v1.11.0

AIO Sandbox 1.11.0 adds browser file-picker policy controls, fuller Codex configuration support, MCP Bash output controls, and image build improvements.

Features

• Browser file selection policy: Added BROWSER_ALLOW_FILE_SELECTION_DIALOGS to control whether web pages can open file picker dialogs.
• Codex custom configuration: Added CODEX_CONFIG_TOML and CODEX_MODEL_CATALOG_JSON for fully custom Codex config and model catalog setup at container startup.
• MCP Bash truncate control: sandbox_execute_bash now supports a truncate parameter, allowing callers to request full output when needed.
• Image build improvements: Restored native arm64 image build support.

Improvements

• Shell output reliability: Improved handling of long terminal output without newlines.
• AIO CLI availability: Restored in-sandbox aio loader commands such as shell, tty, file, and bash aliases.

---

AIO Sandbox 1.11.0 新增浏览器文件选择策略控制、Codex 完整自定义配置、MCP Bash 输出控制，并改进镜像构建与 Shell 可靠性。

主要更新

• 浏览器文件选择弹窗策略：新增 BROWSER_ALLOW_FILE_SELECTION_DIALOGS 控制网页是否可以打开文件选择器。
• Codex 完整自定义配置：新增 CODEX_CONFIG_TOML 与 CODEX_MODEL_CATALOG_JSON，支持在容器启动时写入完整 Codex 配置和模型目录。
• MCP Bash 输出截断控制：sandbox_execute_bash 新增 truncate 参数，调用方可按需获取完整输出。
• 镜像构建改进：恢复原生 arm64 镜像构建支持。

改进

• Shell 输出可靠性：优化无换行长输出场景。
• AIO CLI 可用性：恢复沙箱内 aio shell、tty、file、bash 等基础命令。

v1.10.0

AIO Sandbox 1.10.0 adds built-in OpenAI Codex CLI support and includes reliability improvements for Shell and Browser workflows.

Feature

• Codex CLI integration: The sandbox image now includes the OpenAI Codex CLI. You can open it from /codex in the browser or run codex directly inside the container.

Improvements

• Shell reliability: Improved long-running task handling and Shell session reconnect behavior.
• Large output handling: Shell and Bash can preserve full logs in sandbox files when inline responses are truncated.
• Browser screenshot compatibility: The legacy /screenshot route now respects image format and quality parameters, matching the newer browser screenshot API behavior.

For Codex setup details, see: https://sandbox.agent-infra.com/guide/advanced/codex

---

AIO Sandbox 1.10.0 新增内置 OpenAI Codex CLI 支持，并继续改进 Shell 与 Browser 工作流的可靠性。

主要更新

• Codex CLI 集成：镜像内置 OpenAI Codex CLI，可通过浏览器访问 /codex，或在容器内直接运行 codex 命令使用。

改进

• Shell 稳定性：改进长任务处理与 Shell 会话重连表现。
• 大输出处理：Shell 和 Bash 在内联响应被截断时可保留完整日志文件。
• 浏览器截图兼容性：旧版 /screenshot 路由现在会正确处理图片格式和质量参数，与新版浏览器截图接口保持一致。

Codex 配置方式见文档：https://sandbox.agent-infra.com/zh/guide/advanced/codex

v1.9.3

AIO Sandbox has been aligned to SemVer. This is a cumulative release from 1.0.0.152 to 1.9.3, and also a consolidation release before the 2.0 refactor.

The 1.9.x line syncs the major capabilities, reliability improvements, and documentation updates accumulated over the past releases.

The current image is distributed for linux/amd64 only. ARM machines can run it with --platform linux/amd64 emulation.

Highlights

• Built-in AIO CLI: The sandbox image now includes the aio command, allowing agents and users to access browser, GUI, file, sandbox, MCP, and skills capabilities directly from shell.
• Improved Bash execution: Added /v1/bash/*, a stateful Bash execution path designed for agent workflows, with separated stdout/stderr, stdin writing, process control, output polling, and session management.
• Real-time file watching: Added File Watch with SSE, long polling, and blocking wait modes for file and directory events such as create, write, remove, rename, and chmod.
• Desktop recording: Added Display recording support for capturing the X11 desktop as MP4.
• Proxy Mapping upgrade: Upgraded Proxy Mapping to GOST v3, with support for domain/IP/port mapping, HTTPS mapping, header-based port routing, proxy authentication, and health checks.
• Runtime and tooling improvements: Added or improved support for multiple Node.js versions, Go, bun, yarn, common system dependencies, and fonts for better code execution and browser rendering.
• MCP and Skills improvements: Reworked MCP Hub integration with extensible MCP servers, tool search, skill loading, and skill registration overwrite support.

