dsh-rw

Working with a remote host from your agent used to mean either syncing a mirror back and forth or stitching together one-off SSH calls for every command. dsh-rw takes a different route: pick an SSH host and a remote directory, and that directory becomes a native DeepSeek Harness workspace. The agent reads, edits, builds, and tests directly on the remote filesystem. No mirror, no sync loop, no drift — the remote is the single source of truth.

The best part is that there is nothing new to learn. With shim mode on by default, the agent's existing read, write, edit, glob, grep, and bash tools are intercepted on the tool pipeline and transparently translated to remote execution. You never have to teach it an rw_ prefix. Ask it to run the test suite or fix a bug and it behaves as if the project were sitting on a local disk. When you do want explicit control, the rw_ tool set is still available for file operations and workspace management.

Safety is non-negotiable here. Every file path is confined to the chosen workspace root — symlink escapes, absolute paths outside the root, and directory traversal are all rejected with structured error codes. SSH host keys are verified against known_hosts by default, and a changed key aborts the connection rather than being silently accepted. The connection pool runs keepalives and bounded channel-open timeouts, so a silently dead TCP connection triggers a transparent one-time redial instead of hanging the agent. Credentials live in a 0600 local file and never leak into tool output or API responses.

If your project lives on a remote dev box, a container cluster, or a server behind a jump box, and you want your DeepSeek Harness agent to read code, make edits, run builds, and execute tests in that environment without the back-and-forth of copying files locally, dsh-rw is built for exactly that. Trust the host, point at the directory, and let the agent do its work where the code actually runs.