一保存在 `@docs/` 目录,不要在其他地方保存 -4. shell脚本统一保存在 `@scripts/` 目录,不要在其他地方保存 -5. 执行任何命令前,必须先通过 `pwd` 确认当前工作目录是否是要执行命令所在的路径,避免路径错误