fix: Responses API 工具调用事件格式错误导致 Codex CLI 无法多轮工作

修复 6 个问题:

1. requestId 未传入 convertStreamChunk, 每个 chunk 生成新 UUID,
   导致 state 不共享, response.completed 时 fullText 为空
2. TOTAL_CONTEXT_TOKENS=1000000 但 Claude 实际上下文只有 200K,
   token 数虚报 5 倍触发异常压缩
3. 默认模型名硬编码为 gpt-4.1-2025-04-14
4. 工具调用参数事件类型用了 response.custom_tool_call_input.delta,
   应为 response.function_call_arguments.delta
5. response.completed 的 output 只有空 message, 缺少 function_call 项,
   Codex 收不到工具调用结果无法进入多轮
6. 缺少 response.function_call_arguments.done 和工具专属
   response.output_item.done 事件

src/converters/strategies/ClaudeConverter.js

@@ -28,7 +28,14 @@ import {
     generateOutputTextDone,
     generateContentPartDone,
     generateOutputItemDone,
-    generateResponseCompleted
+    generateResponseCompleted,
+    generateOutputTextDelta,
+    streamStateManager,
+    startToolCall,
+    finishToolCall,
+    generateFunctionCallArgsDelta,
+    generateFunctionCallArgsDone,
+    generateFunctionCallOutputItemDone
 } from '../../providers/openai/openai-responses-core.mjs';
 
 /**
@@ -81,14 +88,14 @@ export class ClaudeConverter extends BaseConverter {
     /**
      * 转换流式响应块
      */
-    convertStreamChunk(chunk, targetProtocol, model) {
+    convertStreamChunk(chunk, targetProtocol, model, requestId) {
         switch (targetProtocol) {
             case MODEL_PROTOCOL_PREFIX.OPENAI:
                 return this.toOpenAIStreamChunk(chunk, model);
             case MODEL_PROTOCOL_PREFIX.GEMINI:
                 return this.toGeminiStreamChunk(chunk, model);
             case MODEL_PROTOCOL_PREFIX.OPENAI_RESPONSES:
-                return this.toOpenAIResponsesStreamChunk(chunk, model);
+                return this.toOpenAIResponsesStreamChunk(chunk, model, requestId);
             case MODEL_PROTOCOL_PREFIX.CODEX:
                 return this.toCodexStreamChunk(chunk, model);
             default:
@@ -1694,9 +1701,11 @@ export class ClaudeConverter extends BaseConverter {
             
             // 对于 tool_use 类型，添加工具调用项
             if (contentBlock && contentBlock.type === 'tool_use') {
+                startToolCall(responseId, contentBlock.id, contentBlock.name);
                 events.push({
                     item: {
                         id: contentBlock.id,
+                        call_id: contentBlock.id,
                         type: "function_call",
                         name: contentBlock.name,
                         arguments: "",
@@ -1715,13 +1724,7 @@ export class ClaudeConverter extends BaseConverter {
             
             // 处理文本增量
             if (delta && delta.type === 'text_delta') {
-                events.push({
+                events.push(generateOutputTextDelta(responseId, delta.text || ""));
-                    delta: delta.text || "",
-                    item_id: `msg_${uuidv4().replace(/-/g, '')}`,
-                    output_index: claudeChunk.index || 0,
-                    sequence_number: 3,
-                    type: "response.output_text.delta"
-                });
             }
             // 处理推理内容增量
             else if (delta && delta.type === 'thinking_delta') {
@@ -1735,61 +1738,51 @@ export class ClaudeConverter extends BaseConverter {
             }
             // 处理工具调用参数增量
             else if (delta && delta.type === 'input_json_delta') {
-                events.push({
+                const state = streamStateManager.getOrCreateState(responseId);
-                    delta: delta.partial_json || "",
+                const itemId = state.currentToolCall ? state.currentToolCall.id : 'unknown';
-                    item_id: `call_${uuidv4().replace(/-/g, '')}`,
+                events.push(generateFunctionCallArgsDelta(
-                    output_index: claudeChunk.index || 0,
+                    responseId, itemId, claudeChunk.index || 0, delta.partial_json || ""
-                    sequence_number: 3,
+                ));
-                    type: "response.custom_tool_call_input.delta"
-                });
             }
         }
 
