[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"similar-aden-hive--hive":3,"tool-aden-hive--hive":62},[4,18,26,36,46,54],{"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",160411,2,"2026-04-18T23:33:24",[14,13,35],"语言模型",{"id":37,"name":38,"github_repo":39,"description_zh":40,"stars":41,"difficulty_score":42,"last_commit_at":43,"category_tags":44,"status":17},8272,"opencode","anomalyco\u002Fopencode","OpenCode 是一款开源的 AI 编程助手（Coding Agent），旨在像一位智能搭档一样融入您的开发流程。它不仅仅是一个代码补全插件，而是一个能够理解项目上下文、自主规划任务并执行复杂编码操作的智能体。无论是生成全新功能、重构现有代码，还是排查难以定位的 Bug，OpenCode 都能通过自然语言交互高效完成，显著减少开发者在重复性劳动和上下文切换上的时间消耗。\n\n这款工具专为软件开发者、工程师及技术研究人员设计，特别适合希望利用大模型能力来提升编码效率、加速原型开发或处理遗留代码维护的专业人群。其核心亮点在于完全开源的架构，这意味着用户可以审查代码逻辑、自定义行为策略，甚至私有化部署以保障数据安全，彻底打破了传统闭源 AI 助手的“黑盒”限制。\n\n在技术体验上，OpenCode 提供了灵活的终端界面（Terminal UI）和正在测试中的桌面应用程序，支持 macOS、Windows 及 Linux 全平台。它兼容多种包管理工具，安装便捷，并能无缝集成到现有的开发环境中。无论您是追求极致控制权的资深极客，还是渴望提升产出的独立开发者，OpenCode 都提供了一个透明、可信",144296,1,"2026-04-16T14:50:03",[13,45],"插件",{"id":47,"name":48,"github_repo":49,"description_zh":50,"stars":51,"difficulty_score":32,"last_commit_at":52,"category_tags":53,"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。",109154,"2026-04-18T11:18:24",[14,15,13],{"id":55,"name":56,"github_repo":57,"description_zh":58,"stars":59,"difficulty_score":32,"last_commit_at":60,"category_tags":61,"status":17},6121,"gemini-cli","google-gemini\u002Fgemini-cli","gemini-cli 是一款由谷歌推出的开源 AI 命令行工具，它将强大的 Gemini 大模型能力直接集成到用户的终端环境中。对于习惯在命令行工作的开发者而言，它提供了一条从输入提示词到获取模型响应的最短路径，无需切换窗口即可享受智能辅助。\n\n这款工具主要解决了开发过程中频繁上下文切换的痛点，让用户能在熟悉的终端界面内直接完成代码理解、生成、调试以及自动化运维任务。无论是查询大型代码库、根据草图生成应用，还是执行复杂的 Git 操作，gemini-cli 都能通过自然语言指令高效处理。\n\n它特别适合广大软件工程师、DevOps 人员及技术研究人员使用。其核心亮点包括支持高达 100 万 token 的超长上下文窗口，具备出色的逻辑推理能力；内置 Google 搜索、文件操作及 Shell 命令执行等实用工具；更独特的是，它支持 MCP（模型上下文协议），允许用户灵活扩展自定义集成，连接如图像生成等外部能力。此外，个人谷歌账号即可享受免费的额度支持，且项目基于 Apache 2.0 协议完全开源，是提升终端工作效率的理想助手。",100752,"2026-04-10T01:20:03",[45,13,15,14],{"id":63,"github_repo":64,"name":65,"description_en":66,"description_zh":67,"ai_summary_zh":67,"readme_en":68,"readme_zh":69,"quickstart_zh":70,"use_case_zh":71,"hero_image_url":72,"owner_login":73,"owner_name":74,"owner_avatar_url":75,"owner_bio":76,"owner_company":77,"owner_location":77,"owner_email":78,"owner_twitter":79,"owner_website":80,"owner_url":81,"languages":82,"stars":120,"forks":121,"last_commit_at":122,"license":123,"difficulty_score":32,"env_os":124,"env_gpu":125,"env_ram":125,"env_deps":126,"category_tags":132,"github_topics":133,"view_count":32,"oss_zip_url":77,"oss_zip_packed_at":77,"status":17,"created_at":148,"updated_at":149,"faqs":150,"releases":179},9500,"aden-hive\u002Fhive","hive","Multi-Agent Harness for Production AI","Hive 是一款专为生产环境打造的多智能体协作框架，旨在让复杂的 AI 任务自动化运行更稳定、高效。它解决了传统多智能体系统在长流程业务中容易出现的状态混乱、错误难恢复、缺乏监控等痛点，无需繁琐配置，只需定义目标，Hive 就能自动构建基于图结构的执行流程，协调多个专用智能体并行完成任务。\n\nHive 特别适合开发者和技术团队使用，尤其是那些希望将 AI 智能体应用于实际业务场景（如数据处理、网页操作、流程自动化）的人群。其独特亮点包括：支持动态生成多智能体拓扑结构、具备基于角色的持久化记忆机制、提供完整的状态可观测性与故障自愈能力，并兼容 OpenAI、Anthropic、Google Gemini 等多种大模型。此外，Hive 还支持“人在回路”（Human-in-the-Loop）机制，确保关键节点有人工干预的可能，提升系统可靠性。\n\n无论是构建自动化工作流，还是探索多智能体协同的新应用，Hive 都能以零样板代码的方式，帮助团队快速部署可信赖的 AI 代理系统。","\u003Cp align=\"center\">\n  \u003Cimg width=\"100%\" alt=\"Hive Banner\" src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Faden-hive_hive_readme_d9fb50469d32.gif\" \u002F>\n\u003C\u002Fp>\n\n\u003Cp align=\"center\">\n  \u003Ca href=\"README.md\">English\u003C\u002Fa> |\n  \u003Ca href=\"docs\u002Fi18n\u002Fzh-CN.md\">简体中文\u003C\u002Fa> |\n  \u003Ca href=\"docs\u002Fi18n\u002Fes.md\">Español\u003C\u002Fa> |\n  \u003Ca href=\"docs\u002Fi18n\u002Fhi.md\">हिन्दी\u003C\u002Fa> |\n  \u003Ca href=\"docs\u002Fi18n\u002Fpt.md\">Português\u003C\u002Fa> |\n  \u003Ca href=\"docs\u002Fi18n\u002Fja.md\">日本語\u003C\u002Fa> |\n  \u003Ca href=\"docs\u002Fi18n\u002Fru.md\">Русский\u003C\u002Fa> |\n  \u003Ca href=\"docs\u002Fi18n\u002Fko.md\">한국어\u003C\u002Fa>\n\u003C\u002Fp>\n\n\u003Cp align=\"center\">\n  \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Faden-hive\u002Fhive\u002Fblob\u002Fmain\u002FLICENSE\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FLicense-Apache%202.0-blue.svg\" alt=\"Apache 2.0 License\" \u002F>\u003C\u002Fa>\n  \u003Ca href=\"https:\u002F\u002Fwww.ycombinator.com\u002Fcompanies\u002Faden\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FY%20Combinator-Aden-orange\" alt=\"Y Combinator\" \u002F>\u003C\u002Fa>\n  \u003Ca href=\"https:\u002F\u002Fdiscord.com\u002Finvite\u002FMXE49hrKDk\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fdiscord\u002F1172610340073242735?logo=discord&labelColor=%235462eb&logoColor=%23f5f5f5&color=%235462eb\" alt=\"Discord\" \u002F>\u003C\u002Fa>\n  \u003Ca href=\"https:\u002F\u002Fx.com\u002Faden_hq\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Ftwitter\u002Ffollow\u002Fteamaden?logo=X&color=%23f5f5f5\" alt=\"Twitter Follow\" \u002F>\u003C\u002Fa>\n  \u003Ca href=\"https:\u002F\u002Fwww.linkedin.com\u002Fcompany\u002Fteamaden\u002F\">\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Faden-hive_hive_readme_006e5e7b2fbb.png\" alt=\"LinkedIn\" \u002F>\u003C\u002Fa>\n  \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FMCP-102_Tools-00ADD8?style=flat-square\" alt=\"MCP\" \u002F>\n\u003C\u002Fp>\n\n\u003Cp align=\"center\">\n  \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FAgent_Harness-Runtime_Layer-ff6600?style=flat-square\" alt=\"Agent Harness\" \u002F>\n  \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FAI_Agents-Self--Improving-brightgreen?style=flat-square\" alt=\"AI Agents\" \u002F>\n  \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FMulti--Agent-Systems-blue?style=flat-square\" alt=\"Multi-Agent\" \u002F>\n  \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FHeadless-Development-purple?style=flat-square\" alt=\"Headless\" \u002F>\n  \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FHuman--in--the--Loop-orange?style=flat-square\" alt=\"HITL\" \u002F>\n  \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FBrowser-Use-red?style=flat-square\" alt=\"Browser Use\" \u002F>\n\u003C\u002Fp>\n\u003Cp align=\"center\">\n  \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FOpenAI-supported-412991?style=flat-square&logo=openai\" alt=\"OpenAI\" \u002F>\n  \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FAnthropic-supported-d4a574?style=flat-square\" alt=\"Anthropic\" \u002F>\n  \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FGoogle_Gemini-supported-4285F4?style=flat-square&logo=google\" alt=\"Gemini\" \u002F>\n\u003C\u002Fp>\n\n\u003Cp align=\"center\">\u003Cem>The agent harness for production workloads — state management, failure recovery, observability, and human oversight so your agents actually run.\u003C\u002Fem>\u003C\u002Fp>\n\n## Overview\n\nOpenHive is a zero-setup, model-agnostic execution harness that dynamically generates multi-agent topologies to tackle complex, long-running business workflows without requiring any orchestration boilerplate. By simply defining your objective, the runtime compiles a strict, graph-based execution DAG that safely coordinates specialized agents to execute concurrent tasks in parallel. Backed by persistent, role-based memory that intelligently evolves with your project's context, OpenHive ensures deterministic fault tolerance, deep state observability, and seamless asynchronous execution across whichever underlying LLMs you choose to plug in.\n\n## Features\n\n- ✅ Multi-Agent Coordination for parallel task execution \n- ✅ Graph-based execution for recurring and complex processes \n- ✅ Role-based memory that evolves with your projects \n- ✅ Zero Setup - No technical configuration required\n- ✅ General Compute Use and Browser Use with Native Extension \n- ✅ Custom Model Support\n\nVisit [adenhq.com](https:\u002F\u002Fadenhq.com) for complete documentation, examples, and guides.\n\nVisit [HoneyComb](http:\u002F\u002Fhoneycomb.open-hive.com\u002F) to see what jobs are being automated by AI. It’s a stock market for jobs, driven by our community’s AI agent progress. You can long and short jobs (with no real money but compute token)based on how much you think a job is going to be replaced by AI.\n\nhttps:\u002F\u002Fgithub.com\u002Fuser-attachments\u002Fassets\u002Fbf10edc3-06ba-48b6-98ba-d069b15fb69d\n\n\n## Who Is Hive For?\n\nHive is the multi-agent harness layer for teams moving AI agents from prototype to production. Single agents like Openclaw and Cowork can finish personal jobs pretty well but lack the rigor to fulfil business processes. \n\nHive is a good fit if you:\n\n- Want AI agents that **execute real business processes**, not demos\n- Need a **runtime that handles state, recovery, and parallel execution** at scale\n- Need **self-healing and adaptive agents** that improve over time\n- Require **human-in-the-loop control**, observability, and cost limits\n- Plan to run agents in **production** where uptime, cost, and auditability matter\n\nHive may not be the best fit if you’re only experimenting with simple agent chains or one-off scripts.\n\n## When Should You Use Hive?\n\nUse Hive when the bottleneck is no longer the model but the harness around it:\n\n- Long-running agents that need **state persistence and crash recovery**\n- Production workloads requiring **cost enforcement, observability, and audit trails**\n- Agents that **self-heal** through failure capture and graph evolution\n- Multi-agent coordination with **session isolation and shared buffers**\n- A framework that **scales with model improvements** rather than fighting them\n\n## Quick Links\n\n- **[Documentation](https:\u002F\u002Fdocs.adenhq.com\u002F)** - Complete guides and API reference\n- **[Self-Hosting Guide](https:\u002F\u002Fdocs.adenhq.com\u002Fgetting-started\u002Fquickstart)** - Deploy Hive on your infrastructure\n- **[Changelog](https:\u002F\u002Fgithub.com\u002Faden-hive\u002Fhive\u002Freleases)** - Latest updates and releases\n- **[Roadmap](docs\u002Froadmap.md)** - Upcoming features and plans\n- **[Report Issues](https:\u002F\u002Fgithub.com\u002Faden-hive\u002Fhive\u002Fissues)** - Bug reports and feature requests\n- **[Contributing](CONTRIBUTING.md)** - How to contribute and submit PRs\n\n## Quick Start\n\n### Prerequisites\n\n- Python 3.11+ for agent development\n- An LLM provider that powers the agents\n- **ripgrep (optional, recommended on Windows):** The `search_files` tool uses ripgrep for faster file search. If not installed, a Python fallback is used. On Windows: `winget install BurntSushi.ripgrep` or `scoop install ripgrep`\n\n> **Windows Users:** Native Windows is supported via `quickstart.ps1` and `hive.ps1`. Run these in PowerShell 5.1+. WSL is also an option but not required.\n\n### Installation\n\n> **Note**\n> Hive uses a `uv` workspace layout and is not installed with `pip install`.\n> Running `pip install -e .` from the repository root will create a placeholder package and Hive will not function correctly.\n> Please use the quickstart script below to set up the environment.\n\n```bash\n# Clone the repository\ngit clone https:\u002F\u002Fgithub.com\u002Faden-hive\u002Fhive.git\ncd hive\n\n# Run quickstart setup (macOS\u002FLinux)\n.\u002Fquickstart.sh\n\n# Windows (PowerShell)\n.\\quickstart.ps1\n```\n\nThis sets up:\n\n- **framework** - Core agent runtime and graph executor (in `core\u002F.venv`)\n- **aden_tools** - MCP tools for agent capabilities (in `tools\u002F.venv`)\n- **credential store** - Encrypted API key storage (`~\u002F.hive\u002Fcredentials`)\n- **LLM provider** - Interactive default model configuration, including Hive LLM and OpenRouter\n- All required Python dependencies with `uv`\n\n- Finally, it will open the Hive interface in your browser\n\n> **Tip:** To reopen the dashboard later, run `hive open` from the project directory.\n\n### Build Your First Agent\n\nType the agent you want to build in the home input box. The queen is going to ask you questions and work out a solution with you.\n\n\u003Cimg width=\"2500\" height=\"1214\" alt=\"Image\" src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Faden-hive_hive_readme_9b14b86ab053.png\" \u002F>\n\n### Use Template Agents\n\nClick \"Try a sample agent\" and check the templates. You can run a template directly or choose to build your version on top of the existing template.\n\n### Run Agents\n\nNow you can run an agent by selecting the agent (either an existing agent or example agent). You can click the Run button on the top left, or talk to the queen agent and it can run the agent for you.\n\n\u003Cimg width=\"2549\" height=\"1174\" alt=\"Screenshot 2026-03-12 at 9 27 36 PM\" src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Faden-hive_hive_readme_d3643f9c5dd4.png\" \u002F>\n\n## Integration\n\n\u003Ca href=\"https:\u002F\u002Fgithub.com\u002Faden-hive\u002Fhive\u002Ftree\u002Fmain\u002Ftools\u002Fsrc\u002Faden_tools\u002Ftools\">\u003Cimg width=\"100%\" alt=\"Integration\" src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Faden-hive_hive_readme_88a8f401ad27.png\" \u002F>\u003C\u002Fa>\nHive is built to be model-agnostic and system-agnostic.