diff --git a/packages/core/src/core/openaiContentGenerator/provider/deepseek.test.ts b/packages/core/src/core/openaiContentGenerator/provider/deepseek.test.ts
index cff092bdc..a38ffb52b 100644
--- a/packages/core/src/core/openaiContentGenerator/provider/deepseek.test.ts
+++ b/packages/core/src/core/openaiContentGenerator/provider/deepseek.test.ts
@@ -180,9 +180,9 @@ describe('DeepSeekOpenAICompatibleProvider', () => {
     });
 
     // https://github.com/QwenLM/qwen-code/issues/3695 — DeepSeek's thinking
-    // mode rejects subsequent requests when a prior tool-calling assistant
-    // turn omits reasoning_content, even if the model itself returned no
-    // reasoning text. The provider must always send the field.
+    // mode rejects subsequent requests when any prior assistant turn omits
+    // reasoning_content, even if the model itself returned no reasoning text.
+    // The provider must always send the field.
     it('injects empty reasoning_content on tool-calling assistant turns missing it', () => {
       const originalRequest: OpenAI.Chat.ChatCompletionCreateParams = {
         model: 'deepseek-v4-flash',
@@ -245,7 +245,7 @@ describe('DeepSeekOpenAICompatibleProvider', () => {
       expect(assistant.reasoning_content).toBe('Let me glob first.');
     });
 
-    it('does not add reasoning_content to assistant turns without tool_calls', () => {
+    it('injects empty reasoning_content on assistant turns without tool_calls', () => {
       const originalRequest: OpenAI.Chat.ChatCompletionCreateParams = {
         model: 'deepseek-v4-flash',
         messages: [
@@ -259,7 +259,7 @@ describe('DeepSeekOpenAICompatibleProvider', () => {
       const assistant = result.messages?.[1] as {
         reasoning_content?: string;
       };
-      expect(assistant.reasoning_content).toBeUndefined();
+      expect(assistant.reasoning_content).toBe('');
     });
   });
 
diff --git a/packages/core/src/core/openaiContentGenerator/provider/deepseek.ts b/packages/core/src/core/openaiContentGenerator/provider/deepseek.ts
index 235c5cf00..0500cc939 100644
--- a/packages/core/src/core/openaiContentGenerator/provider/deepseek.ts
+++ b/packages/core/src/core/openaiContentGenerator/provider/deepseek.ts
@@ -108,19 +108,16 @@ function flattenContentParts(
 }
 
 // DeepSeek's thinking mode requires reasoning_content to be replayed on every
-// prior assistant turn that carried tool_calls. The model may legitimately
-// return a tool round without reasoning text, so the field can be missing
-// when we rebuild the request. Send an empty string in that case so the API
-// contract is satisfied. https://github.com/QwenLM/qwen-code/issues/3695
+// prior assistant turn, including ones without tool_calls. The model may
+// legitimately return a turn without reasoning text, so the field can be
+// missing when we rebuild the request. Send an empty string in that case so
+// the API contract is satisfied. https://github.com/QwenLM/qwen-code/issues/3695
 function ensureReasoningContentOnToolCalls(
   message: OpenAI.Chat.ChatCompletionMessageParam,
 ): OpenAI.Chat.ChatCompletionMessageParam {
   if (message.role !== 'assistant') {
     return message;
   }
-  if (!Array.isArray(message.tool_calls) || message.tool_calls.length === 0) {
-    return message;
-  }
   const extended = message as ExtendedChatCompletionAssistantMessageParam;
   if (
     typeof extended.reasoning_content === 'string' &&