[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"similar-massgen--MassGen":3,"tool-massgen--MassGen":61},[4,18,26,36,44,53],{"id":5,"name":6,"github_repo":7,"description_zh":8,"stars":9,"difficulty_score":10,"last_commit_at":11,"category_tags":12,"status":17},4358,"openclaw","openclaw\u002Fopenclaw","OpenClaw 是一款专为个人打造的本地化 AI 助手，旨在让你在自己的设备上拥有完全可控的智能伙伴。它打破了传统 AI 助手局限于特定网页或应用的束缚，能够直接接入你日常使用的各类通讯渠道，包括微信、WhatsApp、Telegram、Discord、iMessage 等数十种平台。无论你在哪个聊天软件中发送消息，OpenClaw 都能即时响应，甚至支持在 macOS、iOS 和 Android 设备上进行语音交互，并提供实时的画布渲染功能供你操控。\n\n这款工具主要解决了用户对数据隐私、响应速度以及“始终在线”体验的需求。通过将 AI 部署在本地，用户无需依赖云端服务即可享受快速、私密的智能辅助，真正实现了“你的数据，你做主”。其独特的技术亮点在于强大的网关架构，将控制平面与核心助手分离，确保跨平台通信的流畅性与扩展性。\n\nOpenClaw 非常适合希望构建个性化工作流的技术爱好者、开发者，以及注重隐私保护且不愿被单一生态绑定的普通用户。只要具备基础的终端操作能力（支持 macOS、Linux 及 Windows WSL2），即可通过简单的命令行引导完成部署。如果你渴望拥有一个懂你",349277,3,"2026-04-06T06:32:30",[13,14,15,16],"Agent","开发框架","图像","数据工具","ready",{"id":19,"name":20,"github_repo":21,"description_zh":22,"stars":23,"difficulty_score":10,"last_commit_at":24,"category_tags":25,"status":17},3808,"stable-diffusion-webui","AUTOMATIC1111\u002Fstable-diffusion-webui","stable-diffusion-webui 是一个基于 Gradio 构建的网页版操作界面，旨在让用户能够轻松地在本地运行和使用强大的 Stable Diffusion 图像生成模型。它解决了原始模型依赖命令行、操作门槛高且功能分散的痛点，将复杂的 AI 绘图流程整合进一个直观易用的图形化平台。\n\n无论是希望快速上手的普通创作者、需要精细控制画面细节的设计师，还是想要深入探索模型潜力的开发者与研究人员，都能从中获益。其核心亮点在于极高的功能丰富度：不仅支持文生图、图生图、局部重绘（Inpainting）和外绘（Outpainting）等基础模式，还独创了注意力机制调整、提示词矩阵、负向提示词以及“高清修复”等高级功能。此外，它内置了 GFPGAN 和 CodeFormer 等人脸修复工具，支持多种神经网络放大算法，并允许用户通过插件系统无限扩展能力。即使是显存有限的设备，stable-diffusion-webui 也提供了相应的优化选项，让高质量的 AI 艺术创作变得触手可及。",162132,"2026-04-05T11:01:52",[14,15,13],{"id":27,"name":28,"github_repo":29,"description_zh":30,"stars":31,"difficulty_score":32,"last_commit_at":33,"category_tags":34,"status":17},1381,"everything-claude-code","affaan-m\u002Feverything-claude-code","everything-claude-code 是一套专为 AI 编程助手（如 Claude Code、Codex、Cursor 等）打造的高性能优化系统。它不仅仅是一组配置文件，而是一个经过长期实战打磨的完整框架，旨在解决 AI 代理在实际开发中面临的效率低下、记忆丢失、安全隐患及缺乏持续学习能力等核心痛点。\n\n通过引入技能模块化、直觉增强、记忆持久化机制以及内置的安全扫描功能，everything-claude-code 能显著提升 AI 在复杂任务中的表现，帮助开发者构建更稳定、更智能的生产级 AI 代理。其独特的“研究优先”开发理念和针对 Token 消耗的优化策略，使得模型响应更快、成本更低，同时有效防御潜在的攻击向量。\n\n这套工具特别适合软件开发者、AI 研究人员以及希望深度定制 AI 工作流的技术团队使用。无论您是在构建大型代码库，还是需要 AI 协助进行安全审计与自动化测试，everything-claude-code 都能提供强大的底层支持。作为一个曾荣获 Anthropic 黑客大奖的开源项目，它融合了多语言支持与丰富的实战钩子（hooks），让 AI 真正成长为懂上",145895,2,"2026-04-08T11:32:59",[14,13,35],"语言模型",{"id":37,"name":38,"github_repo":39,"description_zh":40,"stars":41,"difficulty_score":32,"last_commit_at":42,"category_tags":43,"status":17},2271,"ComfyUI","Comfy-Org\u002FComfyUI","ComfyUI 是一款功能强大且高度模块化的视觉 AI 引擎，专为设计和执行复杂的 Stable Diffusion 图像生成流程而打造。它摒弃了传统的代码编写模式，采用直观的节点式流程图界面，让用户通过连接不同的功能模块即可构建个性化的生成管线。\n\n这一设计巧妙解决了高级 AI 绘图工作流配置复杂、灵活性不足的痛点。用户无需具备编程背景，也能自由组合模型、调整参数并实时预览效果，轻松实现从基础文生图到多步骤高清修复等各类复杂任务。ComfyUI 拥有极佳的兼容性，不仅支持 Windows、macOS 和 Linux 全平台，还广泛适配 NVIDIA、AMD、Intel 及苹果 Silicon 等多种硬件架构，并率先支持 SDXL、Flux、SD3 等前沿模型。\n\n无论是希望深入探索算法潜力的研究人员和开发者，还是追求极致创作自由度的设计师与资深 AI 绘画爱好者，ComfyUI 都能提供强大的支持。其独特的模块化架构允许社区不断扩展新功能，使其成为当前最灵活、生态最丰富的开源扩散模型工具之一，帮助用户将创意高效转化为现实。",108111,"2026-04-08T11:23:26",[14,15,13],{"id":45,"name":46,"github_repo":47,"description_zh":48,"stars":49,"difficulty_score":32,"last_commit_at":50,"category_tags":51,"status":17},4721,"markitdown","microsoft\u002Fmarkitdown","MarkItDown 是一款由微软 AutoGen 团队打造的轻量级 Python 工具，专为将各类文件高效转换为 Markdown 格式而设计。它支持 PDF、Word、Excel、PPT、图片（含 OCR）、音频（含语音转录）、HTML 乃至 YouTube 链接等多种格式的解析，能够精准提取文档中的标题、列表、表格和链接等关键结构信息。\n\n在人工智能应用日益普及的今天，大语言模型（LLM）虽擅长处理文本，却难以直接读取复杂的二进制办公文档。MarkItDown 恰好解决了这一痛点，它将非结构化或半结构化的文件转化为模型“原生理解”且 Token 效率极高的 Markdown 格式，成为连接本地文件与 AI 分析 pipeline 的理想桥梁。此外，它还提供了 MCP（模型上下文协议）服务器，可无缝集成到 Claude Desktop 等 LLM 应用中。\n\n这款工具特别适合开发者、数据科学家及 AI 研究人员使用，尤其是那些需要构建文档检索增强生成（RAG）系统、进行批量文本分析或希望让 AI 助手直接“阅读”本地文件的用户。虽然生成的内容也具备一定可读性，但其核心优势在于为机器",93400,"2026-04-06T19:52:38",[52,14],"插件",{"id":54,"name":55,"github_repo":56,"description_zh":57,"stars":58,"difficulty_score":10,"last_commit_at":59,"category_tags":60,"status":17},4487,"LLMs-from-scratch","rasbt\u002FLLMs-from-scratch","LLMs-from-scratch 是一个基于 PyTorch 的开源教育项目，旨在引导用户从零开始一步步构建一个类似 ChatGPT 的大型语言模型（LLM）。它不仅是同名技术著作的官方代码库，更提供了一套完整的实践方案，涵盖模型开发、预训练及微调的全过程。\n\n该项目主要解决了大模型领域“黑盒化”的学习痛点。许多开发者虽能调用现成模型，却难以深入理解其内部架构与训练机制。通过亲手编写每一行核心代码，用户能够透彻掌握 Transformer 架构、注意力机制等关键原理，从而真正理解大模型是如何“思考”的。此外，项目还包含了加载大型预训练权重进行微调的代码，帮助用户将理论知识延伸至实际应用。\n\nLLMs-from-scratch 特别适合希望深入底层原理的 AI 开发者、研究人员以及计算机专业的学生。对于不满足于仅使用 API，而是渴望探究模型构建细节的技术人员而言，这是极佳的学习资源。其独特的技术亮点在于“循序渐进”的教学设计：将复杂的系统工程拆解为清晰的步骤，配合详细的图表与示例，让构建一个虽小但功能完备的大模型变得触手可及。无论你是想夯实理论基础，还是为未来研发更大规模的模型做准备",90106,"2026-04-06T11:19:32",[35,15,13,14],{"id":62,"github_repo":63,"name":64,"description_en":65,"description_zh":66,"ai_summary_zh":66,"readme_en":67,"readme_zh":68,"quickstart_zh":69,"use_case_zh":70,"hero_image_url":71,"owner_login":72,"owner_name":72,"owner_avatar_url":73,"owner_bio":74,"owner_company":74,"owner_location":74,"owner_email":74,"owner_twitter":74,"owner_website":74,"owner_url":75,"languages":76,"stars":107,"forks":108,"last_commit_at":109,"license":110,"difficulty_score":32,"env_os":111,"env_gpu":112,"env_ram":113,"env_deps":114,"category_tags":121,"github_topics":122,"view_count":32,"oss_zip_url":74,"oss_zip_packed_at":74,"status":17,"created_at":139,"updated_at":140,"faqs":141,"releases":142},5600,"massgen\u002FMassGen","MassGen","🚀 MassGen is an open-source multi-agent scaling system that runs in your terminal, autonomously orchestrating frontier models and agents to collaborate, reason, and produce high-quality results. | Join us on Discord: discord.massgen.ai","MassGen 是一款运行在终端的开源多智能体扩展系统，旨在通过协作式人工智能解决复杂任务。它不再依赖单个模型“单打独斗”，而是自主编排多个前沿 AI 模型与智能体，让它们并行工作、相互观察、批判并迭代优化彼此的结果。当智能体们认为已找到足够好的答案时，会通过投票机制达成共识，最终输出集体验证的高质量结果。\n\n这一机制有效解决了传统单代理系统在应对高难度问题时容易出现的逻辑断层或质量不稳定痛点，利用冗余处理和集体智慧显著提升了输出的可靠性与速度（实测可达 4 倍加速）。其核心理念源自“思维线程”与“迭代优化”，并扩展了经典的多智能体对话框架。\n\nMassGen 特别适合开发者、AI 研究人员以及需要处理复杂逻辑推理任务的技术团队使用。它不仅支持跨模型协同和实时可视化监控，还能作为技能插件无缝集成到 Claude Code、Cursor 等 40 多种主流开发工具中。对于希望探索多智能体规模化应用、追求更高代码质量或研究共识机制的用户来说，MassGen 提供了一个原则清晰且易于上手的高效平台。","\u003Cp align=\"center\">\n \u003Cpicture>\n \u003Csource media=\"(prefers-color-scheme: dark)\" srcset=\"assets\u002Flogo-dark.png\">\n \u003Csource media=\"(prefers-color-scheme: light)\" srcset=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Fmassgen_MassGen_readme_ade0b41da867.png\">\n \u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Fmassgen_MassGen_readme_ade0b41da867.png\" alt=\"MassGen Logo\" width=\"360\" \u002F>\n \u003C\u002Fpicture>\n\u003C\u002Fp>\n\n\u003Cdiv align=\"center\">\n\n[![PyPI](https:\u002F\u002Fimg.shields.io\u002Fpypi\u002Fv\u002Fmassgen?style=flat-square&logo=pypi&logoColor=white&label=PyPI&color=3775A9)](https:\u002F\u002Fpypi.org\u002Fproject\u002Fmassgen\u002F)\n[![Docs](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fdocs-massgen.ai-blue?style=flat-square&logo=readthedocs&logoColor=white)](https:\u002F\u002Fdocs.massgen.ai)\n[![GitHub Stars](https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FLeezekun\u002FMassGen?style=flat-square&logo=github&color=181717&logoColor=white)](https:\u002F\u002Fgithub.com\u002FLeezekun\u002FMassGen)\n[![Python 3.11+](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fpython-3.11+-3776AB?style=flat-square&logo=python&logoColor=white)](https:\u002F\u002Fwww.python.org\u002Fdownloads\u002F)\n[![License](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Flicense-Apache%202.0-green?style=flat-square)](LICENSE)\n\n\u003C\u002Fdiv>\n\n\u003Cdiv align=\"center\">\n\n[![Follow on X](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FFOLLOW%20ON%20X-000000?style=for-the-badge&logo=x&logoColor=white)](https:\u002F\u002Fx.massgen.ai)\n[![Follow on LinkedIn](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FFOLLOW%20ON%20LINKEDIN-0A66C2?style=for-the-badge&logo=linkedin&logoColor=white)](https:\u002F\u002Fwww.linkedin.com\u002Fcompany\u002Fmassgen-ai)\n[![Join our Discord](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FJOIN%20OUR%20DISCORD-5865F2?style=for-the-badge&logo=discord&logoColor=white)](https:\u002F\u002Fdiscord.massgen.ai)\n\n\u003C\u002Fdiv>\n\n\u003Ch1 align=\"center\">🚀 MassGen: Multi-Agent Scaling System for GenAI\u003C\u002Fh1>\n\n\u003Cp align=\"center\">\n \u003Ci>MassGen is a cutting-edge multi-agent system that leverages the power of collaborative AI to solve complex tasks.\u003C\u002Fi>\n\u003C\u002Fp>\n\n\u003Cp align=\"center\">\n \u003Ca href=\"https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=5JofXWf_Ok8\">\n \u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Fmassgen_MassGen_readme_0433a8f0466f.gif\" alt=\"MassGen example\" width=\"800\">\n \u003C\u002Fa>\n\u003C\u002Fp>\n\n\u003Cp align=\"center\">\n \u003Ci>Scaling AI with collaborative, continuously improving agents (4x speed)\u003C\u002Fi>\n\u003C\u002Fp>\n\nMassGen is a cutting-edge multi-agent framework that coordinates AI agents to solve complex tasks through redundancy and iterative refinement. Every agent tackles the full problem, observing, critiquing, and building on each other's work across cycles of refinement and restarts. When agents believe there is a strong enough answer, they vote, and the best collectively validated answer wins. This approach to parallel refinement and collective validation lays the groundwork for principled multi-agent scaling, where the system continuously improves its outputs by leveraging diverse agent perspectives and enforcing quality through consensus.\n\nThis project started with the \"threads of thought\" and \"iterative refinement\" ideas presented in [The Myth of Reasoning](https:\u002F\u002Fdocs.ag2.ai\u002Flatest\u002Fdocs\u002Fblog\u002F2025\u002F04\u002F16\u002FReasoning\u002F), and extends the classic \"multi-agent conversation\" idea in [AG2](https:\u002F\u002Fgithub.com\u002Fag2ai\u002Fag2). Here is a [video recording](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=xM2Uguw1UsQ) of the background context introduction presented at the Berkeley Agentic AI Summit 2025.\n\n\u003Cp align=\"center\">\n \u003Cb>🧩 Use MassGen as a Skill:\u003C\u002Fb> \u003Ccode>npx skills add massgen\u002Fskills --all\u003C\u002Fcode> — then type invoke the skill in Claude Code, Cursor, Copilot, or 40+ other agents. \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fmassgen\u002Fskills\">Learn more →\u003C\u002Fa>\n\u003C\u002Fp>\n\n\u003Cp align=\"center\">\n \u003Cb>📚 For Contributors:\u003C\u002Fb> See \u003Ca href=\"https:\u002F\u002Fmassgen.github.io\u002FHandbook\u002F\">MassGen Contributor Handbook\u003C\u002Fa> - Centralized policies and resources for development and research teams\n\u003C\u002Fp>\n\n---\n\n## 📋 Table of Contents\n\n\u003Cdetails open>\n\u003Csummary>\u003Ch3>✨ Key Features\u003C\u002Fh3>\u003C\u002Fsummary>\n\n- [Cross-Model\u002FAgent Synergy](#-key-features-1)\n- [Parallel Processing](#-key-features-1)\n- [Intelligence Sharing](#-key-features-1)\n- [Consensus Building](#-key-features-1)\n- [Live Visualization](#-key-features-1)\n\u003C\u002Fdetails>\n\n\u003Cdetails open>\n\u003Csummary>\u003Ch3>🆕 Latest Features\u003C\u002Fh3>\u003C\u002Fsummary>\n\n- [v0.1.73 Features](#-latest-features-v0173)\n\u003C\u002Fdetails>\n\n\u003Cdetails open>\n\u003Csummary>\u003Ch3>🏗️ System Design\u003C\u002Fh3>\u003C\u002Fsummary>\n\n- [System Architecture](#%EF%B8%8F-system-design-1)\n- [Parallel Processing](#%EF%B8%8F-system-design-1)\n- [Real-time Collaboration](#%EF%B8%8F-system-design-1)\n- [Convergence Detection](#%EF%B8%8F-system-design-1)\n- [Adaptive Coordination](#%EF%B8%8F-system-design-1)\n\u003C\u002Fdetails>\n\n\u003Cdetails open>\n\u003Csummary>\u003Ch3>🚀 Quick Start\u003C\u002Fh3>\u003C\u002Fsummary>\n\n- [📥 Installation](#1--installation)\n- [🔐 API Configuration](#2--api-configuration)\n- [🧩 Supported Models and Tools](#3--supported-models-and-tools)\n - [Models](#models)\n - [Tools](#tools)\n- [🏃 Run MassGen](#4--run-massgen)\n - [CLI Configuration Parameters](#cli-configuration-parameters)\n - [1. Single Agent (Easiest Start)](#1-single-agent-easiest-start)\n - [2. Multi-Agent Collaboration (Recommended)](#2-multi-agent-collaboration-recommended)\n - [3. Model Context Protocol (MCP)](#3-model-context-protocol-mcp)\n - [4. File System Operations](#4-file-system-operations--workspace-management)\n - [5. Project Integration (NEW in v0.0.21)](#5-project-integration--user-context-paths-new-in-v0021)\n - [Backend Configuration Reference](#backend-configuration-reference)\n - [Interactive Multi-Turn Mode](#interactive-multi-turn-mode)\n- [📊 View Results](#5--view-results)\n - [Real-time Display](#real-time-display)\n - [Comprehensive Logging](#comprehensive-logging)\n\u003C\u002Fdetails>\n\n\u003Cdetails open>\n\u003Csummary>\u003Ch3>🤖 Automation & LLM Integration\u003C\u002Fh3>\u003C\u002Fsummary>\n\n- [Automation Mode](#-automation--llm-integration)\n- [BackgroundShellManager](#using-backgroundshellmanager)\n- [Status File Reference](#statusjson-structure)\n- [Full Automation Guide](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Fautomation.html)\n\u003C\u002Fdetails>\n\n\u003Cdetails open>\n\u003Csummary>\u003Ch3>💡 Case Studies & Examples\u003C\u002Fh3>\u003C\u002Fsummary>\n\n- [Case Studies](#-case-studies)\n\u003C\u002Fdetails>\n\n\u003Cdetails open>\n\u003Csummary>\u003Ch3>🗺️ Roadmap\u003C\u002Fh3>\u003C\u002Fsummary>\n\n- [Recent Achievements (v0.1.73)](#recent-achievements-v0173)\n- [Previous Achievements (v0.0.3 - v0.1.72)](#previous-achievements-v003---v0172)\n- [Key Future Enhancements](#key-future-enhancements)\n - Bug Fixes & Backend Improvements\n - Advanced Agent Collaboration\n - Expanded Model, Tool & Agent Integrations\n - Improved Performance & Scalability\n - Enhanced Developer Experience\n- [v0.1.74 Roadmap](#v0174-roadmap)\n\u003C\u002Fdetails>\n\n\u003Cdetails open>\n\u003Csummary>\u003Ch3>📚 Additional Resources\u003C\u002Fh3>\u003C\u002Fsummary>\n\n- [🤝 Contributing](#-contributing)\n- [📄 License](#-license)\n- [⭐ Star History](#-star-history)\n\u003C\u002Fdetails>\n\n---\n\n## ✨ Key Features\n\n| Feature | Description |\n|---------|-------------|\n| **🤝 Cross-Model\u002FAgent Synergy** | Harness strengths from diverse frontier model-powered agents |\n| **⚡ Parallel Processing** | Multiple agents tackle problems simultaneously |\n| **👥 Intelligence Sharing** | Agents share and learn from each other's work |\n| **🔄 Consensus Building** | Natural convergence through collaborative refinement |\n| **🖥️ Live Visualization** | Interactive Textual TUI with timeline, agent cards, and vote tracking (default). Also available: Web UI, Rich display. |\n\n---\n\n## 🆕 Latest Features (v0.1.73)\n\n**🎉 Released: April 6, 2026**\n\n**What's New in v0.1.73:**\n- **🧬 Eval Criteria Evolver Subagent** - New subagent type that evolves evaluation criteria across rounds.\n- **🛡️ Checkpoint Objective Mode** - Initial draft of checkpoint MCP `objective` mode for safety planning of irreversible actions.\n- **👁️ Improved Eval Criteria Visibility** - Clearer visibility into what criteria agents are working against.\n\n**Try v0.1.73 Features:**\n```bash\npip install massgen==0.1.73\nuv run massgen --config @examples\u002Ffeatures\u002Ftrace_analyzer_background.yaml \"Create an svg of an AI agent coding.\"\n```\n\n→ [See full release history and examples](massgen\u002Fconfigs\u002FREADME.md#release-history--examples)\n\n---\n\n## 🏗️ System Design\n\nMassGen operates through an architecture designed for **seamless multi-agent collaboration**:\n\n```mermaid\ngraph TB\n O[🚀 MassGen Orchestrator\u003Cbr\u002F>📋 Task Distribution & Coordination]\n\n subgraph Collaborative Agents\n A1[Agent 1\u003Cbr\u002F>🏗️ Anthropic\u002FClaude + Tools]\n A2[Agent 2\u003Cbr\u002F>🌟 Google\u002FGemini + Tools]\n A3[Agent 3\u003Cbr\u002F>🤖 OpenAI\u002FGPT + Tools]\n A4[Agent 4\u003Cbr\u002F>⚡ xAI\u002FGrok + Tools]\n end\n\n H[🔄 Shared Collaboration Hub\u003Cbr\u002F>📡 Real-time Notification & Consensus]\n\n O --> A1 & A2 & A3 & A4\n A1 & A2 & A3 & A4 \u003C--> H\n\n classDef orchestrator fill:#e1f5fe,stroke:#0288d1,stroke-width:3px\n classDef agent fill:#f3e5f5,stroke:#7b1fa2,stroke-width:2px\n classDef hub fill:#e8f5e8,stroke:#388e3c,stroke-width:2px\n\n class O orchestrator\n class A1,A2,A3,A4 agent\n class H hub\n```\n\nThe system's workflow is defined by the following key principles:\n\n**Parallel Processing** - Multiple agents tackle the same task simultaneously, each leveraging their unique capabilities (different models, tools, and specialized approaches).\n\n**Real-time Collaboration** - Agents continuously share their working summaries and insights through a notification system, allowing them to learn from each other's approaches and build upon collective knowledge.\n\n**Convergence Detection** - The system intelligently monitors when agents have reached stability in their solutions and achieved consensus through natural collaboration rather than forced agreement.\n\n**Adaptive Coordination** - Agents can restart and refine their work when they receive new insights from others, creating a dynamic and responsive problem-solving environment.\n\nThis collaborative approach ensures that the final output leverages collective intelligence from multiple AI systems, leading to more robust and well-rounded results than any single agent could achieve alone.\n\n---\n\n> 📖 **Complete Documentation:** For comprehensive guides, API reference, and detailed examples, visit **[MassGen Official Documentation](https:\u002F\u002Fdocs.massgen.ai\u002F)**\n\n---\n\n## 🚀 Quick Start\n\n### 1. 📥 Installation\n\n**Method 1: PyPI Installation** (Recommended - Python 3.11+):\n\n```bash\n# Install MassGen via pip\npip install massgen\n\n# Or with uv (faster)\npip install uv\nuv venv && source .venv\u002Fbin\u002Factivate\nuv pip install massgen\n\n# If you install massgen in uv, make sure you either activate your venv using source .venv\u002Fbin\u002Factivate\n# Or include \"uv run\" before all commands\n```\n\n**Quickstart Setup** (Fastest way to get running):\n\n```bash\n# Step 1: Set up API keys, Docker, and skills\nuv run massgen --setup\n\n# Step 2: Create a simple config and start\nuv run massgen --quickstart\n```\n\nThe `--setup` command will:\n- Configure your API keys (OpenAI, Anthropic, Google, xAI)\n- Offer to set up Docker images for code execution\n- Offer to install skills (openskills, Anthropic\u002FOpenAI\u002FVercel collections, Agent Browser skill, Crawl4AI)\n\nThe `--quickstart` command will:\n- Ask how many agents you want (1-5, default 3)\n- Ask which backend\u002Fmodel for each agent\n- For GPT-5x models, ask for `reasoning.effort` (`low|medium|high`; Codex GPT-5 models also include `xhigh`)\n- Auto-detect Docker availability and configure execution mode\n- If Docker mode is selected, show a Skills step where you can choose package(s) (`openskills`-based Anthropic\u002FOpenAI\u002FVercel\u002FAgent Browser plus Crawl4AI) and install them in-place with live status\n- Create a ready-to-use config and launch into interactive TUI mode\n\n**🤖 Use MassGen from Your AI Coding Agent:**\n\nInstall the [MassGen skill](https:\u002F\u002Fgithub.com\u002Fmassgen\u002Fskills) to invoke MassGen directly from Claude Code, OpenAI Codex, GitHub Copilot, Cursor, and [40+ other agents](https:\u002F\u002Fskills.sh) that support the [Agent Skills](https:\u002F\u002Fagentskills.io\u002Fhome) standard:\n\n```bash\nnpx skills add massgen\u002Fskills\n```\n\nThen use `\u002Fmassgen` (Claude Code) or `$massgen` (Codex) to run multi-agent evaluation, planning, spec writing, or any general task. See the [skills docs](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Fskills.html) for per-agent install options.\n\n**🖥️ Textual TUI (Default Display Mode):**\n\nMassGen launches with an interactive Terminal User Interface (TUI) by default, providing:\n- 📊 **Real-time timeline** of all agent activities\n- 🎯 **Individual agent status cards** for each team member\n- 🗳️ **Vote visualization** and consensus tracking\n- 💬 **Multi-turn conversation** management\n- ⌨️ **Keyboard controls** for navigation (↑\u002F↓ to scroll, 'q' to cancel)\n\n**Legacy Rich display:**\n```bash\nmassgen --display rich \"Your question\"\n```\n\n**Alternative: Full Setup Wizard**\n\nFor more control, use the full configuration wizard:\n```bash\nuv run massgen --init\n```\n\nThis guides you through use case selection (Research, Code, Q&A, etc.) and advanced configuration options.\n\n**After setup:**\n```bash\n# Interactive mode\nuv run massgen\n\n# Single query\nuv run massgen \"Your question here\"\n\n# With example configurations\nuv run massgen --config @examples\u002Fbasic\u002Fmulti\u002Fthree_agents_default \"Your question\"\n```\n\n→ See [Installation Guide](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fquickstart\u002Finstallation.html) for complete setup instructions.\n\n**Method 2: Development Installation** (for contributors):\n\n**Clone the repository**\n```bash\ngit clone https:\u002F\u002Fgithub.com\u002FLeezekun\u002FMassGen.git\ncd MassGen\n```\n\n**Install in editable mode with pip**\n\n**Option 1 (recommended): Installing with uv (faster)**\n\n```bash\nuv venv\nsource .venv\u002Fbin\u002Factivate # Windows: .venv\\Scripts\\activate\nuv pip install -e .\n\n# If you install massgen in uv, make sure you either activate your venv using source .venv\u002Fbin\u002Factivate\n# Or include \"uv run\" before all commands\n\n# Automated setup (works on all platforms) - installs dependencies, skills, Docker images, also sets up API keys\nuv run massgen --setup\n\n# Or use the bash script (Unix\u002FLinux\u002FmacOS only), need manually config API keys, see sections below\nuv run .\u002Fscripts\u002Finit.sh\n\n# If you would like to install other dependencies later\n# Here is a light-weighted setup script which only installs skills (works on all platforms)\nuv run massgen --setup-skills\n\n# Or use the bash script (Unix\u002FLinux\u002FmacOS only)\nuv run .\u002Fscripts\u002Finit_skills.sh\n```\n\n**Option 2: Using traditional Python env**\n\n```bash\npip install -e .\n\n# Optional: External framework integration\npip install -e \".[external]\"\n\n# Automated setup (works on all platforms) - installs dependencies, skills, Docker images, also sets up API keys\nmassgen --setup\n\n# Or use the bash script (Unix\u002FLinux\u002FmacOS only), need manually config API keys, see sections below\n.\u002Fscripts\u002Finit.sh\n\n# If you would like to install other dependencies later\n# Here is a light-weighted setup script which only installs skills (works on all platforms)\nmassgen --setup-skills\n\n# Or use the bash script (Unix\u002FLinux\u002FmacOS only)\n.\u002Fscripts\u002Finit_skills.sh\n```\n\n> **Note:** The `--setup` and `--setup-skills` commands work cross-platform (Windows, macOS, Linux). The bash scripts (`init.sh`, `init_skills.sh`) are Unix-only but provide additional dev setup like Docker image builds.\n\n\u003Cdetails>\n\u003Csummary>\u003Cb>Alternative Installation Methods\u003C\u002Fb> (click to expand)\u003C\u002Fsummary>\n\n**Using uv with venv:**\n```bash\ngit clone https:\u002F\u002Fgithub.com\u002FLeezekun\u002FMassGen.git\ncd MassGen\nuv venv\nsource .venv\u002Fbin\u002Factivate # Windows: .venv\\Scripts\\activate\nuv pip install -e .\n```\n\n**Using traditional Python venv:**\n```bash\ngit clone https:\u002F\u002Fgithub.com\u002FLeezekun\u002FMassGen.git\ncd MassGen\npython -m venv .venv\nsource .venv\u002Fbin\u002Factivate # Windows: .venv\\Scripts\\activate\npip install -e .\n```\n\n**Global installation with uv tool:**\n```bash\ngit clone https:\u002F\u002Fgithub.com\u002FLeezekun\u002FMassGen.git\ncd MassGen\nuv tool install -e .\n# Now run from any directory\nuv tool run massgen --config @examples\u002Fbasic\u002Fmulti\u002Fthree_agents_default \"Question\"\n```\n\n**Backwards compatibility (uv run):**\n```bash\ncd \u002Fpath\u002Fto\u002FMassGen\nuv run massgen --config @examples\u002Fbasic\u002Fmulti\u002Fthree_agents_default \"Question\"\nuv run python -m massgen.cli --config config.yaml \"Question\"\n```\n\n\u003C\u002Fdetails>\n\n**Optional CLI Tools:**\n```bash\n# Claude Code CLI - Advanced coding assistant\nnpm install -g @anthropic-ai\u002Fclaude-code\n\n# LM Studio - Local model inference\n# MacOS\u002FLinux:\nsudo ~\u002F.lmstudio\u002Fbin\u002Flms bootstrap\n# Windows:\ncmd \u002Fc %USERPROFILE%\\.lmstudio\\bin\\lms.exe bootstrap\n```\n\n**After setup:**\n```bash\n# Interactive mode\nuv run massgen\n\n# Single query\nuv run massgen \"Your question here\"\n\n# With example configurations\nuv run massgen --config @examples\u002Fbasic\u002Fmulti\u002Fthree_agents_default \"Your question\"\n```\n\n### 2. 🔐 API Configuration\n\n**Create a `.env` file in your working directory with your API keys:**\n\n```bash\n# Copy this template to .env and add your API keys\nOPENAI_API_KEY=sk-...\nANTHROPIC_API_KEY=sk-ant-...\nGOOGLE_API_KEY=...\nXAI_API_KEY=...\n\n# Optional: Additional providers\nCEREBRAS_API_KEY=...\nTOGETHER_API_KEY=...\nGROQ_API_KEY=...\nOPENROUTER_API_KEY=...\n```\n\nMassGen automatically loads API keys from `.env` in your current directory.\n\n→ **Complete setup guide with all providers:** See [API Key Configuration](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fquickstart\u002Finstallation.html#api-key-configuration) in the docs\n\n**Get API keys:**\n - [OpenAI](https:\u002F\u002Fplatform.openai.com\u002Fapi-keys) | [Claude](https:\u002F\u002Fdocs.anthropic.com\u002Fen\u002Fapi\u002Foverview) | [Gemini](https:\u002F\u002Fai.google.dev\u002Fgemini-api\u002Fdocs) | [Grok](https:\u002F\u002Fdocs.x.ai\u002Fdocs\u002Foverview)\n - [Azure OpenAI](https:\u002F\u002Flearn.microsoft.com\u002Fen-us\u002Fazure\u002Fai-services\u002Fopenai\u002F) | [Cerebras](https:\u002F\u002Finference-docs.cerebras.ai\u002Fintroduction) | [OpenRouter](https:\u002F\u002Fopenrouter.ai\u002Fdocs\u002Fapi\u002Fapi-reference\u002Fapi-keys\u002Fcreate-keys) | [More providers...](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Freference\u002Fsupported_models.html)\n\n### 3. 