From c2a6c9ca956bad5f027ac1f5989d8513274a1513 Mon Sep 17 00:00:00 2001
From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Date: Wed, 8 Jul 2026 15:54:52 +0000
Subject: [PATCH] fix(router): honor AllowedFails=0 policy for
 cooldown-on-first-failure

---
 litellm/router_utils/cooldown_handlers.py     |  8 ++---
 ..._health_check_allowed_fails_integration.py | 30 +++++++++++++++++++
 2 files changed, 34 insertions(+), 4 deletions(-)

diff --git a/litellm/router_utils/cooldown_handlers.py b/litellm/router_utils/cooldown_handlers.py
index 2bc2ed998ca..ec14363d594 100644
--- a/litellm/router_utils/cooldown_handlers.py
+++ b/litellm/router_utils/cooldown_handlers.py
@@ -372,11 +372,11 @@ def should_cooldown_based_on_allowed_fails_policy(
     - True if fails exceed the allowed limit (should cooldown)
     - False if fails are within the allowed limit (should not cooldown)
     """
+    allowed_fails_from_policy = litellm_router_instance.get_allowed_fails_from_policy(
+        exception=original_exception,
+    )
     allowed_fails = (
-        litellm_router_instance.get_allowed_fails_from_policy(
-            exception=original_exception,
-        )
-        or litellm_router_instance.allowed_fails
+        allowed_fails_from_policy if allowed_fails_from_policy is not None else litellm_router_instance.allowed_fails
     )
     cooldown_time = litellm_router_instance.cooldown_time or DEFAULT_COOLDOWN_TIME_SECONDS
 
diff --git a/tests/test_litellm/router_utils/test_health_check_allowed_fails_integration.py b/tests/test_litellm/router_utils/test_health_check_allowed_fails_integration.py
index 64239f33966..934420919e5 100644
--- a/tests/test_litellm/router_utils/test_health_check_allowed_fails_integration.py
+++ b/tests/test_litellm/router_utils/test_health_check_allowed_fails_integration.py
@@ -214,6 +214,36 @@ class TestHealthCheckCooldownIntegration:
         )
         assert result is True
 
+    def test_allowed_fails_policy_zero_cooldowns_on_first_failure(self):
+        """Regression for #32425.
+
+        AllowedFails=0 means cooldown on the very first failure. The value 0 was
+        previously discarded by an `or` fallback (`policy_value or self.allowed_fails`),
+        so the generic allowed_fails default was used instead and the first failure
+        did not cool the deployment down.
+        """
+        from litellm.router_utils.cooldown_handlers import (
+            should_cooldown_based_on_allowed_fails_policy,
+        )
+
+        router = Router(
+            model_list=[_make_model("deploy-1"), _make_model("deploy-2", "gpt-5")],
+            allowed_fails_policy=AllowedFailsPolicy(RateLimitErrorAllowedFails=0),
+            # A permissive generic threshold that must NOT be used when the policy says 0.
+            allowed_fails=100,
+        )
+
+        rate_limit_exc = litellm.RateLimitError(
+            message="rate limited", model="gpt-4", llm_provider="openai"
+        )
+
+        result = should_cooldown_based_on_allowed_fails_policy(
+            litellm_router_instance=router,
+            deployment="deploy-1",
+            original_exception=rate_limit_exc,
+        )
+        assert result is True
+
     def test_health_check_failure_falls_back_to_allowed_fails(self):
         """When policy has no matching field, fall back to generic allowed_fails."""
         from litellm.router_utils.cooldown_handlers import (