# CubeSandbox upstream status / 上游状态 Checked on 2026-11-14 / 核对日期:2026-11-14. ## v0.8.0:Adapter 兼容性和快照实测(2026-11-14) sr1 使用单个 x86_64 计算节点、Pod 网络和 XFS 模板;CubeSandbox 后端及 PyPI Python SDK 均为 v0.8.0。测试启动当前工作区源码的独立本地 Adapter, 经 HTTP 和 MCP stdio 调用真实后端,不修改已有的生产 Adapter 部署。 | 实测项目 | 结果 | | --- | --- | | 鉴权、租约复用、命令、文件和二进制附件 | 通过 | | 异步任务完成和取消、PTY 创建/输入/调整/关闭 | 通过 | | 完整快照与回滚 | 通过,回滚后读取原文件 | | 新沙箱 fs-only 快照和分支 | 通过,快照前文件保留、后续修改隔离,分支未继承原进程 | | MCP stdio 创建、列出、删除 fs-only 快照;远程 MCP 列表调用 | 通过 | | 可信任务执行、清理和签名回执 | 通过 | | 审计记录与资源清理 | 通过,测试创建的三个租约均清理;额外快照按名称核对删除 | `cube_checkpoint` 默认仍为完整快照,可显式传 `mode: "fs-only"`;新加 `cube_checkpoint_list`、`cube_checkpoint_delete`,各自要求 `checkpoint:list`、 `checkpoint:delete` 权限。文件系统快照不保存内存和进程,不可原地回滚; `cube_fork` 从它冷启动新沙箱。外部挂载不随之保存,不能拿它承诺进程断点续跑。 旧记录没有 `mode` 字段时按完整快照读取;服务端未确认 fs-only 模式时会删除返回的 快照并拒绝成功响应。旧模板仍可能携带旧 shim,需在升级后单独验收。 另两次按“创建完整快照 → 回滚 → 在同一沙箱创建 fs-only 快照”的顺序测试, 上游返回 `snapshot_filesystem_only_rootfs_unsupported`;Adapter 返回 502。 失败操作留下的快照记录使结果存在不确定性,必需审计模式随后阻止常规清理; 本次按测试专用标识核对并清理了沙箱和快照。这个现象只针对本次 XFS 和测试顺序, 原因尚未进一步定位;不要自动重试或将其泛化为所有存储后端。遇到类似问题应保留 审计记录,核对实际沙箱和快照状态后再按运维流程处理。 [脱敏验收结果](evidence/v080-adapter-sr1-2026-11-14.json); [可重复执行的实测脚本](../tests/acceptance/cubesandbox_v080_adapter_live.py)。 本轮没有覆盖生产升级、跨节点恢复、所有旧后端和所有原生客户端界面;单元测试 在 Python 3.10、3.12、3.14 上各运行 92 项(85 通过、7 项因未配置存储而跳过), 这 7 项随后在本机独立 Redis 与 PostgreSQL 容器中全部通过。上游升级步骤和实测见 [CubeSandbox v0.8.0 升级实战](https://aik8s.run/ai-k8s/rag-agent/cubesandbox-v080-upgrade-practice/)。 ### English summary Against one sr1 x86_64 compute node with Pod networking and XFS, the current-source Adapter passed real HTTP lifecycle, file and artifact operations, async jobs and cancellation, PTY operations, full snapshot rollback, filesystem-only snapshot fork, MCP stdio snapshot create/list/delete, remote MCP listing, a trusted task and signed receipt, and audit. All three test leases and the run's snapshots were cleaned. This did not exercise production, cross-node recovery, every older backend, or every native client UI. The default checkpoint remains `full`; `fs-only` captures root filesystem state, not running processes or memory. It supports a cold-start fork and is rejected for in-place rollback. External mounts remain separate. A subsequent fs-only snapshot after a full rollback failed twice with `snapshot_filesystem_only_rootfs_unsupported` on the tested XFS backend, and the uncertain audited operation blocked normal cleanup. The test resources were reconciled separately. Do not automatically retry or assume the same outcome across storage backends. ## v0.7.3:S3lvol 单节点验收(2026-10-26) 这次单独启用 CubeS3lvol,新建 `backend=s3` 模板和独立测试 Bucket, 验证 VM 快照的 S3 存储链路。此前的 S3 外部卷测试只验证沙箱挂载对象存储, 不能代替这项验收。测试限于现有一个计算节点,没有扩节点或改生产环境。 后端、CubeS3lvol 和 Python SDK 均为 v0.7.3。沙箱规格为 1 vCPU、512 MiB 内存、1 GiB 可写层;模板产物仍使用原有 S3 artifact store。 测试池 journal / WAL / cache 分别为 128 / 2048 / 4096 MiB。 S3 后端在创建模板时指定,后续沙箱、快照和暂停包继承该选择; 控制面记录的 `remote_status=ready` 是每次恢复前核对的条件。 [官方存储与调度说明](https://docs.cubesandbox.com/guide/cross-node-snapshot) 介绍了后端继承及源节点优先的行为。 | 实测项 | 结果 | | --- | --- | | S3 模板构建、创建沙箱、执行命令 | 通过 | | 写入 2 MiB 文件、创建快照并等待远端就绪 | 通过,远端就绪等待 4.83 秒 | | 修改文件后 Rollback | SHA-256 一致 | | 从快照创建新沙箱 | SHA-256 一致 | | Clone 后修改克隆文件 | 原沙箱文件不变 | | 删除仍被恢复实例使用的快照 | 现有实例可读,再用该快照创建返回 404 | | 连续 3 轮 Pause / Resume | 文件校验均通过;暂停包远端就绪等待 4.91 / 4.79 / 7.09 秒 | | API 清理 | 3 个测试沙箱均返回 404,测试模板已注销,暂停包记录为 0 | | 底层自动回收 | 未完成,见下文 | 这些等待时间从快照或暂停 API 返回后开始计算,包含状态轮询和端口转发, 不是快照或暂停操作的总耗时。本轮恢复留在源节点,读取以本地缓存命中为主, 尚未验证冷拉取、跨节点恢复、宿主机掉电或满环 WAL 重放。 ### 删除返回成功,回收仍有残留 测试结束后只通过正常 API 删除自建沙箱、快照和模板,再持续观察约 301 秒。 `rcow_get_exports` 已为空,`rcow_get_lvstores` 只剩 2 个引擎基础卷, 但 `rcow_get_pending_deletes` 仍有 2 条 `reason=failed` 记录。 