Client Context (User State)

While the server defines who the agent is, the client defines who the user is. This “User Context” is dynamic, changing as the user connects their wallet, buys NFTs, or switches modes.

The Injection Mechanism: ToolAwareTransport

The bridge between the React state and the AI system is the ToolAwareTransport. This custom transport layer wraps the standard Vercel AI SDK transport to intercept requests and inject metadata.

Located in dapp/app/lib/llm/client/ToolAwareTransport.ts, it performs two critical functions:

1. Metadata Injection: It reads the current state (Mode, Auth) and merges it into the request body.
2. Stream Teeing: It splits the incoming SSE stream to parse tool events for the UI (see Streaming Context).

// Simplified logic from ToolAwareTransport.ts
const toolAwareFetch = async (input, init) => {
  if (isPostToChatApi) {
    const metadata = getMetadata(); // Reads fresh state
    const body = JSON.parse(init.body);
    
    // Inject metadata
    init.body = JSON.stringify({
      ...body,
      metadata: { ...body.metadata, ...metadata }
    });
  }
  return fetch(input, init);
};

Gathering User State

The ChatBot component (dapp/components/chatBot/index.tsx) is responsible for aggregating the user’s “World State” into a cohesive object called nftContext.

Components of User Context

Visual Representation

The context is not just hidden metadata; it is visually represented to the user to reinforce trust.

• User Header: UserHeaderWithWagmi.tsx displays the user’s ENS avatar and name. If the LLM addresses the user as “rito.eth”, the user sees that name in their chat bubble, confirming the context is shared.
• WagmiBoundary: A safety wrapper that ensures if the wallet connection fails or flickers, the chat UI degrades gracefully to a “Guest” state rather than crashing.

Context Ingestion Sources

Mode Selection

Mode Selection

The Active Mode is a special piece of context that drives the entire experience.

1. Selection: The user selects a mode via ModeSelectModal.
2. Store Update: useChatModeStore updates the global activeMode.
3. Prompt Composition: The client immediately generates the mode-specific System Prompt and adds it to the chat history.
4. Transport Injection: The ToolAwareTransport injects the mode metadata.
5. Server Enforcement: The server reads the mode metadata to enforce the Tool Whitelist.

Related Docs