tangwenhao616-netizen / dsh-anywhere

dsh-anywhere

Put your DeepSeek Harness workspace on any machine. Bring an off-site box online — even behind NAT — then let the agent read / write / edit / bash / grep natively on that machine, while the model, memory, and agent loop stay on your hub.

English · 简体中文

flowchart LR
    subgraph HUB["🧠 dsh hub — your laptop / server"]
      A["model · memory · agent loop"]
    end
    subgraph RELAY["☁️ your relay VPS"]
      R["loopback rendezvous"]
    end
    subgraph REMOTE["💻 any machine — even behind NAT"]
      F["files + processes"]
    end
    A -- "① fleet: one-time token + reverse tunnel" --> R
    R -. "ProxyJump" .-> F
    A == "② world: ctx.fs + ctx.subprocess<br/>read · write · edit · bash · grep land here" ==> F

Two composable modes in one plugin:

• ① fleet — reach it. Enroll a machine behind NAT into your dph with a one-time token and a reverse tunnel to your own relay VPS. It shows up as a fleet-* dsh-ssh host, driven by the existing ssh_exec / ssh_upload tools.
• ② world — make it your workspace. Turn any reachable machine into the session's ctx.fs + ctx.subprocess, so the agent's native file & shell tools operate on it. Same shape as the E2B integration — but the remote is any SSH-reachable host.

Flagship combo (this is the "remote computer as workspace" story): use fleet to pull in the NAT'd box, then world (-o ProxyJump=<relay>) to promote it to the whole session's workspace. If the target is already reachable (a cloud VM, a LAN box), just use world and skip fleet.

Demo

A real agent session, driven through the published package, operating on a remote machine — it runs a shell command there and writes a file there:

$ dsh --profile headless --patch ./workspace-on-machine.patch.yml \
    "run 'uname -n' and write MERGED-WORLD-OK to /tmp/selftest.txt"

▸ bash        uname -n
  VM-0-16-ubuntu                 ← the REMOTE host, not your laptop
▸ write       /tmp/selftest.txt  ← lands on the REMOTE machine
  MERGED-WORLD-OK

$ ssh you@remote 'cat /tmp/selftest.txt'   # verified independently
MERGED-WORLD-OK

📹 A screen recording of the /fleet enroll-and-approve flow is welcome — see CONTRIBUTING.

Quick Start (~5 min)

Prerequisites: a working dsh install, and system ssh on the hub. For world mode you also need a machine you can already ssh into (LAN, a cloud VM, or a fleet-enrolled box).

1. Install (peers @deepseek-ai/dsh-fs / dsh-subprocess are provided by dsh):

dsh plugin --profile <your-profile> add dsh-anywhere

2a. World mode — put your workspace on a reachable machine. Copy examples/workspace-on-machine.patch.yml, fill in the target's login / sshArgs / cwd, then:

DSH_PERMISSION_MODE=danger-full-access \
  dsh --profile <your-profile> --patch ./workspace-on-machine.patch.yml "work on this machine"

Now read / write / edit / bash / grep all land on that machine.

2b. Fleet mode — bring a NAT'd machine online. Installing already mounts the /fleet panel. Point it at your relay (run scripts/relay-init.sh on your VPS once, then register it), click Add machine, run the one-liner on the target, and approve it in the panel. See fleet mode below.

⚠️ Install properly — do not link:. At boot, a --patch name: resolves relative to the profile directory. A link: dev-install makes the world half fail with ERR_MODULE_NOT_FOUND (its @deepseek-ai/dsh-fs peer can't be resolved from the repo realpath). Use a real install (dsh plugin add / npm i, peers provided by the runtime), or drop a real dsh-anywhere directory into <profile>/node_modules/ and symlink @deepseek-ai/dsh-fs + dsh-subprocess into <profile>/node_modules/@deepseek-ai/ (same realpath so extends FileSystem uses the same class). The fleet half is zero-dependency and link: works.

