Repository navigation
Replies: 2 comments 1 reply
|
On Q6 (graduation path): official, in-repo container execution providers would be a product-level capability, not only an out-of-repo plugin family. DSH 0.2.2-rc.1 already documents the intended architecture in Hermes Agent v0.23.1 treats this as a first-party terminal backend. A community Docker plugin is a reasonable Stage A, but kernel-level isolation, image pinning, and how Proposal
Published by Grok Bot |
Stage A complete: prototype + lifecycle plugin + policy ceiling (2026-11-12)Independent community work — not upstreamed, not affiliated with or endorsed by DeepSeek; both plugins are MIT. As promised in my reply upthread: the default-refuse TL;DR of the arc since the design post
Stage A prototype: green, with two sshd requirementsPrototype shape: stock
The BoringSSL/PSK finding (unchanged, now with a workaround recipe)The Desktop app and the Triple-reproduced: bash tool (stream path) fails with exactly this error while the fs tool (pure RPC) succeeds in the same session; the same command under a plain-OpenSSL node (24.21) completes the full spawn + TLS-PSK stream + wait cycle against the same container. Scope: every reproduction above is from one macOS arm64 machine against 0.2.2-rc.1; the pin is unconditional in the family code, so I expect it to generalize, but that is expectation, not a tested matrix. Update: 0.2.2-rc.2 shipped on 2026-11-11 — I unpacked its Workaround that the lifecycle plugin documents and that carried all subsequent verification: the Desktop install ships a plain-OpenSSL node ( Questions for upstream (repeated from my design post above):
Implementation notes useful to upstream (found while building the plugins)
Measured end-to-end (fresh profile, macOS arm64 + Docker Desktop)
NextNothing blocked on upstream except Desktop usability (the PSK finding). If upstream rules that ssh-exact sandbox semantics are correct for remote transports, the ceiling plugin's default flips or the plugin goes away — that deletion is one commit by design. Corridor status: 0.2.2-rc.2 shipped with no container execution world in its package graph; the pin bump + gate re-run is done same-day — both plugins now pin the rc.2 family ( (Poster's note: independent community plugins; no affiliation with or endorsement by DeepSeek. "DeepSeek Harness" and |
Uh oh!
There was an error while loading. Please reload this page.
你好——我计划为 dsh 生态贡献一套容器执行 provider(以独立 npm 包形式的仓库外插件)。动工之前,想先就六个设计点与上游对齐。共同背景:
packages/sandbox/sandbox/src/index.ts写明:容器、microVM 与远程执行应当替换周边的整套能力接口(capability seam),而不是注册为ctx.sandbox的后端。我的方案遵循这一点:由一个负责连接管理的服务,加上分别面向 shell/subprocess/fs 的 provider,整体以dsh-ssh家族为模板。dsh-ssh的 helper 架构(在远端用摘要校验的 Node helper,以私有 Context 组合本地 provider;控制走 exec 通道上的 JSON-RPC,字节流走带能力令牌认证的 socket 转发)是我的设计模板。在我的设计里,helper 运行在容器内部,因此每次调用的文件访问策略仍由容器内的 sandbox 执行——容器提供外层的内核/文件系统隔离,逐调用的SandboxExecutionPolicy语义(含approveEscalation审批升级流程)原样保留。分阶段计划:Stage A 复用原封不动的
dsh-ssh家族,连接一个由小型生命周期插件管理的、运行 sshd 的容器;Stage B 是原生 Docker 家族,以docker exec的 stdio 作为 RPC 传输。完整设计文档可按需提供。Q1 —— 整套能力接口被替换后,
danger-full-access的语义。 SSH helper 在远端拒绝danger-full-access(ssh/src/helper.ts)。对容器家族,这一档应该 (a) 在容器内同样拒绝,还是 (b) 明确回落到宿主本地 provider 执行(即"离开容器")?我倾向 (b)——它符合该模式"不受限"的本义,且审批升级流程仍可使用——但这是产品决策,希望由上游确定。Q2 ——
dsh-ssh连接临时容器(Stage A)。host配置是 OpenSSH 别名(认证与 known-hosts 全部来自~/.ssh/config),且连接硬编码StrictHostKeyChecking=yes/BatchMode=yes。因此把 SSH 家族指向运行 sshd 的容器,需要:由插件自动写入并维护一段~/.ssh/config配置块,端口固定写在 Config 里;容器的临时 host key 由创建它的本地生命周期插件取出、写入UserKnownHostsFile——信任基础是"本机插件对 Docker 的控制权",而不是网络层的首次信任(TOFU);helper 的摘要按镜像 manifest 固定。这算受支持的配置吗?还是上游更希望在连接配置里直接支持临时主机(例如增加port/user/knownHostsFile/trustedHostKey字段),从而完全不需要管理~/.ssh/config?Q3 —— 宿主与容器的混合组合。 工作区以 bind mount 挂进容器后,宿主与容器的路径一致,因此可以在同一个 profile 里让文件访问留在宿主(
fs-local)、命令执行进入容器——这个组合既能缩小 MVP 范围,也让 LSP 留在宿主侧。是否有使用方或不变式要求ctx.fs与ctx.shell必须在同一个执行环境里?这种混合组合在接口契约上是否成立?Q4 —— 网络可见性。 sandbox 的三档模式刻意不包含网络与进程可见性,而容器做网络隔离的成本很低(
--network none/ 自定义网络)。这个维度应该只作为容器 provider 的Config配置,还是将来有意做跨 provider 的统一策略?目前我倾向保持 Config-only。Q5 —— 第三方的版本兼容约定。 已发布的 seam 包在打包时锁定精确版本。对仓库外的 provider 家族,预期的约定是"跟进已发布的 upgrade guide,并在 README 声明测试过的 dsh 版本",还是等 API 稳定后会放宽 peer 版本范围?
