fix: make vllm tool-calling turns tool-first

test/vllm-rewriter.test.ts

@@ -28,12 +28,13 @@ describe('vLLM Response Rewriter', () => {
     expect(argsObject).toEqual({ path: '/tmp/test.txt' });
   });
 
-  it('does not touch response that already has tool_calls', () => {
+  it('normalizes response that already has tool_calls into a tool-first shape', () => {
     const inputResponse = {
       choices: [{
         message: {
           role: "assistant",
           content: "Here are the calls",
+          thinking: "<think>internal</think>",
           tool_calls: [
             {
               id: "123",
@@ -47,7 +48,8 @@ describe('vLLM Response Rewriter', () => {
 
     const result = rewriteVllmResponse(inputResponse);
 
-    expect(result.choices[0].message.content).toBe("Here are the calls");
+    expect(result.choices[0].message.content).toBe("");
+    expect(result.choices[0].message.thinking).toBe("");
     expect(result.choices[0].message.tool_calls).toHaveLength(1);
   });