diff --git a/src/config.py b/src/config.py
index 8f8a1a78..59bdd857 100644
--- a/src/config.py
+++ b/src/config.py
@@ -11,7 +11,7 @@ class Config(BaseSettings):
     DATABASE_POOL_SIZE: int = 20
     DATABASE_POOL_TTL: int = 60 * 20  # 20 minutes
     DATABASE_POOL_PRE_PING: bool = True
-    DATABASE_POOL_TIMEOUT: int = 5  # seconds to wait for a connection before raising; keep low to fail fast
+    DATABASE_POOL_TIMEOUT: int = 5  # seconds; keep low to fail fast on pool exhaustion
 
     REDIS_URL: RedisDsn
     REDIS_HEALTH_CHECK_INTERVAL: int = 30
diff --git a/src/database.py b/src/database.py
index 54238685..1cab0763 100644
--- a/src/database.py
+++ b/src/database.py
@@ -20,10 +20,12 @@
     Table,
     UniqueConstraint,
     Update,
+    event,
     func,
     text,
 )
 from sqlalchemy.dialects.postgresql import JSONB
+from sqlalchemy.exc import DBAPIError
 from sqlalchemy.ext.asyncio import create_async_engine
 from sqlalchemy.pool import NullPool
 
@@ -58,6 +60,30 @@
 
 engine = create_async_engine(DATABASE_URL, **_engine_kwargs)
 
+
+@event.listens_for(engine.sync_engine, "handle_error")
+def _mark_stale_connection_as_disconnect(context: object) -> None:
+    """Tell SQLAlchemy to invalidate pool connections that asyncpg reports as closed.
+
+    pool_pre_ping catches most stale connections, but a race is possible: the server
+    closes the connection between the ping and the actual query.  When that happens
+    asyncpg raises ConnectionDoesNotExistError (a subclass of InterfaceError).
+    Marking it as a disconnect causes the pool to discard that connection so it is
+    never handed out again.
+    """
+    original = getattr(context, "original_exception", None)
+    if original is not None and type(original).__name__ == "ConnectionDoesNotExistError":
+        context.is_disconnect = True  # type: ignore[union-attr]
+
+
+def _is_stale_connection_error(exc: BaseException) -> bool:
+    return (
+        isinstance(exc, DBAPIError)
+        and exc.__cause__ is not None
+        and type(exc.__cause__).__name__ == "ConnectionDoesNotExistError"
+    )
+
+
 metadata = MetaData(naming_convention=DB_NAMING_CONVENTION)
 
 
@@ -507,6 +533,16 @@ async def fetch_one(
     select_query: Select | Insert | Update,
     params: dict[str, Any] | None = None,
 ) -> dict[str, Any] | None:
+    try:
+        async with engine.begin() as conn:
+            cursor: CursorResult = await conn.execute(select_query, params or {})
+            row = cursor.first()
+            return row._asdict() if row is not None else None
+    except Exception as exc:
+        if not _is_stale_connection_error(exc):
+            raise
+    # Retry once — pool_pre_ping catches most stale connections, but a race can
+    # still return a connection that Postgres closed between the ping and the query.
     async with engine.begin() as conn:
         cursor: CursorResult = await conn.execute(select_query, params or {})
         row = cursor.first()
@@ -517,6 +553,13 @@ async def fetch_all(
     select_query: Select | Insert | Update,
     params: dict[str, Any] | None = None,
 ) -> list[dict[str, Any]]:
+    try:
+        async with engine.begin() as conn:
+            cursor: CursorResult = await conn.execute(select_query, params or {})
+            return [r._asdict() for r in cursor.all()]
+    except Exception as exc:
+        if not _is_stale_connection_error(exc):
+            raise
     async with engine.begin() as conn:
         cursor: CursorResult = await conn.execute(select_query, params or {})
         return [r._asdict() for r in cursor.all()]
@@ -526,5 +569,11 @@ async def execute(
     select_query: Insert | Update,
     params: dict[str, Any] | None = None,
 ) -> CursorResult:
+    try:
+        async with engine.begin() as conn:
+            return await conn.execute(select_query, params or {})
+    except Exception as exc:
+        if not _is_stale_connection_error(exc):
+            raise
     async with engine.begin() as conn:
         return await conn.execute(select_query, params or {})
diff --git a/src/flows/storage/describe_memes.py b/src/flows/storage/describe_memes.py
index 2e0d8214..8e0456d3 100644
--- a/src/flows/storage/describe_memes.py
+++ b/src/flows/storage/describe_memes.py
@@ -38,7 +38,7 @@
 VISION_MODELS = [
     "google/gemma-3-27b-it:free",  # best quality, 140+ languages, 131k context
     "google/gemma-3-12b-it:free",  # good fallback, smaller, 32k context
-    "google/gemma-3-4b-it:free",   # smallest Gemma, lowest rate limits hit
+    "google/gemma-3-4b-it:free",  # smallest Gemma, lowest rate limits hit
     "nvidia/nemotron-nano-12b-v2-vl:free",  # last resort: still listed but unreliable JSON
 ]
 
