simplify Athena runtime and centralize MCP management

router/ai_profile_router.py

@@ -62,6 +62,7 @@ import logging
 import os
 import queue
 import re
+import select
 import socket
 import subprocess
 import sys
@@ -2098,6 +2099,21 @@ class Handler(BaseHTTPRequestHandler):
                 chunk = resp.read1(16384)
                 if not chunk:
                     break
+                # A closed downstream socket may still accept one small write
+                # into the kernel buffer. Detect the peer FIN before writing so
+                # slow upstream streams are cancelled promptly and do not keep
+                # an inference lease occupied until generation completes.
+                try:
+                    readable, _, _ = select.select([self.connection], [], [], 0)
+                    if readable:
+                        flags = socket.MSG_PEEK | getattr(socket, "MSG_DONTWAIT", 0)
+                        if self.connection.recv(1, flags) == b"":
+                            log.info("Client trennte Streaming-Verbindung; Upstream wird abgebrochen")
+                            break
+                except (BlockingIOError, InterruptedError):
+                    pass
+                except OSError:
+                    break
                 self.wfile.write(chunk)
                 self.wfile.flush()
         except (OSError, http.client.HTTPException) as e: