debug: patch parse_tool_call to identify source of string args

Added global monkey-patch of langchain_core's parse_tool_call to log
the type of 'args' it returns. This will definitively show whether:
1. parse_tool_call is returning string args (bug in langchain_core)
2. Something else is modifying the result after parse_tool_call returns
3. AIMessage construction is getting tool_calls from a different source

This is the critical diagnostic to find the root cause.

agent/chat_model_wrapper.py

@@ -32,14 +32,37 @@ class ToolCallArgsParsingWrapper:
             # Model doesn't have this method (e.g., MockChatModel), skip patching
             return
 
+        # CRITICAL: Also patch parse_tool_call to see what it's returning
+        from langchain_core.output_parsers import openai_tools
+        original_parse_tool_call = openai_tools.parse_tool_call
+
+        def patched_parse_tool_call(raw_tool_call, *, partial=False, strict=False, return_id=True):
+            """Patched parse_tool_call to log what it returns"""
+            result = original_parse_tool_call(raw_tool_call, partial=partial, strict=strict, return_id=return_id)
+            if result:
+                args_type = type(result.get('args', None)).__name__
+                print(f"[DIAGNOSTIC] parse_tool_call returned: args type = {args_type}")
+                if args_type == 'str':
+                    print(f"[DIAGNOSTIC] ⚠️ BUG FOUND! parse_tool_call returned STRING args: {result['args']}")
+            return result
+
+        # Replace globally
+        openai_tools.parse_tool_call = patched_parse_tool_call
+
         original_create_chat_result = self.wrapped_model._create_chat_result
 
         @wraps(original_create_chat_result)
         def patched_create_chat_result(response: Any, generation_info: Optional[Dict] = None):
             """Patched version with diagnostic logging and args parsing"""
+            import traceback
             response_dict = response if isinstance(response, dict) else response.model_dump()
 
             # DIAGNOSTIC: Log response structure for debugging
+            print(f"\n[DIAGNOSTIC] _create_chat_result called")
+            print(f"  Response type: {type(response)}")
+            print(f"  Call stack:")
+            for line in traceback.format_stack()[-5:-1]:  # Show last 4 stack frames
+                print(f"    {line.strip()}")
             print(f"\n[DIAGNOSTIC] Response structure:")
             print(f"  Response keys: {list(response_dict.keys())}")
 
@@ -116,7 +139,18 @@ class ToolCallArgsParsingWrapper:
                                         # Keep as-is if serialization fails
 
             # Call original method with fixed response
-            return original_create_chat_result(response_dict, generation_info)
+            print(f"[DIAGNOSTIC] Calling original_create_chat_result...")
+            result = original_create_chat_result(response_dict, generation_info)
+            print(f"[DIAGNOSTIC] original_create_chat_result returned successfully")
+            print(f"[DIAGNOSTIC] Result type: {type(result)}")
+            if hasattr(result, 'generations') and result.generations:
+                gen = result.generations[0]
+                if hasattr(gen, 'message') and hasattr(gen.message, 'tool_calls'):
+                    print(f"[DIAGNOSTIC] Result has {len(gen.message.tool_calls)} tool_calls")
+                    if gen.message.tool_calls:
+                        tc = gen.message.tool_calls[0]
+                        print(f"[DIAGNOSTIC] tool_calls[0]['args'] type in result: {type(tc['args'])}")
+            return result
 
         # Replace the method
         self.wrapped_model._create_chat_result = patched_create_chat_result