wehub-resource-sync
e30e75b5d4
2.3 KiB
@assistant-ui/cloud-ai-sdk
Part of assistant-ui monorepo. See root AGENTS.md for build commands.
React hooks bridging Vercel AI SDK (useChat) with assistant-cloud persistence. Works without the full @assistant-ui/react runtime.
Structure
src/
├── chat/
│ ├── useCloudChat.ts # Main hook — orchestrates cloud + AI SDK
│ ├── useCloudChatCore.ts # React lifecycle wrapper for CloudChatCore
│ ├── useChatRegistry.ts # Manages Chat instances by thread/session key
│ ├── ChatRegistry.ts # Chat instance store (non-React)
│ └── MessagePersistence.ts # Encodes/decodes UIMessage for cloud storage
├── core/
│ ├── CloudChatCore.ts # Orchestrator — persistence, thread creation, title gen
│ ├── ThreadSessionManager.ts # Deduplicates concurrent thread creation
│ └── TitlePolicy.ts # One-time auto-title for new threads
├── threads/
│ ├── useThreads.ts # Thread list CRUD hook
│ └── generateThreadTitle.ts # Cloud API call for title generation
├── types.ts # Shared types (UseCloudChatOptions, ChatMeta, etc.)
└── index.ts # Public exports
Key Concepts
Auto-cloud singleton — Created from NEXT_PUBLIC_ASSISTANT_BASE_URL env var at module load. Enables zero-config anonymous auth.
Two-layer pattern — useCloudChatCore (React lifecycle) wraps CloudChatCore (pure logic). The hook manages creation, syncing, and cleanup of the core instance across renders.
Chat registry — ChatRegistry + useChatRegistry manage multiple chat instances keyed by thread or session ID. Switching threads reuses existing instances.
Thread auto-creation — First message triggers thread creation in the transport layer. ThreadSessionManager deduplicates concurrent creation promises (no duplicates).
External vs internal threads — useCloudChat() creates its own useThreads internally. Pass { threads } from an external useThreads() to share thread state across components (e.g., sidebar + chat).
Title generation — Happens automatically as a side effect of message persistence (onFinish), not an explicit API call. TitlePolicy ensures it fires only once per thread.