diff --git a/src/py/extra/__init__.py b/src/py/extra/__init__.py
index 4093aa0..4260fcb 100644
--- a/src/py/extra/__init__.py
+++ b/src/py/extra/__init__.py
@@ -1,9 +1,9 @@
-from .http.model import (
+from .http.model import (  # NOQA: F401
     HTTPRequest,
     HTTPResponse,
     HTTPResponseLine,
     HTTPRequestError,
-)  # NOQA: F401
+)
 from .decorators import on, expose, pre, post  # NOQA: F401
 from .server import run  # NOQA: F401
 from .model import Service  # NOQA: F401
diff --git a/src/py/extra/server.py b/src/py/extra/server.py
index e1cd13b..99c6cb6 100644
--- a/src/py/extra/server.py
+++ b/src/py/extra/server.py
@@ -33,7 +33,15 @@ def onException(
     ) -> None:
         e = context.get("exception")
         if e:
-            exception(e)
+            if isinstance(e, asyncio.InvalidStateError):
+                # Known race condition between wait_for timeout and sock_accept
+                # callback. Log as warning rather than exception to avoid alarm.
+                warning(
+                    "AsyncIO InvalidStateError (likely accept race)",
+                    Error=str(e),
+                )
+            else:
+                exception(e)
 
 
 class ServerOptions(NamedTuple):
@@ -392,13 +400,48 @@ async def Serve(
         )
 
         try:
+            # We use a persistent accept task instead of creating a new one each
+            # iteration via wait_for. This avoids the known race condition where
+            # wait_for cancels the sock_accept future but the internal _sock_accept
+            # callback still fires, causing InvalidStateError and leaking the
+            # accepted socket file descriptor.
+            accept_task: asyncio.Task | None = None
             while state.isRunning:
                 if options.condition and not options.condition():
                     break
                 try:
-                    res = await asyncio.wait_for(
-                        loop.sock_accept(server), timeout=options.polling or 1.0
+                    if accept_task is None:
+                        accept_task = loop.create_task(
+                            loop.sock_accept(server)
+                        )
+
+                    # asyncio.wait does NOT cancel the future on timeout,
+                    # unlike asyncio.wait_for. The accept_task persists across
+                    # timeout iterations until a connection actually arrives.
+                    done, _ = await asyncio.wait(
+                        {accept_task},
+                        timeout=options.polling or 1.0,
                     )
+
+                    if accept_task not in done:
+                        # Timeout, no connection yet — loop back and reuse
+                        # the same accept_task.
+                        continue
+
+                    # Connection arrived (or error). Consume the result and
+                    # clear accept_task so a new one is created next iteration.
+                    try:
+                        res = accept_task.result()
+                    except OSError as e:
+                        if e.errno == 24:
+                            # Too many open files — backpressure
+                            await asyncio.sleep(0.1)
+                        else:
+                            exception(e)
+                        accept_task = None
+                        continue
+
+                    accept_task = None
                     if res is None:
                         continue
                     else:
@@ -409,17 +452,22 @@ async def Serve(
                     )
                     tasks.add(task)
                     task.add_done_callback(tasks.discard)
-                except asyncio.TimeoutError:
+                except asyncio.InvalidStateError:
+                    # Defensive: should not happen with the new pattern,
+                    # but guard against it to prevent server crash.
+                    warning("InvalidStateError in accept loop (recovered)")
+                    accept_task = None
                     continue
-                except OSError as e:
-                    # This can be: [OSError] [Errno 24] Too many open files
-                    if e.errno == 24:
-                        # Implement backpressure or wait mechanism here
-                        await asyncio.sleep(0.1)  # Short delay before retrying
-                    else:
-                        exception(e)
+                except asyncio.CancelledError:
+                    break
 
         finally:
+            if accept_task and not accept_task.done():
+                accept_task.cancel()
+                try:
+                    await accept_task
+                except (asyncio.CancelledError, Exception):
+                    pass
             server.close()
             for task in tasks:
                 task.cancel()