diff --git a/backend/app/celery/celery_app.py b/backend/app/celery/celery_app.py
index 45364a72f..975948600 100644
--- a/backend/app/celery/celery_app.py
+++ b/backend/app/celery/celery_app.py
@@ -1,7 +1,7 @@
 import logging
 
 from celery import Celery
-from celery.signals import worker_process_init
+from celery.signals import task_prerun, worker_process_init
 from kombu import Exchange, Queue
 
 from app.core.config import settings
@@ -9,6 +9,24 @@
 logger = logging.getLogger(__name__)
 
 
+@task_prerun.connect
+def log_pool_status(task: "celery.Task", **_: object) -> None:  # type: ignore[name-defined]
+    """Log DB connection pool state before each task to detect connection leaks.
+
+    If checked_out equals pool size right when a task starts, connections
+    are being held across tasks (likely across LLM API calls) and leaking.
+    """
+    from app.core.db import engine
+    from sqlalchemy.pool import QueuePool
+
+    pool = engine.pool
+    if isinstance(pool, QueuePool):
+        logger.info(
+            f"[pool] task={task.name} checked_out={pool.checkedout()} "
+            f"size={pool.size()} overflow={pool.overflow()}"
+        )
+
+
 @worker_process_init.connect
 def warm_llm_modules(**_) -> None:
     """Import LLM service modules in each worker process right after fork.