日志中对应的异步删除返回 `Device or resource busy`,SPDK 随后报告 forced removal。 独立 Bucket 从基线 4 个对象、1,073,312 字节,变为 215 个对象、209,761,652 字节 (200.04 MiB),观察期间没有继续下降到基线。 随后执行一次官方 `s3lvol_rpc.py --retry-pending`。命令退出码为 0,提示 `no pending snapshot deletes to retry`,但再次查询仍有 2 条失败意图。 这两个 lvol 已不在卷清单中,而工具从卷清单选择可重试对象,因此这次重试没有 消除残留。此处记录的是观察窗口内的回收不完整,尚未确认残留对象的完整归属及 是否会永久泄漏;没有手工删数据库记录或删除 S3 对象。 [v0.7.3 的 S3lvol 运维说明](https://github.com/TencentCloud/CubeSandbox/blob/v0.7.3/CubeS3lvol/README.md#retrying-a-refused-snapshot-delete---retry-pending) 也说明删除可能延迟,异步失败需要显式处理,集群删除路径不会自动运行重试工具。 本次确实观察到了清理残留,但没有复现“必须删库、清空 S3 才能恢复”的情况。 上游另有 [#1971](https://github.com/TencentCloud/CubeSandbox/issues/1971): 维护者确认 v0.7.3 的 WAL 重放会重新应用已被 checkpoint 覆盖的旧记录, 可能覆盖较新的元数据并导致 `EILSEQ`;不是 issue 最初猜测的 flusher 等待边界。 截至 2026-10-26,相关 [#1996](https://github.com/TencentCloud/CubeSandbox/pull/1996) 仍未合并。这次正常路径测试没有触发该场景。 ### 证据与环境收尾 - [脱敏实测结果](assets/cubesandbox-s3lvol/v073-single-node-2026-10-26.json) 分别保存功能结果、回收观察序列和显式重试结果。 - [Light 模式报告源文件](assets/cubesandbox-s3lvol/v073-single-node-2026-10-26-light.html) 与下图来自同一份结果数据。 - 测试后恢复原部署配置,关闭临时启用的 S3lvol;原有模板、镜像配置和凭据保留。 恢复后再用原模板验证创建、执行、删除。独立测试 Bucket、WAL 和日志保留供排查。 ![CubeSandbox v0.7.3 S3lvol 单节点实测报告,Light 模式](assets/cubesandbox-s3lvol/v073-single-node-2026-10-26-light.png) ## v0.7.3:升级和实测(2026-10-14) [v0.7.3](https://github.com/TencentCloud/CubeSandbox/releases/tag/v0.7.3) 于 2026-10-12 发布。本次测试集群和生产集群均从 v0.7.2 升到 v0.7.3, Helm revision 分别为 9 和 18。生产三个计算节点均健康、已恢复调度。 MySQL、Redis、MinIO、现有模板和 Adapter 审计数据保留,没有清空重装。 ### 新特性和兼容性 - 新增沙箱内 MCP 网关:SDK 创建沙箱时可启动多个 stdio MCP server, 通过一个 HTTP 入口调用聚合工具。需要包含 `mcp-gateway` 的模板,并暴露 `50005` 端口;把程序写入普通代码沙箱后做快照,不会自动增加暴露端口。 [官方 MCP 文档](https://github.com/TencentCloud/CubeSandbox/blob/v0.7.3/docs/guide/mcp-gateway.md) 说明这是预览功能。同一沙箱中的 server 默认以 root 运行、共享一个信任域, 不能代替 Adapter 的权限、审批和持久审计。 - 修复暂停后丢失元数据、模板信息和挂载信息;接受快照恢复时重新发送卷挂载。 已删除快照创建沙箱返回 `404`,不再返回 `500`。 - 网络修复包括端口映射键加入 sandbox IP、网卡复用后的连接问题和连接回收。 克隆资源选择、shim 存活时的资源回收和模板列表查询也有修复。 - 模板分页适用于 CubeMaster `/cube/template` 和 `cubemastercli`。 E2B `/templates?limit=1` 仍返回全量列表;不能把两套接口混用。 - [PyPI SDK 0.7.3](https://pypi.org/project/cubesandbox/0.7.3/) 也已在 2026-10-12 发布。v0.7.3 Git tag 中的 `pyproject.toml` 仍写 `0.7.0`, 因此记录区分源码 SDK、已发布包和后端版本。线上 Adapter 本次保留原镜像及 `cubesandbox==0.7.0`,实际租约、执行、文件和异步任务回归通过;新包单独验证。 生产审计最终核对为:未完成操作 0、待投递记录 0、阻断状态 0。 ### 部署过程 19 个组件镜像通过中转机 Registry → Registry 复制,没有在本地拉镜像层。 固定 linux/amd64 manifest,核对源/目标 config 和 layer digest 后,经镜像服务 同步到生产。生产 WebUI 派生镜像保留原有内部模板镜像地址,不恢复外网地址。 升级前保存 Helm 配置,并在 MySQL 持久卷保留逻辑备份。先升级 CubeOps, 再更新控制面;计算节点在 CubeOps 内隔离超过 60 秒,并确认沙箱库存为空。 生产临时使用 OnDelete,逐个更新 bootstrap、installer 和 cube-node,确认就绪 后解除隔离,最终恢复原 RollingUpdate 策略。节点标签、存储类、宿主机内核和 `cubeNode.hostNetwork: false` 均未改动。 生产第一次预检查失败是 Helm `--reuse-values` 把已置空的默认 cube 标签选择器 重新渲染成空字符串;不是 CPU 不足。使用保存的完整配置配合 `--reset-values` 重新渲染后,只保留原有的节点选择条件,预检查通过。 按照[官方升级指南](https://github.com/TencentCloud/CubeSandbox/blob/v0.7.3/docs/guide/kubernetes/upgrade.md), Pod 网络模式必须先排空沙箱,不能依靠 Pod 重建保留原沙箱网络。 ### 实测结果 | 检查 | 测试集群 | 生产集群 | | --- | --- | --- | | 创建、执行、删除 | 5 轮通过 | 3 轮通过 | | 暂停元数据、恢复文件 | 通过 | 通过 | | Snapshot、Rollback、Restore、Clone 和根文件隔离 | 通过 | 通过 | | S3 外部卷当前数据、克隆共享、重新发送挂载恢复 | 7 项通过,7 项清理完成 | 同左 | | 已删除快照 / 非法 MCP 参数 | 404 / 400 | 404 / 400 | | CLI 模板分页 | limit=1 返回 1 条 | limit=1、offset=1 返回 1 条 | | MCP 网关和工具调用 | 新建离线模板通过 | 本轮未部署 MCP 模板 | | Adapter 可信任务 | 19 项通过 | 租约、执行、文件、异步任务、释放通过 | 测试集群升级前后采用同一组 canary:暂停元数据从丢失变为保留; 已删除快照返回码从 500 变为 404;非法 MCP 参数从被忽略变为 400。 创建 API 的 5 轮中位数为 146.33 ms,删除为 130.53 ms,首条命令为 88.44 ms;均是经过本地端口转发的客户端耗时,测试配置不是官方基准。 MCP 用两个无需外网的 Echo server 验证同名工具聚合为 `echo` 和 `beta_echo`, 两个工具均实际返回预期标记。缺少网关令牌返回 401,缺少代理流量令牌返回 403。 新建 MCP 模板及相关测试沙箱已删除,原有模板未改动。 PVM 宿主机内核更新、S3lvol 热升级和大规模跨节点性能不在本轮运行配置中。 