dsh-remote-sandbox

Production-grade remote execution world for DeepSeek Harness (dsh): move ctx.fs and ctx.subprocess into an E2B sandbox that survives crashes, stays alive under load, and keeps your workspace in sync — the pieces the official @deepseek-ai/dsh-e2b proof-of-concept explicitly leaves out.

Third-party plugin suite. Not affiliated with DeepSeek. MIT licensed.

Why

dsh's capability seams let one provider swap move the whole execution world to a remote sandbox — Bash, PTY, and LSP follow ctx.fs/ctx.subprocess with no forks. The official dsh-e2b package proves the seam works but is, by its own README, a POC: the sandbox is "deliberately ephemeral", dies after five minutes, keeps a process's complete output in host memory, and "neither uploads nor synchronizes the host workspace". This suite makes that execution world usable for real work.

	official dsh-e2b (POC)	dsh-remote-sandbox
Sandbox lifetime	fixed 5 min, then deleted	heartbeat keep-alive, lives as long as the task
Connection drop / pause	none — sandbox death is task death	transparent recovery: resume the paused sandbox (disk retained)
Hard crash	none	recreate + restore workspace from the last synced snapshot
Host workspace	not uploaded or synced (path-spelling mirror only)	tar IN at open, tar OUT on request/teardown, .gitignore-independent excludes
Process output memory	complete stream retained in host memory (POC admits it breaks the seam's bound)	bounded in-memory tail + optional host spill file — host memory is capped
Round trips per fs op	3–8 SDK commands per mutation (realpath, stage, rename, xattr…)	one — the sidecar does read-modify-publish atomically in the sandbox