\n\n- **LLM flexibility** - Hive Framework supports Anthropic, OpenAI, OpenRouter, Hive LLM, and other hosted or local models through LiteLLM-compatible providers.\n- **Business system connectivity** - Hive Framework is designed to connect to all kinds of business systems as tools, such as CRM, support, messaging, data, file, and internal APIs via MCP.\n\n## Why Hive\n\nAs models improve, the upper bound of what agents can do rises — but their reliability and production value are determined by the harness. Hive focuses on generating agents that run real business processes rather than generic agents. Instead of requiring you to manually design workflows, define agent interactions, and handle failures reactively, Hive flips the paradigm: **you describe outcomes, and the system builds itself**—delivering an outcome-driven, adaptive experience with an easy-to-use set of tools and integrations.\n\n```mermaid\nflowchart LR\n    GOAL[\"Define Goal\"] --> GEN[\"Auto-Generate Graph\"]\n    GEN --> EXEC[\"Execute Agents\"]\n    EXEC --> MON[\"Monitor & Observe\"]\n    MON --> CHECK{{\"Pass?\"}}\n    CHECK -- \"Yes\" --> DONE[\"Deliver Result\"]\n    CHECK -- \"No\" --> EVOLVE[\"Evolve Graph\"]\n    EVOLVE --> EXEC\n\n    GOAL -.- V1[\"Natural Language\"]\n    GEN -.- V2[\"Instant Architecture\"]\n    EXEC -.- V3[\"Easy Integrations\"]\n    MON -.- V4[\"Full visibility\"]\n    EVOLVE -.- V5[\"Adaptability\"]\n    DONE -.- V6[\"Reliable outcomes\"]\n\n    style GOAL fill:#ffbe42,stroke:#cc5d00,stroke-width:2px,color:#333\n    style GEN fill:#ffb100,stroke:#cc5d00,stroke-width:2px,color:#333\n    style EXEC fill:#ff9800,stroke:#cc5d00,stroke-width:2px,color:#fff\n    style MON fill:#ff9800,stroke:#cc5d00,stroke-width:2px,color:#fff\n    style CHECK fill:#fff59d,stroke:#ed8c00,stroke-width:2px,color:#333\n    style DONE fill:#4caf50,stroke:#2e7d32,stroke-width:2px,color:#fff\n    style EVOLVE fill:#e8763d,stroke:#cc5d00,stroke-width:2px,color:#fff\n    style V1 fill:#fff,stroke:#ed8c00,stroke-width:1px,color:#cc5d00\n    style V2 fill:#fff,stroke:#ed8c00,stroke-width:1px,color:#cc5d00\n    style V3 fill:#fff,stroke:#ed8c00,stroke-width:1px,color:#cc5d00\n    style V4 fill:#fff,stroke:#ed8c00,stroke-width:1px,color:#cc5d00\n    style V5 fill:#fff,stroke:#ed8c00,stroke-width:1px,color:#cc5d00\n    style V6 fill:#fff,stroke:#ed8c00,stroke-width:1px,color:#cc5d00\n```\n\n### How It Works\n\n1. **[Define Your Goal](docs\u002Fkey_concepts\u002Fgoals_outcome.md)** → Describe what you want to achieve in plain English\n2. **Coding Agent Generates** → Creates the [agent graph](docs\u002Fkey_concepts\u002Fgraph.md), connection code, and test cases\n3. **[Workers Execute](docs\u002Fkey_concepts\u002Fworker_agent.md)** → SDK-wrapped nodes run with full observability and tool access\n4. **Control Plane Monitors** → Real-time metrics, budget enforcement, policy management\n5. **[Adaptiveness](docs\u002Fkey_concepts\u002Fevolution.md)** → On failure, the system evolves the graph and redeploys automatically\n\n## Documentation\n\n- **[Developer Guide](docs\u002Fdeveloper-guide.md)** - Comprehensive guide for developers\n- [Getting Started](docs\u002Fgetting-started.md) - Quick setup instructions\n- [Configuration Guide](docs\u002Fconfiguration.md) - All configuration options\n- [Architecture Overview](docs\u002Farchitecture\u002FREADME.md) - System design and structure\n\n## Contributing\nWe welcome contributions from the community! We’re especially looking for help building tools, integrations, and example agents for the framework ([check #2805](https:\u002F\u002Fgithub.com\u002Faden-hive\u002Fhive\u002Fissues\u002F2805)). If you’re interested in extending its functionality, this is the perfect place to start. Please see [CONTRIBUTING.md](CONTRIBUTING.md) for guidelines.\n\n**Important:** Please get assigned to an issue before submitting a PR. Comment on an issue to claim it, and a maintainer will assign you. Issues with reproducible steps and proposals are prioritized. This helps prevent duplicate work.\n\n1. Find or create an issue and get assigned\n2. Fork the repository\n3. Create your feature branch (`git checkout -b feature\u002Famazing-feature`)\n4. Commit your changes (`git commit -m 'Add amazing feature'`)\n5. Push to the branch (`git push origin feature\u002Famazing-feature`)\n6. Open a Pull Request\n\n## Community & Support\n\nWe use [Discord](https:\u002F\u002Fdiscord.com\u002Finvite\u002FMXE49hrKDk) for support, feature requests, and community discussions.\n\n- Discord - [Join our community](https:\u002F\u002Fdiscord.com\u002Finvite\u002FMXE49hrKDk)\n- Twitter\u002FX - [@adenhq](https:\u002F\u002Fx.com\u002Faden_hq)\n- LinkedIn - [Company Page](https:\u002F\u002Fwww.linkedin.com\u002Fcompany\u002Fteamaden\u002F)\n\n## Join Our Team\n\n**We're hiring!** Join us in engineering, research, and go-to-market roles.\n\n[View Open Positions](https:\u002F\u002Fjobs.adenhq.com\u002Fa8cec478-cdbc-473c-bbd4-f4b7027ec193\u002Fapplicant)\n\n## Security\n\nFor security concerns, please see [SECURITY.md](SECURITY.md).\n\n## License\n\nThis project is licensed under the Apache License 2.0 - see the [LICENSE](LICENSE) file for details.\n\n## Frequently Asked Questions (FAQ)\n\n**Q: What LLM providers does Hive support?**\n\nHive supports 100+ LLM providers through LiteLLM integration, including OpenAI (GPT-4, GPT-4o), Anthropic (Claude models), Google Gemini, DeepSeek, Mistral, Groq, OpenRouter, and Hive LLM. Simply set the appropriate API key environment variable and specify the model name. See [docs\u002Fconfiguration.md](docs\u002Fconfiguration.md) for provider-specific configuration examples.\n\n**Q: Can I use Hive with local AI models like Ollama?**\n\nYes! Hive supports local models through LiteLLM. Simply use the model name format `ollama\u002Fmodel-name` (e.g., `ollama\u002Fllama3`, `ollama\u002Fmistral`) and ensure Ollama is running locally.\n\n**Q: What makes Hive different from other agent frameworks?**\n\nHive is an agent harness, not just an orchestration framework. It provides the production runtime layer — session isolation, checkpoint-based crash recovery, cost enforcement, real-time observability, and human-in-the-loop controls — that makes agents reliable enough to run real workloads. On top of that, Hive generates your entire agent system from natural language goals and automatically [evolves the graph](docs\u002Fkey_concepts\u002Fevolution.md) when agents fail. The combination of a robust harness with self-improving generation is what sets Hive apart.\n\n**Q: Is Hive open-source?**\n\nYes, Hive is fully open-source under the Apache License 2.0. We actively encourage community contributions and collaboration.\n\n**Q: Does Hive support human-in-the-loop workflows?**\n\nYes, Hive fully supports [human-in-the-loop](docs\u002Fkey_concepts\u002Fgraph.md#human-in-the-loop) workflows through intervention nodes that pause execution for human input. These include configurable timeouts and escalation policies, allowing seamless collaboration between human experts and AI agents.\n\n**Q: What programming languages does Hive support?**\n\nThe Hive framework is built in Python. A JavaScript\u002FTypeScript SDK is on the roadmap.\n\n**Q: Can Hive agents interact with external tools and APIs?**\n\nYes. Aden's SDK-wrapped nodes provide built-in tool access, and the framework supports flexible tool ecosystems. Agents can integrate with external APIs, databases, and services through the node architecture.\n\n**Q: How does cost control work in Hive?**\n\nHive provides granular budget controls including spending limits, throttles, and automatic model degradation policies. You can set budgets at the team, agent, or workflow level, with real-time cost tracking and alerts.\n\n**Q: Where can I find examples and documentation?**\n\nVisit [docs.adenhq.com](https:\u002F\u002Fdocs.adenhq.com\u002F) for complete guides, API reference, and getting started tutorials. The repository also includes documentation in the `docs\u002F` folder and a comprehensive [developer guide](docs\u002Fdeveloper-guide.md).\n\n**Q: How can I contribute to Aden?**\n\nContributions are welcome! Fork the repository, create your feature branch, implement your changes, and submit a pull request. See [CONTRIBUTING.md](CONTRIBUTING.md) for detailed guidelines.\n\n## Star History\n\n\u003Ca href=\"https:\u002F\u002Fstar-history.com\u002F#aden-hive\u002Fhive&Date\">\n \u003Cpicture>\n   \u003Csource media=\"(prefers-color-scheme: dark)\" srcset=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Faden-hive_hive_readme_0edb124f1e5c.png&theme=dark\" \u002F>\n   \u003Csource media=\"(prefers-color-scheme: light)\" srcset=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Faden-hive_hive_readme_0edb124f1e5c.png\" \u002F>\n   \u003Cimg alt=\"Star History Chart\" src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Faden-hive_hive_readme_0edb124f1e5c.png\" \u002F>\n \u003C\u002Fpicture>\n\u003C\u002Fa>\n\n---\n\n\u003Cp align=\"center\">\n  Made with 🔥 Passion in San Francisco\n\u003C\u002Fp>\n","\u003Cp align=\"center\">\n  \u003Cimg width=\"100%\" alt=\"Hive Banner\" src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Faden-hive_hive_readme_d9fb50469d32.gif\" \u002F>\n\u003C\u002Fp>\n\n\u003Cp align=\"center\">\n  \u003Ca href=\"README.md\">English\u003C\u002Fa> |\n  \u003Ca href=\"docs\u002Fi18n\u002Fzh-CN.md\">简体中文\u003C\u002Fa> |\n  \u003Ca href=\"docs\u002Fi18n\u002Fes.md\">Español\u003C\u002Fa> |\n  \u003Ca href=\"docs\u002Fi18n\u002Fhi.md\">हिन्दी\u003C\u002Fa> |\n  \u003Ca href=\"docs\u002Fi18n\u002Fpt.md\">Português\u003C\u002Fa> |\n  \u003Ca href=\"docs\u002Fi18n\u002Fja.md\">日本語\u003C\u002Fa> |\n  \u003Ca href=\"docs\u002Fi18n\u002Fru.md\">Русский\u003C\u002Fa> |\n  \u003Ca href=\"docs\u002Fi18n\u002Fko.md\">한국어\u003C\u002Fa>\n\u003C\u002Fp>\n\n\u003Cp align=\"center\">\n  \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Faden-hive\u002Fhive\u002Fblob\u002Fmain\u002FLICENSE\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FLicense-Apache%202.0-blue.svg\" alt=\"Apache 2.0 License\" \u002F>\u003C\u002Fa>\n  \u003Ca href=\"https:\u002F\u002Fwww.ycombinator.com\u002Fcompanies\u002Faden\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FY%20Combinator-Aden-orange\" alt=\"Y Combinator\" \u002F>\u003C\u002Fa>\n  \u003Ca href=\"https:\u002F\u002Fdiscord.com\u002Finvite\u002FMXE49hrKDk\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fdiscord\u002F1172610340073242735?logo=discord&labelColor=%235462eb&logoColor=%23f5f5f5&color=%235462eb\" alt=\"Discord\" \u002F>\u003C\u002Fa>\n  \u003Ca href=\"https:\u002F\u002Fx.com\u002Faden_hq\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Ftwitter\u002Ffollow\u002Fteamaden?logo=X&color=%23f5f5f5\" alt=\"Twitter Follow\" \u002F>\u003C\u002Fa>\n  \u003Ca href=\"https:\u002F\u002Fwww.linkedin.com\u002Fcompany\u002Fteamaden\u002F\">\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Faden-hive_hive_readme_006e5e7b2fbb.png\" alt=\"LinkedIn\" \u002F>\u003C\u002Fa>\n  \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FMCP-102_Tools-00ADD8?style=flat-square\" alt=\"MCP\" \u002F>\n\u003C\u002Fp>\n\n\u003Cp align=\"center\">\n  \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FAgent_Harness-Runtime_Layer-ff6600?style=flat-square\" alt=\"Agent Harness\" \u002F>\n  \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FAI_Agents-Self--Improving-brightgreen?style=flat-square\" alt=\"AI Agents\" \u002F>\n  \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FMulti--Agent-Systems-blue?style=flat-square\" alt=\"Multi-Agent\" \u002F>\n  \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FHeadless-Development-purple?style=flat-square\" alt=\"Headless\" \u002F>\n  \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FHuman--in--the--Loop-orange?style=flat-square\" alt=\"HITL\" \u002F>\n  \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FBrowser-Use-red?style=flat-square\" alt=\"Browser Use\" \u002F>\n\u003C\u002Fp>\n\u003Cp align=\"center\">\n  \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FOpenAI-supported-412991?style=flat-square&logo=openai\" alt=\"OpenAI\" \u002F>\n  \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FAnthropic-supported-d4a574?style=flat-square\" alt=\"Anthropic\" \u002F>\n  \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FGoogle_Gemini-supported-4285F4?style=flat-square&logo=google\" alt=\"Gemini\" \u002F>\n\u003C\u002Fp>\n\n\u003Cp align=\"center\">\u003Cem>面向生产级工作负载的智能体运行时框架——提供状态管理、故障恢复、可观测性以及人工监督功能，确保您的智能体真正落地运行。\u003C\u002Fem>\u003C\u002Fp>\n\n## 概述\n\nOpenHive 是一款无需任何配置、与模型无关的执行框架，能够动态生成多智能体拓扑结构，以高效处理复杂且长期运行的业务流程，而无需编写任何编排模板代码。用户只需定义目标，运行时便会自动生成一个严格的基于图的执行 DAG，安全协调各专业智能体并行执行任务。借助持久化的角色化内存，OpenHive 能够根据项目上下文智能演进，从而实现确定性的容错能力、深度的状态可观测性，以及在任意底层大语言模型上无缝进行异步执行。