Reliability and Compatibility

• Overall browser path improvements: Improved stability across browser connection, restart, cookies, language handling, user-agent behavior, agent-browser integration, and page interactions.
• Core execution path improvements: Shell / Bash, file upload/download, file watching, and proxy routing are more reliable for long-running tasks, large files, and concurrent workloads.

For more details, see the documentation: https://sandbox.agent-infra.com/guide/start/agent-sandbox

---

AIO Sandbox 已统一为 SemVer（三段式语义化版本）。本次为从 1.0.0.152 到 1.9.3 的累积更新，也是 2.0 重构前的一次集中同步。

1.9.x 同步了过去一段时间内的主要能力、稳定性修复和文档更新。

当前镜像仅分发 linux/amd64 版本，ARM 机器可使用 --platform linux/amd64 模拟执行。

主要更新

• 内置 AIO CLI：镜像内置 aio 命令，Agent 和用户可以直接在 shell 中调用 browser、GUI、file、sandbox、MCP 和 skills 等能力。
• Bash 执行能力升级：新增 /v1/bash/*，提供更适合 Agent 使用的有状态 Bash 执行链路，支持 stdout/stderr 分离、stdin 写入、进程控制、输出轮询和会话管理。
• 实时文件监听：新增 File Watch，支持 SSE、长轮询和阻塞等待，可监听文件和目录的 create、write、remove、rename、chmod 等事件。
• 桌面录屏：新增 Display recording 能力，可录制 X11 桌面画面并输出 MP4。
• 代理能力升级：Proxy Mapping 升级为 GOST v3，支持域名/IP/端口映射、HTTPS 映射、请求头端口路由、代理认证和健康检查。
• 运行时与工具链增强：补齐多版本 Node.js、Go、bun、yarn、常用系统依赖和字体支持，提升代码执行、浏览器渲染和多语言场景的可用性。
• MCP 与 Skills 体验增强：重构 MCP Hub，支持扩展 MCP server、工具搜索、skill 加载和 skill 覆盖注册，降低 Agent 接入成本。

稳定性与兼容性

• 浏览器链路整体优化：围绕连接、重启、Cookie、语言、User-Agent、agent-browser 和页面操作稳定性进行了系统性增强。
• 核心执行链路增强：Shell / Bash、文件上传下载、文件监听和代理转发在长任务、大文件和并发场景下更加稳定。

更多使用方式和能力说明可参考文档：https://sandbox.agent-infra.com/guide/start/agent-sandbox

v1.0.0.152

• Established the AIO Sandbox evaluation framework & dataset (!78), generating evaluation reports (Latest.md) and tool optimization reports (improvement_suggestions.md). The /mcp tools will be continuously iterated based on evaluation feedback.
• Services can be disabled on-demand. By setting environment variables DISABLE_JUPYTER=true and DISABLE_CODE_SERVER=true, the Jupyter Lab and Code Server services can be turned off as needed, reducing some sandbox resource usage. Close #29
• Fixed an issue where static resources were blocked by authentication after enabling JWT_PUBLIC_KEY authentication.
• /v1/browser/info now supports the X-Forwarded-Prefix header prefix for retrieving the CDP address.
• Disabled the Google Translate popup prompt in the browser.
• Added weasyprint dependency (commonly used for generating PDFs) and downgraded plotly from version 6.x to 5.x.

