React 中的 WebMCP：向浏览器 AI 智能体开放客户端状态与表单操作

"use client";

import React, { useActionState } from "react";

interface FilterState {
  minAmount: number;
  category: string;
  status: "idle" | "applied";
}

async function applyFilterAction(
  prevState: FilterState,
  formData: FormData
): Promise<FilterState> {
  const minAmount = Number(formData.get("minAmount") || 0);
  const category = String(formData.get("category") || "all");

  // Perform client-side filter computation or query
  return {
    minAmount,
    category,
    status: "applied",
  };
}

export function AgenticOrderFilter() {
  const [state, formAction, isPending] = useActionState(applyFilterAction, {
    minAmount: 0,
    category: "all",
    status: "idle",
  });

  return (
    <form
      action={formAction}
      // WebMCP Declarative Tool Annotations
      toolname="filter-orders"
      tooldescription="Filters the current order ledger by minimum dollar amount and product category."
      className="filter-form"
    >
      <label htmlFor="minAmount">Minimum Amount ($)</label>
      <input
        id="minAmount"
        name="minAmount"
        type="number"
        defaultValue={state.minAmount}
        required
      />

      <label htmlFor="category">Category</label>
      <select id="category" name="category" defaultValue={state.category}>
        <option value="all">All Categories</option>
        <option value="hardware">Hardware</option>
        <option value="software">Software</option>
      </select>

      <button type="submit" disabled={isPending}>
        {isPending ? "Filtering..." : "Apply Filter"}
      </button>

      {state.status === "applied" && (
        <p className="status-text">
          Showing orders &gt; ${state.minAmount} in category &quot;{state.category}&quot;
        </p>
      )}
    </form>
  );
}

"use client";

import { useEffect, useRef } from "react";

export interface ToolDefinition<TInput = Record<string, unknown>, TOutput = Record<string, unknown>> {
  name: string;
  description: string;
  inputSchema: Record<string, unknown>;
  execute: (input: TInput) => Promise<TOutput>;
  readOnlyHint?: boolean;
}

/**
 * Registers an imperative tool with document.modelContext, ensuring
 * safe teardown with AbortSignal and fresh closure references.
 */
export function useWebMCPTool<TInput = Record<string, unknown>, TOutput = Record<string, unknown>>(
  tool: ToolDefinition<TInput, TOutput>,
  enabled: boolean = true
) {
  // Store the latest executor in a ref to avoid re-registering on every state change
  const executeRef = useRef(tool.execute);
  useEffect(() => {
    executeRef.current = tool.execute;
  });

  useEffect(() => {
    // Feature detection: Check for browser WebMCP support
    if (typeof document === "undefined" || !document.modelContext || !enabled) {
      return;
    }

    const abortController = new AbortController();

    try {
      document.modelContext.registerTool(
        {
          name: tool.name,
          description: tool.description,
          inputSchema: tool.inputSchema,
          execute: async (args: Record<string, unknown>) => {
            return await executeRef.current(args as TInput);
          },
          annotations: tool.readOnlyHint ? { readOnlyHint: true } : undefined,
        },
        { signal: abortController.signal }
      );
    } catch (err) {
      console.warn(`[WebMCP] Failed to register tool "${tool.name}":`, err);
    }

    // Teardown: AbortSignal unregisters the tool automatically
    return () => {
      abortController.abort();
    };
  }, [tool.name, tool.description, tool.readOnlyHint, enabled]);
}

"use client";

import React, { useState, useCallback } from "react";
import { useWebMCPTool } from "./useWebMCPTool";

interface DocumentViewerProps {
  totalPages: number;
  documentTitle: string;
}

export function DocumentViewer({ totalPages, documentTitle }: DocumentViewerProps) {
  const [currentPage, setCurrentPage] = useState<number>(1);
  const [selection, setSelection] = useState<string>("");

  // Tool 1: Jump to a specific page (Mutating action)
  useWebMCPTool({
    name: "navigate-document-page",
    description: "Navigates the interactive PDF viewer to a specific page number.",
    inputSchema: {
      type: "object",
      properties: {
        pageNumber: {
          type: "integer",
          minimum: 1,
          maximum: totalPages,
          description: "Target page index to display",
        },
      },
      required: ["pageNumber"],
    },
    execute: async (args: { pageNumber: number }) => {
      if (args.pageNumber < 1 || args.pageNumber > totalPages) {
        return {
          success: false,
          error: `Page ${args.pageNumber} out of bounds (1-${totalPages}).`,
        };
      }
      setCurrentPage(args.pageNumber);
      return {
        success: true,
        activePage: args.pageNumber,
        documentTitle,
      };
    },
  });

