fix: openai response tool use (#7332)

* fix: openai response tool use

- Added OpenAIResponseStreamListener interface for handling OpenAI response streams.
- Implemented attachRawStreamListener method in OpenAIResponseAPIClient to manage raw output.
- Updated RawStreamListenerMiddleware to integrate OpenAI response handling.
- Refactored BaseApiClient to remove unused attachRawStreamListener method.
- Improved buildSdkMessages to handle OpenAI response formats.

* fix: remove logging from StreamAdapterMiddleware

- Removed Logger.info call from StreamAdapterMiddleware to streamline output and reduce unnecessary logging.

* fix: update attachRawStreamListener to return a Promise

- Changed attachRawStreamListener method in OpenAIResponseAPIClient to be asynchronous, returning a Promise for better handling of raw output.
- Updated RawStreamListenerMiddleware to await the result of attachRawStreamListener, ensuring proper flow of data handling.

* refactor: enhance attachRawStreamListener to return a ReadableStream

- Updated the attachRawStreamListener method in OpenAIResponseAPIClient to return a ReadableStream, allowing for more efficient handling of streamed responses.
- Modified RawStreamListenerMiddleware to accommodate the new return type, ensuring proper integration of the transformed stream into the middleware flow.

* refactor: update getResponseChunkTransformer to accept CompletionsContext

- Modified the getResponseChunkTransformer method in BaseApiClient and its implementations to accept a CompletionsContext parameter, enhancing the flexibility of response handling.
- Adjusted related middleware and client classes to ensure compatibility with the new method signature, improving the overall integration of response transformations.

* refactor: update getResponseChunkTransformer to accept CompletionsContext

- Modified the getResponseChunkTransformer method in AihubmixAPIClient to accept a CompletionsContext parameter, enhancing the flexibility of response handling.
- Ensured compatibility with the updated method signature across related client classes.

src/renderer/src/aiCore/clients/AihubmixAPIClient.ts

@@ -20,6 +20,7 @@ import {
   SdkToolCall
 } from '@renderer/types/sdk'
 
+import { CompletionsContext } from '../middleware/types'
 import { AnthropicAPIClient } from './anthropic/AnthropicAPIClient'
 import { BaseApiClient } from './BaseApiClient'
 import { GeminiAPIClient } from './gemini/GeminiAPIClient'
@@ -163,8 +164,8 @@ export class AihubmixAPIClient extends BaseApiClient {
     return this.currentClient.getRequestTransformer()
   }
 
-  getResponseChunkTransformer(): ResponseChunkTransformer<SdkRawChunk> {
-    return this.currentClient.getResponseChunkTransformer()
+  getResponseChunkTransformer(ctx: CompletionsContext): ResponseChunkTransformer<SdkRawChunk> {
+    return this.currentClient.getResponseChunkTransformer(ctx)
   }
 
   convertMcpToolsToSdkTools(mcpTools: MCPTool[]): SdkTool[] {

src/renderer/src/aiCore/clients/BaseApiClient.ts

@@ -42,7 +42,8 @@ import { defaultTimeout } from '@shared/config/constant'
 import Logger from 'electron-log/renderer'
 import { isEmpty } from 'lodash'
 
-import { ApiClient, RawStreamListener, RequestTransformer, ResponseChunkTransformer } from './types'
+import { CompletionsContext } from '../middleware/types'
+import { ApiClient, RequestTransformer, ResponseChunkTransformer } from './types'
 
 /**
  * Abstract base class for API clients.
@@ -95,7 +96,7 @@ export abstract class BaseApiClient<
   // 在 CoreRequestToSdkParamsMiddleware中使用
   abstract getRequestTransformer(): RequestTransformer<TSdkParams, TMessageParam>
   // 在RawSdkChunkToGenericChunkMiddleware中使用
-  abstract getResponseChunkTransformer(): ResponseChunkTransformer<TRawChunk>
+  abstract getResponseChunkTransformer(ctx: CompletionsContext): ResponseChunkTransformer<TRawChunk>
 
