fix(proxy): only release the legacy limiter's slot from an admitted Responses WebSocket (#46822)

The shared disconnect release dispatched to the legacy limiter, which decrements unconditionally, so a rejected realtime attempt freed a slot it never took. The generic release is v3-only again and the Responses WebSocket close releases through its own method

Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

litellm/proxy/response_api_endpoints/endpoints.py

@@ -1906,7 +1906,7 @@ async def _responses_websocket_session(
         )
         await websocket.close(code=1011, reason="Internal server error")
     finally:
-        await proxy_logging_obj.arelease_max_parallel_requests_on_disconnect(user_api_key_dict)
+        await proxy_logging_obj.arelease_responses_websocket_slot(user_api_key_dict)
 
 
 @router.websocket("/v1/responses")

litellm/proxy/utils.py

@@ -4405,12 +4405,20 @@ class ProxyLogging:
         double-decrement under the limiter's in-memory fallback.
         """
         limiter: Final = self.get_proxy_hook("parallel_request_limiter")
-        if not isinstance(limiter, (PROXY_MaxParallelRequestsHandler_v3, _PROXY_MaxParallelRequestsHandler)):
+        if not isinstance(limiter, PROXY_MaxParallelRequestsHandler_v3):
             return
         await limiter.async_release_max_parallel_requests_on_disconnect(user_api_key_dict)
 
     _arelease_max_parallel_requests_on_disconnect = arelease_max_parallel_requests_on_disconnect
 
+    async def arelease_responses_websocket_slot(self, user_api_key_dict: UserAPIKeyAuth) -> None:
+        """Release the slot a Responses WebSocket took at admission. Only call this once that admission succeeded:
+        the legacy limiter keeps no record of its acquisitions and decrements unconditionally."""
+        limiter: Final = self.get_proxy_hook("parallel_request_limiter")
+        if not isinstance(limiter, (PROXY_MaxParallelRequestsHandler_v3, _PROXY_MaxParallelRequestsHandler)):
+            return
+        await limiter.async_release_max_parallel_requests_on_disconnect(user_api_key_dict)
+
     async def enforce_mcp_server_rate_limits(
         self,
         user_api_key_dict: UserAPIKeyAuth | None,

tests/unit/proxy/hooks/test_parallel_request_limiter.py

@@ -15,7 +15,7 @@ from litellm.proxy.common_utils.proxy_rate_limit_error import ProxyRateLimitErro
 from litellm.proxy.hooks.parallel_request_limiter import (
     PROXY_MaxParallelRequestsHandler,
 )
-from litellm.proxy.utils import InternalUsageCache, hash_token
+from litellm.proxy.utils import InternalUsageCache, ProxyLogging, hash_token
 from litellm.types.utils import CallTypes, EmbeddingResponse, ModelResponse, TextCompletionResponse, Usage
 
 FROZEN_INSTANT: Final = datetime(2026, 1, 31, 23, 59, 30)
@@ -353,3 +353,26 @@ async def test_websocket_turn_callbacks_keep_the_sockets_slot_until_the_disconne
     await handler.async_release_max_parallel_requests_on_disconnect(UserAPIKeyAuth(api_key=api_key))
     after_disconnect = await handler.internal_usage_cache.async_get_cache(key=bucket, litellm_parent_otel_span=None)
     assert after_disconnect["current_requests"] == 0, after_disconnect
+
+
+@pytest.mark.asyncio
+async def test_a_rejected_realtime_attempt_does_not_free_a_slot_it_never_took():
+    api_key = hash_token("sk-at-limit")
+    handler = PROXY_MaxParallelRequestsHandler(internal_usage_cache=InternalUsageCache(DualCache()), clock=_frozen_clock)
+    bucket = f"{api_key}::{FROZEN_INSTANT.strftime('%Y-%m-%d-%H-%M')}::request_count"
+    await handler.internal_usage_cache.async_set_cache(
+        key=bucket,
+        value={"current_requests": 1, "current_tpm": 0, "current_rpm": 1},
+        litellm_parent_otel_span=None,
+    )
+    proxy_logging_obj = ProxyLogging(user_api_key_cache=DualCache())
+    proxy_logging_obj.proxy_hook_mapping["parallel_request_limiter"] = handler
+    auth = UserAPIKeyAuth(api_key=api_key, max_parallel_requests=1)
+
+    await proxy_logging_obj.arelease_max_parallel_requests_on_disconnect(auth)
+    after_rejection = await handler.internal_usage_cache.async_get_cache(key=bucket, litellm_parent_otel_span=None)
+    assert after_rejection["current_requests"] == 1, after_rejection
+
+    await proxy_logging_obj.arelease_responses_websocket_slot(auth)
+    after_socket_close = await handler.internal_usage_cache.async_get_cache(key=bucket, litellm_parent_otel_span=None)
+    assert after_socket_close["current_requests"] == 0, after_socket_close

tests/unit/proxy/response_api_endpoints/test_endpoints.py

@@ -1952,7 +1952,7 @@ class TestResponsesWSFirstFrameModelAuth:
 
         proxy_logging_obj = MagicMock()
         proxy_logging_obj.post_call_failure_hook = AsyncMock(return_value=None)
-        proxy_logging_obj.arelease_max_parallel_requests_on_disconnect = AsyncMock(return_value=None)
+        proxy_logging_obj.arelease_responses_websocket_slot = AsyncMock(return_value=None)
         user_api_key_dict = MagicMock()
 
         with (
@@ -2017,7 +2017,7 @@ class TestResponsesWSFirstFrameModelAuth:
         )
         proxy_logging_obj = MagicMock()
         proxy_logging_obj.post_call_failure_hook = AsyncMock(return_value=None)
-        proxy_logging_obj.arelease_max_parallel_requests_on_disconnect = AsyncMock(return_value=None)
+        proxy_logging_obj.arelease_responses_websocket_slot = AsyncMock(return_value=None)
         user_api_key_dict = MagicMock()
 
         with (