DWJZ/dsh-restart

---

description: "dsh-restart: a Settings page that restarts the dsh Web host process with one click."

dsh-restart

English | 中文

A Restart page in the DeepSeek Harness (dsh) Web settings panel: one button replaces the running host process with a fresh one.

What it does

• Host half registers three routes on the Web server, all limited to direct same-origin loopback requests:
  • GET /dsh-restart/status — boot id, PID, start time, and serving port of the current process
  • POST /dsh-restart/restart — replaces the current process (the host exits after answering 202)
  • POST /dsh-restart/stop — ends the current process and starts nothing (same ordering)
• Browser half registers a settings.section page showing that identity, a restart button, and a stop button.
• Shortcut — Cmd+Shift+K (macOS) or Ctrl+Shift+K (Windows) restarts the host without opening settings. Rebind it on the settings Keyboard page like any other command. It fires from the page and from editable surfaces such as the composer, while terminal input keeps it out, and it appears with its label on the settings Keyboard page. Web on Linux is left unbound: the client's shortcut allowlist rejects every primary+alt chord there, so registering one would fail the plugin's activation. After a restart it polls /dsh-restart/status and reloads the page as soon as a different boot id answers; after a stop there is nothing to wait for, so the page reports success once the status route stops answering — and says plainly that starting again means running the launch command in a terminal.

The restart itself is done by a detached helper: the host exits after 400 ms, and the helper waits until the port stops accepting connections before spawning the replacement with exactly the same launch invocation (Node binary, execArgv, absolute entry path, remaining argv, working directory). That ordering is what keeps the replacement from dying of EADDRINUSE while the old socket is still held. Logs land in the system temp directory as dsh-restart-<timestamp>.out.log / .err.log.

A stop is the same 400 ms SIGTERM with no helper behind it, which is what lets the CLI flush sessions and dispose plugins before its own escalation timer forces the exit. Because a stopped host has no replacement, the process you restart from the page is also the one no longer attached to your terminal — after any restart, stopping it means this button, kill <pid>, or curl -X POST http://127.0.0.1:<port>/dsh-restart/stop.

Sessions live on disk and survive the restart, and plugins that are installed but not active yet load on the next boot.

Install

# from GitHub
dsh plugin --profile web add github:DWJZ/dsh-restart

# local development (the profile links the checkout, so edits apply directly)
dsh plugin --profile web add link:/path/to/dsh-restart

The first install needs one restart before the page exists — restart the host from a terminal, or use dshmarket's restart button once. After that, this page is the way to restart.

Endpoint contract

POST /dsh-restart/restart accepts only requests whose peer is loopback and whose Origin matches Host; any forwarding header (Forwarded, X-Forwarded-For, X-Real-IP) or non-loopback peer is refused with 403. It answers 202, then the host exits and the helper starts the replacement.

POST /dsh-restart/stop takes the same guard and answers 202 before exiting, so the browser learns the outcome from the response instead of from a dropped connection. It never sends SIGKILL: the graceful signal is what flushes sessions.

Test

node test/smoke.mjs

Covers: the detached helper really starts a replacement process, all four refusal paths of the same-origin guard, and the client bundle's registration and render. Set DSH_CHECKOUT=<dsh checkout> to add an SSR render assertion using that checkout's React; without it that one assertion is skipped.

License

MIT