🧩 Supported Models and Tools\n\n#### Models\n\nThe system currently supports multiple model providers with advanced capabilities:\n\n**API-based Models:**\n- **OpenAI**: GPT-5.2 (recommended default), GPT-5.1, GPT-5 series (GPT-5, GPT-5-mini, GPT-5-nano), GPT-5.1-Codex series, GPT-4.1 series, GPT-4o, o4-mini with reasoning, web search, code interpreter, and computer-use support\n - **Note**: We recommend GPT-5.2\u002F5.1\u002F5 over Codex models. Codex models are [optimized for shorter system messages](https:\u002F\u002Fcookbook.openai.com\u002Fexamples\u002Fgpt-5-codex_prompting_guide) and may not work well with MassGen's coordination prompts.\n - **Reasoning**: GPT-5.1 and GPT-5.2 default to `reasoning: none`. MassGen automatically sets `reasoning.effort: medium` when no reasoning config is provided, matching GPT-5's default behavior.\n- **Azure OpenAI**: Any Azure-deployed models (GPT-4, GPT-4o, GPT-35-turbo, etc.)\n- **Claude \u002F Anthropic**: Claude Opus 4.5, Claude Haiku 4.5, Claude Sonnet 4.5, Claude Opus 4.1, Claude Sonnet 4\n - Advanced tooling: web search, code execution, Files API, programmatic tool calling, tool search with deferred loading\n- **Claude Code**: Native Claude Code SDK with server-side session persistence and built-in dev tools\n- **Gemini**: Gemini 3 Pro, Gemini 2.5 Flash, Gemini 2.5 Pro with code execution and grounding\n- **Grok \u002F xAI**: Grok-4.1, Grok-4, Grok-3, Grok-3-mini with Grok Live Search\n- **Cerebras AI**: Ultra-fast inference for supported models\n- **Together AI**, **Fireworks AI**, **Groq**: Fast inference for LLaMA, Mistral, Qwen, and other open models\n- **OpenRouter**: Multi-model aggregator with dynamic model listing (400+ models)\n- **Kimi \u002F Moonshot**: Chinese AI models via OpenAI-compatible API\n- **Nebius AI Studio**: Cloud inference platform\n- **POE**: Quora AI platform with dynamic model discovery\n- **Qwen \u002F Alibaba**: DashScope API for Qwen models\n- **Z AI \u002F Zhipu**: GLM-4.5 and related models\n\n**Local Model Support:**\n- **vLLM & SGLang**: Unified inference backend supporting both vLLM and SGLang servers\n - vLLM (port 8000) and SGLang (port 30000) with OpenAI-compatible API\n - Support for `top_k`, `repetition_penalty`, `chat_template_kwargs` parameters\n - SGLang-specific `separate_reasoning` parameter for thinking models\n - Mixed server deployments with configuration example: `two_qwen_vllm_sglang.yaml`\n\n- **LM Studio**: Run open-weight models locally with automatic server management\n - Automatic LM Studio CLI installation\n - Auto-download and loading of models\n - Support for LLaMA, Mistral, Qwen and other open-weight models\n\n→ For complete model list and configuration details, see [Supported Models](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Freference\u002Fsupported_models.html)\n\n#### Tools\n\nMassGen agents can leverage various tools to enhance their problem-solving capabilities:\n\n- **Built-in Tools**: Web search, code execution, bash\u002Fshell (provider-dependent)\n- **Filesystem**: Native file operations or via MCP\n- **MCP Integration**: Connect to any MCP server for extended capabilities\n- **Custom Tools**: Define your own tools via YAML configuration\n- **Multimodal**: Image, audio, video understanding and generation (native or via custom tools)\n\n→ For detailed backend capabilities and tool support matrix, see [User Guide - Backends](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Fbackends.html#backend-capabilities)\n\n---\n\n### 4. 🏃 Run MassGen\n\n> **Complete Usage Guide:** For all usage modes, advanced features, and interactive multi-turn sessions, see [Running MassGen](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fquickstart\u002Frunning-massgen.html)\n\n#### 🚀 Getting Started\n\n#### CLI Configuration Parameters\n\n| Parameter | Description |\n|-------------------|-------------|\n| `--config` | Path to YAML configuration file with agent definitions, model parameters, backend parameters and UI settings |\n| `--backend` | Backend type for quick setup without a config file (`claude`, `claude_code`, `gemini`, `grok`, `openai`, `azure_openai`, `zai`). Optional for [models with default backends](massgen\u002Futils.py).|\n| `--model` | Model name for quick setup (e.g., `gemini-2.5-flash`, `gpt-5-nano`, ...). `--config` and `--model` are mutually exclusive - use one or the other. |\n| `--system-message` | System prompt for the agent in quick setup mode. If `--config` is provided, `--system-message` is omitted. |\n| `--cwd-context` | Add current working directory as runtime context path: `ro`\u002F`read` for read-only, `rw`\u002F`write` for write access. In TUI, this initializes the same state as `Ctrl+P`. |\n| `--plan` | Planning-only mode. Agents create a structured task plan without auto-executing it. |\n| `--plan-depth` | Plan granularity for `--plan`: `dynamic`, `shallow`, `medium`, or `deep`. |\n| `--plan-and-execute` | Run both phases: create a plan, then execute it automatically. |\n| `--execute-plan` | Execute an existing plan by path, plan ID, or `latest`. |\n| `--no-display` | Disable real-time streaming UI coordination display (fallback to simple text output).|\n| `--no-logs` | Disable real-time logging.|\n| `--debug` | Enable debug mode with verbose logging (NEW in v0.0.13). Shows detailed orchestrator activities, agent messages, backend operations, and tool calls. Debug logs are saved to `agent_outputs\u002Flog_{time}\u002Fmassgen_debug.log`. |\n| `\"\u003Cyour question>\"` | Optional single-question input; if omitted, MassGen enters interactive chat mode. |\n\n#### **0. OpenAI-Compatible HTTP Server (NEW)**\n\nRun MassGen as an **OpenAI-compatible** HTTP API (FastAPI + Uvicorn). This is useful for integrating MassGen with existing tooling that expects `POST \u002Fv1\u002Fchat\u002Fcompletions`.\n\n```bash\n# Start server (defaults: host 0.0.0.0, port 4000)\nmassgen serve\n\n# With explicit bind + defaults for model\u002Fconfig\nmassgen serve --host 0.0.0.0 --port 4000 --config path\u002Fto\u002Fconfig.yaml --default-model gpt-5\n```\n\n**Endpoints**\n\n- `GET \u002Fhealth`\n- `POST \u002Fv1\u002Fchat\u002Fcompletions` (supports `stream: true` SSE and OpenAI-style tool calling)\n\n**cURL examples**\n\n```bash\n# Health\ncurl http:\u002F\u002Flocalhost:4000\u002Fhealth\n\n# Non-streaming chat completion\ncurl http:\u002F\u002Flocalhost:4000\u002Fv1\u002Fchat\u002Fcompletions \\\n -H \"Content-Type: application\u002Fjson\" \\\n -d '{\n \"model\": \"massgen\",\n \"messages\": [{\"role\": \"user\", \"content\": \"hi\"}],\n \"stream\": false\n }'\n\n# Streaming (Server-Sent Events)\ncurl -N http:\u002F\u002Flocalhost:4000\u002Fv1\u002Fchat\u002Fcompletions \\\n -H \"Content-Type: application\u002Fjson\" \\\n -d '{\n \"model\": \"massgen\",\n \"messages\": [{\"role\": \"user\", \"content\": \"hi\"}],\n \"stream\": true\n }'\n```\n\n**Notes**\n\n- Client-provided `tools` are supported, but tool names that collide with MassGen workflow tools are rejected.\n- Environment variables (optional): `MASSGEN_SERVER_HOST`, `MASSGEN_SERVER_PORT`, `MASSGEN_SERVER_DEFAULT_CONFIG`, `MASSGEN_SERVER_DEFAULT_MODEL`, `MASSGEN_SERVER_DEBUG`.\n\n\n#### **1. Single Agent (Easiest Start)**\n\n**Quick Start Commands:**\n```bash\n# Quick test with any supported model - no configuration needed\nuv run python -m massgen.cli --model claude-sonnet-4-5-20250929 \"What is machine learning?\"\nuv run python -m massgen.cli --model gemini-3-pro-preview \"Explain quantum computing\"\nuv run python -m massgen.cli --model gpt-5-nano \"Summarize the latest AI developments\"\n```\n\n**Configuration:**\n\nUse the `agent` field to define a single agent with its backend and settings:\n\n```yaml\nagent:\n id: \"\u003Cagent_name>\"\n backend:\n type: \"azure_openai\" | \"chatcompletion\" | \"claude\" | \"claude_code\" | \"gemini\" | \"grok\" | \"openai\" | \"zai\" | \"lmstudio\" #Type of backend\n model: \"\u003Cmodel_name>\" # Model name\n api_key: \"\u003Coptional_key>\" # API key for backend. Uses env vars by default.\n system_message: \"...\" # System Message for Single Agent\n```\n\n→ [See all single agent configs](massgen\u002Fconfigs\u002Fbasic\u002Fsingle\u002F)\n\n\n#### **2. Multi-Agent Collaboration (Recommended)**\n\n**Configuration:**\n\nUse the `agents` field to define multiple agents, each with its own backend and config:\n\n**Quick Start Commands:**\n\n```bash\n# Three powerful agents working together - Gemini, GPT-5, and Grok\nmassgen --config @examples\u002Fbasic\u002Fmulti\u002Fthree_agents_default \\\n \"Analyze the pros and cons of renewable energy\"\n```\n\n**This showcases MassGen's core strength:**\n- **Gemini 3 Pro** - Fast research with web search\n- **GPT-5 Nano** - Advanced reasoning with code execution\n- **Grok-4 Fast** - Real-time information and alternative perspectives\n\n```yaml\nagents: # Multiple agents (alternative to 'agent')\n - id: \"\u003Cagent1 name>\"\n backend:\n type: \"azure_openai\" | \"chatcompletion\" | \"claude\" | \"claude_code\" | \"gemini\" | \"grok\" | \"openai\" | \"zai\" | \"lmstudio\" #Type of backend\n model: \"\u003Cmodel_name>\" # Model name\n api_key: \"\u003Coptional_key>\" # API key for backend. Uses env vars by default.\n system_message: \"...\" # System Message for Single Agent\n - id: \"...\"\n backend:\n type: \"...\"\n model: \"...\"\n ...\n system_message: \"...\"\n```\n\n→ [Explore more multi-agent setups](massgen\u002Fconfigs\u002Fbasic\u002Fmulti\u002F)\n\n\n#### **3. Model context protocol (MCP)**\n\nThe [Model context protocol](https:\u002F\u002Fmodelcontextprotocol.io\u002F) (MCP) standardises how applications expose tools and context to language models. From the official documentation:\n\n>MCP is an open protocol that standardizes how applications provide context to LLMs. Think of MCP like a USB-C port for AI applications. Just as USB-C provides a standardized way to connect your devices to various peripherals and accessories, MCP provides a standardized way to connect AI models to different data sources and tools.\n\n**MCP Configuration Parameters:**\n\n| Parameter | Type | Required | Description |\n|-----------|------|----------|-------------|\n| `mcp_servers` | dict | **Yes** (for MCP) | Container for MCP server definitions |\n| └─ `type` | string | Yes | Transport: `\"stdio\"` or `\"streamable-http\"` |\n| └─ `command` | string | stdio only | Command to run the MCP server |\n| └─ `args` | list | stdio only | Arguments for the command |\n| └─ `url` | string | http only | Server endpoint URL |\n| └─ `env` | dict | No | Environment variables to pass |\n| `allowed_tools` | list | No | Whitelist specific tools (if omitted, all tools available) |\n| `exclude_tools` | list | No | Blacklist dangerous\u002Funwanted tools |\n\n\n**Quick Start Commands ([Check backend MCP support here](#tools)):**\n\n```bash\n# Weather service with GPT-5\nmassgen --config @examples\u002Ftools\u002Fmcp\u002Fgpt5_nano_mcp_example \\\n \"What's the weather forecast for New York this week?\"\n\n# Multi-tool MCP with Gemini - Search + Weather + Filesystem (Requires BRAVE_API_KEY in .env)\nmassgen --config @examples\u002Ftools\u002Fmcp\u002Fmultimcp_gemini \\\n \"Find the best restaurants in Paris and save the recommendations to a file\"\n```\n\n**Configuration:**\n\n```yaml\nagents:\n # Basic MCP Configuration:\n backend:\n type: \"openai\" # Your backend choice\n model: \"gpt-5-mini\" # Your model choice\n\n # Add MCP servers here\n mcp_servers:\n weather: # Server name (you choose this)\n type: \"stdio\" # Communication type\n command: \"npx\" # Command to run\n args: [\"-y\", \"@modelcontextprotocol\u002Fserver-weather\"] # MCP server package\n\n # That's it! The agent can now check weather.\n\n # Multiple MCP Tools Example:\n backend:\n type: \"gemini\"\n model: \"gemini-3.0-pro-preview\"\n mcp_servers:\n # Web search\n search:\n type: \"stdio\"\n command: \"npx\"\n args: [\"-y\", \"@modelcontextprotocol\u002Fserver-brave-search\"]\n env:\n BRAVE_API_KEY: \"${BRAVE_API_KEY}\" # Set in .env file\n\n # HTTP-based MCP server (streamable-http transport)\n custodm_api:\n type: \"streamable-http\" # For HTTP\u002FSSE servers\n url: \"http:\u002F\u002Flocalhost:8080\u002Fmcp\u002Fsse\" # Server endpoint\n\n\n # Tool configuration (MCP tools are auto-discovered)\n allowed_tools: # Optional: whitelist specific tools\n - \"mcp__weather__get_current_weather\"\n - \"mcp__test_server__mcp_echo\"\n - \"mcp__test_server__add_numbers\"\n\n exclude_tools: # Optional: blacklist specific tools\n - \"mcp__test_server__current_time\"\n```\n\n→ [View more MCP examples](massgen\u002Fconfigs\u002Ftools\u002Fmcp\u002F)\n\n→ For comprehensive MCP integration guide, see [MCP Integration](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Fmcp_integration.html)\n\n#### **4. File System Operations & Workspace Management**\n\nMassGen provides comprehensive file system support through multiple backends, enabling agents to read, write, and manipulate files in organized workspaces.\n\n\n**Filesystem Configuration Parameters:**\n\n| Parameter | Type | Required | Description |\n|-----------|------|----------|-------------|\n| `cwd` | string | **Yes** (for file ops) | Working directory for file operations (agent-specific workspace) |\n| `snapshot_storage` | string | Yes | Directory for workspace snapshots |\n| `agent_temporary_workspace` | string | Yes | Parent directory for temporary workspaces |\n\n\n**Quick Start Commands:**\n\n```bash\n# File operations with Claude Code\nmassgen --config @examples\u002Ftools\u002Ffilesystem\u002Fclaude_code_single \\\n \"Create a Python web scraper and save results to CSV\"\n\n# Multi-agent file collaboration\nmassgen --config @examples\u002Ftools\u002Ffilesystem\u002Fclaude_code_context_sharing \\\n \"Generate a comprehensive project report with charts and analysis\"\n```\n\n**Configuration:**\n\n```yaml\n# Basic Workspace Setup:\nagents:\n - id: \"file-agent\"\n backend:\n type: \"claude_code\" # Backend with file support\n cwd: \"workspace\" # Isolated workspace for file operations\n\n# Multi-Agent Workspace Isolation:\nagents:\n - id: \"agent_a\"\n backend:\n type: \"claude_code\"\n cwd: \"workspace1\" # Agent-specific workspace\n\n - id: \"agent_b\"\n backend:\n type: \"gemini\"\n cwd: \"workspace2\" # Separate workspace\n\norchestrator:\n snapshot_storage: \"snapshots\" # Shared snapshots directory\n agent_temporary_workspace: \"temp_workspaces\" # Temporary workspace management\n```\n**Available File Operations:**\n- **Claude Code**: Built-in tools (Read, Write, Edit, MultiEdit, Bash, Grep, Glob, LS, TodoWrite)\n- **Other Backends**: Via [MCP Filesystem Server](https:\u002F\u002Fgithub.com\u002Fmodelcontextprotocol\u002Fservers\u002Fblob\u002Fmain\u002Fsrc%2Ffilesystem%2FREADME.md)\n\n**Workspace Management:**\n- **Isolated Workspaces**: Each agent's `cwd` is fully isolated and writable\n- **Snapshot Storage**: Share workspace context between Claude Code agents\n- **Temporary Workspaces**: Agents can access previous coordination results\n\n→ [View more filesystem examples](massgen\u002Fconfigs\u002Ftools\u002Ffilesystem\u002F)\n\n> ⚠️ **IMPORTANT SAFETY WARNING**\n>\n> MassGen agents can **autonomously read, write, modify, and delete files** within their permitted directories.\n>\n> **Before running MassGen with filesystem access:**\n> - Only grant access to directories you're comfortable with agents modifying\n> - Use the permission system to restrict write access where needed\n> - Consider testing in an isolated directory or virtual environment first\n> - Back up important files before granting write access\n> - Review the `context_paths` configuration carefully\n>\n> The agents will execute file operations without additional confirmation once permissions are granted.\n\n→ For comprehensive file operations guide, see [File Operations](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Ffile_operations.html)\n\n#### **5. Project Integration & User Context Paths (NEW in v0.0.21)**\n\nWork directly with your existing projects! User Context Paths allow you to share specific directories with all agents while maintaining granular permission control. This enables secure multi-agent collaboration on your real codebases, documentation, and data.\n\nMassGen automatically organizes all its working files under a `.massgen\u002F` directory in your project root, keeping your project clean and making it easy to exclude MassGen's temporary files from version control.\n\n**Project Integration Parameters:**\n\n| Parameter | Type | Required | Description |\n|-----------|------|----------|-------------|\n| `context_paths` | list | **Yes** (for project integration) | Shared directories for all agents |\n| └─ `path` | string | Yes | Absolute or relative path to your project directory (**must be directory, not file**) |\n| └─ `permission` | string | Yes | Access level: `\"read\"` or `\"write\"` (write applies only to final agent) |\n| └─ `protected_paths` | list | No | Files\u002Fdirectories immune from modification (relative to context path) |\n\n**⚠️ Important Notes:**\n- Context paths must point to **directories**, not individual files\n- Paths can be **absolute** or **relative** (resolved against current working directory)\n- **Write permissions** apply only to the **final agent** during presentation phase\n- During coordination, all context paths are **read-only** to protect your files\n- MassGen validates all paths during startup and will show clear error messages for missing paths or file paths\n\n\n**Quick Start Commands:**\n\n```bash\n# Multi-agent collaboration to improve the website in `massgen\u002Fconfigs\u002Fresources\u002Fv0.0.21-example\nmassgen --config @examples\u002Ftools\u002Ffilesystem\u002Fgpt5mini_cc_fs_context_path \"Enhance the website with: 1) A dark\u002Flight theme toggle with smooth transitions, 2) An interactive feature that helps users engage with the blog content (your choice - could be search, filtering by topic, reading time estimates, social sharing, reactions, etc.), and 3) Visual polish with CSS animations or transitions that make the site feel more modern and responsive. Use vanilla JavaScript and be creative with the implementation details.\"\n```\n\n**Configuration:**\n\n```yaml\n# Basic Project Integration:\nagents:\n - id: \"code-reviewer\"\n backend:\n type: \"claude_code\"\n cwd: \"workspace\" # Agent's isolated work area\n\norchestrator:\n context_paths:\n - path: \".\" # Current directory (relative path)\n permission: \"write\" # Final agent can create\u002Fmodify files\n protected_paths: # Optional: files immune from modification\n - \".env\"\n - \"config.json\"\n - path: \"\u002Fhome\u002Fuser\u002Fmy-project\u002Fsrc\" # Absolute path example\n permission: \"read\" # Agents can analyze your code\n\n# Advanced: Multi-Agent Project Collaboration\nagents:\n - id: \"analyzer\"\n backend:\n type: \"gemini\"\n cwd: \"analysis_workspace\"\n\n - id: \"implementer\"\n backend:\n type: \"claude_code\"\n cwd: \"implementation_workspace\"\n\norchestrator:\n context_paths:\n - path: \"..\u002Flegacy-app\u002Fsrc\" # Relative path to existing codebase\n permission: \"read\" # Read existing codebase\n - path: \"..\u002Flegacy-app\u002Ftests\"\n permission: \"write\" # Final agent can write new tests\n protected_paths: # Protect specific test files\n - \"integration_tests\u002Fproduction_data_test.py\"\n - path: \"\u002Fhome\u002Fuser\u002Fmodernized-app\" # Absolute path\n permission: \"write\" # Final agent can create modernized version\n```\n\n**This showcases project integration:**\n- **Real Project Access** - Work with your actual codebases, not copies\n- **Secure Permissions** - Granular control over what agents can read\u002Fmodify\n- **Multi-Agent Collaboration** - Multiple agents safely work on the same project\n- **Context Agents** (during coordination): Always READ-only access to protect your files\n- **Final Agent** (final execution): Gets the configured permission (READ or write)\n\n**Use Cases:**\n- **Code Review**: Agents analyze your source code and suggest improvements\n- **Documentation**: Agents read project docs to understand context and generate updates\n- **Data Processing**: Agents access shared datasets and generate analysis reports\n- **Project Migration**: Agents examine existing projects and create modernized versions\n\n**Clean Project Organization:**\n```\nyour-project\u002F\n├── .massgen\u002F # All MassGen state\n│ ├── sessions\u002F # Multi-turn conversation history (if using interactively)\n│ │ └── session_20240101_143022\u002F\n│ │ ├── turn_1\u002F # Results from turn 1\n│ │ ├── turn_2\u002F # Results from turn 2\n│ │ └── SESSION_SUMMARY.txt # Human-readable summary\n│ ├── workspaces\u002F # Agent working directories\n│ │ ├── agent1\u002F # Individual agent workspaces\n│ │ └── agent2\u002F\n│ ├── snapshots\u002F # Workspace snapshots for coordination\n│ └── temp_workspaces\u002F # Previous turn results for context\n├── massgen\u002F\n└── ...\n```\n\n**Benefits:**\n- ✅ **Clean Projects** - All MassGen files contained in one directory\n- ✅ **Easy Gitignore** - Just add `.massgen\u002F` to `.gitignore`\n- ✅ **Portable** - Move or delete `.massgen\u002F` without affecting your project\n- ✅ **Multi-Turn Sessions** - Conversation history preserved across sessions\n\n**Configuration Auto-Organization:**\n```yaml\norchestrator:\n # User specifies simple names - MassGen organizes under .massgen\u002F\n snapshot_storage: \"snapshots\" # → .massgen\u002Fsnapshots\u002F\n agent_temporary_workspace: \"temp\" # → .massgen\u002Ftemp\u002F\n\nagents:\n - backend:\n cwd: \"workspace1\" # → .massgen\u002Fworkspaces\u002Fworkspace1\u002F\n```\n\n→ For comprehensive project integration guide, see [Project Integration](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Fproject_integration.html)\n\n**Security Considerations:**\n- **Agent ID Safety**: Avoid using agent+incremental digits for IDs (e.g., `agent1`, `agent2`). This may cause ID exposure during voting\n- **File Access Control**: Restrict file access using MCP server configurations when needed\n- **Path Validation**: All context paths are validated to ensure they exist and are directories (not files)\n- **Directory-Only Context Paths**: Context paths must point to directories, not individual files\n\n---\n\n#### Additional Examples by Provider\n\n**Claude (Recursive MCP Execution - v0.0.20+)**\n```bash\n# Claude with advanced tool chaining\nmassgen --config @examples\u002Ftools\u002Fmcp\u002Fclaude_mcp_example \\\n \"Research and compare weather in Beijing and Shanghai\"\n```\n\n**OpenAI (GPT-5 Series with MCP - v0.0.17+)**\n```bash\n# GPT-5 with weather and external tools\nmassgen --config @examples\u002Ftools\u002Fmcp\u002Fgpt5_nano_mcp_example \\\n \"What's the weather of Tokyo\"\n```\n\n**Gemini (Multi-Server MCP - v0.0.15+)**\n```bash\n# Gemini with multiple MCP services\nmassgen --config @examples\u002Ftools\u002Fmcp\u002Fmultimcp_gemini \\\n \"Find accommodations in Paris with neighborhood analysis\" # (requires BRAVE_API_KEY in .env)\n```\n\n**Claude Code (Development Tools)**\n```bash\n# Professional development environment with auto-configured workspace\nuv run python -m massgen.cli \\\n --backend claude_code \\\n --model sonnet \\\n \"Create a Flask web app with authentication\"\n\n# Default workspace directories created automatically:\n# - workspace1\u002F (working directory)\n# - snapshots\u002F (workspace snapshots)\n# - temp_workspaces\u002F (temporary agent workspaces)\n```\n\n**Local Models (LM Studio - v0.0.7+)**\n```bash\n# Run open-source models locally\nmassgen --config @examples\u002Fproviders\u002Flocal\u002Flmstudio \\\n \"Explain machine learning concepts\"\n```\n\n→ [Browse by provider](massgen\u002Fconfigs\u002Fproviders\u002F) | [Browse by tools](massgen\u002Fconfigs\u002Ftools\u002F) | [Browse teams](massgen\u002Fconfigs\u002Fteams\u002F)\n\n#### Additional Use Case Examples\n\n**Question Answering & Research:**\n```bash\n# Complex research with multiple perspectives\nmassgen --config @examples\u002Fbasic\u002Fmulti\u002Fgemini_gpt5_claude \\\n \"What's best to do in Stockholm in October 2025\"\n\n# Specific research requirements\nmassgen --config @examples\u002Fbasic\u002Fmulti\u002Fgemini_gpt5_claude \\\n \"Give me all the talks on agent frameworks in Berkeley Agentic AI Summit 2025\"\n```\n\n**Creative Writing:**\n```bash\n# Story generation with multiple creative agents\nmassgen --config @examples\u002Fbasic\u002Fmulti\u002Fgemini_gpt5_claude \\\n \"Write a short story about a robot who discovers music\"\n```\n\n**Development & Coding:**\n```bash\n# Full-stack development with file operations\nmassgen --config @examples\u002Ftools\u002Ffilesystem\u002Fclaude_code_single \\\n \"Create a Flask web app with authentication\"\n```\n\n**Web Automation:** (still in test)\n```bash\n# Browser automation with screenshots and reporting\n# Prerequisites: npm install @playwright\u002Fmcp@latest (for Playwright MCP server)\nmassgen --config @examples\u002Ftools\u002Fcode-execution\u002Fmulti_agent_playwright_automation \\\n \"Browse three issues in https:\u002F\u002Fgithub.com\u002FLeezekun\u002FMassGen and suggest documentation improvements. Include screenshots and suggestions in a website.