本轮没有重新运行所有 Agent 客户端 TUI,也没有重做 Claude 验证。 源码 SDK 的 288 项非 E2E 测试通过;Adapter 本地 53 项测试中 52 项通过、 1 项可选 Redis 集成跳过,Hermes 2 项、验收解析 3 项、可信执行 5 项及 OpenClaw/DSH 插件、安装脚本、Lint、Helm 检查通过。 已发布 SDK 0.7.3 也通过同一组 288 项 SDK 测试、53 项 Adapter 测试 (同样跳过可选 Redis 集成)及 5 轮真实生命周期/快照回归;没有因此替换线上 Adapter 镜像或发布新的 Adapter 版本。 ### 证据和复现 - [升级前基线](assets/cubesandbox-upstream/v072-before-v073-2026-10-14.json)、 [测试集群结果](assets/cubesandbox-upstream/v073-sr1-results-2026-10-14.json)、 [生产结果](assets/cubesandbox-upstream/v073-production-results-2026-10-14.json)。 - [测试卷快照结果](assets/cubesandbox-upstream/v073-volume-results-2026-10-14.json)、 [生产卷快照结果](assets/cubesandbox-upstream/v073-production-volume-results-2026-10-14.json)。 - [MCP 结果](assets/cubesandbox-upstream/v073-mcp-results-2026-10-14.json)、 [可信任务 19 项验收](assets/cubesandbox-upstream/v073-trusted-acceptance/results.json)、 [生产 Adapter 结果](assets/cubesandbox-upstream/v073-production-adapter-results-2026-10-14.json)、 [已发布 SDK 0.7.3 实测](assets/cubesandbox-upstream/v073-published-sdk-results-2026-10-14.json)。 真实控制台版本页,Light 模式,内部标识已遮盖: ![v0.7.3 控制台版本矩阵](assets/cubesandbox-upstream/v073-webui-versions-light.png) 快照、MCP 和生产 Adapter 的真实结果报告,Light 模式: ![v0.7.3 快照、MCP 与生产 Adapter 实测](assets/cubesandbox-upstream/v073-snapshot-mcp-light.png) 可信任务验收报告,Light 模式;数据来自实际 Adapter 请求: ![v0.7.3 可信任务验收](assets/cubesandbox-upstream/v073-trusted-tasks-light.png) 通用回归脚本为 `tests/acceptance/cubesandbox_release_smoke.py`,使用 SDK 的 `CUBE_API_URL`、`CUBE_PROXY_NODE_IP`、`CUBE_PROXY_PORT_HTTP`、`CUBE_TEMPLATE_ID` 连接变量。卷测试增加 `--backend` 和 `--restore-with-mounts`,保留旧版本默认行为。 MCP 脚本 `tests/acceptance/cubesandbox_mcp_gateway_smoke.py` 要求传入预先构建、 暴露 50005 的离线模板,搭配 `mcp_echo_server.py`;不会在执行时下载 MCP server。 ## Historical release channels / 历史发布通道(2026-10-08) [CubeSandbox v0.7.2](https://github.com/TencentCloud/CubeSandbox/releases/tag/v0.7.2) is the latest stable release as of 2026-10-08. It was published on 2026-09-24; there has been no later stable release. It adds: - faster S3-backed cross-node restore reads through a local object cache and coalesced GETs; upstream reports under one second to `running` with its default MinIO setup, which is not a measurement of this cluster; - a JuiceFS persistent-volume plugin and automatic reclaim of idle guest memory; - fixes for incremental snapshot data loss, snapshot/template references, template-build races, S3lvol hot-upgrade I/O errors and overlay-CNI guest MTU; - the Kubernetes `cube-node` host-network default and a faster guest cold-start path; upstream also documents an OpenCode Bash redirection hook. This Adapter's OpenCode MCP tools retain separate server-side approval and audit. The upstream `master` branch has additional commits through 2026-10-08, but they are **not part of the v0.7.2 tag**. Relevant examples are [#1861](https://github.com/TencentCloud/CubeSandbox/pull/1861) (preserve sandbox metadata across pause), [#1894](https://github.com/TencentCloud/CubeSandbox/pull/1894) (accept re-sent volume mounts during snapshot restore) and [#1810](https://github.com/TencentCloud/CubeSandbox/pull/1810) (avoid reclaiming sandbox resources while its shim may still be alive). Do not rely on these changes in a deployed v0.7.1 backend or a stock v0.7.2 image. [CubeSandbox v0.7.1](https://github.com/TencentCloud/CubeSandbox/releases/tag/v0.7.1) was released on 2026-09-11 and remains this project's last backend with the full 30-check Adapter/client acceptance suite. Its key changes were: - 独立的 CubeTemplateCenter 支持多副本 CubeMaster,生命周期管理支持主备; - 带 Host Mount 或 Volume Plugin 的 Sandbox 可以执行 Snapshot、Restore、 Rollback 和 Clone; - 模板构建产物可以存入 S3/MinIO,并修复并发构建竞争; - 同 Bucket 跨节点恢复加速,Snapshot 删除和回收语义得到完善; - 增加 Kubernetes CubeS3lvol、外部 PostgreSQL 和统一的 `redis.db` 配置; - 修复恢复后日志、非法 CPU/内存、Resume 超时、S3 增量快照、LLM 长响应、 Registry 凭据日志泄露等问题。 [CubeSandbox v0.7.2-rc1](https://github.com/TencentCloud/CubeSandbox/releases/tag/v0.7.2-rc1) was reviewed on 2026-09-23. Its changes are now part of the stable v0.7.2 release, including the following upgrade and SDK considerations: - Kubernetes `cube-node` now defaults to `hostNetwork: true`. Existing releases must either pin `cubeNode.hostNetwork: false`, or drain every compute node and set the one-time `cubeNode.hostNetworkChangeAck: true` migration acknowledgement. - Host networking preserves the sandbox TAP devices and cubevs hooks when the Big Pod is recreated, but Kubernetes NetworkPolicy no longer governs sandbox traffic. Host ports, Service/sandbox CIDR overlap, Cilium/eBPF interaction and Pod Security admission must be validated before adopting the new default. - Snapshot correctness work freezes the VM across memory and rootfs capture and preserves rootfs artifacts referenced by snapshots. S3lvol gains hot-upgrade I/O pausing, faster restore reads and namespace-ID reuse. - The Python SDK source now always uses CubeSandbox's direct Connect API for commands and adds optional user selection to more filesystem calls, removing the optional E2B command implementation path. - The upstream repository now documents an OpenCode `tool.execute.before` hook that redirects the built-in Bash tool. This Adapter's OpenCode integration is different: it exposes explicit MCP tools, server-owned policies, approval, durable audit and receipts instead of transparently rewriting Bash. Although `sr1` is absent from the current gmanctl workspace, it remains reachable through the local `kubectl --context sr1` kubeconfig. On 2026-10-08, sr1 was upgraded from v0.7.1 to v0.7.2 with `cubeNode.hostNetwork: false` preserved; the reachable `gd18-ai-002` production deployment remained on v0.7.1. A future host-network migration requires draining each compute node and explicitly acknowledging the change. The upstream guide still requires a compute-node drain even when host networking is already in use, pending live validation of sandbox survival through a cubelet restart. On 2026-10-09, `gd18-ai-002` was also upgraded to v0.7.2. On 2026-10-10, the Adapter was recovered from one audited operation left unresolved after its five-minute backend sandbox expired before the one-hour lease idle TTL. The original audit event was retained and reconciled only after CubeAPI confirmed the sandbox was gone. A candidate Adapter fix passed a live expired-lease, re-acquire, command and cleanup check on that cluster. This is a targeted regression check, not the full Adapter/client acceptance suite. ### Backend release and SDK package are different versions The v0.7.2 Git tag still declares Python package version `0.7.0`, and [PyPI](https://pypi.org/project/cubesandbox/) still publishes `0.7.0` as the latest SDK on 2026-10-08. The Adapter therefore keeps `cubesandbox==0.7.0`: backend and SDK release numbers are separate. The v0.7.2 SDK changes are only available from source, including direct Connect RPC command execution and Connect timeout handling. An isolated check using the stable v0.7.2 SDK source passed all 47 runnable Adapter Python tests (one optional Redis integration test skipped); the upstream SDK suite passed 269 non-E2E tests. The SDK was also used for the sr1 v0.7.1/v0.7.2 live canaries below. These are not the full Adapter/client acceptance suite. CubeMaster continues to proxy the existing template APIs to CubeTemplateCenter, so the upstream HA split does not require an Adapter API change. ### 2026-10-08 sr1 recovery, upgrade and live canaries The first