[sanchitrk]

Summary

• One status, kept by the app server. A new Health service checks the servers:

  • at launch, then on an interval (300s by default; 0 turns it off);
  • after any change made in the app;
  • when a call to the default server fails.

  It also checks the default server's process every 5s with no RPC and no shell, so a server started or stopped outside the app shows within seconds. The status is pushed over a new servers.watch stream, which replaces servers.list.

• Simpler downtime UI, like t3code.

  • The sidebar shows a muted "Open Provider Settings" button.
  • The composer replaces the model and thinking pickers with the same button.
  • The server-down notices and the "Couldn't Load Models" / catalog notices are gone. Only the settings-file error remains.
  • The session list and models load again by themselves once the server is back.

• Settings › Providers › Advanced gets a "Health check interval" row: reset to default, −/+ in 30s steps, and 0 to disable.

• Splash: it stays over the mounted app until the first status arrives (2s at most), then fades.

Testing

• vp check and vp run typecheck pass.
• New E2E test apps/server/test/health.test.ts. It starts the real app server, and its default server goes Stopped → Unreachable → Stopped as a registration appears and goes. The interval setting saves. The result is written to __artifacts__/health.json.
• Driven by hand and with agent-browser in vp run dev:desktop:
  • down state at launch;
  • Start from settings;
  • codework service stop / start from a terminal (shown in 2s and 4s);
  • the session composer;
  • the interval stepper and reset;
  • the splash fade.