diff --git a/hermes_cli/web_server.py b/hermes_cli/web_server.py
index 583b83b0410..f489d43bcc6 100644
--- a/hermes_cli/web_server.py
+++ b/hermes_cli/web_server.py
@@ -12210,6 +12210,16 @@ async def pty_ws(ws: WebSocket) -> None:
             # handler never reaches its ``finally`` and the PTY's fds leak.
             # With dashboard auto-reconnect (#52962) every dropped socket then
             # stacks a fresh PTY on top of the orphaned one, exhausting fds.
+            #
+            # Reap the bridge here too (close() is idempotent): on child EOF the
+            # writer loop's ``finally`` is the usual closer, but if the handler
+            # task is cancelled the instant we close the WS, that ``finally``
+            # can be skipped, leaking the PTY. Closing from the EOF path makes
+            # the reap independent of that cancellation race (#54028).
+            try:
+                await asyncio.to_thread(bridge.close)
+            except Exception:
+                pass
             try:
                 await ws.close()
             except Exception:
diff --git a/tests/hermes_cli/test_web_server_pty_reconnect.py b/tests/hermes_cli/test_web_server_pty_reconnect.py
index 5a399407084..a4d12687383 100644
--- a/tests/hermes_cli/test_web_server_pty_reconnect.py
+++ b/tests/hermes_cli/test_web_server_pty_reconnect.py
@@ -165,4 +165,12 @@ def test_child_eof_closes_socket_and_bridge(pty_client, monkeypatch):
             conn.receive_bytes()
 
     assert len(bridges) == 1
+    # bridge.close() runs in the handler's `finally` via asyncio.to_thread,
+    # which can lag the client-side context exit by a tick or two. Poll briefly
+    # instead of asserting immediately so the teardown isn't a race.
+    import time
+
+    deadline = time.monotonic() + 5.0
+    while not bridges[0].closed and time.monotonic() < deadline:
+        time.sleep(0.01)
     assert bridges[0].closed is True