【2026犀牛鸟issue】CubeSandbox 与 Claude Code / CodeBuddy / OpenCode 集成

[fslongjin]

【产品名称/功能模块】
CubeSandbox 与 Claude Code / CodeBuddy / OpenCode 集成

【问题描述】
CubeSandbox 定位为面向 AI Agent 的安全沙箱服务，并与 e2b 协议兼容。Claude Code、CodeBuddy、OpenCode 是当前主流的终端型 AI 编码 Agent——它们在隔离环境中运行代码、编辑文件、执行命令，天然适合以 CubeSandbox 沙箱作为执行后端或运行环境：既可以把编码 Agent 本身跑在沙箱里获得隔离、可复现的运行环境，也可以让 Agent 编排系统按需拉起 CubeSandbox 沙箱执行其生成的代码。

目前 CubeSandbox 已有面向 OpenAI Agents、OpenClaw 等的集成示例，但尚未提供 Claude Code / CodeBuddy / OpenCode 这三类终端型编码 Agent 的集成文档与可运行示例，社区缺少可直接参照的"上手即用"材料。本 issue 为开放性文档与示例贡献任务，允许多人并行参与、各自认领一个子方向，以增量方式合入，不要求单次提交完整 PR。

【期望内容/功能改进】（每人可认领其中一个或多个子方向，渐进式贡献）
1）为 Claude Code、CodeBuddy、OpenCode 中至少一个 Agent 编写端到端集成指南，覆盖：在 CubeSandbox 中运行该 Agent 所需的模板构建、运行环境与依赖、API Key 等敏感配置的注入方式、网络与出口策略（Agent 需访问 LLM API）、以及会话保持与状态持久化（借助快照保存 Agent 运行状态、跨会话恢复）；
2）提供可运行的示例工程（置于 examples/ 下），包含模板定义与构建脚本、最小可复现的调用流程，以及能在本地部署环境跑通的 README；
3）给出典型使用场景与最佳实践：如把编码 Agent 跑在沙箱内做隔离开发、用沙箱执行 Agent 生成的代码并回收结果、利用快照实现 Agent 长时任务的断点续跑等；
4）补充常见问题与排错（鉴权失败、出口被拦截、镜像拉取、超时等），并与平台既有出口管控、鉴权、模板机制对齐；
5）文档纳入 docs/guide/integrations 体系，遵循既有集成文档模板与风格。

【参与说明与验收】
1）本题为开放参与：多人可同时认领不同 Agent 或不同子方向，贡献以增量合入，不要求单人完成全部内容；
2）每份贡献应做到：文档表述准确、与当前 CubeSandbox 版本行为一致、示例可在本地部署环境复现；
3）单份贡献的最低门槛：至少一个 Agent 的完整集成指南 + 一个可运行示例 + README，其他同学可依照文档在合理时间内跑通；
4）贡献不应与已有 OpenAI Agents / OpenClaw 等示例重复，建议在 issue 下认领以避免撞车。

【认领规则】
本issue为2026犀牛鸟开源人才培养活动专属issue，仅供已报名参与犀牛鸟活动的同学认领
【认领时间】7月1日～7月31日（7月1日前认领视为无效❗）
【认领方式】在本issue评论区回复“已认领本任务”，即视为认领成功
【活动报名】需提前完成犀牛鸟报名问卷，问卷将用于活动登记和奖励发放：https://wj.qq.com/s2/26888567/gh2q
【微信添加Cube社区小助手进入交流群】：CubeSandbox

---

【Product Name / Feature Module】
CubeSandbox Integration with Claude Code / CodeBuddy / OpenCode

【Problem Description】
CubeSandbox is positioned as a security sandbox service for AI Agents and is compatible with the e2b protocol. Claude Code, CodeBuddy, and OpenCode are mainstream terminal-based AI coding Agents—they run code, edit files, and execute commands in isolated environments, making CubeSandbox a natural fit as an execution backend or runtime: you can run the coding Agent itself inside a sandbox for an isolated, reproducible environment, or have an Agent orchestration system spin up CubeSandbox sandboxes on demand to execute generated code.

CubeSandbox already has integration examples for OpenAI Agents, OpenClaw, and others, but there is no integration documentation or runnable examples yet for these three terminal-based coding Agents. The community lacks ready-to-use reference materials. This issue is an open documentation and example contribution task. Multiple contributors may participate in parallel, each claiming a sub-direction and merging incrementally— a single PR covering everything is not required.

【Expected Content / Feature Improvements】 (Each contributor may claim one or more sub-directions and contribute incrementally)

1. Write an end-to-end integration guide for at least one of Claude Code, CodeBuddy, or OpenCode, covering: template build requirements for running the Agent in CubeSandbox, runtime environment and dependencies, injection of sensitive configuration such as API keys, network and egress policies (Agent needs access to LLM APIs), and session persistence (using snapshots to save Agent runtime state and resume across sessions);
2. Provide a runnable example project (under examples/), including template definitions and build scripts, a minimal reproducible workflow, and a README that works in a local deployment;
3. Document typical use cases and best practices: e.g., running a coding Agent inside a sandbox for isolated development, using sandboxes to execute Agent-generated code and collect results, using snapshots for checkpoint/resume of long-running Agent tasks, etc.;
4. Add FAQ and troubleshooting (auth failures, egress blocked, image pull issues, timeouts, etc.), aligned with the platform's existing egress control, authentication, and template mechanisms;
5. Integrate documentation into the docs/guide/integrations system, following existing integration doc templates and style.

【Participation Guidelines & Acceptance】

1. This is an open contribution: multiple people may claim different Agents or sub-directions concurrently; contributions merge incrementally— no single person is expected to complete everything;
2. Each contribution should be: accurate, consistent with current CubeSandbox behavior, and reproducible in a local deployment;
3. Minimum bar per contribution: a complete integration guide for at least one Agent + one runnable example + README, so others can follow the docs and get it working within a reasonable time;
4. Contributions should not duplicate existing OpenAI Agents / OpenClaw examples; claim your direction in the issue comments to avoid overlap.

【Claim Rules】
This issue is exclusive to the 2026 Tencent Rhino Bird Open Source Talent Training Program and may only be claimed by students who have registered for the Rhino Bird program.
【Claim Period】July 1 – July 31 (claims before July 1 are invalid ❗)
【How to Claim】Reply "已认领本任务" (I have claimed this task) in the comments on this issue to claim successfully
【Program Registration】Complete the Rhino Bird registration questionnaire in advance; it will be used for program enrollment and reward distribution: https://wj.qq.com/s2/26888567/gh2q
【Join the Cube Community Group on WeChat】Add the Cube community assistant on WeChat (WeChat ID: CubeSandbox) to join the discussion group