Q6 —— 转正路径。 如果这套 provider 成熟,上游是否考虑把它收进仓库维护(如同 ssh 家族)?另外,Stage A——原封不动的
dsh-ssh加一个容器生命周期插件——无论是否收编,是否算值得写进文档的受支持配置?如果拆成独立 issue 更合适,我很乐意拆。谢谢!
Hi — I'm planning to contribute a container execution provider family to the dsh ecosystem as out-of-repo plugins. Before writing code, I'd like to align on six design points. Context for all of them:
packages/sandbox/sandbox/src/index.tsstates that containers, microVMs, and remote execution replace the surrounding capability seam rather than registering as actx.sandboxbackend. My proposal follows that: a connection-owner service plus per-seam providers for shell/subprocess/fs, modeled on thedsh-sshfamily.dsh-sshhelper architecture (digest-pinned Node helper composing the local providers in a private Context on the remote side, JSON-RPC over the exec channel, capability-token-authenticated socket forwards) is my template. In my design the helper runs inside the container, so per-call file-effect policy is still enforced by the in-container sandbox — the container adds the outer kernel/filesystem boundary, and per-callSandboxExecutionPolicysemantics (including theapproveEscalationladder) are preserved unchanged.My staged plan: Stage A reuses the stock
dsh-sshfamily against an sshd-in-container managed by a small lifecycle plugin; Stage B is a native Docker family usingdocker execstdio as the RPC transport. Full design doc available on request.Q1 —
danger-full-accessin a replaced-seam world. The SSH helper refusesdanger-full-accessremotely (ssh/src/helper.ts). For a container family, should that mode be (a) refused inside the container as well, or (b) routed back to host-local providers as an explicit "leaves the container" semantic? I lean (b) — it matches the mode's meaning of "not confined" and keeps the escalation ladder usable — but this is a product decision I'd like upstream to own.Q2 —
dsh-sshagainst ephemeral containers (Stage A). Thehostconfig is an OpenSSH alias (auth and known-hosts come from~/.ssh/config), and the connection hardcodesStrictHostKeyChecking=yes/BatchMode=yes. Reusing the SSH family against an sshd-in-container therefore requires a managed~/.ssh/configblock with a Config-pinned fixed port, plus the container's ephemeral host key written into aUserKnownHostsFileby the local lifecycle plugin that created the container (trust anchored in Docker API control rather than network TOFU), and the helper digest pinned from the image manifest. Is this a supported configuration? Or would upstream prefer the connection config to grow explicit support for ephemeral or trusted-at-first-use hosts (e.g.port/user/knownHostsFile/trustedHostKeyfields), which would remove the~/.ssh/configmanagement entirely?Q3 — mixed-world composition. With the workspace bind-mounted, host paths and container paths coincide, enabling
fs-local(host side) + container shell/subprocess in one profile — an attractive MVP surface that also keeps LSP on the host. Does any consumer or invariant assumectx.fsandctx.shellshare one world, or is mixed composition contractually sound?Q4 — network visibility axis. The sandbox mode vocabulary deliberately excludes network/process visibility, and containers make network policy cheap (
--network none/custom networks). Should that stay container-providerConfigonly, or is there appetite for a cross-provider policy extension later? I'm happy to keep it Config-only for now.Q5 — third-party version corridor. Published seam packages pin exact versions at pack time. For an out-of-repo provider family, is the intended contract "track released upgrade guides and declare tested dsh versions in my README", or is a looser peer range planned once APIs stabilize?
Q6 — graduation path. If the provider family matures, would upstream consider adopting it in-repo (as the ssh family is)? And does Stage A — stock
dsh-sshplus a container lifecycle plugin — count as a supported configuration worth documenting either way?Happy to split these into separate issues if that fits the tracker better. Thanks!
All reactions