v0.7.1 canary was blocked before sandbox creation: CubeMaster returned `130400`, because CubeOps marked the only node `HeartbeatExpired` even though the `agent-code` template and its local replica were `READY`. The last persisted heartbeat was 2026-09-23; node-status upserts were failing. The sr1 test MySQL 10 GiB PVC was full, with about 9.5 GiB of binary logs and about 15 MiB of application database files. This was an environment failure, **not** a reproduction of [#1893](https://github.com/TencentCloud/CubeSandbox/issues/1893). With approval for test-cluster cleanup, the MySQL PVC was expanded to 12 GiB after the full disk prevented MySQL from starting. Old binary logs were then purged with MySQL's `PURGE BINARY LOGS` command, leaving the active log and about 2% of the PVC used. Retention was persistently shortened from 30 to 7 days. No binlog files were manually deleted. A compressed logical database backup was validated and retained outside this repository before the upgrade. Node heartbeat recovered to `Healthy: true`. All 19 v0.7.2 Cube-owned images were copied through the image relay to the private test registry, with source/target digest checks; sr1 actually pulled the `cube-api` image in a temporary probe Pod. Helm release `cube` was upgraded from revision 6 / v0.7.1 to revision 7 / v0.7.2 using the current user values on top of new chart defaults. The pre-upgrade dry run and deployed release both kept `hostNetwork: false`. All CubeSandbox Pods became Ready, the node reported Healthy, and all eight enabled upstream Helm test groups passed. Using the v0.7.2 Python SDK source and the existing `agent-code` template, five serial create/first-command/delete rounds passed both before and after the backend upgrade. The fifth round in each set also passed pause, resume and a command after resume. Mean create times were 190 ms on v0.7.1 and 137 ms on v0.7.2, but five samples are a functional canary, **not** a performance claim. On v0.7.2, a separate canary passed snapshot creation/listing, rootfs rollback, creation from that snapshot, and snapshot deletion. The sandbox and snapshot counts returned to their pre-test baselines. The redacted results are in [`v072-results-2026-10-08.json`](assets/cubesandbox-upstream/v072-results-2026-10-08.json). These results do not prove the open issues are closed. The reused `agent-code` template predates v0.7.2; sr1 TemplateCenter cannot anonymously pull its private base image to build a fresh v0.7.2 template, and the affected artifact from #1893 was not available. No guest-agent failure, failed-resume storage fault, or shim-log rotation was injected for #1890, #1847, or #1896. On 2026-09-12, the Kubernetes acceptance environment was upgraded from v0.7.0 to v0.7.1 while the Adapter kept the published `cubesandbox==0.7.0` SDK. All CubeSandbox workloads became Ready, all eight upstream Helm test groups passed, and the Adapter/backend suite passed 30/30 checks. The publication captures and the mounted-volume result are recorded in the [trusted-execution guide](trusted-execution.md#live-acceptance-evidence). The same mounted-volume scenario was rerun on 2026-09-23 without upgrading the backend. It again passed snapshot creation, rootfs rollback, external-volume current-state semantics, clone/remount, referenced-snapshot deletion and all six cleanup checks. Before and after the run the cluster had zero active sandboxes, zero volumes and the same two pre-existing snapshots. The redacted machine-readable result is [`snapshot-results-2026-09-23.json`](assets/cubesandbox-upstream/snapshot-results-2026-09-23.json). ### Mounted-workspace checkpoint semantics v0.7.1 closes [#1599](https://github.com/TencentCloud/CubeSandbox/issues/1599), but a mounted-workspace checkpoint is an **external-reference checkpoint**: - VM memory and rootfs are restored to the checkpoint state; - files below a Host Mount or Volume Plugin mount are not copied into the checkpoint and remain at their latest external state; - upstream preserves origin-node affinity for snapshots with Host Mounts and external volumes; clones can deliberately share writable mounted data; - point-in-time data consistency still needs a storage-provider snapshot or an application freeze coordinated outside this Adapter. The Adapter therefore keeps `allow_checkpoint_with_mounts: false` as the safe default. On a v0.7.1-or-newer backend, an operator may enable it only after validating node affinity, volume remounting, read-only behavior and the desired data-consistency model: ```yaml profiles: durable-checkpoint-code: checkpoints_enabled: true allow_checkpoint_with_mounts: true workspace: mode: per-session-volume retain_on_kill: true ``` The Adapter does not currently perform a backend version handshake. Enabling the gate against an older backend will pass the request upstream and fail there. ### Upgrade notes that matter to operators - For v0.7.1 to v0.7.2 on the current Pod-network deployment, explicitly set `cubeNode.hostNetwork: false` in Helm values. Changing it to `true` requires per-node isolation, at least 60 seconds of drain, destruction of resident sandboxes and the one-time `cubeNode.hostNetworkChangeAck: true`. The flag acknowledges the migration; it does not perform the drain. Check host-port, Cilium/eBPF and policy behavior before choosing host networking. - Canary the control and compute-plane images in the order in the upstream v0.7.2 guide. Rebuild at least one template with the new components, then repeatedly test create, first command, pause/resume, snapshot, clone and cleanup before moving an alias. Check Cubelet/shim process counts and node resources after failed creates, not just API success rates. - The v0.7.2 release's sub-second cross-node restore claim is tied to its default MinIO and S3 cache configuration; benchmark our own storage and network path before using it as a capacity or latency target. - CubeS3lvol is now opt-in. Existing nodes that used S3 implicitly must set `[cow.s3] enable = true`, or `ONE_CLICK_ENABLE_S3LVOL=1` for one-click deployments, before relying on cross-node S3 restore. - CubeTemplateCenter becomes the template build path and needs durable S3/MinIO artifact storage. Existing CubeMaster template endpoints are proxied for client compatibility. - A Kubernetes compute-plane image or Pod-template upgrade can recreate the `cube-node` DaemonSet Pod and interrupt existing sandboxes. Isolate and drain nodes according to the upstream upgrade guide; do not treat v0.7.1 control- plane HA as compute-plane live migration. - The new top-level Helm `redis.db` should be set explicitly when several CubeSandbox clusters share one Redis instance. ### Findings from the live v0.7.0 to v0.7.1 upgrade - The first Helm upgrade hit an immutable Redis StatefulSet `volumeClaimTemplates` change when v0.7.1 introduced stable labels. The test environment recorded the Redis Pod/PVC identities, orphan-deleted only the StatefulSet, reran the forward upgrade, and verified that the same Pod and PVC were adopted. Back up Redis first and repeat those identity checks before and after the change; do not delete the PVC. - Database migrations had already advanced when the first upgrade failed. The v0.7.0 CubeMaster could not start against that newer schema, so automatic binary rollback was not a usable recovery path. Take a database backup and prepare a forward-fix