diff --git a/src/agents/workflows/problem_splitter.py b/src/agents/workflows/problem_splitter.py
index 7aee608..17a4754 100644
--- a/src/agents/workflows/problem_splitter.py
+++ b/src/agents/workflows/problem_splitter.py
@@ -117,7 +117,7 @@ def split_problems(text: str, use_llm: bool = True) -> ProblemSplit:
     print("---SPLITTER: Using LLM for problem splitting---")
 
     try:
-        model = get_model(OpenRouterModelName.GPT_5_MINI)
+        model = get_model(OpenRouterModelName.GPT_5_MINI).with_config(tags=["skip_stream"])
 
         messages = [
             SystemMessage(content=SPLIT_SYSTEM_PROMPT),
diff --git a/src/agents/workflows/review_logic.py b/src/agents/workflows/review_logic.py
index 96a5308..8c804ce 100644
--- a/src/agents/workflows/review_logic.py
+++ b/src/agents/workflows/review_logic.py
@@ -228,7 +228,8 @@ def run_review(state: Dict[str, Any]) -> Dict[str, Any]:
         print(
             f"---REVIEW: MODEL={MODEL_NAME} openrouter_key_set={bool(settings.OPENROUTER_API_KEY)}---"
         )
-        model = get_model(MODEL_NAME)
+        # Review 노드는 내부 JSON만 사용하므로 토큰 스트리밍을 비활성화한다.
+        model = get_model(MODEL_NAME).with_config(tags=["skip_stream"])
         prompt = [
             SystemMessage(content=REVIEW_SYSTEM_PROMPT),
             HumanMessage(
diff --git a/src/service/service.py b/src/service/service.py
index 2faa04a..967846a 100644
--- a/src/service/service.py
+++ b/src/service/service.py
@@ -269,7 +269,12 @@ async def _handle_input(
     )
 
     input: Command | dict[str, Any]
-    input = {"messages": [HumanMessage(content=user_input.message)]}
+    input = {
+        "messages": [HumanMessage(content=user_input.message)],
+        # 이전 실행에서 체크포인터가 복원한 stale 값을 매 실행마다 초기화
+        "final_output": {},
+        "partial_responses": [],
+    }
 
     if user_input.files_url:
         files = list(user_input.files_url)