\"\n\n# Data extraction and analysis\nmassgen --config @examples\u002Ftools\u002Fcode-execution\u002Fmulti_agent_playwright_automation \\\n \"Navigate to https:\u002F\u002Fnews.ycombinator.com, extract the top 10 stories, and create a summary report\"\n```\n\n→ [**See detailed case studies**](docs\u002Fsource\u002Fexamples\u002Fcase_studies\u002FREADME.md) with real session logs and outcomes\n\n#### Interactive Mode & Advanced Usage\n\n**Multi-Turn Conversations:**\n```bash\n# Start interactive chat (no initial question)\nmassgen --config @examples\u002Fbasic\u002Fmulti\u002Fthree_agents_default\n\n# Add CWD context quickly (read-only)\nmassgen --config @examples\u002Fbasic\u002Fmulti\u002Fthree_agents_default --cwd-context ro\n\n# Add CWD context quickly (read+write)\nmassgen --config @examples\u002Fbasic\u002Fmulti\u002Fthree_agents_default --cwd-context rw\n\n# Debug mode for troubleshooting\nmassgen --config @examples\u002Fbasic\u002Fmulti\u002Fthree_agents_default \\\n --debug \"Your question\"\n```\n\n## Configuration Files\n\nMassGen configurations are organized by features and use cases. See the [Configuration Guide](massgen\u002Fconfigs\u002FREADME.md) for detailed organization and examples.\n\n**Quick navigation:**\n- **Basic setups**: [Single agent](massgen\u002Fconfigs\u002Fbasic\u002Fsingle\u002F) | [Multi-agent](massgen\u002Fconfigs\u002Fbasic\u002Fmulti\u002F)\n- **Tool integrations**: [MCP servers](massgen\u002Fconfigs\u002Ftools\u002Fmcp\u002F) | [Web search](massgen\u002Fconfigs\u002Ftools\u002Fweb-search\u002F) | [Filesystem](massgen\u002Fconfigs\u002Ftools\u002Ffilesystem\u002F)\n- **Provider examples**: [OpenAI](massgen\u002Fconfigs\u002Fproviders\u002Fopenai\u002F) | [Claude](massgen\u002Fconfigs\u002Fproviders\u002Fclaude\u002F) | [Gemini](massgen\u002Fconfigs\u002Fproviders\u002Fgemini\u002F)\n- **Specialized teams**: [Creative](massgen\u002Fconfigs\u002Fteams\u002Fcreative\u002F) | [Research](massgen\u002Fconfigs\u002Fteams\u002Fresearch\u002F) | [Development](massgen\u002Fconfigs\u002Fteams\u002Fdevelopment\u002F)\n\nSee MCP server setup guides: [Discord MCP](massgen\u002Fconfigs\u002Fdocs\u002FDISCORD_MCP_SETUP.md) | [Twitter MCP](massgen\u002Fconfigs\u002Fdocs\u002FTWITTER_MCP_ENESCINAR_SETUP.md)\n\n#### Backend Configuration Reference\n\nFor detailed configuration of all supported backends (OpenAI, Claude, Gemini, Grok, etc.), see:\n\n→ **[Backend Configuration Guide](massgen\u002Fconfigs\u002FBACKEND_CONFIGURATION.md)**\n\n#### Interactive Multi-Turn Mode\n\nMassGen supports an interactive mode where you can have ongoing conversations with the system:\n\n```bash\n# Start interactive mode with a single agent (no tool enabled by default)\nuv run python -m massgen.cli --model gpt-5-mini\n\n# Start interactive mode with configuration file\nuv run python -m massgen.cli \\\n --config massgen\u002Fconfigs\u002Fbasic\u002Fmulti\u002Fthree_agents_default.yaml\n```\n\n**Interactive Mode Features:**\n- **Multi-turn conversations**: Multiple agents collaborate to chat with you in an ongoing conversation\n- **Real-time coordination tracking**: Live visualization of agent interactions, votes, and decision-making processes\n- **Real-time feedback**: Displays real-time agent and system status with enhanced coordination visualization\n- **Multi-line input**: Use `\"\"\"` or `'''` to enter multi-line messages\n- **Slash commands**:\n - `\u002Fhelp` or `\u002Fh` - Show available commands\n - `\u002Fstatus` - Display current system status\n - `\u002Fconfig` - Open the configuration file\n - `\u002Fclear` or `\u002Freset` - Clear conversation history and start fresh\n - `\u002Fquit`, `\u002Fexit`, or `\u002Fq` - Exit the session (or press `Ctrl+C`)\n\n**Watch the recorded demo:**\n\n[![MassGen Case Study](https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Fmassgen_MassGen_readme_07732971c916.jpg)](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=h1R7fxFJ0Zc)\n\n### 5. 📊 View Results\n\nThe system provides multiple ways to view and analyze results:\n\n#### Real-time Display\n- **Live Collaboration View**: See agents working in parallel through a multi-region terminal display\n- **Status Updates**: Real-time phase transitions, voting progress, and consensus building\n- **Streaming Output**: Watch agents' reasoning and responses as they develop\n\n**Watch an example here:**\n\n[![MassGen Case Study](https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Fmassgen_MassGen_readme_a17172c9da1c.jpg)](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=Dp2oldJJImw)\n\n#### Comprehensive Logging\n\nAll sessions are automatically logged with detailed information for debugging and analysis.\n\n**Real-time Interaction:**\n- Press `r` during execution to view the coordination table in your terminal\n- Watch agents collaborate, vote, and reach consensus in real-time\n\n##### Logging Storage Structure\n\n```\n.massgen\u002F\n└── massgen_logs\u002F\n └── log_YYYYMMDD_HHMMSS\u002F # Timestamped log directory\n ├── agent_\u003Cid>\u002F # Agent-specific coordination logs\n │ └── YYYYMMDD_HHMMSS_NNNNNN\u002F # Timestamped coordination steps\n │ ├── answer.txt # Agent's answer at this step\n │ ├── context.txt # Context available to agent\n │ └── workspace\u002F # Agent workspace (if filesystem tools used)\n ├── agent_outputs\u002F # Consolidated output files\n │ ├── agent_\u003Cid>.txt # Complete output from each agent\n │ ├── final_presentation_agent_\u003Cid>.txt # Winning agent's final answer\n │ ├── final_presentation_agent_\u003Cid>_latest.txt # Symlink to latest\n │ └── system_status.txt # System status and metadata\n ├── final\u002F # Final presentation phase\n │ └── agent_\u003Cid>\u002F # Winning agent's final work\n │ ├── answer.txt # Final answer\n │ └── context.txt # Final context\n ├── coordination_events.json # Structured coordination events\n ├── coordination_table.txt # Human-readable coordination table\n ├── vote.json # Final vote tallies and consensus data\n ├── massgen.log # Complete debug log (or massgen_debug.log in debug mode)\n ├── snapshot_mappings.json # Workspace snapshot metadata\n └── execution_metadata.yaml # Query, config, and execution details\n```\n\n##### Key Log Files\n\n- **Coordination Table** (`coordination_table.txt`): Complete visualization of multi-agent coordination with event timeline, voting patterns, and consensus building\n- **Coordination Events** (`coordination_events.json`): Structured JSON log of all events (started_streaming, new_answer, vote, restart, final_answer)\n- **Vote Summary** (`vote.json`): Final vote tallies, winning agent, and consensus information\n- **Execution Metadata** (`execution_metadata.yaml`): Original query, timestamp, configuration, and execution context for reproducibility\n- **Agent Outputs** (`agent_outputs\u002F`): Complete output history and final presentations from all agents\n- **Debug Log** (`massgen.log`): Complete system operations, API calls, tool usage, and error traces (use `--debug` for verbose logging)\n\n→ For comprehensive logging guide and debugging techniques, see [Logging & Debugging](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Flogging.html)\n\n---\n\n## 🤖 Automation & LLM Integration\n\n**→ For LLM agents: See [AI_USAGE.md](AI_USAGE.md) for complete command-line usage guide**\n\nMassGen provides **automation mode** designed for LLM agents and programmatic workflows:\n\n### Quick Start - Automation Mode\n\n```bash\n# Run with minimal output and status tracking\nuv run massgen --automation --config your_config.yaml \"Your question\"\n```\n\n### Comprehensive Guide\n\n→ **Full automation guide with examples:** [Automation Guide](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Fautomation.html)\n\nTopics covered:\n- Complete automation patterns with error handling\n- Parallel experiment execution\n- Performance tips and troubleshooting\n\n### Python API & LiteLLM\n\nUse MassGen programmatically with the familiar LiteLLM\u002FOpenAI interface:\n\n```python\nfrom dotenv import load_dotenv\nload_dotenv() # Load API keys from .env\n\nimport litellm\nfrom massgen import register_with_litellm\n\nregister_with_litellm()\n\n# Multi-agent with slash format: \"backend\u002Fmodel\"\nresponse = litellm.completion(\n model=\"massgen\u002Fbuild\",\n messages=[{\"role\": \"user\", \"content\": \"Compare AI approaches\"}],\n optional_params={\"models\": [\"openai\u002Fgpt-5\", \"groq\u002Fllama-3.3-70b\"]}\n)\nprint(response.choices[0].message.content) # Final consensus answer\n```\n\nOr use the direct Python API:\n\n```python\nfrom dotenv import load_dotenv\nload_dotenv()\n\nimport asyncio\nimport massgen\n\nresult = asyncio.run(massgen.run(\n query=\"What is machine learning?\",\n models=[\"openai\u002Fgpt-5\", \"gemini\u002Fgemini-3-pro-preview\"]\n))\nprint(result[\"final_answer\"]) # Consensus answer from winning agent\n```\n\n> **Full API reference:** [Programmatic API Guide](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Fintegration\u002Fpython_api.html)\n\n---\n\n## 💡 Case Studies\n\nTo see how MassGen works in practice, check out these detailed case studies based on real session logs:\n\n**Featured:**\n- [**Multi-Turn Persistent Memory**](docs\u002Fsource\u002Fexamples\u002Fcase_studies\u002Fmulti-turn-persistent-memory.md) - Research-to-implementation workflow demonstrating memory system (v0.1.5) | [📹 Watch Demo](https:\u002F\u002Fyoutu.be\u002FwWxxFgyw40Y)\n\n**All Case Studies:**\n- [**MassGen Case Studies**](docs\u002Fsource\u002Fexamples\u002Fcase_studies\u002FREADME.md)\n- [**Case Studies Documentation**](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fexamples\u002Fcase_studies.html) - Browse case studies online\n\n---\n\n\n## 🗺️ Roadmap\n\nMassGen is currently in its foundational stage, with a focus on parallel, asynchronous multi-agent collaboration and orchestration. Our roadmap is centered on transforming this foundation into a highly robust, intelligent, and user-friendly system, while enabling frontier research and exploration.\n\n⚠️ **Early Stage Notice:** As MassGen is in active development, please expect upcoming breaking architecture changes as we continue to refine and improve the system.\n\n### Recent Achievements (v0.1.73)\n\n**🎉 Released: April 6, 2026**\n\n#### Eval Criteria Evolver & Checkpoint Objectives\n- **Eval Criteria Evolver Subagent** ([#1047](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1047)): New subagent type that evolves evaluation criteria across rounds — sharper, more opinionated criteria as the run progresses\n- **Checkpoint Objective Mode (Initial Draft)** ([#1047](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1047)): Initial draft of checkpoint MCP with `objective` mode for safety planning of irreversible actions\n- **Improved Eval Criteria Visibility**: Clearer visibility into what criteria agents are working against\n\n### Previous Achievements (v0.0.3 - v0.1.72)\n\n✅ **Grok Backend Update & Circuit Breaker Phase 2 (v0.1.72)**: Grok backend update with latest improvements. LLM API circuit breaker extended to ChatCompletions, Response API, and Gemini backends (was Claude-only).\n\n✅ **Trace Memory & Evaluation Polish (v0.1.71)**: Trace analyzer subagents launch in background after each round to write insights from execution traces into memory. Improved evaluation criteria generation and system prompt tuning.\n\n✅ **Evaluation Criteria Redesign (v0.1.70)**: Redesigned three-tier evaluation criteria with anti-pattern definitions and aspiration statements. Improved checklist-gated evaluation. Fast iteration mode, WebUI review modal, and background trace analysis.\n\n✅ **WebUI Automation & Improved Skill (v0.1.69)**: WebUI automation auto-starts without browser interaction. MassGen skill redesign for increased usability and WebUI integration. Quickstart Wizard rework and Workspace Browser expansion.\n\n✅ **Checkpoint Mode (v0.1.68)**: New checkpoint coordination mode with delegator pattern — main agent plans solo then delegates to team via `checkpoint()` tool. LLM API circuit breaker for 429 handling. WebUI checkpoint support. LiteLLM supply chain fix.\n\n✅ **Modernized WebUI (v0.1.67)**: Complete WebUI redesign with inline final answers, keyboard shortcuts, and Zustand state management. RoundBudgetGuardHook for per-round cost control. Unified parallel pre-collab phases. Regression guard.\n\n✅ **Step Mode (v0.1.66)**: New `--step` CLI mode for external orchestrators. Powers massgen-refinery plugin step mode. Codex Windows UTF-8 fixes and console text sanitization.\n\n✅ **MassGen Refinery Plugin (v0.1.65)**: Standalone MCP servers (quality, workflow, media) bring MassGen's checklist-based evaluation to Claude Code through the massgen-refinery plugin. Single-agent refinement working; multi-agent experimental.\n\n✅ **Gemini CLI Backend (v0.1.64)**: Gemini CLI as a first-class backend with session persistence, MCP tools, and Docker support. WebSocket streaming for OpenAI Response API. Execution trace analyzer subagent. Copilot Docker mode.\n\n✅ **Ensemble & Contracts (v0.1.63)**: Subagent ensemble pattern with `disable_injection` and `defer_voting_until_all_answered` as defaults. Round evaluator transformation pressure and success contracts. Lighter refinement for subagents. Killed agent handling.\n\n✅ **MassGen Skill & Viewer (v0.1.62)**: General-purpose multi-agent skill with 4 modes (general, evaluate, plan, spec) for Claude Code and other AI agents. Session viewer for real-time observation. Backend improvements for Claude Code, Codex, and Copilot. Headless and web quickstart modes.\n\n✅ **Round Evaluator Paradigm (v0.1.61)**: New round evaluator subagent type that automatically spawns evaluator subagents after each new answer to provide detailed feedback as input to the next round. Major orchestrator refactoring with improved evaluation prompts, task plan injection, and subagent fixes.\n\n✅ **Multimodal Tools, Subagent Enhancements & GPT-5.4 (v0.1.60)**: Rewritten read_media with clearer schema and MediaCallLedgerHook. Subagent enhancements with inherit_spawning_agent_backend, final_answer_strategy, per-agent subagent_agents. GPT-5.4 as default OpenAI flagship. Decomp mode cooperates with checklist workflow. Codex prompt caching fix.\n\n✅ **Quality Round Improvements (v0.1.59)**: Auto-add improvements to task plan, plan review enhancements. Better eval gen config, checklist fixes, Gemini tool name normalization for MCP. Subagent behavior adjustments, Docker skill write access fixes. Video gen skill adjustments and impact metric restoration.\n\n✅ **Comprehensive Multimodal Revamp (v0.1.58)**: ElevenLabs TTS\u002FSTT, Nano Banana 2 image generation, Grok multimedia generation, media generation skills, and multi-turn image editing. Nvidia NIM backend. Quality rethinking subagent. Smarter checklists with improve\u002Fpreserve listings. CLI mode flags and logging architecture refactor.\n\n✅ **Delegated Subagent Protocol & Builder Subagent (v0.1.57)**: File-based delegation protocol for container-to-host subagent spawning. New builder subagent type for large artifact generation with fresh context. Substantiveness tracking for smarter convergence. Claude Code reasoning parameters for updated SDK.\n\n✅ **Spec Plan Mode & Targeted Messaging (v0.1.56)**: Formal requirements specification with `plan_mode=\"spec\"` and TUI spec mode support. Targeted agent-to-agent messaging via `target_agents` parameter. Critic subagent for quality assessment. Media conversation continuity for follow-up image analysis. Codex OAuth login fix.\n\n✅ **Specialized Subagent Types & Dynamic Evaluation Criteria (v0.1.55)**: Discovery-based subagent roles (evaluator, explorer, researcher, novelty) via `SUBAGENT.md` frontmatter. GEPA-inspired task-specific evaluation criteria with core\u002Fstretch gates. Native backend image routing. Configurable video frame extraction.\n\n✅ **Subagent Messaging & Copilot SDK Backend (v0.1.54)**: Runtime messaging to steer running background subagents. New GitHub Copilot backend via copilot-sdk with native MCP support. Gemini 3.1 Pro support. Per-agent injection targeting.\n\n✅ **Background Tool Execution (v0.1.53)**: Non-blocking lifecycle tools for long-running work (start, monitor, wait, cancel, list). Planning task verification requirements. TUI background job indicators and lifecycle controls. Subagent infrastructure groundwork with Evaluator and Explorer types.\n\n✅ **Final Answer Modal & Coordination Quality Gates (v0.1.52)**: Dedicated final answer modal with tabbed answer and workspace\u002Freview interface. Substantive gate prevents low-value iteration rounds. Novelty injection combats premature convergence. Agent identity versioning for answer provenance tracking.\n\n✅ **Reviewing Coordination & Change Documents (v0.1.51)**: Review modal with multi-file diff visualization. Decision journal system for multi-agent coordination traceability. Changedoc-anchored evaluation checklists with gap reports. Drift conflict policy for safer change application. `--cwd-context` CLI flag.\n\n✅ **Chunked Plan Execution & Skill Lifecycle Management (v0.1.50)**: Chunked plan execution for safer long-form task completion with progress checkpoints. Skill lifecycle management with consolidation, organizer, and previous-session skill loading. Iterative planning review modal. Responsive TUI mode bar. Worktree improvements with branch accumulation and cross-agent diff visibility.\n\n✅ **Coordination Quality: Log Analysis TUI, Fairness Gate & Checklist Voting (v0.1.49)**: Log analysis mode built into TUI mode bar for in-app run analysis. Fairness gate prevents fast agents from dominating coordination. Checklist voting tool for structured quality evaluation. Automated testing infrastructure with CI\u002FCD and SVG snapshot baselines.\n\n\n✅ **Decomposition Mode & Worktree Isolation (v0.1.48)**: New decomposition coordination mode that decomposes tasks into subtasks assigned to individual agents with a presenter role, git worktree-based isolation for agent file writes with review modal, quickstart wizard Docker setup with animated pull progress, stop tool for agent completion signaling\n\n✅ **Codex Backend & TUI Theme Refactoring (v0.1.47)**: New Codex backend for OpenAI Codex CLI with local and Docker execution, NativeToolMixin for shared tool handling, TUI theme system refactored to palette-based architecture with dark and light variants, per-agent voting sensitivity configuration\n\n✅ **Subagent TUI Streaming & Event Architecture Refactor (v0.1.46)**: Interactive preview cards that expand to full timeline views with real-time event streaming, unified event pipeline with single source of truth for display creation, enhanced final presentation with workspace visualization and winning agent highlighting, fixed banner display and tool call ID handling\n\n✅ **TUI as Default & Config Migration (v0.1.45)**: Textual Terminal UI now launches by default with automatic `rich_terminal` to `textual_terminal` migration, setup wizard generates TUI configs, legacy Rich display accessible via `--display rich` flag\n\n✅ **Execute Mode for Independent Plan Selection (v0.1.44)**: Mode cycling through Normal → Planning → Execute via `Shift+Tab` or mode bar, plan selector browsing up to 10 recent plans with timestamps, view full plan modal with complete task breakdown, empty submission for plan execution, context path preservation between planning and execution phases, enhanced case studies with interactive setup guides, TUI performance optimizations with viewport-based rendering\n\n✅ **Tool Call Batching & Interactive Case Studies (v0.1.43)**: Consecutive MCP tool calls grouped into collapsible tree views with \"+N more\" indicators and click-to-expand. New interactive case studies page with side-by-side SVG comparisons. `PlanOptionsPopover` for browsing plans and selecting depth. Quoted path support for paths with spaces. Final presentation display and TUI polish fixes.\n\n✅ **TUI Visual Redesign & Human Input Queue (v0.1.42)**: Modern \"Conversational AI\" aesthetic with rounded corners, redesigned agent tabs with dot indicators, adaptive tool cards, polished modals. New `HumanInputHook` for injecting messages to agents mid-stream with thread-safe per-agent tracking. AG2 single-agent coordination fix.\n\n✅ **Async Subagent Execution (v0.1.41)**: Background subagent execution with `async_=True` for non-blocking parallel work, poll for completion and retrieve results, per-round timeout control with `subagent_round_timeouts` config, extended subagent parameters for timeout and concurrency control\n\n✅ **Textual TUI Interactive Mode (v0.1.40)**: Interactive terminal UI with `--display textual` for real-time agent streaming, comprehensive modals for costs\u002Fvotes\u002Fworkspace\u002Fanswers, context path injection with `@path\u002Fto\u002Ffile` syntax, human feedback integration via prompt modals\n\n✅ **Plan and Execute Workflow (v0.1.39)**: Complete plan-then-execute workflow with `--plan-and-execute` for autonomous planning and execution, `--execute-plan` to run existing plans without re-planning, task verification workflow with `verified` status and verification groups for batch validation, plan storage system in `.massgen\u002Fplans\u002F` with frozen snapshots and execution tracking, Response API function call message sanitization fixes\n\n✅ **Task Planning & Two-Tier Workspaces (v0.1.38)**: Task planning mode with `--plan` flag for structured work breakdown (plan-only, no auto-execution), git-backed two-tier workspaces separating scratch exploration from final deliverables, automatic CLAUDE.md\u002FAGENTS.md discovery for project context, batch image analysis with multi-image comparison, circuit breaker for timeout denial loops, Docker health monitoring\n\n✅ **Execution Traces & Thinking Mode (v0.1.37)**: Full execution history preserved as `execution_trace.md` for compression recovery and cross-agent coordination, Claude Code and Gemini reasoning content streaming buffer integration, standardized agent labeling across all backends\n\n✅ **@path Context Handling & Hook Framework (v0.1.36)**: Inline file picker with `@path` syntax and autocomplete, PreToolUse\u002FPostToolUse hooks for permission validation and content injection, global and per-agent hook registration, built-in `MidStreamInjectionHook` and `HighPriorityTaskReminderHook`, Claude Code hooks compatibility, improved Docker resource management\n\n✅ **Log Analysis CLI & Logfire Observability (v0.1.35)**: `massgen logs analyze` command with prompt mode and multi-agent self-analysis, Logfire workflow attributes for round context and vote reasoning, `direct_mcp_servers` config for keeping specific MCPs as protocol tools, improved tool handling for unknown tools and vote-only mode fixes\n\n✅ **OpenAI-Compatible Server & Model Discovery (v0.1.34)**: Local HTTP server with `massgen serve` compatible with any OpenAI SDK client, dynamic model discovery for Groq and Together backends via authenticated API calls, WebUI file diffs and answer refresh polling, subagent status tracking and cancellation recovery improvements\n\n✅ **Reactive Context Compression & Streaming Buffers (v0.1.33)**: Automatic conversation compression when context length errors occur, streaming buffer system tracking partial responses for recovery, file overwrite protection in `write_file` tool, task plan duplicate prevention, Grok MCP tools visibility fix, Gemini vote-only mode fix, GPT-5 model behavior improvements\n\n✅ **Multi-Turn Session Export & Per-Attempt Logging (v0.1.32)**: Turn range selection for session export (`--turns`), workspace export controls (`--no-workspace`, `--workspace-limit`), Logfire moved to optional `[observability]` extra, per-attempt isolated log files with handler reconfiguration, automatic DOCX\u002FPPTX\u002FXLSX to PDF conversion for session sharing\n\n✅ **Logfire Observability & Azure Tool Streaming (v0.1.31)**: Optional Logfire integration with automatic LLM instrumentation for OpenAI, Claude, and Gemini backends, Azure OpenAI tool calls yielded as structured chunks, `--logfire` CLI flag and `MASSGEN_LOGFIRE_ENABLED` environment variable\n\n✅ **OpenRouter Web Search & Persona Diversity (v0.1.30)**: Native web search via OpenRouter plugins with `enable_web_search`, persona diversity modes (`perspective`\u002F`implementation`) with phase-based adaptation, Azure multi-endpoint auto-detection, environment variable expansion with `${VAR}` syntax\n\n✅ **Subagent System & Tool Metrics (v0.1.29)**: Spawn parallel child MassGen processes with isolated workspaces and automatic result aggregation, enhanced tool metrics with per-call averages and min\u002Fmax\u002Fmedian distribution, CLI per-agent system messages via `massgen --quickstart`\n\n✅ **Unified Multimodal Tools & Artifact Previews (v0.1.28)**: Consolidated `read_media` tool for image\u002Faudio\u002Fvideo analysis, unified `generate_media` tool for media creation (images, videos, audio), Web UI artifact previewer for PDFs\u002FDOCX\u002FPPTX\u002Fimages\u002FHTML\u002FSVG\u002FMarkdown\u002FMermaid, OpenRouter tool-capable model filtering, Azure OpenAI fixes\n\n✅ **Session Sharing & Log Analysis (v0.1.27)**: Session sharing via GitHub Gist with `massgen export`, log analysis CLI with `massgen logs` command, per-LLM call timing metrics, Gemini 3 Flash model support, enhanced CLI config builder with per-agent web search and system messages\n\n✅ **Web UI Setup & Shadow Agent Depth (v0.1.26)**: Docker diagnostics module, Web UI setup wizard with guided first-run experience, shadow agent response depth for test-time compute scaling, GPT-5.1-Codex family models\n\n✅ **UI-TARS & Evolving Skills (v0.1.25)**: ByteDance's UI-TARS-1.5-7B for GUI automation, GPT-5.2 model support, evolving skill creator system with session persistence, enhanced Textual terminal with adaptive layouts\n\n✅ **Multi-Backend Cost Tracking (v0.1.24)**: Real-time token counting for OpenRouter, xAI\u002FGrok, Gemini, and Claude Code