procedure before upgrading. - Templates retain their agent and shim component versions. A template built on v0.7.0 could still start a sandbox, but v0.7.1 rejected a snapshot clone whose metadata named the older agent. Rebuild the template after the control and compute planes are upgraded, verify its component annotations, then move the stable alias. Keep the old template unaliased until rollback decisions are complete. - Private-image template builds require registry credentials. Supply them at execution time and keep them out of values files, command transcripts and screenshots. - Treat CubeMaster and CubeTemplateCenter startup output as sensitive. This run observed effective configuration values in startup logs, so raw logs were not retained or published. Restrict log access and retention, inspect masking in the exact release, and rotate any credential whose value has entered an exported log. ## Issues to follow / 建议持续跟进的 Issue State was re-checked through the GitHub API on 2026-10-08. The newly added open reports below used older backend versions, so their presence on v0.7.2 is not established by the issue reports alone. | Issue | State / 状态 | Why it matters here / 对本项目的影响 | Current handling / 当前处理 | | --- | --- | --- | --- | | [#1395 Declarative runtime profiles](https://github.com/TencentCloud/CubeSandbox/issues/1395) | Open; last updated 2026-08-27. | Upstream-native policy profiles could reduce duplicated policy mapping. | Adapter profiles remain operator-owned YAML and translate into SDK arguments. | | [#1599 Snapshots with volumes/host mounts](https://github.com/TencentCloud/CubeSandbox/issues/1599) | Closed 2026-09-08; delivered by [#1654](https://github.com/TencentCloud/CubeSandbox/pull/1654) and [#1656](https://github.com/TencentCloud/CubeSandbox/pull/1656). | Mounted checkpoint operations now work, but external data stays current rather than rolling back. | Keep the explicit opt-in gate and validate the external-reference semantics on v0.7.1 before enabling it. | | [#1598 Restored stdout/stderr logs](https://github.com/TencentCloud/CubeSandbox/issues/1598) | Closed 2026-09-02; the fix is listed in v0.7.1. | Restored sandbox log output is repaired upstream. | Adapter durable jobs still persist their own stdout/stderr/exit files for independent recovery. | | [#1414 CubeAPI default unauthenticated](https://github.com/TencentCloud/CubeSandbox/issues/1414) | Closed 2026-09-03. | Control-plane authentication remains a deployment boundary even after the upstream issue closes. | Keep `CUBE_API_KEY`, private reachability, TLS/mTLS or OIDC and restrictive NetworkPolicy. | | [#1484 Non-positive SDK command timeout](https://github.com/TencentCloud/CubeSandbox/issues/1484) | Closed 2026-08-28; v0.7.1 contains Node/Python source fixes. | The fix has not been published as a Python 0.7.1 package. | Adapter request and profile validation already require positive command timeouts, so this path is not used. | | [#1521 Pause/resume stuck edge](https://github.com/TencentCloud/CubeSandbox/issues/1521) | Closed 2026-09-02. | Older deployments could leave leases stuck after failed lifecycle operations. | `/readyz`, status refresh, last-error state and GC visibility remain in place. | | [#1565 Concurrent scheduling overcommit](https://github.com/TencentCloud/CubeSandbox/issues/1565) | Closed 2026-09-17 after design discussion; the linked consolidation PR [#1595](https://github.com/TencentCloud/CubeSandbox/pull/1595) was closed without merge. | Closing the report is not evidence that concurrent admission or compounded overcommit was changed in a release. | Keep tenant lease/job quotas and node-capacity alerts; load-test the exact scheduler