   /**
    * 工具转换
@@ -129,17 +130,6 @@ export abstract class BaseApiClient<
    */
   abstract extractMessagesFromSdkPayload(sdkPayload: TSdkParams): TMessageParam[]
 
-  /**
-   * 附加原始流监听器
-   */
-  public attachRawStreamListener<TListener extends RawStreamListener<TRawChunk>>(
-    rawOutput: TRawOutput,
-    // eslint-disable-next-line @typescript-eslint/no-unused-vars
-    _listener: TListener
-  ): TRawOutput {
-    return rawOutput
-  }
-
   /**
    * 通用函数
    **/

src/renderer/src/aiCore/clients/anthropic/AnthropicAPIClient.ts

@@ -367,7 +367,7 @@ export class AnthropicAPIClient extends BaseApiClient<
    * Anthropic专用的原始流监听器
    * 处理MessageStream对象的特定事件
    */
-  override attachRawStreamListener(
+  attachRawStreamListener(
     rawOutput: AnthropicSdkRawOutput,
     listener: RawStreamListener<AnthropicSdkRawChunk>
   ): AnthropicSdkRawOutput {

src/renderer/src/aiCore/clients/openai/OpenAIApiClient.ts

@@ -494,7 +494,7 @@ export class OpenAIAPIClient extends OpenAIBaseClient<
   }
 
   // 在RawSdkChunkToGenericChunkMiddleware中使用
-  getResponseChunkTransformer = (): ResponseChunkTransformer<OpenAISdkRawChunk> => {
+  getResponseChunkTransformer(): ResponseChunkTransformer<OpenAISdkRawChunk> {
     let hasBeenCollectedWebSearch = false
     const collectWebSearchData = (
       chunk: OpenAISdkRawChunk,

src/renderer/src/aiCore/clients/openai/OpenAIResponseAPIClient.ts

@@ -1,4 +1,5 @@
 import { GenericChunk } from '@renderer/aiCore/middleware/schemas'
+import { CompletionsContext } from '@renderer/aiCore/middleware/types'
 import {
   isOpenAIChatCompletionOnlyModel,
   isSupportedReasoningEffortOpenAIModel,
@@ -38,6 +39,7 @@ import { buildSystemPrompt } from '@renderer/utils/prompt'
 import { MB } from '@shared/config/constant'
 import { isEmpty } from 'lodash'
 import OpenAI from 'openai'
+import { ResponseInput } from 'openai/resources/responses/responses'
 
 import { RequestTransformer, ResponseChunkTransformer } from '../types'
 import { OpenAIAPIClient } from './OpenAIApiClient'
@@ -225,9 +227,15 @@ export class OpenAIResponseAPIClient extends OpenAIBaseClient<
     return
   }
 
+  private convertResponseToMessageContent(response: OpenAI.Responses.Response): ResponseInput {
+    const content: OpenAI.Responses.ResponseInput = []
+    content.push(...response.output)
+    return content
+  }
+
   public buildSdkMessages(
     currentReqMessages: OpenAIResponseSdkMessageParam[],
-    output: string | undefined,
+    output: OpenAI.Responses.Response | undefined,
     toolResults: OpenAIResponseSdkMessageParam[],
     toolCalls: OpenAIResponseSdkToolCall[]
   ): OpenAIResponseSdkMessageParam[] {
@@ -239,11 +247,9 @@ export class OpenAIResponseAPIClient extends OpenAIBaseClient<
       return [...currentReqMessages, ...(toolCalls || []), ...(toolResults || [])]
     }
 
-    const assistantMessage: OpenAIResponseSdkMessageParam = {
-      role: 'assistant',
-      content: [{ type: 'input_text', text: output }]
-    }
-    const newReqMessages = [...currentReqMessages, assistantMessage, ...(toolCalls || []), ...(toolResults || [])]
+    const content = this.convertResponseToMessageContent(output)
+
+    const newReqMessages = [...currentReqMessages, ...content, ...(toolResults || [])]
     return newReqMessages
   }
 
