diff --git a/gpttype_adapter.cpp b/gpttype_adapter.cpp
index 7ae78b36b..7ac64aacb 100644
--- a/gpttype_adapter.cpp
+++ b/gpttype_adapter.cpp
@@ -5176,12 +5176,18 @@ std::string gpttype_parse_chat_tool_calls(const std::string & generated_text,
         inputs.tool_choice = common_chat_tool_choice_parse_oaicompat(tool_choice.empty() ? "auto" : tool_choice);
         inputs.parallel_tool_calls = parallel_tool_calls;
         inputs.add_generation_prompt = true;
+        // Consume reasoning (including a <think> in the generation prompt) before tools.
+        inputs.reasoning_format = COMMON_REASONING_FORMAT_AUTO;
 
         if(!chat_template_kwargs_json.empty())
         {
             common_json kwargs = common_json::parse(chat_template_kwargs_json);
             if(kwargs.is_object())
             {
+                if(kwargs.contains("enable_thinking") && kwargs["enable_thinking"].is_boolean())
+                {
+                    inputs.enable_thinking = kwargs["enable_thinking"].get<bool>();
+                }
                 for(const auto & item : kwargs.items())
                 {
                     inputs.chat_template_kwargs[item.key()] = item.value().dump();
diff --git a/koboldcpp.py b/koboldcpp.py
index 6b25e9235..0dc225b01 100644
--- a/koboldcpp.py
+++ b/koboldcpp.py
@@ -5643,6 +5643,8 @@ class KcppServerRequestHandler(http.server.SimpleHTTPRequestHandler):
 
         utfprint("\nOutput: " + recvtxt,1)
 
+        # The native parser needs the full output to match reasoning in the generation prompt.
+        native_toolcall_text = recvtxt
         #handle potential think tags, but only chat completions will return them. the others just drop them
         reasoningtxt = ""
         if api_format==4 or api_format==8 or api_format==9: #chat completions, responses and anthropic messages, but only chat has reasoning returned
@@ -5668,7 +5670,7 @@ class KcppServerRequestHandler(http.server.SimpleHTTPRequestHandler):
             using_openai_tools = genparams.get('using_openai_tools', False)
             if using_openai_tools:
                 # first, let llama.cpp's chat parser handle known template-specific tool formats
-                tool_calls = native_parse_toolcall_tags(recvtxt, genparams)
+                tool_calls = native_parse_toolcall_tags(native_toolcall_text, genparams)
                 # fallback: check and potentially segment multiple tags for multi-tool calls
                 if not tool_calls:
                     tool_calls = repack_toolcall_tags(recvtxt,genparams.get('tools', []))