diff --git a/common/chat.cpp b/common/chat.cpp
index 4350a9ec6be..7f6428dddf1 100644
--- a/common/chat.cpp
+++ b/common/chat.cpp
@@ -2720,20 +2720,20 @@ static common_chat_params common_chat_params_init_laguna(const common_chat_templ
         // "<assistant>\n" followed by the opening <think> (thinking enabled) or
         // </think> (disabled). Consume it, then capture reasoning. The model's
         // output therefore starts with the reasoning text (no leading <think>).
-        const std::string GEN_PROMPT = "<assistant>\n";
+        auto gen_prompt = p.literal("<assistant>") + p.optional(p.space());
         auto head = p.eps();
         if (extract_reasoning && inputs.enable_thinking) {
             // The model normally closes reasoning with </think> before a tool
             // call, but sometimes emits <tool_call> directly without it.
             // Terminate reasoning on whichever marker comes first so the call is
             // not swallowed into reasoning_content; consume </think> if present.
-            head = p.literal(GEN_PROMPT + THINK_START) +
+            head = gen_prompt + p.literal(THINK_START) +
                    p.reasoning(p.until_one_of({ THINK_END, CALL_START })) +
                    p.optional(p.literal(THINK_END));
         } else if (extract_reasoning) {
-            head = p.literal(GEN_PROMPT + THINK_END);
+            head = gen_prompt + p.literal(THINK_END);
         } else {
-            head = p.literal(GEN_PROMPT) +
+            head = gen_prompt +
                    p.optional(p.literal(THINK_START)) + p.optional(p.literal(THINK_END));
         }