configuration before raising concurrency. | | [#1893 READY template may not restore](https://github.com/TencentCloud/CubeSandbox/issues/1893) | Open; reported on v0.5.1. A maintainer [said on 2026-10-08](https://github.com/TencentCloud/CubeSandbox/issues/1893#issuecomment-6051145081) that a wrong-guest-memory snapshot bug was fixed in v0.7.2 and asked the reporter to retry. | A template can report `READY` while repeated sandbox creates fail during guest-time reset; whether the reported case is fixed still needs a reporter or independent retest. | After each template rebuild, test several create-and-command cycles before moving its alias; record failures by template ID. The sr1 run was blocked by a separate MySQL/heartbeat failure before any restore. | | [#1890 Failed-create shim leak](https://github.com/TencentCloud/CubeSandbox/issues/1890) | Open; reported on v0.5.1, updated 2026-10-08. | Repeated failed creates reportedly left orphan shims and exhausted node memory. | Monitor failed creates, shim count and memory together during canary; do not infer a v0.7.2 fix from its release notes. | | [#1847 Failed resume leaves rootfs CoW object](https://github.com/TencentCloud/CubeSandbox/issues/1847) | Open; reported on v0.7.0, updated 2026-10-08. | A mid-resume failure can reportedly make a paused sandbox impossible to resume. | Include forced-failure resume and recovery in storage-specific acceptance before relying on long-lived paused sandboxes. | | [#1896 Rotated shim logs absent from API](https://github.com/TencentCloud/CubeSandbox/issues/1896) | Open; filed 2026-10-08 for a newer one-click logrotate change. | Sandbox log queries may cover only the current log segment after rotation. | Continue treating Adapter-owned durable job output and audit records as the retention source; verify API log history if enabling the new rotation policy. | ## Integration guide follow-up / 集成指南跟进 [#244](https://github.com/TencentCloud/CubeSandbox/issues/244) remains open. The Hermes Agent and DeepSeek Harness guide slots were claimed on 2026-08-29, and bilingual guide drafts have been prepared against the upstream template. They currently cite the v0.2.0 evidence. The v0.4.0 release adds refreshed real CubeSandbox acceptance evidence for trusted tasks across OpenClaw, DSH, Codex, and Hermes; update the upstream drafts to cite it before opening an external pull request. No external pull request is created from this working tree automatically. The v0.5.0 follow-up additionally validates Claude Code through both the trusted-task MCP flow and the Codex-equivalent acquire/exec/status/release flow. Its light native-client captures can be cited from the main README. ## Feature decisions / 功能决策 - Persistent Volume and Checkpoint are separate profile capabilities; neither silently implies the other. - Redis durability protects Adapter ownership data, not CubeSandbox's internal scheduler state. - Kubernetes HA improves the Adapter and control-plane operations, but cannot hide compute-node pause/resume, snapshot or overcommit defects. - `CUBE_API_KEY`, TLS/mTLS or OIDC, strict egress, and centralized audit should be treated as a deployment set rather than independent checkboxes. Re-check the open issues before raising tenant quotas substantially. Before enabling `allow_checkpoint_with_mounts`, run restore, rollback, clone, read-only mount and cleanup tests on the exact storage driver and node topology. Before a production upgrade, also follow the upstream [Kubernetes upgrade guide](https://github.com/TencentCloud/CubeSandbox/blob/v0.7.2/docs/guide/kubernetes/upgrade.md) and the [v0.7.2 release notes](https://github.com/TencentCloud/CubeSandbox/releases/tag/v0.7.2).