fix: make vllm tool-calling turns tool-first

test/integration.vllm.test.ts

@@ -59,4 +59,37 @@ describe('vLLM Proxy Integration Test', () => {
       path: "/tmp/test.txt"
     });
   });
+
+  it('spoofs streaming responses for tool-calling turns without content chunks', async () => {
+    const requestFixturePath = path.join(__dirname, 'fixtures', 'vllm-like-request.json');
+    const responseFixturePath = path.join(__dirname, 'fixtures', 'vllm-xml-response.json');
+
+    const requestJson = JSON.parse(fs.readFileSync(requestFixturePath, 'utf8'));
+    const responseJson = JSON.parse(fs.readFileSync(responseFixturePath, 'utf8'));
+    requestJson.stream = true;
+
+    (global.fetch as any).mockResolvedValue({
+      ok: true,
+      json: async () => responseJson
+    });
+
+    const response = await server.inject({
+      method: 'POST',
+      url: '/v1/chat/completions',
+      payload: requestJson
+    });
+
+    expect(response.statusCode).toBe(200);
+    expect(response.headers['content-type']).toContain('text/event-stream');
+
+    const fetchArgs = (global.fetch as any).mock.calls[0];
+    const upstreamBody = JSON.parse(fetchArgs[1].body);
+    expect(upstreamBody.stream).toBe(false);
+
+    expect(response.payload).toContain('"role":"assistant"');
+    expect(response.payload).toContain('"tool_calls"');
+    expect(response.payload).toContain('"finish_reason":"tool_calls"');
+    expect(response.payload).not.toContain('"content"');
+    expect(response.payload).toContain('data: [DONE]');
+  });
 });

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);
   });