\n\n## 特性\n\n- ✅ 多智能体协同，支持并行任务执行\n- ✅ 基于图的执行架构，适用于重复性和复杂流程\n- ✅ 角色化内存，随项目进展动态演化\n- ✅ 零配置——无需任何技术设置\n- ✅ 支持通用计算及浏览器原生扩展使用\n- ✅ 自定义模型支持\n\n访问 [adenhq.com](https:\u002F\u002Fadenhq.com) 获取完整文档、示例和指南。\n\n访问 [HoneyComb](http:\u002F\u002Fhoneycomb.open-hive.com\u002F) 查看由 AI 自动化的各类工作。这是一个由社区 AI 智能体发展驱动的工作市场，您可以基于对某项工作被 AI 取代程度的判断，用计算代币“做多”或“做空”这些工作。\n\nhttps:\u002F\u002Fgithub.com\u002Fuser-attachments\u002Fassets\u002Fbf10edc3-06ba-48b6-98ba-d069b15fb69d\n\n\n## Hive 适合哪些人群？\n\nHive 是一款用于将 AI 智能体从原型阶段推向生产环境的多智能体运行时框架。像 Openclaw 和 Cowork 这样的单体智能体虽然能很好地完成个人任务，但缺乏满足企业级业务流程所需的严谨性。\n\n如果您符合以下条件，Hive 将非常适合您：\n\n- 希望部署能够**执行真实业务流程**而非仅用于演示的 AI 智能体\n- 需要一个能够在大规模场景下处理**状态管理、故障恢复和并行执行**的运行时框架\n- 需要具备**自我修复和自适应能力**、能够持续优化的智能体\n- 需要**人工介入控制**、可观测性以及成本约束\n- 计划在**生产环境**中运行智能体，其中系统可用性、成本和可审计性至关重要\n\n然而，如果您目前只是在试验简单的智能体链或一次性脚本，则 Hive 可能并不是最佳选择。\n\n## 何时应该使用 Hive？\n\n当瓶颈不再是模型本身，而是围绕模型的运行框架时，就该使用 Hive 了：\n\n- 需要**状态持久化和崩溃恢复**的长期运行智能体\n- 生产级工作负载需要**成本控制、可观测性及审计追踪**\n- 智能体能够通过捕获故障并动态调整执行图来实现**自我修复**\n- 多智能体协作需要**会话隔离和共享缓冲区**\n- 您希望使用一个能够**随着模型能力提升而扩展**，而不是与其对抗的框架\n\n## 快速链接\n\n- **[文档](https:\u002F\u002Fdocs.adenhq.com\u002F)** — 完整指南和 API 参考\n- **[自托管指南](https:\u002F\u002Fdocs.adenhq.com\u002Fgetting-started\u002Fquickstart)** — 在您的基础设施上部署 Hive\n- **[更新日志](https:\u002F\u002Fgithub.com\u002Faden-hive\u002Fhive\u002Freleases)** — 最新更新和发布版本\n- **[路线图](docs\u002Froadmap.md)** — 即将推出的功能和未来规划\n- **[报告问题](https:\u002F\u002Fgithub.com\u002Faden-hive\u002Fhive\u002Fissues)** — Bug 报告和功能请求\n- **[贡献指南](CONTRIBUTING.md)** — 如何参与贡献和提交 PR\n\n## 快速入门\n\n### 前置条件\n\n- Python 3.11 或更高版本，用于智能体开发\n- 一个为智能体提供支持的大语言模型提供商\n- **ripgrep（可选，Windows 推荐）：** `search_files` 工具使用 ripgrep 来加速文件搜索。如果未安装，将回退到 Python 实现。在 Windows 上：运行 `winget install BurntSushi.ripgrep` 或 `scoop install ripgrep`\n\n> **Windows 用户：** 通过 `quickstart.ps1` 和 `hive.ps1` 支持原生 Windows 环境。请在 PowerShell 5.1+ 中运行这些脚本。此外，WSL 也是一种选择，但并非必需。\n\n### 安装\n\n> **注意**\n> Hive 使用 `uv` 工作区布局，无法通过 `pip install` 进行安装。\n> 如果直接在仓库根目录下运行 `pip install -e .`，只会创建一个占位包，Hive 将无法正常工作。\n> 请使用下方的快速入门脚本设置环境。\n\n```bash\n# 克隆仓库\ngit clone https:\u002F\u002Fgithub.com\u002Faden-hive\u002Fhive.git\ncd hive\n\n# 运行快速入门设置（macOS\u002FLinux）\n.\u002Fquickstart.sh\n\n# Windows (PowerShell)\n.\\quickstart.ps1\n```\n\n这将设置：\n\n- **framework** - 核心代理运行时和图执行器（位于 `core\u002F.venv` 中）\n- **aden_tools** - 用于代理能力的 MCP 工具（位于 `tools\u002F.venv` 中）\n- **credential store** - 加密的 API 密钥存储（`~\u002F.hive\u002Fcredentials`）\n- **LLM provider** - 交互式的默认模型配置，包括 Hive LLM 和 OpenRouter\n- 所有必需的 Python 依赖项通过 `uv` 安装\n\n- 最后，它会在您的浏览器中打开 Hive 界面\n\n> **提示：** 若要稍后重新打开仪表板，请从项目目录运行 `hive open`。\n\n### 构建您的第一个代理\n\n在主页输入框中键入您想要构建的代理。女王会向您提问，并与您一起制定解决方案。\n\n\u003Cimg width=\"2500\" height=\"1214\" alt=\"Image\" src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Faden-hive_hive_readme_9b14b86ab053.png\" \u002F>\n\n### 使用模板代理\n\n点击“尝试示例代理”并查看模板。您可以直接运行模板，也可以选择在现有模板的基础上构建您自己的版本。\n\n### 运行代理\n\n现在您可以选择一个代理（无论是现有代理还是示例代理）来运行。您可以点击左上角的“运行”按钮，或者与女王代理对话，它会为您运行该代理。\n\n\u003Cimg width=\"2549\" height=\"1174\" alt=\"Screenshot 2026-03-12 at 9 27 36 PM\" src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Faden-hive_hive_readme_d3643f9c5dd4.png\" \u002F>\n\n## 集成\n\n\u003Ca href=\"https:\u002F\u002Fgithub.com\u002Faden-hive\u002Fhive\u002Ftree\u002Fmain\u002Ftools\u002Fsrc\u002Faden_tools\u002Ftools\">\u003Cimg width=\"100%\" alt=\"Integration\" src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Faden-hive_hive_readme_88a8f401ad27.png\" \u002F>\u003C\u002Fa>\nHive 的设计使其与模型和系统无关。\n\n- **LLM 灵活性** - Hive Framework 支持 Anthropic、OpenAI、OpenRouter、Hive LLM 以及其他通过 LiteLLM 兼容提供商托管或本地运行的模型。\n- **业务系统连接** - Hive Framework 被设计为能够通过 MCP 将各种业务系统作为工具连接起来，例如 CRM、支持、消息传递、数据、文件和内部 API。\n\n## 为什么选择 Hive\n\n随着模型性能的提升，代理的能力上限也在不断提高——但其可靠性和生产价值却取决于框架的设计。Hive 专注于构建能够运行实际业务流程的代理，而非通用型代理。与其让您手动设计工作流、定义代理交互并被动地处理故障，Hive 反其道而行之：**您只需描述期望的结果，系统便会自动构建所需的一切**——从而提供以结果为导向、适应性强的体验，并配备易于使用的工具和集成。\n\n```mermaid\nflowchart LR\n    GOAL[\"定义目标\"] --> GEN[\"自动生成图\"]\n    GEN --> EXEC[\"执行代理\"]\n    EXEC --> MON[\"监控与观测\"]\n    MON --> CHECK{{\"成功？\"}}\n    CHECK -- \"是\" --> DONE[\"交付结果\"]\n    CHECK -- \"否\" --> EVOLVE[\"进化图\"]\n    EVOLVE --> EXEC\n\n    GOAL -.- V1[\"自然语言\"]\n    GEN -.- V2[\"即时架构\"]\n    EXEC -.- V3[\"轻松集成\"]\n    MON -.- V4[\"全面可见性\"]\n    EVOLVE -.- V5[\"适应性\"]\n    DONE -.- V6[\"可靠结果\"]\n\n    style GOAL fill:#ffbe42,stroke:#cc5d00,stroke-width:2px,color:#333\n    style GEN fill:#ffb100,stroke:#cc5d00,stroke-width:2px,color:#333\n    style EXEC fill:#ff9800,stroke:#cc5d00,stroke-width:2px,color:#fff\n    style MON fill:#ff9800,stroke:#cc5d00,stroke-width:2px,color:#fff\n    style CHECK fill:#fff59d,stroke:#ed8c00,stroke-width:2px,color:#333\n    style DONE fill:#4caf50,stroke:#2e7d32,stroke-width:2px,color:#fff\n    style EVOLVE fill:#e8763d,stroke:#cc5d00,stroke-width:2px,color:#fff\n    style V1 fill:#fff,stroke:#ed8c00,stroke-width:1px,color:#cc5d00\n    style V2 fill:#fff,stroke:#ed8c00,stroke-width:1px,color:#cc5d00\n    style V3 fill:#fff,stroke:#ed8c00,stroke-width:1px,color:#cc5d00\n    style V4 fill:#fff,stroke:#ed8c00,stroke-width:1px,color:#cc5d00\n    style V5 fill:#fff,stroke:#ed8c00,stroke-width:1px,color:#cc5d00\n    style V6 fill:#fff,stroke:#ed8c00,stroke-width:1px,color:#cc5d00\n```\n\n### 工作原理\n\n1. **[定义您的目标](docs\u002Fkey_concepts\u002Fgoals_outcome.md)** → 用通俗易懂的英语描述您希望达成的目标\n2. **编码代理生成** → 创建 [代理图](docs\u002Fkey_concepts\u002Fgraph.md)、连接代码和测试用例\n3. **[工作者执行](docs\u002Fkey_concepts\u002Fworker_agent.md)** → SDK 包装的节点在完全可观测性和工具访问权限下运行\n4. **控制平面监控** → 实时指标、预算执行和策略管理\n5. **[适应性](docs\u002Fkey_concepts\u002Fevolution.md)** → 发生故障时，系统会自动进化图并重新部署\n\n## 文档\n\n- **[开发者指南](docs\u002Fdeveloper-guide.md)** - 针对开发者的全面指南\n- [入门指南](docs\u002Fgetting-started.md) - 快速设置说明\n- [配置指南](docs\u002Fconfiguration.md) - 所有配置选项\n- [架构概述](docs\u002Farchitecture\u002FREADME.md) - 系统设计和结构\n\n## 贡献\n\n我们欢迎社区的贡献！我们尤其需要帮助构建框架的工具、集成和示例代理（[请参阅 #2805](https:\u002F\u002Fgithub.com\u002Faden-hive\u002Fhive\u002Fissues\u002F2805)）。如果您有兴趣扩展其功能，这里正是开始的理想之地。请参阅 [CONTRIBUTING.md](CONTRIBUTING.md) 获取相关指南。\n\n**重要提示：** 请务必在提交 PR 之前先被分配到某个问题。在问题下方留言以认领，维护人员会为您分配。具有可复现步骤和详细方案的问题将优先处理。这样可以避免重复劳动。\n\n1. 查找或创建一个问题并被分配\n2. 分支仓库\n3. 创建您的功能分支（`git checkout -b feature\u002Famazing-feature`）\n4. 提交更改（`git commit -m '添加超赞功能'`）\n5. 推送到分支（`git push origin feature\u002Famazing-feature`）\n6. 打开拉取请求\n\n## 社区与支持\n\n我们使用 [Discord](https:\u002F\u002Fdiscord.com\u002Finvite\u002FMXE49hrKDk) 进行支持、功能请求和社区讨论。\n\n- Discord - [加入我们的社区](https:\u002F\u002Fdiscord.com\u002Finvite\u002FMXE49hrKDk)\n- Twitter\u002FX - [@adenhq](https:\u002F\u002Fx.com\u002Faden_hq)\n- LinkedIn - [公司页面](https:\u002F\u002Fwww.linkedin.com\u002Fcompany\u002Fteamaden\u002F)\n\n## 加入我们的团队\n\n**我们正在招聘！** 欢迎加入我们的工程、研究和市场推广团队。\n\n[查看空缺职位](https:\u002F\u002Fjobs.adenhq.com\u002Fa8cec478-cdbc-473c-bbd4-f4b7027ec193\u002Fapplicant)\n\n## 安全\n\n如有安全相关问题，请参阅 [SECURITY.md](SECURITY.md)。\n\n## 许可证\n\n本项目采用 Apache License 2.0 许可证授权——详情请参阅 [LICENSE](LICENSE) 文件。\n\n## 常见问题解答 (FAQ)\n\n**问：Hive 支持哪些大模型提供商？**\n\nHive 通过 LiteLLM 集成支持 100 多家大模型提供商，包括 OpenAI（GPT-4、GPT-4o）、Anthropic（Claude 系列模型）、Google Gemini、DeepSeek、Mistral、Groq、OpenRouter 以及 Hive LLM。只需设置相应的 API 密钥环境变量并指定模型名称即可。有关各提供商的具体配置示例，请参阅 [docs\u002Fconfiguration.md](docs\u002Fconfiguration.md)。\n\n**问：我能否将 Hive 与本地 AI 模型（如 Ollama）一起使用？**\n\n可以！Hive 通过 LiteLLM 支持本地模型。只需使用 `ollama\u002F模型名称` 的格式（例如 `ollama\u002Fllama3`、`ollama\u002Fmistral`），并确保 Ollama 已在本地运行。\n\n**问：Hive 与其他智能体框架有何不同？**\n\nHive 是一个智能体运行时框架，而不仅仅是编排框架。它提供了生产级的运行时层——会话隔离、基于检查点的故障恢复、成本控制、实时可观测性以及人机协作控制——这些功能使智能体能够可靠地处理实际工作负载。此外，Hive 还能根据自然语言目标自动生成整个智能体系统，并在智能体失败时自动[演化图谱](docs\u002Fkey_concepts\u002Fevolution.md)。正是这种强大的运行时框架与自我改进生成能力的结合，使 Hive 脱颖而出。\n\n**问：Hive 是开源的吗？**\n\n是的，Hive 在 Apache License 2.0 许可下完全开源。我们积极鼓励社区贡献和协作。\n\n**问：Hive 是否支持人机协作工作流？**\n\n是的，Hive 完全支持[人机协作](docs\u002Fkey_concepts\u002Fgraph.md#human-in-the-loop)工作流，通过干预节点暂停执行以等待人工输入。这些节点还提供可配置的超时和升级策略，从而实现人类专家与 AI 智能体之间的无缝协作。\n\n**问：Hive 支持哪些编程语言？**\n\nHive 框架是用 Python 构建的。JavaScript\u002FTypeScript SDK 正在开发计划中。\n\n**问：Hive 智能体能否与外部工具和 API 进行交互？**\n\n可以。Aden 的 SDK 封装节点提供了内置的工具访问功能，同时框架也支持灵活的工具生态系统。智能体可以通过节点架构集成外部 API、数据库和其他服务。\n\n**问：Hive 中的成本控制是如何实现的？**\n\nHive 提供细粒度的预算控制，包括支出上限、速率限制以及自动模型降级策略。您可以在团队、智能体或工作流级别设置预算，并获得实时成本跟踪和警报。\n\n**问：在哪里可以找到示例和文档？**\n\n请访问 [docs.adenhq.com](https:\u002F\u002Fdocs.adenhq.com\u002F) 获取完整的指南、API 参考和入门教程。仓库中的 `docs\u002F` 文件夹也包含相关文档，以及一份全面的[开发者指南](docs\u002Fdeveloper-guide.md)。\n\n**问：我如何为 Aden 做出贡献？**\n\n欢迎任何形式的贡献！您可以 fork 该仓库，创建自己的功能分支，实现更改后提交 pull request。详细指南请参阅 [CONTRIBUTING.md](CONTRIBUTING.md)。\n\n## 星标历史\n\n\u003Ca href=\"https:\u002F\u002Fstar-history.com\u002F#aden-hive\u002Fhive&Date\">\n \u003Cpicture>\n   \u003Csource media=\"(prefers-color-scheme: dark)\" srcset=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Faden-hive_hive_readme_0edb124f1e5c.png&theme=dark\" \u002F>\n   \u003Csource media=\"(prefers-color-scheme: light)\" srcset=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Faden-hive_hive_readme_0edb124f1e5c.png\" \u002F>\n   \u003Cimg alt=\"星标历史图表\" src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Faden-hive_hive_readme_0edb124f1e5c.png\" \u002F>\n \u003C\u002Fpicture>\n\u003C\u002Fa>\n\n---\n\n\u003Cp align=\"center\">\n  由旧金山的热情打造\n\u003C\u002Fp>","# Hive 快速上手指南\n\nHive 是一个面向生产环境的零配置、模型无关的多智能体执行框架。它能够根据你定义的目标，动态生成多智能体拓扑图，自动协调并行任务、处理状态持久化及故障恢复，适用于复杂的长周期业务工作流。\n\n## 环境准备\n\n在开始之前，请确保你的开发环境满足以下要求：\n\n*   **操作系统**：macOS, Linux 或 Windows (原生支持)。\n*   **Python 版本**：Python 3.11 或更高版本。\n*   **LLM 提供商**：需要拥有一个可用的大模型 API Key（支持 OpenAI, Anthropic, Google Gemini, OpenRouter 等）。\n*   **可选依赖 (推荐)**：\n    *   **ripgrep**：用于加速文件搜索工具 (`search_files`)。\n    *   *Windows 用户安装命令*：`winget install BurntSushi.ripgrep` 或 `scoop install ripgrep`。\n    *   *macOS\u002FLinux 用户*：通常可通过包管理器安装 (如 `brew install ripgrep`)。\n\n> **注意**：Hive 使用 `uv` 工作区布局管理依赖，**不要**直接使用 `pip install -e .` 安装，否则会导致功能异常。请务必使用官方提供的快速启动脚本。\n\n## 安装步骤\n\nHive 提供了自动化脚本以一键完成环境搭建、依赖安装及凭证配置。\n\n### 1. 克隆仓库\n\n```bash\ngit clone https:\u002F\u002Fgithub.com\u002Faden-hive\u002Fhive.git\ncd hive\n```\n\n### 2. 运行快速启动脚本\n\n根据你的操作系统选择对应的脚本执行：\n\n**macOS \u002F Linux:**\n```bash\n.\u002Fquickstart.sh\n```\n\n**Windows (PowerShell):**\n```powershell\n.\\quickstart.ps1\n```\n> *提示：Windows 用户请使用 PowerShell 5.1+ 运行上述命令。虽然也支持 WSL，但非必须。*\n\n### 3. 初始化配置\n\n脚本执行后将自动完成以下操作：\n*   创建核心框架 (`framework`) 和工具集 (`aden_tools`) 的独立虚拟环境。\n*   设置加密凭证存储 (`~\u002F.hive\u002Fcredentials`)。\n*   交互式配置默认的 LLM 提供商及 API Key。\n*   安装所有必要的 Python 依赖。\n*   自动在浏览器中打开 Hive 仪表盘界面。\n\n> **提示**：若需稍后重新打开仪表盘，可在项目目录下运行 `hive open`。\n\n## 基本使用\n\nHive 的核心交互通过 Web 界面进行，采用“目标驱动”模式：你只需描述想要达成的结果，系统会自动构建并执行智能体网络。\n\n### 1. 创建你的第一个智能体\n\n1.  在浏览器打开的 Hive 主页输入框中，用自然语言描述你想要构建的智能体或任务目标（例如：“分析本周的销售数据并生成报告”）。\n2.  系统中的 **Queen Agent** 会与你互动，询问细节并共同制定解决方案。\n3.  确认后，系统将自动生成执行图谱（Graph）和相应的代码。\n\n### 2. 使用模板智能体\n\n如果你希望快速体验或基于现有方案修改：\n1.  点击界面上的 **\"Try a sample agent\"** 按钮。\n2.  浏览可用的模板列表。\n3.  你可以直接运行模板，或者选择在其基础上构建你自己的版本。\n\n### 3. 运行智能体\n\n*   **手动运行**：在智能体列表中选择已创建的智能体（自定义或示例），点击左上角的 **Run** 按钮。\n*   **对话运行**：直接与 **Queen Agent** 对话，指令它为你运行特定的智能体。\n\n运行过程中，你可以在仪表盘实时观察：\n*   多智能体的并行执行状态。\n*   详细的日志与可观测性数据。\n*   故障自动恢复与图谱演化过程。\n\n---\n*更多详细文档、API 参考及自托管指南，请访问 [Hive 官方文档](https:\u002F\u002Fdocs.adenhq.com\u002F)。*","某电商运营团队需要每日从全球数十个竞品网站抓取价格、库存及促销信息，并生成动态分析报告以调整定价策略。\n\n### 没有 hive 时\n- **流程脆弱易断**：自定义编写的爬虫脚本一旦遇到网站结构微调或反爬机制，整个任务链即刻崩溃，需人工介入修复代码。\n- **并发能力受限**：串行执行导致处理上百个页面耗时数小时，无法在早会前产出最新数据，错失黄金调价窗口。\n- **状态黑盒难溯**：任务失败后难以定位是网络波动、解析错误还是模型幻觉，缺乏可视化的执行链路排查问题。\n- **上下文丢失严重**：多轮任务间无法共享历史数据，每次运行都是“从零开始”，无法基于过往趋势优化抓取策略。\n\n### 使用 hive 后\n- **自愈容错稳定**：hive 自动构建基于图的执行拓扑，当某个代理节点因反爬失败时，系统自动切换代理 IP 或重试，无需人工干预。\n- **高效并行处理**：通过多智能体协同，hive 将数百个网站的抓取任务拆解为并行子任务，将原本数小时的流程压缩至十分钟完成。\n- **全链路可观测**：提供深度的状态观察面板，清晰展示每个智能体的决策路径、工具调用及中间状态，故障根因一目了然。\n- **记忆持续进化**：基于角色的持久化记忆让智能体记住特定网站的反爬规律和历史价格波动，随项目运行越用越聪明。\n\nhive 将原本脆弱、缓慢且不可控的复杂数据工程，转化为具备自愈能力、实时响应且持续进化的自动化生产流。","https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Faden-hive_hive_d9fb5046.gif","aden-hive","Aden","https:\u002F\u002Foss.gittoolsai.com\u002Favatars\u002Faden-hive_0b60b1f2.png","Aden Hive is an open-source, Python-based framework designed for building production-grade, autonomous, and self-improving AI agents.",null,"contact@adenhq.com","aden_hq","https:\u002F\u002Fadenhq.com\u002F","https:\u002F\u002Fgithub.com\u002Faden-hive",[83,87,91,95,99,103,107,110,113,117],{"name":84,"color":85,"percentage":86},"Python","#3572A5",90.5,{"name":88,"color":89,"percentage":90},"TypeScript","#3178c6",6.9,{"name":92,"color":93,"percentage":94},"PowerShell","#012456",1.1,{"name":96,"color":97,"percentage":98},"Shell","#89e051",0.9,{"name":100,"color":101,"percentage":102},"HTML","#e34c26",0.3,{"name":104,"color":105,"percentage":106},"JavaScript","#f1e05a",0.1,{"name":108,"color":109,"percentage":106},"CSS","#663399",{"name":111,"color":112,"percentage":106},"TSQL","#e38c00",{"name":114,"color":115,"percentage":116},"Makefile","#427819",0,{"name":118,"color":119,"percentage":116},"Dockerfile","#384d54",10089,5609,"2026-04-18T21:46:24","Apache-2.0","Linux, macOS, Windows","未说明",{"notes":127,"python":128,"dependencies":129},"该工具使用 uv 工作区布局，不建议直接使用 pip install 安装。