Added
- Per-widget AI mode (GBCA). Each website widget now has an AI agent setting in Settings → Integrations → Channels → Website widget: Off (operators handle every message), Draft (the GBCA agent drafts a reply for operator review on every visitor message), or Auto-send (the agent replies automatically until a human operator takes over). This is the tawk.to replacement path: unlike tawk.to's built-in RAG bot, the widget pipes conversations into your own connected agent with the same draft/approval machinery used for WhatsApp.
- Live inbox delivery for widget messages. Visitor messages now fire the workspace realtime broadcast, so operators watching the inbox see website chats appear instantly instead of on the next refresh.
- Visitor-facing author labels. Replies in the widget show who answered — the operator's author label or the AI agent's name — instead of a generic "Team".
- Unread badge + smarter polling. The launcher bubble shows an unread count when replies arrive while the chat is closed; the widget polls faster while open, slower when closed, and pauses entirely in hidden tabs. Enter sends, Shift+Enter adds a newline.
Security & abuse hardening
- Widget-triggered AI runs enforce a minimum 5-second scheduling debounce regardless of workspace settings, on top of the existing per-workspace public API rate limit, so rapid-fire visitor messages coalesce into a single agent run instead of amplifying model spend.
- Auto-send authority is re-checked against the widget's own settings at send time (fail closed), and any human operator reply in the conversation stops autonomous sends.
- Website-widget conversations never enqueue WhatsApp/Meta dispatch anywhere in the pipeline.
Verification
pnpm --filter @repo/database test,pnpm --filter worker test,pnpm --filter api test, targetedapps/appvitest suites — all green.pnpm --filter app exec tsc --noEmit,pnpm --filter api exec tsc --noEmit,pnpm lint,pnpm audit:i18n— all clean.- Independent adversarial review of the enqueue/auto-send path (workspace scoping, fail-closed gates, legacy-job isolation, flood cost analysis) with findings addressed.