backends with `\u002Finspect c` cost breakdown showing per-agent token usage, aggregated session cost totals with improved display formatting\n\n✅ **Turn History Inspection & Web UI Automation (v0.1.23)**: Interactive `\u002Finspect` commands for reviewing turn details with menu navigation, `AutomationView` component for programmatic monitoring, `SessionMountManager` for Docker container persistence across turns, flag-based cancellation with terminal restoration, `run_async_safely()` for nested event loop handling\n\n✅ **Shadow Agent Architecture (v0.1.22)**: Lightweight shadow agents respond to broadcasts in parallel without interrupting parent work, inheriting full conversation history and current turn context via `asyncio.gather()` parallelization\n\n✅ **Graceful Cancellation & Session Resumption (v0.1.21)**: Ctrl+C saves partial progress during coordination, cancelled sessions resume with `--continue` preserving agent answers and workspaces\n\n✅ **Web UI & Auto Docker Setup (v0.1.20)**: Browser-based real-time visualization with React frontend, WebSocket streaming, timeline views, and workspace browsing. Automatic Docker container setup for computer use agents with pre-configured X11 virtual display, xdotool, Firefox, Chromium, and scrot\n\n✅ **LiteLLM Integration & Claude Strict Tool Use (v0.1.19)**: MassGen as LiteLLM custom provider with `run()` and `build_config()` programmatic API, Claude strict tool use with structured outputs, Gemini exponential backoff for rate limit resilience\n\n✅ **Agent Communication System (v0.1.18)**: Human broadcast Q&A via `ask_others()` tool with three modes, blocking execution with inline response delivery, session-persistent Q&A history\n\n✅ **Claude Advanced Tooling (v0.1.18)**: Programmatic tool calling via `enable_programmatic_flow` flag, server-side tool discovery via `enable_tool_search` with regex or bm25 variants\n\n✅ **Textual Terminal Display (v0.1.17)**: Interactive terminal UI using the Textual library with dark\u002Flight themes, multi-panel layout for agents and orchestrator, real-time streaming with syntax highlighting, content filtering for critical patterns\n\n✅ **Terminal Evaluation & Cost Tracking (v0.1.16)**: Automated VHS recording with AI-powered terminal display evaluation, LiteLLM integration for accurate pricing across 500+ models with reasoning\u002Fcached tokens support, memory archiving for multi-turn session persistence, four self-evolution skills for MassGen development\n\n✅ **Persona Generation & Docker Distribution (v0.1.15)**: Automatic persona generation for agent diversity with multiple strategies (complementary, diverse, specialized, adversarial), GitHub Container Registry integration with ARM support, custom tools in isolated Docker containers for security, MassGen pre-installed in Docker images\n\n✅ **Parallel Tool Execution & Gemini 3 Pro (v0.1.14)**: Configurable concurrent tool execution across all backends with asyncio-based scheduling, Gemini 3 Pro integration with function calling, interactive quickstart workflow, MCP registry client for server metadata\n\n✅ **Code-Based Tools & MCP Registry (v0.1.13)**: CodeAct paradigm implementation with tool integration via importable Python code reducing token usage by 98%, MCP server registry with auto-discovery and on-demand loading, TOOL.md documentation standard\n\n✅ **NLIP Integration & Skills System (v0.1.13)**: Advanced tool routing with Natural Language Interface Protocol across Claude, Gemini, and OpenAI backends, cross-platform automated skills installer for openskills CLI, Anthropic skills, and Crawl4AI\n\n✅ **System Prompt Architecture Refactoring (v0.1.12)**: Hierarchical system prompt structure with XML-based formatting for Claude, improved LLM attention management\n\n✅ **Semtools & Serena Skills (v0.1.12)**: Semantic search via embedding-based similarity, symbol-level code understanding via LSP integration, local execution mode for non-Docker environments\n\n✅ **Multi-Agent Computer Use (v0.1.12)**: Enhanced Gemini computer use with Docker integration, VNC visualization, multi-agent coordination combining Claude (Docker\u002FLinux) and Gemini (Browser)\n\n✅ **Skills System (v0.1.11)**: Modular prompting framework with SkillsManager for dynamic skill loading, automatic discovery with always\u002Foptional categories, file search skill, Docker-compatible mounting\n\n✅ **Memory MCP Tool & Filesystem Integration (v0.1.11)**: MCP server for memory management with markdown-based storage, short-term\u002Flong-term memory tiers, automatic workspace persistence, orchestrator integration for cross-agent memory sharing, enhanced Windows support for long system prompts\n\n✅ **Rate Limiting System (v0.1.11)**: Multi-dimensional limiting (RPM, TPM, RPD) for Gemini models with configurable thresholds, YAML-based configuration, CLI integration with --enable-rate-limiting flag, asyncio lock fix for event loop reuse\n\n✅ **Framework Interoperability Streaming (v0.1.10)**: Real-time intermediate step streaming for LangGraph and SmoLAgent with log\u002Foutput distinction, enhanced debugging for external framework reasoning steps\n\n✅ **Docker Configuration Enhancements (v0.1.10)**: Nested authentication with separate mount and environment variable arrays, custom image support via Dockerfile.custom-example, automatic package installation\n\n✅ **Universal Workspace Isolation (v0.1.10)**: Instance ID generation extended to all execution modes ensuring safe parallel execution, enhanced workspace path uniqueness across concurrent sessions\n\n✅ **Session Management System (v0.1.9)**: Complete session state tracking and restoration with SessionState dataclass and SessionRegistry for multi-turn persistence across CLI invocations, workspace continuity preserving agent states and coordination history between turns\n\n✅ **Computer Use Tools (v0.1.9)**: Native Claude and Gemini computer use API integration for browser and desktop automation with screenshot analysis and action generation, lightweight browser automation for specific tasks without full computer use overhead\n\n✅ **Fuzzy Model Matching (v0.1.9)**: Intelligent model name search with approximate inputs (e.g., \"sonnet\" → \"claude-sonnet-4-5-20250929\"), model catalog system with curated lists across providers, enhanced config builder with automatic model search\n\n✅ **Backend Capabilities Expansion (v0.1.9)**: Comprehensive backend registry with detailed specifications for all providers, audio\u002Fvideo support, hardware acceleration, unified access across diverse model families, enhanced memory update logic focusing on actionable patterns\n\n✅ **Automation Mode for LLM Agents (v0.1.8)**: Complete infrastructure for running MassGen inside LLM agents with SilentDisplay class for minimal output (~10 lines vs 250-3,000+), real-time status.json monitoring updated every 2 seconds, meaningful exit codes (0=success, 1=config error, 2=execution error, 3=timeout, 4=interrupted), automatic workspace isolation for parallel execution, meta-coordination capabilities allowing MassGen to run MassGen\n\n✅ **DSPy Question Paraphrasing Integration (v0.1.8)**: Intelligent question diversity for multi-agent coordination with semantic-preserving paraphrasing module supporting three strategies (diverse\u002Fbalanced\u002Fconservative), automatic semantic validation to ensure meaning preservation, thread-safe caching system with SHA-256 hashing, support for all backends as paraphrasing engines, orchestrator integration for automatic question variant distribution\n\n✅ **Agent Task Planning System (v0.1.7)**: MCP-based planning server with task lifecycle management, dependency tracking with automatic validation and blocking, status transitions between pending\u002Fin_progress\u002Fcompleted\u002Fblocked states, orchestrator integration for plan-aware multi-agent coordination\n\n✅ **Background Shell Execution (v0.1.7)**: Persistent shell sessions for long-running commands with BackgroundShell class supporting async execution, real-time output streaming and monitoring, automatic timeout handling, enhanced code execution server with background capabilities\n\n✅ **Preemption Coordination (v0.1.7)**: Agents can interrupt ongoing coordination to submit better answers without full restart, partial progress preservation during preemption, enhanced coordination tracker logging preemption events\n\n✅ **Framework Interoperability (v0.1.6)**: AG2 nested chat, LangGraph workflows, AgentScope agents, OpenAI Assistants, and SmoLAgent integrated as custom tools with cross-framework collaboration and streaming support for AG2\n\n✅ **Configuration Validator (v0.1.6)**: Comprehensive YAML validation with ConfigValidator class, pre-commit integration, and detailed error messages with actionable suggestions\n\n✅ **Unified Tool Execution (v0.1.6)**: ToolExecutionConfig dataclass standardizing tool handling across ResponseBackend, ChatCompletionsBackend, and ClaudeBackend with consistent error reporting\n\n✅ **Gemini Backend Simplification (v0.1.6)**: Removed gemini_mcp_manager and gemini_trackers modules, consolidated code reducing codebase by 1,598 lines\n\n✅ **Memory System (v0.1.5)**: Long-term semantic memory via mem0 integration with fact extraction and retrieval across sessions, short-term conversational memory for active context, automatic context compression when approaching token limits, cross-agent memory sharing with turn-aware filtering, session management for memory isolation and continuation, Qdrant vector database integration for semantic search\n\n✅ **Multimodal Generation Tools (v0.1.4)**: Create images from text via DALL-E API, generate videos from descriptions, text-to-speech with audio transcription support, document generation for PDF\u002FDOCX\u002FXLSX\u002FPPTX formats, image transformation capabilities for existing images\n\n✅ **Binary File Protection (v0.1.4)**: Automatic blocking prevents text tools from accessing 40+ binary file types including images, videos, audio, archives, and Office documents, intelligent error messages guide users to appropriate specialized tools for binary content\n\n✅ **Crawl4AI Integration (v0.1.4)**: Intelligent web scraping with LLM-powered content extraction and customizable extraction patterns for structured data retrieval from websites\n\n✅ **Post-Evaluation Workflow (v0.1.3)**: Winning agents evaluate their own answers before submission with submit and restart capabilities, supports answer confirmation and orchestration restart with feedback across all backends\n\n✅ **Multimodal Understanding Tools (v0.1.3)**: Analyze images, transcribe audio, extract video frames, and process documents (PDF\u002FDOCX\u002FXLSX\u002FPPTX) with structured JSON output, works across all backends via OpenAI GPT-4.1 integration\n\n✅ **Docker Sudo Mode (v0.1.3)**: Privileged command execution in Docker containers for system-level operations requiring elevated permissions\n\n✅ **Intelligent Planning Mode (v0.1.2)**: Automatic question analysis determining operation irreversibility via `_analyze_question_irreversibility()` in orchestrator, selective tool blocking with `set_planning_mode_blocked_tools()` and `is_mcp_tool_blocked()` methods, read-only MCP operations during coordination with write operations blocked, zero-configuration transparent operation, multi-workspace support\n\n✅ **Model Updates (v0.1.2)**: Claude 4.5 Haiku model `claude-haiku-4-5-20251001`, reorganized Claude model priorities with `claude-sonnet-4-5-20250929` default, Grok web search fix with `_add_grok_search_params()` method for proper `extra_body` parameter handling\n\n✅ **Custom Tools System (v0.1.1)**: User-defined Python function registration using `ToolManager` class in `massgen\u002Ftool\u002F_manager.py`, cross-backend support alongside MCP servers, builtin\u002FMCP\u002Fcustom tool categories with automatic discovery, 40+ examples in `massgen\u002Fconfigs\u002Ftools\u002Fcustom_tools\u002F`, voting sensitivity controls with three-tier quality system (lenient\u002Fbalanced\u002Fstrict), answer novelty detection preventing duplicates\n\n✅ **Backend Enhancements (v0.1.1)**: Gemini architecture refactoring with extracted MCP management (`gemini_mcp_manager.py`), tracking (`gemini_trackers.py`), and utilities, new capabilities registry in `massgen\u002Fbackend\u002Fcapabilities.py` documenting feature support across all backends\n\n✅ **PyPI Package Release (v0.1.0)**: Official distribution via `pip install massgen` with simplified installation, global `massgen` command accessible from any directory, comprehensive Sphinx documentation at [docs.massgen.ai](https:\u002F\u002Fdocs.massgen.ai\u002F), interactive setup wizard with use case presets and API key management, enhanced CLI with `@examples\u002F` prefix for built-in configurations\n\n✅ **Docker Execution Mode (v0.0.32)**: Container-based isolation with secure command execution in isolated Docker containers preventing host filesystem access, persistent state management with packages and dependencies persisting across conversation turns, multi-agent support with dedicated isolated containers for each agent, configurable security with resource limits (CPU, memory), network isolation modes, and read-only volume mounts\n\n✅ **MCP Architecture Refactoring (v0.0.32)**: Simplified client with renamed `MultiMCPClient` to `MCPClient` reflecting streamlined architecture, code consolidation by removing deprecated modules and consolidating duplicate MCP protocol handling, improved maintainability with standardized type hints, enhanced error handling, and cleaner code organization\n\n✅ **Claude Code Docker Integration (v0.0.32)**: Automatic tool management with Bash tool automatically disabled in Docker mode routing commands through execute_command, MCP auto-permissions with automatic approval for MCP tools while preserving security validation, enhanced guidance with system messages preventing git repository confusion between host and container environments\n\n✅ **Universal Command Execution (v0.0.31)**: MCP-based execute_command tool works across Claude, Gemini, OpenAI, and Chat Completions providers, comprehensive security with permission management and command filtering, code execution in planning mode for safer coordination\n\n✅ **External Framework Integration (v0.0.31)**: Multi-agent conversations using external framework group chat patterns, smart speaker selection (automatic, round-robin, manual) powered by LLMs, enhanced adapter supporting native group chat coordination\n\n✅ **Audio & Video Generation (v0.0.31)**: Audio tools for text-to-speech and transcription, video generation using OpenAI's Sora-2 API, multimodal expansion beyond text and images\n\n✅ **Multimodal Support Extension (v0.0.30)**: Audio and video processing for Chat Completions and Claude backends (WAV, MP3, MP4, AVI, MOV, WEBM formats), flexible media input via local paths or URLs, extended base64 encoding for audio\u002Fvideo files, configurable file size limits\n\n✅ **Claude Agent SDK Migration (v0.0.30)**: Package migration from `claude-code-sdk` to `claude-agent-sdk>=0.0.22`, improved bash tool permission validation, enhanced system message handling\n\n✅ **Qwen API Integration (v0.0.30)**: Added Qwen API provider to Chat Completions ecosystem with `QWEN_API_KEY` support, video understanding configuration examples\n\n✅ **MCP Planning Mode (v0.0.29)**: Strategic planning coordination strategy for safer MCP tool usage, multi-backend support (Response API, Chat Completions, Gemini), agents plan without execution during coordination, 5 planning mode configurations\n\n✅ **File Operation Safety (v0.0.29)**: Read-before-delete enforcement with `FileOperationTracker` class, `PathPermissionManager` integration with operation tracking methods, enhanced file operation safety mechanisms\n\n✅ **External Framework Integration (v0.0.28)**: Adapter system for external agent frameworks with async execution, code execution in multiple environments (Local, Docker, Jupyter, YepCode), ready-to-use configurations for framework integration\n\n✅ **Multimodal Support - Image Processing (v0.0.27)**: New `stream_chunk` module for multimodal content, image generation and understanding capabilities, file upload and search for document Q&A, Claude Sonnet 4.5 support, enhanced workspace multimodal tools\n\n✅ **File Deletion and Workspace Management (v0.0.26)**: New MCP tools (`delete_file`, `delete_files_batch`, `compare_directories`, `compare_files`) for workspace cleanup and file comparison, consolidated `_workspace_tools_server.py`, enhanced path permission manager\n\n✅ **Protected Paths and File-Based Context Paths (v0.0.26)**: Protect specific files within write-permitted directories, grant access to individual files instead of entire directories\n\n✅ **Multi-Turn Filesystem Support (v0.0.25)**: Multi-turn conversation support with persistent context across turns, automatic `.massgen` directory structure, workspace snapshots and restoration, enhanced path permission system with smart exclusions, and comprehensive backend improvements\n\n✅ **SGLang Backend Integration (v0.0.25)**: Unified vLLM\u002FSGLang backend with auto-detection, support for SGLang-specific parameters like `separate_reasoning`, and dual server support for mixed vLLM and SGLang deployments\n\n✅ **vLLM Backend Support (v0.0.24)**: Complete integration with vLLM for high-performance local model serving, POE provider support, GPT-5-Codex model recognition, backend utility modules refactoring, and comprehensive bug fixes including streaming chunk processing\n\n✅ **Backend Architecture Refactoring (v0.0.23)**: Major code consolidation with new `base_with_mcp.py` class reducing ~1,932 lines across backends, extracted formatter module for better code organization, and improved maintainability through unified MCP integration\n\n✅ **Workspace Copy Tools via MCP (v0.0.22)**: Seamless file copying capabilities between workspaces, configuration organization with hierarchical structure, and enhanced file operations for large-scale collaboration\n\n✅ **Grok MCP Integration (v0.0.21)**: Unified backend architecture with full MCP server support, filesystem capabilities through MCP servers, and enhanced configuration files\n\n✅ **Claude Backend MCP Support (v0.0.20)**: Extended MCP integration to Claude backend, full MCP protocol and filesystem support, robust error handling, and comprehensive documentation\n\n✅ **Comprehensive Coordination Tracking (v0.0.19)**: Complete coordination tracking and visualization system with event-based tracking, interactive coordination table display, and advanced debugging capabilities for multi-agent collaboration patterns\n\n✅ **Comprehensive MCP Integration (v0.0.18)**: Extended MCP to all Chat Completions backends (Cerebras AI, Together AI, Fireworks AI, Groq, Nebius AI Studio, OpenRouter), cross-provider function calling compatibility, 9 new MCP configuration examples\n\n✅ **OpenAI MCP Integration (v0.0.17)**: Extended MCP (Model Context Protocol) support to OpenAI backend with full tool discovery and execution capabilities for GPT models, unified MCP architecture across multiple backends, and enhanced debugging\n\n✅ **Unified Filesystem Support with MCP Integration (v0.0.16)**: Complete `FilesystemManager` class providing unified filesystem access for Gemini and Claude Code backends, with MCP-based operations for file manipulation and cross-agent collaboration\n\n✅ **MCP Integration Framework (v0.0.15)**: Complete MCP implementation for Gemini backend with multi-server support, circuit breaker patterns, and comprehensive security framework\n\n✅ **Enhanced Logging (v0.0.14)**: Improved logging system for better agents' answer debugging, new final answer directory structure, and detailed architecture documentation\n\n✅ **Unified Logging System (v0.0.13)**: Centralized logging infrastructure with debug mode and enhanced terminal display formatting\n\n✅ **Windows Platform Support (v0.0.13)**: Windows platform compatibility with improved path handling and process management\n\n✅ **Enhanced Claude Code Agent Context Sharing (v0.0.12)**: Claude Code agents now share workspace context by maintaining snapshots and temporary workspace in orchestrator's side\n\n✅ **Documentation Improvement (v0.0.12)**: Updated README with current features and improved setup instructions\n\n✅ **Custom System Messages (v0.0.11)**: Enhanced system message configuration and preservation with backend-specific system prompt customization\n\n✅ **Claude Code Backend Enhancements (v0.0.11)**: Improved integration with better system message handling, JSON response parsing, and coordination action descriptions\n\n✅ **Azure OpenAI Support (v0.0.10)**: Integration with Azure OpenAI services including GPT-4.1 and GPT-5-chat models with async streaming\n\n✅ **MCP (Model Context Protocol) Support (v0.0.9)**: Integration with MCP for advanced tool capabilities in Claude Code Agent, including Discord and Twitter integration\n\n✅ **Timeout Management System (v0.0.8)**: Orchestrator-level timeout with graceful fallback and enhanced error messages\n\n✅ **Local Model Support (v0.0.7)**: Complete LM Studio integration for running open-weight models locally with automatic server management\n\n✅ **GPT-5 Series Integration (v0.0.6)**: Support for OpenAI's GPT-5, GPT-5-mini, GPT-5-nano with advanced reasoning parameters\n\n✅ **Claude Code Integration (v0.0.5)**: Native Claude Code backend with streaming capabilities and tool support\n\n✅ **GLM-4.5 Model Support (v0.0.4)**: Integration with ZhipuAI's GLM-4.5 model family\n\n✅ **Foundation Architecture (v0.0.3)**: Complete multi-agent orchestration system with async streaming, builtin tools, and multi-backend support\n\n✅ **Extended Provider Ecosystem**: Support for 15+ providers including Cerebras AI, Together AI, Fireworks AI, Groq, Nebius AI Studio, and OpenRouter\n\n### Key Future Enhancements\n\n- **Bug Fixes & Backend Improvements:** Fixing image generation path issues and adding Claude multimodal support\n- **Advanced Agent Collaboration:** Exploring improved communication patterns and consensus-building protocols to improve agent synergy\n- **Expanded Model Integration:** Adding support for more frontier models and local inference engines\n- **Improved Performance & Scalability:** Optimizing the streaming and logging mechanisms for better performance and resource management\n- **Enhanced Developer Experience:** Completing tool registration system and web interface for better visualization\n\nWe welcome community contributions to achieve these goals.\n\n### v0.1.74 Roadmap\n\nVersion 0.1.74 focuses on cloud execution:\n\n#### Planned Features\n- **Cloud Modal MVP** ([#982](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fissues\u002F982)): Run MassGen as a cloud job on Modal — progress streams to terminal, results saved locally under `.massgen\u002Fcloud_jobs\u002F`\n\n---\n\n## 🤝 Contributing\n\nWe welcome contributions! Please see our [Contributing Guidelines](CONTRIBUTING.md) for details.\n\n---\n\n## 🤝 Acknowledge\n\nWe thank AgentWeb\n\n\u003Ca href=\"https:\u002F\u002Fwww.agentweb.pro\u002F\">\n \u003Cimg width=\"196\" height=\"51\" alt=\"68dacef628cd7a44dfb97814_agentweb-logo\" src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Fmassgen_MassGen_readme_4a2ebd177b17.png\" \u002F>\n\u003C\u002Fa>\n\nfor their kind sponsorship.\n\n---\n\n## 📄 License\n\nThis project is licensed under the Apache License 2.0 - see the [LICENSE](LICENSE) file for details.\n\n---\n\n\u003Cdiv align=\"center\">\n\n**⭐ Star this repo if you find it useful! ⭐**\n\nMade with ❤️ by the MassGen team\n\n\u003C\u002Fdiv>\n\n## ⭐ Star History\n\n[![Star History Chart](https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Fmassgen_MassGen_readme_fb6c57c3716f.png)](https:\u002F\u002Fwww.star-history.com\u002F#Leezekun\u002FMassGen&Date)\n","\u003Cp align=\"center\">\n \u003Cpicture>\n \u003Csource media=\"(prefers-color-scheme: dark)\" srcset=\"assets\u002Flogo-dark.png\">\n \u003Csource media=\"(prefers-color-scheme: light)\" srcset=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Fmassgen_MassGen_readme_ade0b41da867.png\">\n \u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Fmassgen_MassGen_readme_ade0b41da867.png\" alt=\"MassGen Logo\" width=\"360\" \u002F>\n \u003C\u002Fpicture>\n\u003C\u002Fp>\n\n\u003Cdiv align=\"center\">\n\n[![PyPI](https:\u002F\u002Fimg.shields.io\u002Fpypi\u002Fv\u002Fmassgen?style=flat-square&logo=pypi&logoColor=white&label=PyPI&color=3775A9)](https:\u002F\u002Fpypi.org\u002Fproject\u002Fmassgen\u002F)\n[![Docs](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fdocs-massgen.ai-blue?style=flat-square&logo=readthedocs&logoColor=white)](https:\u002F\u002Fdocs.massgen.ai)\n[![GitHub Stars](https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002FLeezekun\u002FMassGen?style=flat-square&logo=github&color=181717&logoColor=white)](https:\u002F\u002Fgithub.com\u002FLeezekun\u002FMassGen)\n[![Python 3.11+](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fpython-3.11+-3776AB?style=flat-square&logo=python&logoColor=white)](https:\u002F\u002Fwww.python.org\u002Fdownloads\u002F)\n[![License](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Flicense-Apache%202.0-green?style=flat-square)](LICENSE)\n\n\u003C\u002Fdiv>\n\n\u003Cdiv align=\"center\">\n\n[![Follow on X](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FFOLLOW%20ON%20X-000000?style=for-the-badge&logo=x&logoColor=white)](https:\u002F\u002Fx.massgen.ai)\n[![Follow on LinkedIn](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FFOLLOW%20ON%20LINKEDIN-0A66C2?style=for-the-badge&logo=linkedin&logoColor=white)](https:\u002F\u002Fwww.linkedin.com\u002Fcompany\u002Fmassgen-ai)\n[![Join our Discord](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FJOIN%20OUR%20DISCORD-5865F2?style=for-the-badge&logo=discord&logoColor=white)](https:\u002F\u002Fdiscord.massgen.ai)\n\n\u003C\u002Fdiv>\n\n\u003Ch1 align=\"center\">🚀 MassGen: 多智能体生成式AI扩展系统\u003C\u002Fh1>\n\n\u003Cp align=\"center\">\n \u003Ci>MassGen是一个前沿的多智能体系统，利用协作式AI的力量来解决复杂任务。