diff --git a/common/chat.cpp b/common/chat.cpp
index 21aa524e..49626ff4 100644
--- a/common/chat.cpp
+++ b/common/chat.cpp
@@ -1279,8 +1279,8 @@ static common_chat_params common_chat_params_init_kimi_k2(const common_chat_temp
         form.tool_start  = "<|tool_call_begin|>";
         form.tool_sep    = "<|tool_call_argument_begin|>{";
         form.key_start   = "\"";
-        form.key_val_sep = "\":";
-        form.val_end     = ",";
+        form.key_val_sep = "\": ";
+        form.val_end     = ", ";
         form.tool_end    = "}<|tool_call_end|>";
         form.scope_end   = "<|tool_calls_section_end|>";
         form.raw_argval  = false;
diff --git a/common/sampling.cpp b/common/sampling.cpp
index 0e19de4c..769fc06c 100644
--- a/common/sampling.cpp
+++ b/common/sampling.cpp
@@ -429,6 +429,12 @@ static llama_token llama_sampling_sample_impl(
         GGML_ASSERT(!original_logits.empty());
     }
     llama_token id = 0;
+    // Sample grammar first for resampling
+    if (ctx_sampling->grammar != NULL && is_resampling) {
+        float* logits = llama_get_logits_ith(ctx_main, idx);
+        // Apply grammar constraints to all candidates
+        llama_grammar_sample(ctx_sampling->grammar, ctx_main, &cur_p);        
+    }
 
     if (temp < 0.0) {
         // greedy sampling, with probs