Separate thinking tokens from meta — use __thinking key for display

stream_response() now yields {"__thinking": str} for thinking content
instead of {"__meta": str}. Removed [사고 과정]/[/사고 과정] marker
yields; consumers handle thinking display independently.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

services/agent/agent_service.py

@@ -257,8 +257,6 @@ class AgentService:
                 if isinstance(data, dict) and "__thinking" in data:
                     # thinking 첫 토큰 도착 시 agent 레이블 + prev_node 갱신
                     if "agent" != prev_node:
-                        if thinking_open:
-                            yield {"__meta": "\n[/사고 과정]\n"}
                         thinking_open = False
                         content_started = False
                         if lg:
@@ -267,10 +265,8 @@ class AgentService:
                             yield {"__meta": f"\n[LangGraph → {label}] ({elapsed:.2f}s)\n"}
                         prev_node = "agent"
                     if _think_verbose:
-                        if not thinking_open:
-                            yield {"__meta": "\n[사고 과정]\n"}
                         thinking_open = True
-                        yield {"__meta": data["__thinking"]}
+                        yield {"__thinking": data["__thinking"]}
                 continue
 
             # ── messages 이벤트 ──────────────────────────────────────
@@ -280,8 +276,6 @@ class AgentService:
             # ── 노드 전환 시 플래그 리셋 + 레이블 출력 ──────────────
             # (agent 레이블은 custom 이벤트 핸들러에서 이미 처리될 수 있으므로 중복 방지)
             if node != prev_node:
-                if thinking_open:
-                    yield {"__meta": "\n[/사고 과정]\n"}
                 thinking_open = False
                 content_started = False
                 if lg:
@@ -298,8 +292,6 @@ class AgentService:
             # ── agent 노드 — AIMessageChunk만 처리 (중복 방지) ──────
             if node == "agent" and isinstance(chunk, AIMessageChunk):
                 if chunk.tool_calls:
-                    if thinking_open:
-                        yield {"__meta": "\n[/사고 과정]\n"}
                     thinking_open = False
                     for tc in chunk.tool_calls:
                         pending_tool_calls[tc["id"]] = tc
@@ -314,8 +306,6 @@ class AgentService:
                             yield {"__meta": f"  [tool_call: {tc['name']}({args_str})]\n"}
 
                 elif chunk.content:
-                    if thinking_open:
-                        yield {"__meta": "\n[/사고 과정]\n"}
                     thinking_open = False
                     if lg and not content_started:
                         yield {"__meta": "\n[LangGraph → agent: 최종 답변 생성]\n\n"}
@@ -329,9 +319,7 @@ class AgentService:
                 if not content_started and not thinking_open:
                     thinking = chunk.additional_kwargs.get("thinking", "")
                     if thinking and _think_verbose:
-                        yield {"__meta": "\n[사고 과정]\n"}
-                        yield {"__meta": thinking}
-                        yield {"__meta": "\n[/사고 과정]\n"}
+                        yield {"__thinking": thinking}
                     if chunk.content:
                         if lg:
                             yield {"__meta": "\n[LangGraph → agent: 최종 답변 생성]\n\n"}
@@ -359,9 +347,7 @@ class AgentService:
                     result_lines = [b for b in chunk.content.split("\n\n") if b.strip()]
                     yield {"__meta": f"  [웹 검색 결과: {len(result_lines)}건 → agent 복귀]\n"}
 
-        if thinking_open:
-            yield {"__meta": "\n[/사고 과정]\n"}
-
+        thinking_open = False
         self._last_run_id = str(run_id)
 
         # 대화 내용을 MySQL에 저장