dsh-task2docker-studio

English · 简体中文

Task-to-Docker studio plugin for DeepSeek Harness: offload compute from a weak or security-constrained host onto servers you already control. The host stays a controller — it registers a runner, sends a task, and follows the result.

Vendor-neutral. A runner is any machine you can reach over SSH: a cloud server, a GPU box, a cluster entry node. No specific provider, no API key to a third party.

Built for two scenarios:

• The host is weaker than the target. A model won't fit, a build is too slow, a batch job is too big — run it on a server that can handle it.
• Safe execution. Untrusted or experimental code runs in a throwaway Docker container on that server, isolated from the host you're typing on.

A task runs as a detached container when an image is set (recommended for isolation), or backgrounded on the host otherwise.

Also usable as a general Agent Skill

The skill and its tools ship in the open Agent Skills format under skills/dsh-task2docker-studio/ — a SKILL.md plus a stdlib-only scripts/t2d.mjs CLI. Install it into Claude Code, Cursor, Copilot and other Agent-Skills agents (20+) with no DeepSeek Harness: it is the same dispatch logic, driven from the shell. See the skill README section.

What it gives the agent

Tool	Purpose
t2d_register_runner	Add a runner: name, host, port, user, auth (key/sshpass), optional default image
t2d_list_runners	Show registered runners (credential pointers only, never the secret)
t2d_remove_runner	Forget a runner
t2d_copy_to	Push a local file/dir (code, data, a skill) to a runner over scp
t2d_copy_from	Pull artifacts (checkpoints, reports) from a runner back to the host
t2d_run	Dispatch a command to a runner → a tracked job (container or native); optionally bind host paths into the container via mounts
t2d_status	Poll a job: RUNNING / SUCCEEDED / FAILED / STOPPED + exit code
t2d_logs	Tail a job's output, or live-follow it for a bounded window to watch a running task
t2d_stop	Stop a job

Plus a bundled skill (task2docker-studio) that teaches the loop: list → register → copy_to → run (with mounts) → status → logs → copy_from → stop, with container-vs-native, file-transfer and secrecy rules.

Install (DeepSeek Harness)

dsh plugin --profile web add https://github.com/CarryChang/dsh-task2docker-studio

Configure

Register runners live (recommended), from chat:

> 我有一台服务器 203.0.113.10，root + 公钥免密，名字叫 gpu-box，帮我登记，之后训练都丢它上面跑

Or declare them in the profile's cordis.yml so they are ready on load:

- id: task2docker
  name: dsh-task2docker-studio
  config:
    stateDir: ''                 # default ~/.dsh-task2docker (runners.json + jobs.json)
    sshTimeoutSec: 60
    runners:
      - name: gpu-box
        host: 203.0.113.10
        user: root
        auth: key
        defaultImage: nvidia/cuda:12.4.1-base-ubuntu22.04
      - name: isolated
        host: 203.0.113.20
        auth: sshpass
        passwordEnv: ISOLATED_PW # or passwordFile: /run/secrets/isolated_pw

Credentials

• Key / agent auth stores nothing secret — the normal path.
• Password auth (sshpass) stores only a pointer: passwordEnv names an environment variable, passwordFile names a 0600 file. The password is read at call time, never written to the ledger, and never returned in tool output.

Try it

> 把 ./train.py 丢到 gpu-box 上用 pytorch:2.4 镜像跑起来，跑完告诉我状态
> 这段不受信任的脚本别在我本机跑，去 isolated 上开个容器执行，只看输出
> run npm ci && npm test on my build server and tail the log

Send local code to a runner and mount it into the container, then watch the task as it runs and pull the result back:

> 把本地 ./src 目录传到 gpu-box 的 /data/src，然后挂载进容器跑 python /data/src/train.py，
> 实时看 60 秒日志，跑完把 /data/out 拉回本地

Architecture

src/
  index.ts             plugin entry: Config + skill + 7 t2d_* tools (the only Cordis-facing file)
  skill.ts             embedded task-to-runner methodology
  backend/             DSH-free, offline-testable
    types.ts           Runner, SshTarget, JobRecord, ExecResult, SshClient, Task2DockerError
    ssh.ts             shell quoting, docker/native command composition, ssh/scp/sshpass argv, SSH client
    registry.ts        runner + job JSON store; password-pointer resolution
    engine.ts          Task2DockerEngine: register / list / remove / copyTo / copyFrom / run / status / logs / stop
scripts/
  smoke.ts             offline deterministic smoke (fake SSH client + temp store)
  build-portable.mjs   derive the Agent-Skill CLI runtime from compiled DSH-free modules

The engine talks to runners through an injected SshClient, so the whole dispatch flow is unit-tested offline with a fake. Any SSH-reachable machine is a runner; no vendor SDK anywhere in the backend.

Use it as a standalone Agent Skill

skills/dsh-task2docker-studio/ is a self-contained Agent Skill (open format) — install it into Claude Code / Cursor / Copilot by copying the folder or via your agent's skill installer, then drive it through the shell:

cd skills/dsh-task2docker-studio
node scripts/t2d.mjs register --name gpu-box --host 203.0.113.10 --user root --image pytorch:2.4
node scripts/t2d.mjs run --runner gpu-box --command "python train.py"        # detached container
node scripts/t2d.mjs status --job <id>
node scripts/t2d.mjs logs   --job <id> --tail 80
node scripts/t2d.mjs stop   --job <id>

The CLI's compiled runtime lives in skills/dsh-task2docker-studio/scripts/lib/ (generated from the DSH-free src/backend/ by scripts/build-portable.mjs), so a GitHub clone or a plain folder copy needs no build. src/*.ts stays the single source of truth.

Provenance & license

MIT licensed (see LICENSE). v0.1 targeted OpenBayes / HyperAI GraphQL; v0.2 removed that vendor coupling and generalizes to any SSH runner with optional Docker. Not affiliated with or endorsed by DeepSeek, Docker, or any cloud provider.

Development

pnpm install     # needs access to the @deepseek-ai registry packages
pnpm run build   # tsc -> lib/  + build:portable -> skills/.../scripts/lib
pnpm run test    # offline deterministic smoke (no network, no ssh)
pnpm run check   # build + smoke

Limitations (v0.2)

• Runners must be SSH-reachable from the host. Containerized runs need Docker on the runner; native runs do not.
• Native background runs track the PID and a log file but do not capture a container-style exit code.
• sshpass must be installed on the host for password auth; prefer key auth.
• No file transfer tool yet — copy artifacts with scp/rsync against the runner, or mount a path in the container.