diff --git a/src/tgbot/handlers/chat/agent/runner.py b/src/tgbot/handlers/chat/agent/runner.py
index e224bd8b..e9187972 100644
--- a/src/tgbot/handlers/chat/agent/runner.py
+++ b/src/tgbot/handlers/chat/agent/runner.py
@@ -110,8 +110,16 @@ async def run_chat_agent(
     tool_calls_count = 0
     for raw in result.raw_responses:
         if hasattr(raw, "usage") and raw.usage:
-            total_prompt_tokens += getattr(raw.usage, "input_tokens", None) or getattr(raw.usage, "prompt_tokens", None) or 0
-            total_completion_tokens += getattr(raw.usage, "output_tokens", None) or getattr(raw.usage, "completion_tokens", None) or 0
+            total_prompt_tokens += (
+                getattr(raw.usage, "input_tokens", None)
+                or getattr(raw.usage, "prompt_tokens", None)
+                or 0
+            )
+            total_completion_tokens += (
+                getattr(raw.usage, "output_tokens", None)
+                or getattr(raw.usage, "completion_tokens", None)
+                or 0
+            )
         for choice in getattr(raw, "choices", []):
             if hasattr(choice, "message") and choice.message and choice.message.tool_calls:
                 tool_calls_count += len(choice.message.tool_calls)
diff --git a/src/tgbot/handlers/chat/feedback.py b/src/tgbot/handlers/chat/feedback.py
index cfd5a858..b1fd7390 100644
--- a/src/tgbot/handlers/chat/feedback.py
+++ b/src/tgbot/handlers/chat/feedback.py
@@ -37,6 +37,8 @@ async def handle_feedback_message(update: Update, context: ContextTypes.DEFAULT_
 
 async def handle_feedback_reply(update: Update, context: ContextTypes.DEFAULT_TYPE):
     reply_text = update.message.text
+    if not reply_text:
+        return  # Admin sent a non-text message (sticker, voice, etc.) — can't forward
     header, _ = update.message.reply_to_message.text.split("\n", 1)
     user_id, message_id = header.split(":")
 
diff --git a/src/tgbot/handlers/stats/wrapped.py b/src/tgbot/handlers/stats/wrapped.py
index 73c00a8f..6e5eba8c 100644
--- a/src/tgbot/handlers/stats/wrapped.py
+++ b/src/tgbot/handlers/stats/wrapped.py
@@ -1008,9 +1008,7 @@ async def generate_wrapped_data(
         # Pick oneliner meme (avoid your_meme)
         oneliner_meme_id = None
         if liked:
-            oneliner_candidates = [
-                m for m in liked[:10] if m["meme_id"] not in global_used_memes
-            ]
+            oneliner_candidates = [m for m in liked[:10] if m["meme_id"] not in global_used_memes]
             if oneliner_candidates:
                 oneliner_meme_id = random.choice(oneliner_candidates)["meme_id"]
             else:
@@ -1187,9 +1185,7 @@ def _build_zodiac_slide(p: dict, is_ru: bool = True) -> str:
     return f"{header}\n\n" f"<b>{html_escape(sign)}</b>\n\n" f"<i>{html_escape(why)}</i>"
 
 
-def _attach_memes_to_absurd(
-    p: dict, liked: list, used_ids: set | None = None
-) -> list:
+def _attach_memes_to_absurd(p: dict, liked: list, used_ids: set | None = None) -> list:
     """Attach meme IDs to each absurd comparison, ensuring no duplicates."""
     comparisons = p.get("absurd_comparisons", [])
     result = []