fix(llm): 🐛 修复 max_tokens 为 undefined 时导致的请求问题 (#44)

当调用 LLM 接口时未提供 maxTokens 参数，会导致请求体中包含
max_tokens: null，某些 API 提供商可能将其解释为 0 或 1，
导致只返回极少内容。

修复内容:
- OpenAIClient: 仅在 maxTokens 不为 undefined 时添加 max_tokens
- OpenAIResponsesClient: 同上
- AnthropicClient: 同上
- GeminiClient: 仅在 maxTokens 不为 undefined 时添加 maxOutputTokens

同时添加测试用例验证未提供 maxTokens 时不会包含在请求体中。

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

Author: jorben

src/core/infrastructure/adapters/llm/AnthropicClient.ts

@@ -37,10 +37,14 @@ export class AnthropicClient extends LLMClient {
         model: modelName,
         messages: anthropicMessages,
         temperature: normalizedOptions.temperature ?? 0.7,
-        max_tokens: normalizedOptions.maxTokens,
         stream: normalizedOptions.stream || false
       };
 
+      // 只在提供了 maxTokens 时才添加到请求体
+      if (normalizedOptions.maxTokens !== undefined) {
+        requestBody.max_tokens = normalizedOptions.maxTokens;
+      }
+
       // Anthropic支持System指令，而不是作为一个普通消息
       const systemMessage = normalizedOptions.messages.find(msg => msg.role === 'system');
       if (systemMessage) {

src/core/infrastructure/adapters/llm/GeminiClient.ts

@@ -35,11 +35,15 @@ export class GeminiClient extends LLMClient {
         contents: geminiContents,
         generationConfig: {
           temperature: normalizedOptions.temperature ?? 0.7,
-          maxOutputTokens: normalizedOptions.maxTokens,
           topP: 0.95
         }
       };
 
+      // 只在提供了 maxTokens 时才添加到请求体
+      if (normalizedOptions.maxTokens !== undefined) {
+        requestBody.generationConfig.maxOutputTokens = normalizedOptions.maxTokens;
+      }
+
       // 添加响应格式（如果指定）
       if (normalizedOptions.response_format?.type === 'json_object') {
         // Gemini使用不同的键来指定JSON输出

src/core/infrastructure/adapters/llm/OpenAIClient.ts

@@ -34,10 +34,14 @@ export class OpenAIClient extends LLMClient {
         model: normalizedOptions.model || 'gpt-3.5-turbo',
         messages: openaiMessages,
         temperature: normalizedOptions.temperature ?? 0.7,
-        max_tokens: normalizedOptions.maxTokens,
         stream: normalizedOptions.stream || false
       };
 
+      // 只在提供了 maxTokens 时才添加到请求体
+      if (normalizedOptions.maxTokens !== undefined) {
+        requestBody.max_tokens = normalizedOptions.maxTokens;
+      }
+
       // 添加工具配置（如果有）
       if (normalizedOptions.tools && normalizedOptions.tools.length > 0) {
         requestBody.tools = normalizedOptions.tools;

src/core/infrastructure/adapters/llm/OpenAIResponsesClient.ts

@@ -35,10 +35,14 @@ export class OpenAIResponsesClient extends LLMClient {
         model: normalizedOptions.model || 'gpt-4o',
         input: input,
         temperature: normalizedOptions.temperature ?? 0.7,
-        max_tokens: normalizedOptions.maxTokens,
         stream: normalizedOptions.stream || false
       };
 
+      // 只在提供了 maxTokens 时才添加到请求体
+      if (normalizedOptions.maxTokens !== undefined) {
+        requestBody.max_tokens = normalizedOptions.maxTokens;
+      }
+
       // 添加系统指令（如果有）
       if (instructions) {
         requestBody.instructions = instructions;

src/core/infrastructure/adapters/llm/__tests__/OpenAIClient.test.ts

@@ -425,6 +425,22 @@ describe('OpenAIClient', () => {
       expect(body.temperature).toBe(0.7)
     })
 
+    it('should not include max_tokens when not provided', async () => {
+      const mockResponse = createMockOpenAIResponse('Response')
+      mockFetchSuccess(mockResponse)
+
+      await client.completion({
+        messages: [
+          { role: 'user', content: { type: 'text', text: 'Test' } }
+        ],
+        model: 'gpt-4o'
+      })
+
+      const callArgs = (global.fetch as any).mock.calls[0][1]
+      const body = JSON.parse(callArgs.body)
+      expect(body.max_tokens).toBeUndefined()
+    })
+
     it('should include response_format when specified', async () => {
       const mockResponse = createMockOpenAIResponse('{"key": "value"}')
       mockFetchSuccess(mockResponse)