Windows 用户可通过 PowerShell 脚本运行，可选安装 ripgrep 以加速文件搜索。需要配置 LLM 提供商（如 OpenAI, Anthropic, Google Gemini 等）的 API 密钥。","3.11+",[130,131],"uv","ripgrep (可选)",[14,13],[134,135,136,137,138,139,140,141,142,143,144,145,146,147],"agent","agent-framework","agent-skills","anthropic","automation","autonomous-agents","claude","human-in-the-loop","openai","python","self-hosted","self-improving","harness","harness-engineering","2026-03-27T02:49:30.150509","2026-04-19T15:38:52.624425",[151,156,161,166,171,175],{"id":152,"question_zh":153,"answer_zh":154,"source_url":155},42613,"如何为缺失文档的工具贡献 README 文件？","您可以认领一批工具（例如每次 5-10 个）来编写 README。请遵循现有模式（如 `gmail_tool\u002FREADME.md`），内容需包含：工具名称与描述、可用函数及参数、所需的凭证和环境变量、使用示例。完成后可以提交单个 PR 包含多个工具的文档，或者为每个工具单独提交 PR，具体取决于维护者的偏好。目前已有贡献者通过分批提交 PR（如 #6886, #6887 等）成功完成了 42 个工具的文档补充。","https:\u002F\u002Fgithub.com\u002Faden-hive\u002Fhive\u002Fissues\u002F6486",{"id":157,"question_zh":158,"answer_zh":159,"source_url":160},42614,"TUI 中的“可视化差异模式”功能是否还会开发？","该功能原本计划在 TUI 中实现以展示图拓扑的演变差异，但由于 TUI 已被弃用，原计划已取消。不过，社区已将此功能转化为离线 CLI 工具。您可以使用新实现的 `hive diff \u003Cfile_a> \u003Cfile_b>` 命令，该工具比较两个图 JSON 文件，并在终端中使用颜色高亮（红色表示移除，绿色表示新增）显示节点和边的差异。这实现了从“运行时变异”到“离线\u002F代际分析”的转变。","https:\u002F\u002Fgithub.com\u002Faden-hive\u002Fhive\u002Fissues\u002F4352",{"id":162,"question_zh":163,"answer_zh":164,"source_url":165},42615,"Hive 是否支持 Zoom 会议管理功能？","是的，Hive 已通过 PR #5767 集成了 Zoom 工具。该集成不仅限于生成链接，而是提供了一个“会议智能工具层”，允许 Agent 执行以下操作：1. 调度会议（创建会议、获取加入链接）；2. 发现会议（列出会议、获取详情）；3. 知识摄入（获取会议录音转录稿以供总结和分析）。这使得 Hive 能够支持行政助理、销售自动化等需要处理同步人类协作的工作流。","https:\u002F\u002Fgithub.com\u002Faden-hive\u002Fhive\u002Fissues\u002F2867",{"id":167,"question_zh":168,"answer_zh":169,"source_url":170},42616,"是否有针对贡献者的运行时执行流程详细指南？","目前官方暂无计划编写专门的端到端运行时执行流程指南（涵盖从目标初始化到节点执行、评估及边遍历的详细过程）。虽然社区有提议撰写此类文档以解释 Runner 初始化、GraphExecutor 管理及 HITL 暂停恢复等细节，但维护者明确表示该特定议题不在当前规划中，因此相关 Issue 已关闭。建议贡献者参考现有的高层架构文档。","https:\u002F\u002Fgithub.com\u002Faden-hive\u002Fhive\u002Fissues\u002F5202",{"id":172,"question_zh":173,"answer_zh":174,"source_url":160},42617,"如何在没有权限分配 Issue 的情况下提交 PR？","如果您提交了 PR 但因无法自行分配 Issue 而导致机器人自动关闭 PR，请在评论中@维护者（如 @austin931114 或其他核心团队成员），请求他们手动将关联的 Issue 分配给您。一旦获得分配，通常就可以重新打开或继续推进您的 PR。此外，也可以在评论中明确说明您的实现方案（如切换到 CLI 工具而非运行时修改），以争取维护者的支持。",{"id":176,"question_zh":177,"answer_zh":178,"source_url":160},42618,"Hive 的“自演进”图拓扑变化如何被验证？","虽然可视化的 TUI 差异模式已被弃用，但“自演进”的核心机制（如 `EvolutionGuard`）已在代码中实现并经过本地验证。系统通过快照和差异对比逻辑来拦截和验证实时突变。如果团队倾向于不可变运行时架构，这些差异对比逻辑会被重构用于离线工作流（即通过 `hive diff` 命令比较不同代际的图文件），从而在不改变运行时状态的前提下分析演进效果。",[180,185,190,195,200,205,210,215,220,225,230,235,240,245,250,255,260,265,270,275],{"id":181,"version":182,"summary_zh":183,"released_at":184},342320,"v0.10.2","# 🐝 Hive Agent v0.10.2\n\n> 一款以浏览器自动化为核心的 **v0.10.1** 后续版本。现在，视觉模型与 Chrome 浏览器之间传递的坐标值采用的是视口的 **分数比例**，而非截图像素——因此，相同的 `(x, y)` 坐标可以在 Claude、GPT-4o、Gemini 等任何 VLM 中通用，无论它们如何调整图像大小或进行分块处理。此外，还修复了女王切换、标签页组隔离以及 CI 构建等方面的可靠性问题。\n\n---\n\n## ✨ 亮点\n\n- **与模型无关的视觉点击操作。** 所有需要坐标输入的浏览器工具（`browser_click_coordinate`、`browser_hover_coordinate`、`browser_press_at`）以及所有返回矩形区域的工具（`browser_get_rect`、`browser_shadow_query`，还有 `focused_element` 中的 `rect`），现在都使用视口的 `0..1` 分数比例来表示坐标。这样一来，当你切换后端时，视觉模型对像素尺寸的调整就不会再悄然导致点击失效。\n- **女王在切换配置文件或女王身份时仍能保持运行。** 切换女王身份不再会中断当前活跃女王的运行时环境。\n- **标签页组隔离。** 浏览器标签页组现已按每个配置文件命名空间化，因此当 Chrome 复用某个标签页 ID 时，旧的高亮显示或附加状态不会跨配置文件传播。\n- **远程浏览器调试器。** 新增的 `scripts\u002Fbrowser_remote.py` 脚本及 HTML 用户界面，为 Chrome 扩展桥提供了一个可视化调试界面——包括实时截图、坐标检查器，以及一键式 GCU 工具测试框架。\n- **更绿色的持续集成。** 解决了所有框架和工具测试中的失败问题，并且 Windows 平台的 CI 构建已恢复正常；整个代码库通过了 ruff 的完整 lint 和格式化检查。\n- **Gemini 可靠性调优。** `gemini-3-flash-customtools` 和 `gemini-3.1-pro-preview-customtools` 现在使用 `max_context_tokens: 240000`（之前为 90 万）——Gemini 在长上下文场景下的质量远低于其宣传的最大窗口范围，适当降低上限虽然牺牲了一些扩展性，却能让工具调用更加稳定可靠。\n\n---\n\n## 🆕 新功能\n\n### 浏览器自动化\n\n- **基于分数的坐标系统**——所有点击、悬停、按键及获取矩形区域的工具现在都使用视口的 `(0..1, 0..1)` 分数比例。这些工具内部会在将指令发送至 CDP 之前，先乘以缓存的 `cssWidth` 和 `cssHeight`。四位小数精度（`0.0001` ≈ 在 1717 宽度的视口中约 0.17 CSS 像素）足以应对最精细的目标。（@timothyadenhq）\n- **`browser_type_focused`——专用的聚焦元素输入工具**，从 `browser_type` 中拆分出来。在使用 `browser_click_coordinate` 将焦点定位到目标元素之后调用此工具，相比 `browser_press` 更适合多字符输入。（@RichardTang-Aden）\n- **多模式截图工具**——`browser_screenshot` 新增了视口、整页和选择器裁剪三种模式，并在元数据中返回 `cssWidth` 和 `cssHeight`，以便调用方在需要时能够根据视口尺寸进行推断。（@RichardTang-Aden）\n- **用于输入聚焦事件的虚线高亮标记**——在交互后的截图中，通过实线和虚线高亮来区分点击和输入聚焦的不同效果。（@RichardTang-Aden）\n- **默认 1 毫秒按键延迟 + 提示优化**——`browser_type` 现在默认使用 1 毫秒的延迟（之前更高），这与真实富文本编辑器的预期一致；相关编排器提示","2026-04-17T06:44:55",{"id":186,"version":187,"summary_zh":188,"released_at":189},342321,"v0.10.1","# 🐝 Hive Agent v0.10.1\n\n> 这是 **v0.10.0 The Colony** 的一个小版本更新——优化了女王体验，在长上下文中使代理循环更加可靠，并提升了浏览器自动化能力。无破坏性变更；v0.10.0 的会话仍可继续使用。\n\n---\n\n## ✨ 亮点\n\n- **女王私信更生动。** 聊天界面现在显示消息时间戳和日期分隔行，且在流式更新时 `createdAt` 保持稳定，因此新消息到达时不会抖动。\n- **女王个人资料一键直达。** 女王个人资料现已成为独立面板，可直接从应用头部打开——不仅在组织架构图页面可用，所有页面均可访问。\n- **侧边栏更智能。** 女王按最近的私信活动排序，这样你实际正在协作的对象会自动浮到顶部。“Head of”前缀已被去除，界面更加简洁。\n- **更平和、更精简的女王提示词。** 独立模式\u002FPM 模式的提示词经过大幅精简和重构，推理能力更强。\n- **值得信赖的上下文健康状态。** 修复了代理循环中的上下文跟踪、压缩以及工具结果处理机制——长时间会话能够保持健康，不再陷入驱逐循环。\n- **浏览器自动化升级。** 浏览器、LinkedIn 和 X 的自动化技能新增了更清晰的指导说明，底层 CDP 桥接在点击、截图和检查等操作路径上也更为稳健。\n\n---\n\n## 🆕 新增功能\n\n### 女王与聊天 UX\n\n- **私信中显示消息时间戳和日期分隔线** —— `ChatPanel` 现在会显示每条消息的时间，并按日期分组；同时在流式更新时也能保持稳定的 `createdAt`，避免消息重新排序。（@bryanadenhq）\n- **将 `QueenProfilePanel` 从 `org-chart` 中提取出来** —— 个人资料面板现已成为一个独立组件，可通过 `AppHeader` 打开，并在整个应用布局中全局可用。（@bryanadenhq）\n- **按最后私信活动对女王排序** —— `ColonyContext` 会根据女王最近的互动情况对她们进行排序，而 `SidebarQueenItem` 则去掉了“Head of”这一标题前缀。（@bryanadenhq）\n- **从最新消息推导 `last_active_at`** —— `session_manager` 现在会根据实际的消息流推导女王的活跃度，并按时间倒序排列历史记录，确保侧边栏始终与实际情况同步。（@bryanadenhq）\n- **女王独立提示词重构** —— `queen\u002Fnodes\u002F__init__.py` 文件从约 215 行缩减至约 95 行，独立模式和 PM 模式的提示词构建更加清晰，并新增了 `debug_queen_prompt.py` 脚本。（@RichardTang-Aden）\n- **财务女王头衔优化** —— 在 `queen_profiles.py` 中更新了 Charlotte 的头衔。（@RichardTang-Aden）\n\n### 代理循环与上下文健康\n\n- **上下文健康与驱逐问题修复** —— 对 `agent_loop.py`、`conversation.py`、`compaction.py`、`tool_result_handler.py` 以及 `internals\u002Ftypes.py` 进行了大规模重构，以保持长时间会话的稳定性。如今，上下文压缩、工具结果的统计以及驱逐决策都基于更准确的对话状态视图来进行。（@timothyadenhq）\n\n### 技能与工具\n\n- **浏览器自动化技能指南** —— 更新了 `browser-automation\u002FSKILL.md`，为执行相关任务的代理提供了更清晰的操作指引。","2026-04-16T01:23:09",{"id":191,"version":192,"summary_zh":193,"released_at":194},342322,"v0.10.0","# 🐝 Hive Agent v0.10.0：蜂群\n\n> ⚠️ **重大变更。** 这是对 Hive 中代理工作方式的一次大规模架构重构。**旧版代理不再兼容。** 现有的工作空间、自定义代理以及 v0.10.0 之前的版本中保存的会话都需要重新创建。\n\n---\n\n## ✨ 亮点\n\n“蜂群”引入了一种全新的工作模式：一组专业化的“工蜂”协同运作，以推动并扩展您的业务。\n\n“女王”的角色已发生演变。她不再仅仅负责协调，而是会首先**亲自执行任务**以快速交付价值，随后再围绕这些任务**构建系统**，从而形成稳定且可重复的业务流程。\n\n如今，您拥有一支由八位女王组成的完整领导团队，每位女王都具备独特的身份、专长和风格：\n\n| 女王 | 角色 |\n| --- | --- |\n| **Sophia** | 品牌与设计负责人 |\n| **Charlotte** | 财务与融资负责人 |\n| **Victoria** | 增长负责人 |\n| **Eleanor** | 法务负责人 |\n| **Rachel** | 运营负责人 |\n| **Isabella** | 产品战略负责人 |\n| **Amelia** | 人才负责人 |\n| **Alexandra** | 技术负责人 |\n\n立即与您的女王们一起开启业务流程自动化吧。\n\n---\n\n## 🏛️ 蜂群架构\n\n### 女王作为独立个体，而不仅是协调者\n\n- **女王档案** — 每位女王都是一个基于 YAML 的角色设定文件（`~\u002F.hive\u002Fagents\u002Fqueens\u002F{queen_id}\u002Fprofile.yaml`），包含核心特质、隐藏背景、心理画像、行为触发机制及技能组合。这些档案会在会话开始时注入到系统提示中。\n- **CEO 式女王选择** — 通过 LLM 分类器，根据具体任务将每个新用户请求路由至最匹配的女王，并提供结构化的路由诊断信息（`QueenSelection`）。\n- **女王私信界面** — 为每位女王配备独立的私信页面，内置专属会话流程、会话切换器及提示库集成。\n- **独立模式 \u002F PM 模式** — 女王可在规划阶段以独立模式运行，其“自言自语”的思考过程会通过内部标签呈现出来。\n- **女王记忆 v2** — 简化后的记忆实现，配备反思智能体、冷却时间限制的反思机制、用户身份识别、替身机制以及用于精准检索的回忆选择器。\n- **女王生命周期工具** — 提供一流的升级、女王回复及会话交接工具。\n\n### 蜂群运行时\n\n- **整体架构大改版** — 框架、代理循环、运行时、图谱、流水线、执行器及节点工作进程等各层代码均从零重写。废弃的适配层和遗留的协调路径已被移除。\n- **蜂群创建流程** — 蜂群可通过技能创建，支持可靠的事件总线订阅、工作进程孵化以及创建后的列表刷新。\n- **定时触发器** — 蜂群现在可以按 Cron 定时唤醒，触发器会直接进入所属女王的会话。\n- **代理的简单分叉机制**、稳定的凭证状态以及更可靠的进程执行能力。\n\n---\n\n## 🆕 新功能\n\n### 蜂群与女王\n\n- 8 个默认女王角色","2026-04-15T03:01:59",{"id":196,"version":197,"summary_zh":198,"released_at":199},342323,"v0.9.0","Chrome 扩展替代了 Playwright， episodic queen memory 升级为智能回忆系统，图执行器被重写以支持并行扇出，并新增了一个真实的 LLM 集成测试套件。共修改了 232 个文件，在 116 次提交中增加了 3.32 万行、删除了 1.29 万行代码。\n\n## 亮点：浏览器扩展与桥接架构\n\n基于 Playwright 的浏览器栈已被移除。现在，代理通过 WebSocket 桥接和未打包的 MV3 Chrome 扩展来控制用户现有的 Chrome 浏览器。不再为每个代理单独启动独立的 Chrome 进程——它们直接继承用户的登录状态、Cookie 和已安装的扩展。\n\n**Chrome 扩展**（`tools\u002Fbrowser-extension\u002F`）运行一个后台服务工作线程，通过离屏文档分发命令（如 `context.create`、`tab.create`、`cdp.attach` 等），以实现持久的 WebSocket 连接。每个代理都会获得一个颜色编码的 `chrome.tabGroup`，直接显示在浏览器界面中。\n\n在 Python 端，**`BeelineBridge`**（`bridge.py`）在端口 9229 上托管一个 WebSocket 服务器，在端口 9230 上提供 HTTP 状态接口。它负责 Page\u002FDOM\u002FInput\u002FRuntime 等域的 CDP 直通，管理标签组，支持可配置的导航等待条件，并通过 CDP Overlay 渲染视觉高亮效果（蓝色矩形、红色十字准线）。\n\n所有浏览器工具（如 `click`、`type`、`navigate`、`screenshot` 等）现在都与桥接层通信，而非直接调用 Playwright。旧的 `session.py` 文件从 996 行缩减至 67 行——仅保留了一个用于活跃配置文件上下文变量的存根。\n\n### 设置步骤\n\n```\n1. 打开 chrome:\u002F\u002Fextensions\u002F\n2. 启用“开发者模式”\n3. 点击“加载已解压的扩展程序” → 选择 tools\u002Fbrowser-extension\u002F\n```\n\n---\n\n## 亮点：Queen Memory v2\n\n单体式的跨会话记忆模型已被一种**细粒度的 episodic 系统**所取代。每个 `.md` 文件都存储在 `~\u002F.hive\u002Fqueen\u002Fmemories\u002F` 目录下，带有 YAML 前置元数据，并分为五种类型：目标、环境、技术、参考和日记。每份文件限制在 4 KB 以内，总文件数不超过 200 个，从而保持存储的轻量化。\n\n### 回忆选择器\n\n在每一轮行动之前，一个轻量级的选择器会在一次 LLM 调用中扫描记忆标题，挑选出约 5 条最相关的记忆注入到上下文中。它仅接收当前用户查询（不包含完整对话历史），以降低计算成本；对于超过 1 天的条目，会添加过时警告；若发生错误，则会优雅地退化为空列表。\n\n### 反思代理\n\n该代理会在每次 Queen 行动后异步运行，采用基于游标的增量处理方式。**短周期反思**每轮都会进行（批量读取后再写入）。而**长周期反思**则每 5 次短周期反思、在 `CONTEXT_COMPACTED` 事件触发时以及会话结束时执行——进行整体性的重组、去重和整合。每日的日记记录会在每个周期结束后写入 `MEMORY-YYYY-MM-DD.md` 文件中。\n\nv1 版本的记忆模板还增加了对沟通情况的明确记录——包括技术深度、交流节奏和语气偏好等信息，从最初的表面化描述（第 1 天）逐渐演变为更细腻的表达（第 5 天及以后）。\n\n---\n\n## 亮点：思考标记渲染\n\nGemini 系列模型会输出结构化的思维","2026-04-04T04:18:10",{"id":201,"version":202,"summary_zh":203,"released_at":204},342324,"v0.8.0","技能 CLI 提供安装、卸载和验证功能，MCP 注册表支持代理级别的服务器选择，新增用于覆盖默认技能配置的应用程序，并进行了一轮安全修复。共修改 147 个文件，涉及 120 次提交，新增 16,800 行代码，删除 15,200 行代码。\n\n## 亮点：代理技能\n\nv0.8.0 完成了在 v0.7.4 中引入的代理技能系统。技能是以 SKILL.md 格式打包的组件，可在运行时将操作协议（如记录笔记、批量跟踪、错误恢复等）注入到代理的系统提示中。\n\n### 安装技能\n\n目前尚无公开的技能注册表。要安装技能，可从 Git URL 克隆：\n\n```bash\nhive skill install --from https:\u002F\u002Fgithub.com\u002Fsomeone\u002Fmy-skill.git\n\n# 锁定特定分支或标签\nhive skill install --from https:\u002F\u002Fgithub.com\u002Fsomeone\u002Fmy-skill.git --version v2.0\n\n# 覆盖本地目录名称\nhive skill install --from https:\u002F\u002Fgithub.com\u002Fsomeone\u002Fmy-skill.git --name custom-name\n```\n\n技能会被安装到 `~\u002F.hive\u002Fskills\u002F\u003Cname>\u002F` 目录下，并在代理启动时自动发现。\n\n### 管理技能\n\n```bash\nhive skill list                # 列出已安装的技能（项目级、用户级和框架级）\nhive skill info \u003Cname>         # 显示详细信息、脚本和引用\nhive skill validate \u003Cpath>     # 用于 CI 和贡献者的严格验证\nhive skill remove \u003Cname>       # 卸载技能\n```\n\n### 编写 SKILL.md\n\n一个最小化的技能示例如下：\n\n```markdown\n---\nname: my-skill\ndescription: 这个技能的作用\nlicense: MIT\ncompatibility:\n  - hive\nallowed-tools:\n  - Bash(curl:*)\n---\n\n## 协议\n\n在此处填写您的操作指令。\n```\n\n将此文件放置于 `~\u002F.hive\u002Fskills\u002Fmy-skill\u002FSKILL.md`，它将被自动加载。\n\n---\n\n## 功能特性\n\n**技能 CLI** (#6782 — @levxn)\n\n- `hive skill install` — 从 Git URL 安装技能，支持 `--version` 和 `--name` 参数覆盖\n- `hive skill install --pack \u003Cname>` — 安装入门包（一组技能的集合）\n- `hive skill remove`、`hive skill list`、`hive skill info`、`hive skill validate`\n- `hive skill doctor` — 诊断常见的技能问题\n- 严格的验证流程（检查 1–12），涵盖前言、正文、脚本、允许工具、命名规范等\n\n**技能配置覆盖与运行时启发式** (#6610 — @levxn)\n\n- SKILL.md 正文中支持 `{{placeholder}}` 替换 — 来自 `default_skills` 代理配置的覆盖项现在会在运行时生效（此前仅解析但被静默丢弃）\n- **DS-12**：`hive.batch-ledger` 可根据目标或输入文本自动检测批量场景，并在系统提示前插入一条账本初始化的提醒\n- **DS-13**：`hive.context-preservation` 会监控令牌使用量，当使用比例超过 `warn_at_usage_ratio`（默认 0.45）时，会在框架的 0.6 截断阈值之前发出一次性的上下文保存警告\n\n**MCP 注册表与代理选择** (#6574 — @Antiarin，#6792 — @fermano)\n\n- `MCPRegistry` 核心模块 — 可从中央注册表安装、列出、健康检查并管理 MCP 服务器\n- 每个代理目录下的 `mcp_registry.json` 文件 — 提供声明式的服务器选择功能，支持 `include`、`tags`、`exclude`、`profile`、`max_tools` 和 `versions` 等配置","2026-04-01T02:23:25",{"id":206,"version":207,"summary_zh":208,"released_at":209},342325,"v0.7.6","图像功能端到端支持（上传、截图直通、视觉检测、模型回退）、Antigravity 原生 Google OAuth、结构化技能错误码、符号链接沙箱安全修复以及 PDF URL 支持。仅图像功能相关就有 28 个文件变更；整个技能错误系统新增 8 个文件。\n\n## 功能特性\n\n**图像功能**（PR #6682）\n\n- **用户图片上传** — 在 UI 中可直接将图片附加到聊天消息中；以图片内容块形式转发至 LLM（关闭 #6579）\n- **GCU 截图直通** — 浏览器工具的截图会连同文本结果一起作为图像内容块转发给具备视觉能力的 LLM，从而在网页自动化过程中实现视觉推理（关闭 #6678）\n- **视觉能力检测** — `core\u002Fframework\u002Fllm\u002F` 目录下新增 `capabilities.py` 模块，通过基于提供商和模型名称的分层允许\u002F拒绝规则，判断模型是否支持图像处理；涵盖 ZAI、MiniMax、DeepSeek、Cerebras、Groq 以及本地运行器（Ollama、LM Studio、vLLM、llama.cpp）（关闭 #6679）\n- **自动剥离图像内容** — 在调用非视觉模型之前，会先剥离图像内容块，避免 API 错误（关闭 #6679）\n- **视觉回退机制** — 当主模型不具备视觉能力时，系统会使用具备视觉能力的备用模型将图像描述为文本，并将该描述注入到对话中，使纯文本模型仍能获得视觉上下文信息（关闭 #6680）\n- **Aria 引用系统** — `refs.py` 会在 `aria_snapshot()` 的输出中为每个交互元素添加 `[ref=eN]` 标记；LLM 现在可以通过稳定的引用 ID（如 `selector=\"e5\"`）来定位浏览器元素，而无需依赖易失效的 CSS 选择器（关闭 #6681）\n\n**Antigravity \u002F 订阅**\n\n- 将原生 Google OAuth 流程提取至 `core\u002Fantigravity_auth.py` — 实现与运行器启动流程的更清晰分离，支持正确的凭据缓存，并在重新认证前验证现有凭据\n- 修复 Antigravity 模型工具调用 bug：函数调用部分现已包含 `thought_signature` 字段，以符合模型预期格式，从而解决此前导致 400 错误的问题\n- OAuth 客户端 ID 和密钥现在可通过所有回退路径正确解析\n\n**结构化技能错误码**（关闭 #6366）\n\n- 新增 `skill_errors.py`，定义了带类型的 `SkillError` 错误层级结构（加载、解析、验证、执行等类别）\n- 技能解析器、目录和管理器现在会抛出结构化的错误码及诊断信息，而非原始异常\n- 在 `test_skill_errors.py` 中实现了全面的测试覆盖\n\n## Bug 修复\n\n**安全：符号链接沙箱逃逸**（关闭 #1167）\n\n- 文件系统工具包中的 `get_secure_path` 现在会在进行路径前缀检查之前先解析符号链接，从而防止通过符号链接绕过沙箱边界的情况发生\n\n**图谱清理**\n\n- 移除了图谱中从未实现且只会增加执行日志噪音的空 `check_constraint` 占位符\n\n**工具**\n\n- `pdf_read` 工具现不仅支持本地路径，还支持 HTTP\u002FHTTPS URL — 会先下载到临时文件，验证 `Content-Type: application\u002Fpdf`，并在使用后自动清理；至此完成了 `web_scrape → pdf_read` 的完整链路","2026-03-21T04:27:22",{"id":211,"version":212,"summary_zh":213,"released_at":214},342326,"v0.7.5","v0.7.5 — 并行子代理显示、会话恢复与稳定性修复\n\n修复了六个相互关联的 bug，分别影响并行子代理执行、会话恢复完整性和 GCU 浏览器子代理的终止问题。