Tool-call parsing: Qwen3 native <tool_call> XML fallback in _parse (parser-agnostic, zero overhead when server converts properly)

evalharness/model/adapter.py

@@ -108,9 +108,25 @@ def _key_for(api_base: str) -> str:
 
 
 def _parse_text_tool_calls(text: str) -> list:
-    """Extract [{"name":..,"arguments":{..}}] JSON from a text-protocol reply."""
+    """Extract tool calls from a text reply. Handles both shapes:
+    - JSON array:  [{"name":..,"arguments":{..}}]
+    - Qwen3 native XML: <tool_call>{"name":..,"arguments":{..}}</tool_call>
+      (for vLLM builds whose qwen3_xml parser doesn't convert to tool_calls)
+    """
     import re as _re
 
+    out = []
+    for m in _re.finditer(r'<tool_call>\s*(\{.*?\})\s*</tool_call>', text, _re.S):
+        try:
+            obj = json.loads(m.group(1))
+            if isinstance(obj, dict) and obj.get('name'):
+                out.append({'id': '', 'type': 'function',
+                            'function': {'name': obj['name'],
+                                         'arguments': json.dumps(obj.get('arguments', {}))}})
+        except (ValueError, TypeError):
+            continue
+    if out:
+        return out
     candidates = _re.findall(r'\[[\s\S]*?\]', text) or []
     for cand in candidates:
         try:
@@ -196,7 +212,7 @@ class OpenAICompatible(ModelAdapter):
         msg = choice.get('message') or {}
         calls = []
         raw_calls = list(msg.get('tool_calls') or [])
-        if not raw_calls and self.extra.get('tools_mode') == 'text':
+        if not raw_calls:
             raw_calls = _parse_text_tool_calls(msg.get('content') or '')
         for c in raw_calls:
             fn = c.get('function') or {}