         // content_block_stop 事件
         if (claudeChunk.type === 'content_block_stop') {
-            events.push({
+            const state = streamStateManager.getOrCreateState(responseId);
-                item_id: `msg_${uuidv4().replace(/-/g, '')}`,
+            if (state.currentToolCall) {
-                output_index: claudeChunk.index || 0,
+                const itemId = state.currentToolCall.id;
-                sequence_number: 4,
+                const outputIdx = claudeChunk.index || 0;
-                type: "response.output_item.done"
+                events.push(generateFunctionCallArgsDone(responseId, itemId, outputIdx));
-            });
+                const finished = finishToolCall(responseId);
+                if (finished) {
+                    events.push(generateFunctionCallOutputItemDone(responseId, finished, outputIdx));
+                }
+            }
         }
 
-        // message_delta 事件 - 流结束
+        // message_delta 事件 - 保存 usage 供 message_stop 使用
         if (claudeChunk.type === 'message_delta') {
-            // events.push(
+            if (claudeChunk.usage) {
-            //     generateOutputTextDone(responseId),
+                const state = streamStateManager.getOrCreateState(responseId);
-            //     generateContentPartDone(responseId),
+                state.savedUsage = {
-            //     generateOutputItemDone(responseId),
+                    input_tokens: claudeChunk.usage.input_tokens || 0,
-            //     generateResponseCompleted(responseId)
+                    input_tokens_details: { cached_tokens: claudeChunk.usage.cache_read_input_tokens || 0 },
-            // );
+                    output_tokens: claudeChunk.usage.output_tokens || 0,
-            
+                    output_tokens_details: { reasoning_tokens: 0 },
-            // 如果有 usage 信息，更新最后一个事件
+                    total_tokens: (claudeChunk.usage.input_tokens || 0) + (claudeChunk.usage.output_tokens || 0)
-            if (claudeChunk.usage && events.length > 0) {
+                };
-                const lastEvent = events[events.length - 1];
-                if (lastEvent.response) {
-                    lastEvent.response.usage = {
-                        input_tokens: claudeChunk.usage.input_tokens || 0,
-                        input_tokens_details: {
-                            cached_tokens: claudeChunk.usage.cache_read_input_tokens || 0
-                        },
-                        output_tokens: claudeChunk.usage.output_tokens || 0,
-                        output_tokens_details: {
-                            reasoning_tokens: 0
-                        },
-                        total_tokens: (claudeChunk.usage.input_tokens || 0) + (claudeChunk.usage.output_tokens || 0)
-                    };
-                }
             }
         }
 
         // message_stop 事件
         if (claudeChunk.type === 'message_stop') {
+            const state = streamStateManager.getOrCreateState(responseId);
+            const savedUsage = state.savedUsage || null;
             events.push(
                 generateOutputTextDone(responseId),
                 generateContentPartDone(responseId),
                 generateOutputItemDone(responseId),
-                generateResponseCompleted(responseId)
+                generateResponseCompleted(responseId, savedUsage)
             );
         }
 

src/providers/claude/claude-kiro.js

@@ -44,7 +44,7 @@ const KIRO_CONSTANTS = {
     AUTH_METHOD_SOCIAL: 'social',
     CHAT_TRIGGER_TYPE_MANUAL: 'MANUAL',
     ORIGIN_AI_EDITOR: 'AI_EDITOR',
-    TOTAL_CONTEXT_TOKENS: 1000000, // Kiro now supports 1M context window
+    TOTAL_CONTEXT_TOKENS: 200000, // Claude Sonnet 4.5 actual context is 200K
 };
 
 // 从 provider-models.js 获取支持的模型列表
@@ -1597,6 +1597,7 @@ async saveCredentialsToFile(filePath, newData) {
                 return this.callApi(method, model, body, isRetry, retryCount + 1);
             }
 
+            if (error.response && error.response.data) { logger.error('[Kiro] 400 Response body:', typeof error.response.data === 'string' ? error.response.data.substring(0, 500) : JSON.stringify(error.response.data).substring(0, 500)); }
             logger.error(`[Kiro] API call failed (Status: ${status}, Code: ${errorCode}):`, error.message);
             throw error;
         }