新增了类似 tmux 的并行代理显示功能。共修改 9 个文件，增加 636 行、删除 25 行。\n\n## 功能特性\n\n**并行子代理 tmux 显示** (#6652)\n- 新增 `ParallelSubagentBubble` 组件，将并发运行的子代理以类似 tmux 的终端多路复用器形式呈现，而非分散的独立消息气泡。\n- 每个子代理拥有独立的窗格，包含标题栏（显示代理名称、消息数量及上下文使用条）、可滚动的消息正文（支持 Markdown 渲染）以及闪烁光标。\n- 窗格焦点跟踪：最后活跃的窗格会显示彩色边框，其他窗格则变暗。\n- 单击任意窗格标题栏即可将其放大至全宽；再次单击则恢复原状。\n- 已完成的子代理会显示绿色对勾、窗格变暗且无光标。\n- 头部徽章实时显示计数：“3 个正在运行” → “2 个正在运行” → “5 个已完成”。\n- 智能分组：子代理消息会跨交错的女王节点\u002F系统消息进行分组，仅在硬边界处断开（如用户消息、运行分隔符或下一阶段工作节点消息）。\n\n## Bug 修复\n\n**节点调用其不拥有的工具（连续模式）** (#6653)\n- 在连续对话模式下，所有节点共享同一段对话。当流程图循环时，LLM 会看到来自前一个节点历史记录中的 `delegate_to_sub_agent` 调用，并在当前节点上重放这些调用，而该节点并不具备 `sub_agents`。\n- 在工具调度逻辑中添加了一道检查机制，确保只有在当前节点确实提供了 `delegate_to_sub_agent` 选项时，才会接受该调用。\n- 证据：事件日志显示，“start” 节点（仅拥有 `google_sheets_get_values` 工具）曾三次调用 `delegate_to_sub_agent`。\n\n**无法区分的子代理实例** (#6653)\n- 当多次委托给同一种子代理类型时（例如 3 次委托给 `browser-researcher`），所有实例会共享相同的 `node_id` 和 `stream_id`，导致在事件流中难以区分。\n- 将此前仅用于文件系统路径的实例计数嵌入到 `node_id` 中：第一个实例保持原有格式，后续实例则在末尾添加 `:N` 后缀。\n- 现在事件日志会显示明确的标识符，如 `batch-orchestrator:subagent:browser-researcher`, `:2`, `:3` 等。\n\n**会话恢复时前言提示丢失** (#6653)\n- `EXECUTION_SCOPE_PREAMBLE` 和 `GCU_BROWSER_SYSTEM_PROMPT` 仅在全新会话的代码路径中注入；而在会话恢复路径中，则调用了 `compose_system_prompt()`，该函数并未预留这些前言提示的位置。\n- 向 `compose_system_prompt()` 添加了 `execution_preamble` 和 `node_type_preamble` 参数。\n- 恢复路径现在会应用与全新路径相同的前言条件。\n\n**阻塞内存整合** (#6653)\n- `CONTEXT_COMPACTED` 事件处理器曾 `await` `consolidate_queen_memory()`，从而阻塞事件总线，直到 LLM 的调用完成。\n- 改为使用“发后即忘”的 `asyncio.create_task()`，与清理路径保持一致。\n- 过滤为仅针对女王节点的整合——此前所有整合都会触发（单次会话中多达 17 次，包括","2026-03-20T03:23:51",{"id":216,"version":217,"summary_zh":218,"released_at":219},342327,"v0.7.4","v0.7.4 — 代理技能、工作节点模型与 MCP 传输协议\n\n这是一次重大发布，引入了带有信任门控的代理技能系统，新增对独立工作节点 LLM 模型的支持，扩展了 MCP 传输选项，恢复了 Claude Code 订阅认证，并将 OpenRouter 作为一级提供商纳入其中。本次更新共涉及 66 个文件，新增 7,636 行代码，删除 570 行代码。\n\n## 功能特性\n\n**代理技能系统** (#6624, #6625, #6626, #6362, #6363)\n- **信任门控 (AS-13)** — 来自不受信任仓库的项目级技能，在加载到代理系统提示词之前，现在需要用户明确同意。提供会话仅一次、永久授权和拒绝三种交互式同意选项。信任状态持久化存储于 `~\u002F.hive\u002Ftrusted_repos.json`\n- **技能 CLI** — `hive skill list` 可发现所有作用域（项目、用户、框架）中的技能；`hive skill trust \u003Cpath>` 可永久信任某个项目仓库\n- **第三层资源加载 (AS-6)** — 在目录清单 XML 中暴露 `base_dir`，技能目录通过 `SkillsManager` → `AgentRuntime` → `ExecutionStream` → `GraphExecutor` → `NodeContext` 进行传递\n- **目录白名单机制 (AS-9)** — 文件读取工具（`view_file`、`load_data`、`read_file`）在 `EventLoopNode._execute_tool()` 中被拦截，以从已激活的技能目录中提供文件，绕过会话沙箱\n- **上下文保护 (AS-10)** — 在 `ToolResult` 和 `Message` 中添加 `is_skill_content` 标志，防止在令牌使用率达到 60% 时因上下文修剪而丢弃技能指令\n- **技能数据模型** — 提供 `SkillEntry`、`SkillScope`、`TrustStatus` 等结构化内存中技能跟踪模型\n- **技能用户指南** — 完整的终端用户文档位于 `docs\u002Fskills-user-guide.md`\n\n**独立工作节点 LLM 模型** (#6614, #6615)\n- 在 `~\u002F.hive\u002Fconfiguration.json` 中新增 `worker_llm` 部分，允许用户为工作节点代理配置不同的（更便宜或更快）模型\n- 提供工作节点模型、API 密钥、基础 URL、额外参数及令牌限制的配置辅助工具\n- 会话管理器会在启动工作节点代理时注入特定于工作节点的身份凭证\n- 提供交互式设置脚本：`scripts\u002Fsetup_worker_model.sh` 和 `scripts\u002Fsetup_worker_model.ps1`\n- 快速入门脚本现已提及工作节点模型的设置选项\n\n**MCP Unix 域套接字与 SSE 传输协议** (#6347, #6531)\n- 通过 `httpx.HTTPTransport(uds=...)` 实现 Unix 域套接字传输\n- 使用 MCP Python SDK 的会话生命周期实现 SSE 传输\n- 针对 Unix 和 SSE 的临时连接失败，实现了单次重连逻辑\n- 对所有新引入的传输路径和重试路径进行了重点单元测试覆盖\n\n**MCP 共享连接管理器** (#6348, #6534)\n- 全进程范围的 `MCPConnectionManager`，采用引用计数进行获取与释放\n- 包含健康检查、自动重连、并发访问及清理功能\n- 默认启用连接复用，从而降低多代理工作流中 MCP 服务器的启动开销\n- 在连接复用时记录包含引用计数的调试日志\n\n**实时上下文窗口使用情况显示** (#6627)\n- 工作区 UI 中为女王代理和工作节点代理分别显示实时上下文窗口利用率条形图\n- 采用颜色编码阈值：绿色\u002F金色 \u003C 70%，橙色 70–90%，红色 > 90%\n- 鼠标悬停时会展","2026-03-19T03:35:14",{"id":221,"version":222,"summary_zh":223,"released_at":224},342328,"v0.7.3","v0.7.3 — 数据完整性、工作节点摘要与冷恢复\n\n这是一次聚焦的发布，修复了代理执行流水线中的多条数据丢失路径，引入了用于女王可观测性的持久化工作节点运行摘要，并增强了冷会话恢复功能。共修改48个文件，新增4,225行，删除1,208行。\n\n## 功能特性\n\n**工作节点每轮运行摘要** (#6588, #6589)\n- 新增 `worker_memory.py` 模块，为每次工作节点执行生成由大模型总结的摘要。\n- 存储路径：`~\u002F.hive\u002Fagents\u002F{name}\u002Fruns\u002F{run_id}\u002Fdigest.md` — 摘要可在不同会话间持久保存。\n- 会话管理器订阅 `EXECUTION_COMPLETED`、`EXECUTION_FAILED`、`EXECUTION_PAUSED` 和 `TOOL_CALL_COMPLETED`（用于 `delegate_to_sub_agent`）事件。\n- 运行中快照设置为每5分钟一次；最终摘要总会在完成或失败时写入。\n- 摘要会通过 `[WORKER_DIGEST]` 事件自动注入到女王对话中。\n- 新增 `get_worker_status(focus='diary')` 函数，可从磁盘读取持久化摘要，无需工作节点运行时环境。\n- 关联 #6550。\n\n**工作节点执行范围前言** (#6568)\n- 新增 `EXECUTION_SCOPE_PREAMBLE` 守护机制，注入到工作节点系统提示中。\n- 提示大模型它只是多步图中的一个节点，应仅执行其分配的任务。\n- 在新对话创建及连续模式阶段转换时应用。\n\n**冷会话恢复加固** (#6576)\n- 冷恢复失败时回退至仅女王会话（代理代码不完整）。\n- 阶段感知恢复：读取 `meta.json` 中的阶段信息，以决定是加载工作节点还是继续构建。\n- 创建女王会话时不会覆盖 `meta.json` 中的字段——保留 `phase` 和 `agent_path`。\n- 恢复时按阶段跟踪并恢复会话，使用 `switch_to_staging` 或 `switch_to_building` 方法。\n\n**SDR 代理模板** (#5178)\n- 新建6节点AI流水线：`摄入 → 评分 → 过滤 → 个性化 → 发送触达 → 报告`。\n- 安全优先：仅创建Gmail草稿（从不发送），跳过 `risk_score >= 7` 的联系人。\n- 提供完整的图定义，包括 `agent.json`、`flowchart.json`、各节点实现及配置。\n\n**DraftGraph 中触发节点渲染** (#6545)\n- 在 `DraftGraph.tsx` 中恢复触发节点渲染，并支持SVG形状。\n- 通过触发更新API支持Cron表达式编辑。\n- 保存时实时更新触发药丸和详情面板。\n- 确保用户请求移除调度器时，女王会调用 `remove_trigger`。\n\n---\n\n## 错误修复\n\n**数据披露与完整性** (#6589)\n- **恢复时自动溢出绕过** — 将溢出逻辑移至 `OutputAccumulator.set()`，使其在所有代码路径上（全新、恢复、还原）均触发。修复 #6580。\n- **小型工具结果修剪** — 跳过长度小于100字符的结果修剪，防止大模型重试已成功的 `set_output` 调用。修复 #6581。\n- **子代理结果截断** — 将 `delegate_to_sub_agent` 的结果经由 `_truncate_tool_result` 处理，使大型输出能避免被修剪。修复 #6582。\n- **过渡标记数据丢失** — 在 `build_transition_marker` 中将大值自动溢出至数据文件，而非直接在代码中截断至300字符。修复 #6583。\n- **路径 d","2026-03-18T03:51:14",{"id":226,"version":227,"summary_zh":228,"released_at":229},342329,"v0.7.2","## 变更内容\n\u003Cimg width=\"2707\" height=\"1476\" alt=\"2026年3月16日 20:26:05 截图\" src=\"https:\u002F\u002Fgithub.com\u002Fuser-attachments\u002Fassets\u002Fdbd2ddf8-d667-482f-9954-5a5e859dbb02\" \u002F>\n\n### 功能特性\n- **智能体技能系统** — 新的技能框架，包含发现、目录、解析器、管理器，以及6个默认技能（批量分类账、上下文保持、错误恢复、笔记记录、质量监控、任务分解）\n- **节点间文件通信** — 基于文件路径的图节点间数据传递\n- **Hive LLM 支持** — 在快速入门中新增 Hive LLM 端点配置、验证及 baseURL 支持\n- **流程图工具** — 将流程图分类与重新映射提取到专用模块，并提供回退生成功能；更新架构至9种类型，新增深色主题配色方案\n- **运行按钮组件** — 前端工作区新增运行按钮\n- **HuggingFace 工具** — 推理、嵌入和端点工具 (#6132)\n- **Notion 工具增强** — 新增区块类型，改进页面创建、文档和测试\n\n### 优化改进\n- 图表面板可调整大小；移除旧版 `AgentGraph` 组件\n- 加载智能体时添加过渡动画\n- 屏蔽冗余的 LiteLLM INFO 日志\n- 在并行执行中强制实施分支超时和内存冲突策略 (#6504)\n- 简化开发者代理日志的终端输出 (#6388)\n- 增强命令验证的命令净化模块 (#6217)\n- 原子化的 `state.json` 进度写入\n- Windows 兼容性修复（字符串操作、快速入门文档）\n\n### 错误修复\n- 修复使用现有密钥更换模型时的配置写入问题\n- 修复技能内存键导致无限制节点权限失效的问题\n- 修复技能注入和工具调用超时问题\n- 修复技能生命周期与运行时的绑定问题\n- 修复“蜂后”多响应错误 (#5962)\n- 修复所有阶段中的情景记忆访问问题\n- 修复 Google 表格凭据孤儿处理问题 (#6385)\n- 修复 LLMJudge OpenAI 回退使用 LiteLLM 提供者的问题 (#5674)\n- 在运行时日志中保留自定义会话 ID (#6241)\n- 修复删除 `aden_api_key` 时未找到返回404的问题 (#6340)\n- 清理 `~\u002F.hive\u002Ffailed_requests\u002F` 目录以防止磁盘空间无限制增长 (#5725)\n- 当 litellm monkey-patching 因 ImportError 失败时发出警告 (#5757)\n- 声明 croniter 依赖项 (#6405)\n- 为从磁盘加载的智能体返回暂存阶段，以防止规划加载器误判\n\n### 杂项任务\n- 移除 `core\u002Fdemos\u002F`（已废弃的演示脚本）\n- 代码检查与格式化修复\n- 新增10个流程图模板（竞争情报、深度研究、邮件收件箱、邮件回复、求职者、本地企业提取、会议安排、科技新闻报道、Twitter 新闻、漏洞评估）\n- 更新文档（Windows 快速入门、流程图架构、Notion 工具 README）\n\n**完整变更日志**：https:\u002F\u002Fgithub.com\u002Faden-hive\u002Fhive\u002Fcompare\u002Fv0.7.1...v0.7.2\n\n## 感谢\n\n感谢所有为本次发布做出贡献的开发者！\n\n@Antiarin, @AryanNandanwar, @Gowthamtadikamalla, @Hundao, @MILTONADINA, @Mactavish28, @Nupreeth, @Sri-Likhita-adru, @Vee-27, @Waryjust","2026-03-17T03:46:27",{"id":231,"version":232,"summary_zh":233,"released_at":234},342330,"v0.7.1","## What's Changed\n* fix: generate worker mcp.json correctly in initialize_agent_package by @RichardTang-Aden in https:\u002F\u002Fgithub.com\u002Faden-hive\u002Fhive\u002Fpull\u002F6331\n* Fix: new agent resume and GCU browser improvements by @RichardTang-Aden in https:\u002F\u002Fgithub.com\u002Faden-hive\u002Fhive\u002Fpull\u002F6333\n* Add Level 2 dummy agent end-to-end tests by @RichardTang-Aden in https:\u002F\u002Fgithub.com\u002Faden-hive\u002Fhive\u002Fpull\u002F6342\n* feat: GCU browser cleanup, draft loading state, and inner_turn message fix by @RichardTang-Aden in https:\u002F\u002Fgithub.com\u002Faden-hive\u002Fhive\u002Fpull\u002F6345\n* fix: save json path for the new agent update meta.json when loaded worker by @RichardTang-Aden in https:\u002F\u002Fgithub.com\u002Faden-hive\u002Fhive\u002Fpull\u002F6346\n* bugFix: micro-fix add tab UI by @prasoonmhwr in https:\u002F\u002Fgithub.com\u002Faden-hive\u002Fhive\u002Fpull\u002F6313\n\n## New Contributors\n* @prasoonmhwr made their first contribution in https:\u002F\u002Fgithub.com\u002Faden-hive\u002Fhive\u002Fpull\u002F6313\n\n**Full Changelog**: https:\u002F\u002Fgithub.com\u002Faden-hive\u002Fhive\u002Fcompare\u002Fv0.7.0...v0.7.1","2026-03-14T03:42:45",{"id":236,"version":237,"summary_zh":238,"released_at":239},342331,"v0.7.0","v0.7.0 — Session Refactor, Flowchart Planning & Triggers 🐝 \r\n\r\nA major release that overhauls session management, introduces visual flowchart-based planning for the queen orchestrator, and adds an agent trigger system. 125 files changed, +13,719 \u002F -2,812 lines.\r\n\u003Cimg width=\"2549\" height=\"1174\" alt=\"Screenshot 2026-03-12 at 9 27 36 PM\" src=\"https:\u002F\u002Fgithub.com\u002Fuser-attachments\u002Fassets\u002F7c7d30fa-9ceb-4c23-95af-b1caa405547d\" \u002F>\r\n\r\n## Features\r\n\r\n**Flowchart-Based Planning Phase** (#6282)\r\n- Queen now designs agents as ISO 5807 flowcharts during planning via `save_agent_draft` \u002F `confirm_and_build` tools\r\n- Draft graphs support ~30 flowchart symbol types (start, terminal, process, decision, io, document, subprocess, browser, etc.)\r\n- Dissolution engine converts decision nodes into predecessor `success_criteria` and sub-agent\u002Fbrowser nodes into parent `sub_agents` lists\r\n- BFS validation ensures no orphaned nodes; decision nodes cannot hold subagents\r\n- Original pre-dissolution draft preserved as `original_draft_graph` on `QueenPhaseState` and persisted to `flowchart.json`\r\n- Runtime-to-flowchart node mapping via `flowchart_map` for UI rendering\r\n- Replanning supported: `replan_agent` tool and `save_agent_draft` available during building phase\r\n\r\n**Agent Trigger System** (#6052)\r\n- New `TriggerDefinition` dataclass in `core\u002Fframework\u002Fruntime\u002Ftriggers.py` with timer and webhook trigger types\r\n- Queen lifecycle tools: `set_trigger`, `remove_trigger`, `list_triggers` for runtime trigger management\r\n- Triggers loaded from `triggers.json` in agent export directory at worker load time\r\n- Triggers persisted to `SessionState.active_triggers` and `SessionState.trigger_tasks` for cold-restore\r\n- `TriggerEvent` dataclass and `inject_trigger()` method on `EventLoopNode`\r\n- New event types: `TRIGGER_AVAILABLE`, `TRIGGER_ACTIVATED`, `TRIGGER_DEACTIVATED`, `TRIGGER_FIRED`, `TRIGGER_REMOVED`\r\n\r\n**Multi-Question Planning**\r\n- New `ask_user_multiple` synthetic tool on `EventLoopNode` — queen-only, presents 2+ questions in a single UI interaction\r\n- Frontend: new `MultiQuestionWidget` component (215 lines) with individual option selectors and free-text inputs\r\n- Deferred `tool_call_completed` emission — waits for user response before emitting completion event\r\n\r\n**Cold Session Resume from Event Logs**\r\n- `EventBus.set_session_log()` persists all events to `events.jsonl` in queen session directory\r\n- Streaming delta coalescing: `client_output_delta`, `llm_text_delta`, `llm_reasoning_delta` accumulated and flushed as single snapshots on `llm_turn_complete`\r\n- `iteration_offset` prevents frontend message ID collisions between original and resumed runs\r\n- Auto-load worker on cold restore: reads `meta.json` from queen session dir, loads agent, switches to staging phase\r\n- New API endpoint: `GET \u002Fapi\u002Fsessions\u002F{session_id}\u002Fevents\u002Fhistory` for full replay\r\n\r\n**DraftGraph Frontend Component**\r\n- New `DraftGraph.tsx` (1,187 lines) — full flowchart renderer with SVG shape rendering, bezier edge curves, labeled arrows (Yes\u002FNo for decisions)\r\n- Interactive: click nodes to see tools, inputs, outputs, success criteria\r\n- Auto-layout algorithm for flowchart positioning\r\n\r\n**Workspace Page Rewrite**\r\n- Phase-aware UI: shows `DraftGraph` during planning, `AgentGraph` during running\r\n- Pause\u002Frun button for session control\r\n- Event history replay on cold session restore via `\u002Fevents\u002Fhistory` endpoint\r\n\r\n**Progressive Tool Discovery**\r\n- `list_agent_tools()` reworked