@@ -415,13 +421,17 @@ export class OpenAIResponseAPIClient extends OpenAIBaseClient<
     }
   }
 
-  getResponseChunkTransformer(): ResponseChunkTransformer<OpenAIResponseSdkRawChunk> {
+  getResponseChunkTransformer(ctx: CompletionsContext): ResponseChunkTransformer<OpenAIResponseSdkRawChunk> {
     const toolCalls: OpenAIResponseSdkToolCall[] = []
     const outputItems: OpenAI.Responses.ResponseOutputItem[] = []
+    let hasBeenCollectedToolCalls = false
     return () => ({
       async transform(chunk: OpenAIResponseSdkRawChunk, controller: TransformStreamDefaultController<GenericChunk>) {
         // 处理chunk
         if ('output' in chunk) {
+          if (ctx._internal?.toolProcessingState) {
+            ctx._internal.toolProcessingState.output = chunk
+          }
           for (const output of chunk.output) {
             switch (output.type) {
               case 'message':
@@ -463,6 +473,22 @@ export class OpenAIResponseAPIClient extends OpenAIBaseClient<
                 })
             }
           }
+          if (toolCalls.length > 0) {
+            controller.enqueue({
+              type: ChunkType.MCP_TOOL_CREATED,
+              tool_calls: toolCalls
+            })
+          }
+          controller.enqueue({
+            type: ChunkType.LLM_RESPONSE_COMPLETE,
+            response: {
+              usage: {
+                prompt_tokens: chunk.usage?.input_tokens || 0,
+                completion_tokens: chunk.usage?.output_tokens || 0,
+                total_tokens: chunk.usage?.total_tokens || 0
+              }
+            }
+          })
         } else {
           switch (chunk.type) {
             case 'response.output_item.added':
@@ -510,7 +536,8 @@ export class OpenAIResponseAPIClient extends OpenAIBaseClient<
                 if (outputItem.type === 'function_call') {
                   toolCalls.push({
                     ...outputItem,
-                    arguments: chunk.arguments
+                    arguments: chunk.arguments,
+                    status: 'completed'
                   })
                 }
               }
@@ -526,15 +553,26 @@ export class OpenAIResponseAPIClient extends OpenAIBaseClient<
                   }
                 })
               }
-              if (toolCalls.length > 0) {
+              if (toolCalls.length > 0 && !hasBeenCollectedToolCalls) {
                 controller.enqueue({
                   type: ChunkType.MCP_TOOL_CREATED,
                   tool_calls: toolCalls
                 })
+                hasBeenCollectedToolCalls = true
               }
               break
             }
             case 'response.completed': {
+              if (ctx._internal?.toolProcessingState) {
+                ctx._internal.toolProcessingState.output = chunk.response
+              }
+              if (toolCalls.length > 0 && !hasBeenCollectedToolCalls) {
+                controller.enqueue({
+                  type: ChunkType.MCP_TOOL_CREATED,
+                  tool_calls: toolCalls
+                })
+                hasBeenCollectedToolCalls = true
+              }
               const completion_tokens = chunk.response.usage?.output_tokens || 0
               const total_tokens = chunk.response.usage?.total_tokens || 0
               controller.enqueue({

src/renderer/src/aiCore/clients/types.ts

@@ -3,6 +3,8 @@ import { Assistant, MCPTool, MCPToolResponse, Model, ToolCallResponse } from '@r
 import { Provider } from '@renderer/types'
 import {
   AnthropicSdkRawChunk,
+  OpenAIResponseSdkRawChunk,
+  OpenAIResponseSdkRawOutput,
   OpenAISdkRawChunk,
   SdkMessageParam,
   SdkParams,
@@ -14,6 +16,7 @@ import {
 import OpenAI from 'openai'
 
 import { CompletionsParams, GenericChunk } from '../middleware/schemas'
+import { CompletionsContext } from '../middleware/types'
 
