miisaka19800/dsh-restart-fab

A floating one-click restart button for the DeepSeek Harness web GUI.

Pinned to the bottom-right corner of the page: one click arms a confirmation, a second click restarts DeepSeek Harness in place — the page shows a full-screen "restarting" overlay and automatically reloads onto the new host once it is back up. No manual refresh, no second tab, no launcher script.

Features

• Fully self-contained — the restarted host is relaunched with the exact same command line that started the current one (node …/bin.js web --port N), preserving workspace, port, profile and flags. No dependency on any launcher script (DSH启动器.exe, shell wrappers, etc.).
• Cross-platform — Windows (force kill via taskkill), macOS/Linux (graceful SIGTERM then SIGKILL).
• Fast — the old process is killed ~1.6 s after you confirm; the new host starts immediately.
• Clean UX — full-screen overlay with a spinner and elapsed timer while restarting; the page polls the host every second and reloads itself the moment the new host responds. A "refresh now" fallback appears after 60 s.
• Sessions survive — DeepSeek Harness persists sessions, so your conversations resume automatically after the restart.
• Same-origin only — the restart route rejects requests from other origins.

Install

Once published, install into the web profile with one command:

# from npm (once the package is published)
dsh plugin --profile web add dsh-restart-fab

# or straight from GitHub (works even before the npm release)
dsh plugin --profile web add github:miisaka19800/dsh-restart-fab

Then restart DeepSeek Harness once (close the console window / stop the process and start it again) — from then on the button is always there.

Install from a local checkout (development)

dsh plugin --profile web add /path/to/dsh-restart-fab

The profile keeps a link: dependency, so edits to this folder are live after the next restart.

Uninstall

dsh plugin --profile web remove dsh-restart-fab

How it works

button (2nd click) ──POST /dsh-restart-fab/restart──▶ server route
                                                        │ spawns detached helper
                                                        │   node restart-helper.mjs <pid> <relaunch-spec-json>
helper:  sleep 1.6s ──▶ kill server process ──▶ sleep 400ms ──▶ relaunch
                                                              (same argv, same cwd)
                                                              Windows: new visible console
                                                              (close it to stop; child processes
                                                               attach to it — no stray popups)
         ──▶ poll new host ──▶ exit
page:    overlay + poll /dsh-restart-fab/ping every 1s
         ── boot id changed? ──▶ location.reload() on the new host

• GET /dsh-restart-fab/ping — health check; reports the host boot id (<pid>-<timestamp>). The page reloads when it changes.
• POST /dsh-restart-fab/restart — same-origin only; schedules the restart through lib/restart-helper.mjs, spawned as a detached, hidden child so it survives however the server dies. It kills the server process, relaunches the same command line, and exits once the new host answers.

Project layout

path	purpose
client/client.js	hand-authored CJS bundle, injected into the shell.overlay slot (bottom-right button + restart overlay), zh/en locale
lib/index.js	server-side cordis plugin: ping + restart routes on the webServer service
lib/restart-helper.mjs	detached helper: kill + relaunch (cross-platform)
cordis.patch.yml	profile bundle patch that inserts the plugin entry

No build step — plain ESM server code and a hand-authored client bundle.

Security

The restart route only accepts same-origin POSTs (the page served by the host itself). Restarting stops the current dsh web process and starts a new one — anything running inside it (including model sessions) is interrupted, then resumed from persistence.

License

MIT