dsh-remote-dev
用于 AI 远程开发的 DeepSeek Harness SSH 插件：在 Linux 和 Windows 服务器上运行命令，以及读取、写入或浏览文件。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add dsh-remote-dev@1.0.0Remote Dev 以 dsh-remote-dev 发布，当前版本为 1.0.0。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
点击下方 Badge 复制 Markdown，粘贴到 README 即可。
验证 package.json 声明的 GitHub 仓库，即可管理这个公开页面。认领后，Hub 会优先安排当前版本的安全扫描，并在通过后公开展示结果。
认领这个 Plugin →
AI-native SSH remote development for DeepSeek Harness.
Add a directory on a remote machine as a workspace and develop in it exactly as you would locally—without a remote agent.
English · 简体中文 · GitHub · Changelog
DeepSeek Harness Remote SSH settings UI
This is a DeepSeek Harness bundle. One command installs and registers it:
npx dsh-remote-dev@latest setup # --profile web by default
Then start the Harness (dsh --profile web), open Settings → Remote Connections, add a
password or private-key profile, select Test connection, save, and connect.
dsh plugin add forwards to pnpm, and pnpm 11 fails an install whose dependency tree contains any
package with a build script (ERR_PNPM_IGNORED_BUILDS) — after which dsh never registers the
bundle, so the plugin is downloaded but never loaded. The only dependency here, ssh2, ships two
OPTIONAL native build scripts (its own install, and node-gyp for cpu-features); ssh2 is a pure
JavaScript SSH client that falls back to Node's crypto, so both are denied — no compiler needed.
The installer records that decision in the profile, runs the install, and verifies the result:
# ~/.dsh/profiles/web/pnpm-workspace.yaml
allowBuilds:
ssh2: false
cpu-features: false
With those two lines in place, dsh plugin --profile web add dsh-remote-dev works directly.
Options: --profile <name>, --package <spec>, --dsh "<command>", --allow-native,
--dry-run, --help.
In the sidebar, choose Add workspace → Remote machines, pick a machine, browse to a directory, and confirm.
The directory becomes an ordinary workspace row (app [SSH: buildbox]). Every session you start
under it reads, writes, edits, globs, greps and runs commands inside that directory on that
machine — with the same tools, the same tool cards, and the same relative paths a local
workspace gives you. Nothing has to be installed on the remote host, and no session-by-session
setup is involved: opening an old remote session restores the same remote world.
A DeepSeek Harness workspace is a real host directory (the registry canonicalizes it through
fs.realpath and groups sessions by their header cwd), so a remote:// path can never be one.
The plugin therefore keeps an empty local anchor directory per remote root — the stable
identity the sidebar groups by — and puts the actual work on the remote machine by composing the
session from a generated agent preset whose fs and shell services are this plugin's SSH
implementations, published inside one isolate realm.
The preset is derived from your default preset rather than hand-listed, so a remote session keeps
every capability a local one has (persona, instructions, skills, todos, plan mode, compaction,
delegation). Rows that would reach the local machine — the host filesystem, the host shell, the
local pty backend — are disabled inside the realm, and {{cwd}} resolves to the remote directory.
Editing your default preset regenerates the remote ones on next use.
Removing a remote workspace removes the sidebar row and keeps the generated preset, so existing sessions of that workspace still open; Also delete the generated preset under Settings → Remote Connections is the full cleanup.
cmd.exe context to the model.| Tool | Purpose |
|---|---|
remote_status | List profiles, states, platforms, and recent errors. |
remote_connect | Connect with a saved profile or one-off credentials. |
remote_disconnect | Close a profile's live connection. |
remote_exec | Execute a command and return stdout, stderr, exit status, and platform. |
remote_read | Read a UTF-8 text file over SFTP. |
remote_write | Write a complete UTF-8 text file over SFTP. |
remote_list | List directory entries with type, size, and modification time. |
Try:
Connect to my staging profile, inspect
/srv/app, run the tests, and explain the failures before changing anything.
The bundle automatically contributes tool guidance to the DeepSeek Harness system prompt.
ctx.credentials in standard Harness compositions. Minimal compositions fall back to ~/.dsh/remote/profiles.json with 0600 permissions.remote_* toolsbash in a kept-alive terminal) are disabled — the remote shell runs one command per callknown_hosts integration yetThe GitHub repository contains the complete quick start, Windows guidance, security model, Docker test environment, architecture, publishing guide, and contribution instructions.
This is a community-maintained plugin and is not an official DeepSeek AI product.
MIT © 2026 dsh-remote contributors