with 4-level progressive disclosure: `summary` → `names` → `simple` → `full`\r\n- New filters: `credentials` (\"available\"\u002F\"all\"), `service` (filter by service within provider)\r\n- ~18x efficiency improvement by avoiding full schema dump on first call\r\n\r\n**Other Features**\r\n- Unique run ID tracking: `run_id` field on `AgentEvent`, propagated through execution lifecycle\r\n- Run count tracking: `AgentEntry.run_count` via `_count_runs()` scanning `runs.jsonl`\r\n- Cached token handling through entire pipeline; `cached_tokens` on `FinishEvent` and `LLM_TURN_COMPLETE` events\r\n- `max_context_tokens` replaces `max_history_tokens` with resolution priority: agent config → global config → hardcoded 32000\r\n- CSS variable theming for agent graph components with trigger countdown color states\r\n- Session creation honors explicit `session_id` for cold-restore continuity (#6233)\r\n- MoonShot AI Kimi LLM provider support (#6165)\r\n- Google OAuth credential alignment for Google Docs tool (#6230, #6075)\r\n- Windows atomic file writes via `ReplaceFileW` with DACL restoration (#5849)\r\n\r\n---\r\n\r\n## New API Endpoints\r\n\r\n| Method | Path | Description |\r\n|--------|------|-------------|\r\n| `GET` | `\u002Fapi\u002Fsessions\u002F{session_id}\u002Fevents\u002Fhistory` | Persisted event log for replay |\r\n| `PATCH` | `\u002Fapi\u002Fsessions\u002F{session_id}\u002Ftriggers\u002F{trigger_id}` | Update trigger task text |\r\n| `GET` | `\u002Fapi\u002Fsessions\u002F{session_id}\u002Fdraft-graph` | Current planning-phase draft graph |\r\n| `GET` | `\u002Fapi\u002Fsessions\u002F{session_id}\u002Fflowchart-map` | Runtime-to-flowch","2026-03-13T04:24:34",{"id":241,"version":242,"summary_zh":243,"released_at":244},342332,"v0.6.7","A patch release for the problems in the release 0.6.6 and some features\u002Frefactor to make the contribution easier\r\n\r\n## Bug Fixes\r\n\r\n- Add KIMI LLM provider support\r\n- Fix subagent `_EscalationReceiver` causing reply stalls (#6159)\r\n- Fix queen agent system prompt hooks not applying correctly\r\n- Fix `SessionManager._cleanup_stale_active_sessions` indiscriminately cancelling healthy concurrent agent sessions (#6081)\r\n- Fix MCP errlog file handle resource leak (#6094)\r\n- Fix server using `session.mode_state` instead of `session.phase_state` in `handle_pause` (#6069)\r\n- Fix subagent judge passing no feedback on evaluated ACCEPT verdicts (#6108)\r\n- Fix test assertions for newly added tools\r\n- Track `reported_to_parent` to prevent false empty-turn detection (#6102)\r\n\r\n---\r\n\r\n## Refactoring\r\n\r\n- Remove deprecated codes(including TUI) and AI-generated temp doc\r\n- Add minimax key check\r\n- Lint cleanup\r\n\r\n---\r\n\r\n## Docs\r\n\r\n- Mark TUI as deprecated in roadmap to match CLAUDE.md (#5988)\r\n\r\n---\r\n\r\n## Community Contributors\r\n\r\n- **Navya Bijoy** (@navyabijoy) — Stale session cleanup fix (#6081)\r\n- **Emmanuel Nwanguma** (@Emart29) — MCP errlog resource leak fix (#6094)\r\n- **Pushkal** (@xpushkal) — Server phase_state fix (#6069)\r\n- **Robert Hallers** (@roberthallers) — TUI deprecation docs (#5988)\r\n\r\n---\r\n\r\n## Upgrading\r\n\r\n```bash\r\ngit pull origin main\r\nuv sync\r\n\r\n# For the web frontend\r\ncd core\u002Ffrontend\r\nnpm install\r\nnpm run build\r\n```\r\nor, simply run\r\n```bash\r\n.\u002Fquickstart.sh (linux\u002Fmac)\r\n.\\quickstart.ps1 (windows)\r\n```\r\n","2026-03-11T02:48:28",{"id":246,"version":247,"summary_zh":248,"released_at":249},342333,"v0.6.6","## Highlights\n\n### Queen Planning Phase\n\nThe queen agent now starts in a dedicated PLANNING phase before building. This enforces a disciplined workflow: discover → design → confirm → build.\n\n- **Read-only exploration** — Planning phase has no write\u002Fedit tools; queen can only read files, search, and discover tools\n- **Explicit user approval** — Queen MUST get user confirmation before calling `initialize_and_build_agent`\n- **Diagnosis mode** — When returning from staging\u002Frunning via `stop_worker_and_plan`, queen enters diagnosis mode instead of restarting the full discovery workflow\n- **Phase-switching tools** — New `stop_worker_and_plan` and `replan_agent` tools for transitioning between phases\n\n### Queen Global Memory\n\nQueens now maintain episodic memory across sessions. Memory files are stored in `~\u002F.hive\u002Fqueen\u002F`.\n\n- **Diary system** — `write_to_diary()` tool available in all phases for recording important events\n- **Cross-session persistence** — Memory consolidation hooks generate and maintain memory files in `~\u002F.hive\u002Fqueen\u002F`\n- **Context continuity** — Queen reads past memory at session start and picks up where it left off\n\n### 120 New Tool Integrations\n\n120 new tools across 40 existing integrations — 3 new tools per vendor. Bulk operations, write operations, metadata introspection, and convenience wrappers. (#6097)\n\n| Round | Integrations |\n|-------|-------------|\n| 1 | HubSpot, Jira, Notion, Discord, Salesforce, Telegram, GitHub, Trello |\n| 2 | Asana, Zendesk, Shopify, Twilio, Zoom, Linear, Stripe, Slack |\n| 3 | GitLab, Intercom, Confluence, Pipedrive, Reddit, Airtable, PagerDuty, Calendly |\n| 4 | Apollo, Docker Hub, Google Analytics, Twitter\u002FX, Cloudinary, QuickBooks, Brevo, Greenhouse |\n| 5 | Exa Search, SerpAPI, News, Pushover, AWS S3, Google Search Console, Lusha, Postgres |\n\n---\n\n## What's New\n\n### Core Framework\n\n- **Queen orchestrator** — Extracted `queen_orchestrator.py` from `session_manager.py` for cleaner separation of concerns\n- **Rename hive_coder → queen** — Agent package renamed with dedicated agent.py, config.py, and node definitions\n- **`initialize_and_build_agent`** — Improved with module import validation, better error messages, and edit-not-rewrite guidance\n- **Event loop node refactor** — Clearer execution flow and improved cancellation handling\n- **Judge improvements** — Refactored evaluate logic; handle empty message edge case\n- **Agent package validation** — Detect orphaned nodes during validation (#5955)\n- **Stall threshold increase** — Raised stall detection threshold to reduce false positives\n\n### Email Tools\n\n- **Draft email** — Added reply-in-thread support (#6048)\n- **Reply email** — Email thread body included in replies (#6047)\n\n### CI\u002FCD\n\n- **PR requirements workflow** — Automated PR requirements warning and enforcement (#5955)\n- **UV caching** — CI now uses `enable-cache` for faster builds (#5955)\n\n---\n\n## Bug Fixes\n\n- Fix cancellation button in frontend\n- Fix event loop node test\n- Fix judge with empty message handling\n- Fix memory consolidation hook and simplify generated files\n- Fix queen memory health check\n- Fix orphaned google_sheets.py credential spec (#6079)\n- Fix E501 line-too-long in coder_tools_server.py (#6044)\n- Fix CI test blockage\n- Fix ruff lint issues\n\n---\n\n## Refactoring\n\n- Drop deprecated `hive code` CLI entry point\n- Extract shared building knowledge between planning and building phases\n- Separate queen and worker tools in prompts\n- Simplify session_manager by moving queen logic to queen_orchestrator\n\n---\n\n## Community Contributors\n\n- **Akshat Tiwari** (@akshajtiwari) — CI caching and PR requirements workflow (#5955)\n- **Amdev-5** (@Amdev-5) — E501 lint fix (#6044)\n- **Shaurya Singh** (@Waryjustice) — Remove orphaned google_sheets credential spec (#6079)\n- **Karthik Kotra** (@karthik-kotra) — WSL setup troubleshooting docs (#5029)\n\n---\n\n## Upgrading\n\n```bash\ngit pull origin main\nuv sync\n\n# For the web frontend\ncd core\u002Ffrontend\nnpm install\nnpm run build\n```\nor, simply run\n```bash\n.\u002Fquickstart.sh (linux\u002Fmac)\n.\\quickstart.ps1 (windows)\n```","2026-03-10T03:03:00",{"id":251,"version":252,"summary_zh":253,"released_at":254},342334,"v0.6.5","## Highlights\n\n### Queen Phase Separation\n\nThe queen agent's responsibilities are now cleanly separated into three phases, each with its own scoped system prompt and toolset.\n\n- **Building phase** — Full coding tools for writing and editing agent code, designing graphs, and constructing worker agents\n- **Staging phase** — Agent is loaded but not running; queen can inspect, configure, and launch the worker\n- **Running phase** — Worker is executing; queen monitors, controls, and can stop or restart\n- **Dynamic prompt swapping** — System prompt and available tools change on each phase transition, reducing prompt size and improving instruction adherence\n- **QueenPhaseState** — New state machine managing phase lifecycle, tool availability, and phase-change event notifications\n- **Smaller model support** — Phase-scoped prompts significantly improve instruction following on smaller models\n\n### Queen Thinking Hook\n\n- **Thinking visibility** — New thinking hook surfaces queen reasoning during agent construction\n- **Max token control** — Configurable token limits for thinking output\n\n### Telegram Tool Expansion\n\nExpanded Telegram integration with message management, media, and chat info operations. (#5403)\n\n---\n\n## What's New\n\n### Core Framework\n\n- **Queen prompt optimization** — Condensed building phase prompts, removed unused prompts, reorganized tool documentation\n- **Escalation system** — New escalation tool for worker-to-queen communication during execution\n- **Progressive disclosure** — Runtime data revealed progressively to avoid information overload\n- **Dynamic system prompts** — System prompts loaded dynamically into context\n- **Validation improvements** — GCU and validation added to `initialize_agent_package`; agent.json validated before parsing (#5846)\n- **Judge improvements** — Judge can now wait for queen input; skip-judge logic improved\n- **BOM-safe JSON loading** — Handles byte-order marks in JSON files (#5901)\n- **LLM logger on by default** — LLM debug logging enabled by default\n- **Config error logging** — Parse errors now logged for better debugging (#4955)\n- **Provider key detection** — Improved indirect variable expansion for API key detection (#5504)\n- **Minimax provider** — Added Minimax provider mapping and stream fallback\n\n### Frontend & Server\n\n- **Duplicate session fix** — Prevent duplicate session creation when starting from home\n- **CLI validation** — Validate `--output` path before agent execution (#5838)\n\n### Refactoring\n\n- **Remove old TUI CodingAgent** — Deprecated TUI-specific coding agent removed\n- **Remove unused builder functions** — Cleaned up legacy builder code and output cleaner\n- **Remove old skills** — All unused agent skills removed\n- **Rename queen mode → phase** — Consistent \"phase\" terminology throughout codebase\n- **Reorganize tools** — Coder tools and agent initialization tools restructured\n\n---\n\n## Bug Fixes\n\n- Fix back-to-back compaction edge case\n- Fix turn signal edge case\n- Fix output key that terminated the queen\n- Fix duplicated session calls\n- Fix `Cannot write to closing transport` error\n- Fix missing GCU prompts and instructions\n- Fix thinking hook max token limits\n- Fix MCP tools and templates loading\n- Fix logger schema mismatch\n- Fix legacy agent.json loading error handling\n- Fix Aden client import duplication after rebase\n- Skip POSIX permission tests on Windows (#5842, #5847)\n\n## Documentation\n\n- Add README for brevo, csv, runtime_logs, account_info tools (#5602)\n- Remove old agent skill references\n- Restructure docs layout\n\n---\n\n## Community Contributors\n\n- **Jack** (@jackthepunished) — Telegram tool expansion with message management, media, and chat info (#5403)\n- **VasuBansal7576** — Minimax provider integration\n\n---\n\n## Upgrading\n\n```bash\ngit pull origin main\nuv sync\n\n# For the web frontend\ncd core\u002Ffrontend\nnpm install\nnpm run build\n```\nor, simply run\n```bash\n.\u002Fquickstart.sh (linux\u002Fmac)\n.\\quickstart.ps1 (windows)\n```","2026-03-07T03:58:17",{"id":256,"version":257,"summary_zh":258,"released_at":259},342335,"v0.6.4","## Highlights\r\n\r\n### 60+ New Tool Integrations\r\n\r\nThe largest integration expansion in Hive's history — over 60 new tool integrations (300+ individual tools) landed in a single release, many contributed by the community. Agents can now connect to CRMs, databases, cloud services, messaging platforms, and more out of the box.\r\n\r\n**CRM & Sales** — Salesforce, Zoho CRM, Attio, Pipedrive, Lusha\r\n**Project Management** — Jira, Asana, Linear, Trello, Notion\r\n**Developer Tools** — GitLab, Docker Hub, Vercel, Databricks, Redis, Supabase, HuggingFace, Pinecone, Terraform Cloud\r\n**Communication** — Twilio, Zoom, Microsoft Graph (Outlook, Teams, OneDrive), Pushover, Tines, Brevo\r\n**Data & Analytics** — Snowflake, Amazon Redshift, Azure SQL, Apache Kafka, Power BI, Google Sheets, Google Search Console, MongoDB, Airtable, MSSQL\r\n**Finance** — Plaid, QuickBooks, Yahoo Finance\r\n**Cloud & Storage** — AWS S3, Cloudinary\r\n**Media & Content** — YouTube Data API, YouTube Transcript, Reddit, Confluence, Obsidian, Langfuse\r\n**Automation** — n8n, Calendly, Apify, PagerDuty\r\n**Other** — Twitter\u002FX, Greenhouse, DuckDuckGo, SAP S\u002F4HANA, Shopify\r\n\r\nEach integration includes credential specs, health checks, and unit tests.