feat: implement Stream Spoofing interceptor and robust format preservation patches

test/vllm-rewriter.test.ts

@@ -0,0 +1,53 @@
+import { describe, it, expect } from 'vitest';
+import { rewriteVllmResponse } from '../src/proxy/vllm-response-rewriter';
+
+describe('vLLM Response Rewriter', () => {
+  it('rewrites XML tool call in OpenAI choices content into structured tool_calls', () => {
+    const inputResponse = {
+      id: "chatcmpl-123",
+      choices: [{
+        index: 0,
+        message: {
+          role: "assistant",
+          content: "<function=read>\n<parameter=path>\n/tmp/test.txt\n</parameter>\n</function>"
+        }
+      }]
+    };
+
+    const result = rewriteVllmResponse(inputResponse);
+
+    expect(result.choices[0].message.content).toBe("");
+    expect(result.choices[0].message.tool_calls).toBeDefined();
+    expect(result.choices[0].message.tool_calls).toHaveLength(1);
+    
+    const toolCall = result.choices[0].message.tool_calls![0];
+    expect(toolCall.type).toBe('function');
+    expect(toolCall.function.name).toBe('read');
+    
+    const argsObject = JSON.parse(toolCall.function.arguments);
+    expect(argsObject).toEqual({ path: '/tmp/test.txt' });
+  });
+
+  it('does not touch response that already has tool_calls', () => {
+    const inputResponse = {
+      choices: [{
+        message: {
+          role: "assistant",
+          content: "Here are the calls",
+          tool_calls: [
+            {
+              id: "123",
+              type: "function",
+              function: { name: "read", arguments: "{}" }
+            }
+          ]
+        }
+      }]
+    };
+
+    const result = rewriteVllmResponse(inputResponse);
+
+    expect(result.choices[0].message.content).toBe("Here are the calls");
+    expect(result.choices[0].message.tool_calls).toHaveLength(1);
+  });
+});