diff --git a/apps/notes/src/main.rs b/apps/notes/src/main.rs
index 153ac8c..3eabfba 100644
--- a/apps/notes/src/main.rs
+++ b/apps/notes/src/main.rs
@@ -114,6 +114,43 @@ async fn main() -> std::io::Result<()> {
         .route("/health", get(|| async { "ok" }))
         .merge(protected_api);
 
+    // 启动时把上次 pod 死前卡在中间状态的 recording 重新喂给 worker。
+    // 状态 transcribing/summarizing 是 worker 进程内存的，pod 重启就丢，
+    // db 还停留在原状态 → 不 resume 永远不会再动。
+    {
+        let stuck: Vec<i64> = {
+            let conn = state.db.lock().unwrap();
+            let mut stmt = conn
+                .prepare(
+                    "SELECT id FROM recordings
+                     WHERE status IN ('pending', 'transcribing', 'summarizing')
+                     ORDER BY id ASC",
+                )
+                .expect("prepare resume query");
+            stmt.query_map([], |r| r.get::<_, i64>(0))
+                .expect("run resume query")
+                .collect::<Result<Vec<_>, _>>()
+                .expect("collect resume ids")
+        };
+        if !stuck.is_empty() {
+            tracing::info!(count = stuck.len(), ids = ?stuck, "resuming stuck recordings");
+            for id in stuck {
+                let s = state.clone();
+                tokio::spawn(async move {
+                    // 改回 pending 让 worker 从头跑（idempotent）
+                    {
+                        let conn = s.db.lock().unwrap();
+                        let _ = conn.execute(
+                            "UPDATE recordings SET status = 'pending', error = NULL WHERE id = ?1",
+                            params![id],
+                        );
+                    }
+                    process_recording(s, id).await;
+                });
+            }
+        }
+    }
+
     let app = cube_core::base(dist).nest("/api", api);
     cube_core::serve(app, 8080).await
 }