---

• 建立 AIO Sandbox 评测框架 & 数据集（!78），生成评估报告（Latest.md）、工具优化报告（improvement_suggestions.md），/mcp工具将基于评测反馈持续迭代。
• 服务按需关闭，通过环境变量 DISABLE_JUPYTER=true、DISABLE_CODE_SERVER=true可按需关闭 Jupyter Lab 和 Code Server 服务，减少一些沙箱资源占用。
• 修复 JWT_PUBLIC_KEY鉴权开启后，静态资源被鉴权拦截问题。
• /v1/browser/info 获取 CDP 地址支持 Header X-Forwarded-Prefix 前缀
• 浏览器禁用提示 Google Translate 翻译弹窗
• 新增 weasyprint依赖（生成 PDF 时常用），降级 plotly 6.x → 5.x

v1.0.0.151

• Fixed the issue where the homepage showed a 404 error after deploying AIO via a subpath.
• Added the AIO_SKILLS_PATH environment variable to support mounting/registering Skills during service startup, while also parsing the installation commands for Skill dependencies.
• /mcp now supports stateless requests, eliminating the need to call initialize first to establish a session.

---

• 修复通过子路径部署 AIO 后，首页 404 无法展示问题
• 新增AIO_SKILLS_PATH环境变量，支持服务启动时挂载/注册 Skills，同时解析出 Skill 依赖安装命令。
• /mcp 支持无状态请求，无须先调用 initialize 来建立 Session

v1.0.0.150

• Optimized service startup time, reducing duration from 16s to 8s for 1c2g configuration, and from 11s to 4s for 2c4g configuration.

• The system now supports registration, mounting, retrieval, and deletion of /v1/skills/*(API) Claude Skills. Users can upload and mount skills to the sandbox environment, integrating seamlessly with the existing Sandbox API.
• The number of /mcp tools has been reduced (from 60 to 30). Methods from chrome-devtools-mcp are hidden by default but can be accessed via /mcp?search=chrome_devtools.
• Added QR code generation qrcode dependency

---

• 优化服务启动时长，1c2g 配置耗时从 16s → 8s，2c4g 从 11s → 4s
• 新增 /v1/skills/*（API） Claude Skills 注册/挂载/获取/删除，支持注册、挂载、查询与删除，可将 Skills 上传/挂载到沙箱环境，与现有 Sandbox API 结合使用。
• /mcp tools 数量压缩（从 60 → 30） ，默认隐藏 chrome-devtools-mcp 工具方法，可通过 /mcp?search=chrome_devtools 使用；
• 新增二维码生成 qrcode 依赖

v1.0.0.145

• Fixed the status determination for long-running shell commands to prevent misidentification as terminated.
• Resolved the lag issue caused by an excessive number of files in /home/${USER} during startup.
• Browser remove --proxy-bypass args

---

• 修复 shell 长时间执行命令时的 status 状态判定，避免被误判成终止。
• 修复启动时 /home/${USER} 文件过多时造成的卡顿问题
• 浏览器移除 --proxy-bypass 参数

v1.0.0.144

• Fixed proxy log file could not be written
• Append user-level executable directory '~/.local/bin' to'PATH '
• Security enhancements, removal of browser '--no-sandbox' parameter
• Added system dependencies related to ffmpeg, latex, and graphics rendering, which can be used to generate video animations using manim

• Added some code examples used by AIO examples

---

• 修复 proxy 日志文件无法写入问题
• 追加用户级可执行目录 ~/.local/bin 到 PATH
• 安全性增强，移除浏览器 --no-sandbox 参数
• 增加 ffmpeg、latex、图形渲染相关的系统依赖，可使用 manim 于生成视频动画
• 新增 AIO 使用的一些代码示例 examples

v1.0.0.143

• Add pyecharts dependency and include the DejaVu Sans font.
• Fixed error reported after container restart mv: No such file or directory

---

• 增加 pyecharts 依赖，新增 DejaVu Sans 字体
• 修复容器重启后报错 mv: No such file or directory 问题

v1.0.0.142

• POST /v1/shell/exec
  • Add a timeout parameter, which triggers a soft exit on timeout (returns status: running)
  • Fix the issue where exec_dir (execution directory) was not taking effect.
• POST /v1/shell/wait
  • Fix the inconsistency between the status and success return values for running tasks
• Fix the problem where the browser proxy PROXY_SERVER environment variable was not taking effect

---

• POST /v1/shell/exec
  • 增加 timeout 参数，超时软退出（返回 status: running）
  • 修复 exec_dir（执行目录）无法生效的问题。
• POST /v1/shell/wait
  • 修复运行中任务的 status 与 success 返回值不一致问题
• 修复浏览器代理PROXY_SERVER环境变量不生效问题