fix(aiCore): update tool call status and enhance execution flow

- Changed tool call status from 'invoking' to 'pending' for better clarity in execution state.
- Updated the tool execution logic to include user confirmation for non-auto-approved tools, improving user interaction.
- Refactored the handling of experimental context in the tool execution parameters to support chunk streaming.
- Commented out unused tool input event cases in AiSdkToChunkAdapter for cleaner code.

src/renderer/src/aiCore/chunk/AiSdkToChunkAdapter.ts

@@ -135,11 +135,11 @@ export class AiSdkToChunkAdapter {
 
       // === 工具调用相关事件（原始 AI SDK 事件，如果没有被中间件处理） ===
 
-      case 'tool-input-start':
-      case 'tool-input-delta':
-      case 'tool-input-end':
-        this.toolCallHandler.handleToolCallCreated(chunk)
-        break
+      // case 'tool-input-start':
+      // case 'tool-input-delta':
+      // case 'tool-input-end':
+      //   this.toolCallHandler.handleToolCallCreated(chunk)
+      //   break
 
       // case 'tool-input-delta':
       //   this.toolCallHandler.handleToolCallCreated(chunk)

src/renderer/src/aiCore/chunk/handleToolCallChunk.ts

@@ -208,14 +208,14 @@ export class ToolCallChunkHandler {
       id: toolCallId,
       tool: tool,
       arguments: args,
-      status: 'invoking',
+      status: 'pending',
       toolCallId: toolCallId
     }
 
     // 调用 onChunk
     if (this.onChunk) {
       this.onChunk({
-        type: ChunkType.MCP_TOOL_IN_PROGRESS,
+        type: ChunkType.MCP_TOOL_PENDING,
         responses: [toolResponse]
       })
     }

src/renderer/src/aiCore/index_new.ts

@@ -236,7 +236,7 @@ export default class ModernAiProvider {
 
       const streamResult = await executor.streamText(
         modelId,
-        params,
+        { ...params, experimental_context: { onChunk: config.onChunk } },
         middlewares.length > 0 ? { middlewares } : undefined
       )
 

src/renderer/src/aiCore/utils/mcp.ts

@@ -1,9 +1,10 @@
-import { Tool } from '@cherrystudio/ai-core'
 import { loggerService } from '@logger'
 // import { AiSdkTool, ToolCallResult } from '@renderer/aiCore/tools/types'
 import { MCPTool, MCPToolResponse } from '@renderer/types'
-import { callMCPTool } from '@renderer/utils/mcp-tools'
-import { jsonSchema, tool } from 'ai'
+import { Chunk, ChunkType } from '@renderer/types/chunk'
+import { callMCPTool, getMcpServerByTool, isToolAutoApproved } from '@renderer/utils/mcp-tools'
+import { requestToolConfirmation } from '@renderer/utils/userConfirmation'
+import { jsonSchema, type Tool, tool } from 'ai'
 import { JSONSchema7 } from 'json-schema'
 
 const logger = loggerService.withContext('MCP-utils')
@@ -31,18 +32,52 @@ export function convertMcpToolsToAiSdkTools(mcpTools: MCPTool[]): Record<string,
     tools[mcpTool.name] = tool({
       description: mcpTool.description || `Tool from ${mcpTool.serverName}`,
       inputSchema: jsonSchema(mcpTool.inputSchema as JSONSchema7),
-      execute: async (params) => {
+      execute: async (params, { toolCallId, experimental_context }) => {
+        const { onChunk } = experimental_context as { onChunk: (chunk: Chunk) => void }
         // 创建适配的 MCPToolResponse 对象
         const toolResponse: MCPToolResponse = {
-          id: `tool_${Date.now()}_${Math.random().toString(36).substr(2, 9)}`,
+          id: toolCallId,
           tool: mcpTool,
           arguments: params,
-          status: 'invoking',
-          toolCallId: `call_${Date.now()}_${Math.random().toString(36).substr(2, 9)}`
+          status: 'pending',
+          toolCallId
         }
 
         try {
-          // 复用现有的 callMCPTool 函数
+          // 检查是否启用自动批准
+          const server = getMcpServerByTool(mcpTool)
+          const isAutoApproveEnabled = isToolAutoApproved(mcpTool, server)
+
+          let confirmed = true
+          if (!isAutoApproveEnabled) {
+            // 请求用户确认
+            logger.debug(`Requesting user confirmation for tool: ${mcpTool.name}`)
+            confirmed = await requestToolConfirmation(toolResponse.id)
+          }
+
+          if (!confirmed) {
+            // 用户拒绝执行工具
+            logger.debug(`User cancelled tool execution: ${mcpTool.name}`)
+            return {
+              content: [
+                {
+                  type: 'text',
+                  text: `User declined to execute tool "${mcpTool.name}".`
+                }
+              ],
+              isError: false
+            }
+          }
+
+          // 用户确认或自动批准，执行工具
+          toolResponse.status = 'invoking'
+          logger.debug(`Executing tool: ${mcpTool.name}`)
+
+          onChunk({
+            type: ChunkType.MCP_TOOL_IN_PROGRESS,
+            responses: [toolResponse]
+          })
+
           const result = await callMCPTool(toolResponse)
 
           // 返回结果，AI SDK 会处理序列化