\r\n\r\n### Hashline Edit Tool\r\n\r\nA new cross-tool hashing anchor system for precise file editing.\r\n\r\n- **Hash-based line anchoring** — Unique hashes per line for stable references across tool calls\r\n- **Grep search integration** — `grep_search` with large file skip reporting\r\n- **Coder tools support** — Hive Coder agent tools now support hashline editing\r\n- **CRLF handling** — Prevents double-conversion in hashline edits\r\n\r\n### Windows Platform Support\r\n\r\nComprehensive Windows compatibility improvements across the entire stack.\r\n\r\n- **Filesystem support** — Windows file system paths and operations (#5677)\r\n- **UTF-8 encoding** — Systematic enforcement across tools and core to fix charmap decode errors\r\n- **Browser auto-open** — Fixed quickstart browser launch on Windows\r\n- **`os.fchmod` guard** — Safely handles missing POSIX calls on Windows\r\n\r\n### Session Management & Continuity\r\n\r\n- **Session resume** — Conversations resume from where they were left off\r\n- **Session loading** — All sessions load properly without errors\r\n- **Pause\u002Fstop fix** — Pipeline pause now correctly uses `stop_worker` like queen\r\n\r\n---\r\n\r\n## What's New\r\n\r\n### Core Framework\r\n\r\n- **Verified\u002Funverified tool tiers** — Tool loading split into verified and unverified tiers for clearer trust boundaries. (`core\u002Fframework\u002Ftools\u002F`)\r\n- **Google OAuth unification** — Single credential flow for all Google services (Calendar, Sheets, Search Console). (`tools\u002F`)\r\n- **Google Scopes expansion** — Extended OAuth scopes for broader Google API access. (#5764)\r\n- **Remove old session tools** — Deprecated `get_agent_session_state` and `get_agent_session_memory` removed. (#5828, #5829)\r\n- **Remove hardcoded Anthropic logics** — LLM provider logic is now fully configurable.\r\n- **Health check system** — Simplified health check architecture with LLM key validation.\r\n- **LLM provider key update** — Easy mechanism to update LLM provider keys.\r\n- **GCU enabled by default** — Graph context updates now on by default.\r\n- **Recommended models update** — Updated default model recommendations.\r\n- **Mac keychain for Claude Code** — Utilize macOS keychain for Claude Code subscription credentials.\r\n\r\n### Frontend & Server\r\n\r\n- **`hive open` command** — New CLI command to launch the browser-based interface directly.\r\n- **Permanent top bar** — Fixed persistent top navigation bar.\r\n- **Quickstart improvements** — Auto-install Node 20, Windows alignment, streamlined setup flow.\r\n\r\n### New Agent Templates\r\n\r\n| Template | Description |\r\n|----------|-------------|\r\n| **Twitter News Agent** | Automated Twitter\u002FX news monitoring and aggregation |\r\n| **Local Business Agent** | Local business discovery and information |\r\n| **Meeting Scheduler Agent** | Calendar-aware meeting scheduling with Google Calendar |\r\n\r\n---\r\n\r\n## Bug Fixes\r\n\r\n- Fix permanent top bar display issue\r\n- Fix pause in pipeline to use `stop_worker` like queen (#5816)\r\n- Fix CRLF double-conversion in hashline edit and add large file skip reporting\r\n- Fix Windows compat — guard `os.fchmod` and remove deleted `LLM_CREDENTIALS` import\r\n- Fix UTF-8 encoding across tools and core to resolve Windows charmap decode errors (#5786)\r\n- Fix browser auto-open after quickstart on Windows (#5797)\r\n- Fix `stdin` conflicts in quickstart flow\r\n- Fix quickstart build failing (#5834)\r\n- Fix session loading and resume continuity (#5815)\r\n- Fix event routes unhandled error\r\n- Remove duplicate `_execute_subagent` method in EventLoopNode (#5785)\r\n- Remove old `get_agent_session_state` and `get_agent_session_memory` tools\r\n- Fix tool tests and credential wiring for community integrations\r\n- Remove hardcoded Anthropic logics from LLM provider\r\n\r\n## Documentation\r\n\r\n- Update roadmap to reflect completed features (refs #4780)\r\n- Update `AGENTS.md` with la","2026-03-05T04:13:08",{"id":261,"version":262,"summary_zh":263,"released_at":264},342336,"v0.6.3","## Highlights\n\n### Queen Mode Separation\n\nThe queen agent now operates in three distinct modes with dynamic tool provisioning.\n\n- **Building mode** -- Queen has full coding tools for editing and creating agents\n- **Staging mode** -- Worker agent is loaded and ready, queen gains lifecycle tools (start, stop, pause, inject messages)\n- **Running mode** -- Worker is executing, queen switches to monitoring and control tools (get status, stop worker, inject messages)\n- **Frontend mode sync** -- Mode tag displayed on queen messages with visual styling per mode; frontend can trigger mode transitions\n\n### Interactive Question Widget\n\nA new structured question UI for agent-to-human interactions.\n\n- **Option-based questions** -- Agents can present up to 3 predefined options plus an auto-appended \"Other\" free-text option\n- **Keyboard shortcuts** -- Number keys (1-4) to select options, Enter to submit\n- **`ask_user` tool upgrade** -- Options support added to the `ask_user` tool for structured user input collection\n- **Queen context endpoint** -- New `\u002Fqueen-context` API for queuing non-intrusive messages to the queen between iterations\n\n### Email Reply Agent Template\n\nA new sample agent demonstrating email automation workflows.\n\n- **Email filtering** -- Intake node for user criteria collection\n- **Search and reply** -- Automated unreplied email search with draft composition\n- **Confirmation gate** -- Recipient approval required before sending\n- **Continuous conversation mode** -- Multi-node conversation threading support\n\n---\n\n## What's New\n\n### Core Framework\n\n- **Queen mode state machine** -- Three-stage mode switching (building\u002Fstaging\u002Frunning) with dynamic tool provisioning and event-driven transitions. (`core\u002Fframework\u002Ftools\u002Fqueen_lifecycle_tools.py`)\n- **Execution quality tracking** -- Granular metrics: clean\u002Fdegraded\u002Ffailed execution quality, retry counts, node visit tracking, and partial failure detection. (`core\u002Fframework\u002Fgraph\u002Fexecutor.py`)\n- **Continuous conversation mode** -- Phase-aware conversation compaction with `phase_id`, transition markers, and client input tagging. (`core\u002Fframework\u002Fgraph\u002Fconversation.py`)\n- **Consolidated MCP tool discovery** -- Refactored `discover_mcp_tools` and `list_agent_tools` into a single flow. (`core\u002Fframework\u002Fgraph\u002Fevent_loop_node.py`)\n- **Agent collaboration guidelines** -- New `AGENTS.md` for multi-agent workspace conventions. (`AGENTS.md`)\n- **uv instructions** -- Added `uv` usage guidance to agent system prompts. (`core\u002Fframework\u002Fagents\u002Fhive_coder\u002Fnodes\u002F__init__.py`)\n\n### Frontend & Server\n\n- **QuestionWidget component** -- Interactive option-based question UI with keyboard navigation and free-text fallback. (`core\u002Ffrontend\u002Fsrc\u002Fcomponents\u002FQuestionWidget.tsx`)\n- **Queen mode display** -- Mode tag on queen messages with per-mode styling and frontend mode switching controls. (`core\u002Ffrontend\u002Fsrc\u002Fpages\u002Fworkspace.tsx`, `core\u002Ffrontend\u002Fsrc\u002Fcomponents\u002FChatPanel.tsx`)\n- **Queen context endpoint** -- `POST \u002Fapi\u002Fsessions\u002F{id}\u002Fqueen-context` for queuing external events without triggering immediate LLM response. (`core\u002Fframework\u002Fserver\u002Froutes_execution.py`)\n- **Queen status on session start** -- Returns queen bee mode state when starting a session. (`core\u002Fframework\u002Fserver\u002Froutes_sessions.py`)\n- **Credential modal fix** -- Hide delete button for Aden-managed credentials. (`core\u002Ffrontend\u002Fsrc\u002Fcomponents\u002FCredentialsModal.tsx`)\n\n### New Agent Template\n\n| Template | Description |\n|----------|-------------|\n| **Email Reply Agent** | Email filtering, search, reply composition, and confirmation-gated sending with continuous conversation mode |\n\n---\n\n## Bug Fixes\n\n- Fix pause\u002Fstop to cancel all running tasks across all graphs in TUI (#5234)\n- Fix queen auto-block from overwriting pending worker questions\n- Fix queen returning on empty stream\n- Fix conversation compaction logics for multi-phase execution\n- Fix LLM conversation cleansing for clean context windows\n- Fix MCP server cwd resolution from repo root instead of agent path\n- Fix credential store sharing between server and runner to avoid redundant syncs\n- Fix hide delete button for Aden-managed credentials\n- Fix load new session from home page with proper session ID generation\n- Fix execution recovery for interrupted sessions\n- Fix initial state condition for queen mode transitions\n- Fix agent generation guidelines for hive coder\n\n## Documentation\n\n- Remove TUI references from primary README\n- Add running screenshot and update coding agent instructions\n- Sync all i18n READMEs with primary README (es, hi, ja, ko, pt, ru, zh-CN)\n- Reorder documentation sections\n- Update latest features in README\n\n---\n\n## Community Contributors\n\n- **Aaryann Chandola** (@Antiarin) -- Fix pause\u002Fstop to cancel all running tasks across all graphs (#5234)\n\n---\n\n## Upgrading\n\n```bash\ngit pull origin main\nuv sync\n\n# For the web frontend\ncd core\u002Ffrontend\nnpm install\nnpm run build\n```\n","2026-03-04T05:43:09",{"id":266,"version":267,"summary_zh":268,"released_at":269},342337,"v0.6.2","## Highlights\n\n### Separate Queen and Worker Input\n\nA significant UX redesign for multi-agent workflows.\n\n- **Dedicated input channels** -- The main chat input is now permanently connected to the queen agent, eliminating routing ambiguity\n- **Inline worker reply boxes** -- When a worker requests user input, a dedicated reply box appears inline in the chat thread\n- **Clearer multi-agent interactions** -- Users always know which agent they're communicating with\n\n### Subagent Framework\n\nA major architectural addition enabling hierarchical agent orchestration with queen-worker communication patterns.\n\n- **Queen-worker communication** -- Parent agents can spawn and coordinate child subagents with bidirectional messaging\n- **Subagent lifecycle tools** -- New tools for spawning, monitoring, and managing subagent execution\n- **Progressive feedback** -- SubagentJudge provides incremental feedback during long-running subagent tasks\n- **Worker status reporting** -- Consolidated status updates from workers to parent agents\n- **Subagent logging** -- Organized log structure with subagent logs stored in dedicated node folders\n\n### GCU (General Computer Use)\n\nNew browser automation capabilities built into the framework as a first-class node type.\n\n- **Browser control tools** -- Complete toolkit for navigation, interaction, inspection, and tab management\n- **Interactive element highlighting** -- Visual highlighting of actionable elements on web pages\n- **Snapshot tools** -- Page state capture for debugging and verification\n- **Health checks** -- Browser session validation at startup\n- **GCU node type** -- Built-in browser tools automatically available for GCU-enabled nodes\n- **Quickstart integration** -- GCU option added to the quickstart menu\n\n### Security Hardening\n\nMultiple security improvements for safer agent execution.\n\n- **Agent path validation** -- Restricts agent loading to allowed directories only\n- **OAuth token permissions** -- Enforces 0600 permissions on OAuth token files\n- **Preload validation** -- Enhanced pre-start validation for GCU subagents and credentials\n\n### Frontend & UX Improvements\n\nSignificant updates to the web workspace experience.