dsh-restart-button

English | 中文

A restart row in Settings → General for the DeepSeek Harness Web GUI: one click restarts the host process that serves the page, then the tab reloads itself once the replacement host answers. No terminal, no systemctl, no hunting for the window that started dsh web.

Sessions are persisted per event under ~/.dsh/sessions, so a restart does not lose them; a running turn is interrupted, and in-flight approval questions do not survive (the row's hint says so).

Features

• One row in Settings → General, next to the shipped rows: title, hint, and a 36px pill button (Restart now / Restarting…).
• Restart semantics that fit how the host is actually supervised:
  • Under systemd (the unit sets INVOCATION_ID/JOURNAL_STREAM): exit cleanly after the response has flushed, and let Restart=always bring the host back. The plugin never spawns a replacement here — a detached child would race the supervisor for the port and escape its cgroup.
  • Without a supervisor (a bare terminal, pnpm dev): a detached node helper waits for the port to free, relaunches the exact same command line, and logs to /tmp/dsh-restart-button-*.{out,err}.log.
• The row polls the boot id after the restart and reloads the page when the replacement host is up (60s budget).
• Degrades visibly: on a non-loopback client the row renders read-only instead of a button that cannot work.

Compatibility

DSH	Status
>=0.1.5-rc.1	✅ supported (declared in engines.dsh)
<=0.1.4	❌ not supported — the channel mount below relies on the 0.1.5 Web-route/requestRejection shape

The row talks to the host over one fenced HTTP channel, /_dsh-restart-button, carrying the Connection RPC envelope ({type:'client-request',rpcId,method,payload} → {type:'server-response',rpcId,result}) that the browser half already speaks. Endpoints: status (boot id) and restart.

Why the route is mounted by the plugin instead of ctx.connection.rpc.handle(): on 0.1.5 that registry mounts through owner.webServer.register(...), where owner is the Connection service's own context (inject: credentials only). Service access is inject-gated by cordis 4, so the call throws cannot get property "webServer" without inject; raised inside the ctx.inject() child fiber this plugin needs, that error only fails the child fiber — the plugin entry still reports active and the channel is silently missing. Passing webServer in the caller's inject list does not help, because the gated context is the service's, not the caller's.

So the plugin uses the two public pieces that exist for exactly this:

• webServer.register({ kind: 'prefix', path: '/_dsh-restart-button', handler }) — a feature-owned Web route;
• connection.requestRejection(req) — "apply Connection's Host/Origin checks and browser authentication to another Web route": the same fence /api uses (trusted authority plus the browser session cookie), applied before a single request byte is decoded (401 unauthenticated, 403 cross-site).

Both are consumed structurally (no SDK import at runtime), which is also why the host half keeps working across DSH minor versions.

Install

# into the web profile, as a local checkout:
#   "dsh-restart-button": "file:/path/to/dsh-restart-button"
cd ~/.dsh/profiles/web && pnpm install
# then restart dsh once so the plugin tree picks it up

The plugin registers the channel itself; it needs the connection and webServer services, i.e. the Web profile.

Development

pnpm install
pnpm typecheck   # host + client
pnpm build       # lib/index.js, lib/client.js, lib/types/**

• Host half: src/index.ts (TypeScript → lib/index.js, declarations to lib/types/).
• Browser half: src/client/* (tsdown → lib/client.js, one CJS bundle in the client module table; only react/react-dom stay external).
• cordis.patch.yml is the loader row (dsh.bundle.patch) that mounts the host half.

License

MIT © 2026 allenlogo