src/providers/openai/openai-responses-core.mjs

@@ -16,7 +16,9 @@ class StreamState {
         sequenceNumber: 0,
         model: null,
         status: 'in_progress',
-        startTime: Math.floor(Date.now() / 1000)
+        startTime: Math.floor(Date.now() / 1000),
+        toolCalls: [],
+        currentToolCall: null
       });
     }
     return this.states.get(requestId);
@@ -73,7 +75,7 @@ function generateResponseCreated(requestId, model) {
       incomplete_details: null,
       instructions: '',
       max_output_tokens: null,
-      model: state.model || 'gpt-4.1-2025-04-14',
+      model: state.model || 'claude-sonnet-4-6',
       output: [],
       parallel_tool_calls: true,
       previous_response_id: null,
@@ -110,7 +112,7 @@ function generateResponseInProgress(requestId) {
       incomplete_details: null,
       instructions: '',
       max_output_tokens: null,
-      model: state.model || 'gpt-4.1-2025-04-14',
+      model: state.model || 'claude-sonnet-4-6',
       output: [],
       parallel_tool_calls: true,
       previous_response_id: null,
@@ -269,25 +271,32 @@ function generateResponseCompleted(requestId, usage) {
       max_output_tokens: null,
       max_tool_calls: null,
       metadata: {},
-      model: state.model || 'gpt-4.1-2025-04-14',
+      model: state.model || 'claude-sonnet-4-6',
       object: 'response',
-      output: [
+      output: (() => {
-        {
+        const items = [];
-          id: state.msgId,
+        if (state.fullText) {
-          summary: [],
+          items.push({
-          type: 'message',
+            id: state.msgId, summary: [], type: 'message', role: 'assistant', status: 'completed',
-          role: 'assistant',
+            content: [{ type: 'output_text', text: state.fullText, annotations: [], logprobs: [] }]
-          status: 'completed',
+          });
-          content: [
-            {
-              type: 'output_text',
-              text: state.fullText,
-              annotations: [],
-              logprobs: []
-            }
-          ]
         }
-      ],
+        if (state.toolCalls && state.toolCalls.length > 0) {
+          for (const tc of state.toolCalls) {
+            items.push({
+              id: tc.id, call_id: tc.call_id || tc.id, type: 'function_call',
+              name: tc.name, arguments: tc.arguments || '{}', status: 'completed'
+            });
+          }
+        }
+        if (items.length === 0) {
+          items.push({
+            id: state.msgId, summary: [], type: 'message', role: 'assistant', status: 'completed',
+            content: [{ type: 'output_text', text: '', annotations: [], logprobs: [] }]
+          });
+        }
+        return items;
+      })(),
       parallel_tool_calls: true,
       previous_response_id: null,
       prompt_cache_key: null,
@@ -322,8 +331,50 @@ function generateResponseCompleted(requestId, usage) {
   };
 }
 
+
+function startToolCall(requestId, toolCallId, name) {
+  const state = streamStateManager.getOrCreateState(requestId);
+  state.currentToolCall = { id: toolCallId, call_id: toolCallId, name: name, arguments: '' };
+}
+
+function appendToolCallArgs(requestId, delta) {
+  const state = streamStateManager.getOrCreateState(requestId);
+  if (state.currentToolCall) state.currentToolCall.arguments += delta;
+}
+
+function finishToolCall(requestId) {
+  const state = streamStateManager.getOrCreateState(requestId);
+  if (state.currentToolCall) {
+    state.toolCalls.push({...state.currentToolCall});
+    const finished = state.currentToolCall;
+    state.currentToolCall = null;
+    return finished;
+  }
+  return null;
+}
+
+function generateFunctionCallArgsDelta(requestId, itemId, outputIndex, delta) {
+  appendToolCallArgs(requestId, delta);
+  return { type: 'response.function_call_arguments.delta', item_id: itemId, output_index: outputIndex, delta: delta };
+}
+
+function generateFunctionCallArgsDone(requestId, itemId, outputIndex) {
+  const state = streamStateManager.getOrCreateState(requestId);
+  const args = state.currentToolCall ? state.currentToolCall.arguments : '{}';
+  return { type: 'response.function_call_arguments.done', item_id: itemId, output_index: outputIndex, arguments: args };
+}
+
+function generateFunctionCallOutputItemDone(requestId, toolCall, outputIndex) {
+  return {
+    type: 'response.output_item.done', output_index: outputIndex,
+    item: { id: toolCall.id, call_id: toolCall.call_id || toolCall.id, type: 'function_call',
+      name: toolCall.name, arguments: toolCall.arguments || '{}', status: 'completed' }
+  };
+}
+
 // 导出流式状态管理器以供外部使用
 export { streamStateManager, generateResponseCreated, generateResponseInProgress,
   generateOutputItemAdded, generateContentPartAdded, generateOutputTextDelta,
   generateOutputTextDone, generateContentPartDone, generateOutputItemDone,
-  generateResponseCompleted };
+  generateResponseCompleted, startToolCall, appendToolCallArgs, finishToolCall,
+  generateFunctionCallArgsDelta, generateFunctionCallArgsDone, generateFunctionCallOutputItemDone };