fix(sse): close the event stream on pagehide and reconnect on bfcache restore - #434

Summary

Pages could hang on loading skeletons for 15s or more after a few full-page navigations over plain HTTP.

The app opens one EventSource to /api/sse/stream and never closed it when the page was hidden. Chrome puts the page into the back/forward cache on full-document navigations (address bar, bookmarks, back/forward, the /login redirects), and the cached page's stream stayed open. /api/sse/status reported 4-5 live clients for a single tab. Over HTTP/1.1 Chrome allows only 6 connections per host, so after about 5 navigations the next page load waited for a connection to free up. HTTPS behind an HTTP/2 proxy hides the problem; plain http://host:port access does not.

Fix

OpenCodeEventStream registers window listeners in the browser:

• pagehide closes the live connection through the existing disconnectConnectionOnly(). Subscribers and directory ref counts are kept, and directories are re-queued as pending.
• pageshow with persisted: true (restored from the back/forward cache) calls the existing reconnect(), which resubscribes the pending directories and notifies status like any other reconnect.

Type of Change

• Bug fix
• New feature
• Refactor
• Documentation

Checklist

• Code follows project style (no comments, named imports)
• TypeScript types are properly defined
• Tests added/updated (80% coverage target)
• pnpm lint:frontend passes locally
• Frontend typecheck passes locally

Tests

frontend/src/lib/opencode-event-stream/__tests__/openCodeEventStream.test.ts:

• pagehide closes the active connection and does not schedule a reconnect.
• pageshow with persisted: true opens a new connection carrying the subscribed directories.
• pageshow with persisted: false opens no extra connection.

Validation

• pnpm --filter frontend exec vitest run src/lib/opencode-event-stream src/hooks/useSSE.test.tsx: 38 passed
• pnpm --filter frontend exec tsc --noEmit -p . and pnpm lint:frontend: clean
• Manual check in Chrome against a production build, 15 full-page navigations:
  • Before: /api/sse/status climbed to 4-5 clients and navigations timed out after 15s.
  • After: every page entered the back/forward cache, the server held exactly 1 client throughout, and each load took about 85ms. Going back restored the cached page and reconnected its stream, still with 1 client.

Summary by CodeRabbit

• Bug Fixes
  • Event stream connections now close when a page is hidden, preventing reconnect attempts during the following 60 seconds.
  • When a page is restored from the browser’s back-forward cache, the event stream reconnects to its subscribed directories. Other page-show events do not trigger a reconnect.