diff --git a/server/app.py b/server/app.py
index e5d8fe6..f75eed2 100644
--- a/server/app.py
+++ b/server/app.py
@@ -23,6 +23,7 @@ from capcut_agent.pipeline import process_bg_template, process_paste
 from capcut_agent.paste import parse_paste
 from capcut_agent.draft import DEFAULT_DRAFT_ROOT, list_drafts, repair_layers
 from capcut_agent import comments as hlab
+from capcut_agent import recommend
 from capcut_agent import plan as autoplan
 from capcut_agent import prompts as prompt_store
 from capcut_agent.correct import GeminiQuotaError, has_gemini_key
@@ -530,6 +531,17 @@ async def auto_stream(aid: str) -> StreamingResponse:
             matched = hlab.match_window(comments, h["start"], h["end"])
             h["matched"] = matched
             h["candidates"] = hlab.top_liked(no_ts, set(matched), len(no_ts))
+            # 컷별 추천 — 실패해도 위의 matched/candidates 로 화면이 돌아간다.
+            # Gemini 호출이 섞여 있어 블로킹이므로 스레드로 뺀다.
+            try:
+                cuts, need = await asyncio.to_thread(
+                    recommend.build_highlight_cuts, h, comments)
+                if cuts:
+                    h["cuts"], h["need"] = cuts, need
+            except Exception as exc:  # noqa: BLE001 — 추천 실패가 생성을 막으면 안 된다
+                warnings.append(
+                    f"ID {h.get('id')} 컷별 추천 실패 — 기존 방식으로 표시 "
+                    f"({type(exc).__name__}: {exc})")
         yield _sse({"type": "result", "highlights": highlights,
                     "comments": comments, "warnings": warnings})