\u003C\u002Fi>\n\u003C\u002Fp>\n\n\u003Cp align=\"center\">\n \u003Ca href=\"https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=5JofXWf_Ok8\">\n \u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Fmassgen_MassGen_readme_0433a8f0466f.gif\" alt=\"MassGen示例\" width=\"800\">\n \u003C\u002Fa>\n\u003C\u002Fp>\n\n\u003Cp align=\"center\">\n \u003Ci>通过协作、持续改进的智能体实现AI规模化（速度提升4倍）\u003C\u002Fi>\n\u003C\u002Fp>\n\nMassGen是一个尖端的多智能体框架，它通过冗余和迭代优化来协调AI智能体解决复杂任务。每个智能体都会处理整个问题，在优化和重启的循环中相互观察、批评并在此基础上继续工作。当智能体认为答案足够可靠时，它们会进行投票，最终由集体验证的最佳答案胜出。这种并行优化与集体验证的方法为原则性的多智能体扩展奠定了基础，系统能够通过利用不同智能体的视角，并以共识来确保质量，从而不断改进其输出。\n\n该项目始于《推理的神话》（[The Myth of Reasoning](https:\u002F\u002Fdocs.ag2.ai\u002Flatest\u002Fdocs\u002Fblog\u002F2025\u002F04\u002F16\u002FReasoning\u002F)）中提出的“思维线程”和“迭代优化”理念，并扩展了AG2（[AG2](https:\u002F\u002Fgithub.com\u002Fag2ai\u002Fag2)）中的经典“多智能体对话”思想。这里有一段在2025年伯克利代理型AI峰会上介绍背景情况的[视频记录](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=xM2Uguw1UsQ)。\n\n\u003Cp align=\"center\">\n \u003Cb>🧩 将MassGen用作技能：\u003C\u002Fb> \u003Ccode>npx skills add massgen\u002Fskills --all\u003C\u002Fcode> — 然后在Claude Code、Cursor、Copilot或其他40多个智能体中调用该技能。 \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fmassgen\u002Fskills\">了解更多 →\u003C\u002Fa>\n\u003C\u002Fp>\n\n\u003Cp align=\"center\">\n \u003Cb>📚 致贡献者：\u003C\u002Fb> 请参阅\u003Ca href=\"https:\u002F\u002Fmassgen.github.io\u002FHandbook\u002F\">MassGen贡献者手册\u003C\u002Fa> - 面向开发和研究团队的集中化政策与资源\n\u003C\u002Fp>\n\n---\n\n## 📋 目录\n\n\u003Cdetails open>\n\u003Csummary>\u003Ch3>✨ 核心特性\u003C\u002Fh3>\u003C\u002Fsummary>\n\n- [跨模型\u002F智能体协同](#-key-features-1)\n- [并行处理](#-key-features-1)\n- [智能共享](#-key-features-1)\n- [共识构建](#-key-features-1)\n- [实时可视化](#-key-features-1)\n\u003C\u002Fdetails>\n\n\u003Cdetails open>\n\u003Csummary>\u003Ch3>🆕 最新特性\u003C\u002Fh3>\u003C\u002Fsummary>\n\n- [v0.1.73 特性](#-latest-features-v0173)\n\u003C\u002Fdetails>\n\n\u003Cdetails open>\n\u003Csummary>\u003Ch3>🏗️ 系统设计\u003C\u002Fh3>\u003C\u002Fsummary>\n\n- [系统架构](#%EF%B8%8F-system-design-1)\n- [并行处理](#%EF%B8%8F-system-design-1)\n- [实时协作](#%EF%B8%8F-system-design-1)\n- [收敛检测](#%EF%B8%8F-system-design-1)\n- [适应性协调](#%EF%B8%8F-system-design-1)\n\u003C\u002Fdetails>\n\n\u003Cdetails open>\n\u003Csummary>\u003Ch3>🚀 快速入门\u003C\u002Fh3>\u003C\u002Fsummary>\n\n- [📥 安装](#1--installation)\n- [🔐 API配置](#2--api-configuration)\n- [🧩 支持的模型和工具](#3--supported-models-and-tools)\n - [模型](#models)\n - [工具](#tools)\n- [🏃 运行MassGen](#4--run-massgen)\n - [CLI配置参数](#cli-configuration-parameters)\n - [1. 单智能体（最简单的启动方式）](#1-single-agent-easiest-start)\n - [2. 多智能体协作（推荐）](#2-multi-agent-collaboration-recommended)\n - [3. 模型上下文协议（MCP）](#3-model-context-protocol-mcp)\n - [4. 文件系统操作（工作空间管理）](#4-file-system-operations--workspace-management)\n - [5. 项目集成（v0.0.21版本新增功能）](#5-project-integration--user-context-paths-new-in-v0021)\n - [后台配置参考](#backend-configuration-reference)\n - [交互式多轮模式](#interactive-multi-turn-mode)\n- [📊 查看结果](#5--view-results)\n - [实时显示](#real-time-display)\n - [全面日志记录](#comprehensive-logging)\n\u003C\u002Fdetails>\n\n\u003Cdetails open>\n\u003Csummary>\u003Ch3>🤖 自动化与LLM集成\u003C\u002Fh3>\u003C\u002Fsummary>\n\n- [自动化模式](#-automation--llm-integration)\n- [BackgroundShellManager](#using-backgroundshellmanager)\n- [状态文件参考](#statusjson-structure)\n- [完整自动化指南](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Fautomation.html)\n\u003C\u002Fdetails>\n\n\u003Cdetails open>\n\u003Csummary>\u003Ch3>💡 案例研究与示例\u003C\u002Fh3>\u003C\u002Fsummary>\n\n- [案例研究](#-case-studies)\n\u003C\u002Fdetails>\n\n\u003Cdetails open>\n\u003Csummary>\u003Ch3>🗺️ 路线图\u003C\u002Fh3>\u003C\u002Fsummary>\n\n- [近期成就（v0.1.73）](#recent-achievements-v0173)\n- [之前成就（v0.0.3 - v0.1.72）](#previous-achievements-v003---v0172)\n- [关键未来增强功能](#key-future-enhancements)\n - 错误修复与后端改进\n - 高级智能体协作\n - 扩展模型、工具和智能体集成\n - 性能与可扩展性提升\n - 开发者体验优化\n- [v0.1.74路线图](#v0174-roadmap)\n\u003C\u002Fdetails>\n\n\u003Cdetails open>\n\u003Csummary>\u003Ch3>📚 其他资源\u003C\u002Fh3>\u003C\u002Fsummary>\n\n- [🤝 贡献](#-contributing)\n- [📄 许可证](#-license)\n- [⭐ 星标历史](#-star-history)\n\u003C\u002Fdetails>\n\n---\n\n## ✨ 核心功能\n\n| 功能 | 描述 |\n|---------|-------------|\n| **🤝 跨模型\u002F代理协同** | 充分利用由不同前沿模型驱动的代理的优势 |\n| **⚡ 并行处理** | 多个代理同时解决任务 |\n| **👥 智能共享** | 代理之间共享并相互学习彼此的工作成果 |\n| **🔄 达成共识** | 通过协作式优化自然收敛 |\n| **🖥️ 实时可视化** | 默认提供带有时间线、代理卡片和投票跟踪的交互式文本 TUI。此外还支持 Web UI 和丰富显示模式。 |\n\n---\n\n## 🆕 最新特性（v0.1.73）\n\n**🎉 发布日期：2026年4月6日**\n\n**v0.1.73 新增内容：**\n- **🧬 评估标准进化子代理** - 一种可在多轮中不断进化评估标准的新类型子代理。\n- **🛡️ 检查点目标模式** - 初步实现了检查点 MCP 的 `objective` 模式，用于规划不可逆操作的安全性。\n- **👁️ 评估标准可见性提升** - 更清晰地展示代理当前所依据的评估标准。\n\n**体验 v0.1.73 特性：**\n```bash\npip install massgen==0.1.73\nuv run massgen --config @examples\u002Ffeatures\u002Ftrace_analyzer_background.yaml \"创建一个 AI 代理编码的 SVG 图像。\"\n```\n\n→ [查看完整发布历史与示例](massgen\u002Fconfigs\u002FREADME.md#release-history--examples)\n\n---\n\n## 🏗️ 系统设计\n\nMassGen 采用专为**无缝多代理协作**设计的架构运行：\n\n```mermaid\ngraph TB\n O[🚀 MassGen 协调器\u003Cbr\u002F>📋 任务分配与协调]\n\n subgraph 协作代理\n A1[代理 1\u003Cbr\u002F>🏗️ Anthropic\u002FClaude + 工具]\n A2[代理 2\u003Cbr\u002F>🌟 Google\u002FGemini + 工具]\n A3[代理 3\u003Cbr\u002F>🤖 OpenAI\u002FGPT + 工具]\n A4[代理 4\u003Cbr\u002F>⚡ xAI\u002FGrok + 工具]\n end\n\n H[🔄 共享协作中心\u003Cbr\u002F>📡 实时通知与共识]\n\n O --> A1 & A2 & A3 & A4\n A1 & A2 & A3 & A4 \u003C--> H\n\n classDef orchestrator fill:#e1f5fe,stroke:#0288d1,stroke-width:3px\n classDef agent fill:#f3e5f5,stroke:#7b1fa2,stroke-width:2px\n classDef hub fill:#e8f5e8,stroke:#388e3c,stroke-width:2px\n\n class O orchestrator\n class A1,A2,A3,A4 agent\n class H hub\n```\n\n系统的运行流程基于以下核心原则：\n\n**并行处理** - 多个代理同时处理同一任务，各自发挥其独特能力（不同的模型、工具和专业方法）。\n\n**实时协作** - 代理通过通知系统持续共享工作摘要和见解，从而相互学习并在此基础上构建集体知识。\n\n**收敛检测** - 系统智能监测代理何时在解决方案上达到稳定状态，并通过自然协作而非强制一致达成共识。\n\n**适应性协调** - 当代理从其他成员处获得新见解时，可以重新开始并优化工作，从而形成动态且响应迅速的解决问题环境。\n\n这种协作方式确保最终输出能够充分利用多个 AI 系统的集体智慧，产生比单个代理单独完成时更为稳健和全面的结果。\n\n---\n\n> 📖 **完整文档：** 如需全面指南、API 参考及详细示例，请访问 **[MassGen 官方文档](https:\u002F\u002Fdocs.massgen.ai\u002F)**\n\n---\n\n## 🚀 快速入门\n\n### 1. 📥 安装\n\n**方法 1：PyPI 安装**（推荐 - Python 3.11+）：\n\n```bash\n# 使用 pip 安装 MassGen\npip install massgen\n\n# 或使用 uv（更快）\npip install uv\nuv venv && source .venv\u002Fbin\u002Factivate\nuv pip install massgen\n\n# 如果您使用 uv 安装 massgen，请确保先激活虚拟环境 source .venv\u002Fbin\u002Factivate，\n# 或在所有命令前加上 \"uv run\"。\n```\n\n**快速启动设置**（最快运行方式）：\n\n```bash\n# 步骤 1：设置 API 密钥、Docker 和技能\nuv run massgen --setup\n\n# 步骤 2：创建简单配置并启动\nuv run massgen --quickstart\n```\n\n`--setup` 命令将：\n- 配置您的 API 密钥（OpenAI、Anthropic、Google、xAI）\n- 提供设置 Docker 镜像以执行代码的选项\n- 提供安装技能的选项（openskills、Anthropic\u002FOpenAI\u002FVercel 系列、Agent Browser 技能、Crawl4AI）\n\n`--quickstart` 命令将：\n- 询问您希望使用的代理数量（1–5 个，默认 3 个）\n- 询问每个代理的后端或模型\n- 对于 GPT-5x 模型，会询问 `reasoning.effort` 参数（`low|medium|high`；Codex GPT-5 模型还包括 `xhigh`）\n- 自动检测 Docker 是否可用，并配置执行模式\n- 如果选择 Docker 模式，则会显示技能步骤，您可以从中选择软件包（基于 openskills 的 Anthropic\u002FOpenAI\u002FVercel\u002FAgent Browser 以及 Crawl4AI），并实时安装这些技能\n- 创建一个即用型配置，并启动进入交互式 TUI 模式。\n\n**🤖 在您的 AI 编码代理中使用 MassGen：**\n\n安装 [MassGen 技能](https:\u002F\u002Fgithub.com\u002Fmassgen\u002Fskills)，以便直接从 Claude Code、OpenAI Codex、GitHub Copilot、Cursor 以及其他支持 [Agent Skills](https:\u002F\u002Fagentskills.io\u002Fhome) 标准的 **40 多种代理** 中调用 MassGen：\n\n```bash\nnpx skills add massgen\u002Fskills\n```\n\n然后使用 `\u002Fmassgen`（Claude Code）或 `$massgen`（Codex）来运行多代理评估、规划、编写规格说明或任何通用任务。有关各代理的具体安装选项，请参阅 [技能文档](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Fskills.html)。\n\n**🖥️ 文本 TUI（默认显示模式）：**\n\nMassGen 默认启动交互式终端用户界面 (TUI)，提供：\n- 📊 所有代理活动的**实时时间线**\n- 🎯 每位团队成员的**独立代理状态卡**\n- 🗳️ **投票可视化**及共识跟踪\n- 💬 **多轮对话**管理\n- ⌨️ **键盘控制**导航（↑\u002F↓ 滚动，'q' 取消）\n\n**旧版富文本显示：**\n```bash\nmassgen --display rich \"您的问题\"\n```\n\n**替代方案：完整设置向导**\n\n如需更多控制，可使用完整的配置向导：\n```bash\nuv run massgen --init\n```\n\n该向导将引导您完成用例选择（研究、代码、问答等）以及高级配置选项。\n\n**设置完成后：**\n```bash\n# 交互模式\nuv run massgen\n\n# 单一查询\nuv run massgen \"您的问题在这里\"\n\n# 使用示例配置\nuv run massgen --config @examples\u002Fbasic\u002Fmulti\u002Fthree_agents_default \"您的问题\"\n```\n\n→ 请参阅 [安装指南](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fquickstart\u002Finstallation.html) 获取完整的设置说明。\n\n**方法 2：开发环境安装**（适用于贡献者）：\n\n**克隆仓库**\n```bash\ngit clone https:\u002F\u002Fgithub.com\u002FLeezekun\u002FMassGen.git\ncd MassGen\n```\n\n**以可编辑模式使用 pip 安装**\n\n**选项 1（推荐）：使用 uv 安装（更快）**\n\n```bash\nuv venv\nsource .venv\u002Fbin\u002Factivate # Windows: .venv\\Scripts\\activate\nuv pip install -e .\n\n# 如果您使用 uv 安装 massgen，请确保先激活虚拟环境 source .venv\u002Fbin\u002Factivate，\n# 或在所有命令前加上 \"uv run\"。\n\n# 自动化设置（适用于所有平台）——安装依赖项、技能、Docker 镜像，并配置 API 密钥\nuv run massgen --setup\n\n# 或者使用 Bash 脚本（仅限 Unix\u002FLinux\u002FmacOS），需手动配置 API 密钥，详见下文\nuv run .\u002Fscripts\u002Finit.sh\n\n# 如果您希望稍后再安装其他依赖项\n# 这里有一个轻量级的设置脚本，仅安装技能（适用于所有平台）\nuv run massgen --setup-skills\n\n# 或者使用 Bash 脚本（仅限 Unix\u002FLinux\u002FmacOS）\nuv run .\u002Fscripts\u002Finit_skills.sh\n```\n\n**选项 2：使用传统 Python 环境**\n\n```bash\npip install -e .\n\n# 可选：集成外部框架\npip install -e \".[external]\"\n\n# 自动化设置（适用于所有平台）——安装依赖项、技能、Docker 镜像，并配置 API 密钥\nmassgen --setup\n\n# 或者使用 Bash 脚本（仅限 Unix\u002FLinux\u002FmacOS），需手动配置 API 密钥，详见下文\n.\u002Fscripts\u002Finit.sh\n\n# 如果您希望稍后再安装其他依赖项\n# 这里有一个轻量级的设置脚本，仅安装技能（适用于所有平台）\nmassgen --setup-skills\n\n# 或者使用 Bash 脚本（仅限 Unix\u002FLinux\u002FmacOS）\n.\u002Fscripts\u002Finit_skills.sh\n```\n\n> **注意：** `--setup` 和 `--setup-skills` 命令支持跨平台运行（Windows、macOS、Linux）。Bash 脚本（`init.sh`、`init_skills.sh`）仅适用于 Unix 系统，但提供了额外的开发环境设置，例如构建 Docker 镜像。\n\n\u003Cdetails>\n\u003Csummary>\u003Cb>替代安装方法\u003C\u002Fb>（点击展开）\u003C\u002Fsummary>\n\n**使用 uv 和 venv：**\n```bash\ngit clone https:\u002F\u002Fgithub.com\u002FLeezekun\u002FMassGen.git\ncd MassGen\nuv venv\nsource .venv\u002Fbin\u002Factivate # Windows: .venv\\Scripts\\activate\nuv pip install -e .\n```\n\n**使用传统 Python venv：**\n```bash\ngit clone https:\u002F\u002Fgithub.com\u002FLeezekun\u002FMassGen.git\ncd MassGen\npython -m venv .venv\nsource .venv\u002Fbin\u002Factivate # Windows: .venv\\Scripts\\activate\npip install -e .\n```\n\n**使用 uv 工具进行全局安装：**\n```bash\ngit clone https:\u002F\u002Fgithub.com\u002FLeezekun\u002FMassGen.git\ncd MassGen\nuv tool install -e .\n# 现在可以在任何目录下运行\nuv tool run massgen --config @examples\u002Fbasic\u002Fmulti\u002Fthree_agents_default \"Question\"\n```\n\n**向后兼容（uv run）：**\n```bash\ncd \u002Fpath\u002Fto\u002FMassGen\nuv run massgen --config @examples\u002Fbasic\u002Fmulti\u002Fthree_agents_default \"Question\"\nuv run python -m massgen.cli --config config.yaml \"Question\"\n```\n\n\u003C\u002Fdetails>\n\n**可选 CLI 工具：**\n```bash\n# Claude Code CLI — 高级编码助手\nnpm install -g @anthropic-ai\u002Fclaude-code\n\n# LM Studio — 本地模型推理\n# macOS\u002FLinux：\nsudo ~\u002F.lmstudio\u002Fbin\u002Flms bootstrap\n# Windows：\ncmd \u002Fc %USERPROFILE%\\.lmstudio\\bin\\lms.exe bootstrap\n```\n\n**设置完成后：**\n```bash\n# 交互模式\nuv run massgen\n\n# 单次查询\nuv run massgen \"您的问题\"\n\n# 使用示例配置\nuv run massgen --config @examples\u002Fbasic\u002Fmulti\u002Fthree_agents_default \"您的问题\"\n```\n\n### 2. 🔐 API 配置\n\n**在工作目录中创建一个 `.env` 文件，并填写您的 API 密钥：**\n\n```bash\n# 将此模板复制到 .env 文件中，并添加您的 API 密钥\nOPENAI_API_KEY=sk-...\nANTHROPIC_API_KEY=sk-ant-...\nGOOGLE_API_KEY=...\nXAI_API_KEY=...\n\n# 可选：其他提供商的密钥\nCEREBRAS_API_KEY=...\nTOGETHER_API_KEY=...\nGROQ_API_KEY=...\nOPENROUTER_API_KEY=...\n```\n\nMassGen 会自动从当前目录下的 `.env` 文件中加载 API 密钥。\n\n→ **完整设置指南（包含所有提供商）：** 请参阅文档中的 [API 密钥配置](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fquickstart\u002Finstallation.html#api-key-configuration)\n\n**获取 API 密钥：**\n - [OpenAI](https:\u002F\u002Fplatform.openai.com\u002Fapi-keys) | [Claude](https:\u002F\u002Fdocs.anthropic.com\u002Fen\u002Fapi\u002Foverview) | [Gemini](https:\u002F\u002Fai.google.dev\u002Fgemini-api\u002Fdocs) | [Grok](https:\u002F\u002Fdocs.x.ai\u002Fdocs\u002Foverview)\n - [Azure OpenAI](https:\u002F\u002Flearn.microsoft.com\u002Fen-us\u002Fazure\u002Fai-services\u002Fopenai\u002F) | [Cerebras](https:\u002F\u002Finference-docs.cerebras.ai\u002Fintroduction) | [OpenRouter](https:\u002F\u002Fopenrouter.ai\u002Fdocs\u002Fapi\u002Fapi-reference\u002Fapi-keys\u002Fcreate-keys) | [更多提供商...](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Freference\u002Fsupported_models.html)\n\n### 3. 🧩 支持的模型与工具\n\n#### 模型\n\n当前系统支持多家具备先进能力的模型提供商：\n\n**基于 API 的模型：**\n- **OpenAI**：GPT-5.2（推荐默认）、GPT-5.1、GPT-5 系列（GPT-5、GPT-5-mini、GPT-5-nano）、GPT-5.1-Codex 系列、GPT-4.1 系列、GPT-4o、o4-mini，支持推理、网络搜索、代码解释器及计算机使用功能\n - **注意**：我们推荐使用 GPT-5.2\u002F5.1\u002F5 而非 Codex 模型。Codex 模型[针对较短的系统消息进行了优化](https:\u002F\u002Fcookbook.openai.com\u002Fexamples\u002Fgpt-5-codex_prompting_guide)，可能无法很好地配合 MassGen 的协调提示。\n - **推理设置**：GPT-5.1 和 GPT-5.2 默认为 `reasoning: none`。MassGen 在未提供推理配置时会自动设置 `reasoning.effort: medium`，这与 GPT-5 的默认行为一致。\n- **Azure OpenAI**：任何 Azure 部署的模型（GPT-4、GPT-4o、GPT-35-turbo 等）\n- **Claude \u002F Anthropic**：Claude Opus 4.5、Claude Haiku 4.5、Claude Sonnet 4.5、Claude Opus 4.1、Claude Sonnet 4\n - 高级工具支持：网络搜索、代码执行、Files API、程序化工具调用、延迟加载的工具搜索\n- **Claude Code**：原生 Claude Code SDK，支持服务器端会话持久化和内置开发工具\n- **Gemini**：Gemini 3 Pro、Gemini 2.5 Flash、Gemini 2.5 Pro，支持代码执行和上下文增强\n- **Grok \u002F xAI**：Grok-4.1、Grok-4、Grok-3、Grok-3-mini，配备 Grok 实时搜索功能\n- **Cerebras AI**：为支持的模型提供超快速推理\n- **Together AI**、**Fireworks AI**、**Groq**：为 LLaMA、Mistral、Qwen 等开源模型提供快速推理\n- **OpenRouter**：多模型聚合平台，动态列出 400 多种模型\n- **Kimi \u002F Moonshot**：通过兼容 OpenAI 的 API 使用中文 AI 模型\n- **Nebius AI Studio**：云端推理平台\n- **POE**：Quora 的 AI 平台，可动态发现模型\n- **Qwen \u002F Alibaba**：通过 DashScope API 使用 Qwen 模型\n- **Z AI \u002F Zhipu**：GLM-4.5 及相关模型\n\n**本地模型支持：**\n- **vLLM & SGLang**：统一的推理后端，同时支持 vLLM 和 SGLang 服务器\n - vLLM（端口 8000）和 SGLang（端口 30000），提供兼容 OpenAI 的 API\n - 支持 `top_k`、`repetition_penalty`、`chat_template_kwargs` 参数\n - SGLang 特有的 `separate_reasoning` 参数，适用于需要独立思考的模型\n - 混合服务器部署示例配置文件：`two_qwen_vllm_sglang.yaml`\n\n- **LM Studio**：在本地运行开放权重模型，并实现自动服务器管理\n - 自动安装 LM Studio CLI\n - 自动下载并加载模型\n - 支持 LLaMA、Mistral、Qwen 等开放权重模型\n\n→ 完整模型列表及配置详情，请参阅 [支持的模型](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Freference\u002Fsupported_models.html)\n\n#### 工具\n\nMassGen 代理可以利用多种工具来提升其问题解决能力：\n\n- **内置工具**：网络搜索、代码执行、Bash\u002FShell（取决于提供商）\n- **文件系统**：原生文件操作或通过 MCP\n- **MCP 集成**：连接任意 MCP 服务器以扩展功能\n- **自定义工具**：通过 YAML 配置定义您自己的工具\n- **多模态**：图像、音频、视频的理解与生成（原生或通过自定义工具）\n\n→ 有关详细后端能力和工具支持矩阵，请参阅 [用户指南 - 后端](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Fbackends.html#backend-capabilities)\n\n---\n\n### 4. 🏃 运行 MassGen\n\n> **完整使用指南**：关于所有使用模式、高级功能以及交互式多轮会话，请参阅 [运行 MassGen](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fquickstart\u002Frunning-massgen.html)\n\n#### 🚀 入门\n\n#### CLI 配置参数\n\n| 参数 | 描述 |\n|-------------------|-------------|\n| `--config` | 指向包含代理定义、模型参数、后端参数和 UI 设置的 YAML 配置文件路径 |\n| `--backend` | 快速设置时使用的后端类型，无需配置文件（`claude`、`claude_code`、`gemini`、`grok`、`openai`、`azure_openai`、`zai`）。对于[具有默认后端的模型](massgen\u002Futils.py)，此选项为可选。|\n| `--model` | 快速设置时使用的模型名称（例如，`gemini-2.5-flash`、`gpt-5-nano`等）。`--config` 和 `--model` 互斥——只能选择其中之一。|\n| `--system-message` | 快速设置模式下代理的系统提示语。如果提供了 `--config`，则忽略 `--system-message`。|\n| `--cwd-context` | 将当前工作目录作为运行时上下文路径添加：`ro`\u002F`read` 表示只读，`rw`\u002F`write` 表示读写权限。在 TUI 中，这将初始化与 `Ctrl+P` 相同的状态。|\n| `--plan` | 仅规划模式。代理会创建结构化的任务计划，但不会自动执行。|\n| `--plan-depth` | `--plan` 的规划粒度：`dynamic`、`shallow`、`medium` 或 `deep`。|\n| `--plan-and-execute` | 同时运行规划和执行两个阶段：先创建计划，再自动执行。|\n| `--execute-plan` | 根据路径、计划 ID 或 `latest` 执行现有计划。|\n| `--no-display` | 禁用实时流媒体 UI 协调显示（回退到简单的文本输出）。|\n| `--no-logs` | 禁用实时日志记录。|\n| `--debug` | 启用调试模式并进行详细日志记录（v0.0.13 新增）。显示详细的编排活动、代理消息、后端操作和工具调用。调试日志保存至 `agent_outputs\u002Flog_{time}\u002Fmassgen_debug.log`。|\n| `\"\u003Cyour question>\"` | 可选的单次提问输入；若省略，MassGen 将进入交互式聊天模式。|\n\n#### **0. 兰开普兼容 HTTP 服务器（新增）**\n\n将 MassGen 作为**OpenAI 兼容**的 HTTP API 运行（FastAPI + Uvicorn）。这对于将 MassGen 集成到预期接收 `POST \u002Fv1\u002Fchat\u002Fcompletions` 请求的现有工具中非常有用。\n\n```bash\n# 启动服务器（默认：主机 0.0.0.0，端口 4000）\nmassgen serve\n\n# 显式绑定 + 模型\u002F配置默认值\nmassgen serve --host 0.0.0.0 --port 4000 --config path\u002Fto\u002Fconfig.yaml --default-model gpt-5\n```\n\n**端点**\n\n- `GET \u002Fhealth`\n- `POST \u002Fv1\u002Fchat\u002Fcompletions`（支持 `stream: true` SSE 和 OpenAI 风格的工具调用）\n\n**cURL 示例**\n\n```bash\n# 健康检查\ncurl http:\u002F\u002Flocalhost:4000\u002Fhealth\n\n# 非流式对话完成\ncurl http:\u002F\u002Flocalhost:4000\u002Fv1\u002Fchat\u002Fcompletions \\\n -H \"Content-Type: application\u002Fjson\" \\\n -d '{\n \"model\": \"massgen\",\n \"messages\": [{\"role\": \"user\", \"content\": \"hi\"}],\n \"stream\": false\n }'\n\n# 流式（Server-Sent Events）\ncurl -N http:\u002F\u002Flocalhost:4000\u002Fv1\u002Fchat\u002Fcompletions \\\n -H \"Content-Type: application\u002Fjson\" \\\n -d '{\n \"model\": \"massgen\",\n \"messages\": [{\"role\": \"user\", \"content\": \"hi\"}],\n \"stream\": true\n }'\n```\n\n**注意事项**\n\n- 支持客户端提供的 `tools`，但与 MassGen 工作流工具名称冲突的工具将被拒绝。\n- 环境变量（可选）：`MASSGEN_SERVER_HOST`、`MASSGEN_SERVER_PORT`、`MASSGEN_SERVER_DEFAULT_CONFIG`、`MASSGEN_SERVER_DEFAULT_MODEL`、`MASSGEN_SERVER_DEBUG`。\n\n\n#### **1. 单个代理（最简单启动方式）**\n\n**快速启动命令：**\n\n# 使用任何支持的模型进行快速测试 - 无需配置\nuv run python -m massgen.cli --model claude-sonnet-4-5-20250929 \"什么是机器学习？\"\nuv run python -m massgen.cli --model gemini-3-pro-preview \"解释量子计算\"\nuv run python -m massgen.cli --model gpt-5-nano \"总结最新的AI发展\"\n```\n\n**配置：**\n\n使用 `agent` 字段定义一个具有后端和设置的单一智能体：\n\n```yaml\nagent:\n id: \"\u003Cagent_name>\"\n backend:\n type: \"azure_openai\" | \"chatcompletion\" | \"claude\" | \"claude_code\" | \"gemini\" | \"grok\" | \"openai\" | \"zai\" | \"lmstudio\" #后端类型\n model: \"\u003Cmodel_name>\" # 模型名称\n api_key: \"\u003Coptional_key>\" # 后端的API密钥。默认使用环境变量。\n system_message: \"...\" # 单一智能体的系统消息\n```\n\n→ [查看所有单智能体配置](massgen\u002Fconfigs\u002Fbasic\u002Fsingle\u002F)\n\n\n#### **2. 多智能体协作（推荐）**\n\n**配置：**\n\n使用 `agents` 字段定义多个智能体，每个智能体都有自己的后端和配置：\n\n**快速入门命令：**\n\n```bash\n# 三个强大的智能体协同工作——Gemini、GPT-5 和 Grok\nmassgen --config @examples\u002Fbasic\u002Fmulti\u002Fthree_agents_default \\\n \"分析可再生能源的优缺点\"\n```\n\n**这展示了 MassGen 的核心优势：**\n- **Gemini 3 Pro** - 带有网络搜索的快速研究\n- **GPT-5 Nano** - 具有代码执行能力的高级推理\n- **Grok-4 Fast** - 实时信息与不同视角\n\n```yaml\nagents: # 多个智能体（替代 'agent'）\n - id: \"\u003Cagent1 name>\"\n backend:\n type: \"azure_openai\" | \"chatcompletion\" | \"claude\" | \"claude_code\" | \"gemini\" | \"grok\" | \"openai\" | \"zai\" | \"lmstudio\" #后端类型\n model: \"\u003Cmodel_name>\" # 模型名称\n api_key: \"\u003Coptional_key>\" # 后端的API密钥。默认使用环境变量。\n system_message: \"...\" # 单一智能体的系统消息\n - id: \"...\"\n backend:\n type: \"...\"\n model: \"...\"\n ...\n system_message: \"...\"\n```\n\n→ [探索更多多智能体配置](massgen\u002Fconfigs\u002Fbasic\u002Fmulti\u002F)\n\n\n#### **3. 模型上下文协议 (MCP)**\n\n[模型上下文协议](https:\u002F\u002Fmodelcontextprotocol.io\u002F) (MCP) 标准化了应用程序向语言模型暴露工具和上下文的方式。根据官方文档：\n\n>MCP 是一种开放协议，用于标准化应用程序向 LLM 提供上下文的方式。可以将 MCP 看作是 AI 应用程序的 USB-C 接口。就像 USB-C 为您的设备提供了一种连接各种外设和配件的标准方式一样，MCP 也为 AI 模型提供了连接不同数据源和工具的标准方式。\n\n**MCP 配置参数：**\n\n| 参数 | 类型 | 必需 | 描述 |\n|-----------|------|----------|-------------|\n| `mcp_servers` | 字典 | **是**（对于 MCP） | MCP 服务器定义的容器 |\n| └─ `type` | 字符串 | 是 | 传输方式：“stdio” 或 “streamable-http” |\n| └─ `command` | 字符串 | 仅 stdio | 运行 MCP 服务器的命令 |\n| └─ `args` | 列表 | 仅 stdio | 命令的参数 |\n| └─ `url` | 字符串 | 仅 http | 服务器端点 URL |\n| └─ `env` | 字典 | 否 | 要传递的环境变量 |\n| `allowed_tools` | 列表 | 否 | 白名单特定工具（若省略，则所有工具可用） |\n| `exclude_tools` | 列表 | 否 | 黑名单危险或不需要的工具 |\n\n\n**快速入门命令（[在此处查看后端是否支持 MCP](#tools)）：**\n\n```bash\n# GPT-5 天气服务\nmassgen --config @examples\u002Ftools\u002Fmcp\u002Fgpt5_nano_mcp_example \\\n \"纽约本周的天气预报是什么？\"\n\n# Gemini 多工具 MCP —— 搜索 + 天气 + 文件系统（需要在 .env 中设置 BRAVE_API_KEY）\nmassgen --config @examples\u002Ftools\u002Fmcp\u002Fmultimcp_gemini \\\n \"查找巴黎最好的餐厅，并将推荐保存到文件中\"\n```\n\n**配置：**\n\n```yaml\nagents:\n # 基本 MCP 配置：\n backend:\n type: \"openai\" # 您选择的后端\n model: \"gpt-5-mini\" # 您选择的模型\n\n # 在此处添加 MCP 服务器\n mcp_servers:\n weather: # 服务器名称（由您命名）\n type: \"stdio\" # 通信类型\n command: \"npx\" # 要运行的命令\n args: [\"-y\", \"@modelcontextprotocol\u002Fserver-weather\"] # MCP 服务器包\n\n # 就这样！智能体现在可以查询天气了。\n\n # 多工具 MCP 示例：\n backend:\n type: \"gemini\"\n model: \"gemini-3.0-pro-preview\"\n mcp_servers:\n # 网络搜索\n search:\n type: \"stdio\"\n command: \"npx\"\n args: [\"-y\", \"@modelcontextprotocol\u002Fserver-brave-search\"]\n env:\n BRAVE_API_KEY: \"${BRAVE_API_KEY}\" # 在 .env 文件中设置\n\n # 基于 HTTP 的 MCP 服务器（streamable-http 传输）\n custodm_api:\n type: \"streamable-http\" # 适用于 HTTP\u002FSSE 服务器\n url: \"http:\u002F\u002Flocalhost:8080\u002Fmcp\u002Fsse\" # 服务器端点\n\n\n # 工具配置（MCP 工具会自动发现）\n allowed_tools: # 可选：白名单特定工具\n - \"mcp__weather__get_current_weather\"\n - \"mcp__test_server__mcp_echo\"\n - \"mcp__test_server__add_numbers\"\n\n exclude_tools: # 可选：黑名单特定工具\n - \"mcp__test_server__current_time\"\n```\n\n→ [查看更多 MCP 示例](massgen\u002Fconfigs\u002Ftools\u002Fmcp\u002F)\n\n→ 如需全面的 MCP 集成指南，请参阅 [MCP 集成](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Fmcp_integration.html)\n\n#### **4. 文件系统操作与工作区管理**\n\nMassGen 通过多种后端提供了全面的文件系统支持，使智能体能够在有序的工作区中读取、写入和操作文件。\n\n\n**文件系统配置参数：**\n\n| 参数 | 类型 | 必需 | 描述 |\n|-----------|------|----------|-------------|\n| `cwd` | 字符串 | **是**（用于文件操作） | 文件操作的工作目录（智能体专属的工作区） |\n| `snapshot_storage` | 字符串 | 是 | 工作区快照的存储目录 |\n| `agent_temporary_workspace` | 字符串 | 是 | 临时工作区的父目录 |\n\n\n**快速入门命令：**\n\n```bash\n# Claude Code 文件操作\nmassgen --config @examples\u002Ftools\u002Ffilesystem\u002Fclaude_code_single \\\n \"创建一个 Python 网络爬虫，并将结果保存为 CSV 文件\"\n\n# 多智能体文件协作\nmassgen --config @examples\u002Ftools\u002Ffilesystem\u002Fclaude_code_context_sharing \\\n \"生成包含图表和分析的综合项目报告\"\n```\n\n**配置：**\n\n```yaml\n# 基本工作区设置：\nagents:\n - id: \"file-agent\"\n backend:\n type: \"claude_code\" # 支持文件操作的后端\n cwd: \"workspace\" # 文件操作的隔离工作区\n\n# 多智能体工作空间隔离：\nagents:\n - id: \"agent_a\"\n backend:\n type: \"claude_code\"\n cwd: \"workspace1\" # 智能体专用工作空间\n\n - id: \"agent_b\"\n backend:\n type: \"gemini\"\n cwd: \"workspace2\" # 独立工作空间\n\norchestrator:\n snapshot_storage: \"snapshots\" # 共享快照目录\n agent_temporary_workspace: \"temp_workspaces\" # 临时工作空间管理\n```\n**可用文件操作：**\n- **Claude Code**: 内置工具（读取、写入、编辑、多编辑、Bash、Grep、Glob、LS、TodoWrite）\n- **其他后端**：通过 [MCP 文件系统服务器](https:\u002F\u002Fgithub.com\u002Fmodelcontextprotocol\u002Fservers\u002Fblob\u002Fmain\u002Fsrc%2Ffilesystem%2FREADME.md)\n\n**工作空间管理：**\n- **隔离的工作空间**：每个智能体的 `cwd` 完全隔离且可写\n- **快照存储**：在 Claude Code 智能体之间共享工作空间上下文\n- **临时工作空间**：智能体可以访问之前的协调结果\n\n→ [查看更多文件系统示例](massgen\u002Fconfigs\u002Ftools\u002Ffilesystem\u002F)\n\n> ⚠️ **重要安全警告**\n>\n> MassGen 智能体可以在其被允许的目录内**自主地读取、写入、修改和删除文件**。\n>\n> **在运行具有文件系统访问权限的 MassGen 之前：**\n> - 仅授予对您愿意让智能体修改的目录的访问权限\n> - 使用权限系统在必要时限制写入权限\n> - 建议先在一个隔离的目录或虚拟环境中进行测试\n> - 在授予写入权限前备份重要文件\n> - 仔细检查 `context_paths` 配置\n>\n> 一旦授予权限，智能体将无需额外确认即可执行文件操作。\n\n→ 如需全面的文件操作指南，请参阅 [文件操作](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Ffile_operations.html)\n\n#### **5. 项目集成与用户上下文路径（v0.0.21 新增）**\n\n直接与您现有的项目协作！用户上下文路径允许您在保持细粒度权限控制的同时，与所有智能体共享特定目录。这使得在您的实际代码库、文档和数据上进行安全的多智能体协作成为可能。\n\nMassGen 会自动将其所有工作文件组织到项目根目录下的 `.massgen\u002F` 目录中，从而保持项目整洁，并便于将 MassGen 的临时文件从版本控制系统中排除。\n\n**项目集成参数：**\n\n| 参数 | 类型 | 必需 | 描述 |\n|-----------|------|----------|-------------|\n| `context_paths` | 列表 | **是**（用于项目集成） | 所有智能体共享的目录 |\n| └─ `path` | 字符串 | 是 | 您项目目录的绝对或相对路径（**必须是目录，不能是文件**） |\n| └─ `permission` | 字符串 | 是 | 访问级别：“read” 或 “write”（写权限仅适用于最终智能体） |\n| └─ `protected_paths` | 列表 | 否 | 不受修改影响的文件\u002F目录（相对于上下文路径） |\n\n**⚠️ 重要提示：**\n- 上下文路径必须指向**目录**，而非单个文件\n- 路径可以是**绝对路径**或**相对路径**（相对于当前工作目录解析）\n- **写权限**仅在展示阶段适用于**最终智能体**\n- 在协调过程中，所有上下文路径均为**只读**，以保护您的文件\n- MassGen 会在启动时验证所有路径，对于缺失路径或文件路径，将显示明确的错误信息\n\n\n**快速入门命令：**\n\n```bash\n# 多智能体协作，改进位于 `massgen\u002Fconfigs\u002Fresources\u002Fv0.0.21-example` 的网站\nmassgen --config @examples\u002Ftools\u002Ffilesystem\u002Fgpt5mini_cc_fs_context_path \"请对网站进行以下优化：1) 添加带有平滑过渡效果的暗色\u002F浅色主题切换功能；2) 实现一项交互式功能，帮助用户更好地参与博客内容（您可以选择搜索、按主题筛选、阅读时间估算、社交分享、评论互动等）；3) 通过 CSS 动画或过渡效果提升视觉效果，使网站更具现代感和响应性。请使用原生 JavaScript 实现，并在具体实现细节上发挥创意。\"\n```\n\n**配置：**\n\n```yaml\n# 基本项目集成：\nagents:\n - id: \"code-reviewer\"\n backend:\n type: \"claude_code\"\n cwd: \"workspace\" # 智能体的隔离工作区\n\norchestrator:\n context_paths:\n - path: \".\" # 当前目录（相对路径）\n permission: \"write\" # 最终智能体可以创建\u002F修改文件\n protected_paths: # 可选：不受修改影响的文件\n - \".env\"\n - \"config.json\"\n - path: \"\u002Fhome\u002Fuser\u002Fmy-project\u002Fsrc\" # 绝对路径示例\n permission: \"read\" # 智能体可以分析您的代码\n\n# 高级：多智能体项目协作\nagents:\n - id: \"analyzer\"\n backend:\n type: \"gemini\"\n cwd: \"analysis_workspace\"\n\n - id: \"implementer\"\n backend:\n type: \"claude_code\"\n cwd: \"implementation_workspace\"\n\norchestrator:\n context_paths:\n - path: \"..\u002Flegacy-app\u002Fsrc\" # 相对路径，指向现有代码库\n permission: \"read\" # 只读现有代码库\n - path: \"..\u002Flegacy-app\u002Ftests\"\n permission: \"write\" # 最终智能体可以编写新测试\n protected_paths: # 保护特定测试文件\n - \"integration_tests\u002Fproduction_data_test.py\"\n - path: \"\u002Fhome\u002Fuser\u002Fmodernized-app\" # 绝对路径\n permission: \"write\" # 最终智能体可以创建现代化版本\n```\n\n**这展示了项目集成：**\n- **真实项目访问** - 使用您的实际代码库，而非副本\n- **安全权限控制** - 对智能体可读\u002F可修改的内容进行细粒度控制\n- **多智能体协作** - 多个智能体可安全地协同处理同一项目\n- **上下文智能体**（协调期间）：始终为只读访问，以保护您的文件\n- **最终智能体**（最终执行）：获得配置的权限（只读或写入）\n\n**使用场景：**\n- **代码审查**：智能体分析源代码并提出改进建议\n- **文档生成**：智能体阅读项目文档以理解上下文，并生成更新内容\n- **数据处理**：智能体访问共享数据集并生成分析报告\n- **项目迁移**：智能体检查现有项目并创建现代化版本\n\n**整洁的项目组织结构：**\n```\nyour-project\u002F\n├── .massgen\u002F # 所有 MassGen 状态文件\n│ ├── sessions\u002F # 多轮对话历史（若交互式使用）\n│ │ └── session_20240101_143022\u002F\n│ │ ├── turn_1\u002F # 第一轮结果\n│ │ ├── turn_2\u002F # 第二轮结果\n│ │ └── SESSION_SUMMARY.txt # 人类可读摘要\n│ ├── workspaces\u002F # 智能体工作目录\n│ │ ├── agent1\u002F # 单个智能体的工作空间\n│ │ └── agent2\u002F\n│ ├── snapshots\u002F # 工作空间快照，用于协调\n│ └── temp_workspaces\u002F # 上一轮结果，用作上下文\n├── massgen\u002F\n└── ...