fix(health-check-routing): properly isolate exceptions from health response

Return exceptions_by_model_id as a separate third value from
_perform_health_check / perform_health_check so exception objects
(which contain non-JSON-serializable httpx URL types) never appear
in the endpoint dicts that get serialized by the /health response.

Callers updated: _health_endpoints.py, shared_health_check_manager.py,
proxy_server.py background loop. All use the exceptions dict only for
cooldown integration, not for display.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

litellm/proxy/health_check.py

@@ -211,6 +211,10 @@ async def _perform_health_check(
 
     healthy_endpoints = []
     unhealthy_endpoints = []
+    # Exceptions keyed by model_id; returned separately so callers can use
+    # them for cooldown integration without risking JSON-serialization errors
+    # in the /health response.
+    exceptions_by_model_id: dict = {}
 
     for is_healthy, model in zip(results, model_list):
         litellm_params = model["litellm_params"]
@@ -225,14 +229,18 @@ async def _perform_health_check(
             cleaned = _clean_endpoint_data({**litellm_params, **is_healthy}, details)
             if _model_id:
                 cleaned["model_id"] = _model_id
+                if "exception" in is_healthy:
+                    exceptions_by_model_id[_model_id] = is_healthy["exception"]
             unhealthy_endpoints.append(cleaned)
         else:
             cleaned = _clean_endpoint_data(litellm_params, details)
             if _model_id:
                 cleaned["model_id"] = _model_id
+                if isinstance(is_healthy, Exception):
+                    exceptions_by_model_id[_model_id] = is_healthy
             unhealthy_endpoints.append(cleaned)
 
-    return healthy_endpoints, unhealthy_endpoints
+    return healthy_endpoints, unhealthy_endpoints, exceptions_by_model_id
 
 
 def build_deployment_health_states(
@@ -413,7 +421,11 @@ async def perform_health_check(
         )
 
     try:
-        healthy_endpoints, unhealthy_endpoints = await _perform_health_check(
+        (
+            healthy_endpoints,
+            unhealthy_endpoints,
+            exceptions_by_model_id,
+        ) = await _perform_health_check(
             model_list,
             details,
             max_concurrency=max_concurrency,
@@ -445,4 +457,4 @@ async def perform_health_check(
             _rss_mb_for_log(),
         )
 
-    return healthy_endpoints, unhealthy_endpoints
+    return healthy_endpoints, unhealthy_endpoints, exceptions_by_model_id

litellm/proxy/health_check_utils/shared_health_check_manager.py

@@ -231,7 +231,11 @@ class SharedHealthCheckManager:
                     len(model_list),
                 )
 
-                healthy_endpoints, unhealthy_endpoints = await perform_health_check(
+                (
+                    healthy_endpoints,
+                    unhealthy_endpoints,
+                    exceptions_by_model_id,
+                ) = await perform_health_check(
                     model_list=model_list,
                     details=details,
                     max_concurrency=max_concurrency,
@@ -242,7 +246,7 @@ class SharedHealthCheckManager:
                     healthy_endpoints, unhealthy_endpoints
                 )
 
-                return healthy_endpoints, unhealthy_endpoints
+                return healthy_endpoints, unhealthy_endpoints, exceptions_by_model_id
 
             finally:
                 # Always release the lock

litellm/proxy/health_endpoints/_health_endpoints.py

@@ -771,7 +771,7 @@ async def _perform_health_check_and_save(
     max_concurrency=None,
 ):
     """Helper function to perform health check and save results to database"""
-    healthy_endpoints, unhealthy_endpoints = await perform_health_check(
+    healthy_endpoints, unhealthy_endpoints, _ = await perform_health_check(
         model_list=model_list,
         cli_model=cli_model,
         model=target_model,

litellm/proxy/proxy_server.py

@@ -2287,6 +2287,7 @@ async def _run_background_health_check():
                 (
                     healthy_endpoints,
                     unhealthy_endpoints,
+                    _exceptions_by_model_id,
                 ) = await shared_health_manager.perform_shared_health_check(
                     model_list=_llm_model_list,
                     details=details_bool,
@@ -2300,6 +2301,7 @@ async def _run_background_health_check():
                 (
                     healthy_endpoints,
                     unhealthy_endpoints,
+                    _exceptions_by_model_id,
                 ) = await _run_direct_health_check_with_instrumentation(
                     _llm_model_list,
                     health_check_details,
@@ -2310,6 +2312,7 @@ async def _run_background_health_check():
             (
                 healthy_endpoints,
                 unhealthy_endpoints,
+                _exceptions_by_model_id,
             ) = await _run_direct_health_check_with_instrumentation(
                 _llm_model_list,
                 health_check_details,
@@ -2352,7 +2355,9 @@ async def _run_background_health_check():
         )
 
         # Write health state to router cache for health-check-driven routing
-        _write_health_state_to_router_cache(healthy_endpoints, unhealthy_endpoints)
+        _write_health_state_to_router_cache(
+            healthy_endpoints, unhealthy_endpoints, _exceptions_by_model_id
+        )
 
         await asyncio.sleep(health_check_interval)