Keep SDK loopback requests direct and disable shared cookies

backend/services/llm_transport.py

@@ -4,9 +4,18 @@ from __future__ import annotations
 import os
 import atexit
 from functools import lru_cache
+from http.cookiejar import CookieJar, DefaultCookiePolicy
 from types import SimpleNamespace
 
 
+class _NoCookies(DefaultCookiePolicy):
+    def set_ok(self, cookie, request):
+        return False
+
+    def return_ok(self, cookie, request):
+        return False
+
+
 def _check_sdk_request(request):
     """Validate the effective URL after SDK environment/provider resolution."""
     from ipaddress import ip_address
@@ -28,7 +37,12 @@ def _sdk_http_client():
     # A shared thread-safe pool also keeps streaming responses alive after
     # completion() returns. It carries no provider credentials of its own.
     import httpx
-    client = httpx.Client(follow_redirects=False, event_hooks={"request": [_check_sdk_request]})
+    client = httpx.Client(
+        follow_redirects=False,
+        trust_env=False,
+        cookies=CookieJar(policy=_NoCookies()),
+        event_hooks={"request": [_check_sdk_request]},
+    )
     atexit.register(client.close)
     return client
 

docs/electron-llm-providers.md

@@ -150,3 +150,5 @@ LM Studio automatic discovery accepts only loaded `llm`/`vlm` entries from its n
 SDK transport validates the final request URL, including environment-derived
 OpenAI endpoints, and disables redirects for OpenAI-compatible and native
 Anthropic, Bedrock, and Vertex requests.
+The shared HTTP client ignores environment proxies so loopback prompts stay local,
+and never stores response cookies between provider requests.

tests/test_llm_transports.py

@@ -159,3 +159,33 @@ def test_anthropic_does_not_follow_credentialed_redirect(registry):
         assert seen == ["/v1/messages"]
     finally:
         server.shutdown(); server.server_close(); thread.join(timeout=2)
+
+
+@pytest.mark.parametrize("proxy_enabled", [False, True])
+def test_sdk_pool_keeps_loopback_direct_and_does_not_persist_cookies(monkeypatch, proxy_enabled):
+    from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
+    from threading import Thread
+    from services.llm_transport import _sdk_http_client
+    seen = []
+    class Handler(BaseHTTPRequestHandler):
+        def log_message(self, *args): pass
+        def do_GET(self):
+            seen.append(self.headers.get("Cookie"))
+            self.send_response(200)
+            self.send_header("Set-Cookie", "session=private; Path=/")
+            self.send_header("Content-Length", "0")
+            self.end_headers()
+    for key in ("HTTP_PROXY", "HTTPS_PROXY", "ALL_PROXY", "http_proxy", "https_proxy", "all_proxy"):
+        monkeypatch.setenv(key, "http://127.0.0.1:1" if proxy_enabled else "")
+    monkeypatch.setenv("NO_PROXY", "")
+    monkeypatch.setenv("no_proxy", "")
+    server = ThreadingHTTPServer(("127.0.0.1", 0), Handler)
+    thread = Thread(target=server.serve_forever, daemon=True); thread.start()
+    try:
+        with _sdk_http_client.__wrapped__() as client:
+            for _ in range(2):
+                assert client.get(f"http://127.0.0.1:{server.server_port}/", timeout=2).status_code == 200
+            assert not list(client.cookies.jar)
+            assert seen == [None, None]
+    finally:
+        server.shutdown(); server.server_close(); thread.join(timeout=2)