\n\n- **Colorful tool pills** -- Visual distinction with color-coded tool call indicators\n- **Agent building animation** -- Animated feedback during agent construction\n- **Subagents in node panel** -- Subagent activity displayed in the node detail panel\n- **Scheduler countdown** -- Frontend display of agent idling timer\n- **Separate worker\u002Fqueen input** -- Distinct input handling for orchestrated agents\n\n---\n\n## What's New\n\n### Core Framework\n\n- **Subagent framework** -- Queen-worker communication with bidirectional messaging and lifecycle management. (`core\u002Fframework\u002Ftools\u002Fqueen_lifecycle_tools.py`, `core\u002Fframework\u002Fgraph\u002Fevent_loop_node.py`)\n- **GCU node type** -- Browser automation as a built-in node capability with tool highlighting. (`core\u002Fframework\u002Fgraph\u002Fgcu.py`, `tools\u002Fsrc\u002Fgcu\u002F`)\n- **Agent idling detection** -- Runtime detection of idle agents with configurable timeouts. (`core\u002Fframework\u002Fruntime\u002Fagent_runtime.py`)\n- **Preload validation** -- Enhanced validation pipeline for subagents and credentials before agent start. (`core\u002Fframework\u002Frunner\u002Fpreload_validation.py`)\n- **LLM debug logger** -- Human-friendly logging for LLM calls and tool invocations. (`core\u002Fframework\u002Fruntime\u002Fllm_debug_logger.py`)\n- **Gemini 3.1 Pro support** -- Added support for Gemini 3.1 Pro model. (`core\u002Fframework\u002Fllm\u002Flitellm.py`)\n\n### Frontend & Server\n\n- **Tool pills v2** -- Colorful, per-call tool status indicators with turn tracking. (`core\u002Ffrontend\u002Fsrc\u002Fcomponents\u002FChatPanel.tsx`)\n- **Agent building animation** -- Visual feedback during agent graph construction. (`core\u002Ffrontend\u002Fsrc\u002Fpages\u002Fworkspace.tsx`)\n- **Node panel subagents** -- Subagent activity visible in node detail panel. (`core\u002Ffrontend\u002Fsrc\u002Fcomponents\u002FNodeDetailPanel.tsx`)\n- **Scheduler countdown** -- Timer display for agent idling detection. (`core\u002Ffrontend\u002Fsrc\u002Fpages\u002Fworkspace.tsx`)\n- **SSE event handling** -- Improved event bus to prevent critical event queue blocking. (`core\u002Fframework\u002Fruntime\u002Fevent_bus.py`)\n\n### Tools & Integrations\n\n- **GCU browser tools** -- Full browser automation suite:\n  - Navigation: `goto_url`, `go_back`, `go_forward`, `refresh_page`\n  - Interactions: `click_element`, `type_text`, `scroll`, `drag_and_drop`\n  - Inspection: `get_page_content`, `get_element_text`, `take_screenshot`\n  - Tabs: `new_tab`, `switch_tab`, `close_tab`, `list_tabs`\n  - Advanced: `execute_javascript`, `wait_for_element`, `extract_data`\n- **File operation tools** -- Upgraded file operations with better error handling. (`tools\u002Fsrc\u002Faden_tools\u002Ffile_ops.py`)\n\n### Documentation\n\n- **Runtime initialization docs** -- Comprehensive documentation for runtime startup flow. (`docs\u002Fruntime_initialization.md`)\n- **Architecture updates** -- Expanded architecture documentation. (`docs\u002Farchitecture\u002FREADME.md`)\n- **GCU guide** -- Reference guide for brow","2026-03-03T07:03:05",{"id":271,"version":272,"summary_zh":273,"released_at":274},342338,"v0.6.1","## Highlights\r\n\r\n### Tool Calling Revamp\r\n\r\nThe event loop node's tool calling logic is rewritten for more reliable and predictable execution.\r\n\r\n- **New tool call routing** -- restructured edge routing with improved conversation flow management\r\n- **Conversation management overhaul** -- refactored conversation handling with better context tracking across iterations\r\n- **Event loop iteration fixes** -- resolved iteration counting and session update spamming issues\r\n- **Pipeline visual update** -- improved execution pipeline visualization in the frontend\r\n\r\n### Credential System Hardening\r\n\r\nThe credential subsystem gains a dedicated key storage layer and cleaner lifecycle management.\r\n\r\n- **Key storage module** -- new `key_storage.py` for structured local encrypted credential management\r\n- **Migrated to `~\u002F.hive`** -- credentials moved from shell config to `~\u002F.hive` directory\r\n- **Improved validation pipeline** -- enhanced credential validation with better error reporting\r\n- **Session-start loading** -- credentials checked and loaded into new agent sessions automatically\r\n- **Dismissable banners** -- credential error banners can now be dismissed without blocking\r\n\r\n### Web Workspace Fixes\r\n\r\nStability and UX improvements across the frontend and server.\r\n\r\n- **SSE reconnect on session change** -- stable reconnection when switching between sessions\r\n- **Tool pill per-call tracking** -- individual tool call status indicators\r\n- **Cancel\u002Fpause event emission** -- proper event signals for session control\r\n- **Credential modal UX** -- improved add\u002Fvalidate\u002Fdismiss flows\r\n- **Graph visualization** -- layout and node rendering refinements\r\n\r\n---\r\n\r\n## What's New\r\n\r\n### Core Framework\r\n\r\n- **Tool call revamp** -- New tool calling logic in event loop node with refactored conversation management. (`core\u002Fframework\u002Fgraph\u002Fevent_loop_node.py`, `core\u002Fframework\u002Fgraph\u002Fconversation.py`)\r\n- **Edge routing** -- New edge routing capabilities for conditional execution flow. (`core\u002Fframework\u002Fgraph\u002Fedge.py`)\r\n- **LiteLLM improvements** -- Better tool call handling and provider compatibility. (`core\u002Fframework\u002Fllm\u002Flitellm.py`)\r\n- **Executor updates** -- Improved graph execution with better session handling. (`core\u002Fframework\u002Fgraph\u002Fexecutor.py`)\r\n- **Execution stream** -- Refined event streaming for tool call tracking. (`core\u002Fframework\u002Fruntime\u002Fexecution_stream.py`)\r\n- **Key storage** -- Dedicated local encrypted credential storage module. (`core\u002Fframework\u002Fcredentials\u002Fkey_storage.py`)\r\n\r\n### Frontend & Server\r\n\r\n- **SSE reconnect** -- Stable reconnection on session change with tool pill per-call tracking. (`core\u002Ffrontend\u002Fsrc\u002Fhooks\u002Fuse-sse.ts`)\r\n- **Credential modal** -- Improved add\u002Fvalidate\u002Fdismiss UX with dismissable error banners. (`core\u002Ffrontend\u002Fsrc\u002Fcomponents\u002FCredentialsModal.tsx`)\r\n- **Chat panel** -- Better message rendering and interaction flow. (`core\u002Ffrontend\u002Fsrc\u002Fcomponents\u002FChatPanel.tsx`)\r\n- **Graph visualization** -- Agent graph layout and node rendering improvements. (`core\u002Ffrontend\u002Fsrc\u002Fcomponents\u002FAgentGraph.tsx`)\r\n- **Home page** -- Centered text and Open Hive branding. (`core\u002Ffrontend\u002Fsrc\u002Fpages\u002Fhome.tsx`)\r\n- **Server credential routes** -- Updated credential and event API routes. (`core\u002Fframework\u002Fserver\u002Froutes_credentials.py`, `core\u002Fframework\u002Fserver\u002Froutes_events.py`)\r\n\r\n### New Tool Integrations\r\n\r\n| Tool | Description | Contributor |\r\n|------|-------------|-------------|\r\n| **Intercom** | Conversations, contacts, and tags management with health checker | @Ttian18 |\r\n| **Google Analytics 4** | Property listing, report generation, and GA4 data access | @Ttian18 |\r\n\r\n---\r\n\r\n## Bug Fixes\r\n\r\n- Fix event loop iteration counting\r\n- Fix spamming session update on every iteration\r\n- Fix dismiss credential banner not persisting\r\n- Fix credential loading into new agent sessions\r\n- Fix mock_mode removal from queen\u002Fcoder system prompt templates\r\n- Fix graph summary for intake nodes\r\n- Fix incorrect CLI commands and docstrings in core docs (#5457)\r\n- Fix wrong credential path and env var references in docs (#5458)\r\n- Fix runtime logger test assertions for unified session run IDs (#5480)\r\n- Fix deep research agent node initialization\r\n- Fix SSE reconnect on session change\r\n- Remove unused load agent code\r\n\r\n## Documentation\r\n\r\n- Fix incorrect CLI commands and docstrings in core README\r\n- Fix credential path and env var references in environment setup docs\r\n- Fix configuration documentation paths\r\n- Updated file templates and anti-patterns in hive coder reference\r\n- Windows quickstart script update\r\n\r\n---\r\n\r\n## Upgrading\r\n\r\n```\r\ngit pull origin main\r\nuv sync\r\n\r\n# For the web frontend\r\ncd core\u002Ffrontend\r\nnpm install\r\nnpm run build\r\n```","2026-02-28T03:49:07",{"id":276,"version":277,"summary_zh":278,"released_at":279},342339,"v0.6.0","# Release Notes\r\n\r\n**Release Date:** February 26, 2026\r\n**Tag:** v0.6.0\r\n\r\n## Open Hive\r\n\r\nv0.6.0 is the biggest release in Hive's history. The framework breaks out of the terminal with a **full web workspace** -- a React SPA backed by a new FastAPI HTTP server where you can run agents, chat, visualize graphs, and manage credentials from the browser. Under the hood, a new **queen\u002Fworker\u002Fjudge multi-agent runtime** enables agents that spawn, monitor, and coordinate other agents. The credential system gets a ground-up rewrite with local encrypted storage, health checks, and mid-session installation. And **OpenAI Codex** joins as a first-class LLM provider with OAuth subscription support.\r\n\r\n---\r\n\r\n## Highlights\r\n\r\n### Web Workspace\r\n\r\nA complete browser-based workspace replaces the TUI as the primary interface for running agents. Built with React, TypeScript, Vite, shadcn\u002Fui, and Tailwind CSS.\r\n\r\n- **Home page** -- browse agents and start sessions with an initial prompt\r\n\r\n\u003Cimg width=\"2544\" height=\"1196\" alt=\"Screenshot 2026-02-26 at 8 54 36 PM\" src=\"https:\u002F\u002Fgithub.com\u002Fuser-attachments\u002Fassets\u002Fe7cedfce-1307-401b-a0e1-c6e7bb441d9f\" \u002F>\r\n\r\n\r\n- **Workspace** -- interactive agent graph visualization (ReactFlow), live chat panel, node detail inspector, and streaming log pane\r\n\r\n\u003Cimg width=\"2544\" height=\"1196\" alt=\"Screenshot 2026-02-26 at 8 54 12 PM\" src=\"https:\u002F\u002Fgithub.com\u002Fuser-attachments\u002Fassets\u002F7b5c38dd-863a-414c-910b-cc93381314de\" \u002F>\r\n\r\n- **Credentials modal** -- add, validate, and revoke credentials mid-session without restarting\r\n- **Real-time streaming** -- SSE-based event streaming for live execution updates\r\n- **Markdown rendering** -- rich message formatting in the chat panel\r\n- **Tab persistence** -- multi-session support with persistent workspace tabs\r\n\r\n### HTTP API Server\r\n\r\nA FastAPI server at `core\u002Fframework\u002Fserver\u002F` provides the backend for the web workspace and exposes a full REST + SSE API:\r\n\r\n- **Session lifecycle** -- create, list, resume, cancel, reconnect sessions\r\n- **Execution control** -- start, stop, send input, cancel workers\r\n- **Graph inspection** -- load and query agent graph definitions\r\n- **Credential management** -- validate, install, and revoke credentials at runtime\r\n- **SSE event streaming** -- real-time execution events for frontend consumption\r\n- **Log queries** -- structured runtime log access\r\n\r\nFull API documentation in `core\u002Fframework\u002Fserver\u002FREADME.md` and `docs\u002Fserver-cli-arch.md`.\r\n\r\n### Queen \u002F Worker \u002F Judge Multi-Agent Runtime\r\n\r\nA new orchestration layer enables hierarchical multi-agent coordination:\r\n\r\n- **Queen lifecycle tools** -- spawn, monitor, stop, and cancel worker agents with escalation ticket support\r\n- **Worker health monitoring** -- stall detection, iteration tracking, and automatic cleanup of stale sessions\r\n- **Concurrent judge** -- event-bus-integrated evaluation that runs alongside agent execution\r\n- **Session Manager** -- replaces the old Agent Manager with full concurrent session support, reconnection, and session isolation\r\n- **3-layer resume prompts** -- robust checkpoint recovery across session restarts\r\n- **Trigger node visualization** -- schedule info displayed when clicking trigger nodes in the graph\r\n\r\n### Credential System v3\r\n\r\nThe credential subsystem is rewritten again with local-first encrypted storage:\r\n\r\n- **Local credential registry** (`credentials\u002Flocal\u002F`) -- encrypted storage with namespace support (`{credential_name}\u002F{alias}`)\r\n- **Health check framework** -- live API validation for stored credentials (Brevo, PostgreSQL, and more)\r\n- **Mid-session credential management** -- install missing credentials and resync MCP servers without restarting\r\n- **Deferred validation** -- credentials validated on use rather than on load, with dismissable error banners\r\n- **Revoke and re-validate** -- full lifecycle management in both TUI and web UI\r\n\r\n### OpenAI Codex LLM Provider\r\n\r\nCodex joins as a supported LLM provider with full OAuth integration:\r\n\r\n- **OAuth consent flow** -- browser pop-out for subscription authentication\r\n- **Streaming support** -- through LiteLLM with proper tool call handling\r\n- **Weak model fixes** -- skip auto-block when models output text instead of calling tools, remove implementation hints from judge feedback\r\n\r\n---\r\n\r\n## What's New\r\n\r\n### Architecture & Runtime\r\n\r\n- **Web frontend** -- Full React + TypeScript SPA with agent workspace, graph visualization, chat, and credential management. (`core\u002Ffrontend\u002F`)\r\n- **HTTP API server** -- FastAPI backend with route modules for sessions, execution, graphs, credentials, events, and logs. (`core\u002Fframework\u002Fserver\u002F`)\r\n- **Session Manager** -- Replaces Agent Manager with concurrent session support, reconnection, and isolation. (`core\u002Fframework\u002Fserver\u002Fsession_manager.py`)\r\n- **Queen lifecycle tools** -- 6 tools for spawning, monitoring, stopping, and canceling worker agents with escalation support. (`core\u002Fframework\u002Ftools\u002Fqueen_lifecycle_tools.py`)\r\n- **Worker monitoring tools** --","2026-02-27T04:52:06"]