  // Tool 2: Read current selection (Read-only query)
  useWebMCPTool({
    name: "get-active-selection",
    description: "Retrieves the currently highlighted text snippet in the viewer.",
    inputSchema: {
      type: "object",
      properties: {},
    },
    readOnlyHint: true,
    execute: async () => {
      return {
        hasSelection: selection.length > 0,
        text: selection,
        pageNumber: currentPage,
      };
    },
  });

  return (
    <div className="viewer-container">
      <header>
        <h3>{documentTitle}</h3>
        <span>Page {currentPage} of {totalPages}</span>
      </header>
      <main
        onMouseUp={() => {
          const selectedText = window.getSelection()?.toString() || "";
          setSelection(selectedText);
        }}
        className="document-canvas"
      >
        <p>Displaying page content for page {currentPage}...</p>
      </main>
    </div>
  );
}

"use client";

import React, { useState, useRef } from "react";
import { useWebMCPTool } from "./useWebMCPTool";

interface PendingAction {
  id: string;
  description: string;
  resolve: (value: { approved: boolean }) => void;
}

export function DestructiveActionShield() {
  const [pendingAction, setPendingAction] = useState<PendingAction | null>(null);

  useWebMCPTool({
    name: "delete-active-workspace",
    description: "Permanently deletes the current active workspace. Requires user confirmation.",
    inputSchema: {
      type: "object",
      properties: {
        reason: { type: "string", description: "Reason for deletion" },
      },
      required: ["reason"],
    },
    execute: async (input: { reason: string }) => {
      // Pause tool execution and wait for manual user approval
      return new Promise((resolve) => {
        setPendingAction({
          id: crypto.randomUUID(),
          description: `Delete workspace: "${input.reason}"`,
          resolve,
        });
      });
    },
  });

  return (
    <>
      {pendingAction && (
        <aside className="confirmation-modal" role="alertdialog">
          <h4>Agent Action Confirmation</h4>
          <p>An AI assistant requested permission to:</p>
          <blockquote>{pendingAction.description}</blockquote>
          <div className="button-row">
            <button
              type="button"
              onClick={() => {
                pendingAction.resolve({ approved: false });
                setPendingAction(null);
              }}
            >
              Reject
            </button>
            <button
              type="button"
              className="danger-btn"
              onClick={() => {
                pendingAction.resolve({ approved: true });
                setPendingAction(null);
              }}
            >
              Confirm Deletion
            </button>
          </div>
        </aside>
      )}
    </>
  );
}

// Client-side agent loop executing in the browser tab
export async function runLocalTabAgent(userPrompt: string): Promise<string> {
  // 1. Verify availability of on-device LLM and WebMCP
  if (!("ai" in window) || !document.modelContext) {
    throw new Error("On-device AI or WebMCP is not supported in this browser.");
  }

  // 2. Discover available tools exposed by mounted React components
  const availableTools = await document.modelContext.listTools();
  
  // Format tool descriptions for system prompt
  const toolDeclarations = availableTools.map((t) => ({
    name: t.name,
    description: t.description,
    parameters: t.inputSchema,
  }));

  // 3. Instantiate local session with Gemini Nano
  const session = await (window as unknown as {
    ai: {
      languageModel: {
        create: (options: { systemPrompt: string }) => Promise<{
          prompt: (msg: string) => Promise<string>;
          destroy: () => void;
        }>;
      };
    };
  }).ai.languageModel.create({
    systemPrompt: `You are an in-tab browser assistant. You can control the active page using tools: ${JSON.stringify(
      toolDeclarations
    )}. Output JSON tool invocations as {"tool": string, "args": object}.`,
  });

  try {
    const response = await session.prompt(userPrompt);
    
    // Parse structured tool call
    const action = JSON.parse(response);
    const targetTool = availableTools.find((t) => t.name === action.tool);

    if (targetTool) {
      const result = await targetTool.execute(action.args);
      return `Tool executed successfully: ${JSON.stringify(result)}`;
    }

    return response;
  } finally {
    session.destroy();
  }
}