dsh-remote-workspace

When AI agents inside DeepSeek Harness orchestrate code, a recurring bottleneck is the local boundary: agent tools can only reach the machine where DSH is deployed, while the real builds, GPU inference, and container clusters live on remote hosts. Hand-rolled SSH wrappers lack connection pooling and drop under latency; naive SCP uploads leave half-written files when the network hiccups; and a simple file overwrite silently discards concurrent changes made on the other side. dsh-remote-workspace folds all of this into a first-class subsystem: persistent SSH2 connection pooling with dual key and password authentication, heartbeat keep-alive, and latency probing; atomic SFTP writes with streaming reads; SHA-256 baseline 3-way conflict-aware synchronization supporting pull, push, dry-run, and forced override; and programmatically managed local port-forwarding tunnels.

For the agent, the plugin exposes four orthogonal tools: remote_exec, remote_fs, remote_sync, and remote_tunnel, so an LLM can issue instructions like running pytest on the remote host, pushing local changes to staging, or tunneling port 8080 to the inference service without relying on external shell scripts. The sync layer ships sensible defaults for excluding .git, node_modules, and .dsh, produces a detailed conflict report instead of silently overwriting, and supports a dry-run preview so developers or agents can inspect every add, modify, and delete before anything hits disk. Version 0.3.1 adds high-throughput tarball streaming, one-shot remote diagnostics for occupied ports, OOM-killer events, disk usage, and service crash logs, one-click import of hosts and ProxyJump rules from ssh config, and proactive disk and memory alerts pushed through the Cordis event bus.

If your workflow requires DSH agents to drive remote VMs, GPU instances, or Kubernetes pods, or if you want to promote SSH, SFTP, and tunneling from ad-hoc terminal commands to a programmable, UI-observable capability, this plugin is built for that role.