Merge pull request #1415 from QwenLM/fix/openai-reasoning-config

fix(core): don’t force reasoning/topP defaults for OpenAI-compatible APIs

packages/core/src/core/openaiContentGenerator/pipeline.ts

@@ -317,15 +317,22 @@ export class ContentGenerationPipeline {
   }
 
   private buildReasoningConfig(): Record<string, unknown> {
-    const reasoning = this.contentGeneratorConfig.reasoning;
+    // Reasoning configuration for OpenAI-compatible endpoints is highly fragmented.
+    // For example, across common providers and models:
+    //
+    //   - deepseek-reasoner   — thinking is enabled by default and cannot be disabled
+    //   - glm-4.7             — thinking is enabled by default; can be disabled via `extra_body.thinking.enabled`
+    //   - kimi-k2-thinking    — thinking is enabled by default and cannot be disabled
+    //   - gpt-5.x series      — thinking is enabled by default; can be disabled via `reasoning.effort`
+    //   - qwen3 series        — model-dependent; can be manually disabled via `extra_body.enable_thinking`
+    //
+    // Given this inconsistency, we choose not to set any reasoning config here and
+    // instead rely on each model’s default behavior.
 
-    if (reasoning === false) {
+    // We plan to introduce provider- and model-specific settings to enable more
-      return {};
+    // fine-grained control over reasoning configuration.
-    }
 
-    return {
+    return {};
-      reasoning_effort: reasoning?.effort ?? 'medium',
-    };
   }
 
   /**

packages/core/src/core/openaiContentGenerator/provider/default.ts

@@ -58,8 +58,6 @@ export class DefaultOpenAICompatibleProvider
   }
 
   getDefaultGenerationConfig(): GenerateContentConfig {
-    return {
+    return {};
-      topP: 0.95,
-    };
   }
 }