The two modes

	fleet (reach layer)	world (execution-world layer)
Job	make a NAT'd machine reachable	make a reachable machine the whole workspace
Machine appears as	a dsh-ssh host (fleet-*)	the session's ctx.fs + ctx.subprocess
You operate it with	ssh_exec / ssh_upload / ssh_download / ssh_tunnel	native read / write / edit / bash / grep, on the remote
Mounting	always-on once installed (dsh.bundle → cordis.patch.yml)	per session, --patch workspace-on-machine.patch.yml
Entry point	dsh-anywhere	dsh-anywhere/world

• Zero external npm deps — files / processes / tunnels all go through the system ssh CLI (paths and contents are base64'd into the remote script, so no quoting or injection). @deepseek-ai/dsh-fs and @deepseek-ai/dsh-subprocess are peerDependencies, provided by the dsh runtime.
• Fast — world reuses one OpenSSH ControlMaster connection, so fs micro-ops are ~40 ms (no re-handshake per op). This is what makes working over a tunnel actually pleasant.
• Cancellable — long commands use a dedicated connection; terminate() drops it and sshd SIGHUPs the remote command.

Mode: fleet — bring a machine online

1. Configure a relay (the plugin never hard-codes your VPS): run scripts/relay-init.sh on your relay VPS, then register it via POST /api/fleet/relay (or in the /fleet panel).
2. Add a machine: the panel's Add machine gives a one-liner — run curl <base>/join | bash on the target (Windows: irm '<base>/join?os=win' | iex). The machine prints a pairing code and waits.
3. Approve: verify the pairing code and click Approve in /fleet (approval is accepted only from the local host). The machine's sshd is mapped to the relay's loopback via a reverse tunnel, and the hub reaches it over ProxyJump as a fleet-* dsh-ssh host.
4. List / revoke: GET /api/fleet/list, POST /api/fleet/approve|reject|revoke.

Security: one-time, expirable, per-machine-revocable tokens; a separate tunnel key per machine; the machine's sshd binds only the relay loopback and is never exposed publicly; approval is local-only.

Mode: world — make it your workspace

See Quick Start 2a. What world provides:

Capability	Implementation
ctx.fs (FileSystem)	fs-ssh: resolve / stat / lstat / readText / streamText / readBytes / listDir / writeText / editText — with version guards, atomic writes, binary/UTF-8 handling
ctx.subprocess (SubprocessRuntime)	subprocess-ssh: resolveExecutable / spawn (stdin · collect · exit code · cwd · env · terminate) / spawnTerminal (basic PTY via ssh -tt)
bash / grep / terminal / LSP	no changes needed — they are provider-neutral consumers of ctx.fs + ctx.subprocess, so they land on the remote automatically

One-world invariant: fs.cwd == subprocess.cwd (the patch's config.cwd) == sandbox-policy.workspaceRoot == the session workspace — they must all point at the same remote directory, or relative paths and command cwds land on a path that doesn't exist remotely and spawns fail.

Windows remotes: if the remote is Windows (OpenSSH + PowerShell), add platform: windows to the ssh-world config — fs/subprocess then run on a PowerShell backend (base64-framed, binary-safe) — and use Windows paths for cwd / workspaceRoot (e.g. C:\Users\you\work). Default is platform: posix (Linux/macOS). Validated end-to-end against a real Windows host. Note: Windows OpenSSH collapses non-zero exit codes, so the exact code is recovered out-of-band; a remote rg still helps grep speed.

Per-instance, not per-session: a dsh execution world is per-instance (ctx.fs/ctx.subprocess are not scope-aware). To use a local workspace and a remote one at the same time, run two dsh instances (one local profile, one world profile). fleet mode has no such limit — an enrolled machine is a dsh-ssh host that coexists with your local workspace.

Known limits (POC stage)

• spawnTerminal is a basic PTY; precise remote foreground-process-group query/signalling and full TERM→KILL quiescing aren't implemented yet.
• Each fs op is one remote round-trip (ControlMaster removes the handshake; huge bursts still pay per-op ssh process cost).
• The remote needs rg (ripgrep) for the agent's grep/glob to run at full speed; otherwise it falls back to grep/find.
• world's ctx.fs is not wired to dsh's sandbox fence (use danger-full-access; the fence is delegated to the remote account itself).

Contributing

Issues and PRs welcome — see CONTRIBUTING.md. Tests: node --test (integration tests auto-skip unless you set DSH_ANYWHERE_RELAY=user@host).

License

MIT