\n```\n\n**优势：**\n- ✅ **项目整洁** - 所有 MassGen 文件都集中在一个目录中\n- ✅ **轻松 Gitignore** - 只需将 `.massgen\u002F` 添加到 `.gitignore`\n- ✅ **便携性** - 移动或删除 `.massgen\u002F` 不会影响您的项目\n- ✅ **多轮会话** - 对话历史在不同会话间得以保留\n\n**配置自动整理：**\n```yaml\norchestrator:\n # 用户指定简单名称 - MassGen 自动在 .massgen\u002F 下组织\n snapshot_storage: \"snapshots\" # → .massgen\u002Fsnapshots\u002F\n agent_temporary_workspace: \"temp\" # → .massgen\u002Ftemp\u002F\n\nagents:\n - backend:\n cwd: \"workspace1\" # → .massgen\u002Fworkspaces\u002Fworkspace1\u002F\n```\n\n→ 如需全面的项目集成指南，请参阅 [项目集成](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Fproject_integration.html)\n\n**安全注意事项：**\n- **智能体 ID 安全性**：避免使用带有递增数字的智能体 ID（如 `agent1`、`agent2`）。这可能导致投票过程中 ID 泄露\n- **文件访问控制**：必要时使用 MCP 服务器配置限制文件访问\n- **路径验证**：所有上下文路径都会被验证，确保其存在且为目录（而非文件）\n- **仅目录型上下文路径**：上下文路径必须指向目录，而非单个文件\n\n---\n\n#### 各提供商的额外示例\n\n**Claude（递归 MCP 执行 - v0.0.20+）**\n```bash\n# Claude 配合高级工具链\nmassgen --config @examples\u002Ftools\u002Fmcp\u002Fclaude_mcp_example \\\n \"研究并比较北京和上海的天气情况\"\n```\n\n**OpenAI（GPT-5 系列配合 MCP - v0.0.17+）**\n```bash\n# GPT-5 结合天气信息与外部工具\nmassgen --config @examples\u002Ftools\u002Fmcp\u002Fgpt5_nano_mcp_example \\\n \"东京现在的天气如何\"\n```\n\n**Gemini（多服务器 MCP - v0.0.15+）**\n```bash\n# Gemini 配合多个 MCP 服务\nmassgen --config @examples\u002Ftools\u002Fmcp\u002Fmultimcp_gemini \\\n \"在巴黎寻找住宿，并进行社区分析\" # （需要在 .env 中设置 BRAVE_API_KEY）\n```\n\n**Claude Code（开发工具）**\n```bash\n# 专业开发环境，自动配置工作空间\nuv run python -m massgen.cli \\\n --backend claude_code \\\n --model sonnet \\\n \"创建一个带有身份验证功能的 Flask Web 应用\"\n\n# 默认工作空间目录会自动生成：\n# - workspace1\u002F （工作目录）\n# - snapshots\u002F （工作空间快照）\n# - temp_workspaces\u002F （临时智能体工作空间）\n```\n\n**本地模型（LM Studio - v0.0.7+）**\n```bash\n# 在本地运行开源模型\nmassgen --config @examples\u002Fproviders\u002Flocal\u002Flmstudio \\\n \"解释机器学习的概念\"\n```\n\n→ [按提供商浏览](massgen\u002Fconfigs\u002Fproviders\u002F) | [按工具浏览](massgen\u002Fconfigs\u002Ftools\u002F) | [按团队浏览](massgen\u002Fconfigs\u002Fteams\u002F)\n\n#### 其他使用案例示例\n\n**问答与研究：**\n```bash\n# 多视角复杂研究\nmassgen --config @examples\u002Fbasic\u002Fmulti\u002Fgemini_gpt5_claude \\\n \"2025年10月在斯德哥尔摩有哪些值得体验的活动\"\n\n# 特定研究需求\nmassgen --config @examples\u002Fbasic\u002Fmulti\u002Fgemini_gpt5_claude \\\n \"请列出2025年伯克利代理智能峰会中关于代理框架的所有演讲\"\n```\n\n**创意写作：**\n```bash\n# 多个创意智能体共同创作故事\nmassgen --config @examples\u002Fbasic\u002Fmulti\u002Fgemini_gpt5_claude \\\n \"写一篇关于机器人发现音乐的短篇小说\"\n```\n\n**开发与编码：**\n```bash\n# 全栈开发，包含文件操作\nmassgen --config @examples\u002Ftools\u002Ffilesystem\u002Fclaude_code_single \\\n \"创建一个带有身份验证功能的 Flask Web 应用\"\n```\n\n**网页自动化：**（仍在测试中）\n```bash\n# 浏览器自动化，结合截图与报告\n# 前提条件：安装 npm install @playwright\u002Fmcp@latest（用于 Playwright MCP 服务器）\nmassgen --config @examples\u002Ftools\u002Fcode-execution\u002Fmulti_agent_playwright_automation \\\n \"浏览 https:\u002F\u002Fgithub.com\u002FLeezekun\u002FMassGen 中的三个问题，并提出文档改进建议。将截图和建议整合到一个网站中。\"\n\n# 数据提取与分析\nmassgen --config @examples\u002Ftools\u002Fcode-execution\u002Fmulti_agent_playwright_automation \\\n \"前往 https:\u002F\u002Fnews.ycombinator.com，提取前10条新闻，并制作一份总结报告\"\n```\n\n→ [查看详细案例研究](docs\u002Fsource\u002Fexamples\u002Fcase_studies\u002FREADME.md)，包含真实的会话记录和成果\n\n#### 交互模式与进阶用法\n\n**多轮对话：**\n```bash\n# 开始交互式聊天（无需初始问题）\nmassgen --config @examples\u002Fbasic\u002Fmulti\u002Fthree_agents_default\n\n# 快速添加只读 CWD 上下文\nmassgen --config @examples\u002Fbasic\u002Fmulti\u002Fthree_agents_default --cwd-context ro\n\n# 快速添加当前工作目录上下文（读+写）\nmassgen --config @examples\u002Fbasic\u002Fmulti\u002Fthree_agents_default --cwd-context rw\n\n# 调试模式用于故障排除\nmassgen --config @examples\u002Fbasic\u002Fmulti\u002Fthree_agents_default \\\n --debug \"你的问题\"\n```\n\n## 配置文件\n\nMassGen 的配置按功能和使用场景进行组织。详细的组织结构和示例请参阅 [配置指南](massgen\u002Fconfigs\u002FREADME.md)。\n\n**快速导航：**\n- **基础设置**：[单智能体](massgen\u002Fconfigs\u002Fbasic\u002Fsingle\u002F) | [多智能体](massgen\u002Fconfigs\u002Fbasic\u002Fmulti\u002F)\n- **工具集成**：[MCP 服务器](massgen\u002Fconfigs\u002Ftools\u002Fmcp\u002F) | [网络搜索](massgen\u002Fconfigs\u002Ftools\u002Fweb-search\u002F) | [文件系统](massgen\u002Fconfigs\u002Ftools\u002Ffilesystem\u002F)\n- **提供商示例**：[OpenAI](massgen\u002Fconfigs\u002Fproviders\u002Fopenai\u002F) | [Claude](massgen\u002Fconfigs\u002Fproviders\u002Fclaude\u002F) | [Gemini](massgen\u002Fconfigs\u002Fproviders\u002Fgemini\u002F)\n- **专业团队**：[创意团队](massgen\u002Fconfigs\u002Fteams\u002Fcreative\u002F) | [研究团队](massgen\u002Fconfigs\u002Fteams\u002Fresearch\u002F) | [开发团队](massgen\u002Fconfigs\u002Fteams\u002Fdevelopment\u002F)\n\n查看 MCP 服务器设置指南：[Discord MCP](massgen\u002Fconfigs\u002Fdocs\u002FDISCORD_MCP_SETUP.md) | [Twitter MCP](massgen\u002Fconfigs\u002Fdocs\u002FTWITTER_MCP_ENESCINAR_SETUP.md)\n\n#### 后端配置参考\n\n有关所有支持的后端（OpenAI、Claude、Gemini、Grok 等）的详细配置，请参阅：\n\n→ **[后端配置指南](massgen\u002Fconfigs\u002FBACKEND_CONFIGURATION.md)**\n\n#### 交互式多轮模式\n\nMassGen 支持交互式模式，您可以在其中与系统进行持续对话：\n\n```bash\n# 使用单个智能体启动交互式模式（默认未启用工具）\nuv run python -m massgen.cli --model gpt-5-mini\n\n# 使用配置文件启动交互式模式\nuv run python -m massgen.cli \\\n --config massgen\u002Fconfigs\u002Fbasic\u002Fmulti\u002Fthree_agents_default.yaml\n```\n\n**交互式模式功能：**\n- **多轮对话**：多个智能体协作与您进行持续对话\n- **实时协调跟踪**：实时可视化智能体之间的交互、投票及决策过程\n- **实时反馈**：显示智能体和系统的实时状态，并增强协调可视化效果\n- **多行输入**：使用 `\"\"\"` 或 `'''` 输入多行消息\n- **斜杠命令**：\n - `\u002Fhelp` 或 `\u002Fh` - 显示可用命令\n - `\u002Fstatus` - 显示当前系统状态\n - `\u002Fconfig` - 打开配置文件\n - `\u002Fclear` 或 `\u002Freset` - 清除对话历史并重新开始\n - `\u002Fquit`、`\u002Fexit` 或 `\u002Fq` - 退出会话（或按 `Ctrl+C`）\n\n**观看录制演示：**\n\n[![MassGen 案例研究](https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Fmassgen_MassGen_readme_07732971c916.jpg)](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=h1R7fxFJ0Zc)\n\n### 5. 📊 查看结果\n\n系统提供了多种方式来查看和分析结果：\n\n#### 实时显示\n- **实时协作视图**：通过多区域终端显示，查看智能体并行工作的情况\n- **状态更新**：实时显示阶段转换、投票进展和共识形成过程\n- **流式输出**：观看智能体的推理和响应逐步展开\n\n**在此处观看示例：**\n\n[![MassGen 案例研究](https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Fmassgen_MassGen_readme_a17172c9da1c.jpg)](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=Dp2oldJJImw)\n\n#### 全面日志记录\n\n所有会话都会自动记录，包含详细的调试和分析信息。\n\n**实时交互：**\n- 在执行过程中按 `r` 键，可在终端中查看协调表格\n- 实时观察智能体协作、投票并达成共识的过程\n\n##### 日志存储结构\n\n```\n.massgen\u002F\n└── massgen_logs\u002F\n └── log_YYYYMMDD_HHMMSS\u002F # 带时间戳的日志目录\n ├── agent_\u003Cid>\u002F # 智能体特定的协调日志\n │ └── YYYYMMDD_HHMMSS_NNNNNN\u002F # 带时间戳的协调步骤\n │ ├── answer.txt # 智能体在该步骤的回答\n │ ├── context.txt # 智能体可用的上下文\n │ └── workspace\u002F # 智能体工作空间（如果使用了文件系统工具）\n ├── agent_outputs\u002F # 整合后的输出文件\n │ ├── agent_\u003Cid>.txt # 每个智能体的完整输出\n │ ├── final_presentation_agent_\u003Cid>.txt # 获胜智能体的最终答案\n │ ├── final_presentation_agent_\u003Cid>_latest.txt # 最新版本的符号链接\n │ └── system_status.txt # 系统状态及元数据\n ├── final\u002F # 最终展示阶段\n │ └── agent_\u003Cid>\u002F # 获胜智能体的最终成果\n │ ├── answer.txt # 最终答案\n │ └── context.txt # 最终上下文\n ├── coordination_events.json # 结构化的协调事件日志\n ├── coordination_table.txt # 人类可读的协调表格\n ├── vote.json # 最终投票统计及共识信息\n ├── massgen.log # 完整的调试日志（或在调试模式下为 massgen_debug.log）\n ├── snapshot_mappings.json # 工作空间快照元数据\n └── execution_metadata.yaml # 查询、配置及执行详情\n```\n\n##### 重要日志文件\n\n- **协调表格**（`coordination_table.txt`）：完整展示多智能体协调情况，包括事件时间线、投票模式和共识形成过程。\n- **协调事件**（`coordination_events.json`）：所有事件的结构化 JSON 日志（started_streaming、new_answer、vote、restart、final_answer）。\n- **投票汇总**（`vote.json`）：最终投票统计、获胜智能体及共识信息。\n- **执行元数据**（`execution_metadata.yaml`）：原始查询、时间戳、配置及执行环境，便于复现。\n- **智能体输出**（`agent_outputs\u002F`）：所有智能体的完整输出历史及最终展示。\n- **调试日志**（`massgen.log`）：系统运行的完整操作记录、API 调用、工具使用情况及错误追踪信息（使用 `--debug` 可获取详细日志）。\n\n→ 如需全面的日志记录指南及调试技巧，请参阅 [日志与调试](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Flogging.html)\n\n---\n\n## 🤖 自动化与 LLM 集成\n\n**→ 对于 LLM 智能体：请参阅 [AI_USAGE.md](AI_USAGE.md)，获取完整的命令行使用指南**\n\nMassGen 提供了专为 LLM 智能体和程序化工作流设计的 **自动化模式**：\n\n### 快速入门 - 自动化模式\n\n```bash\n# 以最小化输出和状态跟踪运行\nuv run massgen --automation --config your_config.yaml \"你的问题\"\n```\n\n### 综合指南\n\n→ **包含示例的完整自动化指南：** [自动化指南](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Fautomation.html)\n\n涵盖主题：\n- 完整的自动化模式及错误处理\n- 并行实验执行\n- 性能优化建议及故障排除\n\n### Python API 和 LiteLLM\n\n通过熟悉的 LiteLLM\u002FOpenAI 接口以编程方式使用 MassGen：\n\n```python\nfrom dotenv import load_dotenv\nload_dotenv() # 从 .env 文件加载 API 密钥\n\nimport litellm\nfrom massgen import register_with_litellm\n\nregister_with_litellm()\n\n# 多智能体调用，采用斜杠格式：“后端\u002F模型”\nresponse = litellm.completion(\n model=\"massgen\u002Fbuild\",\n messages=[{\"role\": \"user\", \"content\": \"比较人工智能的不同方法\"}],\n optional_params={\"models\": [\"openai\u002Fgpt-5\", \"groq\u002Fllama-3.3-70b\"]}\n)\nprint(response.choices[0].message.content) # 最终共识答案\n```\n\n或者直接使用 Python API：\n\n```python\nfrom dotenv import load_dotenv\nload_dotenv()\n\nimport asyncio\nimport massgen\n\nresult = asyncio.run(massgen.run(\n query=\"什么是机器学习？\",\n models=[\"openai\u002Fgpt-5\", \"gemini\u002Fgemini-3-pro-preview\"]\n))\nprint(result[\"final_answer\"]) # 获胜智能体的共识答案\n```\n\n> **完整 API 参考：** [程序化 API 指南](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Fintegration\u002Fpython_api.html)\n\n---\n\n## 💡 案例研究\n\n要了解 MassGen 在实际中的运作方式，请查看这些基于真实会话日志的详细案例研究：\n\n**精选：**\n- [**多轮持久化记忆**](docs\u002Fsource\u002Fexamples\u002Fcase_studies\u002Fmulti-turn-persistent-memory.md) - 从研究到落地的工作流，展示了记忆系统（v0.1.5）| [📹 观看演示](https:\u002F\u002Fyoutu.be\u002FwWxxFgyw40Y)\n\n**全部案例研究：**\n- [**MassGen 案例研究**](docs\u002Fsource\u002Fexamples\u002Fcase_studies\u002FREADME.md)\n- [**案例研究文档**](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fexamples\u002Fcase_studies.html) - 在线浏览案例研究\n\n---\n\n\n## 🗺️ 路线图\n\nMassGen 目前处于基础阶段，重点是并行、异步的多智能体协作与编排。我们的路线图旨在将这一基础转化为一个高度稳健、智能且用户友好的系统，同时支持前沿研究与探索。\n\n⚠️ **早期阶段提示：** 由于 MassGen 处于积极开发中，随着我们不断优化和改进系统，预计未来会有重大架构变更。\n\n### 最新进展（v0.1.73）\n\n**🎉 发布日期：2026年4月6日**\n\n#### 评估标准进化器与检查点目标\n- **评估标准进化子代理**（[#1047](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1047)）：一种新型子代理，可在各轮中不断进化评估标准——随着运行推进，评估标准将更加精准、更具倾向性。\n- **检查点目标模式（初稿）**（[#1047](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1047)）：检查点 MCP 的初稿，包含 `objective` 模式，用于规划不可逆操作的安全性。\n- **评估标准可见性提升**：更清晰地展示智能体当前依据的评估标准。\n\n### 前期进展（v0.0.3 - v0.1.72）\n\n✅ **Grok 后端更新及熔断器第二阶段（v0.1.72）**：Grok 后端更新至最新版本。LLM API 熔断器现已扩展至 ChatCompletions、Response API 和 Gemini 后端（此前仅适用于 Claude）。\n \n✅ **追踪记忆与评估优化（v0.1.71）**：每轮结束后，后台会启动追踪分析子代理，将执行轨迹中的洞察写入记忆。评估标准生成及系统提示词调优得到改进。\n\n✅ **评估标准重新设计（v0.1.70）**：重新设计了三层评估标准，加入反模式定义和期望陈述。检查清单驱动的评估机制进一步完善。新增快速迭代模式、WebUI 审核模态窗口以及后台追踪分析功能。\n\n✅ **WebUI 自动化与技能优化（v0.1.69）**：WebUI 自动化功能无需浏览器交互即可自动启动。MassGen 技能重新设计，提升易用性并与 WebUI 更好集成。快速入门向导重构，工作区浏览器功能扩展。\n\n✅ **检查点模式（v0.1.68）**：新增检查点协调模式，采用委托者模式——主智能体先单独规划，再通过 `checkpoint()` 工具委派给团队。LLM API 熔断器新增对 429 错误的处理功能。WebUI 支持检查点功能。LiteLLM 供应链问题修复。\n\n✅ **现代化 WebUI（v0.1.67）**：全面改版 WebUI，新增内联最终答案、键盘快捷键和 Zustand 状态管理。引入 RoundBudgetGuardHook 实现每轮成本控制。统一并行预协作阶段。增加回归防护机制。\n\n✅ **步骤模式（v0.1.66）**：新增 `--step` CLI 模式，供外部编排工具使用。支持 massgen-refinery 插件的步骤模式。修复 Codex Windows UTF-8 编码问题，并对控制台文本进行净化处理。\n\n✅ **MassGen Refinery 插件（v0.1.65）**：独立的 MCP 服务器（质量、工作流、媒体），通过 massgen-refinery 插件将 MassGen 的检查清单式评估引入 Claude Code。单智能体精炼已可运行；多智能体仍在实验阶段。\n\n✅ **Gemini CLI 后端（v0.1.64）**：Gemini CLI 成为一流后端，支持会话持久化、MCP 工具和 Docker 部署。OpenAI Response API 支持 WebSocket 流式传输。新增执行轨迹分析子代理。Copilot Docker 模式启用。\n\n✅ **子代理集合与合约（v0.1.63）**：子代理集合模式，默认启用 `disable_injection` 和 `defer_voting_until_all_answered`。推动本轮评估者的转变，并引入成功合约。子代理精简优化，增强抗杀伤能力。\n\n✅ **MassGen 技能与观察器（v0.1.62）**：通用型多智能体技能，具备四种模式（通用、评估、计划、规格），适用于 Claude Code 及其他 AI 代理。提供会话观察器，实时监控运行情况。针对 Claude Code、Codex 和 Copilot 进行后端优化。新增无头模式和 Web 快速入门模式。\n\n✅ **本轮评估者范式（v0.1.61）**：新增一轮评估者子代理类型，每次有新答案时自动启动评估子代理，提供详细反馈作为下一轮输入。大规模重构编排器，优化评估提示词、任务计划注入及子代理相关修复。\n\n✅ **多模态工具、子代理增强及 GPT-5.4（v0.1.60）**：重写 read_media 函数，明确数据结构并引入 MediaCallLedgerHook。子代理增强功能包括 inherit_spawning_agent_backend、final_answer_strategy 和 per-agent subagent_agents。GPT-5.4 成为默认的 OpenAI 旗舰模型。分解模式与检查清单工作流协同。修复 Codex 提示词缓存问题。\n\n✅ **质量回合改进（v0.1.59）**：自动添加任务计划改进内容，优化计划审查流程。评估生成配置更好，检查清单修复，Gemini 工具名称规范化以适应 MCP。调整子代理行为，修复 Docker 技能写入权限问题。调整视频生成技能，并恢复影响指标。\n\n✅ **全面多模态升级（v0.1.58）**：引入 ElevenLabs TTS\u002FSTT、Nano Banana 2 图像生成、Grok 多媒体生成、媒体生成技能以及多轮图像编辑功能。Nvidia NIM 后端上线。重新思考质量子代理。检查清单更智能，包含改进\u002F保留条目。CLI 模式标志位及日志架构重构。\n\n✅ **委托子代理协议与构建者子代理（v0.1.57）**：基于文件的委托协议，用于在容器内启动主机端子代理。新增构建者子代理类型，支持以全新上下文生成大型工件。引入实质性跟踪机制，实现更智能的收敛。更新SDK中的Claude Code推理参数。\n\n✅ **规范计划模式与定向消息传递（v0.1.56）**：通过`plan_mode=\"spec\"`正式定义需求规格，并支持TUI规范模式。利用`target_agents`参数实现代理间的定向消息传递。新增批评者子代理用于质量评估。保持媒体对话连续性，便于后续图像分析。修复Codex OAuth登录问题。\n\n✅ **专业化子代理类型与动态评估标准（v0.1.55）**：通过`SUBAGENT.md`前端元数据定义基于发现的子代理角色（评估员、探索者、研究员、新颖性）。借鉴GEPA理念，设置任务特定的评估标准，包含核心与扩展门控。原生后端图像路由功能。支持可配置的视频帧提取。\n\n✅ **子代理消息传递与Copilot SDK后端（v0.1.54）**：运行时消息可用于引导正在执行的后台子代理。新增GitHub Copilot后端，基于copilot-sdk并原生支持MCP。支持Gemini 3.1 Pro。提供按代理注入的目标定位功能。\n\n✅ **后台工具执行（v0.1.53）**：为长时间运行的任务提供非阻塞的生命周期工具（启动、监控、等待、取消、列出）。规划任务验证要求。TUI提供后台作业指示器和生命周期控制。奠定子代理基础设施基础，包括评估员和探索者类型。\n\n✅ **最终答案模态与协调质量门控（v0.1.52）**：专用的最终答案模态，配备标签页式答案及工作区\u002F评审界面。实质性门控可防止低价值迭代循环。新颖性注入机制对抗过早收敛。代理身份版本化，用于追踪答案来源。\n\n✅ **评审协调与变更文档（v0.1.51）**：具备多文件差异可视化的评审模态。决策日志系统用于多代理协调的可追溯性。基于Changedoc锚定的评估清单，附带差距报告。漂移冲突策略确保更安全的变更应用。新增`--cwd-context` CLI标志。\n\n✅ **分块计划执行与技能生命周期管理（v0.1.50）**：分块计划执行，通过进度检查点更安全地完成长篇任务。技能生命周期管理，支持整合、整理以及加载上一 session 的技能。迭代式计划评审模态。响应式TUI模式栏。工作树改进，支持分支累积及跨代理差异可见性。\n\n✅ **协调质量：日志分析TUI、公平性门控与清单投票（v0.1.49）**：内置日志分析模式于TUI模式栏，便于应用内运行分析。公平性门控防止快速代理主导协调过程。清单投票工具用于结构化质量评估。自动化测试基础设施，配备CI\u002FCD及SVG快照基准。\n\n\n✅ **分解模式与工作树隔离（v0.1.48）**：新增分解协调模式，将任务分解为子任务分配给各代理，并指定一名汇报人；基于Git工作树的隔离机制，用于代理文件写入，并配备评审模态；快速入门向导提供Docker部署，带有动画拉取进度；停止工具用于代理完成信号传递。\n\n✅ **Codex后端与TUI主题重构（v0.1.47）**：新增Codex后端，支持OpenAI Codex CLI的本地及Docker运行；NativeToolMixin用于共享工具处理；TUI主题系统重构为基于调色板的架构，提供深色与浅色两种方案；支持按代理配置投票敏感度。\n\n✅ **子代理TUI流式传输与事件架构重构（v0.1.46）**：交互式预览卡片可展开为完整时间线视图，支持实时事件流；统一事件管道，提供单一事实源用于展示创建；增强最终呈现效果，加入工作区可视化及优胜代理高亮；修复横幅显示及工具调用ID处理问题。\n\n✅ **TUI默认启用与配置迁移（v0.1.45）**：文本终端UI现默认启动，自动从`rich_terminal`迁移到`textual_terminal`；设置向导会生成TUI配置；可通过`--display rich`标志访问旧版Rich显示。\n\n✅ **独立计划选择的执行模式（v0.1.44）**：通过`Shift+Tab`或模式栏在正常→规划→执行之间循环切换；计划选择器可浏览最多10个近期计划及其时间戳；查看完整计划模态，包含全面的任务分解；空提交即可执行计划；规划与执行阶段间保留上下文路径；增强案例研究，提供交互式设置指南；TUI性能优化，采用视口渲染。\n\n✅ **工具调用批处理与交互式案例研究（v0.1.43）**：连续的MCP工具调用被归组为可折叠的树形视图，带有“+N更多”提示和点击展开功能。新增交互式案例研究页面，支持并排SVG对比。`PlanOptionsPopover`用于浏览计划并选择深度。支持带空格的路径引用。优化最终展示及TUI细节。\n\n✅ **TUI视觉重新设计与人工输入队列（v0.1.42）**：采用现代“对话型AI”美学，圆角设计；重新设计代理标签，增加点状指示器；自适应工具卡片；打磨后的模态窗口。新增`HumanInputHook`，可在流程中向代理注入消息，并提供线程安全的代理级追踪。修复AG2单代理协调问题。\n\n✅ **异步子代理执行（v0.1.41）**：通过`async_=True`实现后台子代理执行，支持非阻塞的并行工作；轮询执行状态并获取结果；通过`subagent_round_timeouts`配置进行每轮超时控制；扩展子代理参数，用于超时和并发控制。\n\n✅ **Textual TUI交互模式（v0.1.40）**：交互式终端UI，通过`--display textual`实现实时代理流；提供全面的模态窗口，涵盖成本\u002F投票\u002F工作区\u002F答案等信息；支持使用`@path\u002Fto\u002Ffile`语法注入上下文路径；集成人工反馈，通过提示模态实现。\n\n✅ **计划与执行工作流（v0.1.39）**：完整的先计划后执行工作流，通过`--plan-and-execute`实现自主规划与执行；`--execute-plan`则可直接运行现有计划而无需重新规划；任务验证工作流，配备`verified`状态及验证小组，用于批量校验；计划存储系统位于`.massgen\u002Fplans\u002F`，保存冻结快照并记录执行情况；修复Response API函数调用消息净化问题。\n\n✅ **任务规划与双层工作空间（v0.1.38）**：通过`--plan`标志进入任务规划模式，进行结构化的工作分解（仅规划，不自动执行）；基于Git的双层工作空间，将临时探索与最终交付物分离；自动发现CLAUDE.md\u002FAGENTS.md以确定项目上下文；支持多张图片对比的批量图像分析；设置断路器以防止超时死循环；新增Docker健康监测功能。\n\n✅ **执行轨迹与思维模式 (v0.1.37)**：完整执行历史保存为 `execution_trace.md`，用于压缩恢复和跨智能体协调；集成 Claude Code 和 Gemini 推理内容流式缓冲；所有后端统一智能体标签规范。\n\n✅ **@path 上下文处理与钩子框架 (v0.1.36)**：支持 `@path` 语法及自动补全的内联文件选择器；提供 PreToolUse\u002FPostToolUse 钩子用于权限校验和内容注入；支持全局及单个智能体的钩子注册；内置 `MidStreamInjectionHook` 和 `HighPriorityTaskReminderHook`；兼容 Claude Code 钩子；优化 Docker 资源管理。\n\n✅ **日志分析 CLI 与 Logfire 可观测性 (v0.1.35)**：新增 `massgen logs analyze` 命令，支持提示模式及多智能体自我分析；Logfire 工作流属性用于记录回合上下文和投票推理过程；新增 `direct_mcp_servers` 配置，可将特定 MCP 作为协议工具保留；改进未知工具处理逻辑，并修复仅投票模式相关问题。\n\n✅ **OpenAI 兼容服务器与模型发现 (v0.1.34)**：本地 HTTP 服务器通过 `massgen serve` 命令启动，兼容任何 OpenAI SDK 客户端；通过认证 API 调用实现 Groq 和 Together 后端的动态模型发现；WebUI 支持文件差异对比及答案刷新轮询；优化子智能体状态跟踪与取消恢复功能。\n\n✅ **响应式上下文压缩与流式缓冲 (v0.1.33)**：当上下文长度超出限制时自动进行对话压缩；引入流式缓冲系统，追踪部分响应以实现恢复；在 `write_file` 工具中增加文件覆盖保护机制；防止任务计划重复；修复 Grok MCP 工具可见性问题及 Gemini 仅投票模式问题；优化 GPT-5 模型行为。\n\n✅ **多回合会话导出与每次尝试独立日志记录 (v0.1.32)**：支持按回合范围导出会话内容（`--turns`）；新增工作区导出控制选项（`--no-workspace`、`--workspace-limit`）；将 Logfire 移至可选的 `[observability]` 组件；每次尝试生成独立日志文件并重新配置处理器；自动将 DOCX\u002FPPTX\u002FXLSX 文件转换为 PDF 格式，便于会话分享。\n\n✅ **Logfire 可观测性与 Azure 工具流式传输 (v0.1.31)**：可选集成 Logfire，自动为 OpenAI、Claude 和 Gemini 后端进行 LLM 性能监控；Azure OpenAI 工具调用结果以结构化分块形式输出；新增 `--logfire` CLI 标志及 `MASSGEN_LOGFIRE_ENABLED` 环境变量。\n\n✅ **OpenRouter 网络搜索与角色多样性 (v0.1.30)**：通过 OpenRouter 插件原生支持网络搜索，启用 `enable_web_search` 参数；提供角色多样性模式（`perspective`\u002F`implementation`），可根据阶段动态调整；支持 Azure 多端点自动检测；新增 `${VAR}` 语法用于环境变量扩展。\n\n✅ **子智能体系统与工具指标 (v0.1.29)**：可并行启动多个 MassGen 子进程，每个进程拥有独立的工作空间并自动汇总结果；增强工具指标功能，提供每次调用的平均值以及最小\u002F最大\u002F中位数分布统计；通过 `massgen --quickstart` 命令向各智能体发送系统消息。\n\n✅ **统一多模态工具与产物预览 (v0.1.28)**：整合 `read_media` 工具，用于图像\u002F音频\u002F视频分析；统一 `generate_media` 工具，用于创建图像、视频和音频内容；Web UI 新增产物预览功能，支持 PDF\u002FDOCX\u002FPPTX\u002F图片\u002FHTML\u002FSVG\u002FMarkdown\u002FMermaid 等格式；支持 OpenRouter 工具能力模型筛选；修复 Azure OpenAI 相关问题。\n\n✅ **会话共享与日志分析 (v0.1.27)**：通过 GitHub Gist 实现会话共享，使用 `massgen export` 命令；新增日志分析 CLI 命令 `massgen logs`；提供每 LLM 调用的时间计量指标；支持 Gemini 3 Flash 模型；增强 CLI 配置构建器功能，支持针对不同智能体的网络搜索及系统消息设置。\n\n✅ **WebUI 设置与影子智能体深度 (v0.1.26)**：新增 Docker 诊断模块；推出 WebUI 设置向导，提供引导式首次使用体验；支持影子智能体响应深度调节，以实现测试时计算资源的弹性伸缩；支持 GPT-5.1-Codex 系列模型。\n\n✅ **UI-TARS 与技能进化 (v0.1.25)**：采用字节跳动的 UI-TARS-1.5-7B 模型进行 GUI 自动化操作；支持 GPT-5.2 模型；引入技能进化创建系统，结合会话持久化功能；增强 Textual 终端的自适应布局功能。\n\n✅ **多后端成本追踪 (v0.1.24)**：实时统计 OpenRouter、xAI\u002FGrok、Gemini 和 Claude Code 后端的 token 使用量；通过 `\u002Finspect c` 命令查看成本明细，显示各智能体的 token 使用情况；同时汇总会话总成本，并优化显示格式。\n\n✅ **回合历史检查与 WebUI 自动化 (v0.1.23)**：新增交互式 `\u002Finspect` 命令，可通过菜单导航查看回合详情；引入 `AutomationView` 组件，实现程序化监控；新增 `SessionMountManager`，确保 Docker 容器在各回合间持续挂载；支持基于标志的取消操作，并恢复终端状态；提供 `run_async_safely()` 函数，用于处理嵌套事件循环。\n\n✅ **影子智能体架构 (v0.1.22)**：轻量级影子智能体可并行响应广播消息，不会中断父智能体的工作；通过 `asyncio.gather()` 并行化机制继承完整的对话历史和当前回合上下文。\n\n✅ **优雅取消与会话恢复 (v0.1.21)**：按下 Ctrl+C 可保存协作中的部分进度；被取消的会话可通过 `--continue` 参数恢复，保留智能体的回答和工作空间。\n\n✅ **WebUI 与自动 Docker 设置 (v0.1.20)**：基于浏览器的实时可视化界面，采用 React 前端、WebSocket 流式传输、时间线视图和工作区浏览功能。为计算机使用型智能体自动设置 Docker 容器，预配置 X11 虚拟显示器、xdotool、Firefox、Chromium 和 scrot 工具。\n\n✅ **LiteLLM 集成与 Claude 严格工具使用 (v0.1.19)**：MassGen 作为 LiteLLM 自定义提供商，支持 `run()` 和 `build_config()` 程序化 API；Claude 提供严格工具使用模式，输出结构化结果；Gemini 实现指数退避策略，提升对速率限制的韧性。\n\n✅ **智能体沟通系统 (v0.1.18)**：通过 `ask_others()` 工具实现人类广播式问答，提供三种模式；支持内联回复交付以阻断执行流程；保持会话持久化的问答历史。\n\n✅ **Claude 高级工具功能 (v0.1.18)**：通过 `enable_programmatic_flow` 标志实现程序化工具调用；通过 `enable_tool_search` 标志在服务器端进行工具发现，支持正则表达式或 BM25 变体。\n\n✅ **Textual 终端界面 (v0.1.17)**：使用 Textual 库构建交互式终端 UI，支持深色\u002F浅色主题；提供多面板布局，分别用于智能体和编排器；支持实时流式传输及语法高亮；具备关键模式的内容过滤功能。\n\n✅ **终端评估与成本追踪 (v0.1.16)**：利用 AI 技术自动评估终端显示效果，生成 VHS 录像；集成 LiteLLM，实现对 500 多种模型的精准计价，支持推理及缓存 token 的计算；引入内存归档机制，确保多回合会话的持久化；为 MassGen 发展规划四项自我进化技能。\n\n✅ **角色生成与 Docker 分发 (v0.1.15)**：根据多种策略（互补型、多样化、专业化、对抗型）自动为智能体生成角色，以提升多样性；集成 GitHub Container Registry，支持 ARM 架\n\n✅ **并行工具执行与 Gemini 3 Pro (v0.1.14)**：基于 asyncio 的调度机制，可在所有后端上配置并发工具执行；Gemini 3 Pro 集成函数调用功能；交互式快速入门工作流；用于服务器元数据的 MCP 注册客户端。\n\n✅ **代码化工具与 MCP 注册中心 (v0.1.13)**：实现 CodeAct 范式，通过可导入的 Python 代码集成工具，将令牌使用量减少 98%；MCP 服务器注册中心支持自动发现与按需加载；遵循 TOOL.md 文档标准。\n\n✅ **NLIP 集成与技能系统 (v0.1.13)**：在 Claude、Gemini 和 OpenAI 后端之间，利用自然语言接口协议实现高级工具路由；跨平台自动化技能安装程序，适用于 openskills CLI、Anthropic 技能和 Crawl4AI。\n\n✅ **系统提示架构重构 (v0.1.12)**：采用基于 XML 的格式化方式为 Claude 构建分层系统提示结构，优化大模型注意力管理。\n\n✅ **Semtools 与 Serena 技能 (v0.1.12)**：通过嵌入式相似度实现语义搜索；借助 LSP 集成实现符号级代码理解；支持非 Docker 环境下的本地执行模式。\n\n✅ **多智能体计算机使用 (v0.1.12)**：增强 Gemini 的计算机使用能力，集成 Docker、VNC 可视化，并实现多智能体协作，结合 Claude（Docker\u002FLinux）与 Gemini（浏览器）。\n\n✅ **技能系统 (v0.1.11)**：模块化提示框架，配备 SkillsManager 动态加载技能；自动发现功能，支持“始终”与“可选”类别；新增文件搜索技能；兼容 Docker 挂载。\n\n✅ **内存 MCP 工具与文件系统集成 (v0.1.11)**：MCP 服务器用于内存管理，采用 Markdown 格式存储，划分短期与长期记忆层级；自动保存工作空间状态；集成编排器实现跨智能体内存共享；增强对 Windows 系统的支持，以应对长系统提示。\n\n✅ **速率限制系统 (v0.1.11)**：针对 Gemini 模型实施多维度限制（RPM、TPM、RPD），支持自定义阈值；基于 YAML 的配置；CLI 集成 --enable-rate-limiting 标志；修复 asyncio 锁问题，确保事件循环可重用。\n\n✅ **框架互操作性流式传输 (v0.1.10)**：为 LangGraph 和 SmoLAgent 提供实时中间步骤流式传输，区分日志与输出；增强对外部框架推理步骤的调试功能。\n\n✅ **Docker 配置增强 (v0.1.10)**：支持嵌套认证，提供独立的挂载数组与环境变量数组；通过 Dockerfile.custom-example 支持自定义镜像；自动安装软件包。\n\n✅ **通用工作空间隔离 (v0.1.10)**：扩展实例 ID 生成至所有执行模式，确保安全并行执行；增强工作空间路径的唯一性，避免并发会话冲突。\n\n✅ **会话管理系统 (v0.1.9)**：完整记录与恢复会话状态，采用 SessionState 数据类及 SessionRegistry 实现 CLI 调用间的多轮持久化；保持工作空间连续性，保留智能体状态与各轮之间的协作历史。\n\n✅ **计算机使用工具 (v0.1.9)**：原生集成 Claude 和 Gemini 的计算机使用 API，实现浏览器与桌面自动化，包括截图分析与动作生成；轻量级浏览器自动化适用于特定任务，无需完整计算机使用开销。\n\n✅ **模糊模型匹配 (v0.1.9)**：智能模型名称搜索，支持近似输入（如“sonnet”→“claude-sonnet-4-5-20250929”）；构建模型目录系统，收录各提供商精选列表；增强配置生成器的自动模型搜索功能。\n\n✅ **后端能力扩展 (v0.1.9)**：全面的后端注册表，包含所有提供商的详细规格；支持音频\u002F视频处理、硬件加速；统一访问各类模型家族；优化内存更新逻辑，聚焦于可行动模式。\n\n✅ **LLM 智能体自动化模式 (v0.1.8)**：为在 LLM 智能体中运行 MassGen 构建完整基础设施，采用 SilentDisplay 类实现极简输出（约 10 行，而非 250–3,000+ 行）；实时监控 status.json 文件，每 2 秒更新一次；定义有意义的退出码（0=成功，1=配置错误，2=执行错误，3=超时，4=中断）；自动隔离工作空间以支持并行执行；具备元协调能力，允许 MassGen 运行 MassGen。\n\n✅ **DSPy 问句改写集成 (v0.1.8)**：为多智能体协作提供智能化问句多样性，搭载语义保留型改写模块，支持三种策略（多样\u002F平衡\u002F保守）；自动进行语义验证以确保意义不丢失；采用线程安全的缓存系统，基于 SHA-256 哈希；作为改写引擎支持所有后端；集成编排器实现问句变体的自动分发。\n\n✅ **智能体任务规划系统 (v0.1.7)**：基于 MCP 的规划服务器，管理任务生命周期；跟踪依赖关系并自动验证与阻塞；支持待处理\u002F进行中\u002F已完成\u002F已阻塞等状态转换；集成编排器实现计划感知的多智能体协作。\n\n✅ **后台 Shell 执行 (v0.1.7)**：为长时间运行的命令提供持久化 Shell 会话，BackgroundShell 类支持异步执行、实时输出流式传输与监控、自动超时处理；增强代码执行服务器的后台能力。\n\n✅ **抢占式协调 (v0.1.7)**：智能体可在不完全重启的情况下中断正在进行的协调，提交更优答案；抢占过程中部分进度得以保留；增强协调跟踪器，记录抢占事件。\n\n✅ **框架互操作性 (v0.1.6)**：AG2 嵌套聊天、LangGraph 工作流、AgentScope 智能体、OpenAI 助手以及 SmoLAgent 均被整合为自定义工具，支持跨框架协作与 AG2 的流式传输。\n\n✅ **配置校验器 (v0.1.6)**：采用 ConfigValidator 类进行全面的 YAML 校验，集成 pre-commit 并提供包含可操作建议的详细错误信息。\n\n✅ **统一工具执行 (v0.1.6)**：ToolExecutionConfig 数据类标准化 ResponseBackend、ChatCompletionsBackend 和 ClaudeBackend 中的工具处理流程，确保一致的错误报告。\n\n✅ **Gemini 后端简化 (v0.1.6)**：移除 gemini_mcp_manager 和 gemini_trackers 模块，整合代码后代码库减少 1,598 行。\n\n✅ **内存系统 (v0.1.5)**：通过 mem0 集成实现长期语义记忆，支持跨会话的事实提取与检索；提供短期对话记忆以维持活跃上下文；接近令牌限制时自动压缩上下文；支持跨智能体内存共享，按轮次过滤；通过会话管理实现内存隔离与延续；集成 Qdrant 向量数据库以支持语义搜索。\n\n✅ **多模态生成工具 (v0.1.4)**：通过 DALL-E API 根据文本创建图像；根据描述生成视频；支持文本转语音及音频转录；可生成 PDF\u002FDOCX\u002FXLSX\u002FPPTX 格式的文档；具备对现有图像进行变换的能力。\n\n✅ **二进制文件保护（v0.1.4）**：自动拦截功能可防止文本工具访问40多种二进制文件类型，包括图片、视频、音频、压缩包及Office文档；智能错误提示会引导用户使用适合处理二进制内容的专业工具。\n\n✅ **Crawl4AI集成（v0.1.4）**：基于大语言模型的内容提取与智能网页爬取功能，支持自定义提取模式，用于从网站中获取结构化数据。\n\n✅ **后评估工作流（v0.1.3）**：获胜的智能体在提交答案前会自行评估，具备提交与重启功能；支持答案确认，并可在所有后端系统中根据反馈重新启动协调流程。\n\n✅ **多模态理解工具（v0.1.3）**：可分析图像、转录音频、提取视频帧并处理文档（PDF\u002FDOCX\u002FXLSX\u002FPPTX），输出结构化的JSON数据；通过OpenAI GPT-4.1集成，在所有后端系统中均可使用。\n\n✅ **Docker Sudo模式（v0.1.3）**：在Docker容器中以特权模式执行命令，适用于需要高权限的系统级操作。\n\n✅ **智能规划模式（v0.1.2）**：自动分析问题，通过orchestrator中的`_analyze_question_irreversibility()`方法判断操作是否不可逆；可通过`set_planning_mode_blocked_tools()`和`is_mcp_tool_blocked()`方法选择性阻止工具使用；在协调过程中，MCP仅允许只读操作，禁止写入；零配置即可透明运行，支持多工作空间。\n\n✅ **模型更新（v0.1.2）**：新增Claude 4.5 Haiku模型`claude-haiku-4-5-20251001`；重新调整Claude模型优先级，将`claude-sonnet-4-5-20250929`设为默认模型；修复Grok网络搜索问题，通过`_add_grok_search_params()`方法正确处理`extra_body`参数。\n\n✅ **自定义工具系统（v0.1.1）**：用户可通过`massgen\u002Ftool\u002F_manager.py`中的`ToolManager`类注册自定义Python函数；支持跨后端与MCP服务器协同工作，内置\u002FMCP\u002F自定义工具分类并实现自动发现；`massgen\u002Fconfigs\u002Ftools\u002Fcustom_tools\u002F`目录下提供40余种示例；通过三档质量体系（宽松\u002F均衡\u002F严格）控制投票敏感度，检测答案新颖性以避免重复。\n\n✅ **后端增强（v0.1.1）**：重构Gemini架构，将MCP管理（`gemini_mcp_manager.py`）、追踪（`gemini_trackers.py`）及工具类提取出来；新增`massgen\u002Fbackend\u002Fcapabilities.py`能力注册表，记录各后端的功能支持情况。\n\n✅ **PyPI软件包发布（v0.1.0）**：通过`pip install massgen`正式发布，安装更简便；全局`massgen`命令可在任意目录下使用；提供详尽的Sphinx文档，网址为[docs.massgen.ai](https:\u002F\u002Fdocs.massgen.ai\u002F)；配备交互式设置向导，包含用例预设和API密钥管理；增强CLI功能，内置配置以`@examples\u002F`为前缀。\n\n✅ **Docker执行模式（v0.0.32）**：基于容器的隔离机制，确保安全地在独立Docker容器中执行命令，防止访问宿主机文件系统；持久化状态管理，使包和依赖项在多次对话轮次间保持不变；支持多智能体协作，每个智能体拥有独立的隔离容器；可配置资源限制（CPU、内存）、网络隔离模式以及只读挂载卷，以提升安全性。\n\n✅ **MCP架构重构（v0.0.32）**：简化客户端，将`MultiMCPClient`更名为`MCPClient`，反映架构精简；移除已废弃模块并整合重复的MCP协议处理代码，提高可维护性；通过标准化类型注解、改进错误处理及更清晰的代码组织，优化代码结构。\n\n✅ **Claude Code Docker集成（v0.0.32）**：自动工具管理功能，Docker模式下Bash工具将被禁用，所有命令均通过execute_command路由执行；MCP工具自动获得权限批准，同时保留安全验证机制；通过系统消息强化指导，避免混淆宿主机与容器环境中的Git仓库。\n\n✅ **通用命令执行（v0.0.31）**：基于MCP的execute_command工具适用于Claude、Gemini、OpenAI及Chat Completions等提供商；全面的安全措施包括权限管理和命令过滤；在规划模式下执行代码，以实现更安全的协调。\n\n✅ **外部框架集成（v0.0.31）**：利用外部框架的群聊模式进行多智能体对话，由LLM驱动实现智能扬声器选择（自动、轮询、手动）；增强适配器功能，支持原生群聊协调。\n\n✅ **音视频生成（v0.0.31）**：提供文本转语音及转录的音频工具，同时利用OpenAI的Sora-2 API生成视频，进一步拓展多模态能力，超越文本与图像范畴。\n\n✅ **多模态支持扩展（v0.0.30）**：为Chat Completions和Claude后端增加音视频处理功能（支持WAV、MP3、MP4、AVI、MOV、WEBM格式）；灵活支持本地路径或URL作为媒体输入；扩展音频\u002F视频文件的Base64编码；可配置文件大小限制。\n\n✅ **Claude Agent SDK迁移（v0.0.30）**：将软件包从`claude-code-sdk`迁移到`claude-agent-sdk>=0.0.22`；改进Bash工具的权限验证；增强系统消息处理能力。\n\n✅ **Qwen API集成（v0.0.30）**：将Qwen API提供商加入Chat Completions生态系统，支持`QWEN_API_KEY`；提供视频理解配置示例。\n\n✅ **MCP规划模式（v0.0.29）**：一种更安全的MCP工具使用协调策略；支持多后端（Response API、Chat Completions、Gemini）；智能体在协调期间仅规划而不执行；提供5种规划模式配置。\n\n✅ **文件操作安全（v0.0.29）**：通过`FileOperationTracker`类强制执行“先读后删”原则；集成`PathPermissionManager`，提供操作跟踪方法；进一步强化文件操作安全机制。\n\n✅ **外部框架集成（v0.0.28）**：建立外部智能体框架适配器系统，支持异步执行；可在多种环境中（本地、Docker、Jupyter、YepCode）执行代码；提供开箱即用的框架集成配置。\n\n✅ **多模态支持——图像处理（v0.0.27）**：新增`stream_chunk`模块，用于多模态内容处理；具备图像生成与理解能力；支持文件上传及文档问答检索；Claude Sonnet 4.5版本支持；增强工作空间内的多模态工具。\n\n✅ **文件删除与工作空间管理（v0.0.26）**：新增MCP工具（`delete_file`、`delete_files_batch`、`compare_directories`、`compare_files`），用于清理工作空间和文件比较；整合`_workspace_tools_server.py`，增强路径权限管理器。\n\n✅ **受保护路径与基于文件的上下文路径（v0.0.26）**：可在允许写入的目录中保护特定文件，仅授予对单个文件的访问权限，而非整个目录。\n\n✅ **多轮文件系统支持（v0.0.25）**：支持多轮对话，跨轮次保持上下文；自动创建 `.massgen` 目录结构；支持工作区快照与恢复；增强路径权限系统，加入智能排除功能；以及全面的后端改进。\n\n✅ **SGLang 后端集成（v0.0.25）**：统一 vLLM\u002FSGLang 后端架构，具备自动检测功能；支持 SGLang 特有的参数如 `separate_reasoning`；并提供双服务器支持，适用于 vLLM 和 SGLang 混合部署场景。