 /**
  * 原始流监听器接口
@@ -33,6 +36,14 @@ export interface OpenAIStreamListener extends RawStreamListener<OpenAISdkRawChun
   onFinishReason?: (reason: string) => void
 }
 
+/**
+ * OpenAI Response 专用的流监听器
+ */
+export interface OpenAIResponseStreamListener<TChunk extends OpenAIResponseSdkRawChunk = OpenAIResponseSdkRawChunk>
+  extends RawStreamListener<TChunk> {
+  onMessage?: (response: OpenAIResponseSdkRawOutput) => void
+}
+
 /**
  * Anthropic 专用的流监听器
  */
@@ -101,7 +112,7 @@ export interface ApiClient<
   // SDK相关方法
   getSdkInstance(): Promise<TSdkInstance> | TSdkInstance
   getRequestTransformer(): RequestTransformer<TSdkParams, TMessageParam>
-  getResponseChunkTransformer(): ResponseChunkTransformer<TRawChunk>
+  getResponseChunkTransformer(ctx: CompletionsContext): ResponseChunkTransformer<TRawChunk>
 
   // 原始流监听方法
   attachRawStreamListener?(rawOutput: TRawOutput, listener: RawStreamListener<TRawChunk>): TRawOutput

src/renderer/src/aiCore/index.ts

@@ -76,7 +76,7 @@ export default class AiProvider {
       if (!(this.apiClient instanceof OpenAIAPIClient)) {
         builder.remove(ThinkingTagExtractionMiddlewareName)
       }
-      if (!(this.apiClient instanceof AnthropicAPIClient)) {
+      if (!(this.apiClient instanceof AnthropicAPIClient) && !(this.apiClient instanceof OpenAIResponseAPIClient)) {
         builder.remove(RawStreamListenerMiddlewareName)
       }
       if (!params.enableWebSearch) {

src/renderer/src/aiCore/middleware/core/RawStreamListenerMiddleware.ts

@@ -15,8 +15,6 @@ export const RawStreamListenerMiddleware: CompletionsMiddleware =
 
     // 在这里可以监听到从SDK返回的最原始流
     if (result.rawOutput) {
-      console.log(`[${MIDDLEWARE_NAME}] 检测到原始SDK输出，准备附加监听器`)
-
       const providerType = ctx.apiClientInstance.provider.type
       // TODO: 后面下放到AnthropicAPIClient
       if (providerType === 'anthropic') {

src/renderer/src/aiCore/middleware/core/ResponseTransformMiddleware.ts

@@ -37,7 +37,7 @@ export const ResponseTransformMiddleware: CompletionsMiddleware =
         }
 
         // 获取响应转换器
-        const responseChunkTransformer = apiClient.getResponseChunkTransformer?.()
+        const responseChunkTransformer = apiClient.getResponseChunkTransformer(ctx)
         if (!responseChunkTransformer) {
           Logger.warn(`[${MIDDLEWARE_NAME}] No ResponseChunkTransformer available, skipping transformation`)
           return result

src/renderer/src/aiCore/middleware/core/StreamAdapterMiddleware.ts

@@ -25,7 +25,6 @@ export const StreamAdapterMiddleware: CompletionsMiddleware =
     // 但是这个中间件的职责是流适配，是否在这调用优待商榷
     // 调用下游中间件
     const result = await next(ctx, params)
-
     if (
       result.rawOutput &&
       !(result.rawOutput instanceof ReadableStream) &&

src/renderer/src/aiCore/middleware/core/TransformCoreToSdkParamsMiddleware.ts

@@ -14,8 +14,6 @@ export const TransformCoreToSdkParamsMiddleware: CompletionsMiddleware =
   () =>
   (next) =>
   async (ctx: CompletionsContext, params: CompletionsParams): Promise<CompletionsResult> => {
-    Logger.debug(`🔄 [${MIDDLEWARE_NAME}] Starting core to SDK params transformation:`, ctx)
-
     const internal = ctx._internal
 
     // 🔧 检测递归调用：检查 params 中是否携带了预处理的 SDK 消息