\n\n✅ **vLLM 后端支持（v0.0.24）**：与 vLLM 完全集成，实现高性能本地模型服务；支持 POE 提供者；可识别 GPT-5-Codex 模型；重构后端工具模块；并修复包括流式分块处理在内的多项 bug。\n\n✅ **后端架构重构（v0.0.23）**：大幅整合代码，新增 `base_with_mcp.py` 类，减少各后端总计约 1,932 行代码；提取格式化模块以提升代码组织性；通过统一的 MCP 集成提高可维护性。\n\n✅ **通过 MCP 的工作区复制工具（v0.0.22）**：实现工作区间无缝文件复制；采用层级结构组织配置；并增强面向大规模协作的文件操作功能。\n\n✅ **Grok MCP 集成（v0.0.21）**：统一后端架构，全面支持 MCP 服务器；通过 MCP 服务器实现文件系统功能；并优化配置文件。\n\n✅ **Claude 后端的 MCP 支持（v0.0.20）**：将 MCP 集成扩展至 Claude 后端；全面支持 MCP 协议及文件系统；强化错误处理机制；并完善文档说明。\n\n✅ **全面协调跟踪（v0.0.19）**：构建完整的协调跟踪与可视化系统，基于事件的跟踪机制、交互式协调表格展示，以及针对多智能体协作模式的高级调试能力。\n\n✅ **全面 MCP 集成（v0.0.18）**：将 MCP 扩展至所有 Chat Completions 后端（Cerebras AI、Together AI、Fireworks AI、Groq、Nebius AI Studio、OpenRouter）；实现跨提供商的函数调用兼容性；新增 9 个 MCP 配置示例。\n\n✅ **OpenAI 的 MCP 集成（v0.0.17）**：将 MCP（模型上下文协议）支持扩展至 OpenAI 后端，实现 GPT 模型的完整工具发现与执行能力；统一多后端的 MCP 架构；并加强调试功能。\n\n✅ **结合 MCP 集成的统一文件系统支持（v0.0.16）**：完成 `FilesystemManager` 类，为 Gemini 和 Claude Code 后端提供统一的文件系统访问；基于 MCP 进行文件操作和跨智能体协作。\n\n✅ **MCP 集成框架（v0.0.15）**：为 Gemini 后端实现完整的 MCP 实现，支持多服务器部署、断路器模式，并构建全面的安全框架。\n\n✅ **增强日志记录（v0.0.14）**：改进日志系统，便于调试智能体的回答；新增最终答案目录结构；并提供详细的架构文档。\n\n✅ **统一日志系统（v0.0.13）**：建立集中式日志基础设施，配备调试模式和更完善的终端显示格式。\n\n✅ **Windows 平台支持（v0.0.13）**：实现 Windows 平台兼容性，优化路径处理与进程管理。\n\n✅ **增强 Claude Code 智能体上下文共享（v0.0.12）**：Claude Code 智能体现在可通过在编排器端维护快照和临时工作区来共享工作区上下文。\n\n✅ **文档改进（v0.0.12）**：更新 README 文件，包含当前功能并改进安装说明。\n\n✅ **自定义系统消息（v0.0.11）**：增强系统消息的配置与保存功能，支持后端特定的系统提示词定制。\n\n✅ **Claude Code 后端增强（v0.0.11）**：改进集成，优化系统消息处理、JSON 响应解析及协调动作描述。\n\n✅ **Azure OpenAI 支持（v0.0.10）**：集成 Azure OpenAI 服务，包括 GPT-4.1 和 GPT-5-chat 模型，并支持异步流式传输。\n\n✅ **MCP（模型上下文协议）支持（v0.0.9）**：集成 MCP，为 Claude Code 智能体提供高级工具能力，包括 Discord 和 Twitter 集成。\n\n✅ **超时管理系统（v0.0.8）**：引入编排器级别的超时机制，具备优雅降级和更丰富的错误信息提示。\n\n✅ **本地模型支持（v0.0.7）**：完全集成 LM Studio，可在本地运行开放权重模型，并实现自动服务器管理。\n\n✅ **GPT-5 系列集成（v0.0.6）**：支持 OpenAI 的 GPT-5、GPT-5-mini 和 GPT-5-nano 模型，并提供高级推理参数。\n\n✅ **Claude Code 集成（v0.0.5）**：原生 Claude Code 后端，具备流式传输能力和工具支持。\n\n✅ **GLM-4.5 模型支持（v0.0.4）**：集成智谱 AI 的 GLM-4.5 系列模型。\n\n✅ **基础架构（v0.0.3）**：构建完整的多智能体编排系统，支持异步流式传输、内置工具及多后端支持。\n\n✅ **扩展的提供商生态**：支持超过 15 家提供商，包括 Cerebras AI、Together AI、Fireworks AI、Groq、Nebius AI Studio 和 OpenRouter。\n\n\n\n### 未来重点改进方向\n\n- **Bug 修复与后端优化**：修复图像生成路径问题，并增加对 Claude 多模态的支持。\n- **高级智能体协作**：探索更高效的通信模式和共识构建协议，以提升智能体间的协同效应。\n- **扩展模型集成**：增加对更多前沿模型和本地推理引擎的支持。\n- **性能与 scalability 优化**：优化流式传输和日志记录机制，提升性能与资源管理效率。\n- **开发者体验提升**：完善工具注册系统和 Web 界面，以提供更好的可视化效果。\n\n我们欢迎社区贡献，共同实现这些目标。\n\n### v0.1.74 路线图\n\n版本 0.1.74 专注于云端执行：\n\n#### 计划功能\n- **Cloud Modal MVP**（[#982](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fissues\u002F982)）：将 MassGen 作为云任务在 Modal 上运行——进度流会显示在终端上，结果会保存在本地的 `.massgen\u002Fcloud_jobs\u002F` 目录下。\n\n---\n\n## 🤝 贡献方式\n\n我们热烈欢迎各位的贡献！详情请参阅我们的 [贡献指南](CONTRIBUTING.md)。\n\n---\n\n## 🤝 致谢\n\n我们感谢 AgentWeb\n\n\u003Ca href=\"https:\u002F\u002Fwww.agentweb.pro\u002F\">\n \u003Cimg width=\"196\" height=\"51\" alt=\"68dacef628cd7a44dfb97814_agentweb-logo\" src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Fmassgen_MassGen_readme_4a2ebd177b17.png\" \u002F>\n\u003C\u002Fa>\n\n提供的慷慨赞助。\n\n---\n\n## 📄 许可证\n\n本项目采用 Apache License 2.0 许可证——详细信息请参阅 [LICENSE](LICENSE) 文件。\n\n---\n\n\u003Cdiv align=\"center\">\n\n**⭐ 如果你觉得这个项目有用，请给它点个 Star 吧！⭐**\n\n由 MassGen 团队用心打造 ❤️\n\n\u003C\u002Fdiv>\n\n## ⭐ Star 历史\n\n[![Star History Chart](https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Fmassgen_MassGen_readme_fb6c57c3716f.png)](https:\u002F\u002Fwww.star-history.com\u002F#Leezekun\u002FMassGen&Date)","# MassGen 快速上手指南\n\nMassGen 是一个前沿的多智能体协作框架，通过并行处理、冗余计算和迭代优化，协调多个 AI 智能体共同解决复杂任务。系统利用不同模型的优势，通过共识机制输出高质量结果。\n\n## 环境准备\n\n在开始之前，请确保您的开发环境满足以下要求：\n\n* **操作系统**：Linux, macOS 或 Windows (WSL2 推荐)\n* **Python 版本**：Python 3.11 或更高版本\n* **依赖工具**：\n * `pip` 或 `uv` (推荐使用 `uv` 以获得更快的安装和运行速度)\n * Docker (可选，用于安全的代码执行环境)\n* **API Keys**：需准备至少一个主流大模型服务的 API Key (如 OpenAI, Anthropic, Google Gemini, xAI Grok 等)\n\n## 安装步骤\n\n推荐使用 `uv` 进行环境管理和安装，也可使用标准的 `pip`。\n\n### 方法一：使用 uv 安装（推荐）\n\n```bash\n# 1. 安装 uv (如果尚未安装)\npip install uv\n\n# 2. 创建虚拟环境并激活\nuv venv && source .venv\u002Fbin\u002Factivate\n\n# 3. 安装 MassGen\nuv pip install massgen\n```\n\n### 方法二：使用 pip 安装\n\n```bash\npip install massgen\n```\n\n### 初始化配置\n\n安装完成后，运行以下命令进行快速设置。该向导将帮助您配置 API Keys、设置 Docker 环境以及安装必要的技能包（Skills）。\n\n```bash\nuv run massgen --setup\n```\n\n执行 `--setup` 后，系统将引导您：\n1. 输入各大模型厂商的 API Keys。\n2. 选择是否启用 Docker 进行代码沙箱执行。\n3. 选择并安装预置技能包（如 `openskills`, `Crawl4AI` 等）。\n\n## 基本使用\n\nMassGen 提供了多种运行模式，从单智能体测试到多智能体协作。\n\n### 1. 快速启动（推荐新手）\n\n使用 `--quickstart` 命令，系统将交互式地询问您需要的智能体数量、后端模型选择及推理强度，并自动启动交互式终端界面 (TUI)。\n\n```bash\nuv run massgen --quickstart\n```\n\n**交互流程说明：**\n* 输入智能体数量（1-5，默认为 3）。\n* 为每个智能体选择模型后端。\n* 若使用 GPT-5x 系列模型，需设置 `reasoning.effort` (low|medium|high|xhigh)。\n* 系统自动检测 Docker 状态并配置执行模式。\n* 启动后进入实时可视化的 TUI 界面，观察智能体协作过程。\n\n### 2. 单智能体模式（最简单示例）\n\n如果您只想测试单个模型的能力，可以直接传入任务描述：\n\n```bash\nuv run massgen \"用 Python 写一个简单的爬虫脚本\"\n```\n\n### 3. 多智能体协作模式\n\n指定配置文件或直接通过命令行参数启动多个智能体协同工作。以下示例展示如何调用不同模型共同完成任务：\n\n```bash\n# 使用默认配置启动多智能体协作\nuv run massgen --config @examples\u002Ffeatures\u002Ftrace_analyzer_background.yaml \"Create an svg of an AI agent coding.\"\n```\n\n或者在交互模式下，系统会自动调度多个智能体并行处理任务，它们会互相观察、批评并基于彼此的工作进行迭代，最终通过投票机制选出最佳方案。\n\n### 4. 在 AI 编程助手间调用\n\n如果您使用 Claude Code, Cursor, Copilot 等支持 Agent Skills 的工具，可以先安装 MassGen 技能包：\n\n```bash\nnpx skills add massgen\u002Fskills --all\n```\n\n安装后，直接在您的编程助手中输入 `\u002Fmassgen` (Claude Code) 或 `$massgen` (Codex) 即可触发多智能体工作流。\n\n### 查看结果\n\n* **实时显示**：默认启动 Textual TUI 界面，可实时查看时间线、智能体卡片状态及投票追踪。\n* **日志记录**：所有协作过程、中间产物及最终结论均会被详细记录，可在运行结束后的输出目录中查看。","某金融科技团队需要在极短时间内，基于最新监管文件重构核心风控系统的代码逻辑，并确保零逻辑漏洞。\n\n### 没有 MassGen 时\n- **单模型局限明显**：资深工程师依赖单个大模型生成代码，常因模型“幻觉”遗漏边缘情况，导致风控规则覆盖不全。\n- **人工审查耗时**：团队成员需轮流手动审查每一段生成代码，反复进行多轮对话修正错误，开发周期被拉长数倍。\n- **缺乏一致性验证**：不同成员对监管条款理解存在偏差，生成的代码逻辑前后矛盾，难以在部署前发现深层逻辑冲突。\n- **迭代效率低下**：遇到复杂逻辑死胡同时，只能人工重启会话或更换提示词，无法自动利用之前的失败经验进行优化。\n\n### 使用 MassGen 后\n- **多智能体协同纠错**：MassGen 自动调度多个异构 Agent 并行处理同一任务，它们互相观察、批判并构建彼此的代码方案，自动填补单一模型的盲区。\n- **共识机制保障质量**：系统通过多轮迭代 refinement，只有当多个 Agent 对某段风控逻辑投票达成共识时才输出结果，显著降低逻辑漏洞率。\n- **自动化深度推理**：面对复杂监管条款，Agent 们自主进行“思维链”推演和重启尝试，无需人工干预即可突破逻辑死角，输出高鲁棒性代码。\n- **终端可视化监控**：工程师在终端即可实时看到多个 Agent 的协作辩论过程和收敛路径，将原本数天的审查工作压缩至小时级。\n\nMassGen 通过将“单人独斗”升级为“群体智慧共识”，在确保金融级代码严谨性的同时，实现了研发效率的倍数级跃升。","https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Fmassgen_MassGen_07732971.jpg","massgen","https:\u002F\u002Foss.gittoolsai.com\u002Favatars\u002Fmassgen_d6e0b2ec.png",null,"https:\u002F\u002Fgithub.com\u002Fmassgen",[77,81,85,89,93,97,100,104],{"name":78,"color":79,"percentage":80},"Python","#3572A5",88,{"name":82,"color":83,"percentage":84},"TypeScript","#3178c6",8.7,{"name":86,"color":87,"percentage":88},"HTML","#e34c26",2.8,{"name":90,"color":91,"percentage":92},"Shell","#89e051",0.3,{"name":94,"color":95,"percentage":96},"CSS","#663399",0.1,{"name":98,"color":99,"percentage":96},"Dockerfile","#384d54",{"name":101,"color":102,"percentage":103},"JavaScript","#f1e05a",0,{"name":105,"color":106,"percentage":103},"Makefile","#427819",928,146,"2026-04-07T07:13:10","NOASSERTION","Linux, macOS, Windows","未说明 (主要基于 API 调用云端模型，本地仅需运行协调逻辑)","未说明",{"notes":115,"python":116,"dependencies":117},"该工具是一个多智能体协调框架，主要通过 API 调用外部大模型（如 OpenAI, Anthropic, Google, xAI 等），因此对本地 GPU 和内存无特殊高要求。建议使用 'uv' 进行环境管理和安装以获得更快速度。支持通过 Docker 隔离代码执行环境。可通过 '--setup' 命令自动配置 API 密钥和可选的 Docker 环境。","3.11+",[72,118,119,120],"uv","textual (TUI 界面)","docker (可选，用于代码执行沙箱)",[52,14,35,13],[123,124,125,126,127,128,129,130,131,132,133,134,135,136,137,138],"agent","llm","multi-agent","test-time-scaling","agentic-ai","autonomous-agents","cli","collaborative-ai","conversational-ai","genai","generative-ai","llm-orchestration","model-context-protocol","python","terminal-ui","tool-calling","2026-03-27T02:49:30.150509","2026-04-09T01:24:38.390004",[],[143,148,153,158,163,168,173,178,183,188,193,198,203,208,213,218,223,228,233,238],{"id":144,"version":145,"summary_zh":146,"released_at":147},162742,"v0.1.73","# 🚀 发布亮点 — v0.1.73 (2026-04-06)\n\n### 🧬 [评估标准进化子代理](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Fconcepts.html)\n- **可自我改进的评估标准** ([#1047](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1047))：一种新型子代理，能够在多轮迭代中不断进化评估标准——随着运行的推进，评估标准会变得更加敏锐、更具倾向性。\n\n### 🛡️ [检查点目标模式](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fblob\u002Fmain\u002Fdocs\u002Fmodules\u002Fcheckpoint.md)\n- **针对不可逆操作的安全规划** ([#1047](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1047))：初步设计了带有 `objective` 模式的检查点 MCP——在执行删除、部署、财务操作等不可逆任务之前，先制定安全的计划。\n- **结构化计划输出**：返回包含每一步约束条件及递归恢复树的有序计划。\n\n### 👁️ 评估标准可见性提升\n- **更清晰的评估标准显示**：能够更清楚地查看代理正在依据哪些评估标准进行工作。\n\n---\n\n### 📖 入门指南\n- [**快速入门指南**](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen?tab=readme-ov-file#1--installation)\n- **立即体验**：\n ```bash\n pip install massgen==0.1.73\n uv run massgen --config @examples\u002Ffeatures\u002Ftrace_analyzer_background.yaml \"创建一个AI代理编码过程的SVG图。\"\n ```\n\n## 变更内容\n* 功能：评估标准进化子代理；由 @ncrispino 在 https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1047 中提交的初始检查点 MCP 草案。\n* 文档：由 @Henry-811 在 https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1048 中编写的 v0.1.73 版本文档。\n* 功能：v0.1.73 版本由 @Henry-811 在 https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1046 中完成。\n\n\n**完整变更日志**：https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fcompare\u002Fv0.1.72...v0.1.73","2026-04-06T17:35:07",{"id":149,"version":150,"summary_zh":151,"released_at":152},162743,"v0.1.72","# 🚀 发布亮点 — v0.1.72 (2026-04-03)\n\n### 🦎 [Grok 后端更新](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Fbackends.html)\n- **后端改进** ([#1044](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1044)): 更新了 Grok 后端，引入最新改进\n\n### ⚡ [熔断器第二阶段](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Fbackends.html)\n- **扩展至所有主要后端** ([#1038](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1038)): LLM API 熔断器现已覆盖 ChatCompletions、Response API 和 Gemini（此前在 v0.1.68 中仅支持 Claude）\n- **Gemini 503 错误处理**: Gemini 后端的熔断器也会在遇到 503 错误时触发\n- **配置连通性冒烟测试**: 验证所有后端的熔断器配置是否正确\n\n---\n\n### 📖 入门指南\n- [**快速入门指南**](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen?tab=readme-ov-file#1--installation)\n- **立即试用**:\n ```bash\n pip install massgen==0.1.72\n uv run massgen --config @examples\u002Fproviders\u002Fothers\u002Fgrok_x_search.yaml \"研究过去一周关于 AI 代理的最新帖子和新闻，并总结关键趋势与见解。\"\n ```\n\n## 变更内容\n* 功能：将 LLM 熔断器扩展至 ChatCompletions、Response API 和 Gemini（第二阶段），由 @amabito 在 https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1038 中实现\n* 功能：Grok 后端更新，由 @ncrispino 在 https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1044 中完成\n* 文档：v0.1.72 版本文档，由 @Henry-811 在 https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1045 中编写\n* 功能：v0.1.72 版本发布，由 @Henry-811 在 https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1043 中完成\n\n\n**完整变更日志**：https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fcompare\u002Fv0.1.71...v0.1.72","2026-04-03T17:43:23",{"id":154,"version":155,"summary_zh":156,"released_at":157},162744,"v0.1.71","# 🚀 发布亮点 — v0.1.71（2026-04-01）\n\n### 🔍 [跟踪分析子代理](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Flogging.html)\n- **后台执行跟踪分析**：每轮结束后，跟踪分析子代理会自动在后台启动，分析上一轮的执行跟踪，并将洞察写入内存，以便下一轮继续使用。\n- **修复跟踪内存问题**：修正了执行跟踪中的内存处理逻辑。\n- **修复执行跟踪分析器启动问题**：现在跟踪分析器能够正确启动。\n\n### 📋 [更完善的评估标准](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Fconcepts.html)\n- **改进标准生成**：生成的标准质量更高，更具主观性。\n\n### 🧠 [系统提示词调优](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Fconcepts.html)\n- **提升智能体性能**：调整了系统提示词，以改善各轮之间的协作效果。\n\n### 🔧 修复\n- **修复最终注入问题**：修正了最后阶段的注入行为。\n- **修复预协作阶段GPT模型的评估标准问题**：解决了与GPT模型在预协作阶段相关的评估标准问题。\n- **修复内存的自动轮次处理**：修正了内存方面的自动轮次处理逻辑。\n\n---\n\n### 📖 入门指南\n- [**快速入门指南**](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen?tab=readme-ov-file#1--installation)\n- **试用方法**：\n ```bash\n pip install massgen==0.1.71\n uv run massgen --config @examples\u002Ffeatures\u002Ftrace_analyzer_background.yaml \"创建一个AI智能体编程的SVG图像。\"\n ```\n\n## 变更内容\n* feat: 添加后台执行跟踪子代理，由@ncrispino在 https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1041 中实现。\n* docs: v0.1.71版本的文档，由@Henry-811在 https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1042 中完成。\n* feat: v0.1.71版本，由@Henry-811在 https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1040 中发布。\n\n\n**完整变更日志**：https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fcompare\u002Fv0.1.70...v0.1.71","2026-04-01T17:20:21",{"id":159,"version":160,"summary_zh":161,"released_at":162},162745,"v0.1.70","# 🚀 发布亮点 — v0.1.70 (2026-03-30)\n\n### 📋 [评估标准重新设计](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Fconcepts.html)\n- **三级分类**：`primary`（ONE — 模型最需要改进的地方）、`standard`（必须通过）、`stretch`（锦上添花），并为每个标准定义反模式\n- **愿景陈述**：用一句话描述理想结果，明确质量目标\n- **改进的标准生成**：现在生成的标准更具主观性，且更贴合具体任务\n\n### 🔄 [改进的检查清单式评估](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Fconcepts.html)\n- **更紧凑的迭代提交周期**（[#1035](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1035)）：通过改进评分、差距分析和改进建议，在最终投票前推动更有意义的迭代\n\n### ✨ 其他\n- **快速迭代模式**（[#1035](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1035)）：通过 `fast_iteration.yaml` 简化多轮提交流程\n- **WebUI 审核模态框**（[#1035](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1035)）：在使用 Git 工作时，可直接在浏览器中批准和评论输出结果\n- **后台追踪分析**（[#1035](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1035)）：从第二轮开始自动启动执行轨迹分析器\n- **工作空间清理**（[#1035](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1035)）：增强了各轮之间的隔离性\n\n---\n\n### 📖 入门指南\n- [**快速入门指南**](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen?tab=readme-ov-file#1--installation)\n- **试用**：\n ```bash\n pip install massgen==0.1.70\n # 尝试使用重新设计的评估标准进行快速迭代\n uv run massgen --config @examples\u002Ffeatures\u002Ffast_iteration.yaml \"创建一个 AI 助手编程的 SVG 图像。\"\n ```\n\n## 变更内容\n* 功能：改进评估标准及改进步骤的清晰度，由 @ncrispino 在 [#1035](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1035) 中实现\n* 文档：v0.1.70 的文档，由 @Henry-811 在 [#1037](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1037) 中完成\n* 功能：v0.1.70 版本，由 @Henry-811 在 [#1036](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1036) 中发布\n\n\n**完整变更日志**：https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fcompare\u002Fv0.1.69...v0.1.70","2026-03-30T17:10:24",{"id":164,"version":165,"summary_zh":166,"released_at":167},162746,"v0.1.69","# 🚀 发布亮点 — v0.1.69（2026-03-27）\n\n### 🌐 [WebUI 自动化自动启动](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Fwebui.html)\n- **无需浏览器交互即可自动启动**：`massgen --web --automation --config config.yaml \"您的问题\"` 会立即开始运行——您可以在任何时候打开 URL 来监控运行进度\n- **自动解析配置文件**：在未指定配置文件时，自动化模式会自动解析配置\n- **任务完成后自动结束**：Web 自动化会在技能执行完毕时正确地自动结束\n\n### 🤖 [MassGen 技能重新设计](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Fskills.html)\n- **提升易用性与 WebUI 集成**（[#1032](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1032)）：MassGen 技能现在会启动 WebUI，以便实时跟踪和监控会话\n\n### ✨ 其他\n- **快速入门向导重构**（[#1032](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1032)）：新增欢迎、技能、API 密钥、Docker 和设置模式等步骤，使上手更加顺畅\n- **工作区浏览器扩展**（[#1032](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1032)）：WorkspaceModal 及改进的工作区连接功能\n- **灵活的评判标准字段**（[#1032](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1032)）：在评估标准 JSON 中，`description` 或 `name` 可作为 `text` 的替代选项\n\n---\n\n### 📖 开始使用\n- [**快速入门指南**](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen?tab=readme-ov-file#1--installation)\n- **试一试**：\n ```bash\n pip install massgen==0.1.69\n # 自动启动一次运行，并在 WebUI 中监控\n uv run massgen --web --automation --config config.yaml \"您的问题\"\n ```\n\n## 变更内容\n* 功能：WebUI 技能改进，由 @ncrispino 在 https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1032 中完成\n* 文档：v0.1.69 版本文档，由 @Henry-811 在 https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1033 中完成\n* 功能：v0.1.69 版本，由 @Henry-811 在 https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1031 中完成\n\n\n**完整变更日志**：https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fcompare\u002Fv0.1.68...v0.1.69","2026-03-27T18:01:34",{"id":169,"version":170,"summary_zh":171,"released_at":172},162747,"v0.1.68","# 🚀 发布亮点 — v0.1.68（2026-03-25）\n\n### 🔀 [检查点模式](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fblob\u002Fmain\u002Fdocs\u002Fmodules\u002Fcheckpoint.md)\n- **委托模式**：主智能体先以单人模式执行，随后调用 `checkpoint()` 将任务交由完整的多智能体团队协作完成。\n- **全新智能体实例**：为协作执行提供干净的后端环境和克隆的工作空间。\n- **无缝交接**：团队达成共识后，主智能体将继续执行，并将结果及交付文件复制到其工作空间中。\n- **WebUI 支持**：检查点模式的显示已集成到现代化的 WebUI 中。\n\n### ⚡ LLM API 熔断机制\n- **429 速率限制处理**（[#1024](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1024)）：针对 Claude 后端实现自动熔断模式——检测到 429 速率限制时，系统会自动退避并优雅地恢复。\n\n### ✅ 修复\n- **LiteLLM 供应链修复**（[#1025](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1025)）：锁定 litellm 版本至 1.82.6，并提交 uv.lock 文件，以防止依赖性攻击。\n\n---\n\n### 📖 入门指南\n- [**快速入门指南**](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen?tab=readme-ov-file#1--installation)\n- **立即体验**：\n ```bash\n pip install massgen==0.1.68\n # 尝试检查点模式——在输入框上方的模式栏中点击“COORD”，然后勾选检查点复选框\n uv run massgen --web\n ```\n\n## 变更内容\n* 修复：锁定 litellm 版本至 1.82.6，并提交 uv.lock 文件，以防止供应链攻击，由 @ncrispino 在 https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1025 中完成。\n* 功能：检查点模式，由 @ncrispino 在 https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1028 中实现。\n* 功能：LLM API 熔断机制，支持 429 错误分类（Claude 后端），由 @amabito 在 https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1024 中实现。\n* 文档：v0.1.68 的文档，由 @Henry-811 在 https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1029 中完成。\n* 功能：v0.1.68 版本，由 @Henry-811 在 https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1022 中完成。\n\n\n**完整变更日志**：https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fcompare\u002Fv0.1.67...v0.1.68","2026-03-25T17:57:06",{"id":174,"version":175,"summary_zh":176,"released_at":177},162748,"v0.1.67","# 🚀 发布亮点 — v0.1.67 (2026-03-23)\n\n### 🖥️ [现代化 WebUI](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Fwebui.html)\n- **内联最终答案**: 最终答案直接在 AgentChannel 中显示，不再使用模态弹窗\n- **键盘快捷键**: 通过 `useV2KeyboardShortcuts` 实现响应式导航\n- **现代化架构**: 使用 Zustand 状态管理库（消息、模式、面板、代理、主题）替代临时性的状态管理方式\n\n### 💰 RoundBudgetGuardHook\n- **每轮费用强制执行**: 实时跟踪累计及每轮 API 费用，超出预算时自动阻止执行\n- **可配置警告**: 默认阈值为预算的 50%、75% 和 90%，超限时会优雅地终止流程\n\n### 🎭 统一预协作阶段\n- **并行执行**: 角色生成、评估标准和提示优化现在同时进行\n- **统一批量展示**: 单个 TUI 屏幕同时显示所有预协作阶段\n- **角色多样性模式**: 提供三种模式——视角、实现方式和方法论\n\n### 🛡️ 回归防护\n- **盲测 A\u002FB 验证**: 专用子代理比较当前与上一次的回答，但不透露哪一个是新的\n- **基于标准的评估**: 根据完整评估标准列表（E1..EN）进行评估，以捕捉潜在的无声回归问题\n\n---\n\n### 📖 入门指南\n- [**快速入门指南**](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen?tab=readme-ov-file#1--installation)\n- **立即体验**:\n ```bash\n pip install massgen==0.1.67\n # 体验现代化 WebUI\n uv run massgen --web\n ```\n\n## 变更内容\n* feat: 由 @ncrispino 在 https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1016 中实现 WebUI 现代化\n* feat: 由 @amabito 在 https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1013 中添加用于成本控制的 RoundBudgetGuardHook\n* docs: 由 @Henry-811 在 https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1017 中编写 v0.1.67 版本文档\n* feat: 由 @Henry-811 在 https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1015 中完成 v0.1.67 版本开发\n\n## 新贡献者\n* @amabito 在 https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1013 中完成了首次贡献\n\n**完整变更日志**: https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fcompare\u002Fv0.1.66...v0.1.67","2026-03-23T18:06:33",{"id":179,"version":180,"summary_zh":181,"released_at":182},162749,"v0.1.66","# 🚀 发布亮点 — v0.1.66（2026-03-20）\n\n### 🔄 步进模式\n- **一个智能体，执行一步，然后退出**：新增 `--step` CLI 标志，供外部编排器运行单次智能体迭代\n- **会话目录**：加载先前的回答\u002F工作空间，并将更新后的状态写回——与来源无关（MassGen、Claude Code、Shell 脚本）\n- **massgen-refinery 集成**：为 [massgen-refinery Claude Code 插件](https:\u002F\u002Fgithub.com\u002Fmassgen\u002Fmassgen-refinery) 提供逐步执行能力\n\n### ✅ 修复\n- **Codex Windows UTF-8**：确保在 Codex 后端写入文件时使用 UTF-8 编码\n- **控制台文本净化**：新增 `sanitize_console_text` 工具函数，用于在日志记录器和 TUI 事件管道中安全渲染文本\n\n---\n\n### 📖 入门指南\n- [**快速入门指南**](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen?tab=readme-ov-file#1--installation)\n- **试用**：\n ```bash\n pip install massgen==0.1.66\n # 运行已配置智能体的一步\n uv run massgen --step --config your_config.yaml --session-dir .\u002Fmy_session \"您的任务\"\n ```\n\n## 变更内容\n* 功能：由 @ncrispino 在 https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1011 中实现的步进模式\n* 修复：由 @praneeth999 在 https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1010 中完成的 Codex Windows 问题修复\n* 文档：由 @Henry-811 在 https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1012 中编写的 v0.1.66 版本文档\n* 功能：由 @Henry-811 在 https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1009 中完成的 v0.1.66 版本功能开发\n\n\n**完整变更日志**：https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fcompare\u002Fv0.1.65...v0.1.66","2026-03-20T18:42:41",{"id":184,"version":185,"summary_zh":186,"released_at":187},162750,"v0.1.65","# 🚀 发布亮点 — v0.1.65（2026-03-18）\n\n### 🔧 [MassGen 精炼插件](https:\u002F\u002Fgithub.com\u002Fmassgen\u002Fmassgen-refinery)\n独立的 MCP 服务器，可将 MassGen 的质量门控和多轮工作流引入 Claude Code——无需完整的编排器。单代理精炼功能已完全可用；多代理协调目前仍处于实验阶段。\n\n### ✅ [质量服务器](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fblob\u002Fmain\u002Fmassgen\u002Fmcp_tools\u002Fstandalone\u002Fquality_server.py)（`massgen_quality_tools`）\n- **基于会话的评估**：带有时间戳的质量会话，评估标准存储在 `.massgen-quality\u002F` 目录中\n- **检查清单评分**：根据预设标准对回答进行评估，并支持自定义阈值\n- **改进建议**：覆盖度验证功能可识别未满足的标准，并提出需要改进的空白点\n\n### 📋 [工作流服务器](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fblob\u002Fmain\u002Fmassgen\u002Fmcp_tools\u002Fstandalone\u002Fworkflow_server.py)（`massgen_workflow_tools`）\n- **多轮回答**：提交答案时会自动创建交付物快照，并将其存入各轮目录，便于追踪迭代过程\n- **投票支持**：无状态透传功能，用于支持代理间的共识机制\n\n### 🖼️ [媒体服务器](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fblob\u002Fmain\u002Fmassgen\u002Fmcp_tools\u002Fstandalone\u002Fmedia_server.py)（`massgen_media_tools`）\n- **生成媒体**：支持文本转图像、视频和音频，可选择性地提供输入媒体；自动检测可用后端（Google Gemini、OpenAI、Grok、ElevenLabs）\n- **读取媒体**：以关键优先的顺序分析媒体文件，并支持多文件对比\n\n---\n\n### 📖 入门指南\n- [**快速入门指南**](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen?tab=readme-ov-file#1--installation)\n- **试用**：\n ```bash\n pip install massgen==0.1.65\n # 独立的 MCP 服务器现已适用于 massgen-refinery Claude Code 插件\n ```\n\n## 变更内容\n* 功能：新增通用的独立 MCP 服务器，尤其适用于 @ncrispino 在 https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1007 中开发的 massgen-refinery 插件\n* 文档：v0.1.65 版本的文档由 @Henry-811 在 https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1008 中完成\n* 功能：v0.1.65 版本由 @Henry-811 在 https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1005 中发布\n\n\n**完整变更日志**：https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fcompare\u002Fv0.1.64...v0.1.65","2026-03-18T18:55:30",{"id":189,"version":190,"summary_zh":191,"released_at":192},162751,"v0.1.64","# 🚀 发布亮点 — v0.1.64 (2026-03-16)\n\n### 🔌 [Gemini CLI 后端](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Fbackends.html)\n- **Gemini CLI 作为原生后端**：基于子进程的流式传输，集成 Google 的 Gemini CLI\n- **会话持久化**：通过 CLI 会话 ID 实现多轮对话\n- **MCP 工具**：通过 `.gemini\u002Fsettings.json` 配置，并使用原生钩子适配器执行工具\n- **Docker 支持**：通过 `gemini_cli_docker.yaml` 配置实现容器化运行\n\n### 🔍 [执行轨迹分析器](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fblob\u002Fmain\u002Fmassgen\u002Fsubagent_types\u002Fexecution_trace_analyzer\u002FSUBAGENT.md)\n- **全新子代理类型**：对智能体执行轨迹进行机制性分析，提取可复用的学习成果\n- **七维度评估**：错误学习、精力分配、方法有效性、工具策略、推理模式、上下文健康度、验证完整性\n- **输出**：`process_report.md`（叙述性报告）和 `process_verdict.json`（结构化评分）\n\n### ⚡ WebSocket 流式传输\n- **持久化 WebSocket 传输**：使用 `wss:\u002F\u002F` 连接 OpenAI 响应 API，实现实时事件流\n- **自动重连**：支持可配置的指数退避重试逻辑\n- **YAML 配置**：在 OpenAI 后端启用 `websocket_mode: true`\n\n### 🐳 Copilot Docker 模式\n- **容器化工具执行**：为 Copilot 后端设置 `command_line_execution_mode: \"docker\"`\n- **配置选项**：支持 Docker 的 sudo 权限及网络模式选择（桥接或主机模式）\n\n### ✅ 修复\n- **Response API 重复项问题**：防止在递归工具循环中出现重复项错误 ([#1000](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1000))\n\n---\n\n### 📖 快速入门\n- [**快速入门指南**](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen?tab=readme-ov-file#1--installation)\n- **试用**：\n ```bash\n pip install massgen==0.1.64\n # 尝试 Gemini CLI 后端\n uv run massgen --config @examples\u002Fproviders\u002Fgemini\u002Fgemini_cli_local \"解释量子计算\"\n ```\n\n## 变更内容\n* 轨迹评估子代理，由 @ncrispino 在 https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1002 中完成\n* 功能：为 Response API 添加 websocket_mode，由 @praneeth999 在 https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F990 中实现\n* 修复：Response API 重复项错误及 Windows 平台 MCP 服务器崩溃问题，由 @db-ol 在 https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1000 中解决\n* 功能：Gemini CLI 代码清理，由 @ncrispino 在 https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F999 中完成\n* 文档：v0.1.64 版本文档，由 @Henry-811 在 https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F1003 中编写\n* 功能：v0.1.64 版本发布，由 @Henry-811 在 https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F998 中完成\n\n\n**完整变更日志**：https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fcompare\u002Fv0.1.63...v0.1.64","2026-03-16T17:13:39",{"id":194,"version":195,"summary_zh":196,"released_at":197},162752,"v0.1.63","# 🚀 Release Highlights — v0.1.63 (2026-03-13)\r\n\r\n### 🎯 [Ensemble Pattern](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Freference\u002Fyaml_schema.html#ensemble-pattern-recommended-defaults)\r\n- **Ensemble defaults for subagents**: `disable_injection` and `defer_voting_until_all_answered` now default to true, so subagents work independently before voting for more diverse, higher-quality results\r\n- **Automatic ensemble orchestration**: Defaults apply when spawning subagent orchestrators without explicit override\r\n\r\n### 🔄 [Round Evaluator Improvements](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fblob\u002Fmain\u002Fdocs\u002Fmodules\u002Fcoordination_workflow.md#transformation-pressure)\r\n- **Transformation pressure**: Evaluator pushes agents toward meaningful structural changes rather than surface-level edits\r\n- **Success contracts**: Explicit quality gates agents must satisfy before the round evaluator allows convergence\r\n- **Verification replay**: Evaluation consistency across rounds via replayed verification context\r\n\r\n### ⚡ [Lighter Refinement](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Fadvanced\u002Fsubagents.html#refine-mode)\r\n- **Reduced subagent overhead**: Lighter refinement prompts for subagent workflows cut token usage and latency\r\n- **Killed agent handling**: Graceful management of agents that time out or fail mid-round\r\n\r\n### ✅ Fixes\r\n- **Timeout fallback**: More robust coordination when agents hit timeout boundaries\r\n\r\n---\r\n\r\n### 📖 Getting Started\r\n- [**Quick Start Guide**](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen?tab=readme-ov-file#1--installation)\r\n- **Try It**:\r\n ```bash\r\n pip install massgen==0.1.63\r\n # Try the round evaluator with ensemble defaults\r\n uv run massgen --config @examples\u002Ffeatures\u002Fround_evaluator_example.yaml \"Create a polished landing page for an AI product\"\r\n ```\r\n\r\n## What's Changed\r\n* feat: Better subagent contracts; lighter refinement for subagents too by @ncrispino in https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F996\r\n* docs: docs for v0.1.63 by @Henry-811 in https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F997\r\n* feat: v0.1.63 by @Henry-811 in https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F995\r\n\r\n\r\n**Full Changelog**: https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fcompare\u002Fv0.1.62...v0.1.63","2026-03-13T18:26:53",{"id":199,"version":200,"summary_zh":201,"released_at":202},162753,"v0.1.62","# 🚀 Release Highlights — v0.1.62 (2026-03-11)\r\n\r\n### 🧩 [MassGen Skill](https:\u002F\u002Fgithub.com\u002Fmassgen\u002Fskills)\r\n- **Multi-agent collaboration as a skill**: Install with `npx skills add massgen\u002Fskills --all` and use MassGen directly from Claude Code, Cursor, Copilot, and 40+ other AI agents\r\n- **Four modes**: General (any task), Evaluate (critique existing work), Plan (structured project plans), Spec (requirements specifications)\r\n- **Auto-distributed**: Skill automatically syncs to a [dedicated repository](https:\u002F\u002Fgithub.com\u002Fmassgen\u002Fskills) for easy installation\r\n\r\n### 👁️ [Session Viewer](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Freference\u002Fcli.html)\r\n- **Watch automation runs in real-time**: New `massgen viewer` command opens a TUI to observe running or completed sessions\r\n- **Session picker**: `--pick` flag for browsing and selecting specific sessions, `--web` for browser-based viewing\r\n\r\n### ⚡ [Backend & Quickstart Improvements](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Fbackends.html)\r\n- **Claude Code backend**: Background task execution and native MCP support via the SDK\r\n- **Codex backend**: Native filesystem access and MCP tool integration\r\n- **Copilot backend**: Runtime model discovery with automatic capability detection\r\n- **Headless quickstart**: Non-interactive setup via `--quickstart --headless` for CI\u002FCD pipelines\r\n- **Web quickstart**: Browser-based setup via `--web-quickstart`\r\n\r\n### ✅ Fixes\r\n- **Evaluation criteria**: Removed should\u002Fcould criteria that caused agents to produce overly similar outputs\r\n- **Planning prompts**: Improved planning prompts with configurable thoroughness levels\r\n\r\n---\r\n\r\n### 📖 Getting Started\r\n- [**Quick Start Guide**](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen?tab=readme-ov-file#1--installation)\r\n- **Try It**:\r\n ```bash\r\n # Install the MassGen Skill for your AI agent\r\n npx skills add massgen\u002Fskills --all\r\n # Then use MassGen from Claude Code, Cursor, Copilot, etc.\r\n\r\n # Or install MassGen directly and try the Session Viewer\r\n pip install massgen==0.1.62\r\n uv run massgen viewer --pick\r\n ```\r\n\r\n## What's Changed\r\n* feat: MassGen skill by @ncrispino in https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F992\r\n* docs: docs for v0.1.62 by @Henry-811 in https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F993\r\n* feat: v0.1.62 by @Henry-811 in https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F991\r\n\r\n\r\n**Full Changelog**: https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fcompare\u002Fv0.1.61...v0.1.62","2026-03-11T18:43:25",{"id":204,"version":205,"summary_zh":206,"released_at":207},162754,"v0.1.61","# 🚀 Release Highlights — v0.1.61 (2026-03-09)\r\n\r\n### 🔄 [Round Evaluator Paradigm](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Fadvanced\u002Fsubagents.html)\r\n- **Automatic post-answer evaluation**: New `round_evaluator` subagent type that automatically spawns evaluator subagents after each new answer, feeding detailed feedback into the next round\r\n- **Configurable evaluation flow**: Control whether evaluation runs before or after checklist grading, whether to skip synthesis, and whether evaluators refine their feedback\r\n- **Example config**: New `round_evaluator_example.yaml` — one agent builds while three agents evaluate in parallel\r\n\r\n### 📝 [Evaluation & Prompt Improvements](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Fconcepts.html#multi-agent-coordination)\r\n- **Task plan injection**: Evaluation prompts now include the current task plan for context-aware quality assessment\r\n- **Clearer evaluation prompts**: Rewritten round evaluation prompts for more actionable, focused feedback\r\n\r\n### ✅ Fixes\r\n- **Session resumption**: Fixed crash when resuming from an already-resumed log\r\n- **Timeout fallback**: When coordination times out, the latest answer is used directly without an extra final presentation step\r\n- **Subagent compatibility**: Improved SUBAGENT.md template for broader subagent type support\r\n- **Codex backend**: Added Codex backend support for new orchestrator features\r\n\r\n---\r\n\r\n### 📖 Getting Started\r\n- [**Quick Start Guide**](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen?tab=readme-ov-file#1--installation)\r\n- **Try It**:\r\n ```bash\r\n # Install or upgrade to v0.1.61\r\n pip install --upgrade massgen\r\n\r\n # One agent builds, 3 agents evaluate — round evaluator in action\r\n uv run massgen --config @examples\u002Ffeatures\u002Fround_evaluator_example.yaml \"Create a website about a fictional AI product that is visually stunning and has at least one unique interactive element\"\r\n ```\r\n\r\n## What's Changed\r\n* feat: Add round evaluator paradigm by @ncrispino in https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F986\r\n* docs: docs for v0.1.61 by @Henry-811 in https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F987\r\n* feat: v0.1.61 by @Henry-811 in https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F985\r\n\r\n\r\n**Full Changelog**: https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fcompare\u002Fv0.1.60...v0.1.61","2026-03-09T17:43:17",{"id":209,"version":210,"summary_zh":211,"released_at":212},162755,"v0.1.60","# 🚀 Release Highlights — v0.1.60 (2026-03-06)\r\n\r\n### 🛠️ [Multimodal Tool Improvements](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Fadvanced\u002Fmultimodal.html)\r\n- **Rewritten `read_media`**: Clearer tool schema, better error handling, and improved naming for more reliable media understanding\r\n- **Media Call Ledger**: Automatic tracking of all `read_media` and `generate_media` calls for observability\r\n\r\n### 🤖 [Subagent Enhancements](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Fadvanced\u002Fsubagents.html)\r\n- **Backend Inheritance**: New `inherit_spawning_agent_backend` option lets subagents automatically use the same backend as the agent that spawned them\r\n- **Final Answer Strategy**: New `final_answer_strategy` option controls how subagent orchestrators select the final answer (reuse winner, have winner present, or synthesize)\r\n- **Per-Agent Subagent Configs**: Each agent can now define its own `subagent_agents` override for fine-grained subagent control\r\n\r\n### 🧠 [GPT-5.4 Support](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Fbackends.html)\r\n- **New default OpenAI flagship**: GPT-5.4 added to the model registry, ready to use across all coordination modes\r\n\r\n### 🔄 [Decomposition + Checklist Cooperation](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Fconcepts.html#multi-agent-coordination)\r\n- **Unified Quality Workflow**: Decomposition mode now cooperates with the checklist workflow, enabling quality-gated subtask iteration\r\n- **Faster Verification Rounds**: Improved prompts for verification replay, reducing verification round time\r\n\r\n### ✅ Fixes\r\n- **Checklist & Prompt Injection**: More reliable checklist behavior with improved proposal injection and system prompt refocused on entire output quality\r\n- **Codex Pricing Accuracy**: Fixed prompt caching calculation for correct cost tracking\r\n- **Task Plan Refresh**: Fixed plan refresh during quality rounds\r\n- **Skill Prefix Handling**: Fixed edge cases in skill prefix resolution\r\n\r\n---\r\n\r\n### 📖 Getting Started\r\n- [**Quick Start Guide**](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen?tab=readme-ov-file#1--installation)\r\n- **Try It**:\r\n ```bash\r\n # Install or upgrade to v0.1.60\r\n pip install --upgrade massgen\r\n\r\n # Choose backend 'openai' with model 'gpt-5.4' in the setup wizard to start using GPT-5.4\r\n uv run massgen --quickstart\r\n ```\r\n\r\n## What's Changed\r\n* feat: Improve verification time by @ncrispino in https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F978\r\n* docs: docs for v0.1.60 by @Henry-811 in https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F979\r\n* feat: v0.1.60 by @Henry-811 in https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F974\r\n\r\n\r\n**Full Changelog**: https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fcompare\u002Fv0.1.59...v0.1.60","2026-03-06T19:06:36",{"id":214,"version":215,"summary_zh":216,"released_at":217},162756,"v0.1.59","# 🚀 Release Highlights — v0.1.59 (2026-03-04) \r\n ### 🔄 [Smarter Quality Rounds](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Fconcepts.html#multi-agent-coordination) \r\n - **Verification Replay Memories**: Agents save replayable verification steps (commands, scripts, artifacts) to `verification_latest.md`, auto-injected into future rounds so the next agent can replay the exact verification pipeline \r\n - **Plan-Tracked Improvements**: Improvements from each round are auto-added to the task plan, so agents build on prior progress instead of starting fresh\r\n - **Enhanced Plan Review**: More thorough quality evaluation during plan review phases\r\n\r\n ### ✅ [Checklist & Evaluation Fixes](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Fconcepts.html#multi-agent-coordination)\r\n - **More Accurate Evaluations**: Improved evaluation generation config for higher-quality assessments\r\n - **Consistent Checklist Behavior**: Fixed checklist handling across rounds\r\n - **Gemini MCP Compatibility**: Tool name normalization for Gemini backends using MCP tools\r\n\r\n ### 🔧 Infrastructure\r\n - **Subagent Enhancements**: Improved coordination, task delegation, and Docker skill write access\r\n - **Video Generation Fixes**: Cleaner error handling (no silent fallback to animated), restored impact metrics\r\n - **Bug Fixes**: Answer anonymization fix, quickstart updates\r\n\r\n ---\r\n\r\n ### 📖 Getting Started\r\n - [**Quick Start Guide**](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen?tab=readme-ov-file#1--installation)\r\n - **Try It**:\r\n ```bash\r\n # Install or upgrade to v0.1.59\r\n pip install --upgrade massgen\r\n\r\n # Try checklist-gated quality rounds with verification replay\r\n uv run massgen --config @examples\u002Ffeatures\u002Fsubagent_checklist.yaml \\\r\n \"Create a website for an AI company selling a creative sci-fi style product. Ensure polished visuals and cool interactive elements\"\r\n ```\r\n\r\n## What's Changed\r\n* feat: Improve quality rounds by @ncrispino in https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F969\r\n* docs: docs for v0.1.59 by @Henry-811 in https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F970\r\n* feat: v0.1.59 by @Henry-811 in https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F968\r\n\r\n\r\n**Full Changelog**: https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fcompare\u002Fv0.1.58...v0.1.59","2026-03-04T18:28:28",{"id":219,"version":220,"summary_zh":221,"released_at":222},162757,"v0.1.58","# 🚀 Release Highlights — v0.1.58 (2026-03-02)\r\n\r\n### 🖼️ [Comprehensive Multimodal Revamp](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Fadvanced\u002Fmultimodal.html)\r\n- **New Media Providers**: ElevenLabs (TTS\u002FSTT), Nano Banana 2 (default image gen), and Grok Imagine (image\u002Fvideo) join existing providers\r\n- **Media Generation Skills**: Reusable skills for image, video, and audio generation workflows\r\n- **Multi-Turn Image Editing**: Iterative image editing for supported providers — agents can refine images across rounds\r\n\r\n### 🟢 [Nvidia NIM Backend](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Fbackends.html)\r\n- **NVIDIA Inference Microservices**: First-class provider for NVIDIA-hosted models via NIM API\r\n\r\n### 🔍 [Quality Rethinking Subagent](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Fadvanced\u002Fsubagents.html)\r\n- **Per-Element Craft Improvements**: New `quality_rethinking` subagent type that targets specific elements for refinement\r\n- **Improve\u002FPreserve Checklists**: Checklists now explicitly separate what to improve vs. what to preserve\r\n\r\n### 🖥️ [CLI Mode Flags](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fquickstart\u002Frunning-massgen.html)\r\n- **New Flags**: `--quick`, `--single-agent`, `--coordination-mode`, `--personas` flags mirror TUI toggles\r\n- **Plan Mode from CLI**: Start plan mode directly from the command line\r\n\r\n### 🔧 Infrastructure\r\n- **Logging Refactor**: Fixed concurrent logging for parallel multi-agent execution — each agent gets isolated log context\r\n- **Subagent Hardening**: Better error handling for malformed inputs and repeated tool calls\r\n- **Evaluation Criteria Defaults**: Sensible defaults when criteria are not explicitly specified\r\n\r\n---\r\n\r\n### 📖 Getting Started \r\n- [**Quick Start Guide**](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen?tab=readme-ov-file#1--installation)\r\n- **Try It**:\r\n ```bash\r\n # Install or upgrade to v0.1.58\r\n pip install --upgrade massgen\r\n\r\n # Try checklist-driven refinement with quality rethinking\r\n uv run massgen --config @examples\u002Ffeatures\u002Fsubagent_checklist.yaml \\\r\n \"Create a website for an AI company selling a creative sci-fi style product. Ensure polished visuals and cool interactive elements\"\r\n ```\r\n\r\n## What's Changed\r\n* feat(backend): add Nvidia NIM as a first-class backend provider by @AbhimanyuAryan in https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F962\r\n* feat: Improve coordination with improvements; improve and expand multimedia generation by @ncrispino in https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F964\r\n* docs: docs for v0.1.58 by @Henry-811 in https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F965\r\n* feat: v0.1.58 by @Henry-811 in https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F957\r\n\r\n\r\n**Full Changelog**: https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fcompare\u002Fv0.1.57...v0.1.58","2026-03-02T19:26:33",{"id":224,"version":225,"summary_zh":226,"released_at":227},162758,"v0.1.57","# 🚀 Release Highlights — v0.1.57 (2026-02-27) \r\n \r\n ### 🔗 [Subagent Delegation Protocol](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Fconcepts.html#multi-agent-coordination)\r\n - **Container-to-Host Spawning**: File-based delegation via `SubagentLaunchWatcher` with atomic JSON request\u002Fresponse exchange and workspace path validation\r\n - **Auto-Mounted Parent Workspace**: Subagents get parent workspace (read-only) by default \r\n \r\n ### 🏗️ [Builder Subagent](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Fconcepts.html#multi-agent-coordination)\r\n - **Fresh-Context Artifact Generation**: New subagent type for transformative redesigns and complex multi-file rewrites with prescriptive specs\r\n\r\n ### 📊 Smarter Convergence\r\n - **Substantiveness Tracking**: Planned changes classified as transformative\u002Fstructural\u002Fincremental — triggers builder or novelty subagent accordingly\r\n - **Diagnostic Report Gating**: Optional quality gate — agents must submit a structured diagnostic report before checklist passes\r\n - **Per-Agent Checklist Scoring**: Evaluate multiple agents separately with automatic format detection\r\n\r\n ### 🔧 Infrastructure\r\n - **Claude Code Reasoning**: Unified `reasoning` config (type, effort, budget_tokens) replacing deprecated `max_thinking_tokens`\r\n - **Bug Fixes**: Fixed Codex subagent spawning, subagent synchronization\u002Ftimeout handling, temp workspace directories\r\n\r\n ---\r\n\r\n ### 📖 Getting Started\r\n - [**Quick Start Guide**](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen?tab=readme-ov-file#1--installation)\r\n - **Try It**:\r\n ```bash\r\n # Install or upgrade to v0.1.57\r\n pip install --upgrade massgen\r\n\r\n # Try builder subagent with checklist-driven refinement\r\n uv run massgen --config @examples\u002Ffeatures\u002Fsubagent_checklist.yaml \\\r\n \"Create a website for an AI company selling a creative sci-fi style product. Ensure polished visuals and cool interactive elements\"\r\n ```\r\n\r\n## What's Changed\r\n* feat: Improve subagent calling, eval criteria by @ncrispino in https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F955\r\n* docs: docs for v0.1.57 by @Henry-811 in https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F956\r\n* feat: v0.1.57 by @Henry-811 in https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F947\r\n\r\n\r\n**Full Changelog**: https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fcompare\u002Fv0.1.56...v0.1.57","2026-02-27T18:46:19",{"id":229,"version":230,"summary_zh":231,"released_at":232},162759,"v0.1.56","# 🚀 Release Highlights — v0.1.56 (2026-02-25) \r\n\r\n ### 📋 [Spec Plan Mode](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Ftask_planning.html) \r\n - **Formal Requirements Before Execution**: `plan_mode=\"spec\"` — spec creation, approval modal, and execution pipeline with changedoc integration. TUI spec mode (Shift+Tab twice to enter)\r\n\r\n ### 🎯 [ask_others Targeted Messaging](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Fconcepts.html#multi-agent-coordination)\r\n - **Focused Agent-to-Agent Communication**: `target_agents` parameter for directed queries instead of broadcast, with per-target validation and response counting\r\n\r\n ### 🔧 Infrastructure\r\n - **Critic Subagent**: New subagent type for honest quality assessment — detects genuine vs incremental improvement, quality ceiling. Enhanced novelty subagent guidance for growth-oriented refinement\r\n - **read_media Conversation Continuity**: Follow-up image conversations via `continue_from` conversation_id\r\n - **Codex OAuth Login Fix**: Codex backend always available in WebUI regardless of OPENAI_API_KEY\r\n - **Docker Configuration Mounting**: Claude and Codex configuration mounting for Docker containers\r\n\r\n ---\r\n\r\n ### 📖 Getting Started\r\n - [**Quick Start Guide**](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen?tab=readme-ov-file#1--installation)\r\n - **Try It**:\r\n ```bash\r\n # Install or upgrade to v0.1.56\r\n pip install --upgrade massgen\r\n\r\n # Launch MassGen, then press Shift+Tab twice to enter 'spec' mode\r\n uv run massgen\r\n ````\r\n\r\n## What's Changed\r\n* fix: adding codex for OAuth login by @MuL1ian in https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F937\r\n* feat: Add spec mode by @ncrispino in https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F945\r\n* docs: docs for v0.1.56 by @Henry-811 in https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F946\r\n* feat: v0.1.56 by @Henry-811 in https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F944\r\n\r\n\r\n**Full Changelog**: https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fcompare\u002Fv0.1.55...v0.1.56","2026-02-25T18:31:57",{"id":234,"version":235,"summary_zh":236,"released_at":237},162760,"v0.1.55","# 🚀 Release Highlights — v0.1.55 (2026-02-23) \r\n ### 🧩 [Specialized Subagent Types](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Fadvanced\u002Fsubagents.html) \r\n - **Discovery-Based Roles**: Specialized subagent roles via `SUBAGENT.md` frontmatter — evaluator (programmatic verification), explorer (investigation), researcher (deep analysis), novelty (breaks refinement plateaus)\r\n - **TUI Visualization**: Subagent roles displayed in agent status indicators\r\n\r\n ### 📊 [Dynamic Evaluation Criteria](docs\u002Fmodules\u002Fcomposition.md)\r\n - **Task-Specific Quality Gates**: GEPA-inspired evaluation criteria generation replacing static checklist items, with domain-specific presets (persona, decomposition, evaluation, prompt, analysis)\r\n - **Core\u002FStretch Convergence**: Items categorized as core or stretch — convergence off-ramp triggers when all core items pass. Score scale 0-10. Config: `evaluation_criteria_generator`\r\n\r\n ### 🔧 Infrastructure\r\n - **Native Backend Image Routing**: `understand_image` routes to agent's own backend (Claude, Gemini, Grok, Claude Code, Codex) with OpenAI fallback\r\n - **Configurable Video Frame Extraction**: Scene-based (PySceneDetect) or uniform modes with `max_frames` cost guardrail (default 30, max 60). Config: `multimodal_config.video` \r\n - **Remotion Skill**: Video generation\u002Fediting skill available in quickstart wizard\r\n - **Unified Pre-Collaboration**: Persona generation, decomposition, and eval criteria generation unified as composable primitives\r\n\r\n ---\r\n\r\n ### 📖 Getting Started\r\n - [**Quick Start Guide**](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen?tab=readme-ov-file#1--installation)\r\n - **Try It**:\r\n ```bash\r\n # Install or upgrade to v0.1.55\r\n pip install --upgrade massgen\r\n\r\n # Multi-agent coordination with specialized subagents\r\n uv run massgen --config massgen\u002Fconfigs\u002Ffeatures\u002Fbackground_subagent_example.yaml --cwd-context ro \"Use an explorer subagent to analyze this repo\"\r\n ```\r\n\r\n## What's Changed\r\n* feat: Add subagent roles by @ncrispino in https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F938\r\n* docs: docs for v0.1.55 by @Henry-811 in https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F939\r\n* feat: v0.1.55 by @Henry-811 in https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F935\r\n\r\n\r\n**Full Changelog**: https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fcompare\u002Fv0.1.54...v0.1.55","2026-02-23T18:50:24",{"id":239,"version":240,"summary_zh":241,"released_at":242},162761,"v0.1.54","# 🚀 Release Highlights — v0.1.54 (2026-02-20)\r\n \r\n ### 💬 [Subagent Runtime Messaging](https:\u002F\u002Fdocs.massgen.ai\u002Fen\u002Flatest\u002Fuser_guide\u002Fadvanced\u002Fsubagents.html) \r\n - **Steer Running Subagents**: Send messages to background subagents mid-execution via the TUI — target a specific agent or broadcast to all, with queued message management and cancel\u002Fclear controls\r\n\r\n ### 🤖 [Copilot SDK Backend](massgen\u002Fconfigs\u002Fbasic\u002Fsingle\u002Fcopilot.yaml)\r\n - **New `copilot` Backend**: Powered by `github-copilot-sdk` with native MCP server integration and custom tool handling\r\n - **Session Management**: Cache invalidation and session lifecycle; auth via GitHub subscription\r\n\r\n ### 🔧 Infrastructure\r\n - **Gemini 3.1 Pro**: `gemini-3.1-pro-preview` model added to capabilities registry\r\n - **MCP Hooks Improvements**: Hook middleware for subagent MCP servers, `InjectionDeliveryStatus` enum for delivery tracking, hook-dir argument for PostToolUse injection\r\n - **Type Annotation Modernization**: Codebase-wide migration from `typing.Dict\u002FList\u002FOptional\u002FUnion` to modern `dict\u002Flist\u002FX | None` syntax\r\n\r\n ---\r\n\r\n ### 📖 Getting Started\r\n - [**Quick Start Guide**](https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen?tab=readme-ov-file#1--installation)\r\n - **Try It**:\r\n ```bash\r\n # Install or upgrade to v0.1.54\r\n pip install --upgrade massgen\r\n\r\n # Launch with the new Copilot backend\r\n uv run massgen --config massgen\u002Fconfigs\u002Fbasic\u002Fsingle\u002Fcopilot.yaml \"What is GitHub copilot?\"\r\n\r\n # Multi-agent coordination with subagent messaging\r\n uv run massgen --config @examples\u002Ffeatures\u002Ftest_subagent_orchestrator_code_mode.yaml \"Use subagents to research Bob Dylan\"\r\n ```\r\n\r\n## What's Changed\r\n* feat: Adding support for Copilot SDK by @AbhimanyuAryan in https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F862\r\n* docs: survey cloud platforms and provide suggestion for MassGen by @int-chaos in https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F921\r\n* Fixes MAS-299 by @MuL1ian in https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F920\r\n* feat: Subagent messaging by @ncrispino in https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F926\r\n* docs: docs for v0.1.54 by @Henry-811 in https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F928\r\n* feat: v0.1.54 by @Henry-811 in https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F924\r\n\r\n## New Contributors\r\n* @int-chaos made their first contribution in https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fpull\u002F921\r\n\r\n**Full Changelog**: https:\u002F\u002Fgithub.com\u002Fmassgen\u002FMassGen\u002Fcompare\u002Fv0.1.53...v0.1.54","2026-02-20T19:38:19"]